| blob | a8bf9c7cbd0c6c85918fdd50b0fad503e06a0dca |
1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997,
3 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
10 any later version.
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
57 #define LOOP_REG_LIFETIME(LOOP, REGNO) \
58 ((REGNO_LAST_LUID (REGNO) - REGNO_FIRST_LUID (REGNO)))
60 #define LOOP_REG_GLOBAL_P(LOOP, REGNO) \
61 ((REGNO_LAST_LUID (REGNO) > INSN_LUID ((LOOP)->end) \
62 || REGNO_FIRST_LUID (REGNO) < INSN_LUID ((LOOP)->start)))
65 /* Vector mapping INSN_UIDs to luids.
66 The luids are like uids but increase monotonically always.
67 We use them to see whether a jump comes from outside a given loop. */
71 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
72 number the insn is contained in. */
76 /* 1 + largest uid of any insn. */
80 /* 1 + luid of last insn. */
84 /* Number of loops detected in current function. Used as index to the
85 next few tables. */
89 /* Bound on pseudo register number before loop optimization.
90 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
93 /* The value to pass to the next call of reg_scan_update. */
96 #define obstack_chunk_alloc xmalloc
97 #define obstack_chunk_free free
98 \f
99 /* During the analysis of a loop, a chain of `struct movable's
100 is made to record all the movable insns found.
101 Then the entire chain can be scanned to decide which to move. */
103 struct movable
104 {
109 of any registers used within the LIBCALL. */
111 that must be moved with this one. */
114 may be adjusted when matching movables
115 that load the same value are found. */
117 including other movables that force this
118 or match this one. */
122 /* If PARTIAL is 1, GLOBAL means something different:
123 that the reg is live outside the range from where it is set
124 to the following label. */
128 In particular, moving it does not make it
129 invariant. */
131 load SRC, rather than copying INSN. */
133 first insn of a consecutive sets group. */
136 that we should avoid changing when clearing
137 the rest of the reg. */
141 };
146 /* Forward declarations. */
160 #if 0
162 #endif
192 rtx *));
273 {
274 rtx r1;
275 rtx r2;
279 {
280 rtx match;
281 rtx replacement;
282 rtx insn;
285 /* Nonzero iff INSN is between START and END, inclusive. */
286 #define INSN_IN_RANGE_P(INSN, START, END) \
287 (INSN_UID (INSN) < max_uid_for_loop \
288 && INSN_LUID (INSN) >= INSN_LUID (START) \
289 && INSN_LUID (INSN) <= INSN_LUID (END))
291 /* Indirect_jump_in_function is computed once per function. */
299 rtx));
300 \f
301 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
302 copy the value of the strength reduced giv to its original register. */
305 /* Cost of using a register, to normalize the benefits of a giv. */
308 void
310 {
316 }
317 \f
318 /* Compute the mapping from uids to luids.
319 LUIDs are numbers assigned to insns, like uids,
320 except that luids increase monotonically through the code.
321 Start at insn START and stop just before END. Assign LUIDs
322 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
323 static int
327 {
329 rtx insn;
332 {
335 /* Don't assign luids to line-number NOTEs, so that the distance in
336 luids between two insns is not affected by -g. */
340 else
341 /* Give a line number note the same luid as preceding insn. */
343 }
345 }
346 \f
347 /* Entry point of this file. Perform loop optimization
348 on the current function. F is the first insn of the function
349 and DUMPFILE is a stream for output of a trace of actions taken
350 (or 0 if none should be output). */
352 void
354 /* f is the first instruction of a chain of insns for one function */
355 rtx f;
358 {
374 /* Count the number of loops. */
378 {
381 max_loop_num++;
382 }
384 /* Don't waste time if no loops. */
390 /* Get size to use for tables indexed by uids.
391 Leave some space for labels allocated by find_and_verify_loops. */
398 /* Allocate storage for array of loops. */
402 /* Find and process each loop.
403 First, find them, and record them in order of their beginnings. */
406 /* Allocate and initialize auxiliary loop information. */
411 /* Now find all register lifetimes. This must be done after
412 find_and_verify_loops, because it might reorder the insns in the
413 function. */
416 /* This must occur after reg_scan so that registers created by gcse
417 will have entries in the register tables.
419 We could have added a call to reg_scan after gcse_main in toplev.c,
420 but moving this call to init_alias_analysis is more efficient. */
423 /* See if we went too far. Note that get_max_uid already returns
424 one more that the maximum uid of all insn. */
427 /* Now reset it to the actual size we need. See above. */
430 /* find_and_verify_loops has already called compute_luids, but it
431 might have rearranged code afterwards, so we need to recompute
432 the luids now. */
435 /* Don't leave gaps in uid_luid for insns that have been
436 deleted. It is possible that the first or last insn
437 using some register has been deleted by cross-jumping.
438 Make sure that uid_luid for that former insn's uid
439 points to the general area where that insn used to be. */
441 {
445 }
450 /* Determine if the function has indirect jump. On some systems
451 this prevents low overhead loop instructions from being used. */
454 /* Now scan the loops, last ones first, since this means inner ones are done
455 before outer ones. */
457 {
462 }
464 /* If there were lexical blocks inside the loop, they have been
465 replicated. We will now have more than one NOTE_INSN_BLOCK_BEG
466 and NOTE_INSN_BLOCK_END for each such block. We must duplicate
467 the BLOCKs as well. */
473 /* Clean up. */
478 }
479 \f
480 /* Returns the next insn, in execution order, after INSN. START and
481 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
482 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
483 insn-stream; it is used with loops that are entered near the
484 bottom. */
486 static rtx
489 rtx insn;
490 {
494 {
496 /* Go to the top of the loop, and continue there. */
498 else
499 /* We're done. */
501 }
504 /* We're done. */
508 }
510 /* Optimize one loop described by LOOP. */
512 /* ??? Could also move memory writes out of loops if the destination address
513 is invariant, the source is invariant, the memory write is not volatile,
514 and if we can prove that no read inside the loop can read this address
515 before the write occurs. If there is a read of this address after the
516 write, then we can also mark the memory read as invariant. */
518 static void
522 {
528 rtx p;
529 /* 1 if we are scanning insns that could be executed zero times. */
531 /* 1 if we are scanning insns that might never be executed
532 due to a subroutine call which might exit before they are reached. */
534 /* Jump insn that enters the loop, or 0 if control drops in. */
536 /* Number of insns in the loop. */
540 /* The SET from an insn, if it is the only SET in the insn. */
542 /* Chain describing insns movable in current loop. */
544 /* Ratio of extra register life span we can justify
545 for saving an instruction. More if loop doesn't call subroutines
546 since in that case saving an insn makes more difference
547 and more registers are available. */
549 /* Nonzero if we are scanning instructions in a sub-loop. */
557 /* Determine whether this loop starts with a jump down to a test at
558 the end. This will occur for a small number of loops with a test
559 that is too complex to duplicate in front of the loop.
561 We search for the first insn or label in the loop, skipping NOTEs.
562 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
563 (because we might have a loop executed only once that contains a
564 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
565 (in case we have a degenerate loop).
567 Note that if we mistakenly think that a loop is entered at the top
568 when, in fact, it is entered at the exit test, the only effect will be
569 slightly poorer optimization. Making the opposite error can generate
570 incorrect code. Since very few loops now start with a jump to the
571 exit test, the code here to detect that case is very conservative. */
574 p != loop_end
580 ;
584 /* If loop end is the end of the current function, then emit a
585 NOTE_INSN_DELETED after loop_end and set loop->sink to the dummy
586 note insn. This is the position we use when sinking insns out of
587 the loop. */
590 else
593 /* Set up variables describing this loop. */
597 /* If loop has a jump before the first label,
598 the true entry is the target of that jump.
599 Start scan from there.
600 But record in LOOP->TOP the place where the end-test jumps
601 back to so we can scan that after the end of the loop. */
603 {
606 /* Loop entry must be unconditional jump (and not a RETURN) */
609 /* Check to see whether the jump actually
610 jumps out of the loop (meaning it's no loop).
611 This case can happen for things like
612 do {..} while (0). If this label was generated previously
613 by loop, we can't tell anything about it and have to reject
614 the loop. */
616 {
619 }
620 }
622 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
623 as required by loop_reg_used_before_p. So skip such loops. (This
624 test may never be true, but it's best to play it safe.)
626 Also, skip loops where we do not start scanning at a label. This
627 test also rejects loops starting with a JUMP_INSN that failed the
628 test above. */
632 {
637 }
639 /* Allocate extra space for REGs that might be created by load_mems.
640 We allocate a little extra slop as well, in the hopes that we
641 won't have to reallocate the regs array. */
646 {
652 }
654 /* Scan through the loop finding insns that are safe to move.
655 Set REGS->ARRAY[I].SET_IN_LOOP negative for the reg I being set, so that
656 this reg will be considered invariant for subsequent insns.
657 We consider whether subsequent insns use the reg
658 in deciding whether it is worth actually moving.
660 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
661 and therefore it is possible that the insns we are scanning
662 would never be executed. At such times, we must make sure
663 that it is safe to execute the insn once instead of zero times.
664 When MAYBE_NEVER is 0, all insns will be executed at least once
665 so that is not a problem. */
670 {
675 {
682 /* Figure out what to use as a source of this insn. If a REG_EQUIV
683 note is given or if a REG_EQUAL note with a constant operand is
684 specified, use it as the source and mark that we should move
685 this insn by calling emit_move_insn rather that duplicating the
686 insn.
688 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
689 is present. */
693 else
694 {
699 {
701 /* A libcall block can use regs that don't appear in
702 the equivalent expression. To move the libcall,
703 we must move those regs too. */
705 }
706 }
708 /* For parallels, add any possible uses to the depencies, as we can't move
709 the insn without resolving them first. */
711 {
713 {
717 }
718 }
720 /* Don't try to optimize a register that was made
721 by loop-optimization for an inner loop.
722 We don't know its life-span, so we can't compute the benefit. */
724 ;
726 than the one where it is set (meaning that
727 something after this point in the loop might
728 depend on its value before the set). */
730 /* And the set is not guaranteed to be executed one
731 the loop starts, or the value before the set is
732 needed before the set occurs...
734 ??? Note we have quadratic behaviour here, mitigated
735 by the fact that the previous test will often fail for
736 large loops. Rather than re-scanning the entire loop
737 each time for register usage, we should build tables
738 of the register usage and use them here instead. */
739 && (maybe_never
741 /* It is unsafe to move the set.
743 This code used to consider it OK to move a set of a variable
744 which was not created by the user and not used in an exit test.
745 That behavior is incorrect and was removed. */
746 ;
751 || (tem1
752 = consec_sets_invariant_p
755 p)))
756 /* If the insn can cause a trap (such as divide by zero),
757 can't move it unless it's guaranteed to be executed
758 once loop is entered. Even a function call might
759 prevent the trap insn from being reached
760 (since it might exit!) */
763 {
767 /* A potential lossage is where we have a case where two insns
768 can be combined as long as they are both in the loop, but
769 we move one of them outside the loop. For large loops,
770 this can lose. The most common case of this is the address
771 of a function being called.
773 Therefore, if this register is marked as being used exactly
774 once if we are in a loop with calls (a "large loop"), see if
775 we can replace the usage of this register with the source
776 of this SET. If we can, delete this insn.
778 Don't do this if P has a REG_RETVAL note or if we have
779 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
791 && (! SMALL_REGISTER_CLASSES
794 /* This test is not redundant; SET_SRC (set) might be
795 a call-clobbered register and the life of REGNO
796 might span a call. */
802 {
803 /* Replace any usage in a REG_EQUAL note. Must copy the
804 new source, so that we don't get rtx sharing between the
805 SET_SOURCE and REG_NOTES of insn p. */
815 }
833 /* Set M->cond if either loop_invariant_p
834 or consec_sets_invariant_p returned 2
835 (only conditionally invariant). */
844 /* Add M to the end of the chain MOVABLES. */
848 {
849 /* It is possible for the first instruction to have a
850 REG_EQUAL note but a non-invariant SET_SRC, so we must
851 remember the status of the first instruction in case
852 the last instruction doesn't have a REG_EQUAL note. */
855 /* Skip this insn, not checking REG_LIBCALL notes. */
857 /* Skip the consecutive insns, if there are any. */
859 /* Back up to the last insn of the consecutive group. */
862 /* We must now reset m->move_insn, m->is_equiv, and possibly
863 m->set_src to correspond to the effects of all the
864 insns. */
868 else
869 {
873 else
876 }
878 }
879 }
880 /* If this register is always set within a STRICT_LOW_PART
881 or set to zero, then its high bytes are constant.
882 So clear them outside the loop and within the loop
883 just load the low bytes.
884 We must check that the machine has an instruction to do so.
885 Also, if the value loaded into the register
886 depends on the same register, this cannot be done. */
896 {
899 {
913 /* If the insn may not be executed on some cycles,
914 we can't clear the whole reg; clear just high part.
915 Not even if the reg is used only within this loop.
916 Consider this:
917 while (1)
918 while (s != t) {
919 if (foo ()) x = *s;
920 use (x);
921 }
922 Clearing x before the inner loop could clobber a value
923 being saved from the last time around the outer loop.
924 However, if the reg is not used outside this loop
925 and all uses of the register are in the same
926 basic block as the store, there is no problem.
928 If this insn was made by loop, we don't know its
929 INSN_LUID and hence must make a conservative
930 assumption. */
933 || (labels_in_range_p
937 else
945 /* Add M to the end of the chain MOVABLES. */
947 }
948 }
949 }
950 /* Past a call insn, we get to insns which might not be executed
951 because the call might exit. This matters for insns that trap.
