1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 88, 89, 91-97, 1998 Free Software Foundation, Inc.
4 This file is part of GNU CC.
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
42 #include "insn-config.h"
43 #include "insn-flags.h"
45 #include "hard-reg-set.h"
53 /* Vector mapping INSN_UIDs to luids.
54 The luids are like uids but increase monotonically always.
55 We use them to see whether a jump comes from outside a given loop. */
59 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
60 number the insn is contained in. */
64 /* 1 + largest uid of any insn. */
68 /* 1 + luid of last insn. */
72 /* Number of loops detected in current function. Used as index to the
75 static int max_loop_num;
77 /* Indexed by loop number, contains the first and last insn of each loop. */
79 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
81 /* For each loop, gives the containing loop number, -1 if none. */
85 #ifdef HAVE_decrement_and_branch_on_count
86 /* Records whether resource in use by inner loop. */
88 int *loop_used_count_register;
89 #endif /* HAVE_decrement_and_branch_on_count */
91 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
92 really a loop (an insn outside the loop branches into it). */
94 static char *loop_invalid;
96 /* Indexed by loop number, links together all LABEL_REFs which refer to
97 code labels outside the loop. Used by routines that need to know all
98 loop exits, such as final_biv_value and final_giv_value.
100 This does not include loop exits due to return instructions. This is
101 because all bivs and givs are pseudos, and hence must be dead after a
102 return, so the presense of a return does not affect any of the
103 optimizations that use this info. It is simpler to just not include return
104 instructions on this list. */
106 rtx *loop_number_exit_labels;
108 /* Indexed by loop number, counts the number of LABEL_REFs on
109 loop_number_exit_labels for this loop and all loops nested inside it. */
111 int *loop_number_exit_count;
113 /* Nonzero if there is a subroutine call in the current loop. */
115 static int loop_has_call;
117 /* Nonzero if there is a volatile memory reference in the current
120 static int loop_has_volatile;
122 /* Nonzero if there is a tablejump in the current loop. */
124 static int loop_has_tablejump;
126 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
127 current loop. A continue statement will generate a branch to
128 NEXT_INSN (loop_continue). */
130 static rtx loop_continue;
132 /* Indexed by register number, contains the number of times the reg
133 is set during the loop being scanned.
134 During code motion, a negative value indicates a reg that has been
135 made a candidate; in particular -2 means that it is an candidate that
136 we know is equal to a constant and -1 means that it is an candidate
137 not known equal to a constant.
138 After code motion, regs moved have 0 (which is accurate now)
139 while the failed candidates have the original number of times set.
141 Therefore, at all times, == 0 indicates an invariant register;
142 < 0 a conditionally invariant one. */
144 static varray_type set_in_loop;
146 /* Original value of set_in_loop; same except that this value
147 is not set negative for a reg whose sets have been made candidates
148 and not set to 0 for a reg that is moved. */
150 static varray_type n_times_set;
152 /* Index by register number, 1 indicates that the register
153 cannot be moved or strength reduced. */
155 static varray_type may_not_optimize;
157 /* Nonzero means reg N has already been moved out of one loop.
158 This reduces the desire to move it out of another. */
160 static char *moved_once;
162 /* Array of MEMs that are stored in this loop. If there are too many to fit
163 here, we just turn on unknown_address_altered. */
165 #define NUM_STORES 30
166 static rtx loop_store_mems[NUM_STORES];
168 /* Index of first available slot in above array. */
169 static int loop_store_mems_idx;
171 typedef struct loop_mem_info {
172 rtx mem; /* The MEM itself. */
173 rtx reg; /* Corresponding pseudo, if any. */
174 int optimize; /* Nonzero if we can optimize access to this MEM. */
177 /* Array of MEMs that are used (read or written) in this loop, but
178 cannot be aliased by anything in this loop, except perhaps
179 themselves. In other words, if loop_mems[i] is altered during the
180 loop, it is altered by an expression that is rtx_equal_p to it. */
182 static loop_mem_info *loop_mems;
184 /* The index of the next available slot in LOOP_MEMS. */
186 static int loop_mems_idx;
188 /* The number of elements allocated in LOOP_MEMs. */
190 static int loop_mems_allocated;
192 /* Nonzero if we don't know what MEMs were changed in the current loop.
193 This happens if the loop contains a call (in which case `loop_has_call'
194 will also be set) or if we store into more than NUM_STORES MEMs. */
196 static int unknown_address_altered;
198 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
199 static int num_movables;
201 /* Count of memory write instructions discovered in the loop. */
202 static int num_mem_sets;
204 /* Number of loops contained within the current one, including itself. */
205 static int loops_enclosed;
207 /* Bound on pseudo register number before loop optimization.
208 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
209 int max_reg_before_loop;
211 /* This obstack is used in product_cheap_p to allocate its rtl. It
212 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
213 If we used the same obstack that it did, we would be deallocating
216 static struct obstack temp_obstack;
218 /* This is where the pointer to the obstack being used for RTL is stored. */
220 extern struct obstack *rtl_obstack;
222 #define obstack_chunk_alloc xmalloc
223 #define obstack_chunk_free free
225 /* During the analysis of a loop, a chain of `struct movable's
226 is made to record all the movable insns found.
227 Then the entire chain can be scanned to decide which to move. */
231 rtx insn; /* A movable insn */
232 rtx set_src; /* The expression this reg is set from. */
233 rtx set_dest; /* The destination of this SET. */
234 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
235 of any registers used within the LIBCALL. */
236 int consec; /* Number of consecutive following insns
237 that must be moved with this one. */
238 int regno; /* The register it sets */
239 short lifetime; /* lifetime of that register;
240 may be adjusted when matching movables
241 that load the same value are found. */
242 short savings; /* Number of insns we can move for this reg,
243 including other movables that force this
244 or match this one. */
245 unsigned int cond : 1; /* 1 if only conditionally movable */
246 unsigned int force : 1; /* 1 means MUST move this insn */
247 unsigned int global : 1; /* 1 means reg is live outside this loop */
248 /* If PARTIAL is 1, GLOBAL means something different:
249 that the reg is live outside the range from where it is set
250 to the following label. */
251 unsigned int done : 1; /* 1 inhibits further processing of this */
253 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
254 In particular, moving it does not make it
256 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
257 load SRC, rather than copying INSN. */
258 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
259 first insn of a consecutive sets group. */
260 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
261 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
262 that we should avoid changing when clearing
263 the rest of the reg. */
264 struct movable *match; /* First entry for same value */
265 struct movable *forces; /* An insn that must be moved if this is */
266 struct movable *next;
269 static struct movable *the_movables;
271 FILE *loop_dump_stream;
273 /* Forward declarations. */
275 static void find_and_verify_loops PROTO((rtx));
276 static void mark_loop_jump PROTO((rtx, int));
277 static void prescan_loop PROTO((rtx, rtx));
278 static int reg_in_basic_block_p PROTO((rtx, rtx));
279 static int consec_sets_invariant_p PROTO((rtx, int, rtx));
280 static rtx libcall_other_reg PROTO((rtx, rtx));
281 static int labels_in_range_p PROTO((rtx, int));
282 static void count_one_set PROTO((rtx, rtx, varray_type, rtx *));
284 static void count_loop_regs_set PROTO((rtx, rtx, varray_type, varray_type,
286 static void note_addr_stored PROTO((rtx, rtx));
287 static int loop_reg_used_before_p PROTO((rtx, rtx, rtx, rtx, rtx));
288 static void scan_loop PROTO((rtx, rtx, int, int));
290 static void replace_call_address PROTO((rtx, rtx, rtx));
292 static rtx skip_consec_insns PROTO((rtx, int));
293 static int libcall_benefit PROTO((rtx));
294 static void ignore_some_movables PROTO((struct movable *));
295 static void force_movables PROTO((struct movable *));
296 static void combine_movables PROTO((struct movable *, int));
297 static int regs_match_p PROTO((rtx, rtx, struct movable *));
298 static int rtx_equal_for_loop_p PROTO((rtx, rtx, struct movable *));
299 static void add_label_notes PROTO((rtx, rtx));
300 static void move_movables PROTO((struct movable *, int, int, rtx, rtx, int));
301 static int count_nonfixed_reads PROTO((rtx));
302 static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx, int, int));
303 static void find_single_use_in_loop PROTO((rtx, rtx, varray_type));
304 static int valid_initial_value_p PROTO((rtx, rtx, int, rtx));
305 static void find_mem_givs PROTO((rtx, rtx, int, rtx, rtx));
306 static void record_biv PROTO((struct induction *, rtx, rtx, rtx, rtx, int, int));
307 static void check_final_value PROTO((struct induction *, rtx, rtx,
308 unsigned HOST_WIDE_INT));
309 static void record_giv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx, int, enum g_types, int, rtx *, rtx, rtx));
310 static void update_giv_derive PROTO((rtx));
311 static int basic_induction_var PROTO((rtx, enum machine_mode, rtx, rtx, rtx *, rtx *));
312 static rtx simplify_giv_expr PROTO((rtx, int *));
313 static int general_induction_var PROTO((rtx, rtx *, rtx *, rtx *, int, int *));
314 static int consec_sets_giv PROTO((int, rtx, rtx, rtx, rtx *, rtx *, rtx *));
315 static int check_dbra_loop PROTO((rtx, int, rtx, struct loop_info *));
316 static rtx express_from_1 PROTO((rtx, rtx, rtx));
317 static rtx express_from PROTO((struct induction *, struct induction *));
318 static rtx combine_givs_p PROTO((struct induction *, struct induction *));
319 static void combine_givs PROTO((struct iv_class *));
320 static int product_cheap_p PROTO((rtx, rtx));
321 static int maybe_eliminate_biv PROTO((struct iv_class *, rtx, rtx, int, int, int));
322 static int maybe_eliminate_biv_1 PROTO((rtx, rtx, struct iv_class *, int, rtx));
323 static int last_use_this_basic_block PROTO((rtx, rtx));
324 static void record_initial PROTO((rtx, rtx));
325 static void update_reg_last_use PROTO((rtx, rtx));
326 static rtx next_insn_in_loop PROTO((rtx, rtx, rtx, rtx));
327 static void load_mems_and_recount_loop_regs_set PROTO((rtx, rtx, rtx,
330 static void load_mems PROTO((rtx, rtx, rtx, rtx));
331 static int insert_loop_mem PROTO((rtx *, void *));
332 static int replace_loop_mem PROTO((rtx *, void *));
333 static int replace_label PROTO((rtx *, void *));
335 typedef struct rtx_and_int {
340 typedef struct rtx_pair {
345 /* Nonzero iff INSN is between START and END, inclusive. */
346 #define INSN_IN_RANGE_P(INSN, START, END) \
347 (INSN_UID (INSN) < max_uid_for_loop \
348 && INSN_LUID (INSN) >= INSN_LUID (START) \
349 && INSN_LUID (INSN) <= INSN_LUID (END))
351 #ifdef HAVE_decrement_and_branch_on_count
352 /* Test whether BCT applicable and safe. */
353 static void insert_bct PROTO((rtx, rtx, struct loop_info *));
355 /* Auxiliary function that inserts the BCT pattern into the loop. */
356 static void instrument_loop_bct PROTO((rtx, rtx, rtx));
357 #endif /* HAVE_decrement_and_branch_on_count */
359 /* Indirect_jump_in_function is computed once per function. */
360 int indirect_jump_in_function = 0;
361 static int indirect_jump_in_function_p PROTO((rtx));
364 /* Relative gain of eliminating various kinds of operations. */
367 static int shift_cost;
368 static int mult_cost;
371 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
372 copy the value of the strength reduced giv to its original register. */
373 static int copy_cost;
375 /* Cost of using a register, to normalize the benefits of a giv. */
376 static int reg_address_cost;
382 char *free_point = (char *) oballoc (1);
383 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
385 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
388 reg_address_cost = ADDRESS_COST (reg);
390 reg_address_cost = rtx_cost (reg, MEM);
393 /* We multiply by 2 to reconcile the difference in scale between
394 these two ways of computing costs. Otherwise the cost of a copy
395 will be far less than the cost of an add. */
399 /* Free the objects we just allocated. */
402 /* Initialize the obstack used for rtl in product_cheap_p. */
403 gcc_obstack_init (&temp_obstack);
406 /* Entry point of this file. Perform loop optimization
407 on the current function. F is the first insn of the function
408 and DUMPFILE is a stream for output of a trace of actions taken
409 (or 0 if none should be output). */
412 loop_optimize (f, dumpfile, unroll_p, bct_p)
413 /* f is the first instruction of a chain of insns for one function */
422 loop_dump_stream = dumpfile;
424 init_recog_no_volatile ();
426 max_reg_before_loop = max_reg_num ();
428 moved_once = (char *) alloca (max_reg_before_loop);
429 bzero (moved_once, max_reg_before_loop);
433 /* Count the number of loops. */
436 for (insn = f; insn; insn = NEXT_INSN (insn))
438 if (GET_CODE (insn) == NOTE
439 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
443 /* Don't waste time if no loops. */
444 if (max_loop_num == 0)
447 /* Get size to use for tables indexed by uids.
448 Leave some space for labels allocated by find_and_verify_loops. */
449 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
451 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
452 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
454 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
455 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
457 /* Allocate tables for recording each loop. We set each entry, so they need
459 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
460 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
461 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
462 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
463 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
464 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
466 #ifdef HAVE_decrement_and_branch_on_count
467 /* Allocate for BCT optimization */
468 loop_used_count_register = (int *) alloca (max_loop_num * sizeof (int));
469 bzero ((char *) loop_used_count_register, max_loop_num * sizeof (int));
470 #endif /* HAVE_decrement_and_branch_on_count */
472 /* Find and process each loop.
473 First, find them, and record them in order of their beginnings. */
474 find_and_verify_loops (f);
476 /* Now find all register lifetimes. This must be done after
477 find_and_verify_loops, because it might reorder the insns in the
479 reg_scan (f, max_reg_num (), 1);
481 /* This must occur after reg_scan so that registers created by gcse
482 will have entries in the register tables.
484 We could have added a call to reg_scan after gcse_main in toplev.c,
485 but moving this call to init_alias_analysis is more efficient. */
486 init_alias_analysis ();
488 /* See if we went too far. */
489 if (get_max_uid () > max_uid_for_loop)
491 /* Now reset it to the actual size we need. See above. */
492 max_uid_for_loop = get_max_uid () + 1;
494 /* Compute the mapping from uids to luids.
495 LUIDs are numbers assigned to insns, like uids,
496 except that luids increase monotonically through the code.
497 Don't assign luids to line-number NOTEs, so that the distance in luids
498 between two insns is not affected by -g. */
500 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
503 if (GET_CODE (insn) != NOTE
504 || NOTE_LINE_NUMBER (insn) <= 0)
505 uid_luid[INSN_UID (insn)] = ++i;
507 /* Give a line number note the same luid as preceding insn. */
508 uid_luid[INSN_UID (insn)] = i;
513 /* Don't leave gaps in uid_luid for insns that have been
514 deleted. It is possible that the first or last insn
515 using some register has been deleted by cross-jumping.
516 Make sure that uid_luid for that former insn's uid
517 points to the general area where that insn used to be. */
518 for (i = 0; i < max_uid_for_loop; i++)
520 uid_luid[0] = uid_luid[i];
521 if (uid_luid[0] != 0)
524 for (i = 0; i < max_uid_for_loop; i++)
525 if (uid_luid[i] == 0)
526 uid_luid[i] = uid_luid[i - 1];
528 /* Create a mapping from loops to BLOCK tree nodes. */
529 if (unroll_p && write_symbols != NO_DEBUG)
530 find_loop_tree_blocks ();
532 /* Determine if the function has indirect jump. On some systems
533 this prevents low overhead loop instructions from being used. */
534 indirect_jump_in_function = indirect_jump_in_function_p (f);
536 /* Now scan the loops, last ones first, since this means inner ones are done
537 before outer ones. */
538 for (i = max_loop_num-1; i >= 0; i--)
539 if (! loop_invalid[i] && loop_number_loop_ends[i])
540 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
543 /* If debugging and unrolling loops, we must replicate the tree nodes
544 corresponding to the blocks inside the loop, so that the original one
545 to one mapping will remain. */
546 if (unroll_p && write_symbols != NO_DEBUG)
547 unroll_block_trees ();
549 end_alias_analysis ();
552 /* Returns the next insn, in execution order, after INSN. START and
553 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
554 respectively. LOOP_TOP, if non-NULL, is the top of the loop in the
555 insn-stream; it is used with loops that are entered near the
559 next_insn_in_loop (insn, start, end, loop_top)
565 insn = NEXT_INSN (insn);
570 /* Go to the top of the loop, and continue there. */
584 /* Optimize one loop whose start is LOOP_START and end is END.
585 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
586 NOTE_INSN_LOOP_END. */
588 /* ??? Could also move memory writes out of loops if the destination address
589 is invariant, the source is invariant, the memory write is not volatile,
590 and if we can prove that no read inside the loop can read this address
591 before the write occurs. If there is a read of this address after the
592 write, then we can also mark the memory read as invariant. */
595 scan_loop (loop_start, end, unroll_p, bct_p)
601 /* 1 if we are scanning insns that could be executed zero times. */
603 /* 1 if we are scanning insns that might never be executed
604 due to a subroutine call which might exit before they are reached. */
606 /* For a rotated loop that is entered near the bottom,
607 this is the label at the top. Otherwise it is zero. */
609 /* Jump insn that enters the loop, or 0 if control drops in. */
610 rtx loop_entry_jump = 0;
611 /* Place in the loop where control enters. */
613 /* Number of insns in the loop. */
618 /* The SET from an insn, if it is the only SET in the insn. */
620 /* Chain describing insns movable in current loop. */
621 struct movable *movables = 0;
622 /* Last element in `movables' -- so we can add elements at the end. */
623 struct movable *last_movable = 0;
624 /* Ratio of extra register life span we can justify
625 for saving an instruction. More if loop doesn't call subroutines
626 since in that case saving an insn makes more difference
627 and more registers are available. */
629 /* If we have calls, contains the insn in which a register was used
630 if it was used exactly once; contains const0_rtx if it was used more
632 varray_type reg_single_usage = 0;
633 /* Nonzero if we are scanning instructions in a sub-loop. */
637 /* Determine whether this loop starts with a jump down to a test at
638 the end. This will occur for a small number of loops with a test
639 that is too complex to duplicate in front of the loop.
641 We search for the first insn or label in the loop, skipping NOTEs.
642 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
643 (because we might have a loop executed only once that contains a
644 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
645 (in case we have a degenerate loop).
647 Note that if we mistakenly think that a loop is entered at the top
648 when, in fact, it is entered at the exit test, the only effect will be
649 slightly poorer optimization. Making the opposite error can generate
650 incorrect code. Since very few loops now start with a jump to the
651 exit test, the code here to detect that case is very conservative. */
653 for (p = NEXT_INSN (loop_start);
655 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
656 && (GET_CODE (p) != NOTE
657 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
658 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
664 /* Set up variables describing this loop. */
665 prescan_loop (loop_start, end);
666 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
668 /* If loop has a jump before the first label,
669 the true entry is the target of that jump.
670 Start scan from there.
671 But record in LOOP_TOP the place where the end-test jumps
672 back to so we can scan that after the end of the loop. */
673 if (GET_CODE (p) == JUMP_INSN)
677 /* Loop entry must be unconditional jump (and not a RETURN) */
679 && JUMP_LABEL (p) != 0
680 /* Check to see whether the jump actually
681 jumps out of the loop (meaning it's no loop).
682 This case can happen for things like
683 do {..} while (0). If this label was generated previously
684 by loop, we can't tell anything about it and have to reject
686 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, end))
688 loop_top = next_label (scan_start);
689 scan_start = JUMP_LABEL (p);
693 /* If SCAN_START was an insn created by loop, we don't know its luid
694 as required by loop_reg_used_before_p. So skip such loops. (This
695 test may never be true, but it's best to play it safe.)
697 Also, skip loops where we do not start scanning at a label. This
698 test also rejects loops starting with a JUMP_INSN that failed the
701 if (INSN_UID (scan_start) >= max_uid_for_loop
702 || GET_CODE (scan_start) != CODE_LABEL)
704 if (loop_dump_stream)
705 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
706 INSN_UID (loop_start), INSN_UID (end));
710 /* Count number of times each reg is set during this loop.
711 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
712 the setting of register I. If this loop has calls, set
713 VARRAY_RTX (reg_single_usage, I). */
715 /* Allocate extra space for REGS that might be created by
716 load_mems. We allocate a little extra slop as well, in the hopes
717 that even after the moving of movables creates some new registers
718 we won't have to reallocate these arrays. However, we do grow
719 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
720 nregs = max_reg_num () + loop_mems_idx + 16;
721 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
722 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
723 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
726 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
728 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
729 may_not_optimize, reg_single_usage, &insn_count, nregs);
731 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
733 VARRAY_CHAR (may_not_optimize, i) = 1;
734 VARRAY_INT (set_in_loop, i) = 1;
737 #ifdef AVOID_CCMODE_COPIES
738 /* Don't try to move insns which set CC registers if we should not
739 create CCmode register copies. */
740 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
741 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
742 VARRAY_CHAR (may_not_optimize, i) = 1;
745 bcopy ((char *) &set_in_loop->data,
746 (char *) &n_times_set->data, nregs * sizeof (int));
748 if (loop_dump_stream)
750 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
751 INSN_UID (loop_start), INSN_UID (end), insn_count);
753 fprintf (loop_dump_stream, "Continue at insn %d.\n",
754 INSN_UID (loop_continue));
757 /* Scan through the loop finding insns that are safe to move.
758 Set set_in_loop negative for the reg being set, so that
759 this reg will be considered invariant for subsequent insns.
760 We consider whether subsequent insns use the reg
761 in deciding whether it is worth actually moving.
763 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
764 and therefore it is possible that the insns we are scanning
765 would never be executed. At such times, we must make sure
766 that it is safe to execute the insn once instead of zero times.
767 When MAYBE_NEVER is 0, all insns will be executed at least once
768 so that is not a problem. */
770 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
772 p = next_insn_in_loop (p, scan_start, end, loop_top))
774 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
775 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
777 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
778 && find_reg_note (p, REG_RETVAL, NULL_RTX))
781 if (GET_CODE (p) == INSN
782 && (set = single_set (p))
783 && GET_CODE (SET_DEST (set)) == REG
784 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
789 rtx src = SET_SRC (set);
790 rtx dependencies = 0;
792 /* Figure out what to use as a source of this insn. If a REG_EQUIV
793 note is given or if a REG_EQUAL note with a constant operand is
794 specified, use it as the source and mark that we should move
795 this insn by calling emit_move_insn rather that duplicating the
798 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
800 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
802 src = XEXP (temp, 0), move_insn = 1;
805 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
806 if (temp && CONSTANT_P (XEXP (temp, 0)))
807 src = XEXP (temp, 0), move_insn = 1;
808 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
810 src = XEXP (temp, 0);
811 /* A libcall block can use regs that don't appear in
812 the equivalent expression. To move the libcall,
813 we must move those regs too. */
814 dependencies = libcall_other_reg (p, src);
818 /* Don't try to optimize a register that was made
819 by loop-optimization for an inner loop.
820 We don't know its life-span, so we can't compute the benefit. */
821 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
823 else if (/* The set is not guaranteed to be executed one
824 the loop starts, or the value before the set is
825 needed before the set occurs... */
827 || loop_reg_used_before_p (set, p, loop_start,
829 /* And the register is used in basic blocks other
830 than the one where it is set (meaning that
831 something after this point in the loop might
832 depend on its value before the set). */
833 && !reg_in_basic_block_p (p, SET_DEST (set)))
834 /* It is unsafe to move the set.
836 This code used to consider it OK to move a set of a variable
837 which was not created by the user and not used in an exit test.
838 That behavior is incorrect and was removed. */
840 else if ((tem = invariant_p (src))
841 && (dependencies == 0
842 || (tem2 = invariant_p (dependencies)) != 0)
843 && (VARRAY_INT (set_in_loop,
844 REGNO (SET_DEST (set))) == 1
846 = consec_sets_invariant_p
848 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
850 /* If the insn can cause a trap (such as divide by zero),
851 can't move it unless it's guaranteed to be executed
852 once loop is entered. Even a function call might
853 prevent the trap insn from being reached
854 (since it might exit!) */
855 && ! ((maybe_never || call_passed)
856 && may_trap_p (src)))
858 register struct movable *m;
859 register int regno = REGNO (SET_DEST (set));
861 /* A potential lossage is where we have a case where two insns
862 can be combined as long as they are both in the loop, but
863 we move one of them outside the loop. For large loops,
864 this can lose. The most common case of this is the address
865 of a function being called.
867 Therefore, if this register is marked as being used exactly
868 once if we are in a loop with calls (a "large loop"), see if
869 we can replace the usage of this register with the source
870 of this SET. If we can, delete this insn.
872 Don't do this if P has a REG_RETVAL note or if we have
873 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
875 if (reg_single_usage && VARRAY_RTX (reg_single_usage, regno) != 0
876 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
877 && REGNO_FIRST_UID (regno) == INSN_UID (p)
878 && (REGNO_LAST_UID (regno)
879 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
880 && VARRAY_INT (set_in_loop, regno) == 1
881 && ! side_effects_p (SET_SRC (set))
882 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
883 && (! SMALL_REGISTER_CLASSES
884 || (! (GET_CODE (SET_SRC (set)) == REG
885 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
886 /* This test is not redundant; SET_SRC (set) might be
887 a call-clobbered register and the life of REGNO
888 might span a call. */
889 && ! modified_between_p (SET_SRC (set), p,
891 (reg_single_usage, regno))
892 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
893 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
895 (reg_single_usage, regno)))
897 /* Replace any usage in a REG_EQUAL note. Must copy the
898 new source, so that we don't get rtx sharing between the
899 SET_SOURCE and REG_NOTES of insn p. */
900 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
901 = replace_rtx (REG_NOTES (VARRAY_RTX
902 (reg_single_usage, regno)),
903 SET_DEST (set), copy_rtx (SET_SRC (set)));
906 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
907 NOTE_SOURCE_FILE (p) = 0;
908 VARRAY_INT (set_in_loop, regno) = 0;
912 m = (struct movable *) alloca (sizeof (struct movable));
916 m->dependencies = dependencies;
917 m->set_dest = SET_DEST (set);
919 m->consec = VARRAY_INT (set_in_loop,
920 REGNO (SET_DEST (set))) - 1;
924 m->move_insn = move_insn;
925 m->move_insn_first = 0;
926 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
927 m->savemode = VOIDmode;
929 /* Set M->cond if either invariant_p or consec_sets_invariant_p
930 returned 2 (only conditionally invariant). */
931 m->cond = ((tem | tem1 | tem2) > 1);
932 m->global = (uid_luid[REGNO_LAST_UID (regno)] > INSN_LUID (end)
933 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
935 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
936 - uid_luid[REGNO_FIRST_UID (regno)]);
937 m->savings = VARRAY_INT (n_times_set, regno);
938 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
939 m->savings += libcall_benefit (p);
940 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
941 /* Add M to the end of the chain MOVABLES. */
945 last_movable->next = m;
950 /* It is possible for the first instruction to have a
951 REG_EQUAL note but a non-invariant SET_SRC, so we must
952 remember the status of the first instruction in case
953 the last instruction doesn't have a REG_EQUAL note. */
954 m->move_insn_first = m->move_insn;
956 /* Skip this insn, not checking REG_LIBCALL notes. */
957 p = next_nonnote_insn (p);
958 /* Skip the consecutive insns, if there are any. */
959 p = skip_consec_insns (p, m->consec);
960 /* Back up to the last insn of the consecutive group. */
961 p = prev_nonnote_insn (p);
963 /* We must now reset m->move_insn, m->is_equiv, and possibly
964 m->set_src to correspond to the effects of all the
966 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
968 m->set_src = XEXP (temp, 0), m->move_insn = 1;
971 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
972 if (temp && CONSTANT_P (XEXP (temp, 0)))
973 m->set_src = XEXP (temp, 0), m->move_insn = 1;
978 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
981 /* If this register is always set within a STRICT_LOW_PART
982 or set to zero, then its high bytes are constant.
983 So clear them outside the loop and within the loop
984 just load the low bytes.
985 We must check that the machine has an instruction to do so.
986 Also, if the value loaded into the register
987 depends on the same register, this cannot be done. */
988 else if (SET_SRC (set) == const0_rtx
989 && GET_CODE (NEXT_INSN (p)) == INSN
990 && (set1 = single_set (NEXT_INSN (p)))
991 && GET_CODE (set1) == SET
992 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
993 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
994 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
996 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
998 register int regno = REGNO (SET_DEST (set));
999 if (VARRAY_INT (set_in_loop, regno) == 2)
1001 register struct movable *m;
1002 m = (struct movable *) alloca (sizeof (struct movable));
1005 m->set_dest = SET_DEST (set);
1006 m->dependencies = 0;
1012 m->move_insn_first = 0;
1014 /* If the insn may not be executed on some cycles,
1015 we can't clear the whole reg; clear just high part.
1016 Not even if the reg is used only within this loop.
1023 Clearing x before the inner loop could clobber a value
1024 being saved from the last time around the outer loop.
1025 However, if the reg is not used outside this loop
1026 and all uses of the register are in the same
1027 basic block as the store, there is no problem.
1029 If this insn was made by loop, we don't know its
1030 INSN_LUID and hence must make a conservative
1032 m->global = (INSN_UID (p) >= max_uid_for_loop
1033 || (uid_luid[REGNO_LAST_UID (regno)]
1035 || (uid_luid[REGNO_FIRST_UID (regno)]
1037 || (labels_in_range_p
1038 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1039 if (maybe_never && m->global)
1040 m->savemode = GET_MODE (SET_SRC (set1));
1042 m->savemode = VOIDmode;
1046 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1047 - uid_luid[REGNO_FIRST_UID (regno)]);
1049 VARRAY_INT (set_in_loop, regno) = -1;
1050 /* Add M to the end of the chain MOVABLES. */
1054 last_movable->next = m;
1059 /* Past a call insn, we get to insns which might not be executed
1060 because the call might exit. This matters for insns that trap.
1061 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1062 so they don't count. */
1063 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1065 /* Past a label or a jump, we get to insns for which we
1066 can't count on whether or how many times they will be
1067 executed during each iteration. Therefore, we can
1068 only move out sets of trivial variables
1069 (those not used after the loop). */
1070 /* Similar code appears twice in strength_reduce. */
1071 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1072 /* If we enter the loop in the middle, and scan around to the
1073 beginning, don't set maybe_never for that. This must be an
1074 unconditional jump, otherwise the code at the top of the
1075 loop might never be executed. Unconditional jumps are
1076 followed a by barrier then loop end. */
1077 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
1078 && NEXT_INSN (NEXT_INSN (p)) == end
1079 && simplejump_p (p)))
1081 else if (GET_CODE (p) == NOTE)
1083 /* At the virtual top of a converted loop, insns are again known to
1084 be executed: logically, the loop begins here even though the exit
1085 code has been duplicated. */
1086 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1087 maybe_never = call_passed = 0;
1088 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1090 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1095 /* If one movable subsumes another, ignore that other. */
1097 ignore_some_movables (movables);
1099 /* For each movable insn, see if the reg that it loads
1100 leads when it dies right into another conditionally movable insn.
1101 If so, record that the second insn "forces" the first one,
1102 since the second can be moved only if the first is. */
1104 force_movables (movables);
1106 /* See if there are multiple movable insns that load the same value.
1107 If there are, make all but the first point at the first one
1108 through the `match' field, and add the priorities of them
1109 all together as the priority of the first. */
1111 combine_movables (movables, nregs);
1113 /* Now consider each movable insn to decide whether it is worth moving.
1114 Store 0 in set_in_loop for each reg that is moved.
1116 Generally this increases code size, so do not move moveables when
1117 optimizing for code size. */
1119 if (! optimize_size)
1120 move_movables (movables, threshold,
1121 insn_count, loop_start, end, nregs);
1123 /* Now candidates that still are negative are those not moved.
1124 Change set_in_loop to indicate that those are not actually invariant. */
1125 for (i = 0; i < nregs; i++)
1126 if (VARRAY_INT (set_in_loop, i) < 0)
1127 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1129 /* Now that we've moved some things out of the loop, we able to
1130 hoist even more memory references. There's no need to pass
1131 reg_single_usage this time, since we're done with it. */
1132 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top,
1136 /* set_in_loop is still used by invariant_p, so we can't free it now. */
1137 VARRAY_FREE (reg_single_usage);
1139 if (flag_strength_reduce)
1141 the_movables = movables;
1142 strength_reduce (scan_start, end, loop_top,
1143 insn_count, loop_start, end, unroll_p, bct_p);
1146 VARRAY_FREE (set_in_loop);
1147 VARRAY_FREE (n_times_set);
1148 VARRAY_FREE (may_not_optimize);
1151 /* Add elements to *OUTPUT to record all the pseudo-regs
1152 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1155 record_excess_regs (in_this, not_in_this, output)
1156 rtx in_this, not_in_this;
1163 code = GET_CODE (in_this);
1177 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1178 && ! reg_mentioned_p (in_this, not_in_this))
1179 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1186 fmt = GET_RTX_FORMAT (code);
1187 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1194 for (j = 0; j < XVECLEN (in_this, i); j++)
1195 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1199 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1205 /* Check what regs are referred to in the libcall block ending with INSN,
1206 aside from those mentioned in the equivalent value.
1207 If there are none, return 0.
1208 If there are one or more, return an EXPR_LIST containing all of them. */
1211 libcall_other_reg (insn, equiv)
1214 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1215 rtx p = XEXP (note, 0);
1218 /* First, find all the regs used in the libcall block
1219 that are not mentioned as inputs to the result. */
1223 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1224 || GET_CODE (p) == CALL_INSN)
1225 record_excess_regs (PATTERN (p), equiv, &output);
1232 /* Return 1 if all uses of REG
1233 are between INSN and the end of the basic block. */
1236 reg_in_basic_block_p (insn, reg)
1239 int regno = REGNO (reg);
1242 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1245 /* Search this basic block for the already recorded last use of the reg. */
1246 for (p = insn; p; p = NEXT_INSN (p))
1248 switch (GET_CODE (p))
1255 /* Ordinary insn: if this is the last use, we win. */
1256 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1261 /* Jump insn: if this is the last use, we win. */
1262 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1264 /* Otherwise, it's the end of the basic block, so we lose. */
1269 /* It's the end of the basic block, so we lose. */
1277 /* The "last use" doesn't follow the "first use"?? */
1281 /* Compute the benefit of eliminating the insns in the block whose
1282 last insn is LAST. This may be a group of insns used to compute a
1283 value directly or can contain a library call. */
1286 libcall_benefit (last)
1292 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1293 insn != last; insn = NEXT_INSN (insn))
1295 if (GET_CODE (insn) == CALL_INSN)
1296 benefit += 10; /* Assume at least this many insns in a library
1298 else if (GET_CODE (insn) == INSN
1299 && GET_CODE (PATTERN (insn)) != USE
1300 && GET_CODE (PATTERN (insn)) != CLOBBER)
1307 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1310 skip_consec_insns (insn, count)
1314 for (; count > 0; count--)
1318 /* If first insn of libcall sequence, skip to end. */
1319 /* Do this at start of loop, since INSN is guaranteed to
1321 if (GET_CODE (insn) != NOTE
1322 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1323 insn = XEXP (temp, 0);
1325 do insn = NEXT_INSN (insn);
1326 while (GET_CODE (insn) == NOTE);
1332 /* Ignore any movable whose insn falls within a libcall
1333 which is part of another movable.
1334 We make use of the fact that the movable for the libcall value
1335 was made later and so appears later on the chain. */
1338 ignore_some_movables (movables)
1339 struct movable *movables;
1341 register struct movable *m, *m1;
1343 for (m = movables; m; m = m->next)
1345 /* Is this a movable for the value of a libcall? */
1346 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1350 /* Check for earlier movables inside that range,
1351 and mark them invalid. We cannot use LUIDs here because
1352 insns created by loop.c for prior loops don't have LUIDs.
1353 Rather than reject all such insns from movables, we just
1354 explicitly check each insn in the libcall (since invariant
1355 libcalls aren't that common). */
1356 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1357 for (m1 = movables; m1 != m; m1 = m1->next)
1358 if (m1->insn == insn)
1364 /* For each movable insn, see if the reg that it loads
1365 leads when it dies right into another conditionally movable insn.
1366 If so, record that the second insn "forces" the first one,
1367 since the second can be moved only if the first is. */
1370 force_movables (movables)
1371 struct movable *movables;
1373 register struct movable *m, *m1;
1374 for (m1 = movables; m1; m1 = m1->next)
1375 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1376 if (!m1->partial && !m1->done)
1378 int regno = m1->regno;
1379 for (m = m1->next; m; m = m->next)
1380 /* ??? Could this be a bug? What if CSE caused the
1381 register of M1 to be used after this insn?
1382 Since CSE does not update regno_last_uid,
1383 this insn M->insn might not be where it dies.
1384 But very likely this doesn't matter; what matters is
1385 that M's reg is computed from M1's reg. */
1386 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1389 if (m != 0 && m->set_src == m1->set_dest
1390 /* If m->consec, m->set_src isn't valid. */
1394 /* Increase the priority of the moving the first insn
1395 since it permits the second to be moved as well. */
1399 m1->lifetime += m->lifetime;
1400 m1->savings += m->savings;
1405 /* Find invariant expressions that are equal and can be combined into
1409 combine_movables (movables, nregs)
1410 struct movable *movables;
1413 register struct movable *m;
1414 char *matched_regs = (char *) alloca (nregs);
1415 enum machine_mode mode;
1417 /* Regs that are set more than once are not allowed to match
1418 or be matched. I'm no longer sure why not. */
1419 /* Perhaps testing m->consec_sets would be more appropriate here? */
1421 for (m = movables; m; m = m->next)
1422 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1424 register struct movable *m1;
1425 int regno = m->regno;
1427 bzero (matched_regs, nregs);
1428 matched_regs[regno] = 1;
1430 /* We want later insns to match the first one. Don't make the first
1431 one match any later ones. So start this loop at m->next. */
1432 for (m1 = m->next; m1; m1 = m1->next)
1433 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1434 /* A reg used outside the loop mustn't be eliminated. */
1436 /* A reg used for zero-extending mustn't be eliminated. */
1438 && (matched_regs[m1->regno]
1441 /* Can combine regs with different modes loaded from the
1442 same constant only if the modes are the same or
1443 if both are integer modes with M wider or the same
1444 width as M1. The check for integer is redundant, but
1445 safe, since the only case of differing destination
1446 modes with equal sources is when both sources are
1447 VOIDmode, i.e., CONST_INT. */
1448 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1449 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1450 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1451 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1452 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1453 /* See if the source of M1 says it matches M. */
1454 && ((GET_CODE (m1->set_src) == REG
1455 && matched_regs[REGNO (m1->set_src)])
1456 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1458 && ((m->dependencies == m1->dependencies)
1459 || rtx_equal_p (m->dependencies, m1->dependencies)))
1461 m->lifetime += m1->lifetime;
1462 m->savings += m1->savings;
1465 matched_regs[m1->regno] = 1;
1469 /* Now combine the regs used for zero-extension.
1470 This can be done for those not marked `global'
1471 provided their lives don't overlap. */
1473 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1474 mode = GET_MODE_WIDER_MODE (mode))
1476 register struct movable *m0 = 0;
1478 /* Combine all the registers for extension from mode MODE.
1479 Don't combine any that are used outside this loop. */
1480 for (m = movables; m; m = m->next)
1481 if (m->partial && ! m->global
1482 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1484 register struct movable *m1;
1485 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1486 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1490 /* First one: don't check for overlap, just record it. */
1495 /* Make sure they extend to the same mode.
1496 (Almost always true.) */
1497 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1500 /* We already have one: check for overlap with those
1501 already combined together. */
1502 for (m1 = movables; m1 != m; m1 = m1->next)
1503 if (m1 == m0 || (m1->partial && m1->match == m0))
1504 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1505 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1508 /* No overlap: we can combine this with the others. */
1509 m0->lifetime += m->lifetime;
1510 m0->savings += m->savings;
1519 /* Return 1 if regs X and Y will become the same if moved. */
1522 regs_match_p (x, y, movables)
1524 struct movable *movables;
1528 struct movable *mx, *my;
1530 for (mx = movables; mx; mx = mx->next)
1531 if (mx->regno == xn)
1534 for (my = movables; my; my = my->next)
1535 if (my->regno == yn)
1539 && ((mx->match == my->match && mx->match != 0)
1541 || mx == my->match));
1544 /* Return 1 if X and Y are identical-looking rtx's.
1545 This is the Lisp function EQUAL for rtx arguments.
1547 If two registers are matching movables or a movable register and an
1548 equivalent constant, consider them equal. */
1551 rtx_equal_for_loop_p (x, y, movables)
1553 struct movable *movables;
1557 register struct movable *m;
1558 register enum rtx_code code;
1563 if (x == 0 || y == 0)
1566 code = GET_CODE (x);
1568 /* If we have a register and a constant, they may sometimes be
1570 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1573 for (m = movables; m; m = m->next)
1574 if (m->move_insn && m->regno == REGNO (x)
1575 && rtx_equal_p (m->set_src, y))
1578 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1581 for (m = movables; m; m = m->next)
1582 if (m->move_insn && m->regno == REGNO (y)
1583 && rtx_equal_p (m->set_src, x))
1587 /* Otherwise, rtx's of different codes cannot be equal. */
1588 if (code != GET_CODE (y))
1591 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1592 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1594 if (GET_MODE (x) != GET_MODE (y))
1597 /* These three types of rtx's can be compared nonrecursively. */
1599 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1601 if (code == LABEL_REF)
1602 return XEXP (x, 0) == XEXP (y, 0);
1603 if (code == SYMBOL_REF)
1604 return XSTR (x, 0) == XSTR (y, 0);
1606 /* Compare the elements. If any pair of corresponding elements
1607 fail to match, return 0 for the whole things. */
1609 fmt = GET_RTX_FORMAT (code);
1610 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1615 if (XWINT (x, i) != XWINT (y, i))
1620 if (XINT (x, i) != XINT (y, i))
1625 /* Two vectors must have the same length. */
1626 if (XVECLEN (x, i) != XVECLEN (y, i))
1629 /* And the corresponding elements must match. */
1630 for (j = 0; j < XVECLEN (x, i); j++)
1631 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1636 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1641 if (strcmp (XSTR (x, i), XSTR (y, i)))
1646 /* These are just backpointers, so they don't matter. */
1652 /* It is believed that rtx's at this level will never
1653 contain anything but integers and other rtx's,
1654 except for within LABEL_REFs and SYMBOL_REFs. */
1662 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1663 insns in INSNS which use thet reference. */
1666 add_label_notes (x, insns)
1670 enum rtx_code code = GET_CODE (x);
1675 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1677 /* This code used to ignore labels that referred to dispatch tables to
1678 avoid flow generating (slighly) worse code.
1680 We no longer ignore such label references (see LABEL_REF handling in
1681 mark_jump_label for additional information). */
1682 for (insn = insns; insn; insn = NEXT_INSN (insn))
1683 if (reg_mentioned_p (XEXP (x, 0), insn))
1684 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1688 fmt = GET_RTX_FORMAT (code);
1689 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1692 add_label_notes (XEXP (x, i), insns);
1693 else if (fmt[i] == 'E')
1694 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1695 add_label_notes (XVECEXP (x, i, j), insns);
1699 /* Scan MOVABLES, and move the insns that deserve to be moved.
1700 If two matching movables are combined, replace one reg with the
1701 other throughout. */
1704 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1705 struct movable *movables;
1713 register struct movable *m;
1715 /* Map of pseudo-register replacements to handle combining
1716 when we move several insns that load the same value
1717 into different pseudo-registers. */
1718 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1719 char *already_moved = (char *) alloca (nregs);
1721 bzero (already_moved, nregs);
1722 bzero ((char *) reg_map, nregs * sizeof (rtx));
1726 for (m = movables; m; m = m->next)
1728 /* Describe this movable insn. */
1730 if (loop_dump_stream)
1732 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1733 INSN_UID (m->insn), m->regno, m->lifetime);
1735 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1737 fprintf (loop_dump_stream, "cond ");
1739 fprintf (loop_dump_stream, "force ");
1741 fprintf (loop_dump_stream, "global ");
1743 fprintf (loop_dump_stream, "done ");
1745 fprintf (loop_dump_stream, "move-insn ");
1747 fprintf (loop_dump_stream, "matches %d ",
1748 INSN_UID (m->match->insn));
1750 fprintf (loop_dump_stream, "forces %d ",
1751 INSN_UID (m->forces->insn));
1754 /* Count movables. Value used in heuristics in strength_reduce. */
1757 /* Ignore the insn if it's already done (it matched something else).
1758 Otherwise, see if it is now safe to move. */
1762 || (1 == invariant_p (m->set_src)
1763 && (m->dependencies == 0
1764 || 1 == invariant_p (m->dependencies))
1766 || 1 == consec_sets_invariant_p (m->set_dest,
1769 && (! m->forces || m->forces->done))
1773 int savings = m->savings;
1775 /* We have an insn that is safe to move.
1776 Compute its desirability. */
1781 if (loop_dump_stream)
1782 fprintf (loop_dump_stream, "savings %d ", savings);
1784 if (moved_once[regno] && loop_dump_stream)
1785 fprintf (loop_dump_stream, "halved since already moved ");
1787 /* An insn MUST be moved if we already moved something else
1788 which is safe only if this one is moved too: that is,
1789 if already_moved[REGNO] is nonzero. */
1791 /* An insn is desirable to move if the new lifetime of the
1792 register is no more than THRESHOLD times the old lifetime.
1793 If it's not desirable, it means the loop is so big
1794 that moving won't speed things up much,
1795 and it is liable to make register usage worse. */
1797 /* It is also desirable to move if it can be moved at no
1798 extra cost because something else was already moved. */
1800 if (already_moved[regno]
1801 || flag_move_all_movables
1802 || (threshold * savings * m->lifetime) >=
1803 (moved_once[regno] ? insn_count * 2 : insn_count)
1804 || (m->forces && m->forces->done
1805 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1808 register struct movable *m1;
1811 /* Now move the insns that set the reg. */
1813 if (m->partial && m->match)
1817 /* Find the end of this chain of matching regs.
1818 Thus, we load each reg in the chain from that one reg.
1819 And that reg is loaded with 0 directly,
1820 since it has ->match == 0. */
1821 for (m1 = m; m1->match; m1 = m1->match);
1822 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1823 SET_DEST (PATTERN (m1->insn)));
1824 i1 = emit_insn_before (newpat, loop_start);
1826 /* Mark the moved, invariant reg as being allowed to
1827 share a hard reg with the other matching invariant. */
1828 REG_NOTES (i1) = REG_NOTES (m->insn);
1829 r1 = SET_DEST (PATTERN (m->insn));
1830 r2 = SET_DEST (PATTERN (m1->insn));
1832 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1833 gen_rtx_EXPR_LIST (VOIDmode, r2,
1835 delete_insn (m->insn);
1840 if (loop_dump_stream)
1841 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1843 /* If we are to re-generate the item being moved with a
1844 new move insn, first delete what we have and then emit
1845 the move insn before the loop. */
1846 else if (m->move_insn)
1850 for (count = m->consec; count >= 0; count--)
1852 /* If this is the first insn of a library call sequence,
1854 if (GET_CODE (p) != NOTE
1855 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1858 /* If this is the last insn of a libcall sequence, then
1859 delete every insn in the sequence except the last.
1860 The last insn is handled in the normal manner. */
1861 if (GET_CODE (p) != NOTE
1862 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1864 temp = XEXP (temp, 0);
1866 temp = delete_insn (temp);
1870 p = delete_insn (p);
1872 /* simplify_giv_expr expects that it can walk the insns
1873 at m->insn forwards and see this old sequence we are
1874 tossing here. delete_insn does preserve the next
1875 pointers, but when we skip over a NOTE we must fix
1876 it up. Otherwise that code walks into the non-deleted
1878 while (p && GET_CODE (p) == NOTE)
1879 p = NEXT_INSN (temp) = NEXT_INSN (p);
1883 emit_move_insn (m->set_dest, m->set_src);
1884 temp = get_insns ();
1887 add_label_notes (m->set_src, temp);
1889 i1 = emit_insns_before (temp, loop_start);
1890 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1892 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1893 m->set_src, REG_NOTES (i1));
1895 if (loop_dump_stream)
1896 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1898 /* The more regs we move, the less we like moving them. */
1903 for (count = m->consec; count >= 0; count--)
1907 /* If first insn of libcall sequence, skip to end. */
1908 /* Do this at start of loop, since p is guaranteed to
1910 if (GET_CODE (p) != NOTE
1911 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1914 /* If last insn of libcall sequence, move all
1915 insns except the last before the loop. The last
1916 insn is handled in the normal manner. */
1917 if (GET_CODE (p) != NOTE
1918 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1922 rtx fn_address_insn = 0;
1925 for (temp = XEXP (temp, 0); temp != p;
1926 temp = NEXT_INSN (temp))
1932 if (GET_CODE (temp) == NOTE)
1935 body = PATTERN (temp);
1937 /* Find the next insn after TEMP,
1938 not counting USE or NOTE insns. */
1939 for (next = NEXT_INSN (temp); next != p;
1940 next = NEXT_INSN (next))
1941 if (! (GET_CODE (next) == INSN
1942 && GET_CODE (PATTERN (next)) == USE)
1943 && GET_CODE (next) != NOTE)
1946 /* If that is the call, this may be the insn
1947 that loads the function address.
1949 Extract the function address from the insn
1950 that loads it into a register.
1951 If this insn was cse'd, we get incorrect code.
1953 So emit a new move insn that copies the
1954 function address into the register that the
1955 call insn will use. flow.c will delete any
1956 redundant stores that we have created. */
1957 if (GET_CODE (next) == CALL_INSN
1958 && GET_CODE (body) == SET
1959 && GET_CODE (SET_DEST (body)) == REG
1960 && (n = find_reg_note (temp, REG_EQUAL,
1963 fn_reg = SET_SRC (body);
1964 if (GET_CODE (fn_reg) != REG)
1965 fn_reg = SET_DEST (body);
1966 fn_address = XEXP (n, 0);
1967 fn_address_insn = temp;
1969 /* We have the call insn.
1970 If it uses the register we suspect it might,
1971 load it with the correct address directly. */
1972 if (GET_CODE (temp) == CALL_INSN
1974 && reg_referenced_p (fn_reg, body))
1975 emit_insn_after (gen_move_insn (fn_reg,
1979 if (GET_CODE (temp) == CALL_INSN)
1981 i1 = emit_call_insn_before (body, loop_start);
1982 /* Because the USAGE information potentially
1983 contains objects other than hard registers
1984 we need to copy it. */
1985 if (CALL_INSN_FUNCTION_USAGE (temp))
1986 CALL_INSN_FUNCTION_USAGE (i1)
1987 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1990 i1 = emit_insn_before (body, loop_start);
1993 if (temp == fn_address_insn)
1994 fn_address_insn = i1;
1995 REG_NOTES (i1) = REG_NOTES (temp);
2001 if (m->savemode != VOIDmode)
2003 /* P sets REG to zero; but we should clear only
2004 the bits that are not covered by the mode
2006 rtx reg = m->set_dest;
2012 (GET_MODE (reg), and_optab, reg,
2013 GEN_INT ((((HOST_WIDE_INT) 1
2014 << GET_MODE_BITSIZE (m->savemode)))
2016 reg, 1, OPTAB_LIB_WIDEN);
2020 emit_move_insn (reg, tem);
2021 sequence = gen_sequence ();
2023 i1 = emit_insn_before (sequence, loop_start);
2025 else if (GET_CODE (p) == CALL_INSN)
2027 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2028 /* Because the USAGE information potentially
2029 contains objects other than hard registers
2030 we need to copy it. */
2031 if (CALL_INSN_FUNCTION_USAGE (p))
2032 CALL_INSN_FUNCTION_USAGE (i1)
2033 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2035 else if (count == m->consec && m->move_insn_first)
2037 /* The SET_SRC might not be invariant, so we must
2038 use the REG_EQUAL note. */
2040 emit_move_insn (m->set_dest, m->set_src);
2041 temp = get_insns ();
2044 add_label_notes (m->set_src, temp);
2046 i1 = emit_insns_before (temp, loop_start);
2047 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2049 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2051 m->set_src, REG_NOTES (i1));
2054 i1 = emit_insn_before (PATTERN (p), loop_start);
2056 if (REG_NOTES (i1) == 0)
2058 REG_NOTES (i1) = REG_NOTES (p);
2060 /* If there is a REG_EQUAL note present whose value
2061 is not loop invariant, then delete it, since it
2062 may cause problems with later optimization passes.
2063 It is possible for cse to create such notes
2064 like this as a result of record_jump_cond. */
2066 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2067 && ! invariant_p (XEXP (temp, 0)))
2068 remove_note (i1, temp);
2074 if (loop_dump_stream)
2075 fprintf (loop_dump_stream, " moved to %d",
2078 /* If library call, now fix the REG_NOTES that contain
2079 insn pointers, namely REG_LIBCALL on FIRST
2080 and REG_RETVAL on I1. */
2081 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2083 XEXP (temp, 0) = first;
2084 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2085 XEXP (temp, 0) = i1;
2092 /* simplify_giv_expr expects that it can walk the insns
2093 at m->insn forwards and see this old sequence we are
2094 tossing here. delete_insn does preserve the next
2095 pointers, but when we skip over a NOTE we must fix
2096 it up. Otherwise that code walks into the non-deleted
2098 while (p && GET_CODE (p) == NOTE)
2099 p = NEXT_INSN (temp) = NEXT_INSN (p);
2102 /* The more regs we move, the less we like moving them. */
2106 /* Any other movable that loads the same register
2108 already_moved[regno] = 1;
2110 /* This reg has been moved out of one loop. */
2111 moved_once[regno] = 1;
2113 /* The reg set here is now invariant. */
2115 VARRAY_INT (set_in_loop, regno) = 0;
2119 /* Change the length-of-life info for the register
2120 to say it lives at least the full length of this loop.