952 Constant and pure call insns always return, so they don't count. */
955 /* Past a label or a jump, we get to insns for which we
956 can't count on whether or how many times they will be
957 executed during each iteration. Therefore, we can
958 only move out sets of trivial variables
959 (those not used after the loop). */
960 /* Similar code appears twice in strength_reduce. */
962 /* If we enter the loop in the middle, and scan around to the
963 beginning, don't set maybe_never for that. This must be an
964 unconditional jump, otherwise the code at the top of the
965 loop might never be executed. Unconditional jumps are
966 followed a by barrier then loop end. */
972 {
973 /* At the virtual top of a converted loop, insns are again known to
974 be executed: logically, the loop begins here even though the exit
975 code has been duplicated. */
979 loop_depth++;
981 loop_depth--;
982 }
983 }
985 /* If one movable subsumes another, ignore that other. */
989 /* For each movable insn, see if the reg that it loads
990 leads when it dies right into another conditionally movable insn.
991 If so, record that the second insn "forces" the first one,
992 since the second can be moved only if the first is. */
996 /* See if there are multiple movable insns that load the same value.
997 If there are, make all but the first point at the first one
998 through the `match' field, and add the priorities of them
999 all together as the priority of the first. */
1003 /* Now consider each movable insn to decide whether it is worth moving.
1004 Store 0 in regs->array[I].set_in_loop for each reg I that is moved.
1006 Generally this increases code size, so do not move moveables when
1007 optimizing for code size. */
1012 /* Now candidates that still are negative are those not moved.
1013 Change regs->array[I].set_in_loop to indicate that those are not actually
1014 invariant. */
1019 /* Now that we've moved some things out of the loop, we might be able to
1020 hoist even more memory references. */
1023 /* Recalculate regs->array if load_mems has created new registers. */
1031 ;
1038 {
1040 /* Ensure our label doesn't go away. */
1051 }
1054 /* The movable information is required for strength reduction. */
1060 }
1061 \f
1062 /* Add elements to *OUTPUT to record all the pseudo-regs
1063 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1065 void
1069 {
1077 {
1095 }
1099 {
1103 {
1112 }
1113 }
1114 }
1115 \f
1116 /* Check what regs are referred to in the libcall block ending with INSN,
1117 aside from those mentioned in the equivalent value.
1118 If there are none, return 0.
1119 If there are one or more, return an EXPR_LIST containing all of them. */
1121 rtx
1124 {
1129 /* First, find all the regs used in the libcall block
1130 that are not mentioned as inputs to the result. */
1133 {
1138 }
1141 }
1142 \f
1143 /* Return 1 if all uses of REG
1144 are between INSN and the end of the basic block. */
1146 static int
1149 {
1151 rtx p;
1156 /* Search this basic block for the already recorded last use of the reg. */
1158 {
1160 {
1166 /* Ordinary insn: if this is the last use, we win. */
1172 /* Jump insn: if this is the last use, we win. */
1175 /* Otherwise, it's the end of the basic block, so we lose. */
1180 /* It's the end of the basic block, so we lose. */
1185 }
1186 }
1188 /* The "last use" that was recorded can't be found after the first
1189 use. This can happen when the last use was deleted while
1190 processing an inner loop, this inner loop was then completely
1191 unrolled, and the outer loop is always exited after the inner loop,
1192 so that everything after the first use becomes a single basic block. */
1194 }
1195 \f
1196 /* Compute the benefit of eliminating the insns in the block whose
1197 last insn is LAST. This may be a group of insns used to compute a
1198 value directly or can contain a library call. */
1200 static int
1202 rtx last;
1203 {
1204 rtx insn;
1209 {
1212 routine. */
1216 benefit++;
1217 }
1220 }
1221 \f
1222 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1224 static rtx
1226 rtx insn;
1228 {
1230 {
1231 rtx temp;
1233 /* If first insn of libcall sequence, skip to end. */
1234 /* Do this at start of loop, since INSN is guaranteed to
1235 be an insn here. */
1240 do
1243 }
1246 }
1248 /* Ignore any movable whose insn falls within a libcall
1249 which is part of another movable.
1250 We make use of the fact that the movable for the libcall value
1251 was made later and so appears later on the chain. */
1253 static void
1256 {
1260 {
1261 /* Is this a movable for the value of a libcall? */
1264 {
1265 rtx insn;
1266 /* Check for earlier movables inside that range,
1267 and mark them invalid. We cannot use LUIDs here because
1268 insns created by loop.c for prior loops don't have LUIDs.
1269 Rather than reject all such insns from movables, we just
1270 explicitly check each insn in the libcall (since invariant
1271 libcalls aren't that common). */
1276 }
1277 }
1278 }
1280 /* For each movable insn, see if the reg that it loads
1281 leads when it dies right into another conditionally movable insn.
1282 If so, record that the second insn "forces" the first one,
1283 since the second can be moved only if the first is. */
1285 static void
1288 {
1291 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1293 {
1296 /* ??? Could this be a bug? What if CSE caused the
1297 register of M1 to be used after this insn?
1298 Since CSE does not update regno_last_uid,
1299 this insn M->insn might not be where it dies.
1300 But very likely this doesn't matter; what matters is
1301 that M's reg is computed from M1's reg. */
1306 /* If m->consec, m->set_src isn't valid. */
1310 /* Increase the priority of the moving the first insn
1311 since it permits the second to be moved as well. */
1313 {
1317 }
1318 }
1319 }
1320 \f
1321 /* Find invariant expressions that are equal and can be combined into
1322 one register. */
1324 static void
1328 {
1333 /* Regs that are set more than once are not allowed to match
1334 or be matched. I'm no longer sure why not. */
1335 /* Perhaps testing m->consec_sets would be more appropriate here? */
1340 {
1347 /* We want later insns to match the first one. Don't make the first
1348 one match any later ones. So start this loop at m->next. */
1352 /* A reg used outside the loop mustn't be eliminated. */
1354 /* A reg used for zero-extending mustn't be eliminated. */
1357 ||
1358 (
1359 /* Can combine regs with different modes loaded from the
1360 same constant only if the modes are the same or
1361 if both are integer modes with M wider or the same
1362 width as M1. The check for integer is redundant, but
1363 safe, since the only case of differing destination
1364 modes with equal sources is when both sources are
1365 VOIDmode, i.e., CONST_INT. */
1371 /* See if the source of M1 says it matches M. */
1378 {
1384 }
1385 }
1387 /* Now combine the regs used for zero-extension.
1388 This can be done for those not marked `global'
1389 provided their lives don't overlap. */
1393 {
1396 /* Combine all the registers for extension from mode MODE.
1397 Don't combine any that are used outside this loop. */
1401 {
1407 {
1408 /* First one: don't check for overlap, just record it. */
1411 }
1413 /* Make sure they extend to the same mode.
1414 (Almost always true.) */
1418 /* We already have one: check for overlap with those
1419 already combined together. */
1426 /* No overlap: we can combine this with the others. */
1432 overlap:
1433 ;
1434 }
1435 }
1437 /* Clean up. */
1439 }
1441 /* Returns the number of movable instructions in LOOP that were not
1442 moved outside the loop. */
1444 static int
1447 {
1456 }
1458 \f
1459 /* Return 1 if regs X and Y will become the same if moved. */
1461 static int
1465 {
1482 }
1484 /* Return 1 if X and Y are identical-looking rtx's.
1485 This is the Lisp function EQUAL for rtx arguments.
1487 If two registers are matching movables or a movable register and an
1488 equivalent constant, consider them equal. */
1490 static int
1495 {
1509 /* If we have a register and a constant, they may sometimes be
1510 equal. */
1513 {
1518 }
1521 {
1526 }
1528 /* Otherwise, rtx's of different codes cannot be equal. */
1532 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1533 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1538 /* These three types of rtx's can be compared nonrecursively. */
1547 /* Compare the elements. If any pair of corresponding elements
1548 fail to match, return 0 for the whole things. */
1552 {
1554 {
1566 /* Two vectors must have the same length. */
1570 /* And the corresponding elements must match. */
1589 /* These are just backpointers, so they don't matter. */
1595 /* It is believed that rtx's at this level will never
1596 contain anything but integers and other rtx's,
1597 except for within LABEL_REFs and SYMBOL_REFs. */
1600 }
1601 }
1603 }
1604 \f
1605 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1606 insns in INSNS which use the reference. LABEL_NUSES for CODE_LABEL
1607 references is incremented once for each added note. */
1609 static void
1611 rtx x;
1612 rtx insns;
1613 {
1617 rtx insn;
1620 {
1621 /* This code used to ignore labels that referred to dispatch tables to
1622 avoid flow generating (slighly) worse code.
1624 We no longer ignore such label references (see LABEL_REF handling in
1625 mark_jump_label for additional information). */
1628 {
1633 }
1634 }
1638 {
1644 }
1645 }
1646 \f
1647 /* Scan MOVABLES, and move the insns that deserve to be moved.
1648 If two matching movables are combined, replace one reg with the
1649 other throughout. */
1651 static void
1657 {
1665 /* Map of pseudo-register replacements to handle combining
1666 when we move several insns that load the same value
1667 into different pseudo-registers. */
1672 {
1673 /* Describe this movable insn. */
1676 {
1697 }
1699 /* Ignore the insn if it's already done (it matched something else).
1700 Otherwise, see if it is now safe to move. */
1712 {
1717 /* We have an insn that is safe to move.
1718 Compute its desirability. */
1729 /* An insn MUST be moved if we already moved something else
1730 which is safe only if this one is moved too: that is,
1731 if already_moved[REGNO] is nonzero. */
1733 /* An insn is desirable to move if the new lifetime of the
1734 register is no more than THRESHOLD times the old lifetime.
1735 If it's not desirable, it means the loop is so big
1736 that moving won't speed things up much,
1737 and it is liable to make register usage worse. */
1739 /* It is also desirable to move if it can be moved at no
1740 extra cost because something else was already moved. */
1743 || flag_move_all_movables
1748 {
1753 /* Now move the insns that set the reg. */
1756 {
1759 /* Find the end of this chain of matching regs.
1760 Thus, we load each reg in the chain from that one reg.
1761 And that reg is loaded with 0 directly,
1762 since it has ->match == 0. */
1768 /* Mark the moved, invariant reg as being allowed to
1769 share a hard reg with the other matching invariant. */
1773 regs_may_share
1776 regs_may_share));
1784 }
1785 /* If we are to re-generate the item being moved with a
1786 new move insn, first delete what we have and then emit
1787 the move insn before the loop. */
1789 {
1793 {
1794 /* If this is the first insn of a library call sequence,
1795 skip to the end. */
1800 /* If this is the last insn of a libcall sequence, then
1801 delete every insn in the sequence except the last.
1802 The last insn is handled in the normal manner. */
1805 {
1809 }
1814 /* simplify_giv_expr expects that it can walk the insns
1815 at m->insn forwards and see this old sequence we are
1816 tossing here. delete_insn does preserve the next
1817 pointers, but when we skip over a NOTE we must fix
1818 it up. Otherwise that code walks into the non-deleted
1819 insn stream. */
1822 }
1841 /* The more regs we move, the less we like moving them. */
1843 }
1844 else
1845 {
1847 {
1850 /* If first insn of libcall sequence, skip to end. */
1851 /* Do this at start of loop, since p is guaranteed to
1852 be an insn here. */
1857 /* If last insn of libcall sequence, move all
1858 insns except the last before the loop. The last
1859 insn is handled in the normal manner. */
1862 {
1870 {
1871 rtx body;
1872 rtx n;
1873 rtx next;
1880 /* Find the next insn after TEMP,
1881 not counting USE or NOTE insns. */
1889 /* If that is the call, this may be the insn
1890 that loads the function address.
1892 Extract the function address from the insn
1893 that loads it into a register.
1894 If this insn was cse'd, we get incorrect code.
1896 So emit a new move insn that copies the
1897 function address into the register that the
1898 call insn will use. flow.c will delete any
1899 redundant stores that we have created. */
1904 NULL_RTX)))
1905 {
1911 }
1912 /* We have the call insn.
1913 If it uses the register we suspect it might,
1914 load it with the correct address directly. */
1919 gen_move_insn
1923 {
1925 /* Because the USAGE information potentially
1926 contains objects other than hard registers
1927 we need to copy it. */
1931 }
1932 else
1940 }
1943 }
1945 {
1946 /* P sets REG to zero; but we should clear only
1947 the bits that are not covered by the mode
1948 m->savemode. */
1950 rtx sequence;
1951 rtx tem;
1954 tem = expand_simple_binop
1967 }
1969 {
1971 /* Because the USAGE information potentially
1972 contains objects other than hard registers
1973 we need to copy it. */
1977 }
1979 {
1980 rtx seq;
1981 /* The SET_SRC might not be invariant, so we must
1982 use the REG_EQUAL note. */
1997 }
1998 else
2002 {
2005 /* If there is a REG_EQUAL note present whose value
2006 is not loop invariant, then delete it, since it
2007 may cause problems with later optimization passes.
2008 It is possible for cse to create such notes
2009 like this as a result of record_jump_cond. */
2014 }
2023 /* If library call, now fix the REG_NOTES that contain
2024 insn pointers, namely REG_LIBCALL on FIRST
2025 and REG_RETVAL on I1. */
2027 {
2031 }
2037 /* simplify_giv_expr expects that it can walk the insns
2038 at m->insn forwards and see this old sequence we are
2039 tossing here. delete_insn does preserve the next
2040 pointers, but when we skip over a NOTE we must fix
2041 it up. Otherwise that code walks into the non-deleted
2042 insn stream. */
2045 }
2047 /* The more regs we move, the less we like moving them. */
2049 }
2051 /* Any other movable that loads the same register
2052 MUST be moved. */
2055 /* This reg has been moved out of one loop. */
2058 /* The reg set here is now invariant. */
2064 /* Change the length-of-life info for the register
2065 to say it lives at least the full length of this loop.
2066 This will help guide optimizations in outer loops. */
2069 /* This is the old insn before all the moved insns.
2070 We can't use the moved insn because it is out of range
2071 in uid_luid. Only the old insns have luids. */
2076 /* Combine with this moved insn any other matching movables. */
2081 {
2082 rtx temp;
2084 /* Schedule the reg loaded by M1
2085 for replacement so that shares the reg of M.
2086 If the modes differ (only possible in restricted
2087 circumstances, make a SUBREG.
2089 Note this assumes that the target dependent files
2090 treat REG and SUBREG equally, including within
2091 GO_IF_LEGITIMATE_ADDRESS and in all the
2092 predicates since we never verify that replacing the
2093 original register with a SUBREG results in a
2094 recognizable insn. */
2097 else
2102 /* Get rid of the matching insn
2103 and prevent further processing of it. */
2106 /* if library call, delete all insn except last, which
2107 is deleted below */
2109 NULL_RTX)))
2110 {
2114 }
2117 /* Any other movable that loads the same register
2118 MUST be moved. */
2121 /* The reg merged here is now invariant,
2122 if the reg it matches is invariant. */
2125 }
2126 }
2129 }
2135 }
2140 /* Go through all the instructions in the loop, making
2141 all the register substitutions scheduled in REG_MAP. */
2145 {
2149 }
2151 /* Clean up. */
2154 }
2157 static void
2161 {
2164 else
2167 }
2170 static void
2173 {
2178 {
2181 }
2182 }
2183 \f
2184 #if 0
2185 /* Scan X and replace the address of any MEM in it with ADDR.
2186 REG is the address that MEM should have before the replacement. */
2188 static void
2191 {
2200 {
2212 /* Short cut for very common case. */
2217 /* Short cut for very common case. */
2222 /* If this MEM uses a reg other than the one we expected,
2223 something is wrong. */
2231 }
2235 {
2239 {
2243 }
2244 }
2245 }
2246 #endif
2247 \f
2248 /* Return the number of memory refs to addresses that vary
2249 in the rtx X. */
2251 static int
2254 rtx x;
2255 {
2266 {
2283 }
2288 {
2292 {
2296 }
2297 }
2299 }
2300 \f
2301 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2302 `has_call', `has_nonconst_call', `has_volatile', `has_tablejump',
2303 `unknown_address_altered', `unknown_constant_address_altered', and
2304 `num_mem_sets' in LOOP. Also, fill in the array `mems' and the
2305 list `store_mems' in LOOP. */
2307 static void
2310 {
2312 rtx insn;
2316 /* The label after END. Jumping here is just like falling off the
2317 end of the loop. We use next_nonnote_insn instead of next_label
2318 as a hedge against the (pathological) case where some actual insn
2319 might end up between the two. */
2341 {
2343 {
2346 }
2347 }
2351 {
2353 {
2355 {
2357 /* Count number of loops contained in this one. */
2359 }
2361 {
2363 }
2364 }
2366 {
2368 {
2371 }
2373 }
2375 {
2395 {
2397 {
2400 }
2401 else
2402 {
2404 }
2406 do
2407 {
2409 {
2411 {
2412 /* Something tricky. */
2415 }
2418 {
2419 /* A jump outside the current loop. */
2422 }
2423 }
2427 }
2429 }
2430 }
2433 }
2435 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2437 anywhere. */
2439 /* We don't want loads for MEMs moved to a location before the
2440 one at which their stack memory becomes allocated. (Note
2441 that this is not a problem for malloc, etc., since those
2442 require actual function calls. */
2443 && ! current_function_calls_alloca
2444 /* There are ways to leave the loop other than falling off the
2445 end. */
2451 /* BLKmode MEMs are added to LOOP_STORE_MEM as necessary so
2452 that loop_invariant_p and load_mems can use true_dependence
2453 to determine what is really clobbered. */
2455 {
2458 loop_info->store_mems
2460 }
2462 {
2466 loop_info->store_mems
2468 }
2469 }
2470 \f
2471 /* Scan the function looking for loops. Record the start and end of each loop.
2472 Also mark as invalid loops any loops that contain a setjmp or are branched
2473 to from outside the loop. */
2475 static void
2477 rtx f;
2479 {
2480 rtx insn;
2481 rtx label;
2491 /* If there are jumps to undefined labels,
2492 treat them as jumps out of any/all loops.
2493 This also avoids writing past end of tables when there are no loops. */
2496 /* Find boundaries of loops, mark which loops are contained within
2497 loops, and invalidate loops that have setjmp. */
2502 {
2505 {
2509 num_loops++;
2533 }
2537 {
2538 /* In this case, we must invalidate our current loop and any
2539 enclosing loop. */
2541 {
2547 }
2548 }
2550 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2551 enclosing loop, but this doesn't matter. */
2553 }
2555 /* Any loop containing a label used in an initializer must be invalidated,
2556 because it can be jumped into from anywhere. */
2559 {
2563 }
2565 /* Any loop containing a label used for an exception handler must be
2566 invalidated, because it can be jumped into from anywhere. */
2569 {
2573 }
2575 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2576 loop that it is not contained within, that loop is marked invalid.
2577 If any INSN or CALL_INSN uses a label's address, then the loop containing
2578 that label is marked invalid, because it could be jumped into from
2579 anywhere.
2581 Also look for blocks of code ending in an unconditional branch that
2582 exits the loop. If such a block is surrounded by a conditional
2583 branch around the block, move the block elsewhere (see below) and
2584 invert the jump to point to the code block. This may eliminate a
2585 label in our loop and will simplify processing by both us and a
2586 possible second cse pass. */
2590 {
2594 {
2597 {
2601 }
2602 }
2609 /* See if this is an unconditional branch outside the loop. */
2617 {
2618 rtx p;
2624 /* Go backwards until we reach the start of the loop, a label,
2625 or a JUMP_INSN. */
2632 ;
2634 /* Check for the case where we have a jump to an inner nested
2635 loop, and do not perform the optimization in that case. */
2638 {
2641 {
2646 }
2647 }
2649 /* Make sure that the target of P is within the current loop. */
2655 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2656 we have a block of code to try to move.
2658 We look backward and then forward from the target of INSN
2659 to find a BARRIER at the same loop depth as the target.
2660 If we find such a BARRIER, we make a new label for the start
2661 of the block, invert the jump in P and point it to that label,
2662 and move the block of code to the spot we found. */
2667 /* Just ignore jumps to labels that were never emitted.
2668 These always indicate compilation errors. */
2672 /* If it's not safe to move the sequence, then we
2673 mustn't try. */
2676 {
2677 rtx target
2681 rtx tmp;
2683 /* Search for possible garbage past the conditional jumps
2684 and look for the last barrier. */
2692 /* Don't move things inside a tablejump. */
2705 /* Don't move things inside a tablejump. */
2716 {
2720 /* Ensure our label doesn't go away. */
2723 /* Verify that uid_loop is large enough and that
2724 we can invert P. */
2726 {
2729 /* If no suitable BARRIER was found, create a suitable
2730 one before TARGET. Since TARGET is a fall through
2731 path, we'll need to insert an jump around our block
2732 and a add a BARRIER before TARGET.
2734 This creates an extra unconditional jump outside
2735 the loop. However, the benefits of removing rarely
2736 executed instructions from inside the loop usually
2737 outweighs the cost of the extra unconditional jump
2738 outside the loop. */
2740 {
2741 rtx temp;
2748 }
2750 /* Include the BARRIER after INSN and copy the
2751 block after LOC. */
2753 last_insn_to_move);
2756 /* All those insns are now in TARGET_LOOP. */
2762 /* The label jumped to by INSN is no longer a loop
2763 exit. Unless INSN does not have a label (e.g.,
2764 it is a RETURN insn), search loop->exit_labels
2765 to find its label_ref, and remove it. Also turn
2766 off LABEL_OUTSIDE_LOOP_P bit. */
2768 {
2770 r;
2773 {
2777 else
2780 }
2786 /* If we didn't find it, then something is
2787 wrong. */
2790 }
2792 /* P is now a jump outside the loop, so it must be put
2793 in loop->exit_labels, and marked as such.
2794 The easiest way to do this is to just call
2795 mark_loop_jump again for P. */
2798 /* If INSN now jumps to the insn after it,
2799 delete INSN. */
2804 }
2806 /* Continue the loop after where the conditional
2807 branch used to jump, since the only branch insn
2808 in the block (if it still remains) is an inter-loop
2809 branch and hence needs no processing. */
2815 /* This loop will be continued with NEXT_INSN (insn). */
2817 }
2818 }
2819 }
2820 }
2821 }
2823 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2824 loops it is contained in, mark the target loop invalid.
2826 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2828 static void
2830 rtx x;
2832 {
2838 {
2850 /* There could be a label reference in here. */
2862 /* This may refer to a LABEL_REF or SYMBOL_REF. */
2874 /* Link together all labels that branch outside the loop. This
2875 is used by final_[bg]iv_value and the loop unrolling code. Also
2876 mark this LABEL_REF so we know that this branch should predict
2877 false. */
2879 /* A check to make sure the label is not in an inner nested loop,
2880 since this does not count as a loop exit. */
2882 {
2887 }
2888 else
2892 {
2901 }
2903 /* If this is inside a loop, but not in the current loop or one enclosed
2904 by it, it invalidates at least one loop. */
2909 /* We must invalidate every nested loop containing the target of this
2910 label, except those that also contain the jump insn. */
2913 {
2914 /* Stop when we reach a loop that also contains the jump insn. */
2919 /* If we get here, we know we need to invalidate a loop. */
2926 }
2930 /* If this is not setting pc, ignore. */
2952 /* Strictly speaking this is not a jump into the loop, only a possible
2953 jump out of the loop. However, we have no way to link the destination
2954 of this jump onto the list of exit labels. To be safe we mark this
2955 loop and any containing loops as invalid. */
2957 {
2959 {
2965 }
2966 }
2968 }
2969 }
2970 \f
2971 /* Return nonzero if there is a label in the range from
2972 insn INSN to and including the insn whose luid is END
2973 INSN must have an assigned luid (i.e., it must not have
2974 been previously created by loop.c). */
2976 static int
2978 rtx insn;
2980 {
2982 {
2986 }
2989 }
2991 /* Record that a memory reference X is being set. */
2993 static void
2995 rtx x;
2996 rtx y ATTRIBUTE_UNUSED;
2998 {
3004 /* Count number of memory writes.
3005 This affects heuristics in strength_reduce. */
3008 /* BLKmode MEM means all memory is clobbered. */
3010 {
3013 else
3017 }
3021 }
3023 /* X is a value modified by an INSN that references a biv inside a loop
3024 exit test (ie, X is somehow related to the value of the biv). If X
3025 is a pseudo that is used more than once, then the biv is (effectively)
3026 used more than once. DATA is a pointer to a loop_regs structure. */
3028 static void
3030 rtx x;
3031 rtx y ATTRIBUTE_UNUSED;
3033 {
3048 /* If we do not have usage information, or if we know the register
3049 is used more than once, note that fact for check_dbra_loop. */
3054 }
3055 \f
3056 /* Return nonzero if the rtx X is invariant over the current loop.
3058 The value is 2 if we refer to something only conditionally invariant.
3060 A memory ref is invariant if it is not volatile and does not conflict
3061 with anything stored in `loop_info->store_mems'. */
3063 int
3067 {
3074 rtx mem_list_entry;
3080 {
3088 /* A LABEL_REF is normally invariant, however, if we are unrolling
3089 loops, and this label is inside the loop, then it isn't invariant.
3090 This is because each unrolled copy of the loop body will have
3091 a copy of this label. If this was invariant, then an insn loading
3092 the address of this label into a register might get moved outside
3093 the loop, and then each loop body would end up using the same label.
3095 We don't know the loop bounds here though, so just fail for all
3096 labels. */
3099 else
3108 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3109 since the reg might be set by initialization within the loop. */
3126 /* Volatile memory references must be rejected. Do this before
3127 checking for read-only items, so that volatile read-only items
3128 will be rejected also. */
3132 /* See if there is any dependence between a store and this load. */
3135 {
3141 }
3143 /* It's not invalidated by a store in memory
3144 but we must still verify the address is invariant. */
3148 /* Don't mess with insns declared volatile. */
3155 }
3159 {
3161 {
3167 }
3169 {
3172 {
3178 }
3180 }
3181 }
3184 }
3185 \f
3186 /* Return nonzero if all the insns in the loop that set REG
3187 are INSN and the immediately following insns,
3188 and if each of those insns sets REG in an invariant way
3189 (not counting uses of REG in them).
3191 The value is 2 if some of these insns are only conditionally invariant.
3193 We assume that INSN itself is the first set of REG
3194 and that its source is invariant. */
3196 static int
3201 {
3205 rtx temp;
3206 /* Number of sets we have to insist on finding after INSN. */
3212 /* If N_SETS hit the limit, we can't rely on its value. */
3219 {
3221 rtx set;
3226 /* If library call, skip to end of it. */
3235 {
3240 {
3241 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3242 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3243 notes are OK. */
3249 }
3250 }
3252 count--;
3254 {
3257 }
3258 }
3261 /* If loop_invariant_p ever returned 2, we return 2. */
3263 }
3265 #if 0
3266 /* I don't think this condition is sufficient to allow INSN
3267 to be moved, so we no longer test it. */
3269 /* Return 1 if all insns in the basic block of INSN and following INSN
3270 that set REG are invariant according to TABLE. */
3272 static int
3276 {
3281 {
3290 {
3293 }
3294 }
3295 }
3297 \f
3298 /* Look at all uses (not sets) of registers in X. For each, if it is
3299 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3300 a different insn, set USAGE[REGNO] to const0_rtx. */
3302 static void
3305 rtx insn;
3306 rtx x;
3307 {
3319 {
3320 /* Don't count SET_DEST if it is a REG; otherwise count things
3321 in SET_DEST because if a register is partially modified, it won't
3322 show up as a potential movable so we don't care how USAGE is set
3323 for it. */
3327 }
3328 else
3330 {
3336 }
3337 }
3338 \f
3339 /* Count and record any set in X which is contained in INSN. Update
3340 REGS->array[I].MAY_NOT_OPTIMIZE and LAST_SET for any register I set
3341 in X. */
3343 static void
3348 {
3350 /* Don't move a reg that has an explicit clobber.
3351 It's not worth the pain to try to do it correctly. */
3355 {
3363 {
3365 /* If this is the first setting of this reg
3366 in current basic block, and it was set before,
3367 it must be set in two basic blocks, so it cannot
3368 be moved out of the loop. */
3372 /* If this is not first setting in current basic block,
3373 see if reg was used in between previous one and this.