2121 This will help guide optimizations in outer loops. */
2123 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2124 /* This is the old insn before all the moved insns.
2125 We can't use the moved insn because it is out of range
2126 in uid_luid. Only the old insns have luids. */
2127 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2128 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (end))
2129 REGNO_LAST_UID (regno) = INSN_UID (end);
2131 /* Combine with this moved insn any other matching movables. */
2134 for (m1 = movables; m1; m1 = m1->next)
2139 /* Schedule the reg loaded by M1
2140 for replacement so that shares the reg of M.
2141 If the modes differ (only possible in restricted
2142 circumstances, make a SUBREG. */
2143 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2144 reg_map[m1->regno] = m->set_dest;
2147 = gen_lowpart_common (GET_MODE (m1->set_dest),
2150 /* Get rid of the matching insn
2151 and prevent further processing of it. */
2154 /* if library call, delete all insn except last, which
2156 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2159 for (temp = XEXP (temp, 0); temp != m1->insn;
2160 temp = NEXT_INSN (temp))
2163 delete_insn (m1->insn);
2165 /* Any other movable that loads the same register
2167 already_moved[m1->regno] = 1;
2169 /* The reg merged here is now invariant,
2170 if the reg it matches is invariant. */
2172 VARRAY_INT (set_in_loop, m1->regno) = 0;
2175 else if (loop_dump_stream)
2176 fprintf (loop_dump_stream, "not desirable");
2178 else if (loop_dump_stream && !m->match)
2179 fprintf (loop_dump_stream, "not safe");
2181 if (loop_dump_stream)
2182 fprintf (loop_dump_stream, "\n");
2186 new_start = loop_start;
2188 /* Go through all the instructions in the loop, making
2189 all the register substitutions scheduled in REG_MAP. */
2190 for (p = new_start; p != end; p = NEXT_INSN (p))
2191 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2192 || GET_CODE (p) == CALL_INSN)
2194 replace_regs (PATTERN (p), reg_map, nregs, 0);
2195 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2201 /* Scan X and replace the address of any MEM in it with ADDR.
2202 REG is the address that MEM should have before the replacement. */
2205 replace_call_address (x, reg, addr)
2208 register enum rtx_code code;
2214 code = GET_CODE (x);
2228 /* Short cut for very common case. */
2229 replace_call_address (XEXP (x, 1), reg, addr);
2233 /* Short cut for very common case. */
2234 replace_call_address (XEXP (x, 0), reg, addr);
2238 /* If this MEM uses a reg other than the one we expected,
2239 something is wrong. */
2240 if (XEXP (x, 0) != reg)
2249 fmt = GET_RTX_FORMAT (code);
2250 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2253 replace_call_address (XEXP (x, i), reg, addr);
2257 for (j = 0; j < XVECLEN (x, i); j++)
2258 replace_call_address (XVECEXP (x, i, j), reg, addr);
2264 /* Return the number of memory refs to addresses that vary
2268 count_nonfixed_reads (x)
2271 register enum rtx_code code;
2279 code = GET_CODE (x);
2293 return ((invariant_p (XEXP (x, 0)) != 1)
2294 + count_nonfixed_reads (XEXP (x, 0)));
2301 fmt = GET_RTX_FORMAT (code);
2302 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2305 value += count_nonfixed_reads (XEXP (x, i));
2309 for (j = 0; j < XVECLEN (x, i); j++)
2310 value += count_nonfixed_reads (XVECEXP (x, i, j));
2318 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2319 Replace it with an instruction to load just the low bytes
2320 if the machine supports such an instruction,
2321 and insert above LOOP_START an instruction to clear the register. */
2324 constant_high_bytes (p, loop_start)
2328 register int insn_code_number;
2330 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2331 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2333 new = gen_rtx_SET (VOIDmode,
2334 gen_rtx_STRICT_LOW_PART (VOIDmode,
2335 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2336 SET_DEST (PATTERN (p)),
2338 XEXP (SET_SRC (PATTERN (p)), 0));
2339 insn_code_number = recog (new, p);
2341 if (insn_code_number)
2345 /* Clear destination register before the loop. */
2346 emit_insn_before (gen_rtx_SET (VOIDmode, SET_DEST (PATTERN (p)),
2350 /* Inside the loop, just load the low part. */
2356 /* Scan a loop setting the variables `unknown_address_altered',
2357 `num_mem_sets', `loop_continue', `loops_enclosed', `loop_has_call',
2358 `loop_has_volatile', and `loop_has_tablejump'.
2359 Also, fill in the arrays `loop_mems' and `loop_store_mems'. */
2362 prescan_loop (start, end)
2365 register int level = 1;
2367 int loop_has_multiple_exit_targets = 0;
2368 /* The label after END. Jumping here is just like falling off the
2369 end of the loop. We use next_nonnote_insn instead of next_label
2370 as a hedge against the (pathological) case where some actual insn
2371 might end up between the two. */
2372 rtx exit_target = next_nonnote_insn (end);
2373 if (exit_target == NULL_RTX || GET_CODE (exit_target) != CODE_LABEL)
2374 loop_has_multiple_exit_targets = 1;
2376 unknown_address_altered = 0;
2378 loop_has_volatile = 0;
2379 loop_has_tablejump = 0;
2380 loop_store_mems_idx = 0;
2387 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2388 insn = NEXT_INSN (insn))
2390 if (GET_CODE (insn) == NOTE)
2392 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2395 /* Count number of loops contained in this one. */
2398 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2407 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2410 loop_continue = insn;
2413 else if (GET_CODE (insn) == CALL_INSN)
2415 if (! CONST_CALL_P (insn))
2416 unknown_address_altered = 1;
2419 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2421 rtx label1 = NULL_RTX;
2422 rtx label2 = NULL_RTX;
2424 if (volatile_refs_p (PATTERN (insn)))
2425 loop_has_volatile = 1;
2427 if (GET_CODE (insn) == JUMP_INSN
2428 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2429 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2430 loop_has_tablejump = 1;
2432 note_stores (PATTERN (insn), note_addr_stored);
2434 if (! loop_has_multiple_exit_targets
2435 && GET_CODE (insn) == JUMP_INSN
2436 && GET_CODE (PATTERN (insn)) == SET
2437 && SET_DEST (PATTERN (insn)) == pc_rtx)
2439 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2441 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2442 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2446 label1 = SET_SRC (PATTERN (insn));
2450 if (label1 && label1 != pc_rtx)
2452 if (GET_CODE (label1) != LABEL_REF)
2454 /* Something tricky. */
2455 loop_has_multiple_exit_targets = 1;
2458 else if (XEXP (label1, 0) != exit_target
2459 && LABEL_OUTSIDE_LOOP_P (label1))
2461 /* A jump outside the current loop. */
2462 loop_has_multiple_exit_targets = 1;
2472 else if (GET_CODE (insn) == RETURN)
2473 loop_has_multiple_exit_targets = 1;
2476 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2477 if (/* We can't tell what MEMs are aliased by what. */
2478 !unknown_address_altered
2479 /* An exception thrown by a called function might land us
2482 /* We don't want loads for MEMs moved to a location before the
2483 one at which their stack memory becomes allocated. (Note
2484 that this is not a problem for malloc, etc., since those
2485 require actual function calls. */
2486 && !current_function_calls_alloca
2487 /* There are ways to leave the loop other than falling off the
2489 && !loop_has_multiple_exit_targets)
2490 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2491 insn = NEXT_INSN (insn))
2492 for_each_rtx (&insn, insert_loop_mem, 0);
2495 /* Scan the function looking for loops. Record the start and end of each loop.
2496 Also mark as invalid loops any loops that contain a setjmp or are branched
2497 to from outside the loop. */
2500 find_and_verify_loops (f)
2504 int current_loop = -1;
2508 /* If there are jumps to undefined labels,
2509 treat them as jumps out of any/all loops.
2510 This also avoids writing past end of tables when there are no loops. */
2511 uid_loop_num[0] = -1;
2513 /* Find boundaries of loops, mark which loops are contained within
2514 loops, and invalidate loops that have setjmp. */
2516 for (insn = f; insn; insn = NEXT_INSN (insn))
2518 if (GET_CODE (insn) == NOTE)
2519 switch (NOTE_LINE_NUMBER (insn))
2521 case NOTE_INSN_LOOP_BEG:
2522 loop_number_loop_starts[++next_loop] = insn;
2523 loop_number_loop_ends[next_loop] = 0;
2524 loop_outer_loop[next_loop] = current_loop;
2525 loop_invalid[next_loop] = 0;
2526 loop_number_exit_labels[next_loop] = 0;
2527 loop_number_exit_count[next_loop] = 0;
2528 current_loop = next_loop;
2531 case NOTE_INSN_SETJMP:
2532 /* In this case, we must invalidate our current loop and any
2534 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2536 loop_invalid[loop] = 1;
2537 if (loop_dump_stream)
2538 fprintf (loop_dump_stream,
2539 "\nLoop at %d ignored due to setjmp.\n",
2540 INSN_UID (loop_number_loop_starts[loop]));
2544 case NOTE_INSN_LOOP_END:
2545 if (current_loop == -1)
2548 loop_number_loop_ends[current_loop] = insn;
2549 current_loop = loop_outer_loop[current_loop];
2556 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2557 enclosing loop, but this doesn't matter. */
2558 uid_loop_num[INSN_UID (insn)] = current_loop;
2561 /* Any loop containing a label used in an initializer must be invalidated,
2562 because it can be jumped into from anywhere. */
2564 for (label = forced_labels; label; label = XEXP (label, 1))
2568 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2570 loop_num = loop_outer_loop[loop_num])
2571 loop_invalid[loop_num] = 1;
2574 /* Any loop containing a label used for an exception handler must be
2575 invalidated, because it can be jumped into from anywhere. */
2577 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2581 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2583 loop_num = loop_outer_loop[loop_num])
2584 loop_invalid[loop_num] = 1;
2587 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2588 loop that it is not contained within, that loop is marked invalid.
2589 If any INSN or CALL_INSN uses a label's address, then the loop containing
2590 that label is marked invalid, because it could be jumped into from
2593 Also look for blocks of code ending in an unconditional branch that
2594 exits the loop. If such a block is surrounded by a conditional
2595 branch around the block, move the block elsewhere (see below) and
2596 invert the jump to point to the code block. This may eliminate a
2597 label in our loop and will simplify processing by both us and a
2598 possible second cse pass. */
2600 for (insn = f; insn; insn = NEXT_INSN (insn))
2601 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2603 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2605 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2607 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2612 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2614 loop_num = loop_outer_loop[loop_num])
2615 loop_invalid[loop_num] = 1;
2619 if (GET_CODE (insn) != JUMP_INSN)
2622 mark_loop_jump (PATTERN (insn), this_loop_num);
2624 /* See if this is an unconditional branch outside the loop. */
2625 if (this_loop_num != -1
2626 && (GET_CODE (PATTERN (insn)) == RETURN
2627 || (simplejump_p (insn)
2628 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2630 && get_max_uid () < max_uid_for_loop)
2633 rtx our_next = next_real_insn (insn);
2635 int outer_loop = -1;
2637 /* Go backwards until we reach the start of the loop, a label,
2639 for (p = PREV_INSN (insn);
2640 GET_CODE (p) != CODE_LABEL
2641 && ! (GET_CODE (p) == NOTE
2642 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2643 && GET_CODE (p) != JUMP_INSN;
2647 /* Check for the case where we have a jump to an inner nested
2648 loop, and do not perform the optimization in that case. */
2650 if (JUMP_LABEL (insn))
2652 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2653 if (dest_loop != -1)
2655 for (outer_loop = dest_loop; outer_loop != -1;
2656 outer_loop = loop_outer_loop[outer_loop])
2657 if (outer_loop == this_loop_num)
2662 /* Make sure that the target of P is within the current loop. */
2664 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2665 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2666 outer_loop = this_loop_num;
2668 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2669 we have a block of code to try to move.
2671 We look backward and then forward from the target of INSN
2672 to find a BARRIER at the same loop depth as the target.
2673 If we find such a BARRIER, we make a new label for the start
2674 of the block, invert the jump in P and point it to that label,
2675 and move the block of code to the spot we found. */
2677 if (outer_loop == -1
2678 && GET_CODE (p) == JUMP_INSN
2679 && JUMP_LABEL (p) != 0
2680 /* Just ignore jumps to labels that were never emitted.
2681 These always indicate compilation errors. */
2682 && INSN_UID (JUMP_LABEL (p)) != 0
2684 && ! simplejump_p (p)
2685 && next_real_insn (JUMP_LABEL (p)) == our_next)
2688 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2689 int target_loop_num = uid_loop_num[INSN_UID (target)];
2692 for (loc = target; loc; loc = PREV_INSN (loc))
2693 if (GET_CODE (loc) == BARRIER
2694 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2698 for (loc = target; loc; loc = NEXT_INSN (loc))
2699 if (GET_CODE (loc) == BARRIER
2700 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2705 rtx cond_label = JUMP_LABEL (p);
2706 rtx new_label = get_label_after (p);
2708 /* Ensure our label doesn't go away. */
2709 LABEL_NUSES (cond_label)++;
2711 /* Verify that uid_loop_num is large enough and that
2713 if (invert_jump (p, new_label))
2717 /* If no suitable BARRIER was found, create a suitable
2718 one before TARGET. Since TARGET is a fall through
2719 path, we'll need to insert an jump around our block
2720 and a add a BARRIER before TARGET.
2722 This creates an extra unconditional jump outside
2723 the loop. However, the benefits of removing rarely
2724 executed instructions from inside the loop usually
2725 outweighs the cost of the extra unconditional jump
2726 outside the loop. */
2731 temp = gen_jump (JUMP_LABEL (insn));
2732 temp = emit_jump_insn_before (temp, target);
2733 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2734 LABEL_NUSES (JUMP_LABEL (insn))++;
2735 loc = emit_barrier_before (target);
2738 /* Include the BARRIER after INSN and copy the
2740 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2741 reorder_insns (new_label, NEXT_INSN (insn), loc);
2743 /* All those insns are now in TARGET_LOOP_NUM. */
2744 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2746 uid_loop_num[INSN_UID (q)] = target_loop_num;
2748 /* The label jumped to by INSN is no longer a loop exit.
2749 Unless INSN does not have a label (e.g., it is a
2750 RETURN insn), search loop_number_exit_labels to find
2751 its label_ref, and remove it. Also turn off
2752 LABEL_OUTSIDE_LOOP_P bit. */
2753 if (JUMP_LABEL (insn))
2758 r = loop_number_exit_labels[this_loop_num];
2759 r; q = r, r = LABEL_NEXTREF (r))
2760 if (XEXP (r, 0) == JUMP_LABEL (insn))
2762 LABEL_OUTSIDE_LOOP_P (r) = 0;
2764 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2766 loop_number_exit_labels[this_loop_num]
2767 = LABEL_NEXTREF (r);
2771 for (loop_num = this_loop_num;
2772 loop_num != -1 && loop_num != target_loop_num;
2773 loop_num = loop_outer_loop[loop_num])
2774 loop_number_exit_count[loop_num]--;
2776 /* If we didn't find it, then something is wrong. */
2781 /* P is now a jump outside the loop, so it must be put
2782 in loop_number_exit_labels, and marked as such.
2783 The easiest way to do this is to just call
2784 mark_loop_jump again for P. */
2785 mark_loop_jump (PATTERN (p), this_loop_num);
2787 /* If INSN now jumps to the insn after it,
2789 if (JUMP_LABEL (insn) != 0
2790 && (next_real_insn (JUMP_LABEL (insn))
2791 == next_real_insn (insn)))
2795 /* Continue the loop after where the conditional
2796 branch used to jump, since the only branch insn
2797 in the block (if it still remains) is an inter-loop
2798 branch and hence needs no processing. */
2799 insn = NEXT_INSN (cond_label);
2801 if (--LABEL_NUSES (cond_label) == 0)
2802 delete_insn (cond_label);
2804 /* This loop will be continued with NEXT_INSN (insn). */
2805 insn = PREV_INSN (insn);
2812 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2813 loops it is contained in, mark the target loop invalid.
2815 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2818 mark_loop_jump (x, loop_num)
2826 switch (GET_CODE (x))
2839 /* There could be a label reference in here. */
2840 mark_loop_jump (XEXP (x, 0), loop_num);
2846 mark_loop_jump (XEXP (x, 0), loop_num);
2847 mark_loop_jump (XEXP (x, 1), loop_num);
2852 mark_loop_jump (XEXP (x, 0), loop_num);
2856 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2858 /* Link together all labels that branch outside the loop. This
2859 is used by final_[bg]iv_value and the loop unrolling code. Also
2860 mark this LABEL_REF so we know that this branch should predict
2863 /* A check to make sure the label is not in an inner nested loop,
2864 since this does not count as a loop exit. */
2865 if (dest_loop != -1)
2867 for (outer_loop = dest_loop; outer_loop != -1;
2868 outer_loop = loop_outer_loop[outer_loop])
2869 if (outer_loop == loop_num)
2875 if (loop_num != -1 && outer_loop == -1)
2877 LABEL_OUTSIDE_LOOP_P (x) = 1;
2878 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2879 loop_number_exit_labels[loop_num] = x;
2881 for (outer_loop = loop_num;
2882 outer_loop != -1 && outer_loop != dest_loop;
2883 outer_loop = loop_outer_loop[outer_loop])
2884 loop_number_exit_count[outer_loop]++;
2887 /* If this is inside a loop, but not in the current loop or one enclosed
2888 by it, it invalidates at least one loop. */
2890 if (dest_loop == -1)
2893 /* We must invalidate every nested loop containing the target of this
2894 label, except those that also contain the jump insn. */
2896 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2898 /* Stop when we reach a loop that also contains the jump insn. */
2899 for (outer_loop = loop_num; outer_loop != -1;
2900 outer_loop = loop_outer_loop[outer_loop])
2901 if (dest_loop == outer_loop)
2904 /* If we get here, we know we need to invalidate a loop. */
2905 if (loop_dump_stream && ! loop_invalid[dest_loop])
2906 fprintf (loop_dump_stream,
2907 "\nLoop at %d ignored due to multiple entry points.\n",
2908 INSN_UID (loop_number_loop_starts[dest_loop]));
2910 loop_invalid[dest_loop] = 1;
2915 /* If this is not setting pc, ignore. */
2916 if (SET_DEST (x) == pc_rtx)
2917 mark_loop_jump (SET_SRC (x), loop_num);
2921 mark_loop_jump (XEXP (x, 1), loop_num);
2922 mark_loop_jump (XEXP (x, 2), loop_num);
2927 for (i = 0; i < XVECLEN (x, 0); i++)
2928 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2932 for (i = 0; i < XVECLEN (x, 1); i++)
2933 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2937 /* Treat anything else (such as a symbol_ref)
2938 as a branch out of this loop, but not into any loop. */
2942 #ifdef HAVE_decrement_and_branch_on_count
2943 LABEL_OUTSIDE_LOOP_P (x) = 1;
2944 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2945 #endif /* HAVE_decrement_and_branch_on_count */
2947 loop_number_exit_labels[loop_num] = x;
2949 for (outer_loop = loop_num; outer_loop != -1;
2950 outer_loop = loop_outer_loop[outer_loop])
2951 loop_number_exit_count[outer_loop]++;
2957 /* Return nonzero if there is a label in the range from
2958 insn INSN to and including the insn whose luid is END
2959 INSN must have an assigned luid (i.e., it must not have
2960 been previously created by loop.c). */
2963 labels_in_range_p (insn, end)
2967 while (insn && INSN_LUID (insn) <= end)
2969 if (GET_CODE (insn) == CODE_LABEL)
2971 insn = NEXT_INSN (insn);
2977 /* Record that a memory reference X is being set. */
2980 note_addr_stored (x, y)
2982 rtx y ATTRIBUTE_UNUSED;
2986 if (x == 0 || GET_CODE (x) != MEM)
2989 /* Count number of memory writes.
2990 This affects heuristics in strength_reduce. */
2993 /* BLKmode MEM means all memory is clobbered. */
2994 if (GET_MODE (x) == BLKmode)
2995 unknown_address_altered = 1;
2997 if (unknown_address_altered)
3000 for (i = 0; i < loop_store_mems_idx; i++)
3001 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
3002 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
3004 /* We are storing at the same address as previously noted. Save the
3006 if (GET_MODE_SIZE (GET_MODE (x))
3007 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i])))
3008 loop_store_mems[i] = x;
3012 if (i == NUM_STORES)
3013 unknown_address_altered = 1;
3015 else if (i == loop_store_mems_idx)
3016 loop_store_mems[loop_store_mems_idx++] = x;
3019 /* Return nonzero if the rtx X is invariant over the current loop.
3021 The value is 2 if we refer to something only conditionally invariant.
3023 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3024 Otherwise, a memory ref is invariant if it does not conflict with
3025 anything stored in `loop_store_mems'. */
3032 register enum rtx_code code;
3034 int conditional = 0;
3038 code = GET_CODE (x);
3048 /* A LABEL_REF is normally invariant, however, if we are unrolling
3049 loops, and this label is inside the loop, then it isn't invariant.
3050 This is because each unrolled copy of the loop body will have
3051 a copy of this label. If this was invariant, then an insn loading
3052 the address of this label into a register might get moved outside
3053 the loop, and then each loop body would end up using the same label.
3055 We don't know the loop bounds here though, so just fail for all
3057 if (flag_unroll_loops)
3064 case UNSPEC_VOLATILE:
3068 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3069 since the reg might be set by initialization within the loop. */
3071 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3072 || x == arg_pointer_rtx)
3073 && ! current_function_has_nonlocal_goto)
3077 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3080 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3083 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3086 /* Volatile memory references must be rejected. Do this before
3087 checking for read-only items, so that volatile read-only items
3088 will be rejected also. */
3089 if (MEM_VOLATILE_P (x))
3092 /* Read-only items (such as constants in a constant pool) are
3093 invariant if their address is. */
3094 if (RTX_UNCHANGING_P (x))
3097 /* If we filled the table (or had a subroutine call), any location
3098 in memory could have been clobbered. */
3099 if (unknown_address_altered)
3102 /* See if there is any dependence between a store and this load. */
3103 for (i = loop_store_mems_idx - 1; i >= 0; i--)
3104 if (true_dependence (loop_store_mems[i], VOIDmode, x, rtx_varies_p))
3107 /* It's not invalidated by a store in memory
3108 but we must still verify the address is invariant. */
3112 /* Don't mess with insns declared volatile. */
3113 if (MEM_VOLATILE_P (x))
3121 fmt = GET_RTX_FORMAT (code);
3122 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3126 int tem = invariant_p (XEXP (x, i));
3132 else if (fmt[i] == 'E')
3135 for (j = 0; j < XVECLEN (x, i); j++)
3137 int tem = invariant_p (XVECEXP (x, i, j));
3147 return 1 + conditional;
3151 /* Return nonzero if all the insns in the loop that set REG
3152 are INSN and the immediately following insns,
3153 and if each of those insns sets REG in an invariant way
3154 (not counting uses of REG in them).
3156 The value is 2 if some of these insns are only conditionally invariant.
3158 We assume that INSN itself is the first set of REG
3159 and that its source is invariant. */
3162 consec_sets_invariant_p (reg, n_sets, insn)
3166 register rtx p = insn;
3167 register int regno = REGNO (reg);
3169 /* Number of sets we have to insist on finding after INSN. */
3170 int count = n_sets - 1;
3171 int old = VARRAY_INT (set_in_loop, regno);
3175 /* If N_SETS hit the limit, we can't rely on its value. */
3179 VARRAY_INT (set_in_loop, regno) = 0;
3183 register enum rtx_code code;
3187 code = GET_CODE (p);
3189 /* If library call, skip to end of it. */
3190 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3195 && (set = single_set (p))
3196 && GET_CODE (SET_DEST (set)) == REG
3197 && REGNO (SET_DEST (set)) == regno)
3199 this = invariant_p (SET_SRC (set));
3202 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3204 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3205 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3207 this = (CONSTANT_P (XEXP (temp, 0))
3208 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3209 && invariant_p (XEXP (temp, 0))));
3216 else if (code != NOTE)
3218 VARRAY_INT (set_in_loop, regno) = old;
3223 VARRAY_INT (set_in_loop, regno) = old;
3224 /* If invariant_p ever returned 2, we return 2. */
3225 return 1 + (value & 2);
3229 /* I don't think this condition is sufficient to allow INSN
3230 to be moved, so we no longer test it. */
3232 /* Return 1 if all insns in the basic block of INSN and following INSN
3233 that set REG are invariant according to TABLE. */
3236 all_sets_invariant_p (reg, insn, table)
3240 register rtx p = insn;
3241 register int regno = REGNO (reg);
3245 register enum rtx_code code;
3247 code = GET_CODE (p);
3248 if (code == CODE_LABEL || code == JUMP_INSN)
3250 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3251 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3252 && REGNO (SET_DEST (PATTERN (p))) == regno)
3254 if (!invariant_p (SET_SRC (PATTERN (p)), table))
3261 /* Look at all uses (not sets) of registers in X. For each, if it is
3262 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3263 a different insn, set USAGE[REGNO] to const0_rtx. */
3266 find_single_use_in_loop (insn, x, usage)
3271 enum rtx_code code = GET_CODE (x);
3272 char *fmt = GET_RTX_FORMAT (code);
3276 VARRAY_RTX (usage, REGNO (x))
3277 = (VARRAY_RTX (usage, REGNO (x)) != 0
3278 && VARRAY_RTX (usage, REGNO (x)) != insn)
3279 ? const0_rtx : insn;
3281 else if (code == SET)
3283 /* Don't count SET_DEST if it is a REG; otherwise count things
3284 in SET_DEST because if a register is partially modified, it won't
3285 show up as a potential movable so we don't care how USAGE is set
3287 if (GET_CODE (SET_DEST (x)) != REG)
3288 find_single_use_in_loop (insn, SET_DEST (x), usage);
3289 find_single_use_in_loop (insn, SET_SRC (x), usage);
3292 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3294 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3295 find_single_use_in_loop (insn, XEXP (x, i), usage);
3296 else if (fmt[i] == 'E')
3297 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3298 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3302 /* Count and record any set in X which is contained in INSN. Update
3303 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3306 count_one_set (insn, x, may_not_move, last_set)
3308 varray_type may_not_move;
3311 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3312 /* Don't move a reg that has an explicit clobber.
3313 It's not worth the pain to try to do it correctly. */
3314 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3316 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3318 rtx dest = SET_DEST (x);
3319 while (GET_CODE (dest) == SUBREG
3320 || GET_CODE (dest) == ZERO_EXTRACT
3321 || GET_CODE (dest) == SIGN_EXTRACT
3322 || GET_CODE (dest) == STRICT_LOW_PART)
3323 dest = XEXP (dest, 0);
3324 if (GET_CODE (dest) == REG)
3326 register int regno = REGNO (dest);
3327 /* If this is the first setting of this reg
3328 in current basic block, and it was set before,
3329 it must be set in two basic blocks, so it cannot
3330 be moved out of the loop. */
3331 if (VARRAY_INT (set_in_loop, regno) > 0
3332 && last_set[regno] == 0)
3333 VARRAY_CHAR (may_not_move, regno) = 1;
3334 /* If this is not first setting in current basic block,
3335 see if reg was used in between previous one and this.