3374 If so, neither one can be moved. */
3381 }
3382 }
3383 }
3384 \f
3385 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3386 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3387 contained in insn INSN is used by any insn that precedes INSN in
3388 cyclic order starting from the loop entry point.
3390 We don't want to use INSN_LUID here because if we restrict INSN to those
3391 that have a valid INSN_LUID, it means we cannot move an invariant out
3392 from an inner loop past two loops. */
3394 static int
3398 {
3400 rtx p;
3402 /* Scan forward checking for register usage. If we hit INSN, we
3403 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3405 {
3411 }
3414 }
3415 \f
3416 /* A "basic induction variable" or biv is a pseudo reg that is set
3417 (within this loop) only by incrementing or decrementing it. */
3418 /* A "general induction variable" or giv is a pseudo reg whose
3419 value is a linear function of a biv. */
3421 /* Bivs are recognized by `basic_induction_var';
3422 Givs by `general_induction_var'. */
3424 /* Communication with routines called via `note_stores'. */
3428 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3432 /* ??? Unfinished optimizations, and possible future optimizations,
3433 for the strength reduction code. */
3435 /* ??? The interaction of biv elimination, and recognition of 'constant'
3436 bivs, may cause problems. */
3438 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3439 performance problems.
3441 Perhaps don't eliminate things that can be combined with an addressing
3442 mode. Find all givs that have the same biv, mult_val, and add_val;
3443 then for each giv, check to see if its only use dies in a following
3444 memory address. If so, generate a new memory address and check to see
3445 if it is valid. If it is valid, then store the modified memory address,
3446 otherwise, mark the giv as not done so that it will get its own iv. */
3448 /* ??? Could try to optimize branches when it is known that a biv is always
3449 positive. */
3451 /* ??? When replace a biv in a compare insn, we should replace with closest
3452 giv so that an optimized branch can still be recognized by the combiner,
3453 e.g. the VAX acb insn. */
3455 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3456 was rerun in loop_optimize whenever a register was added or moved.
3457 Also, some of the optimizations could be a little less conservative. */
3458 \f
3459 /* Scan the loop body and call FNCALL for each insn. In the addition to the
3460 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
3461 callback.
3463 NOT_EVERY_ITERATION if current insn is not executed at least once for every
3464 loop iteration except for the last one.
3466 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
3467 loop iteration.
3468 */
3469 void
3472 loop_insn_callback fncall;
3473 {
3474 /* This is 1 if current insn is not executed at least once for every loop
3475 iteration. */
3480 rtx p;
3482 /* If loop_scan_start points to the loop exit test, we have to be wary of
3483 subversive use of gotos inside expression statements. */
3487 /* Scan through loop to find all possible bivs. */
3492 {
3495 /* Past CODE_LABEL, we get to insns that may be executed multiple
3496 times. The only way we can be sure that they can't is if every
3497 jump insn between here and the end of the loop either
3498 returns, exits the loop, is a jump to a location that is still
3499 behind the label, or is a jump to the loop start. */
3502 {
3508 {
3513 {
3516 else
3520 }
3528 {
3531 }
3532 }
3533 }
3535 /* Past a jump, we get to insns for which we can't count
3536 on whether they will be executed during each iteration. */
3537 /* This code appears twice in strength_reduce. There is also similar
3538 code in scan_loop. */
3540 /* If we enter the loop in the middle, and scan around to the
3541 beginning, don't set not_every_iteration for that.
3542 This can be any kind of jump, since we want to know if insns
3543 will be executed if the loop is executed. */
3548 {
3551 /* If this is a jump outside the loop, then it also doesn't
3552 matter. Check to see if the target of this branch is on the
3553 loop->exits_labels list. */
3561 }
3564 {
3565 /* At the virtual top of a converted loop, insns are again known to
3566 be executed each iteration: logically, the loop begins here
3567 even though the exit code has been duplicated.
3569 Insns are also again known to be executed each iteration at
3570 the LOOP_CONT note. */
3576 loop_depth++;
3578 loop_depth--;
3579 }
3581 /* Note if we pass a loop latch. If we do, then we can not clear
3582 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3583 a loop since a jump before the last CODE_LABEL may have started
3584 a new loop iteration.
3586 Note that LOOP_TOP is only set for rotated loops and we need
3587 this check for all loops, so compare against the CODE_LABEL
3588 which immediately follows LOOP_START. */
3593 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3594 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3595 or not an insn is known to be executed each iteration of the
3596 loop, whether or not any iterations are known to occur.
3598 Therefore, if we have just passed a label and have no more labels
3599 between here and the test insn of the loop, and we have not passed
3600 a jump to the top of the loop, then we know these insns will be
3601 executed each iteration. */
3604 && !past_loop_latch
3609 }
3610 }
3611 \f
3612 static void
3615 {
3618 /* Temporary list pointers for traversing ivs->list. */
3625 /* Scan ivs->list to remove all regs that proved not to be bivs.
3626 Make a sanity check against regs->n_times_set. */
3628 {
3630 /* Above happens if register modified by subreg, etc. */
3631 /* Make sure it is not recognized as a basic induction var: */
3633 /* If never incremented, it is invariant that we decided not to
3634 move. So leave it alone. */
3636 {
3647 }
3648 else
3649 {
3654 }
3655 }
3656 }
3659 /* Determine how BIVS are initialised by looking through pre-header
3660 extended basic block. */
3661 static void
3664 {
3666 /* Temporary list pointers for traversing ivs->list. */
3669 rtx p;
3671 /* Find initial value for each biv by searching backwards from loop_start,
3672 halting at first label. Also record any test condition. */
3676 {
3677 rtx test;
3687 /* Record any test of a biv that branches around the loop if no store
3688 between it and the start of loop. We only care about tests with
3689 constants and registers and only certain of those. */
3699 {
3700 /* If an NE test, we have an initial value! */
3702 {
3706 }
3707 else
3709 }
3710 }
3711 }
3714 /* Look at the each biv and see if we can say anything better about its
3715 initial value from any initializing insns set up above. (This is done
3716 in two passes to avoid missing SETs in a PARALLEL.) */
3717 static void
3720 {
3722 /* Temporary list pointers for traversing ivs->list. */
3727 {
3728 rtx src;
3729 rtx note;
3734 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3735 is a constant, use the value of that. */
3741 else
3754 {
3758 {
3761 }
3762 }
3763 /* If we can't make it a giv,
3764 let biv keep initial value of "itself". */
3767 }
3768 }
3771 /* Search the loop for general induction variables. */
3773 static void
3776 {
3778 }
3781 /* For each giv for which we still don't know whether or not it is
3782 replaceable, check to see if it is replaceable because its final value
3783 can be calculated. */
3785 static void
3788 {
3793 {
3799 }
3800 }
3803 /* Return non-zero if it is possible to eliminate the biv BL provided
3804 all givs are reduced. This is possible if either the reg is not
3805 used outside the loop, or we can compute what its final value will
3806 be. */
3808 static int
3814 {
3815 /* For architectures with a decrement_and_branch_until_zero insn,
3816 don't do this if we put a REG_NONNEG note on the endtest for this
3817 biv. */
3819 #ifdef HAVE_decrement_and_branch_until_zero
3821 {
3826 }
3827 #endif
3829 /* Check that biv is used outside loop or if it has a final value.
3830 Compare against bl->init_insn rather than loop->start. We aren't
3831 concerned with any uses of the biv between init_insn and
3832 loop->start since these won't be affected by the value of the biv
3833 elsewhere in the function, so long as init_insn doesn't use the
3834 biv itself. */
3845 {
3853 }
3855 }
3858 /* Reduce each giv of BL that we have decided to reduce. */
3860 static void
3864 {
3868 {
3871 {
3874 /* If the code for derived givs immediately below has already
3875 allocated a new_reg, we must keep it. */
3879 #ifdef AUTO_INC_DEC
3880 /* If the target has auto-increment addressing modes, and
3881 this is an address giv, then try to put the increment
3882 immediately after its use, so that flow can create an
3883 auto-increment addressing mode. */
3886 /* We don't handle reversed biv's because bl->biv->insn
3887 does not have a valid INSN_LUID. */
3891 {
3892 /* If other giv's have been combined with this one, then
3893 this will work only if all uses of the other giv's occur
3894 before this giv's insn. This is difficult to check.
3896 We simplify this by looking for the common case where
3897 there is one DEST_REG giv, and this giv's insn is the
3898 last use of the dest_reg of that DEST_REG giv. If the
3899 increment occurs after the address giv, then we can
3900 perform the optimization. (Otherwise, the increment
3901 would have to go before other_giv, and we would not be
3902 able to combine it with the address giv to get an
3903 auto-inc address.) */
3905 {
3910 {
3913 else
3915 }
3922 }
3923 /* Check for case where increment is before the address
3924 giv. Do this test in "loop order". */
3933 else
3936 #ifdef HAVE_cc0
3937 {
3938 rtx prev;
3940 /* We can't put an insn immediately after one setting
3941 cc0, or immediately before one using cc0. */
3948 }
3949 #endif
3953 }
3954 #endif
3956 /* For each place where the biv is incremented, add an insn
3957 to increment the new, reduced reg for the giv. */
3959 {
3960 rtx insert_before;
3966 else
3974 /* A multiply is acceptable here
3975 since this is presumed to be seldom executed. */
3979 }
3981 /* Add code at loop start to initialize giv's reduced reg. */
3986 }
3987 }
3988 }
3991 /* Check for givs whose first use is their definition and whose
3992 last use is the definition of another giv. If so, it is likely
3993 dead and should not be used to derive another giv nor to
3994 eliminate a biv. */
3996 static void
4000 {
4004 {
4011 {
4017 }
4018 }
4019 }
4022 static void
4027 {
4031 {
4038 /* Update expression if this was combined, in case other giv was
4039 replaced. */
4044 /* See if this register is known to be a pointer to something. If
4045 so, see if we can find the alignment. First see if there is a
4046 destination register that is a pointer. If so, this shares the
4047 alignment too. Next see if we can deduce anything from the
4048 computational information. If not, and this is a DEST_ADDR
4049 giv, at least we know that it's a pointer, though we don't know
4050 the alignment. */
4058 {
4067 }
4071 {
4079 }
4084 /* Store reduced reg as the address in the memref where we found
4085 this giv. */
4088 {
4090 }
4091 else
4092 {
4093 /* Not replaceable; emit an insn to set the original giv reg from
4094 the reduced giv, same as above. */
4097 }
4099 /* When a loop is reversed, givs which depend on the reversed
4100 biv, and which are live outside the loop, must be set to their
4101 correct final value. This insn is only needed if the giv is
4102 not replaceable. The correct final value is the same as the
4103 value that the giv starts the reversed loop with. */
4113 {
4118 }
4119 }
4120 }
4123 static int
4128 rtx test_reg;
4129 {
4138 /* Reduce benefit if not replaceable, since we will insert a
4139 move-insn to replace the insn that calculates this giv. Don't do
4140 this unless the giv is a user variable, since it will often be
4141 marked non-replaceable because of the duplication of the exit
4142 code outside the loop. In such a case, the copies we insert are
4143 dead and will be deleted. So they don't have a cost. Similar
4144 situations exist. */
4145 /* ??? The new final_[bg]iv_value code does a much better job of
4146 finding replaceable giv's, and hence this code may no longer be
4147 necessary. */
4152 /* Decrease the benefit to count the add-insns that we will insert
4153 to increment the reduced reg for the giv. ??? This can
4154 overestimate the run-time cost of the additional insns, e.g. if
4155 there are multiple basic blocks that increment the biv, but only
4156 one of these blocks is executed during each iteration. There is
4157 no good way to detect cases like this with the current structure
4158 of the loop optimizer. This code is more accurate for
4159 determining code size than run-time benefits. */
4162 /* Decide whether to strength-reduce this giv or to leave the code
4163 unchanged (recompute it from the biv each time it is used). This
4164 decision can be made independently for each giv. */
4166 #ifdef AUTO_INC_DEC
4167 /* Attempt to guess whether autoincrement will handle some of the
4168 new add insns; if so, increase BENEFIT (undo the subtraction of
4169 add_cost that was done above). */
4171 /* Increasing the benefit is risky, since this is only a guess.
4172 Avoid increasing register pressure in cases where there would
4173 be no other benefit from reducing this giv. */
4176 {
4191 }
4192 #endif
4195 }
4198 /* Free IV structures for LOOP. */
4200 static void
4203 {
4210 {
4216 {
4219 }
4221 {
4224 }
4228 }
4229 }
4232 /* Perform strength reduction and induction variable elimination.
4234 Pseudo registers created during this function will be beyond the
4235 last valid index in several tables including
4236 REGS->ARRAY[I].N_TIMES_SET and REGNO_LAST_UID. This does not cause a
4237 problem here, because the added registers cannot be givs outside of
4238 their loop, and hence will never be reconsidered. But scan_loop
4239 must check regnos to make sure they are in bounds. */
4241 static void
4245 {
4249 rtx p;
4250 /* Temporary list pointer for traversing ivs->list. */
4252 /* Ratio of extra register life span we can justify
4253 for saving an instruction. More if loop doesn't call subroutines
4254 since in that case saving an insn makes more difference
4255 and more registers are available. */
4256 /* ??? could set this to last value of threshold in move_movables */
4258 /* Map of pseudo-register replacements. */
4270 /* Find all BIVs in loop. */
4273 /* Exit if there are no bivs. */
4275 {
4276 /* Can still unroll the loop anyways, but indicate that there is no
4277 strength reduction info available. */
4283 }
4285 /* Determine how BIVS are initialised by looking through pre-header
4286 extended basic block. */
4289 /* Look at the each biv and see if we can say anything better about its
4290 initial value from any initializing insns set up above. */
4293 /* Search the loop for general induction variables. */
4296 /* Try to calculate and save the number of loop iterations. This is
4297 set to zero if the actual number can not be calculated. This must
4298 be called after all giv's have been identified, since otherwise it may
4299 fail if the iteration variable is a giv. */
4302 /* Now for each giv for which we still don't know whether or not it is
4303 replaceable, check to see if it is replaceable because its final value
4304 can be calculated. This must be done after loop_iterations is called,
4305 so that final_giv_value will work correctly. */
4308 /* Try to prove that the loop counter variable (if any) is always
4309 nonnegative; if so, record that fact with a REG_NONNEG note
4310 so that "decrement and branch until zero" insn can be used. */
4313 /* Create reg_map to hold substitutions for replaceable giv regs.
4314 Some givs might have been made from biv increments, so look at
4315 ivs->reg_iv_type for a suitable size. */
4319 /* Examine each iv class for feasibility of strength reduction/induction
4320 variable elimination. */
4323 {
4327 /* Test whether it will be possible to eliminate this biv
4328 provided all givs are reduced. */
4331 /* This will be true at the end, if all givs which depend on this
4332 biv have been strength reduced.
4333 We can't (currently) eliminate the biv unless this is so. */
4336 /* Check each extension dependent giv in this class to see if its
4337 root biv is safe from wrapping in the interior mode. */
4340 /* Combine all giv's for this iv_class. */
4344 {
4352 /* If an insn is not to be strength reduced, then set its ignore
4353 flag, and clear bl->all_reduced. */
4355 /* A giv that depends on a reversed biv must be reduced if it is
4356 used after the loop exit, otherwise, it would have the wrong
4357 value after the loop exit. To make it simple, just reduce all
4358 of such giv's whether or not we know they are used after the loop
4359 exit. */
4364 {
4372 }
4373 else
4374 {
4375 /* Check that we can increment the reduced giv without a
4376 multiply insn. If not, reject it. */
4381 {
4389 }
4390 }
4391 }
4393 /* Check for givs whose first use is their definition and whose
4394 last use is the definition of another giv. If so, it is likely
4395 dead and should not be used to derive another giv nor to
4396 eliminate a biv. */
4399 /* Reduce each giv that we decided to reduce. */
4402 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4403 as not reduced.
4405 For each giv register that can be reduced now: if replaceable,
4406 substitute reduced reg wherever the old giv occurs;
4407 else add new move insn "giv_reg = reduced_reg". */
4410 /* All the givs based on the biv bl have been reduced if they
4411 merit it. */
4413 /* For each giv not marked as maybe dead that has been combined with a
4414 second giv, clear any "maybe dead" mark on that second giv.
4415 v->new_reg will either be or refer to the register of the giv it
4416 combined with.
4418 Doing this clearing avoids problems in biv elimination where
4419 a giv's new_reg is a complex value that can't be put in the
4420 insn but the giv combined with (with a reg as new_reg) is
4421 marked maybe_dead. Since the register will be used in either
4422 case, we'd prefer it be used from the simpler giv. */
4428 /* Try to eliminate the biv, if it is a candidate.
4429 This won't work if ! bl->all_reduced,
4430 since the givs we planned to use might not have been reduced.
4432 We have to be careful that we didn't initially think we could
4433 eliminate this biv because of a giv that we now think may be
4434 dead and shouldn't be used as a biv replacement.
4436 Also, there is the possibility that we may have a giv that looks
4437 like it can be used to eliminate a biv, but the resulting insn
4438 isn't valid. This can happen, for example, on the 88k, where a
4439 JUMP_INSN can compare a register only with zero. Attempts to
4440 replace it with a compare with a constant will fail.
4442 Note that in cases where this call fails, we may have replaced some
4443 of the occurrences of the biv with a giv, but no harm was done in
4444 doing so in the rare cases where it can occur. */
4448 {
4449 /* ?? If we created a new test to bypass the loop entirely,
4450 or otherwise drop straight in, based on this test, then
4451 we might want to rewrite it also. This way some later
4452 pass has more hope of removing the initialization of this
4453 biv entirely. */
4455 /* If final_value != 0, then the biv may be used after loop end
4456 and we must emit an insn to set it just in case.
4458 Reversed bivs already have an insn after the loop setting their
4459 value, so we don't need another one. We can't calculate the
4460 proper final value for such a biv here anyways. */
4468 }
4469 }
4471 /* Go through all the instructions in the loop, making all the
4472 register substitutions scheduled in REG_MAP. */
4477 {
4481 }
4484 {
4485 /* When we completely unroll a loop we will likely not need the increment
4486 of the loop BIV and we will not need the conditional branch at the
4487 end of the loop. */
4490 #ifdef HAVE_cc0
4491 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
4492 need the comparison before the conditional branch at the end of the
4493 loop. */
4495 #endif
4497 /* We'll need one copy for each loop iteration. */
4500 /* A little slop to account for the ability to remove initialization
4501 code, better CSE, and other secondary benefits of completely
4502 unrolling some loops. */
4505 /* Clamp the value. */
4508 }
4510 /* Unroll loops from within strength reduction so that we can use the
4511 induction variable information that strength_reduce has already
4512 collected. Always unroll loops that would be as small or smaller
4513 unrolled than when rolled. */
4519 #ifdef HAVE_doloop_end
4524 /* In case number of iterations is known, drop branch prediction note
4525 in the branch. Do that only in second loop pass, as loop unrolling
4526 may change the number of iterations performed. */
4529 {
4532 PRED_LOOP_ITERATIONS,
4534 }
4542 }
4543 \f
4544 /*Record all basic induction variables calculated in the insn. */
4545 static rtx
4548 rtx p;
4551 {
4553 rtx set;
4554 rtx dest_reg;
4555 rtx inc_val;
4556 rtx mult_val;
4562 {
4567 {
4572 {
4573 /* It is a possible basic induction variable.
4574 Create and initialize an induction structure for it. */
4582 }
4585 }
4586 }
4588 }
4589 \f
4590 /* Record all givs calculated in the insn.
4591 A register is a giv if: it is only set once, it is a function of a
4592 biv and a constant (or invariant), and it is not a biv. */
4593 static rtx
4596 rtx p;
4599 {
4602 rtx set;
4603 /* Look for a general induction variable in a register. */
4608 {
4609 rtx src_reg;
4610 rtx dest_reg;
4611 rtx add_val;
4612 rtx mult_val;
4613 rtx ext_val;
4616 rtx last_consec_insn;
4625 /* Equivalent expression is a giv. */
4630 /* Don't try to handle any regs made by loop optimization.
4631 We have nothing on them in regno_first_uid, etc. */
4633 /* Don't recognize a BASIC_INDUCT_VAR here. */
4635 /* This must be the only place where the register is set. */
4637 /* or all sets must be consecutive and make a giv. */
4642 {
4646 /* If this is a library call, increase benefit. */
4650 /* Skip the consecutive insns, if there are any. */
4658 }
4659 }
4661 #ifndef DONT_REDUCE_ADDR
4662 /* Look for givs which are memory addresses. */
4663 /* This resulted in worse code on a VAX 8600. I wonder if it
4664 still does. */
4667 maybe_multiple);
4668 #endif
4670 /* Update the status of whether giv can derive other givs. This can
4671 change when we pass a label or an insn that updates a biv. */
4676 }
4677 \f
4678 /* Return 1 if X is a valid source for an initial value (or as value being
4679 compared against in an initial test).
4681 X must be either a register or constant and must not be clobbered between
4682 the current insn and the start of the loop.
4684 INSN is the insn containing X. */
4686 static int
4688 rtx x;
4689 rtx insn;
4691 rtx loop_start;
4692 {
4696 /* Only consider pseudos we know about initialized in insns whose luids
4697 we know. */
4702 /* Don't use call-clobbered registers across a call which clobbers it. On
4703 some machines, don't use any hard registers at all. */
4705 && (SMALL_REGISTER_CLASSES
4709 /* Don't use registers that have been clobbered before the start of the
4710 loop. */
4715 }
4716 \f
4717 /* Scan X for memory refs and check each memory address
4718 as a possible giv. INSN is the insn whose pattern X comes from.
4719 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
4720 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
4721 more thanonce in each loop iteration. */
4723 static void
4726 rtx x;
4727 rtx insn;
4729 {
4739 {
4755 {
4756 rtx src_reg;
4757 rtx add_val;
4758 rtx mult_val;
4759 rtx ext_val;
4762 /* This code used to disable creating GIVs with mult_val == 1 and
4763 add_val == 0. However, this leads to lost optimizations when
4764 it comes time to combine a set of related DEST_ADDR GIVs, since
4765 this one would not be seen. */
4770 {
4771 /* Found one; record it. */
4780 }
4781 }
4786 }
4788 /* Recursively scan the subexpressions for other mem refs. */
4794 maybe_multiple);
4798 maybe_multiple);
4799 }
4800 \f
4801 /* Fill in the data about one biv update.
4802 V is the `struct induction' in which we record the biv. (It is
4803 allocated by the caller, with alloca.)
4804 INSN is the insn that sets it.
4805 DEST_REG is the biv's reg.
4807 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4808 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4809 being set to INC_VAL.
4811 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4812 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4813 can be executed more than once per iteration. If MAYBE_MULTIPLE
4814 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4815 executed exactly once per iteration. */
4817 static void
4822 rtx insn;
4823 rtx dest_reg;
4824 rtx inc_val;
4825 rtx mult_val;
4829 {
4845 /* Add this to the reg's iv_class, creating a class
4846 if this is the first incrementation of the reg. */
4850 {
4851 /* Create and initialize new iv_class. */
4861 /* Set initial value to the reg itself. */
4864 /* We haven't seen the initializing insn yet */
4874 /* Add this class to ivs->list. */
4878 /* Put it in the array of biv register classes. */
4880 }
4882 /* Update IV_CLASS entry for this biv. */
4891 }
4892 \f
4893 /* Fill in the data about one giv.
4894 V is the `struct induction' in which we record the giv. (It is
4895 allocated by the caller, with alloca.)
4896 INSN is the insn that sets it.
4897 BENEFIT estimates the savings from deleting this insn.
4898 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4899 into a register or is used as a memory address.
4901 SRC_REG is the biv reg which the giv is computed from.
4902 DEST_REG is the giv's reg (if the giv is stored in a reg).
4903 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4904 LOCATION points to the place where this giv's value appears in INSN. */
4906 static void
4911 rtx insn;
4912 rtx src_reg;
4913 rtx dest_reg;
4919 {
4924 rtx temp;
4926 /* Attempt to prove constantness of the values. */
4954 /* The v->always_computable field is used in update_giv_derive, to
4955 determine whether a giv can be used to derive another giv. For a
4956 DEST_REG giv, INSN computes a new value for the giv, so its value
4957 isn't computable if INSN insn't executed every iteration.
4958 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4959 it does not compute a new value. Hence the value is always computable
4960 regardless of whether INSN is executed each iteration. */
4964 else
4970 {
4973 }
4975 {
4980 /* If the lifetime is zero, it means that this register is
4981 really a dead store. So mark this as a giv that can be
4982 ignored. This will not prevent the biv from being eliminated. */
4988 }
4990 /* Add the giv to the class of givs computed from one biv. */
4994 {
4997 /* Don't count DEST_ADDR. This is supposed to count the number of
4998 insns that calculate givs. */
5002 }
5003 else
5004 /* Fatal error, biv missing for this giv? */
5009 else
5010 {
5011 /* The giv can be replaced outright by the reduced register only if all
5012 of the following conditions are true:
5013 - the insn that sets the giv is always executed on any iteration
5014 on which the giv is used at all
5015 (there are two ways to deduce this:
5016 either the insn is executed on every iteration,
5017 or all uses follow that insn in the same basic block),
5018 - the giv is not used outside the loop
5019 - no assignments to the biv occur during the giv's lifetime. */
5022 /* Previous line always fails if INSN was moved by loop opt. */
5025 && (! not_every_iteration
5027 {
5028 /* Now check that there are no assignments to the biv within the
5029 giv's lifetime. This requires two separate checks. */
5031 /* Check each biv update, and fail if any are between the first
5032 and last use of the giv.
5034 If this loop contains an inner loop that was unrolled, then
5035 the insn modifying the biv may have been emitted by the loop
5036 unrolling code, and hence does not have a valid luid. Just
5037 mark the biv as not replaceable in this case. It is not very
5038 useful as a biv, because it is used in two different loops.
5039 It is very unlikely that we would be able to optimize the giv
5040 using this biv anyways. */
5044 {
5050 {
5054 }
5055 }
5057 /* If there are any backwards branches that go from after the
5058 biv update to before it, then this giv is not replaceable. */
5062 {
5066 }
5067 }
5068 else
5069 {
5070 /* May still be replaceable, we don't have enough info here to
5071 decide. */
5074 }
5075 }
5077 /* Record whether the add_val contains a const_int, for later use by
5078 combine_givs. */
5079 {
5084 ;
5088 {
5090 {
5095 else
5097 }
5100 }
5101 }
5105 }
5107 /* All this does is determine whether a giv can be made replaceable because
5108 its final value can be calculated. This code can not be part of record_giv
5109 above, because final_giv_value requires that the number of loop iterations
5110 be known, and that can not be accurately calculated until after all givs
5111 have been identified. */
5113 static void
5117 {
5124 /* DEST_ADDR givs will never reach here, because they are always marked
5125 replaceable above in record_giv. */
5127 /* The giv can be replaced outright by the reduced register only if all
5128 of the following conditions are true:
5129 - the insn that sets the giv is always executed on any iteration
5130 on which the giv is used at all
5131 (there are two ways to deduce this:
5132 either the insn is executed on every iteration,
5133 or all uses follow that insn in the same basic block),
5134 - its final value can be calculated (this condition is different
5135 than the one above in record_giv)
5136 - it's not used before the it's set
5137 - no assignments to the biv occur during the giv's lifetime. */
5139 #if 0
5140 /* This is only called now when replaceable is known to be false. */
5141 /* Clear replaceable, so that it won't confuse final_giv_value. */
5143 #endif
5147 {
5150 rtx last_giv_use;
5154 /* When trying to determine whether or not a biv increment occurs
5155 during the lifetime of the giv, we can ignore uses of the variable
5156 outside the loop because final_value is true. Hence we can not
5157 use regno_last_uid and regno_first_uid as above in record_giv. */
5159 /* Search the loop to determine whether any assignments to the
5160 biv occur during the giv's lifetime. Start with the insn
5161 that sets the giv, and search around the loop until we come
5162 back to that insn again.