3336 If so, neither one can be moved. */
3337 if (last_set[regno] != 0
3338 && reg_used_between_p (dest, last_set[regno], insn))
3339 VARRAY_CHAR (may_not_move, regno) = 1;
3340 if (VARRAY_INT (set_in_loop, regno) < 127)
3341 ++VARRAY_INT (set_in_loop, regno);
3342 last_set[regno] = insn;
3347 /* Increment SET_IN_LOOP at the index of each register
3348 that is modified by an insn between FROM and TO.
3349 If the value of an element of SET_IN_LOOP becomes 127 or more,
3350 stop incrementing it, to avoid overflow.
3352 Store in SINGLE_USAGE[I] the single insn in which register I is
3353 used, if it is only used once. Otherwise, it is set to 0 (for no
3354 uses) or const0_rtx for more than one use. This parameter may be zero,
3355 in which case this processing is not done.
3357 Store in *COUNT_PTR the number of actual instruction
3358 in the loop. We use this to decide what is worth moving out. */
3360 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3361 In that case, it is the insn that last set reg n. */
3364 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3365 register rtx from, to;
3366 varray_type may_not_move;
3367 varray_type single_usage;
3371 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3373 register int count = 0;
3375 bzero ((char *) last_set, nregs * sizeof (rtx));
3376 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3378 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3382 /* If requested, record registers that have exactly one use. */
3385 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3387 /* Include uses in REG_EQUAL notes. */
3388 if (REG_NOTES (insn))
3389 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3392 if (GET_CODE (PATTERN (insn)) == SET
3393 || GET_CODE (PATTERN (insn)) == CLOBBER)
3394 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3395 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3398 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3399 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3400 may_not_move, last_set);
3404 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3405 bzero ((char *) last_set, nregs * sizeof (rtx));
3410 /* Given a loop that is bounded by LOOP_START and LOOP_END
3411 and that is entered at SCAN_START,
3412 return 1 if the register set in SET contained in insn INSN is used by
3413 any insn that precedes INSN in cyclic order starting
3414 from the loop entry point.
3416 We don't want to use INSN_LUID here because if we restrict INSN to those
3417 that have a valid INSN_LUID, it means we cannot move an invariant out
3418 from an inner loop past two loops. */
3421 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3422 rtx set, insn, loop_start, scan_start, loop_end;
3424 rtx reg = SET_DEST (set);
3427 /* Scan forward checking for register usage. If we hit INSN, we
3428 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3429 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3431 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3432 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3442 /* A "basic induction variable" or biv is a pseudo reg that is set
3443 (within this loop) only by incrementing or decrementing it. */
3444 /* A "general induction variable" or giv is a pseudo reg whose
3445 value is a linear function of a biv. */
3447 /* Bivs are recognized by `basic_induction_var';
3448 Givs by `general_induction_var'. */
3450 /* Indexed by register number, indicates whether or not register is an
3451 induction variable, and if so what type. */
3453 enum iv_mode *reg_iv_type;
3455 /* Indexed by register number, contains pointer to `struct induction'
3456 if register is an induction variable. This holds general info for
3457 all induction variables. */
3459 struct induction **reg_iv_info;
3461 /* Indexed by register number, contains pointer to `struct iv_class'
3462 if register is a basic induction variable. This holds info describing
3463 the class (a related group) of induction variables that the biv belongs
3466 struct iv_class **reg_biv_class;
3468 /* The head of a list which links together (via the next field)
3469 every iv class for the current loop. */
3471 struct iv_class *loop_iv_list;
3473 /* Communication with routines called via `note_stores'. */
3475 static rtx note_insn;
3477 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3479 static rtx addr_placeholder;
3481 /* ??? Unfinished optimizations, and possible future optimizations,
3482 for the strength reduction code. */
3484 /* ??? The interaction of biv elimination, and recognition of 'constant'
3485 bivs, may cause problems. */
3487 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3488 performance problems.
3490 Perhaps don't eliminate things that can be combined with an addressing
3491 mode. Find all givs that have the same biv, mult_val, and add_val;
3492 then for each giv, check to see if its only use dies in a following
3493 memory address. If so, generate a new memory address and check to see
3494 if it is valid. If it is valid, then store the modified memory address,
3495 otherwise, mark the giv as not done so that it will get its own iv. */
3497 /* ??? Could try to optimize branches when it is known that a biv is always
3500 /* ??? When replace a biv in a compare insn, we should replace with closest
3501 giv so that an optimized branch can still be recognized by the combiner,
3502 e.g. the VAX acb insn. */
3504 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3505 was rerun in loop_optimize whenever a register was added or moved.
3506 Also, some of the optimizations could be a little less conservative. */
3508 /* Perform strength reduction and induction variable elimination.
3510 Pseudo registers created during this function will be beyond the last
3511 valid index in several tables including n_times_set and regno_last_uid.
3512 This does not cause a problem here, because the added registers cannot be
3513 givs outside of their loop, and hence will never be reconsidered.
3514 But scan_loop must check regnos to make sure they are in bounds.
3516 SCAN_START is the first instruction in the loop, as the loop would
3517 actually be executed. END is the NOTE_INSN_LOOP_END. LOOP_TOP is
3518 the first instruction in the loop, as it is layed out in the
3519 instruction stream. LOOP_START is the NOTE_INSN_LOOP_BEG. */
3522 strength_reduce (scan_start, end, loop_top, insn_count,
3523 loop_start, loop_end, unroll_p, bct_p)
3530 int unroll_p, bct_p ATTRIBUTE_UNUSED;
3537 /* This is 1 if current insn is not executed at least once for every loop
3539 int not_every_iteration = 0;
3540 /* This is 1 if current insn may be executed more than once for every
3542 int maybe_multiple = 0;
3543 /* Temporary list pointers for traversing loop_iv_list. */
3544 struct iv_class *bl, **backbl;
3545 /* Ratio of extra register life span we can justify
3546 for saving an instruction. More if loop doesn't call subroutines
3547 since in that case saving an insn makes more difference
3548 and more registers are available. */
3549 /* ??? could set this to last value of threshold in move_movables */
3550 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3551 /* Map of pseudo-register replacements. */
3555 rtx end_insert_before;
3557 struct loop_info loop_iteration_info;
3558 struct loop_info *loop_info = &loop_iteration_info;
3560 /* If scan_start points to the loop exit test, we have to be wary of
3561 subversive use of gotos inside expression statements. */
3562 if (prev_nonnote_insn (scan_start) != prev_nonnote_insn (loop_start))
3563 maybe_multiple = back_branch_in_range_p (scan_start, loop_start, loop_end);
3565 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3566 * sizeof (enum iv_mode));
3567 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode));
3568 reg_iv_info = (struct induction **)
3569 alloca (max_reg_before_loop * sizeof (struct induction *));
3570 bzero ((char *) reg_iv_info, (max_reg_before_loop
3571 * sizeof (struct induction *)));
3572 reg_biv_class = (struct iv_class **)
3573 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3574 bzero ((char *) reg_biv_class, (max_reg_before_loop
3575 * sizeof (struct iv_class *)));
3578 addr_placeholder = gen_reg_rtx (Pmode);
3580 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3581 must be put before this insn, so that they will appear in the right
3582 order (i.e. loop order).
3584 If loop_end is the end of the current function, then emit a
3585 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3587 if (NEXT_INSN (loop_end) != 0)
3588 end_insert_before = NEXT_INSN (loop_end);
3590 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3592 /* Scan through loop to find all possible bivs. */
3594 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
3596 p = next_insn_in_loop (p, scan_start, end, loop_top))
3598 if (GET_CODE (p) == INSN
3599 && (set = single_set (p))
3600 && GET_CODE (SET_DEST (set)) == REG)
3602 dest_reg = SET_DEST (set);
3603 if (REGNO (dest_reg) < max_reg_before_loop
3604 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3605 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3607 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3608 dest_reg, p, &inc_val, &mult_val))
3610 /* It is a possible basic induction variable.
3611 Create and initialize an induction structure for it. */
3614 = (struct induction *) alloca (sizeof (struct induction));
3616 record_biv (v, p, dest_reg, inc_val, mult_val,
3617 not_every_iteration, maybe_multiple);
3618 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3620 else if (REGNO (dest_reg) < max_reg_before_loop)
3621 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3625 /* Past CODE_LABEL, we get to insns that may be executed multiple
3626 times. The only way we can be sure that they can't is if every
3627 jump insn between here and the end of the loop either
3628 returns, exits the loop, is a jump to a location that is still
3629 behind the label, or is a jump to the loop start. */
3631 if (GET_CODE (p) == CODE_LABEL)
3639 insn = NEXT_INSN (insn);
3640 if (insn == scan_start)
3648 if (insn == scan_start)
3652 if (GET_CODE (insn) == JUMP_INSN
3653 && GET_CODE (PATTERN (insn)) != RETURN
3654 && (! condjump_p (insn)
3655 || (JUMP_LABEL (insn) != 0
3656 && JUMP_LABEL (insn) != scan_start
3657 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3658 || (INSN_UID (p) < max_uid_for_loop
3659 ? (INSN_LUID (JUMP_LABEL (insn))
3661 : (INSN_UID (insn) >= max_uid_for_loop
3662 || (INSN_LUID (JUMP_LABEL (insn))
3663 < INSN_LUID (insn))))))))
3671 /* Past a jump, we get to insns for which we can't count
3672 on whether they will be executed during each iteration. */
3673 /* This code appears twice in strength_reduce. There is also similar
3674 code in scan_loop. */
3675 if (GET_CODE (p) == JUMP_INSN
3676 /* If we enter the loop in the middle, and scan around to the
3677 beginning, don't set not_every_iteration for that.
3678 This can be any kind of jump, since we want to know if insns
3679 will be executed if the loop is executed. */
3680 && ! (JUMP_LABEL (p) == loop_top
3681 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3682 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3686 /* If this is a jump outside the loop, then it also doesn't
3687 matter. Check to see if the target of this branch is on the
3688 loop_number_exits_labels list. */
3690 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3692 label = LABEL_NEXTREF (label))
3693 if (XEXP (label, 0) == JUMP_LABEL (p))
3697 not_every_iteration = 1;
3700 else if (GET_CODE (p) == NOTE)
3702 /* At the virtual top of a converted loop, insns are again known to
3703 be executed each iteration: logically, the loop begins here
3704 even though the exit code has been duplicated. */
3705 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3706 not_every_iteration = 0;
3707 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3709 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3713 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3714 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3715 or not an insn is known to be executed each iteration of the
3716 loop, whether or not any iterations are known to occur.
3718 Therefore, if we have just passed a label and have no more labels
3719 between here and the test insn of the loop, we know these insns
3720 will be executed each iteration. */
3722 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3723 && no_labels_between_p (p, loop_end))
3724 not_every_iteration = 0;
3727 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3728 Make a sanity check against n_times_set. */
3729 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3731 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3732 /* Above happens if register modified by subreg, etc. */
3733 /* Make sure it is not recognized as a basic induction var: */
3734 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3735 /* If never incremented, it is invariant that we decided not to
3736 move. So leave it alone. */
3737 || ! bl->incremented)
3739 if (loop_dump_stream)
3740 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3742 (reg_iv_type[bl->regno] != BASIC_INDUCT
3743 ? "not induction variable"
3744 : (! bl->incremented ? "never incremented"
3747 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3754 if (loop_dump_stream)
3755 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3759 /* Exit if there are no bivs. */
3762 /* Can still unroll the loop anyways, but indicate that there is no
3763 strength reduction info available. */
3765 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
3771 /* Find initial value for each biv by searching backwards from loop_start,
3772 halting at first label. Also record any test condition. */
3775 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3779 if (GET_CODE (p) == CALL_INSN)
3782 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3783 || GET_CODE (p) == CALL_INSN)
3784 note_stores (PATTERN (p), record_initial);
3786 /* Record any test of a biv that branches around the loop if no store
3787 between it and the start of loop. We only care about tests with
3788 constants and registers and only certain of those. */
3789 if (GET_CODE (p) == JUMP_INSN
3790 && JUMP_LABEL (p) != 0
3791 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3792 && (test = get_condition_for_loop (p)) != 0
3793 && GET_CODE (XEXP (test, 0)) == REG
3794 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3795 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3796 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3797 && bl->init_insn == 0)
3799 /* If an NE test, we have an initial value! */
3800 if (GET_CODE (test) == NE)
3803 bl->init_set = gen_rtx_SET (VOIDmode,
3804 XEXP (test, 0), XEXP (test, 1));
3807 bl->initial_test = test;
3811 /* Look at the each biv and see if we can say anything better about its
3812 initial value from any initializing insns set up above. (This is done
3813 in two passes to avoid missing SETs in a PARALLEL.) */
3814 for (bl = loop_iv_list; bl; bl = bl->next)
3819 if (! bl->init_insn)
3822 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3823 is a constant, use the value of that. */
3824 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
3825 && CONSTANT_P (XEXP (note, 0)))
3826 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
3827 && CONSTANT_P (XEXP (note, 0))))
3828 src = XEXP (note, 0);
3830 src = SET_SRC (bl->init_set);
3832 if (loop_dump_stream)
3833 fprintf (loop_dump_stream,
3834 "Biv %d initialized at insn %d: initial value ",
3835 bl->regno, INSN_UID (bl->init_insn));
3837 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3838 || GET_MODE (src) == VOIDmode)
3839 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3841 bl->initial_value = src;
3843 if (loop_dump_stream)
3845 if (GET_CODE (src) == CONST_INT)
3847 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
3848 fputc ('\n', loop_dump_stream);
3852 print_rtl (loop_dump_stream, src);
3853 fprintf (loop_dump_stream, "\n");
3859 /* Biv initial value is not simple move,
3860 so let it keep initial value of "itself". */
3862 if (loop_dump_stream)
3863 fprintf (loop_dump_stream, "is complex\n");
3867 /* Search the loop for general induction variables. */
3869 /* A register is a giv if: it is only set once, it is a function of a
3870 biv and a constant (or invariant), and it is not a biv. */
3872 not_every_iteration = 0;
3878 /* At end of a straight-in loop, we are done.
3879 At end of a loop entered at the bottom, scan the top. */
3880 if (p == scan_start)
3888 if (p == scan_start)
3892 /* Look for a general induction variable in a register. */
3893 if (GET_CODE (p) == INSN
3894 && (set = single_set (p))
3895 && GET_CODE (SET_DEST (set)) == REG
3896 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
3903 rtx last_consec_insn;
3905 dest_reg = SET_DEST (set);
3906 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3909 if (/* SET_SRC is a giv. */
3910 (general_induction_var (SET_SRC (set), &src_reg, &add_val,
3911 &mult_val, 0, &benefit)
3912 /* Equivalent expression is a giv. */
3913 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
3914 && general_induction_var (XEXP (regnote, 0), &src_reg,
3915 &add_val, &mult_val, 0,
3917 /* Don't try to handle any regs made by loop optimization.
3918 We have nothing on them in regno_first_uid, etc. */
3919 && REGNO (dest_reg) < max_reg_before_loop
3920 /* Don't recognize a BASIC_INDUCT_VAR here. */
3921 && dest_reg != src_reg
3922 /* This must be the only place where the register is set. */
3923 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
3924 /* or all sets must be consecutive and make a giv. */
3925 || (benefit = consec_sets_giv (benefit, p,
3927 &add_val, &mult_val,
3928 &last_consec_insn))))
3931 = (struct induction *) alloca (sizeof (struct induction));
3933 /* If this is a library call, increase benefit. */
3934 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
3935 benefit += libcall_benefit (p);
3937 /* Skip the consecutive insns, if there are any. */
3938 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
3939 p = last_consec_insn;
3941 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3942 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
3948 #ifndef DONT_REDUCE_ADDR
3949 /* Look for givs which are memory addresses. */
3950 /* This resulted in worse code on a VAX 8600. I wonder if it
3952 if (GET_CODE (p) == INSN)
3953 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3957 /* Update the status of whether giv can derive other givs. This can
3958 change when we pass a label or an insn that updates a biv. */
3959 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3960 || GET_CODE (p) == CODE_LABEL)
3961 update_giv_derive (p);
3963 /* Past a jump, we get to insns for which we can't count
3964 on whether they will be executed during each iteration. */
3965 /* This code appears twice in strength_reduce. There is also similar
3966 code in scan_loop. */
3967 if (GET_CODE (p) == JUMP_INSN
3968 /* If we enter the loop in the middle, and scan around to the
3969 beginning, don't set not_every_iteration for that.
3970 This can be any kind of jump, since we want to know if insns
3971 will be executed if the loop is executed. */
3972 && ! (JUMP_LABEL (p) == loop_top
3973 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3974 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3978 /* If this is a jump outside the loop, then it also doesn't
3979 matter. Check to see if the target of this branch is on the
3980 loop_number_exits_labels list. */
3982 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3984 label = LABEL_NEXTREF (label))
3985 if (XEXP (label, 0) == JUMP_LABEL (p))
3989 not_every_iteration = 1;
3992 else if (GET_CODE (p) == NOTE)
3994 /* At the virtual top of a converted loop, insns are again known to
3995 be executed each iteration: logically, the loop begins here
3996 even though the exit code has been duplicated. */
3997 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3998 not_every_iteration = 0;
3999 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
4001 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
4005 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
4006 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
4007 or not an insn is known to be executed each iteration of the
4008 loop, whether or not any iterations are known to occur.
4010 Therefore, if we have just passed a label and have no more labels
4011 between here and the test insn of the loop, we know these insns
4012 will be executed each iteration. */
4014 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
4015 && no_labels_between_p (p, loop_end))
4016 not_every_iteration = 0;
4019 /* Try to calculate and save the number of loop iterations. This is
4020 set to zero if the actual number can not be calculated. This must
4021 be called after all giv's have been identified, since otherwise it may
4022 fail if the iteration variable is a giv. */
4024 loop_iterations (loop_start, loop_end, loop_info);
4026 /* Now for each giv for which we still don't know whether or not it is
4027 replaceable, check to see if it is replaceable because its final value
4028 can be calculated. This must be done after loop_iterations is called,
4029 so that final_giv_value will work correctly. */
4031 for (bl = loop_iv_list; bl; bl = bl->next)
4033 struct induction *v;
4035 for (v = bl->giv; v; v = v->next_iv)
4036 if (! v->replaceable && ! v->not_replaceable)
4037 check_final_value (v, loop_start, loop_end, loop_info->n_iterations);
4040 /* Try to prove that the loop counter variable (if any) is always
4041 nonnegative; if so, record that fact with a REG_NONNEG note
4042 so that "decrement and branch until zero" insn can be used. */
4043 check_dbra_loop (loop_end, insn_count, loop_start, loop_info);
4045 /* Create reg_map to hold substitutions for replaceable giv regs. */
4046 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
4047 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
4049 /* Examine each iv class for feasibility of strength reduction/induction
4050 variable elimination. */
4052 for (bl = loop_iv_list; bl; bl = bl->next)
4054 struct induction *v;
4057 rtx final_value = 0;
4059 /* Test whether it will be possible to eliminate this biv
4060 provided all givs are reduced. This is possible if either
4061 the reg is not used outside the loop, or we can compute
4062 what its final value will be.
4064 For architectures with a decrement_and_branch_until_zero insn,
4065 don't do this if we put a REG_NONNEG note on the endtest for
4068 /* Compare against bl->init_insn rather than loop_start.
4069 We aren't concerned with any uses of the biv between
4070 init_insn and loop_start since these won't be affected
4071 by the value of the biv elsewhere in the function, so
4072 long as init_insn doesn't use the biv itself.
4073 March 14, 1989 -- self@bayes.arc.nasa.gov */
4075 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4077 && INSN_UID (bl->init_insn) < max_uid_for_loop
4078 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4079 #ifdef HAVE_decrement_and_branch_until_zero
4082 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4083 || ((final_value = final_biv_value (bl, loop_start, loop_end,
4084 loop_info->n_iterations))
4085 #ifdef HAVE_decrement_and_branch_until_zero
4089 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
4090 threshold, insn_count);
4093 if (loop_dump_stream)
4095 fprintf (loop_dump_stream,
4096 "Cannot eliminate biv %d.\n",
4098 fprintf (loop_dump_stream,
4099 "First use: insn %d, last use: insn %d.\n",
4100 REGNO_FIRST_UID (bl->regno),
4101 REGNO_LAST_UID (bl->regno));
4105 /* Combine all giv's for this iv_class. */
4108 /* This will be true at the end, if all givs which depend on this
4109 biv have been strength reduced.
4110 We can't (currently) eliminate the biv unless this is so. */
4113 /* Check each giv in this class to see if we will benefit by reducing
4114 it. Skip giv's combined with others. */
4115 for (v = bl->giv; v; v = v->next_iv)
4117 struct induction *tv;
4119 if (v->ignore || v->same)
4122 benefit = v->benefit;
4124 /* Reduce benefit if not replaceable, since we will insert
4125 a move-insn to replace the insn that calculates this giv.
4126 Don't do this unless the giv is a user variable, since it
4127 will often be marked non-replaceable because of the duplication
4128 of the exit code outside the loop. In such a case, the copies
4129 we insert are dead and will be deleted. So they don't have
4130 a cost. Similar situations exist. */
4131 /* ??? The new final_[bg]iv_value code does a much better job
4132 of finding replaceable giv's, and hence this code may no longer
4134 if (! v->replaceable && ! bl->eliminable
4135 && REG_USERVAR_P (v->dest_reg))
4136 benefit -= copy_cost;
4138 /* Decrease the benefit to count the add-insns that we will
4139 insert to increment the reduced reg for the giv. */
4140 benefit -= add_cost * bl->biv_count;
4142 /* Decide whether to strength-reduce this giv or to leave the code
4143 unchanged (recompute it from the biv each time it is used).
4144 This decision can be made independently for each giv. */
4147 /* Attempt to guess whether autoincrement will handle some of the
4148 new add insns; if so, increase BENEFIT (undo the subtraction of
4149 add_cost that was done above). */
4150 if (v->giv_type == DEST_ADDR
4151 && GET_CODE (v->mult_val) == CONST_INT)
4153 if (HAVE_POST_INCREMENT
4154 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4155 benefit += add_cost * bl->biv_count;
4156 else if (HAVE_PRE_INCREMENT
4157 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4158 benefit += add_cost * bl->biv_count;
4159 else if (HAVE_POST_DECREMENT
4160 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4161 benefit += add_cost * bl->biv_count;
4162 else if (HAVE_PRE_DECREMENT
4163 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4164 benefit += add_cost * bl->biv_count;
4168 /* If an insn is not to be strength reduced, then set its ignore
4169 flag, and clear all_reduced. */
4171 /* A giv that depends on a reversed biv must be reduced if it is
4172 used after the loop exit, otherwise, it would have the wrong
4173 value after the loop exit. To make it simple, just reduce all
4174 of such giv's whether or not we know they are used after the loop
4177 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4180 if (loop_dump_stream)
4181 fprintf (loop_dump_stream,
4182 "giv of insn %d not worth while, %d vs %d.\n",
4184 v->lifetime * threshold * benefit, insn_count);
4190 /* Check that we can increment the reduced giv without a
4191 multiply insn. If not, reject it. */
4193 for (tv = bl->biv; tv; tv = tv->next_iv)
4194 if (tv->mult_val == const1_rtx
4195 && ! product_cheap_p (tv->add_val, v->mult_val))
4197 if (loop_dump_stream)
4198 fprintf (loop_dump_stream,
4199 "giv of insn %d: would need a multiply.\n",
4200 INSN_UID (v->insn));
4208 /* Reduce each giv that we decided to reduce. */
4210 for (v = bl->giv; v; v = v->next_iv)
4212 struct induction *tv;
4213 if (! v->ignore && v->same == 0)
4215 int auto_inc_opt = 0;
4217 v->new_reg = gen_reg_rtx (v->mode);
4220 /* If the target has auto-increment addressing modes, and
4221 this is an address giv, then try to put the increment
4222 immediately after its use, so that flow can create an
4223 auto-increment addressing mode. */
4224 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4225 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4226 /* We don't handle reversed biv's because bl->biv->insn
4227 does not have a valid INSN_LUID. */
4229 && v->always_executed && ! v->maybe_multiple
4230 && INSN_UID (v->insn) < max_uid_for_loop)
4232 /* If other giv's have been combined with this one, then
4233 this will work only if all uses of the other giv's occur
4234 before this giv's insn. This is difficult to check.
4236 We simplify this by looking for the common case where
4237 there is one DEST_REG giv, and this giv's insn is the
4238 last use of the dest_reg of that DEST_REG giv. If the
4239 increment occurs after the address giv, then we can
4240 perform the optimization. (Otherwise, the increment
4241 would have to go before other_giv, and we would not be
4242 able to combine it with the address giv to get an
4243 auto-inc address.) */
4244 if (v->combined_with)
4246 struct induction *other_giv = 0;
4248 for (tv = bl->giv; tv; tv = tv->next_iv)
4256 if (! tv && other_giv
4257 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4258 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4259 == INSN_UID (v->insn))
4260 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4263 /* Check for case where increment is before the address
4264 giv. Do this test in "loop order". */
4265 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4266 && (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4267 || (INSN_LUID (bl->biv->insn)
4268 > INSN_LUID (scan_start))))
4269 || (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4270 && (INSN_LUID (scan_start)
4271 < INSN_LUID (bl->biv->insn))))
4280 /* We can't put an insn immediately after one setting
4281 cc0, or immediately before one using cc0. */
4282 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4283 || (auto_inc_opt == -1
4284 && (prev = prev_nonnote_insn (v->insn)) != 0
4285 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
4286 && sets_cc0_p (PATTERN (prev))))
4292 v->auto_inc_opt = 1;
4296 /* For each place where the biv is incremented, add an insn
4297 to increment the new, reduced reg for the giv. */
4298 for (tv = bl->biv; tv; tv = tv->next_iv)
4303 insert_before = tv->insn;
4304 else if (auto_inc_opt == 1)
4305 insert_before = NEXT_INSN (v->insn);
4307 insert_before = v->insn;
4309 if (tv->mult_val == const1_rtx)
4310 emit_iv_add_mult (tv->add_val, v->mult_val,
4311 v->new_reg, v->new_reg, insert_before);
4312 else /* tv->mult_val == const0_rtx */
4313 /* A multiply is acceptable here
4314 since this is presumed to be seldom executed. */
4315 emit_iv_add_mult (tv->add_val, v->mult_val,
4316 v->add_val, v->new_reg, insert_before);
4319 /* Add code at loop start to initialize giv's reduced reg. */
4321 emit_iv_add_mult (bl->initial_value, v->mult_val,
4322 v->add_val, v->new_reg, loop_start);
4326 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4329 For each giv register that can be reduced now: if replaceable,
4330 substitute reduced reg wherever the old giv occurs;
4331 else add new move insn "giv_reg = reduced_reg".
4333 Also check for givs whose first use is their definition and whose
4334 last use is the definition of another giv. If so, it is likely
4335 dead and should not be used to eliminate a biv. */
4336 for (v = bl->giv; v; v = v->next_iv)
4338 if (v->same && v->same->ignore)
4344 if (v->giv_type == DEST_REG
4345 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4347 struct induction *v1;
4349 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4350 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4354 /* Update expression if this was combined, in case other giv was
4357 v->new_reg = replace_rtx (v->new_reg,
4358 v->same->dest_reg, v->same->new_reg);
4360 if (v->giv_type == DEST_ADDR)
4361 /* Store reduced reg as the address in the memref where we found
4363 validate_change (v->insn, v->location, v->new_reg, 0);
4364 else if (v->replaceable)
4366 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4369 /* I can no longer duplicate the original problem. Perhaps
4370 this is unnecessary now? */
4372 /* Replaceable; it isn't strictly necessary to delete the old
4373 insn and emit a new one, because v->dest_reg is now dead.
4375 However, especially when unrolling loops, the special
4376 handling for (set REG0 REG1) in the second cse pass may
4377 make v->dest_reg live again. To avoid this problem, emit
4378 an insn to set the original giv reg from the reduced giv.
4379 We can not delete the original insn, since it may be part
4380 of a LIBCALL, and the code in flow that eliminates dead
4381 libcalls will fail if it is deleted. */
4382 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4388 /* Not replaceable; emit an insn to set the original giv reg from
4389 the reduced giv, same as above. */
4390 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4394 /* When a loop is reversed, givs which depend on the reversed
4395 biv, and which are live outside the loop, must be set to their
4396 correct final value. This insn is only needed if the giv is
4397 not replaceable. The correct final value is the same as the
4398 value that the giv starts the reversed loop with. */
4399 if (bl->reversed && ! v->replaceable)
4400 emit_iv_add_mult (bl->initial_value, v->mult_val,
4401 v->add_val, v->dest_reg, end_insert_before);
4402 else if (v->final_value)
4406 /* If the loop has multiple exits, emit the insn before the
4407 loop to ensure that it will always be executed no matter
4408 how the loop exits. Otherwise, emit the insn after the loop,
4409 since this is slightly more efficient. */
4410 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4411 insert_before = loop_start;
4413 insert_before = end_insert_before;
4414 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4418 /* If the insn to set the final value of the giv was emitted
4419 before the loop, then we must delete the insn inside the loop
4420 that sets it. If this is a LIBCALL, then we must delete
4421 every insn in the libcall. Note, however, that
4422 final_giv_value will only succeed when there are multiple
4423 exits if the giv is dead at each exit, hence it does not
4424 matter that the original insn remains because it is dead
4426 /* Delete the insn inside the loop that sets the giv since
4427 the giv is now set before (or after) the loop. */
4428 delete_insn (v->insn);
4432 if (loop_dump_stream)
4434 fprintf (loop_dump_stream, "giv at %d reduced to ",
4435 INSN_UID (v->insn));
4436 print_rtl (loop_dump_stream, v->new_reg);
4437 fprintf (loop_dump_stream, "\n");
4441 /* All the givs based on the biv bl have been reduced if they
4444 /* For each giv not marked as maybe dead that has been combined with a
4445 second giv, clear any "maybe dead" mark on that second giv.
4446 v->new_reg will either be or refer to the register of the giv it
4449 Doing this clearing avoids problems in biv elimination where a
4450 giv's new_reg is a complex value that can't be put in the insn but
4451 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4452 Since the register will be used in either case, we'd prefer it be
4453 used from the simpler giv. */
4455 for (v = bl->giv; v; v = v->next_iv)
4456 if (! v->maybe_dead && v->same)
4457 v->same->maybe_dead = 0;
4459 /* Try to eliminate the biv, if it is a candidate.
4460 This won't work if ! all_reduced,
4461 since the givs we planned to use might not have been reduced.
4463 We have to be careful that we didn't initially think we could eliminate
4464 this biv because of a giv that we now think may be dead and shouldn't
4465 be used as a biv replacement.
4467 Also, there is the possibility that we may have a giv that looks
4468 like it can be used to eliminate a biv, but the resulting insn
4469 isn't valid. This can happen, for example, on the 88k, where a
4470 JUMP_INSN can compare a register only with zero. Attempts to
4471 replace it with a compare with a constant will fail.
4473 Note that in cases where this call fails, we may have replaced some
4474 of the occurrences of the biv with a giv, but no harm was done in
4475 doing so in the rare cases where it can occur. */
4477 if (all_reduced == 1 && bl->eliminable
4478 && maybe_eliminate_biv (bl, loop_start, end, 1,
4479 threshold, insn_count))
4482 /* ?? If we created a new test to bypass the loop entirely,
4483 or otherwise drop straight in, based on this test, then
4484 we might want to rewrite it also. This way some later
4485 pass has more hope of removing the initialization of this
4488 /* If final_value != 0, then the biv may be used after loop end
4489 and we must emit an insn to set it just in case.
4491 Reversed bivs already have an insn after the loop setting their
4492 value, so we don't need another one. We can't calculate the
4493 proper final value for such a biv here anyways. */
4494 if (final_value != 0 && ! bl->reversed)
4498 /* If the loop has multiple exits, emit the insn before the
4499 loop to ensure that it will always be executed no matter
4500 how the loop exits. Otherwise, emit the insn after the
4501 loop, since this is slightly more efficient. */
4502 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4503 insert_before = loop_start;
4505 insert_before = end_insert_before;
4507 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4512 /* Delete all of the instructions inside the loop which set
4513 the biv, as they are all dead. If is safe to delete them,
4514 because an insn setting a biv will never be part of a libcall. */
4515 /* However, deleting them will invalidate the regno_last_uid info,
4516 so keeping them around is more convenient. Final_biv_value
4517 will only succeed when there are multiple exits if the biv
4518 is dead at each exit, hence it does not matter that the original
4519 insn remains, because it is dead anyways. */
4520 for (v = bl->biv; v; v = v->next_iv)
4521 delete_insn (v->insn);
4524 if (loop_dump_stream)
4525 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4530 /* Go through all the instructions in the loop, making all the
4531 register substitutions scheduled in REG_MAP. */
4533 for (p = loop_start; p != end; p = NEXT_INSN (p))
4534 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4535 || GET_CODE (p) == CALL_INSN)
4537 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
4538 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
4542 /* Unroll loops from within strength reduction so that we can use the
4543 induction variable information that strength_reduce has already
4547 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
4550 #ifdef HAVE_decrement_and_branch_on_count
4551 /* Instrument the loop with BCT insn. */
4552 if (HAVE_decrement_and_branch_on_count && bct_p
4553 && flag_branch_on_count_reg)
4554 insert_bct (loop_start, loop_end, loop_info);
4555 #endif /* HAVE_decrement_and_branch_on_count */
4557 if (loop_dump_stream)
4558 fprintf (loop_dump_stream, "\n");
4561 /* Return 1 if X is a valid source for an initial value (or as value being
4562 compared against in an initial test).
4564 X must be either a register or constant and must not be clobbered between
4565 the current insn and the start of the loop.
4567 INSN is the insn containing X. */
4570 valid_initial_value_p (x, insn, call_seen, loop_start)
4579 /* Only consider pseudos we know about initialized in insns whose luids
4581 if (GET_CODE (x) != REG
4582 || REGNO (x) >= max_reg_before_loop)
4585 /* Don't use call-clobbered registers across a call which clobbers it. On
4586 some machines, don't use any hard registers at all. */
4587 if (REGNO (x) < FIRST_PSEUDO_REGISTER
4588 && (SMALL_REGISTER_CLASSES
4589 || (call_used_regs[REGNO (x)] && call_seen)))
4592 /* Don't use registers that have been clobbered before the start of the
4594 if (reg_set_between_p (x, insn, loop_start))
4600 /* Scan X for memory refs and check each memory address
4601 as a possible giv. INSN is the insn whose pattern X comes from.
4602 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
4603 every loop iteration. */
4606 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
4609 int not_every_iteration;
4610 rtx loop_start, loop_end;
4613 register enum rtx_code code;
4619 code = GET_CODE (x);
4643 /* This code used to disable creating GIVs with mult_val == 1 and
4644 add_val == 0. However, this leads to lost optimizations when
4645 it comes time to combine a set of related DEST_ADDR GIVs, since
4646 this one would not be seen. */
4648 if (general_induction_var (XEXP (x, 0), &src_reg, &add_val,
4649 &mult_val, 1, &benefit))
4651 /* Found one; record it. */
4653 = (struct induction *) oballoc (sizeof (struct induction));
4655 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4656 add_val, benefit, DEST_ADDR, not_every_iteration,
4657 &XEXP (x, 0), loop_start, loop_end);
4659 v->mem_mode = GET_MODE (x);
4668 /* Recursively scan the subexpressions for other mem refs. */
4670 fmt = GET_RTX_FORMAT (code);
4671 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4673 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4675 else if (fmt[i] == 'E')
4676 for (j = 0; j < XVECLEN (x, i); j++)
4677 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4678 loop_start, loop_end);
4681 /* Fill in the data about one biv update.
4682 V is the `struct induction' in which we record the biv. (It is
4683 allocated by the caller, with alloca.)
4684 INSN is the insn that sets it.
4685 DEST_REG is the biv's reg.
4687 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4688 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4689 being set to INC_VAL.
4691 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4692 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4693 can be executed more than once per iteration. If MAYBE_MULTIPLE
4694 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4695 executed exactly once per iteration. */
4698 record_biv (v, insn, dest_reg, inc_val, mult_val,
4699 not_every_iteration, maybe_multiple)
4700 struct induction *v;
4705 int not_every_iteration;
4708 struct iv_class *bl;
4711 v->src_reg = dest_reg;
4712 v->dest_reg = dest_reg;
4713 v->mult_val = mult_val;
4714 v->add_val = inc_val;
4715 v->mode = GET_MODE (dest_reg);
4716 v->always_computable = ! not_every_iteration;
4717 v->always_executed = ! not_every_iteration;
4718 v->maybe_multiple = maybe_multiple;
4720 /* Add this to the reg's iv_class, creating a class
4721 if this is the first incrementation of the reg. */
4723 bl = reg_biv_class[REGNO (dest_reg)];
4726 /* Create and initialize new iv_class. */
4728 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4730 bl->regno = REGNO (dest_reg);
4736 /* Set initial value to the reg itself. */
4737 bl->initial_value = dest_reg;
4738 /* We haven't seen the initializing insn yet */
4741 bl->initial_test = 0;
4742 bl->incremented = 0;
4746 bl->total_benefit = 0;
4748 /* Add this class to loop_iv_list. */
4749 bl->next = loop_iv_list;
4752 /* Put it in the array of biv register classes. */
4753 reg_biv_class[REGNO (dest_reg)] = bl;
4756 /* Update IV_CLASS entry for this biv. */
4757 v->next_iv = bl->biv;
4760 if (mult_val == const1_rtx)
4761 bl->incremented = 1;
4763 if (loop_dump_stream)
4765 fprintf (loop_dump_stream,
4766 "Insn %d: possible biv, reg %d,",
4767 INSN_UID (insn), REGNO (dest_reg));
4768 if (GET_CODE (inc_val) == CONST_INT)
4770 fprintf (loop_dump_stream, " const =");
4771 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
4772 fputc ('\n', loop_dump_stream);
4776 fprintf (loop_dump_stream, " const = ");
4777 print_rtl (loop_dump_stream, inc_val);
4778 fprintf (loop_dump_stream, "\n");
4783 /* Fill in the data about one giv.
4784 V is the `struct induction' in which we record the giv. (It is
4785 allocated by the caller, with alloca.)
4786 INSN is the insn that sets it.
4787 BENEFIT estimates the savings from deleting this insn.
4788 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4789 into a register or is used as a memory address.
4791 SRC_REG is the biv reg which the giv is computed from.
4792 DEST_REG is the giv's reg (if the giv is stored in a reg).
4793 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4794 LOCATION points to the place where this giv's value appears in INSN. */
4797 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4798 type, not_every_iteration, location, loop_start, loop_end)
4799 struct induction *v;
4803 rtx mult_val, add_val;
4806 int not_every_iteration;
4808 rtx loop_start, loop_end;
4810 struct induction *b;
4811 struct iv_class *bl;
4812 rtx set = single_set (insn);
4815 v->src_reg = src_reg;
4817 v->dest_reg = dest_reg;
4818 v->mult_val = mult_val;
4819 v->add_val = add_val;
4820 v->benefit = benefit;
4821 v->location = location;
4823 v->combined_with = 0;
4824 v->maybe_multiple = 0;
4826 v->derive_adjustment = 0;
4832 v->auto_inc_opt = 0;
4836 /* The v->always_computable field is used in update_giv_derive, to
4837 determine whether a giv can be used to derive another giv. For a
4838 DEST_REG giv, INSN computes a new value for the giv, so its value
4839 isn't computable if INSN insn't executed every iteration.
4840 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4841 it does not compute a new value. Hence the value is always computable
4842 regardless of whether INSN is executed each iteration. */
4844 if (type == DEST_ADDR)
4845 v->always_computable = 1;
4847 v->always_computable = ! not_every_iteration;
4849 v->always_executed = ! not_every_iteration;
4851 if (type == DEST_ADDR)
4853 v->mode = GET_MODE (*location);
4856 else /* type == DEST_REG */
4858 v->mode = GET_MODE (SET_DEST (set));
4860 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
4861 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
4863 /* If the lifetime is zero, it means that this register is
4864 really a dead store. So mark this as a giv that can be
4865 ignored. This will not prevent the biv from being eliminated. */
4866 if (v->lifetime == 0)
4869 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4870 reg_iv_info[REGNO (dest_reg)] = v;
4873 /* Add the giv to the class of givs computed from one biv. */
4875 bl = reg_biv_class[REGNO (src_reg)];
4878 v->next_iv = bl->giv;
4880 /* Don't count DEST_ADDR. This is supposed to count the number of
4881 insns that calculate givs. */
4882 if (type == DEST_REG)
4884 bl->total_benefit += benefit;
4887 /* Fatal error, biv missing for this giv? */
4890 if (type == DEST_ADDR)
4894 /* The giv can be replaced outright by the reduced register only if all
4895 of the following conditions are true:
4896 - the insn that sets the giv is always executed on any iteration
4897 on which the giv is used at all
4898 (there are two ways to deduce this:
4899 either the insn is executed on every iteration,
4900 or all uses follow that insn in the same basic block),
4901 - the giv is not used outside the loop
4902 - no assignments to the biv occur during the giv's lifetime. */
4904 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
4905 /* Previous line always fails if INSN was moved by loop opt. */
4906 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] < INSN_LUID (loop_end)
4907 && (! not_every_iteration
4908 || last_use_this_basic_block (dest_reg, insn)))
4910 /* Now check that there are no assignments to the biv within the
4911 giv's lifetime. This requires two separate checks. */
4913 /* Check each biv update, and fail if any are between the first
4914 and last use of the giv.
4916 If this loop contains an inner loop that was unrolled, then
4917 the insn modifying the biv may have been emitted by the loop
4918 unrolling code, and hence does not have a valid luid. Just
4919 mark the biv as not replaceable in this case. It is not very
4920 useful as a biv, because it is used in two different loops.
4921 It is very unlikely that we would be able to optimize the giv
4922 using this biv anyways. */
4925 for (b = bl->biv; b; b = b->next_iv)
4927 if (INSN_UID (b->insn) >= max_uid_for_loop
4928 || ((uid_luid[INSN_UID (b->insn)]
4929 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
4930 && (uid_luid[INSN_UID (b->insn)]
4931 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
4934 v->not_replaceable = 1;
4939 /* If there are any backwards branches that go from after the
4940 biv update to before it, then this giv is not replaceable. */
4942 for (b = bl->biv; b; b = b->next_iv)
4943 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
4946 v->not_replaceable = 1;
4952 /* May still be replaceable, we don't have enough info here to
4955 v->not_replaceable = 0;
4959 /* Record whether the add_val contains a const_int, for later use by
4964 v->no_const_addval = 1;
4965 if (tem == const0_rtx)
4967 else if (GET_CODE (tem) == CONST_INT)
4968 v->no_const_addval = 0;
4969 else if (GET_CODE (tem) == PLUS)
4973 if (GET_CODE (XEXP (tem, 0)) == PLUS)
4974 tem = XEXP (tem, 0);
4975 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
4976 tem = XEXP (tem, 1);
4980 if (GET_CODE (XEXP (tem, 1)) == CONST_INT)
4981 v->no_const_addval = 0;
4985 if (loop_dump_stream)
4987 if (type == DEST_REG)
4988 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4989 INSN_UID (insn), REGNO (dest_reg));
4991 fprintf (loop_dump_stream, "Insn %d: dest address",
4994 fprintf (loop_dump_stream, " src reg %d benefit %d",
4995 REGNO (src_reg), v->benefit);
4996 fprintf (loop_dump_stream, " lifetime %d",
5000 fprintf (loop_dump_stream, " replaceable");
5002 if (v->no_const_addval)
5003 fprintf (loop_dump_stream, " ncav");
5005 if (GET_CODE (mult_val) == CONST_INT)
5007 fprintf (loop_dump_stream, " mult ");
5008 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5012 fprintf (loop_dump_stream, " mult ");
5013 print_rtl (loop_dump_stream, mult_val);
5016 if (GET_CODE (add_val) == CONST_INT)
5018 fprintf (loop_dump_stream, " add ");
5019 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5023 fprintf (loop_dump_stream, " add ");
5024 print_rtl (loop_dump_stream, add_val);
5028 if (loop_dump_stream)
5029 fprintf (loop_dump_stream, "\n");
5034 /* All this does is determine whether a giv can be made replaceable because
5035 its final value can be calculated. This code can not be part of record_giv
5036 above, because final_giv_value requires that the number of loop iterations
5037 be known, and that can not be accurately calculated until after all givs
5038 have been identified. */
5041 check_final_value (v, loop_start, loop_end, n_iterations)
5042 struct induction *v;
5043 rtx loop_start, loop_end;
5044 unsigned HOST_WIDE_INT n_iterations;
5046 struct iv_class *bl;
5047 rtx final_value = 0;
5049 bl = reg_biv_class[REGNO (v->src_reg)];
5051 /* DEST_ADDR givs will never reach here, because they are always marked
5052 replaceable above in record_giv. */
5054 /* The giv can be replaced outright by the reduced register only if all
5055 of the following conditions are true:
5056 - the insn that sets the giv is always executed on any iteration
5057 on which the giv is used at all
5058 (there are two ways to deduce this:
5059 either the insn is executed on every iteration,
5060 or all uses follow that insn in the same basic block),
5061 - its final value can be calculated (this condition is different
5062 than the one above in record_giv)
5063 - no assignments to the biv occur during the giv's lifetime. */
5066 /* This is only called now when replaceable is known to be false. */
5067 /* Clear replaceable, so that it won't confuse final_giv_value. */
5071 if ((final_value = final_giv_value (v, loop_start, loop_end, n_iterations))
5072 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5074 int biv_increment_seen = 0;
5080 /* When trying to determine whether or not a biv increment occurs
5081 during the lifetime of the giv, we can ignore uses of the variable
5082 outside the loop because final_value is true. Hence we can not
5083 use regno_last_uid and regno_first_uid as above in record_giv. */
5085 /* Search the loop to determine whether any assignments to the
5086 biv occur during the giv's lifetime. Start with the insn
5087 that sets the giv, and search around the loop until we come
5088 back to that insn again.
5090 Also fail if there is a jump within the giv's lifetime that jumps
5091 to somewhere outside the lifetime but still within the loop. This
5092 catches spaghetti code where the execution order is not linear, and
5093 hence the above test fails. Here we assume that the giv lifetime
5094 does not extend from one iteration of the loop to the next, so as
5095 to make the test easier. Since the lifetime isn't known yet,
5096 this requires two loops. See also record_giv above. */
5098 last_giv_use = v->insn;
5104 p = NEXT_INSN (loop_start);
5108 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5109 || GET_CODE (p) == CALL_INSN)
5111 if (biv_increment_seen)
5113 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5116 v->not_replaceable = 1;
5120 else if (reg_set_p (v->src_reg, PATTERN (p)))
5121 biv_increment_seen = 1;
5122 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5127 /* Now that the lifetime of the giv is known, check for branches
5128 from within the lifetime to outside the lifetime if it is still
5138 p = NEXT_INSN (loop_start);
5139 if (p == last_giv_use)
5142 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5143 && LABEL_NAME (JUMP_LABEL (p))
5144 && ((INSN_UID (JUMP_LABEL (p)) >= max_uid_for_loop)
5145 || (INSN_UID (v->insn) >= max_uid_for_loop)
5146 || (INSN_UID (last_giv_use) >= max_uid_for_loop)
5147 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
5148 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
5149 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
5150 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
5153 v->not_replaceable = 1;
5155 if (loop_dump_stream)
5156 fprintf (loop_dump_stream,
5157 "Found branch outside giv lifetime.\n");
5164 /* If it is replaceable, then save the final value. */
5166 v->final_value = final_value;
5169 if (loop_dump_stream && v->replaceable)
5170 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5171 INSN_UID (v->insn), REGNO (v->dest_reg));
5174 /* Update the status of whether a giv can derive other givs.
5176 We need to do something special if there is or may be an update to the biv
5177 between the time the giv is defined and the time it is used to derive
5180 In addition, a giv that is only conditionally set is not allowed to
5181 derive another giv once a label has been passed.
5183 The cases we look at are when a label or an update to a biv is passed. */
5186 update_giv_derive (p)
5189 struct iv_class *bl;
5190 struct induction *biv, *giv;
5194 /* Search all IV classes, then all bivs, and finally all givs.
5196 There are three cases we are concerned with. First we have the situation
5197 of a giv that is only updated conditionally. In that case, it may not
5198 derive any givs after a label is passed.
5200 The second case is when a biv update occurs, or may occur, after the
5201 definition of a giv. For certain biv updates (see below) that are
5202 known to occur between the giv definition and use, we can adjust the
5203 giv definition. For others, or when the biv update is conditional,
5204 we must prevent the giv from deriving any other givs. There are two
5205 sub-cases within this case.
5207 If this is a label, we are concerned with any biv update that is done
5208 conditionally, since it may be done after the giv is defined followed by
5209 a branch here (actually, we need to pass both a jump and a label, but
5210 this extra tracking doesn't seem worth it).
5212 If this is a jump, we are concerned about any biv update that may be
5213 executed multiple times. We are actually only concerned about
5214 backward jumps, but it is probably not worth performing the test
5215 on the jump again here.
5217 If this is a biv update, we must adjust the giv status to show that a
5218 subsequent biv update was performed. If this adjustment cannot be done,
5219 the giv cannot derive further givs. */
5221 for (bl = loop_iv_list; bl; bl = bl->next)
5222 for (biv = bl->biv; biv; biv = biv->next_iv)
5223 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5226 for (giv = bl->giv; giv; giv = giv->next_iv)
5228 /* If cant_derive is already true, there is no point in
5229 checking all of these conditions again. */
5230 if (giv->cant_derive)
5233 /* If this giv is conditionally set and we have passed a label,
5234 it cannot derive anything. */
5235 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5236 giv->cant_derive = 1;
5238 /* Skip givs that have mult_val == 0, since
5239 they are really invariants. Also skip those that are
5240 replaceable, since we know their lifetime doesn't contain
5242 else if (giv->mult_val == const0_rtx || giv->replaceable)
5245 /* The only way we can allow this giv to derive another
5246 is if this is a biv increment and we can form the product
5247 of biv->add_val and giv->mult_val. In this case, we will
5248 be able to compute a compensation. */
5249 else if (biv->insn == p)
5253 if (biv->mult_val == const1_rtx)
5254 tem = simplify_giv_expr (gen_rtx_MULT (giv->mode,
5259 if (tem && giv->derive_adjustment)
5260 tem = simplify_giv_expr (gen_rtx_PLUS (giv->mode, tem,
5261 giv->derive_adjustment),
5264 giv->derive_adjustment = tem;
5266 giv->cant_derive = 1;
5268 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5269 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5270 giv->cant_derive = 1;
5275 /* Check whether an insn is an increment legitimate for a basic induction var.
5276 X is the source of insn P, or a part of it.
5277 MODE is the mode in which X should be interpreted.
5279 DEST_REG is the putative biv, also the destination of the insn.
5280 We accept patterns of these forms:
5281 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5282 REG = INVARIANT + REG
5284 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5285 and store the additive term into *INC_VAL.
5287 If X is an assignment of an invariant into DEST_REG, we set
5288 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5290 We also want to detect a BIV when it corresponds to a variable
5291 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5292 of the variable may be a PLUS that adds a SUBREG of that variable to
5293 an invariant and then sign- or zero-extends the result of the PLUS
5296 Most GIVs in such cases will be in the promoted mode, since that is the
5297 probably the natural computation mode (and almost certainly the mode
5298 used for addresses) on the machine. So we view the pseudo-reg containing
5299 the variable as the BIV, as if it were simply incremented.
5301 Note that treating the entire pseudo as a BIV will result in making
5302 simple increments to any GIVs based on it. However, if the variable
5303 overflows in its declared mode but not its promoted mode, the result will
5304 be incorrect. This is acceptable if the variable is signed, since
5305 overflows in such cases are undefined, but not if it is unsigned, since
5306 those overflows are defined. So we only check for SIGN_EXTEND and
5309 If we cannot find a biv, we return 0. */
5312 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val)
5314 enum machine_mode mode;
5320 register enum rtx_code code;
5324 code = GET_CODE (x);
5328 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5329 || (GET_CODE (XEXP (x, 0)) == SUBREG
5330 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5331 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5333 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5334 || (GET_CODE (XEXP (x, 1)) == SUBREG
5335 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5336 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5341 if (invariant_p (arg) != 1)
5344 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5345 *mult_val = const1_rtx;
5349 /* If this is a SUBREG for a promoted variable, check the inner
5351 if (SUBREG_PROMOTED_VAR_P (x))
5352 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
5353 dest_reg, p, inc_val, mult_val);
5357 /* If this register is assigned in a previous insn, look at its
5358 source, but don't go outside the loop or past a label. */
5364 insn = PREV_INSN (insn);
5365 } while (insn && GET_CODE (insn) == NOTE
5366 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5370 set = single_set (insn);
5374 if ((SET_DEST (set) == x
5375 || (GET_CODE (SET_DEST (set)) == SUBREG
5376 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5378 && SUBREG_REG (SET_DEST (set)) == x))
5379 && basic_induction_var (SET_SRC (set),
5380 (GET_MODE (SET_SRC (set)) == VOIDmode
5382 : GET_MODE (SET_SRC (set))),
5387 /* ... fall through ... */
5389 /* Can accept constant setting of biv only when inside inner most loop.