5164 Also fail if there is a jump within the giv's lifetime that jumps
5165 to somewhere outside the lifetime but still within the loop. This
5166 catches spaghetti code where the execution order is not linear, and
5167 hence the above test fails. Here we assume that the giv lifetime
5168 does not extend from one iteration of the loop to the next, so as
5169 to make the test easier. Since the lifetime isn't known yet,
5170 this requires two loops. See also record_giv above. */
5175 {
5178 {
5181 }
5187 {
5188 /* It is possible for the BIV increment to use the GIV if we
5189 have a cycle. Thus we must be sure to check each insn for
5190 both BIV and GIV uses, and we must check for BIV uses
5191 first. */
5198 {
5200 {
5204 }
5206 }
5207 }
5208 }
5210 /* Now that the lifetime of the giv is known, check for branches
5211 from within the lifetime to outside the lifetime if it is still
5212 replaceable. */
5215 {
5218 {
5231 {
5240 }
5241 }
5242 }
5244 /* If it is replaceable, then save the final value. */
5247 }
5252 }
5253 \f
5254 /* Update the status of whether a giv can derive other givs.
5256 We need to do something special if there is or may be an update to the biv
5257 between the time the giv is defined and the time it is used to derive
5258 another giv.
5260 In addition, a giv that is only conditionally set is not allowed to
5261 derive another giv once a label has been passed.
5263 The cases we look at are when a label or an update to a biv is passed. */
5265 static void
5268 rtx p;
5269 {
5273 rtx tem;
5276 /* Search all IV classes, then all bivs, and finally all givs.
5278 There are three cases we are concerned with. First we have the situation
5279 of a giv that is only updated conditionally. In that case, it may not
5280 derive any givs after a label is passed.
5282 The second case is when a biv update occurs, or may occur, after the
5283 definition of a giv. For certain biv updates (see below) that are
5284 known to occur between the giv definition and use, we can adjust the
5285 giv definition. For others, or when the biv update is conditional,
5286 we must prevent the giv from deriving any other givs. There are two
5287 sub-cases within this case.
5289 If this is a label, we are concerned with any biv update that is done
5290 conditionally, since it may be done after the giv is defined followed by
5291 a branch here (actually, we need to pass both a jump and a label, but
5292 this extra tracking doesn't seem worth it).
5294 If this is a jump, we are concerned about any biv update that may be
5295 executed multiple times. We are actually only concerned about
5296 backward jumps, but it is probably not worth performing the test
5297 on the jump again here.
5299 If this is a biv update, we must adjust the giv status to show that a
5300 subsequent biv update was performed. If this adjustment cannot be done,
5301 the giv cannot derive further givs. */
5307 {
5309 {
5310 /* If cant_derive is already true, there is no point in
5311 checking all of these conditions again. */
5315 /* If this giv is conditionally set and we have passed a label,
5316 it cannot derive anything. */
5320 /* Skip givs that have mult_val == 0, since
5321 they are really invariants. Also skip those that are
5322 replaceable, since we know their lifetime doesn't contain
5323 any biv update. */
5327 /* The only way we can allow this giv to derive another
5328 is if this is a biv increment and we can form the product
5329 of biv->add_val and giv->mult_val. In this case, we will
5330 be able to compute a compensation. */
5332 {
5333 rtx ext_val_dummy;
5344 tem = simplify_giv_expr
5351 else
5353 }
5357 }
5358 }
5359 }
5360 \f
5361 /* Check whether an insn is an increment legitimate for a basic induction var.
5362 X is the source of insn P, or a part of it.
5363 MODE is the mode in which X should be interpreted.
5365 DEST_REG is the putative biv, also the destination of the insn.
5366 We accept patterns of these forms:
5367 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5368 REG = INVARIANT + REG
5370 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5371 store the additive term into *INC_VAL, and store the place where
5372 we found the additive term into *LOCATION.
5374 If X is an assignment of an invariant into DEST_REG, we set
5375 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5377 We also want to detect a BIV when it corresponds to a variable
5378 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5379 of the variable may be a PLUS that adds a SUBREG of that variable to
5380 an invariant and then sign- or zero-extends the result of the PLUS
5381 into the variable.
5383 Most GIVs in such cases will be in the promoted mode, since that is the
5384 probably the natural computation mode (and almost certainly the mode
5385 used for addresses) on the machine. So we view the pseudo-reg containing
5386 the variable as the BIV, as if it were simply incremented.
5388 Note that treating the entire pseudo as a BIV will result in making
5389 simple increments to any GIVs based on it. However, if the variable
5390 overflows in its declared mode but not its promoted mode, the result will
5391 be incorrect. This is acceptable if the variable is signed, since
5392 overflows in such cases are undefined, but not if it is unsigned, since
5393 those overflows are defined. So we only check for SIGN_EXTEND and
5394 not ZERO_EXTEND.
5396 If we cannot find a biv, we return 0. */
5398 static int
5403 rtx dest_reg;
5404 rtx p;
5408 {
5416 {
5422 {
5424 }
5429 {
5431 }
5432 else
5445 /* If this is a SUBREG for a promoted variable, check the inner
5446 value. */
5454 /* If this register is assigned in a previous insn, look at its
5455 source, but don't go outside the loop or past a label. */
5457 /* If this sets a register to itself, we would repeat any previous
5458 biv increment if we applied this strategy blindly. */
5464 {
5465 rtx dest;
5466 do
5467 {
5469 }
5498 }
5499 /* Fall through. */
5501 /* Can accept constant setting of biv only when inside inner most loop.
5502 Otherwise, a biv of an inner loop may be incorrectly recognized
5503 as a biv of the outer loop,
5504 causing code to be moved INTO the inner loop. */
5511 /* convert_modes aborts if we try to convert to or from CCmode, so just
5512 exclude that case. It is very unlikely that a condition code value
5513 would be a useful iterator anyways. */
5517 {
5518 /* Possible bug here? Perhaps we don't know the mode of X. */
5522 }
5523 else
5531 /* Similar, since this can be a sign extension. */
5536 ;
5550 location);
5555 }
5556 }
5557 \f
5558 /* A general induction variable (giv) is any quantity that is a linear
5559 function of a basic induction variable,
5560 i.e. giv = biv * mult_val + add_val.
5561 The coefficients can be any loop invariant quantity.
5562 A giv need not be computed directly from the biv;
5563 it can be computed by way of other givs. */
5565 /* Determine whether X computes a giv.
5566 If it does, return a nonzero value
5567 which is the benefit from eliminating the computation of X;
5568 set *SRC_REG to the register of the biv that it is computed from;
5569 set *ADD_VAL and *MULT_VAL to the coefficients,
5570 such that the value of X is biv * mult + add; */
5572 static int
5576 rtx x;
5584 {
5588 /* If this is an invariant, forget it, it isn't a giv. */
5599 {
5602 /* Since this is now an invariant and wasn't before, it must be a giv
5603 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5604 with. */
5611 /* This is equivalent to a BIV. */
5618 /* Either (plus (biv) (invar)) or
5619 (plus (mult (biv) (invar_1)) (invar_2)). */
5621 {
5624 }
5625 else
5626 {
5629 }
5634 /* ADD_VAL is zero. */
5642 }
5644 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5645 unless they are CONST_INT). */
5653 else
5656 /* Always return true if this is a giv so it will be detected as such,
5657 even if the benefit is zero or negative. This allows elimination
5658 of bivs that might otherwise not be eliminated. */
5660 }
5661 \f
5662 /* Given an expression, X, try to form it as a linear function of a biv.
5663 We will canonicalize it to be of the form
5664 (plus (mult (BIV) (invar_1))
5665 (invar_2))
5666 with possible degeneracies.
5668 The invariant expressions must each be of a form that can be used as a
5669 machine operand. We surround then with a USE rtx (a hack, but localized
5670 and certainly unambiguous!) if not a CONST_INT for simplicity in this
5671 routine; it is the caller's responsibility to strip them.
5673 If no such canonicalization is possible (i.e., two biv's are used or an
5674 expression that is neither invariant nor a biv or giv), this routine
5675 returns 0.
5677 For a non-zero return, the result will have a code of CONST_INT, USE,
5678 REG (for a BIV), PLUS, or MULT. No other codes will occur.
5680 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
5685 static rtx
5688 rtx x;
5691 {
5696 rtx tem;
5698 /* If this is not an integer mode, or if we cannot do arithmetic in this
5699 mode, this can't be a giv. */
5706 {
5713 /* Put constant last, CONST_INT last if both constant. */
5721 /* Handle addition of zero, then addition of an invariant. */
5726 {
5729 /* Adding two invariants must result in an invariant, so enclose
5730 addition operation inside a USE and return it. */
5740 else
5749 /* biv + invar or mult + invar. Return sum. */
5753 /* (a + invar_1) + invar_2. Associate. */
5754 return
5760 arg1)),
5765 }
5767 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
5768 MULT to reduce cases. */
5774 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
5775 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
5776 Recurse to associate the second PLUS. */
5781 return
5789 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
5805 /* Handle "a - b" as "a + b * (-1)". */
5811 constm1_rtx)),
5820 /* Put constant last, CONST_INT last if both constant. */
5825 /* If second argument is not now constant, not giv. */
5829 /* Handle multiply by 0 or 1. */
5837 {
5839 /* biv * invar. Done. */
5843 /* Product of two constants. */
5847 /* invar * invar is a giv, but attempt to simplify it somehow. */
5853 {
5854 /* (invar_0 * invar_1) * invar_2. Associate. */
5861 arg1)),
5863 }
5864 /* Porpagate the MULT expressions to the intermost nodes. */
5866 {
5867 /* (invar_0 + invar_1) * invar_2. Distribute. */
5873 arg1),
5877 arg1)),
5879 }
5883 /* (a * invar_1) * invar_2. Associate. */
5889 arg1)),
5893 /* (a + invar_1) * invar_2. Distribute. */
5898 arg1),
5901 arg1)),
5906 }
5909 /* Shift by constant is multiply by power of two. */
5913 return
5922 /* "-a" is "a * (-1)" */
5928 /* "~a" is "-a - 1". Silly, but easy. */
5932 const1_rtx),
5936 /* Already in proper form for invariant. */
5942 /* Conditionally recognize extensions of simple IVs. After we've
5943 computed loop traversal counts and verified the range of the
5944 source IV, we'll reevaluate this as a GIV. */
5946 {
5949 {
5952 }
5953 }
5957 /* If this is a new register, we can't deal with it. */
5961 /* Check for biv or giv. */
5963 {
5967 {
5970 /* Form expression from giv and add benefit. Ensure this giv
5971 can derive another and subtract any needed adjustment if so. */
5973 /* Increasing the benefit here is risky. The only case in which it
5974 is arguably correct is if this is the only use of V. In other
5975 cases, this will artificially inflate the benefit of the current
5976 giv, and lead to suboptimal code. Thus, it is disabled, since
5977 potentially not reducing an only marginally beneficial giv is
5978 less harmful than reducing many givs that are not really
5979 beneficial. */
5980 {
5984 }
5997 {
6000 }
6001 else
6002 {
6005 }
6007 }
6010 do_default:
6011 /* If it isn't an induction variable, and it is invariant, we
6012 may be able to simplify things further by looking through
6013 the bits we just moved outside the loop. */
6015 {
6021 {
6022 /* Ok, we found a match. Substitute and simplify. */
6024 /* If we match another movable, we must use that, as
6025 this one is going away. */
6030 /* If consec is non-zero, this is a member of a group of
6031 instructions that were moved together. We handle this
6032 case only to the point of seeking to the last insn and
6033 looking for a REG_EQUAL. Fail if we don't find one. */
6035 {
6038 do
6039 {
6041 }
6047 }
6048 else
6049 {
6053 }
6056 {
6057 /* What we are most interested in is pointer
6058 arithmetic on invariants -- only take
6059 patterns we may be able to do something with. */
6065 {
6067 benefit);
6070 }
6075 {
6080 }
6081 }
6083 }
6084 }
6086 }
6088 /* Fall through to general case. */
6090 /* If invariant, return as USE (unless CONST_INT).
6091 Otherwise, not giv. */
6096 {
6105 }
6106 else
6108 }
6109 }
6111 /* This routine folds invariants such that there is only ever one
6112 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6114 static rtx
6117 {
6123 {
6126 }
6129 {
6132 }
6133 else
6134 {
6137 }
6138 }
6140 static rtx
6144 {
6146 {
6150 else
6153 }
6156 else
6159 }
6160 \f
6161 /* Help detect a giv that is calculated by several consecutive insns;
6162 for example,
6163 giv = biv * M
6164 giv = giv + A
6165 The caller has already identified the first insn P as having a giv as dest;
6166 we check that all other insns that set the same register follow
6167 immediately after P, that they alter nothing else,
6168 and that the result of the last is still a giv.
6170 The value is 0 if the reg set in P is not really a giv.
6171 Otherwise, the value is the amount gained by eliminating
6172 all the consecutive insns that compute the value.
6174 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6175 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6177 The coefficients of the ultimate giv value are stored in
6178 *MULT_VAL and *ADD_VAL. */
6180 static int
6185 rtx p;
6186 rtx src_reg;
6187 rtx dest_reg;
6192 {
6198 rtx temp;
6199 rtx set;
6201 /* Indicate that this is a giv so that we can update the value produced in
6202 each insn of the multi-insn sequence.
6204 This induction structure will be used only by the call to
6205 general_induction_var below, so we can allocate it on our stack.