5390 Otherwise, a biv of an inner loop may be incorrectly recognized
5391 as a biv of the outer loop,
5392 causing code to be moved INTO the inner loop. */
5394 if (invariant_p (x) != 1)
5399 /* convert_modes aborts if we try to convert to or from CCmode, so just
5400 exclude that case. It is very unlikely that a condition code value
5401 would be a useful iterator anyways. */
5402 if (loops_enclosed == 1
5403 && GET_MODE_CLASS (mode) != MODE_CC
5404 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
5406 /* Possible bug here? Perhaps we don't know the mode of X. */
5407 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
5408 *mult_val = const0_rtx;
5415 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5416 dest_reg, p, inc_val, mult_val);
5419 /* Similar, since this can be a sign extension. */
5420 for (insn = PREV_INSN (p);
5421 (insn && GET_CODE (insn) == NOTE
5422 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5423 insn = PREV_INSN (insn))
5427 set = single_set (insn);
5429 if (set && SET_DEST (set) == XEXP (x, 0)
5430 && GET_CODE (XEXP (x, 1)) == CONST_INT
5431 && INTVAL (XEXP (x, 1)) >= 0
5432 && GET_CODE (SET_SRC (set)) == ASHIFT
5433 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
5434 return basic_induction_var (XEXP (SET_SRC (set), 0),
5435 GET_MODE (XEXP (x, 0)),
5436 dest_reg, insn, inc_val, mult_val);
5444 /* A general induction variable (giv) is any quantity that is a linear
5445 function of a basic induction variable,
5446 i.e. giv = biv * mult_val + add_val.
5447 The coefficients can be any loop invariant quantity.
5448 A giv need not be computed directly from the biv;
5449 it can be computed by way of other givs. */
5451 /* Determine whether X computes a giv.
5452 If it does, return a nonzero value
5453 which is the benefit from eliminating the computation of X;
5454 set *SRC_REG to the register of the biv that it is computed from;
5455 set *ADD_VAL and *MULT_VAL to the coefficients,
5456 such that the value of X is biv * mult + add; */
5459 general_induction_var (x, src_reg, add_val, mult_val, is_addr, pbenefit)
5470 /* If this is an invariant, forget it, it isn't a giv. */
5471 if (invariant_p (x) == 1)
5474 /* See if the expression could be a giv and get its form.
5475 Mark our place on the obstack in case we don't find a giv. */
5476 storage = (char *) oballoc (0);
5478 x = simplify_giv_expr (x, pbenefit);
5485 switch (GET_CODE (x))
5489 /* Since this is now an invariant and wasn't before, it must be a giv
5490 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5492 *src_reg = loop_iv_list->biv->dest_reg;
5493 *mult_val = const0_rtx;
5498 /* This is equivalent to a BIV. */
5500 *mult_val = const1_rtx;
5501 *add_val = const0_rtx;
5505 /* Either (plus (biv) (invar)) or
5506 (plus (mult (biv) (invar_1)) (invar_2)). */
5507 if (GET_CODE (XEXP (x, 0)) == MULT)
5509 *src_reg = XEXP (XEXP (x, 0), 0);
5510 *mult_val = XEXP (XEXP (x, 0), 1);
5514 *src_reg = XEXP (x, 0);
5515 *mult_val = const1_rtx;
5517 *add_val = XEXP (x, 1);
5521 /* ADD_VAL is zero. */
5522 *src_reg = XEXP (x, 0);
5523 *mult_val = XEXP (x, 1);
5524 *add_val = const0_rtx;
5531 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5532 unless they are CONST_INT). */
5533 if (GET_CODE (*add_val) == USE)
5534 *add_val = XEXP (*add_val, 0);
5535 if (GET_CODE (*mult_val) == USE)
5536 *mult_val = XEXP (*mult_val, 0);
5541 *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
5543 *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
5547 *pbenefit += rtx_cost (orig_x, SET);
5549 /* Always return true if this is a giv so it will be detected as such,
5550 even if the benefit is zero or negative. This allows elimination
5551 of bivs that might otherwise not be eliminated. */
5555 /* Given an expression, X, try to form it as a linear function of a biv.
5556 We will canonicalize it to be of the form
5557 (plus (mult (BIV) (invar_1))
5559 with possible degeneracies.
5561 The invariant expressions must each be of a form that can be used as a
5562 machine operand. We surround then with a USE rtx (a hack, but localized
5563 and certainly unambiguous!) if not a CONST_INT for simplicity in this
5564 routine; it is the caller's responsibility to strip them.
5566 If no such canonicalization is possible (i.e., two biv's are used or an
5567 expression that is neither invariant nor a biv or giv), this routine
5570 For a non-zero return, the result will have a code of CONST_INT, USE,
5571 REG (for a BIV), PLUS, or MULT. No other codes will occur.
5573 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
5575 static rtx sge_plus PROTO ((enum machine_mode, rtx, rtx));
5576 static rtx sge_plus_constant PROTO ((rtx, rtx));
5579 simplify_giv_expr (x, benefit)
5583 enum machine_mode mode = GET_MODE (x);
5587 /* If this is not an integer mode, or if we cannot do arithmetic in this
5588 mode, this can't be a giv. */
5589 if (mode != VOIDmode
5590 && (GET_MODE_CLASS (mode) != MODE_INT
5591 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
5594 switch (GET_CODE (x))
5597 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5598 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5599 if (arg0 == 0 || arg1 == 0)
5602 /* Put constant last, CONST_INT last if both constant. */
5603 if ((GET_CODE (arg0) == USE
5604 || GET_CODE (arg0) == CONST_INT)
5605 && ! ((GET_CODE (arg0) == USE
5606 && GET_CODE (arg1) == USE)
5607 || GET_CODE (arg1) == CONST_INT))
5608 tem = arg0, arg0 = arg1, arg1 = tem;
5610 /* Handle addition of zero, then addition of an invariant. */
5611 if (arg1 == const0_rtx)
5613 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
5614 switch (GET_CODE (arg0))
5618 /* Adding two invariants must result in an invariant, so enclose
5619 addition operation inside a USE and return it. */
5620 if (GET_CODE (arg0) == USE)
5621 arg0 = XEXP (arg0, 0);
5622 if (GET_CODE (arg1) == USE)
5623 arg1 = XEXP (arg1, 0);
5625 if (GET_CODE (arg0) == CONST_INT)
5626 tem = arg0, arg0 = arg1, arg1 = tem;
5627 if (GET_CODE (arg1) == CONST_INT)
5628 tem = sge_plus_constant (arg0, arg1);
5630 tem = sge_plus (mode, arg0, arg1);
5632 if (GET_CODE (tem) != CONST_INT)
5633 tem = gen_rtx_USE (mode, tem);
5638 /* biv + invar or mult + invar. Return sum. */
5639 return gen_rtx_PLUS (mode, arg0, arg1);
5642 /* (a + invar_1) + invar_2. Associate. */
5643 return simplify_giv_expr (
5644 gen_rtx_PLUS (mode, XEXP (arg0, 0),
5645 gen_rtx_PLUS (mode, XEXP (arg0, 1), arg1)),
5652 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
5653 MULT to reduce cases. */
5654 if (GET_CODE (arg0) == REG)
5655 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
5656 if (GET_CODE (arg1) == REG)
5657 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
5659 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
5660 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
5661 Recurse to associate the second PLUS. */
5662 if (GET_CODE (arg1) == MULT)
5663 tem = arg0, arg0 = arg1, arg1 = tem;
5665 if (GET_CODE (arg1) == PLUS)
5666 return simplify_giv_expr (gen_rtx_PLUS (mode,
5667 gen_rtx_PLUS (mode, arg0,
5672 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
5673 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
5676 if (!rtx_equal_p (arg0, arg1))
5679 return simplify_giv_expr (gen_rtx_MULT (mode,
5687 /* Handle "a - b" as "a + b * (-1)". */
5688 return simplify_giv_expr (gen_rtx_PLUS (mode,
5690 gen_rtx_MULT (mode, XEXP (x, 1),
5695 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5696 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5697 if (arg0 == 0 || arg1 == 0)
5700 /* Put constant last, CONST_INT last if both constant. */
5701 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
5702 && GET_CODE (arg1) != CONST_INT)
5703 tem = arg0, arg0 = arg1, arg1 = tem;
5705 /* If second argument is not now constant, not giv. */
5706 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
5709 /* Handle multiply by 0 or 1. */
5710 if (arg1 == const0_rtx)
5713 else if (arg1 == const1_rtx)
5716 switch (GET_CODE (arg0))
5719 /* biv * invar. Done. */
5720 return gen_rtx_MULT (mode, arg0, arg1);
5723 /* Product of two constants. */
5724 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
5727 /* invar * invar. It is a giv, but very few of these will
5728 actually pay off, so limit to simple registers. */
5729 if (GET_CODE (arg1) != CONST_INT)
5732 arg0 = XEXP (arg0, 0);
5733 if (GET_CODE (arg0) == REG)
5734 tem = gen_rtx_MULT (mode, arg0, arg1);
5735 else if (GET_CODE (arg0) == MULT
5736 && GET_CODE (XEXP (arg0, 0)) == REG
5737 && GET_CODE (XEXP (arg0, 1)) == CONST_INT)
5739 tem = gen_rtx_MULT (mode, XEXP (arg0, 0),
5740 GEN_INT (INTVAL (XEXP (arg0, 1))
5745 return gen_rtx_USE (mode, tem);
5748 /* (a * invar_1) * invar_2. Associate. */
5749 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (arg0, 0),
5756 /* (a + invar_1) * invar_2. Distribute. */
5757 return simplify_giv_expr (gen_rtx_PLUS (mode,
5771 /* Shift by constant is multiply by power of two. */
5772 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5775 return simplify_giv_expr (gen_rtx_MULT (mode,
5777 GEN_INT ((HOST_WIDE_INT) 1
5778 << INTVAL (XEXP (x, 1)))),
5782 /* "-a" is "a * (-1)" */
5783 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
5787 /* "~a" is "-a - 1". Silly, but easy. */
5788 return simplify_giv_expr (gen_rtx_MINUS (mode,
5789 gen_rtx_NEG (mode, XEXP (x, 0)),
5794 /* Already in proper form for invariant. */
5798 /* If this is a new register, we can't deal with it. */
5799 if (REGNO (x) >= max_reg_before_loop)
5802 /* Check for biv or giv. */
5803 switch (reg_iv_type[REGNO (x)])
5807 case GENERAL_INDUCT:
5809 struct induction *v = reg_iv_info[REGNO (x)];
5811 /* Form expression from giv and add benefit. Ensure this giv
5812 can derive another and subtract any needed adjustment if so. */
5813 *benefit += v->benefit;
5817 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode, v->src_reg,
5820 if (v->derive_adjustment)
5821 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
5822 return simplify_giv_expr (tem, benefit);
5826 /* If it isn't an induction variable, and it is invariant, we
5827 may be able to simplify things further by looking through
5828 the bits we just moved outside the loop. */
5829 if (invariant_p (x) == 1)
5833 for (m = the_movables; m ; m = m->next)
5834 if (rtx_equal_p (x, m->set_dest))
5836 /* Ok, we found a match. Substitute and simplify. */
5838 /* If we match another movable, we must use that, as
5839 this one is going away. */
5841 return simplify_giv_expr (m->match->set_dest, benefit);
5843 /* If consec is non-zero, this is a member of a group of
5844 instructions that were moved together. We handle this
5845 case only to the point of seeking to the last insn and
5846 looking for a REG_EQUAL. Fail if we don't find one. */
5851 do { tem = NEXT_INSN (tem); } while (--i > 0);
5853 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
5855 tem = XEXP (tem, 0);
5859 tem = single_set (m->insn);
5861 tem = SET_SRC (tem);
5866 /* What we are most interested in is pointer
5867 arithmetic on invariants -- only take
5868 patterns we may be able to do something with. */
5869 if (GET_CODE (tem) == PLUS
5870 || GET_CODE (tem) == MULT
5871 || GET_CODE (tem) == ASHIFT
5872 || GET_CODE (tem) == CONST_INT
5873 || GET_CODE (tem) == SYMBOL_REF)
5875 tem = simplify_giv_expr (tem, benefit);
5879 else if (GET_CODE (tem) == CONST
5880 && GET_CODE (XEXP (tem, 0)) == PLUS
5881 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
5882 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
5884 tem = simplify_giv_expr (XEXP (tem, 0), benefit);
5895 /* Fall through to general case. */
5897 /* If invariant, return as USE (unless CONST_INT).
5898 Otherwise, not giv. */
5899 if (GET_CODE (x) == USE)
5902 if (invariant_p (x) == 1)
5904 if (GET_CODE (x) == CONST_INT)
5906 if (GET_CODE (x) == CONST
5907 && GET_CODE (XEXP (x, 0)) == PLUS
5908 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
5909 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
5911 return gen_rtx_USE (mode, x);
5918 /* This routine folds invariants such that there is only ever one
5919 CONST_INT in the summation. It is only used by simplify_giv_expr. */
5922 sge_plus_constant (x, c)
5925 if (GET_CODE (x) == CONST_INT)
5926 return GEN_INT (INTVAL (x) + INTVAL (c));
5927 else if (GET_CODE (x) != PLUS)
5928 return gen_rtx_PLUS (GET_MODE (x), x, c);
5929 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
5931 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
5932 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
5934 else if (GET_CODE (XEXP (x, 0)) == PLUS
5935 || GET_CODE (XEXP (x, 1)) != PLUS)
5937 return gen_rtx_PLUS (GET_MODE (x),
5938 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
5942 return gen_rtx_PLUS (GET_MODE (x),
5943 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
5948 sge_plus (mode, x, y)
5949 enum machine_mode mode;
5952 while (GET_CODE (y) == PLUS)
5954 rtx a = XEXP (y, 0);
5955 if (GET_CODE (a) == CONST_INT)
5956 x = sge_plus_constant (x, a);
5958 x = gen_rtx_PLUS (mode, x, a);
5961 if (GET_CODE (y) == CONST_INT)
5962 x = sge_plus_constant (x, y);
5964 x = gen_rtx_PLUS (mode, x, y);
5968 /* Help detect a giv that is calculated by several consecutive insns;
5972 The caller has already identified the first insn P as having a giv as dest;
5973 we check that all other insns that set the same register follow
5974 immediately after P, that they alter nothing else,
5975 and that the result of the last is still a giv.
5977 The value is 0 if the reg set in P is not really a giv.
5978 Otherwise, the value is the amount gained by eliminating
5979 all the consecutive insns that compute the value.
5981 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5982 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5984 The coefficients of the ultimate giv value are stored in
5985 *MULT_VAL and *ADD_VAL. */
5988 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5989 add_val, mult_val, last_consec_insn)
5996 rtx *last_consec_insn;
6004 /* Indicate that this is a giv so that we can update the value produced in
6005 each insn of the multi-insn sequence.
6007 This induction structure will be used only by the call to
6008 general_induction_var below, so we can allocate it on our stack.
6009 If this is a giv, our caller will replace the induct var entry with
6010 a new induction structure. */
6012 = (struct induction *) alloca (sizeof (struct induction));
6013 v->src_reg = src_reg;
6014 v->mult_val = *mult_val;
6015 v->add_val = *add_val;
6016 v->benefit = first_benefit;
6018 v->derive_adjustment = 0;
6020 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
6021 reg_iv_info[REGNO (dest_reg)] = v;
6023 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6028 code = GET_CODE (p);
6030 /* If libcall, skip to end of call sequence. */
6031 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6035 && (set = single_set (p))
6036 && GET_CODE (SET_DEST (set)) == REG
6037 && SET_DEST (set) == dest_reg
6038 && (general_induction_var (SET_SRC (set), &src_reg,
6039 add_val, mult_val, 0, &benefit)
6040 /* Giv created by equivalent expression. */
6041 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6042 && general_induction_var (XEXP (temp, 0), &src_reg,
6043 add_val, mult_val, 0, &benefit)))
6044 && src_reg == v->src_reg)
6046 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6047 benefit += libcall_benefit (p);
6050 v->mult_val = *mult_val;
6051 v->add_val = *add_val;
6052 v->benefit = benefit;
6054 else if (code != NOTE)
6056 /* Allow insns that set something other than this giv to a
6057 constant. Such insns are needed on machines which cannot
6058 include long constants and should not disqualify a giv. */
6060 && (set = single_set (p))
6061 && SET_DEST (set) != dest_reg
6062 && CONSTANT_P (SET_SRC (set)))
6065 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
6070 *last_consec_insn = p;
6074 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6075 represented by G1. If no such expression can be found, or it is clear that
6076 it cannot possibly be a valid address, 0 is returned.
6078 To perform the computation, we note that
6081 where `v' is the biv.
6083 So G2 = (y/b) * G1 + (b - a*y/x).
6085 Note that MULT = y/x.
6087 Update: A and B are now allowed to be additive expressions such that
6088 B contains all variables in A. That is, computing B-A will not require
6089 subtracting variables. */
6092 express_from_1 (a, b, mult)
6095 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6097 if (mult == const0_rtx)
6100 /* If MULT is not 1, we cannot handle A with non-constants, since we
6101 would then be required to subtract multiples of the registers in A.
6102 This is theoretically possible, and may even apply to some Fortran
6103 constructs, but it is a lot of work and we do not attempt it here. */
6105 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6108 /* In general these structures are sorted top to bottom (down the PLUS
6109 chain), but not left to right across the PLUS. If B is a higher
6110 order giv than A, we can strip one level and recurse. If A is higher
6111 order, we'll eventually bail out, but won't know that until the end.
6112 If they are the same, we'll strip one level around this loop. */
6114 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6116 rtx ra, rb, oa, ob, tmp;
6118 ra = XEXP (a, 0), oa = XEXP (a, 1);
6119 if (GET_CODE (ra) == PLUS)
6120 tmp = ra, ra = oa, oa = tmp;
6122 rb = XEXP (b, 0), ob = XEXP (b, 1);
6123 if (GET_CODE (rb) == PLUS)
6124 tmp = rb, rb = ob, ob = tmp;
6126 if (rtx_equal_p (ra, rb))
6127 /* We matched: remove one reg completely. */
6129 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6130 /* An alternate match. */
6132 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6133 /* An alternate match. */
6137 /* Indicates an extra register in B. Strip one level from B and
6138 recurse, hoping B was the higher order expression. */
6139 ob = express_from_1 (a, ob, mult);
6142 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6146 /* Here we are at the last level of A, go through the cases hoping to
6147 get rid of everything but a constant. */
6149 if (GET_CODE (a) == PLUS)
6153 ra = XEXP (a, 0), oa = XEXP (a, 1);
6154 if (rtx_equal_p (oa, b))
6156 else if (!rtx_equal_p (ra, b))
6159 if (GET_CODE (oa) != CONST_INT)
6162 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6164 else if (GET_CODE (a) == CONST_INT)
6166 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6168 else if (GET_CODE (b) == PLUS)
6170 if (rtx_equal_p (a, XEXP (b, 0)))
6172 else if (rtx_equal_p (a, XEXP (b, 1)))
6177 else if (rtx_equal_p (a, b))
6184 express_from (g1, g2)
6185 struct induction *g1, *g2;
6189 /* The value that G1 will be multiplied by must be a constant integer. Also,
6190 the only chance we have of getting a valid address is if b*c/a (see above
6191 for notation) is also an integer. */
6192 if (GET_CODE (g1->mult_val) == CONST_INT
6193 && GET_CODE (g2->mult_val) == CONST_INT)
6195 if (g1->mult_val == const0_rtx
6196 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6198 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6200 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6204 /* ??? Find out if the one is a multiple of the other? */
6208 add = express_from_1 (g1->add_val, g2->add_val, mult);
6209 if (add == NULL_RTX)
6212 /* Form simplified final result. */
6213 if (mult == const0_rtx)
6215 else if (mult == const1_rtx)
6216 mult = g1->dest_reg;
6218 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6220 if (add == const0_rtx)
6224 if (GET_CODE (add) == PLUS
6225 && CONSTANT_P (XEXP (add, 1)))
6227 rtx tem = XEXP (add, 1);
6228 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6232 return gen_rtx_PLUS (g2->mode, mult, add);
6237 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6238 represented by G1. This indicates that G2 should be combined with G1 and
6239 that G2 can use (either directly or via an address expression) a register
6240 used to represent G1. */
6243 combine_givs_p (g1, g2)
6244 struct induction *g1, *g2;
6246 rtx tem = express_from (g1, g2);
6248 /* If these givs are identical, they can be combined. We use the results
6249 of express_from because the addends are not in a canonical form, so
6250 rtx_equal_p is a weaker test. */
6251 if (tem == g1->dest_reg)
6253 return g1->dest_reg;
6256 /* If G2 can be expressed as a function of G1 and that function is valid
6257 as an address and no more expensive than using a register for G2,
6258 the expression of G2 in terms of G1 can be used. */
6260 && g2->giv_type == DEST_ADDR
6261 && memory_address_p (g2->mem_mode, tem)
6262 /* ??? Looses, especially with -fforce-addr, where *g2->location
6263 will always be a register, and so anything more complicated
6267 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
6269 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
6280 struct combine_givs_stats
6287 cmp_combine_givs_stats (x, y)
6288 struct combine_givs_stats *x, *y;
6291 d = y->total_benefit - x->total_benefit;
6292 /* Stabilize the sort. */
6294 d = x->giv_number - y->giv_number;
6298 /* If one of these givs is a DEST_REG that was used by the other giv,
6299 this is actually a single use. Return 0 if this is not
6300 the case, -1 if g1 is the DEST_REG involved, and 1 if it was g2. */
6303 combine_givs_used_by_other (g1, g2)
6304 struct induction *g1, *g2;
6306 if (g1->giv_type == DEST_REG
6307 && reg_mentioned_p (g1->dest_reg, PATTERN (g2->insn)))
6310 if (g2->giv_type == DEST_REG
6311 && reg_mentioned_p (g2->dest_reg, PATTERN (g1->insn)))
6318 combine_givs_benefit_from (g1, g2)
6319 struct induction *g1, *g2;
6321 int tmp = combine_givs_used_by_other (g1, g2);
6325 return g2->benefit - g1->benefit;
6330 /* Check all pairs of givs for iv_class BL and see if any can be combined with
6331 any other. If so, point SAME to the giv combined with and set NEW_REG to
6332 be an expression (in terms of the other giv's DEST_REG) equivalent to the
6333 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
6337 struct iv_class *bl;
6339 struct induction *g1, *g2, **giv_array;
6340 int i, j, k, giv_count;
6341 struct combine_givs_stats *stats;
6344 /* Count givs, because bl->giv_count is incorrect here. */
6346 for (g1 = bl->giv; g1; g1 = g1->next_iv)
6351 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
6353 for (g1 = bl->giv; g1; g1 = g1->next_iv)
6355 giv_array[i++] = g1;
6357 stats = (struct combine_givs_stats *) alloca (giv_count * sizeof (*stats));
6358 bzero ((char *) stats, giv_count * sizeof (*stats));
6360 can_combine = (rtx *) alloca (giv_count * giv_count * sizeof(rtx));
6361 bzero ((char *) can_combine, giv_count * giv_count * sizeof(rtx));
6363 for (i = 0; i < giv_count; i++)
6369 this_benefit = g1->benefit;
6370 /* Add an additional weight for zero addends. */
6371 if (g1->no_const_addval)
6373 for (j = 0; j < giv_count; j++)
6379 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
6381 can_combine[i*giv_count + j] = this_combine;
6382 this_benefit += combine_givs_benefit_from (g1, g2);
6383 /* Add an additional weight for being reused more times. */
6387 stats[i].giv_number = i;
6388 stats[i].total_benefit = this_benefit;
6391 /* Iterate, combining until we can't. */
6393 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
6395 if (loop_dump_stream)
6397 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
6398 for (k = 0; k < giv_count; k++)
6400 g1 = giv_array[stats[k].giv_number];
6401 if (!g1->combined_with && !g1->same)
6402 fprintf (loop_dump_stream, " {%d, %d}",
6403 INSN_UID (giv_array[stats[k].giv_number]->insn),
6404 stats[k].total_benefit);
6406 putc ('\n', loop_dump_stream);
6409 for (k = 0; k < giv_count; k++)
6411 int g1_add_benefit = 0;
6413 i = stats[k].giv_number;
6416 /* If it has already been combined, skip. */
6417 if (g1->combined_with || g1->same)
6420 for (j = 0; j < giv_count; j++)
6423 if (g1 != g2 && can_combine[i*giv_count + j]
6424 /* If it has already been combined, skip. */
6425 && ! g2->same && ! g2->combined_with)
6429 g2->new_reg = can_combine[i*giv_count + j];
6431 g1->combined_with = 1;
6432 g1->lifetime += g2->lifetime;
6434 g1_add_benefit += combine_givs_benefit_from (g1, g2);
6436 /* ??? The new final_[bg]iv_value code does a much better job
6437 of finding replaceable giv's, and hence this code may no
6438 longer be necessary. */
6439 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
6440 g1_add_benefit -= copy_cost;
6442 /* To help optimize the next set of combinations, remove
6443 this giv from the benefits of other potential mates. */
6444 for (l = 0; l < giv_count; ++l)
6446 int m = stats[l].giv_number;
6447 if (can_combine[m*giv_count + j])
6449 /* Remove additional weight for being reused. */
6450 stats[l].total_benefit -= 3 +
6451 combine_givs_benefit_from (giv_array[m], g2);
6455 if (loop_dump_stream)
6456 fprintf (loop_dump_stream,
6457 "giv at %d combined with giv at %d\n",
6458 INSN_UID (g2->insn), INSN_UID (g1->insn));
6462 /* To help optimize the next set of combinations, remove
6463 this giv from the benefits of other potential mates. */
6464 if (g1->combined_with)
6466 for (j = 0; j < giv_count; ++j)
6468 int m = stats[j].giv_number;
6469 if (can_combine[m*giv_count + j])
6471 /* Remove additional weight for being reused. */
6472 stats[j].total_benefit -= 3 +
6473 combine_givs_benefit_from (giv_array[m], g1);
6477 g1->benefit += g1_add_benefit;
6479 /* We've finished with this giv, and everything it touched.
6480 Restart the combination so that proper weights for the
6481 rest of the givs are properly taken into account. */
6482 /* ??? Ideally we would compact the arrays at this point, so
6483 as to not cover old ground. But sanely compacting
6484 can_combine is tricky. */
6490 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
6493 emit_iv_add_mult (b, m, a, reg, insert_before)
6494 rtx b; /* initial value of basic induction variable */
6495 rtx m; /* multiplicative constant */
6496 rtx a; /* additive constant */
6497 rtx reg; /* destination register */
6503 /* Prevent unexpected sharing of these rtx. */
6507 /* Increase the lifetime of any invariants moved further in code. */
6508 update_reg_last_use (a, insert_before);
6509 update_reg_last_use (b, insert_before);
6510 update_reg_last_use (m, insert_before);
6513 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
6515 emit_move_insn (reg, result);
6516 seq = gen_sequence ();
6519 emit_insn_before (seq, insert_before);
6521 /* It is entirely possible that the expansion created lots of new
6522 registers. Iterate over the sequence we just created and
6525 if (GET_CODE (seq) == SEQUENCE)
6528 for (i = 0; i < XVECLEN (seq, 0); ++i)
6530 rtx set = single_set (XVECEXP (seq, 0, i));
6531 if (set && GET_CODE (SET_DEST (set)) == REG)
6532 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
6535 else if (GET_CODE (seq) == SET
6536 && GET_CODE (SET_DEST (seq)) == REG)
6537 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
6540 /* Test whether A * B can be computed without
6541 an actual multiply insn. Value is 1 if so. */
6544 product_cheap_p (a, b)
6550 struct obstack *old_rtl_obstack = rtl_obstack;
6551 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
6554 /* If only one is constant, make it B. */
6555 if (GET_CODE (a) == CONST_INT)
6556 tmp = a, a = b, b = tmp;
6558 /* If first constant, both constant, so don't need multiply. */
6559 if (GET_CODE (a) == CONST_INT)
6562 /* If second not constant, neither is constant, so would need multiply. */
6563 if (GET_CODE (b) != CONST_INT)
6566 /* One operand is constant, so might not need multiply insn. Generate the
6567 code for the multiply and see if a call or multiply, or long sequence
6568 of insns is generated. */
6570 rtl_obstack = &temp_obstack;
6572 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
6573 tmp = gen_sequence ();
6576 if (GET_CODE (tmp) == SEQUENCE)
6578 if (XVEC (tmp, 0) == 0)
6580 else if (XVECLEN (tmp, 0) > 3)
6583 for (i = 0; i < XVECLEN (tmp, 0); i++)
6585 rtx insn = XVECEXP (tmp, 0, i);
6587 if (GET_CODE (insn) != INSN
6588 || (GET_CODE (PATTERN (insn)) == SET
6589 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
6590 || (GET_CODE (PATTERN (insn)) == PARALLEL
6591 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
6592 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
6599 else if (GET_CODE (tmp) == SET
6600 && GET_CODE (SET_SRC (tmp)) == MULT)
6602 else if (GET_CODE (tmp) == PARALLEL
6603 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
6604 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
6607 /* Free any storage we obtained in generating this multiply and restore rtl
6608 allocation to its normal obstack. */
6609 obstack_free (&temp_obstack, storage);
6610 rtl_obstack = old_rtl_obstack;
6615 /* Check to see if loop can be terminated by a "decrement and branch until
6616 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
6617 Also try reversing an increment loop to a decrement loop
6618 to see if the optimization can be performed.
6619 Value is nonzero if optimization was performed. */
6621 /* This is useful even if the architecture doesn't have such an insn,
6622 because it might change a loops which increments from 0 to n to a loop
6623 which decrements from n to 0. A loop that decrements to zero is usually
6624 faster than one that increments from zero. */
6626 /* ??? This could be rewritten to use some of the loop unrolling procedures,
6627 such as approx_final_value, biv_total_increment, loop_iterations, and
6628 final_[bg]iv_value. */
6631 check_dbra_loop (loop_end, insn_count, loop_start, loop_info)
6635 struct loop_info *loop_info;
6637 struct iv_class *bl;
6644 rtx before_comparison;
6648 int compare_and_branch;
6650 /* If last insn is a conditional branch, and the insn before tests a
6651 register value, try to optimize it. Otherwise, we can't do anything. */
6653 jump = PREV_INSN (loop_end);
6654 comparison = get_condition_for_loop (jump);
6655 if (comparison == 0)
6658 /* Try to compute whether the compare/branch at the loop end is one or
6659 two instructions. */
6660 get_condition (jump, &first_compare);
6661 if (first_compare == jump)
6662 compare_and_branch = 1;
6663 else if (first_compare == prev_nonnote_insn (jump))
6664 compare_and_branch = 2;
6668 /* Check all of the bivs to see if the compare uses one of them.
6669 Skip biv's set more than once because we can't guarantee that
6670 it will be zero on the last iteration. Also skip if the biv is
6671 used between its update and the test insn. */
6673 for (bl = loop_iv_list; bl; bl = bl->next)
6675 if (bl->biv_count == 1
6676 && bl->biv->dest_reg == XEXP (comparison, 0)
6677 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
6685 /* Look for the case where the basic induction variable is always
6686 nonnegative, and equals zero on the last iteration.
6687 In this case, add a reg_note REG_NONNEG, which allows the
6688 m68k DBRA instruction to be used. */
6690 if (((GET_CODE (comparison) == GT
6691 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
6692 && INTVAL (XEXP (comparison, 1)) == -1)
6693 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
6694 && GET_CODE (bl->biv->add_val) == CONST_INT
6695 && INTVAL (bl->biv->add_val) < 0)
6697 /* Initial value must be greater than 0,
6698 init_val % -dec_value == 0 to ensure that it equals zero on
6699 the last iteration */
6701 if (GET_CODE (bl->initial_value) == CONST_INT
6702 && INTVAL (bl->initial_value) > 0
6703 && (INTVAL (bl->initial_value)
6704 % (-INTVAL (bl->biv->add_val))) == 0)
6706 /* register always nonnegative, add REG_NOTE to branch */
6707 REG_NOTES (PREV_INSN (loop_end))
6708 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
6709 REG_NOTES (PREV_INSN (loop_end)));
6715 /* If the decrement is 1 and the value was tested as >= 0 before
6716 the loop, then we can safely optimize. */
6717 for (p = loop_start; p; p = PREV_INSN (p))
6719 if (GET_CODE (p) == CODE_LABEL)
6721 if (GET_CODE (p) != JUMP_INSN)
6724 before_comparison = get_condition_for_loop (p);
6725 if (before_comparison
6726 && XEXP (before_comparison, 0) == bl->biv->dest_reg
6727 && GET_CODE (before_comparison) == LT
6728 && XEXP (before_comparison, 1) == const0_rtx
6729 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
6730 && INTVAL (bl->biv->add_val) == -1)
6732 REG_NOTES (PREV_INSN (loop_end))
6733 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
6734 REG_NOTES (PREV_INSN (loop_end)));
6741 else if (INTVAL (bl->biv->add_val) > 0)
6743 /* Try to change inc to dec, so can apply above optimization. */
6745 all registers modified are induction variables or invariant,
6746 all memory references have non-overlapping addresses
6747 (obviously true if only one write)
6748 allow 2 insns for the compare/jump at the end of the loop. */
6749 /* Also, we must avoid any instructions which use both the reversed
6750 biv and another biv. Such instructions will fail if the loop is
6751 reversed. We meet this condition by requiring that either
6752 no_use_except_counting is true, or else that there is only
6754 int num_nonfixed_reads = 0;
6755 /* 1 if the iteration var is used only to count iterations. */
6756 int no_use_except_counting = 0;
6757 /* 1 if the loop has no memory store, or it has a single memory store
6758 which is reversible. */
6759 int reversible_mem_store = 1;
6761 if (bl->giv_count == 0
6762 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
6764 rtx bivreg = regno_reg_rtx[bl->regno];
6766 /* If there are no givs for this biv, and the only exit is the
6767 fall through at the end of the loop, then
6768 see if perhaps there are no uses except to count. */
6769 no_use_except_counting = 1;
6770 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
6771 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
6773 rtx set = single_set (p);
6775 if (set && GET_CODE (SET_DEST (set)) == REG
6776 && REGNO (SET_DEST (set)) == bl->regno)
6777 /* An insn that sets the biv is okay. */
6779 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
6780 || p == prev_nonnote_insn (loop_end))
6781 /* Don't bother about the end test. */
6783 else if (reg_mentioned_p (bivreg, PATTERN (p)))
6785 no_use_except_counting = 0;
6791 if (no_use_except_counting)
6792 ; /* no need to worry about MEMs. */
6793 else if (num_mem_sets <= 1)
6795 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
6796 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
6797 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
6799 /* If the loop has a single store, and the destination address is
6800 invariant, then we can't reverse the loop, because this address
6801 might then have the wrong value at loop exit.
6802 This would work if the source was invariant also, however, in that
6803 case, the insn should have been moved out of the loop. */
6805 if (num_mem_sets == 1)
6806 reversible_mem_store
6807 = (! unknown_address_altered
6808 && ! invariant_p (XEXP (loop_store_mems[0], 0)));
6813 /* This code only acts for innermost loops. Also it simplifies
6814 the memory address check by only reversing loops with
6815 zero or one memory access.
6816 Two memory accesses could involve parts of the same array,
6817 and that can't be reversed.
6818 If the biv is used only for counting, than we don't need to worry
6819 about all these things. */
6821 if ((num_nonfixed_reads <= 1
6823 && !loop_has_volatile
6824 && reversible_mem_store
6825 && (bl->giv_count + bl->biv_count + num_mem_sets
6826 + num_movables + compare_and_branch == insn_count)
6827 && (bl == loop_iv_list && bl->next == 0))
6828 || no_use_except_counting)
6832 /* Loop can be reversed. */
6833 if (loop_dump_stream)
6834 fprintf (loop_dump_stream, "Can reverse loop\n");
6836 /* Now check other conditions:
6838 The increment must be a constant, as must the initial value,
6839 and the comparison code must be LT.
6841 This test can probably be improved since +/- 1 in the constant
6842 can be obtained by changing LT to LE and vice versa; this is
6846 /* for constants, LE gets turned into LT */
6847 && (GET_CODE (comparison) == LT
6848 || (GET_CODE (comparison) == LE
6849 && no_use_except_counting)))
6851 HOST_WIDE_INT add_val, add_adjust, comparison_val;
6852 rtx initial_value, comparison_value;
6854 enum rtx_code cmp_code;
6855 int comparison_const_width;
6856 unsigned HOST_WIDE_INT comparison_sign_mask;
6858 add_val = INTVAL (bl->biv->add_val);
6859 comparison_value = XEXP (comparison, 1);
6860 if (GET_MODE (comparison_value) == VOIDmode)
6861 comparison_const_width
6862 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
6864 comparison_const_width
6865 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
6866 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
6867 comparison_const_width = HOST_BITS_PER_WIDE_INT;
6868 comparison_sign_mask
6869 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
6871 /* If the comparison value is not a loop invariant, then we
6872 can not reverse this loop.
6874 ??? If the insns which initialize the comparison value as
6875 a whole compute an invariant result, then we could move
6876 them out of the loop and proceed with loop reversal. */
6877 if (!invariant_p (comparison_value))
6880 if (GET_CODE (comparison_value) == CONST_INT)
6881 comparison_val = INTVAL (comparison_value);
6882 initial_value = bl->initial_value;
6884 /* Normalize the initial value if it is an integer and
6885 has no other use except as a counter. This will allow
6886 a few more loops to be reversed. */
6887 if (no_use_except_counting
6888 && GET_CODE (comparison_value) == CONST_INT
6889 && GET_CODE (initial_value) == CONST_INT)
6891 comparison_val = comparison_val - INTVAL (bl->initial_value);
6892 /* The code below requires comparison_val to be a multiple
6893 of add_val in order to do the loop reversal, so
6894 round up comparison_val to a multiple of add_val.
6895 Since comparison_value is constant, we know that the
6896 current comparison code is LT. */
6897 comparison_val = comparison_val + add_val - 1;
6899 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
6900 /* We postpone overflow checks for COMPARISON_VAL here;
6901 even if there is an overflow, we might still be able to
6902 reverse the loop, if converting the loop exit test to
6904 initial_value = const0_rtx;
6907 /* First check if we can do a vanilla loop reversal. */
6908 if (initial_value == const0_rtx
6909 /* If we have a decrement_and_branch_on_count, prefer
6910 the NE test, since this will allow that instruction to
6911 be generated. Note that we must use a vanilla loop
6912 reversal if the biv is used to calculate a giv or has
6913 a non-counting use. */
6914 #if ! defined (HAVE_decrement_and_branch_until_zero) && defined (HAVE_decrement_and_branch_on_count)
6915 && (! (add_val == 1 && loop_info->vtop
6916 && (bl->biv_count == 0
6917 || no_use_except_counting)))
6919 && GET_CODE (comparison_value) == CONST_INT
6920 /* Now do postponed overflow checks on COMPARISON_VAL. */
6921 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
6922 & comparison_sign_mask))
6924 /* Register will always be nonnegative, with value
6925 0 on last iteration */
6926 add_adjust = add_val;
6930 else if (add_val == 1 && loop_info->vtop
6931 && (bl->biv_count == 0
6932 || no_use_except_counting))
6940 if (GET_CODE (comparison) == LE)
6941 add_adjust -= add_val;
6943 /* If the initial value is not zero, or if the comparison
6944 value is not an exact multiple of the increment, then we
6945 can not reverse this loop. */
6946 if (initial_value == const0_rtx
6947 && GET_CODE (comparison_value) == CONST_INT)
6949 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
6954 if (! no_use_except_counting || add_val != 1)
6958 final_value = comparison_value;
6960 /* Reset these in case we normalized the initial value
6961 and comparison value above. */
6962 if (GET_CODE (comparison_value) == CONST_INT
6963 && GET_CODE (initial_value) == CONST_INT)
6965 comparison_value = GEN_INT (comparison_val);
6967 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
6969 bl->initial_value = initial_value;
6971 /* Save some info needed to produce the new insns. */
6972 reg = bl->biv->dest_reg;
6973 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
6974 if (jump_label == pc_rtx)
6975 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
6976 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
6978 /* Set start_value; if this is not a CONST_INT, we need
6980 Initialize biv to start_value before loop start.
6981 The old initializing insn will be deleted as a
6982 dead store by flow.c. */
6983 if (initial_value == const0_rtx
6984 && GET_CODE (comparison_value) == CONST_INT)
6986 start_value = GEN_INT (comparison_val - add_adjust);
6987 emit_insn_before (gen_move_insn (reg, start_value),
6990 else if (GET_CODE (initial_value) == CONST_INT)
6992 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
6993 enum machine_mode mode = GET_MODE (reg);
6994 enum insn_code icode
6995 = add_optab->handlers[(int) mode].insn_code;
6996 if (! (*insn_operand_predicate[icode][0]) (reg, mode)
6997 || ! ((*insn_operand_predicate[icode][1])
6998 (comparison_value, mode))
6999 || ! (*insn_operand_predicate[icode][2]) (offset, mode))
7002 = gen_rtx_PLUS (mode, comparison_value, offset);
7003 emit_insn_before ((GEN_FCN (icode)
7004 (reg, comparison_value, offset)),
7006 if (GET_CODE (comparison) == LE)
7007 final_value = gen_rtx_PLUS (mode, comparison_value,
7010 else if (! add_adjust)
7012 enum machine_mode mode = GET_MODE (reg);
7013 enum insn_code icode
7014 = sub_optab->handlers[(int) mode].insn_code;
7015 if (! (*insn_operand_predicate[icode][0]) (reg, mode)
7016 || ! ((*insn_operand_predicate[icode][1])
7017 (comparison_value, mode))
7018 || ! ((*insn_operand_predicate[icode][2])
7019 (initial_value, mode)))
7022 = gen_rtx_MINUS (mode, comparison_value, initial_value);
7023 emit_insn_before ((GEN_FCN (icode)
7024 (reg, comparison_value, initial_value)),
7028 /* We could handle the other cases too, but it'll be
7029 better to have a testcase first. */
7032 /* We may not have a single insn which can increment a reg, so
7033 create a sequence to hold all the insns from expand_inc. */
7035 expand_inc (reg, new_add_val);
7036 tem = gen_sequence ();
7039 p = emit_insn_before (tem, bl->biv->insn);
7040 delete_insn (bl->biv->insn);
7042 /* Update biv info to reflect its new status. */
7044 bl->initial_value = start_value;
7045 bl->biv->add_val = new_add_val;
7047 /* Update loop info. */
7048 loop_info->initial_value = reg;
7049 loop_info->initial_equiv_value = reg;
7050 loop_info->final_value = const0_rtx;
7051 loop_info->final_equiv_value = const0_rtx;
7052 loop_info->comparison_value = const0_rtx;
7053 loop_info->comparison_code = cmp_code;
7054 loop_info->increment = new_add_val;
7056 /* Inc LABEL_NUSES so that delete_insn will
7057 not delete the label. */
7058 LABEL_NUSES (XEXP (jump_label, 0)) ++;
7060 /* Emit an insn after the end of the loop to set the biv's
7061 proper exit value if it is used anywhere outside the loop. */
7062 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
7064 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
7065 emit_insn_after (gen_move_insn (reg, final_value),
7068 /* Delete compare/branch at end of loop. */
7069 delete_insn (PREV_INSN (loop_end));
7070 if (compare_and_branch == 2)
7071 delete_insn (first_compare);
7073 /* Add new compare/branch insn at end of loop. */
7075 emit_cmp_insn (reg, const0_rtx, cmp_code, NULL_RTX,
7076 GET_MODE (reg), 0, 0);
7077 emit_jump_insn ((*bcc_gen_fctn[(int) cmp_code])
7078 (XEXP (jump_label, 0)));
7079 tem = gen_sequence ();
7081 emit_jump_insn_before (tem, loop_end);
7083 for (tem = PREV_INSN (loop_end);
7084 tem && GET_CODE (tem) != JUMP_INSN;
7085 tem = PREV_INSN (tem))
7089 JUMP_LABEL (tem) = XEXP (jump_label, 0);
7095 /* Increment of LABEL_NUSES done above. */
7096 /* Register is now always nonnegative,
7097 so add REG_NONNEG note to the branch. */
7098 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7104 /* Mark that this biv has been reversed. Each giv which depends
7105 on this biv, and which is also live past the end of the loop
7106 will have to be fixed up. */
7110 if (loop_dump_stream)
7111 fprintf (loop_dump_stream,
7112 "Reversed loop and added reg_nonneg\n");
7122 /* Verify whether the biv BL appears to be eliminable,
7123 based on the insns in the loop that refer to it.
7124 LOOP_START is the first insn of the loop, and END is the end insn.
7126 If ELIMINATE_P is non-zero, actually do the elimination.
7128 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
7129 determine whether invariant insns should be placed inside or at the
7130 start of the loop. */
7133 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
7134 struct iv_class *bl;
7138 int threshold, insn_count;
7140 rtx reg = bl->biv->dest_reg;
7143 /* Scan all insns in the loop, stopping if we find one that uses the
7144 biv in a way that we cannot eliminate. */
7146 for (p = loop_start; p != end; p = NEXT_INSN (p))
7148 enum rtx_code code = GET_CODE (p);
7149 rtx where = threshold >= insn_count ? loop_start : p;
7151 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
7152 && reg_mentioned_p (reg, PATTERN (p))
7153 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
7155 if (loop_dump_stream)
7156 fprintf (loop_dump_stream,
7157 "Cannot eliminate biv %d: biv used in insn %d.\n",
7158 bl->regno, INSN_UID (p));
7165 if (loop_dump_stream)
7166 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
7167 bl->regno, eliminate_p ? "was" : "can be");
7174 /* If BL appears in X (part of the pattern of INSN), see if we can
7175 eliminate its use. If so, return 1. If not, return 0.
7177 If BIV does not appear in X, return 1.
7179 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
7180 where extra insns should be added. Depending on how many items have been
7181 moved out of the loop, it will either be before INSN or at the start of
7185 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
7187 struct iv_class *bl;
7191 enum rtx_code code = GET_CODE (x);
7192 rtx reg = bl->biv->dest_reg;
7193 enum machine_mode mode = GET_MODE (reg);
7194 struct induction *v;
7206 /* If we haven't already been able to do something with this BIV,
7207 we can't eliminate it. */
7213 /* If this sets the BIV, it is not a problem. */
7214 if (SET_DEST (x) == reg)
7217 /* If this is an insn that defines a giv, it is also ok because
7218 it will go away when the giv is reduced. */
7219 for (v = bl->giv; v; v = v->next_iv)
7220 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
7224 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
7226 /* Can replace with any giv that was reduced and
7227 that has (MULT_VAL != 0) and (ADD_VAL == 0).
7228 Require a constant for MULT_VAL, so we know it's nonzero.
7229 ??? We disable this optimization to avoid potential
7232 for (v = bl->giv; v; v = v->next_iv)
7233 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
7234 && v->add_val == const0_rtx
7235 && ! v->ignore && ! v->maybe_dead && v->always_computable
7239 /* If the giv V had the auto-inc address optimization applied
7240 to it, and INSN occurs between the giv insn and the biv
7241 insn, then we must adjust the value used here.
7242 This is rare, so we don't bother to do so. */
7244 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
7245 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
7246 || (INSN_LUID (v->insn) > INSN_LUID (insn)
7247 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
7253 /* If the giv has the opposite direction of change,
7254 then reverse the comparison. */
7255 if (INTVAL (v->mult_val) < 0)
7256 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
7257 const0_rtx, v->new_reg);
7261 /* We can probably test that giv's reduced reg. */
7262 if (validate_change (insn, &SET_SRC (x), new, 0))
7266 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
7267 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
7268 Require a constant for MULT_VAL, so we know it's nonzero.
7269 ??? Do this only if ADD_VAL is a pointer to avoid a potential
7270 overflow problem. */
7272 for (v = bl->giv; v; v = v->next_iv)
7273 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
7274 && ! v->ignore && ! v->maybe_dead && v->always_computable
7276 && (GET_CODE (v->add_val) == SYMBOL_REF
7277 || GET_CODE (v->add_val) == LABEL_REF
7278 || GET_CODE (v->add_val) == CONST
7279 || (GET_CODE (v->add_val) == REG
7280 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
7282 /* If the giv V had the auto-inc address optimization applied
7283 to it, and INSN occurs between the giv insn and the biv
7284 insn, then we must adjust the value used here.
7285 This is rare, so we don't bother to do so. */
7287 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
7288 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
7289 || (INSN_LUID (v->insn) > INSN_LUID (insn)
7290 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
7296 /* If the giv has the opposite direction of change,
7297 then reverse the comparison. */
7298 if (INTVAL (v->mult_val) < 0)
7299 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
7302 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
7303 copy_rtx (v->add_val));
7305 /* Replace biv with the giv's reduced register. */
7306 update_reg_last_use (v->add_val, insn);
7307 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
7310 /* Insn doesn't support that constant or invariant. Copy it
7311 into a register (it will be a loop invariant.) */
7312 tem = gen_reg_rtx (GET_MODE (v->new_reg));
7314 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
7317 /* Substitute the new register for its invariant value in
7318 the compare expression. */
7319 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
7320 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
7329 case GT: case GE: case GTU: case GEU:
7330 case LT: case LE: case LTU: case LEU:
7331 /* See if either argument is the biv. */
7332 if (XEXP (x, 0) == reg)
7333 arg = XEXP (x, 1), arg_operand = 1;
7334 else if (XEXP (x, 1) == reg)
7335 arg = XEXP (x, 0), arg_operand = 0;
7339 if (CONSTANT_P (arg))
7341 /* First try to replace with any giv that has constant positive
7342 mult_val and constant add_val. We might be able to support
7343 negative mult_val, but it seems complex to do it in general. */
7345 for (v = bl->giv; v; v = v->next_iv)
7346 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
7347 && (GET_CODE (v->add_val) == SYMBOL_REF
7348 || GET_CODE (v->add_val) == LABEL_REF
7349 || GET_CODE (v->add_val) == CONST
7350 || (GET_CODE (v->add_val) == REG
7351 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
7352 && ! v->ignore && ! v->maybe_dead && v->always_computable
7355 /* If the giv V had the auto-inc address optimization applied
7356 to it, and INSN occurs between the giv insn and the biv
7357 insn, then we must adjust the value used here.
7358 This is rare, so we don't bother to do so. */
7360 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
7361 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
7362 || (INSN_LUID (v->insn) > INSN_LUID (insn)
7363 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
7369 /* Replace biv with the giv's reduced reg. */
7370 XEXP (x, 1-arg_operand) = v->new_reg;
7372 /* If all constants are actually constant integers and
7373 the derived constant can be directly placed in the COMPARE,
7375 if (GET_CODE (arg) == CONST_INT
7376 && GET_CODE (v->mult_val) == CONST_INT
7377 && GET_CODE (v->add_val) == CONST_INT
7378 && validate_change (insn, &XEXP (x, arg_operand),
7379 GEN_INT (INTVAL (arg)
7380 * INTVAL (v->mult_val)
7381 + INTVAL (v->add_val)), 0))
7384 /* Otherwise, load it into a register. */
7385 tem = gen_reg_rtx (mode);
7386 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
7387 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
7390 /* If that failed, put back the change we made above. */
7391 XEXP (x, 1-arg_operand) = reg;
7394 /* Look for giv with positive constant mult_val and nonconst add_val.
7395 Insert insns to calculate new compare value.
7396 ??? Turn this off due to possible overflow. */
7398 for (v = bl->giv; v; v = v->next_iv)
7399 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
7400 && ! v->ignore && ! v->maybe_dead && v->always_computable
7406 /* If the giv V had the auto-inc address optimization applied
7407 to it, and INSN occurs between the giv insn and the biv
7408 insn, then we must adjust the value used here.
7409 This is rare, so we don't bother to do so. */
7411 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
7412 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
7413 || (INSN_LUID (v->insn) > INSN_LUID (insn)
7414 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
7420 tem = gen_reg_rtx (mode);
7422 /* Replace biv with giv's reduced register. */
7423 validate_change (insn, &XEXP (x, 1 - arg_operand),
7426 /* Compute value to compare against. */
7427 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
7428 /* Use it in this insn. */
7429 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
7430 if (apply_change_group ())
7434 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
7436 if (invariant_p (arg) == 1)
7438 /* Look for giv with constant positive mult_val and nonconst
7439 add_val. Insert insns to compute new compare value.
7440 ??? Turn this off due to possible overflow. */
7442 for (v = bl->giv; v; v = v->next_iv)
7443 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
7444 && ! v->ignore && ! v->maybe_dead && v->always_computable
7450 /* If the giv V had the auto-inc address optimization applied
7451 to it, and INSN occurs between the giv insn and the biv
7452 insn, then we must adjust the value used here.
7453 This is rare, so we don't bother to do so. */
7455 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
7456 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
7457 || (INSN_LUID (v->insn) > INSN_LUID (insn)
7458 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
7464 tem = gen_reg_rtx (mode);
7466 /* Replace biv with giv's reduced register. */
7467 validate_change (insn, &XEXP (x, 1 - arg_operand),
7470 /* Compute value to compare against. */
7471 emit_iv_add_mult (arg, v->mult_val, v->add_val,
7473 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
7474 if (apply_change_group ())
7479 /* This code has problems. Basically, you can't know when
7480 seeing if we will eliminate BL, whether a particular giv
7481 of ARG will be reduced. If it isn't going to be reduced,
7482 we can't eliminate BL. We can try forcing it to be reduced,
7483 but that can generate poor code.
7485 The problem is that the benefit of reducing TV, below should
7486 be increased if BL can actually be eliminated, but this means
7487 we might have to do a topological sort of the order in which
7488 we try to process biv. It doesn't seem worthwhile to do
7489 this sort of thing now. */
7492 /* Otherwise the reg compared with had better be a biv. */
7493 if (GET_CODE (arg) != REG
7494 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
7497 /* Look for a pair of givs, one for each biv,
7498 with identical coefficients. */
7499 for (v = bl->giv; v; v = v->next_iv)
7501 struct induction *tv;
7503 if (v->ignore || v->maybe_dead || v->mode != mode)
7506 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
7507 if (! tv->ignore && ! tv->maybe_dead
7508 && rtx_equal_p (tv->mult_val, v->mult_val)
7509 && rtx_equal_p (tv->add_val, v->add_val)
7510 && tv->mode == mode)
7512 /* If the giv V had the auto-inc address optimization applied
7513 to it, and INSN occurs between the giv insn and the biv
7514 insn, then we must adjust the value used here.