6206 If this is a giv, our caller will replace the induct var entry with
6207 a new induction structure. */
6228 {
6232 /* If libcall, skip to end of call sequence. */
6243 /* Giv created by equivalent expression. */
6249 {
6253 count--;
6257 }
6259 {
6260 /* Allow insns that set something other than this giv to a
6261 constant. Such insns are needed on machines which cannot
6262 include long constants and should not disqualify a giv. */
6271 }
6272 }
6277 }
6278 \f
6279 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6280 represented by G1. If no such expression can be found, or it is clear that
6281 it cannot possibly be a valid address, 0 is returned.
6283 To perform the computation, we note that
6284 G1 = x * v + a and
6285 G2 = y * v + b
6286 where `v' is the biv.
6288 So G2 = (y/b) * G1 + (b - a*y/x).
6290 Note that MULT = y/x.
6292 Update: A and B are now allowed to be additive expressions such that
6293 B contains all variables in A. That is, computing B-A will not require
6294 subtracting variables. */
6296 static rtx
6299 {
6300 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6305 /* If MULT is not 1, we cannot handle A with non-constants, since we
6306 would then be required to subtract multiples of the registers in A.
6307 This is theoretically possible, and may even apply to some Fortran
6308 constructs, but it is a lot of work and we do not attempt it here. */
6313 /* In general these structures are sorted top to bottom (down the PLUS
6314 chain), but not left to right across the PLUS. If B is a higher
6315 order giv than A, we can strip one level and recurse. If A is higher
6316 order, we'll eventually bail out, but won't know that until the end.
6317 If they are the same, we'll strip one level around this loop. */
6320 {
6332 /* We matched: remove one reg completely. */
6335 /* An alternate match. */
6338 /* An alternate match. */
6340 else
6341 {
6342 /* Indicates an extra register in B. Strip one level from B and
6343 recurse, hoping B was the higher order expression. */
6348 }
6349 }
6351 /* Here we are at the last level of A, go through the cases hoping to
6352 get rid of everything but a constant. */
6355 {
6368 }
6370 {
6372 }
6374 {
6375 return simplify_gen_binary (MINUS, GET_MODE (b) != VOIDmode ? GET_MODE (b) : GET_MODE (a), const0_rtx, a);
6376 }
6378 {
6383 else
6385 }
6390 }
6392 rtx
6395 {
6398 /* The value that G1 will be multiplied by must be a constant integer. Also,
6399 the only chance we have of getting a valid address is if b*c/a (see above
6400 for notation) is also an integer. */
6403 {
6408 }
6411 else
6412 {
6413 /* ??? Find out if the one is a multiple of the other? */
6415 }
6419 {
6420 /* Failed. If we've got a multiplication factor between G1 and G2,
6421 scale G1's addend and try again. */
6423 {
6427 {
6428 HOST_WIDE_INT m;
6432 }
6433 else
6434 {
6436 mult);
6437 }
6440 }
6441 }
6445 /* Form simplified final result. */
6450 else
6455 else
6456 {
6459 {
6463 }
6466 }
6467 }
6468 \f
6469 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6470 represented by G1. This indicates that G2 should be combined with G1 and
6471 that G2 can use (either directly or via an address expression) a register
6472 used to represent G1. */
6474 static rtx
6477 {
6480 /* With the introduction of ext dependant givs, we must care for modes.
6481 G2 must not use a wider mode than G1. */
6491 /* If these givs are identical, they can be combined. We use the results
6492 of express_from because the addends are not in a canonical form, so
6493 rtx_equal_p is a weaker test. */
6494 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6495 combination to be the other way round. */
6498 {
6500 }
6502 /* If G2 can be expressed as a function of G1 and that function is valid
6503 as an address and no more expensive than using a register for G2,
6504 the expression of G2 in terms of G1 can be used. */
6508 /* ??? Looses, especially with -fforce-addr, where *g2->location
6509 will always be a register, and so anything more complicated
6510 gets discarded. */
6511 #if 0
6512 #ifdef ADDRESS_COST
6514 #else
6516 #endif
6517 #endif
6518 )
6519 {
6521 }
6524 }
6525 \f
6526 /* Check each extension dependant giv in this class to see if its
6527 root biv is safe from wrapping in the interior mode, which would
6528 make the giv illegal. */
6530 static void
6534 {
6537 HOST_WIDE_INT start_val;
6543 /* Make sure the iteration data is available. We must have
6544 constants in order to be certain of no overflow. */
6545 /* ??? An unknown iteration count with an increment of +-1
6546 combined with friendly exit tests of against an invariant
6547 value is also ameanable to optimization. Not implemented. */
6553 /* Make sure the host can represent the arithmetic. */
6555 {
6557 HOST_WIDE_INT s_end_val;
6569 /* Check for host arithmatic overflow. */
6571 {
6573 HOST_WIDE_INT s_max;
6580 /* Check zero extension of biv ok. */
6582 /* Check for host arithmatic overflow. */
6583 && (neg_incr
6586 /* Check for target arithmetic overflow. */
6587 && (neg_incr
6590 {
6592 }
6594 /* Check sign extension of biv ok. */
6595 /* ??? While it is true that overflow with signed and pointer
6596 arithmetic is undefined, I fear too many programmers don't
6597 keep this fact in mind -- myself included on occasion.
6598 So leave alone with the signed overflow optimizations. */
6600 /* Check for host arithmatic overflow. */
6601 && (neg_incr
6604 /* Check for target arithmetic overflow. */
6605 && (neg_incr
6608 {
6610 }
6611 }
6612 }
6614 /* Invalidate givs that fail the tests. */
6617 {
6622 {
6631 /* We don't know whether this value is being used as either
6632 signed or unsigned, so to safely truncate we must satisfy
6633 both. The initial check here verifies the BIV itself;
6634 once that is successful we may check its range wrt the
6635 derived GIV. */
6637 {
6641 /* We know from the above that both endpoints are nonnegative,
6642 and that there is no wrapping. Verify that both endpoints
6643 are within the (signed) range of the outer mode. */
6646 }
6651 }
6654 {
6656 {
6660 }
6661 }
6662 else
6663 {
6665 {
6670 else
6671 {
6676 else
6678 }
6683 }
6686 }
6687 }
6688 }
6690 /* Generate a version of VALUE in a mode appropriate for initializing V. */
6692 rtx
6695 rtx value;
6696 {
6702 /* Recall that check_ext_dependant_givs verified that the known bounds
6703 of a biv did not overflow or wrap with respect to the extension for
6704 the giv. Therefore, constants need no additional adjustment. */
6708 /* Otherwise, we must adjust the value to compensate for the
6709 differing modes of the biv and the giv. */
6711 }
6712 \f
6713 struct combine_givs_stats
6714 {
6717 };
6719 static int
6723 {
6730 /* Stabilize the sort. */
6734 }
6736 /* Check all pairs of givs for iv_class BL and see if any can be combined with
6737 any other. If so, point SAME to the giv combined with and set NEW_REG to
6738 be an expression (in terms of the other giv's DEST_REG) equivalent to the
6739 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
6741 static void
6745 {
6746 /* Additional benefit to add for being combined multiple times. */
6754 /* Count givs, because bl->giv_count is incorrect here. */
6758 giv_count++;
6760 giv_array
6771 {
6773 rtx single_use;
6778 /* If a DEST_REG GIV is used only once, do not allow it to combine
6779 with anything, for in doing so we will gain nothing that cannot
6780 be had by simply letting the GIV with which we would have combined
6781 to be reduced on its own. The losage shows up in particular with
6782 DEST_ADDR targets on hosts with reg+reg addressing, though it can
6783 be seen elsewhere as well. */
6790 /* Add an additional weight for zero addends. */
6795 {
6796 rtx this_combine;
6801 {
6804 }
6805 }
6807 }
6809 /* Iterate, combining until we can't. */
6810 restart:
6814 {
6817 {
6823 }
6825 }
6828 {
6834 /* If it has already been combined, skip. */
6839 {
6842 /* If it has already been combined, skip. */
6844 {
6854 /* ??? The new final_[bg]iv_value code does a much better job
6855 of finding replaceable giv's, and hence this code may no
6856 longer be necessary. */
6860 /* To help optimize the next set of combinations, remove
6861 this giv from the benefits of other potential mates. */
6863 {
6867 }
6874 }
6875 }
6877 /* To help optimize the next set of combinations, remove
6878 this giv from the benefits of other potential mates. */
6880 {
6882 {
6886 }
6890 /* We've finished with this giv, and everything it touched.
6891 Restart the combination so that proper weights for the
6892 rest of the givs are properly taken into account. */
6893 /* ??? Ideally we would compact the arrays at this point, so
6894 as to not cover old ground. But sanely compacting
6895 can_combine is tricky. */
6897 }
6898 }
6900 /* Clean up. */
6903 }
6904 \f
6905 /* Generate sequence for REG = B * M + A. */
6907 static rtx
6913 {
6914 rtx seq;
6915 rtx result;
6918 /* Use unsigned arithmetic. */
6926 }
6929 /* Update registers created in insn sequence SEQ. */
6931 static void
6934 rtx seq;
6935 {
6936 /* Update register info for alias analysis. */
6939 {
6942 {
6946 }
6947 }
6948 else
6949 {
6953 }
6954 }
6957 /* EMIT code before BEFORE_BB/BEFORE_INSN to set REG = B * M + A. */
6959 void
6966 basic_block before_bb;
6967 rtx before_insn;
6968 {
6969 rtx seq;
6972 {
6975 }
6977 /* Use copy_rtx to prevent unexpected sharing of these rtx. */
6980 /* Increase the lifetime of any invariants moved further in code. */
6987 /* It is possible that the expansion created lots of new registers.
6988 Iterate over the sequence we just created and record them all. */
6990 }
6993 /* Emit insns in loop pre-header to set REG = B * M + A. */
6995 void
7002 {
7003 rtx seq;
7005 /* Use copy_rtx to prevent unexpected sharing of these rtx. */
7008 /* Increase the lifetime of any invariants moved further in code.
7009 ???? Is this really necessary? */
7016 /* It is possible that the expansion created lots of new registers.
7017 Iterate over the sequence we just created and record them all. */
7019 }
7022 /* Emit insns after loop to set REG = B * M + A. */
7024 void
7031 {
7032 rtx seq;
7034 /* Use copy_rtx to prevent unexpected sharing of these rtx. */
7039 /* It is possible that the expansion created lots of new registers.
7040 Iterate over the sequence we just created and record them all. */
7042 }
7046 /* Similar to gen_add_mult, but compute cost rather than generating
7047 sequence. */
7049 static int
7055 {
7065 {
7070 }
7073 }
7074 \f
7075 /* Test whether A * B can be computed without
7076 an actual multiply insn. Value is 1 if so. */
7078 static int
7080 rtx a;
7081 rtx b;
7082 {
7084 rtx tmp;
7087 /* If only one is constant, make it B. */
7091 /* If first constant, both constant, so don't need multiply. */
7095 /* If second not constant, neither is constant, so would need multiply. */
7099 /* One operand is constant, so might not need multiply insn. Generate the
7100 code for the multiply and see if a call or multiply, or long sequence
7101 of insns is generated. */
7109 {
7114 else
7116 {
7125 {
7128 }
7129 }
7130 }
7140 }
7141 \f
7142 /* Check to see if loop can be terminated by a "decrement and branch until
7143 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7144 Also try reversing an increment loop to a decrement loop
7145 to see if the optimization can be performed.
7146 Value is nonzero if optimization was performed. */
7148 /* This is useful even if the architecture doesn't have such an insn,
7149 because it might change a loops which increments from 0 to n to a loop
7150 which decrements from n to 0. A loop that decrements to zero is usually
7151 faster than one that increments from zero. */
7153 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7154 such as approx_final_value, biv_total_increment, loop_iterations, and
7155 final_[bg]iv_value. */
7157 static int
7161 {
7166 rtx reg;
7167 rtx jump_label;
7168 rtx final_value;
7169 rtx start_value;
7170 rtx new_add_val;
7171 rtx comparison;
7172 rtx before_comparison;
7173 rtx p;
7174 rtx jump;
7175 rtx first_compare;
7180 /* If last insn is a conditional branch, and the insn before tests a
7181 register value, try to optimize it. Otherwise, we can't do anything. */
7190 /* Try to compute whether the compare/branch at the loop end is one or
7191 two instructions. */
7197 else
7200 {
7201 /* If more than one condition is present to control the loop, then
7202 do not proceed, as this function does not know how to rewrite
7203 loop tests with more than one condition.
7205 Look backwards from the first insn in the last comparison
7206 sequence and see if we've got another comparison sequence. */
7208 rtx jump1;
7212 }
7214 /* Check all of the bivs to see if the compare uses one of them.
7215 Skip biv's set more than once because we can't guarantee that
7216 it will be zero on the last iteration. Also skip if the biv is
7217 used between its update and the test insn. */
7220 {
7225 first_compare))
7227 }
7232 /* Look for the case where the basic induction variable is always
7233 nonnegative, and equals zero on the last iteration.
7234 In this case, add a reg_note REG_NONNEG, which allows the
7235 m68k DBRA instruction to be used. */
7243 {
7244 /* Initial value must be greater than 0,
7245 init_val % -dec_value == 0 to ensure that it equals zero on
7246 the last iteration */
7252 {
7253 /* register always nonnegative, add REG_NOTE to branch */
7261 }
7263 /* If the decrement is 1 and the value was tested as >= 0 before
7264 the loop, then we can safely optimize. */
7266 {
7279 {
7287 }
7288 }
7289 }
7292 {
7293 /* Try to change inc to dec, so can apply above optimization. */
7294 /* Can do this if:
7295 all registers modified are induction variables or invariant,
7296 all memory references have non-overlapping addresses
7297 (obviously true if only one write)
7298 allow 2 insns for the compare/jump at the end of the loop. */
7299 /* Also, we must avoid any instructions which use both the reversed
7300 biv and another biv. Such instructions will fail if the loop is
7301 reversed. We meet this condition by requiring that either
7302 no_use_except_counting is true, or else that there is only
7303 one biv. */
7305 /* 1 if the iteration var is used only to count iterations. */
7307 /* 1 if the loop has no memory store, or it has a single memory store
7308 which is reversible. */
7312 {
7316 /* If there are no givs for this biv, and the only exit is the
7317 fall through at the end of the loop, then
7318 see if perhaps there are no uses except to count. */
7322 {
7327 /* An insn that sets the biv is okay. */
7328 ;
7332 {
7333 /* If either of these insns uses the biv and sets a pseudo
7334 that has more than one usage, then the biv has uses
7335 other than counting since it's used to derive a value
7336 that is used more than one time. */
7338 regs);
7340 {
7343 }
7344 }
7346 {
7349 }
7350 }
7352 /* A biv has uses besides counting if it is used to set another biv. */
7355 {
7358 }
7359 }
7362 /* No need to worry about MEMs. */
7363 ;
7365 {
7370 /* If the loop has a single store, and the destination address is
7371 invariant, then we can't reverse the loop, because this address
7372 might then have the wrong value at loop exit.