7515 This is rare, so we don't bother to do so. */
7517 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
7518 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
7519 || (INSN_LUID (v->insn) > INSN_LUID (insn)
7520 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
7526 /* Replace biv with its giv's reduced reg. */
7527 XEXP (x, 1-arg_operand) = v->new_reg;
7528 /* Replace other operand with the other giv's
7530 XEXP (x, arg_operand) = tv->new_reg;
7537 /* If we get here, the biv can't be eliminated. */
7541 /* If this address is a DEST_ADDR giv, it doesn't matter if the
7542 biv is used in it, since it will be replaced. */
7543 for (v = bl->giv; v; v = v->next_iv)
7544 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
7552 /* See if any subexpression fails elimination. */
7553 fmt = GET_RTX_FORMAT (code);
7554 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7559 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
7560 eliminate_p, where))
7565 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7566 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
7567 eliminate_p, where))
7576 /* Return nonzero if the last use of REG
7577 is in an insn following INSN in the same basic block. */
7580 last_use_this_basic_block (reg, insn)
7586 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
7589 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
7595 /* Called via `note_stores' to record the initial value of a biv. Here we
7596 just record the location of the set and process it later. */
7599 record_initial (dest, set)
7603 struct iv_class *bl;
7605 if (GET_CODE (dest) != REG
7606 || REGNO (dest) >= max_reg_before_loop
7607 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
7610 bl = reg_biv_class[REGNO (dest)];
7612 /* If this is the first set found, record it. */
7613 if (bl->init_insn == 0)
7615 bl->init_insn = note_insn;
7620 /* If any of the registers in X are "old" and currently have a last use earlier
7621 than INSN, update them to have a last use of INSN. Their actual last use
7622 will be the previous insn but it will not have a valid uid_luid so we can't
7626 update_reg_last_use (x, insn)
7630 /* Check for the case where INSN does not have a valid luid. In this case,
7631 there is no need to modify the regno_last_uid, as this can only happen
7632 when code is inserted after the loop_end to set a pseudo's final value,
7633 and hence this insn will never be the last use of x. */
7634 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
7635 && INSN_UID (insn) < max_uid_for_loop
7636 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
7637 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
7641 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
7642 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
7645 update_reg_last_use (XEXP (x, i), insn);
7646 else if (fmt[i] == 'E')
7647 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7648 update_reg_last_use (XVECEXP (x, i, j), insn);
7653 /* Given a jump insn JUMP, return the condition that will cause it to branch
7654 to its JUMP_LABEL. If the condition cannot be understood, or is an
7655 inequality floating-point comparison which needs to be reversed, 0 will
7658 If EARLIEST is non-zero, it is a pointer to a place where the earliest
7659 insn used in locating the condition was found. If a replacement test
7660 of the condition is desired, it should be placed in front of that
7661 insn and we will be sure that the inputs are still valid.
7663 The condition will be returned in a canonical form to simplify testing by
7664 callers. Specifically:
7666 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
7667 (2) Both operands will be machine operands; (cc0) will have been replaced.
7668 (3) If an operand is a constant, it will be the second operand.
7669 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
7670 for GE, GEU, and LEU. */
7673 get_condition (jump, earliest)
7682 int reverse_code = 0;
7683 int did_reverse_condition = 0;
7684 enum machine_mode mode;
7686 /* If this is not a standard conditional jump, we can't parse it. */
7687 if (GET_CODE (jump) != JUMP_INSN
7688 || ! condjump_p (jump) || simplejump_p (jump))
7691 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
7692 mode = GET_MODE (XEXP (SET_SRC (PATTERN (jump)), 0));
7693 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
7694 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
7699 /* If this branches to JUMP_LABEL when the condition is false, reverse
7701 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
7702 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
7703 code = reverse_condition (code), did_reverse_condition ^= 1;
7705 /* If we are comparing a register with zero, see if the register is set
7706 in the previous insn to a COMPARE or a comparison operation. Perform
7707 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
7710 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
7712 /* Set non-zero when we find something of interest. */
7716 /* If comparison with cc0, import actual comparison from compare
7720 if ((prev = prev_nonnote_insn (prev)) == 0
7721 || GET_CODE (prev) != INSN
7722 || (set = single_set (prev)) == 0
7723 || SET_DEST (set) != cc0_rtx)
7726 op0 = SET_SRC (set);
7727 op1 = CONST0_RTX (GET_MODE (op0));
7733 /* If this is a COMPARE, pick up the two things being compared. */
7734 if (GET_CODE (op0) == COMPARE)
7736 op1 = XEXP (op0, 1);
7737 op0 = XEXP (op0, 0);
7740 else if (GET_CODE (op0) != REG)
7743 /* Go back to the previous insn. Stop if it is not an INSN. We also
7744 stop if it isn't a single set or if it has a REG_INC note because
7745 we don't want to bother dealing with it. */
7747 if ((prev = prev_nonnote_insn (prev)) == 0
7748 || GET_CODE (prev) != INSN
7749 || FIND_REG_INC_NOTE (prev, 0)
7750 || (set = single_set (prev)) == 0)
7753 /* If this is setting OP0, get what it sets it to if it looks
7755 if (rtx_equal_p (SET_DEST (set), op0))
7757 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
7759 /* ??? We may not combine comparisons done in a CCmode with
7760 comparisons not done in a CCmode. This is to aid targets
7761 like Alpha that have an IEEE compliant EQ instruction, and
7762 a non-IEEE compliant BEQ instruction. The use of CCmode is
7763 actually artificial, simply to prevent the combination, but
7764 should not affect other platforms.
7766 However, we must allow VOIDmode comparisons to match either
7767 CCmode or non-CCmode comparison, because some ports have
7768 modeless comparisons inside branch patterns.
7770 ??? This mode check should perhaps look more like the mode check
7771 in simplify_comparison in combine. */
7773 if ((GET_CODE (SET_SRC (set)) == COMPARE
7776 && GET_MODE_CLASS (inner_mode) == MODE_INT
7777 && (GET_MODE_BITSIZE (inner_mode)
7778 <= HOST_BITS_PER_WIDE_INT)
7779 && (STORE_FLAG_VALUE
7780 & ((HOST_WIDE_INT) 1
7781 << (GET_MODE_BITSIZE (inner_mode) - 1))))
7782 #ifdef FLOAT_STORE_FLAG_VALUE
7784 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
7785 && FLOAT_STORE_FLAG_VALUE < 0)
7788 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
7789 && (((GET_MODE_CLASS (mode) == MODE_CC)
7790 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
7791 || mode == VOIDmode || inner_mode == VOIDmode))
7793 else if (((code == EQ
7795 && (GET_MODE_BITSIZE (inner_mode)
7796 <= HOST_BITS_PER_WIDE_INT)
7797 && GET_MODE_CLASS (inner_mode) == MODE_INT
7798 && (STORE_FLAG_VALUE
7799 & ((HOST_WIDE_INT) 1
7800 << (GET_MODE_BITSIZE (inner_mode) - 1))))
7801 #ifdef FLOAT_STORE_FLAG_VALUE
7803 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
7804 && FLOAT_STORE_FLAG_VALUE < 0)
7807 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
7808 && (((GET_MODE_CLASS (mode) == MODE_CC)
7809 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
7810 || mode == VOIDmode || inner_mode == VOIDmode))
7813 /* We might have reversed a LT to get a GE here. But this wasn't
7814 actually the comparison of data, so we don't flag that we
7815 have had to reverse the condition. */
7816 did_reverse_condition ^= 1;
7824 else if (reg_set_p (op0, prev))
7825 /* If this sets OP0, but not directly, we have to give up. */
7830 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
7831 code = GET_CODE (x);
7834 code = reverse_condition (code);
7835 did_reverse_condition ^= 1;
7839 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
7845 /* If constant is first, put it last. */
7846 if (CONSTANT_P (op0))
7847 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
7849 /* If OP0 is the result of a comparison, we weren't able to find what
7850 was really being compared, so fail. */
7851 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
7854 /* Canonicalize any ordered comparison with integers involving equality
7855 if we can do computations in the relevant mode and we do not
7858 if (GET_CODE (op1) == CONST_INT
7859 && GET_MODE (op0) != VOIDmode
7860 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
7862 HOST_WIDE_INT const_val = INTVAL (op1);
7863 unsigned HOST_WIDE_INT uconst_val = const_val;
7864 unsigned HOST_WIDE_INT max_val
7865 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
7870 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
7871 code = LT, op1 = GEN_INT (const_val + 1);
7874 /* When cross-compiling, const_val might be sign-extended from
7875 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
7877 if ((HOST_WIDE_INT) (const_val & max_val)
7878 != (((HOST_WIDE_INT) 1
7879 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
7880 code = GT, op1 = GEN_INT (const_val - 1);
7884 if (uconst_val < max_val)
7885 code = LTU, op1 = GEN_INT (uconst_val + 1);
7889 if (uconst_val != 0)
7890 code = GTU, op1 = GEN_INT (uconst_val - 1);
7898 /* If this was floating-point and we reversed anything other than an
7899 EQ or NE, return zero. */
7900 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
7901 && did_reverse_condition && code != NE && code != EQ
7903 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
7907 /* Never return CC0; return zero instead. */
7912 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
7915 /* Similar to above routine, except that we also put an invariant last
7916 unless both operands are invariants. */
7919 get_condition_for_loop (x)
7922 rtx comparison = get_condition (x, NULL_PTR);
7925 || ! invariant_p (XEXP (comparison, 0))
7926 || invariant_p (XEXP (comparison, 1)))
7929 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
7930 XEXP (comparison, 1), XEXP (comparison, 0));
7933 #ifdef HAVE_decrement_and_branch_on_count
7934 /* Instrument loop for insertion of bct instruction. We distinguish between
7935 loops with compile-time bounds and those with run-time bounds.
7936 Information from loop_iterations() is used to compute compile-time bounds.
7937 Run-time bounds should use loop preconditioning, but currently ignored.
7941 insert_bct (loop_start, loop_end, loop_info)
7942 rtx loop_start, loop_end;
7943 struct loop_info *loop_info;
7946 unsigned HOST_WIDE_INT n_iterations;
7948 int increment_direction, compare_direction;
7950 /* If the loop condition is <= or >=, the number of iteration
7951 is 1 more than the range of the bounds of the loop. */
7952 int add_iteration = 0;
7954 enum machine_mode loop_var_mode = word_mode;
7956 int loop_num = uid_loop_num [INSN_UID (loop_start)];
7958 /* It's impossible to instrument a competely unrolled loop. */
7959 if (loop_info->unroll_number == -1)
7962 /* Make sure that the count register is not in use. */
7963 if (loop_used_count_register [loop_num])
7965 if (loop_dump_stream)
7966 fprintf (loop_dump_stream,
7967 "insert_bct %d: BCT instrumentation failed: count register already in use\n",
7972 /* Make sure that the function has no indirect jumps. */
7973 if (indirect_jump_in_function)
7975 if (loop_dump_stream)
7976 fprintf (loop_dump_stream,
7977 "insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
7982 /* Make sure that the last loop insn is a conditional jump. */
7983 if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
7984 || ! condjump_p (PREV_INSN (loop_end))
7985 || simplejump_p (PREV_INSN (loop_end)))
7987 if (loop_dump_stream)
7988 fprintf (loop_dump_stream,
7989 "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
7994 /* Make sure that the loop does not contain a function call
7995 (the count register might be altered by the called function). */
7998 if (loop_dump_stream)
7999 fprintf (loop_dump_stream,
8000 "insert_bct %d: BCT instrumentation failed: function call in loop\n",
8005 /* Make sure that the loop does not jump via a table.
8006 (the count register might be used to perform the branch on table). */
8007 if (loop_has_tablejump)
8009 if (loop_dump_stream)
8010 fprintf (loop_dump_stream,
8011 "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
8016 /* Account for loop unrolling in instrumented iteration count. */
8017 if (loop_info->unroll_number > 1)
8018 n_iterations = loop_info->n_iterations / loop_info->unroll_number;
8020 n_iterations = loop_info->n_iterations;
8022 if (n_iterations != 0 && n_iterations < 3)
8024 /* Allow an enclosing outer loop to benefit if possible. */
8025 if (loop_dump_stream)
8026 fprintf (loop_dump_stream,
8027 "insert_bct %d: Too few iterations to benefit from BCT optimization\n",
8032 /* Try to instrument the loop. */
8034 /* Handle the simpler case, where the bounds are known at compile time. */
8035 if (n_iterations > 0)
8037 /* Mark all enclosing loops that they cannot use count register. */
8038 for (i = loop_num; i != -1; i = loop_outer_loop[i])
8039 loop_used_count_register[i] = 1;
8040 instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
8044 /* Handle the more complex case, that the bounds are NOT known
8045 at compile time. In this case we generate run_time calculation
8046 of the number of iterations. */
8048 if (loop_info->iteration_var == 0)
8050 if (loop_dump_stream)
8051 fprintf (loop_dump_stream,
8052 "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
8057 if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
8058 || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
8060 if (loop_dump_stream)
8061 fprintf (loop_dump_stream,
8062 "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
8067 /* With runtime bounds, if the compare is of the form '!=' we give up */
8068 if (loop_info->comparison_code == NE)
8070 if (loop_dump_stream)
8071 fprintf (loop_dump_stream,
8072 "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
8076 /* Use common loop preconditioning code instead. */
8080 /* We rely on the existence of run-time guard to ensure that the
8081 loop executes at least once. */
8083 rtx iterations_num_reg;
8085 unsigned HOST_WIDE_INT increment_value_abs
8086 = INTVAL (increment) * increment_direction;
8088 /* make sure that the increment is a power of two, otherwise (an
8089 expensive) divide is needed. */
8090 if (exact_log2 (increment_value_abs) == -1)
8092 if (loop_dump_stream)
8093 fprintf (loop_dump_stream,
8094 "insert_bct: not instrumenting BCT because the increment is not power of 2\n");
8098 /* compute the number of iterations */
8103 /* Again, the number of iterations is calculated by:
8105 ; compare-val - initial-val + (increment -1) + additional-iteration
8106 ; num_iterations = -----------------------------------------------------------------
8109 /* ??? Do we have to call copy_rtx here before passing rtx to
8111 if (compare_direction > 0)
8113 /* <, <= :the loop variable is increasing */
8114 temp_reg = expand_binop (loop_var_mode, sub_optab,
8115 comparison_value, initial_value,
8116 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8120 temp_reg = expand_binop (loop_var_mode, sub_optab,
8121 initial_value, comparison_value,
8122 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8125 if (increment_value_abs - 1 + add_iteration != 0)
8126 temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
8127 GEN_INT (increment_value_abs - 1
8129 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8131 if (increment_value_abs != 1)
8133 /* ??? This will generate an expensive divide instruction for
8134 most targets. The original authors apparently expected this
8135 to be a shift, since they test for power-of-2 divisors above,
8136 but just naively generating a divide instruction will not give
8137 a shift. It happens to work for the PowerPC target because
8138 the rs6000.md file has a divide pattern that emits shifts.
8139 It will probably not work for any other target. */
8140 iterations_num_reg = expand_binop (loop_var_mode, sdiv_optab,
8142 GEN_INT (increment_value_abs),
8143 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8146 iterations_num_reg = temp_reg;
8148 sequence = gen_sequence ();
8150 emit_insn_before (sequence, loop_start);
8151 instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
8155 #endif /* Complex case */
8158 /* Instrument loop by inserting a bct in it as follows:
8159 1. A new counter register is created.
8160 2. In the head of the loop the new variable is initialized to the value
8161 passed in the loop_num_iterations parameter.
8162 3. At the end of the loop, comparison of the register with 0 is generated.
8163 The created comparison follows the pattern defined for the
8164 decrement_and_branch_on_count insn, so this insn will be generated.
8165 4. The branch on the old variable are deleted. The compare must remain
8166 because it might be used elsewhere. If the loop-variable or condition
8167 register are used elsewhere, they will be eliminated by flow. */
8170 instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
8171 rtx loop_start, loop_end;
8172 rtx loop_num_iterations;
8178 if (HAVE_decrement_and_branch_on_count)
8180 if (loop_dump_stream)
8182 fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
8183 if (GET_CODE (loop_num_iterations) == CONST_INT)
8184 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
8185 INTVAL (loop_num_iterations));
8187 fputs ("runtime", loop_dump_stream);
8188 fputs (" iterations)", loop_dump_stream);
8191 /* Discard original jump to continue loop. Original compare result
8192 may still be live, so it cannot be discarded explicitly. */
8193 delete_insn (PREV_INSN (loop_end));
8195 /* Insert the label which will delimit the start of the loop. */
8196 start_label = gen_label_rtx ();
8197 emit_label_after (start_label, loop_start);
8199 /* Insert initialization of the count register into the loop header. */
8201 counter_reg = gen_reg_rtx (word_mode);
8202 emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
8203 sequence = gen_sequence ();
8205 emit_insn_before (sequence, loop_start);
8207 /* Insert new comparison on the count register instead of the
8208 old one, generating the needed BCT pattern (that will be
8209 later recognized by assembly generation phase). */
8210 emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
8213 LABEL_NUSES (start_label)++;
8217 #endif /* HAVE_decrement_and_branch_on_count */
8219 /* Scan the function and determine whether it has indirect (computed) jumps.
8221 This is taken mostly from flow.c; similar code exists elsewhere
8222 in the compiler. It may be useful to put this into rtlanal.c. */
8224 indirect_jump_in_function_p (start)
8229 for (insn = start; insn; insn = NEXT_INSN (insn))
8230 if (computed_jump_p (insn))
8236 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
8237 documentation for LOOP_MEMS for the definition of `appropriate'.
8238 This function is called from prescan_loop via for_each_rtx. */
8241 insert_loop_mem (mem, data)
8243 void *data ATTRIBUTE_UNUSED;
8251 switch (GET_CODE (m))
8257 /* We're not interested in the MEM associated with a
8258 CONST_DOUBLE, so there's no need to traverse into this. */
8262 /* This is not a MEM. */
8266 /* See if we've already seen this MEM. */
8267 for (i = 0; i < loop_mems_idx; ++i)
8268 if (rtx_equal_p (m, loop_mems[i].mem))
8270 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
8271 /* The modes of the two memory accesses are different. If
8272 this happens, something tricky is going on, and we just
8273 don't optimize accesses to this MEM. */
8274 loop_mems[i].optimize = 0;
8279 /* Resize the array, if necessary. */
8280 if (loop_mems_idx == loop_mems_allocated)
8282 if (loop_mems_allocated != 0)
8283 loop_mems_allocated *= 2;
8285 loop_mems_allocated = 32;
8287 loop_mems = (loop_mem_info*)
8288 xrealloc (loop_mems,
8289 loop_mems_allocated * sizeof (loop_mem_info));
8292 /* Actually insert the MEM. */
8293 loop_mems[loop_mems_idx].mem = m;
8294 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
8295 because we can't put it in a register. We still store it in the
8296 table, though, so that if we see the same address later, but in a
8297 non-BLK mode, we'll not think we can optimize it at that point. */
8298 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
8299 loop_mems[loop_mems_idx].reg = NULL_RTX;
8305 /* Like load_mems, but also ensures that SET_IN_LOOP,
8306 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
8307 values after load_mems. */
8310 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top, start,
8311 reg_single_usage, insn_count)
8316 varray_type reg_single_usage;
8319 int nregs = max_reg_num ();
8321 load_mems (scan_start, end, loop_top, start);
8323 /* Recalculate set_in_loop and friends since load_mems may have
8324 created new registers. */
8325 if (max_reg_num () > nregs)
8331 nregs = max_reg_num ();
8333 if ((unsigned) nregs > set_in_loop->num_elements)
8335 /* Grow all the arrays. */
8336 VARRAY_GROW (set_in_loop, nregs);
8337 VARRAY_GROW (n_times_set, nregs);
8338 VARRAY_GROW (may_not_optimize, nregs);
8339 if (reg_single_usage)
8340 VARRAY_GROW (reg_single_usage, nregs);
8342 /* Clear the arrays */
8343 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
8344 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
8345 if (reg_single_usage)
8346 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
8348 count_loop_regs_set (loop_top ? loop_top : start, end,
8349 may_not_optimize, reg_single_usage,
8352 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
8354 VARRAY_CHAR (may_not_optimize, i) = 1;
8355 VARRAY_INT (set_in_loop, i) = 1;
8358 #ifdef AVOID_CCMODE_COPIES
8359 /* Don't try to move insns which set CC registers if we should not
8360 create CCmode register copies. */
8361 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
8362 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
8363 VARRAY_CHAR (may_not_optimize, i) = 1;
8366 /* Set n_times_set for the new registers. */
8367 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
8368 (char *) (&n_times_set->data.i[0] + old_nregs),
8369 (nregs - old_nregs) * sizeof (int));
8373 /* Move MEMs into registers for the duration of the loop. SCAN_START
8374 is the first instruction in the loop (as it is executed). The
8375 other parameters are as for next_insn_in_loop. */
8378 load_mems (scan_start, end, loop_top, start)
8384 int maybe_never = 0;
8387 rtx label = NULL_RTX;
8390 if (loop_mems_idx > 0)
8392 /* Nonzero if the next instruction may never be executed. */
8393 int next_maybe_never = 0;
8395 /* Check to see if it's possible that some instructions in the
8396 loop are never executed. */
8397 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
8398 p != NULL_RTX && !maybe_never;
8399 p = next_insn_in_loop (p, scan_start, end, loop_top))
8401 if (GET_CODE (p) == CODE_LABEL)
8403 else if (GET_CODE (p) == JUMP_INSN
8404 /* If we enter the loop in the middle, and scan
8405 around to the beginning, don't set maybe_never
8406 for that. This must be an unconditional jump,
8407 otherwise the code at the top of the loop might
8408 never be executed. Unconditional jumps are
8409 followed a by barrier then loop end. */
8410 && ! (GET_CODE (p) == JUMP_INSN
8411 && JUMP_LABEL (p) == loop_top
8412 && NEXT_INSN (NEXT_INSN (p)) == end
8413 && simplejump_p (p)))
8415 if (!condjump_p (p))
8416 /* Something complicated. */
8419 /* If there are any more instructions in the loop, they
8420 might not be reached. */
8421 next_maybe_never = 1;
8423 else if (next_maybe_never)
8427 /* Actually move the MEMs. */
8428 for (i = 0; i < loop_mems_idx; ++i)
8433 rtx mem = loop_mems[i].mem;
8435 if (MEM_VOLATILE_P (mem)
8436 || invariant_p (XEXP (mem, 0)) != 1)
8437 /* There's no telling whether or not MEM is modified. */
8438 loop_mems[i].optimize = 0;
8440 /* Go through the MEMs written to in the loop to see if this
8441 one is aliased by one of them. */
8442 for (j = 0; j < loop_store_mems_idx; ++j)
8444 if (rtx_equal_p (mem, loop_store_mems[j]))
8446 else if (true_dependence (loop_store_mems[j], VOIDmode,
8449 /* MEM is indeed aliased by this store. */
8450 loop_mems[i].optimize = 0;
8455 /* If this MEM is written to, we must be sure that there
8456 are no reads from another MEM that aliases this one. */
8457 if (loop_mems[i].optimize && written)
8461 for (j = 0; j < loop_mems_idx; ++j)
8465 else if (true_dependence (mem,
8470 /* It's not safe to hoist loop_mems[i] out of
8471 the loop because writes to it might not be
8472 seen by reads from loop_mems[j]. */
8473 loop_mems[i].optimize = 0;
8479 if (maybe_never && may_trap_p (mem))
8480 /* We can't access the MEM outside the loop; it might
8481 cause a trap that wouldn't have happened otherwise. */
8482 loop_mems[i].optimize = 0;
8484 if (!loop_mems[i].optimize)
8485 /* We thought we were going to lift this MEM out of the
8486 loop, but later discovered that we could not. */
8489 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
8490 order to keep scan_loop from moving stores to this MEM
8491 out of the loop just because this REG is neither a
8492 user-variable nor used in the loop test. */
8493 reg = gen_reg_rtx (GET_MODE (mem));
8494 REG_USERVAR_P (reg) = 1;
8495 loop_mems[i].reg = reg;
8497 /* Now, replace all references to the MEM with the
8498 corresponding pesudos. */
8499 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
8501 p = next_insn_in_loop (p, scan_start, end, loop_top))
8506 for_each_rtx (&p, replace_loop_mem, &ri);
8509 if (!apply_change_group ())
8510 /* We couldn't replace all occurrences of the MEM. */
8511 loop_mems[i].optimize = 0;
8516 /* Load the memory immediately before START, which is
8517 the NOTE_LOOP_BEG. */
8518 set = gen_rtx_SET (GET_MODE (reg), reg, mem);
8519 emit_insn_before (set, start);
8523 if (label == NULL_RTX)
8525 /* We must compute the former
8526 right-after-the-end label before we insert
8528 end_label = next_label (end);
8529 label = gen_label_rtx ();
8530 emit_label_after (label, end);
8533 /* Store the memory immediately after END, which is
8534 the NOTE_LOOP_END. */
8535 set = gen_rtx_SET (GET_MODE (reg), copy_rtx (mem), reg);
8536 emit_insn_after (set, label);
8539 if (loop_dump_stream)
8541 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
8542 REGNO (reg), (written ? "r/w" : "r/o"));
8543 print_rtl (loop_dump_stream, mem);
8544 fputc ('\n', loop_dump_stream);
8550 if (label != NULL_RTX)
8552 /* Now, we need to replace all references to the previous exit
8553 label with the new one. */
8558 for (p = start; p != end; p = NEXT_INSN (p))
8560 for_each_rtx (&p, replace_label, &rr);
8562 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
8563 field. This is not handled by for_each_rtx because it doesn't
8564 handle unprinted ('0') fields. We need to update JUMP_LABEL
8565 because the immediately following unroll pass will use it.
8566 replace_label would not work anyways, because that only handles
8568 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
8569 JUMP_LABEL (p) = label;
8574 /* Replace MEM with its associated pseudo register. This function is
8575 called from load_mems via for_each_rtx. DATA is actually an
8576 rtx_and_int * describing the instruction currently being scanned
8577 and the MEM we are currently replacing. */
8580 replace_loop_mem (mem, data)
8592 switch (GET_CODE (m))
8598 /* We're not interested in the MEM associated with a
8599 CONST_DOUBLE, so there's no need to traverse into one. */
8603 /* This is not a MEM. */
8607 ri = (rtx_and_int*) data;
8610 if (!rtx_equal_p (loop_mems[i].mem, m))
8611 /* This is not the MEM we are currently replacing. */
8616 /* Actually replace the MEM. */
8617 validate_change (insn, mem, loop_mems[i].reg, 1);
8622 /* Replace occurrences of the old exit label for the loop with the new
8623 one. DATA is an rtx_pair containing the old and new labels,
8627 replace_label (x, data)
8632 rtx old_label = ((rtx_pair*) data)->r1;
8633 rtx new_label = ((rtx_pair*) data)->r2;
8638 if (GET_CODE (l) != LABEL_REF)
8641 if (XEXP (l, 0) != old_label)
8644 XEXP (l, 0) = new_label;
8645 ++LABEL_NUSES (new_label);
8646 --LABEL_NUSES (old_label);