7373 This would work if the source was invariant also, however, in that
7374 case, the insn should have been moved out of the loop. */
7377 {
7380 /* If we could prove that each of the memory locations
7381 written to was different, then we could reverse the
7382 store -- but we don't presently have any way of
7383 knowing that. */
7386 /* If the store depends on a register that is set after the
7387 store, it depends on the initial value, and is thus not
7388 reversible. */
7390 {
7397 }
7398 }
7399 }
7400 else
7403 /* This code only acts for innermost loops. Also it simplifies
7404 the memory address check by only reversing loops with
7405 zero or one memory access.
7406 Two memory accesses could involve parts of the same array,
7407 and that can't be reversed.
7408 If the biv is used only for counting, than we don't need to worry
7409 about all these things. */
7414 && reversible_mem_store
7419 {
7420 rtx tem;
7422 /* Loop can be reversed. */
7426 /* Now check other conditions:
7428 The increment must be a constant, as must the initial value,
7429 and the comparison code must be LT.
7431 This test can probably be improved since +/- 1 in the constant
7432 can be obtained by changing LT to LE and vice versa; this is
7433 confusing. */
7436 /* for constants, LE gets turned into LT */
7440 {
7451 comparison_const_width
7453 else
7454 comparison_const_width
7458 comparison_sign_mask
7461 /* If the comparison value is not a loop invariant, then we
7462 can not reverse this loop.
7464 ??? If the insns which initialize the comparison value as
7465 a whole compute an invariant result, then we could move
7466 them out of the loop and proceed with loop reversal. */
7474 /* Normalize the initial value if it is an integer and
7475 has no other use except as a counter. This will allow
7476 a few more loops to be reversed. */
7480 {
7482 /* The code below requires comparison_val to be a multiple
7483 of add_val in order to do the loop reversal, so
7484 round up comparison_val to a multiple of add_val.
7485 Since comparison_value is constant, we know that the
7486 current comparison code is LT. */
7488 comparison_val
7490 /* We postpone overflow checks for COMPARISON_VAL here;
7491 even if there is an overflow, we might still be able to
7492 reverse the loop, if converting the loop exit test to
7493 NE is possible. */
7495 }
7497 /* First check if we can do a vanilla loop reversal. */
7499 /* If we have a decrement_and_branch_on_count,
7500 prefer the NE test, since this will allow that
7501 instruction to be generated. Note that we must
7502 use a vanilla loop reversal if the biv is used to
7503 calculate a giv or has a non-counting use. */
7504 #if ! defined (HAVE_decrement_and_branch_until_zero) \
7505 && defined (HAVE_decrement_and_branch_on_count)
7509 #endif
7511 /* Now do postponed overflow checks on COMPARISON_VAL. */
7514 {
7515 /* Register will always be nonnegative, with value
7516 0 on last iteration */
7520 }
7524 {
7527 }
7528 else
7534 /* If the initial value is not zero, or if the comparison
7535 value is not an exact multiple of the increment, then we
7536 can not reverse this loop. */
7539 {
7542 }
7543 else
7544 {
7547 }
7551 /* Reset these in case we normalized the initial value
7552 and comparison value above. */
7555 {
7557 final_value
7559 }
7562 /* Save some info needed to produce the new insns. */
7567 /* Set start_value; if this is not a CONST_INT, we need
7568 to generate a SUB.
7569 Initialize biv to start_value before loop start.
7570 The old initializing insn will be deleted as a
7571 dead store by flow.c. */
7574 {
7577 }
7579 {
7587 start_value
7593 }
7595 {
7598 initial_value);
7602 start_value
7605 }
7606 else
7607 /* We could handle the other cases too, but it'll be
7608 better to have a testcase first. */
7611 /* We may not have a single insn which can increment a reg, so
7612 create a sequence to hold all the insns from expand_inc. */
7621 /* Update biv info to reflect its new status. */
7626 /* Update loop info. */
7635 /* Inc LABEL_NUSES so that delete_insn will
7636 not delete the label. */
7639 /* Emit an insn after the end of the loop to set the biv's
7640 proper exit value if it is used anywhere outside the loop. */
7646 /* Delete compare/branch at end of loop. */
7651 /* Add new compare/branch insn at end of loop. */
7663 ;
7669 {
7671 {
7672 /* Increment of LABEL_NUSES done above. */
7673 /* Register is now always nonnegative,
7674 so add REG_NONNEG note to the branch. */
7677 }
7679 }
7681 /* No insn may reference both the reversed and another biv or it
7682 will fail (see comment near the top of the loop reversal
7683 code).
7684 Earlier on, we have verified that the biv has no use except
7685 counting, or it is the only biv in this function.
7686 However, the code that computes no_use_except_counting does
7687 not verify reg notes. It's possible to have an insn that
7688 references another biv, and has a REG_EQUAL note with an
7689 expression based on the reversed biv. To avoid this case,
7690 remove all REG_EQUAL notes based on the reversed biv
7691 here. */
7694 {
7697 /* If this is a set of a GIV based on the reversed biv, any
7698 REG_EQUAL notes should still be correct. */
7705 {
7710 else
7712 }
7713 }
7715 /* Mark that this biv has been reversed. Each giv which depends
7716 on this biv, and which is also live past the end of the loop
7717 will have to be fixed up. */
7722 {
7726 else
7728 }
7731 }
7732 }
7733 }
7736 }
7737 \f
7738 /* Verify whether the biv BL appears to be eliminable,
7739 based on the insns in the loop that refer to it.
7741 If ELIMINATE_P is non-zero, actually do the elimination.
7743 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
7744 determine whether invariant insns should be placed inside or at the
7745 start of the loop. */
7747 static int
7753 {
7756 rtx p;
7758 /* Scan all insns in the loop, stopping if we find one that uses the
7759 biv in a way that we cannot eliminate. */
7762 {
7767 /* If this is a libcall that sets a giv, skip ahead to its end. */
7769 {
7773 {
7778 {
7785 }
7786 }
7787 }
7792 {
7798 }
7799 }
7802 {
7807 }
7810 }
7811 \f
7812 /* INSN and REFERENCE are instructions in the same insn chain.
7813 Return non-zero if INSN is first. */
7815 int
7818 {
7822 {
7823 /* Start with test for not first so that INSN == REFERENCE yields not
7824 first. */
7830 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
7831 previous insn, hence the <= comparison below does not work if
7832 P is a note. */
7843 }
7844 }
7846 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
7847 the offset that we have to take into account due to auto-increment /
7848 div derivation is zero. */
7849 static int
7852 rtx insn;
7853 {
7854 /* If the giv V had the auto-inc address optimization applied
7855 to it, and INSN occurs between the giv insn and the biv
7856 insn, then we'd have to adjust the value used here.
7857 This is rare, so we don't bother to make this possible. */
7866 }
7868 /* If BL appears in X (part of the pattern of INSN), see if we can
7869 eliminate its use. If so, return 1. If not, return 0.
7871 If BIV does not appear in X, return 1.
7873 If ELIMINATE_P is non-zero, actually do the elimination.
7874 WHERE_INSN/WHERE_BB indicate where extra insns should be added.
7875 Depending on how many items have been moved out of the loop, it
7876 will either be before INSN (when WHERE_INSN is non-zero) or at the
7877 start of the loop (when WHERE_INSN is zero). */
7879 static int
7885 basic_block where_bb;
7886 rtx where_insn;
7887 {
7893 #ifdef HAVE_cc0
7895 #endif
7901 {
7903 /* If we haven't already been able to do something with this BIV,
7904 we can't eliminate it. */
7910 /* If this sets the BIV, it is not a problem. */
7914 /* If this is an insn that defines a giv, it is also ok because
7915 it will go away when the giv is reduced. */
7920 #ifdef HAVE_cc0
7922 {
7923 /* Can replace with any giv that was reduced and
7924 that has (MULT_VAL != 0) and (ADD_VAL == 0).
7925 Require a constant for MULT_VAL, so we know it's nonzero.
7926 ??? We disable this optimization to avoid potential
7927 overflows. */
7935 {
7942 /* If the giv has the opposite direction of change,
7943 then reverse the comparison. */
7947 else
7950 /* We can probably test that giv's reduced reg. */
7953 }
7955 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
7956 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
7957 Require a constant for MULT_VAL, so we know it's nonzero.
7958 ??? Do this only if ADD_VAL is a pointer to avoid a potential
7959 overflow problem. */
7971 {
7978 /* If the giv has the opposite direction of change,
7979 then reverse the comparison. */
7983 else
7987 /* Replace biv with the giv's reduced register. */
7992 /* Insn doesn't support that constant or invariant. Copy it
7993 into a register (it will be a loop invariant.) */
8000 /* Substitute the new register for its invariant value in
8001 the compare expression. */
8005 }
8006 }
8007 #endif
8014 /* See if either argument is the biv. */
8019 else
8023 {
8024 /* First try to replace with any giv that has constant positive
8025 mult_val and constant add_val. We might be able to support
8026 negative mult_val, but it seems complex to do it in general. */
8038 {
8045 /* Replace biv with the giv's reduced reg. */
8048 /* If all constants are actually constant integers and
8049 the derived constant can be directly placed in the COMPARE,
8050 do so. */
8054 {
8059 }
8060 else
8061 {
8062 /* Otherwise, load it into a register. */
8068 }
8071 }
8073 /* Look for giv with positive constant mult_val and nonconst add_val.
8074 Insert insns to calculate new compare value.
8075 ??? Turn this off due to possible overflow. */
8083 {
8084 rtx tem;
8094 /* Replace biv with giv's reduced register. */
8098 /* Compute value to compare against. */
8102 /* Use it in this insn. */
8106 }
8107 }
8109 {
8111 {
8112 /* Look for giv with constant positive mult_val and nonconst
8113 add_val. Insert insns to compute new compare value.
8114 ??? Turn this off due to possible overflow. */
8121 {
8122 rtx tem;
8132 /* Replace biv with giv's reduced register. */
8136 /* Compute value to compare against. */
8143 }
8144 }
8146 /* This code has problems. Basically, you can't know when
8147 seeing if we will eliminate BL, whether a particular giv
8148 of ARG will be reduced. If it isn't going to be reduced,
8149 we can't eliminate BL. We can try forcing it to be reduced,
8150 but that can generate poor code.
8152 The problem is that the benefit of reducing TV, below should
8153 be increased if BL can actually be eliminated, but this means
8154 we might have to do a topological sort of the order in which
8155 we try to process biv. It doesn't seem worthwhile to do
8156 this sort of thing now. */
8158 #if 0
8159 /* Otherwise the reg compared with had better be a biv. */
8164 /* Look for a pair of givs, one for each biv,
8165 with identical coefficients. */
8167 {
8179 {
8186 /* Replace biv with its giv's reduced reg. */
8188 /* Replace other operand with the other giv's
8189 reduced reg. */
8192 }
8193 }
8194 #endif
8195 }
8197 /* If we get here, the biv can't be eliminated. */
8201 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8202 biv is used in it, since it will be replaced. */
8210 }
8212 /* See if any subexpression fails elimination. */
8215 {
8217 {
8230 }
8231 }
8234 }
8235 \f
8236 /* Return nonzero if the last use of REG
8237 is in an insn following INSN in the same basic block. */
8239 static int
8241 rtx reg;
8242 rtx insn;
8243 {
8244 rtx n;
8248 {
8251 }
8253 }
8254 \f
8255 /* Called via `note_stores' to record the initial value of a biv. Here we
8256 just record the location of the set and process it later. */
8258 static void
8260 rtx dest;
8261 rtx set;
8263 {
8274 /* If this is the first set found, record it. */
8276 {
8279 }
8280 }
8281 \f
8282 /* If any of the registers in X are "old" and currently have a last use earlier
8283 than INSN, update them to have a last use of INSN. Their actual last use
8284 will be the previous insn but it will not have a valid uid_luid so we can't
8285 use it. X must be a source expression only. */
8287 static void
8289 rtx x;
8290 rtx insn;
8291 {
8292 /* Check for the case where INSN does not have a valid luid. In this case,
8293 there is no need to modify the regno_last_uid, as this can only happen
8294 when code is inserted after the loop_end to set a pseudo's final value,
8295 and hence this insn will never be the last use of x.
8296 ???? This comment is not correct. See for example loop_givs_reduce.
8297 This may insert an insn before another new insn. */
8301 {
8303 }
8304 else
8305 {
8309 {
8315 }
8316 }
8317 }
8318 \f
8319 /* Given an insn INSN and condition COND, return the condition in a
8320 canonical form to simplify testing by callers. Specifically:
8322 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8323 (2) Both operands will be machine operands; (cc0) will have been replaced.
8324 (3) If an operand is a constant, it will be the second operand.
8325 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8326 for GE, GEU, and LEU.
8328 If the condition cannot be understood, or is an inequality floating-point
8329 comparison which needs to be reversed, 0 will be returned.
8331 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
8333 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8334 insn used in locating the condition was found. If a replacement test
8335 of the condition is desired, it should be placed in front of that
8336 insn and we will be sure that the inputs are still valid.
8338 If WANT_REG is non-zero, we wish the condition to be relative to that
8339 register, if possible. Therefore, do not canonicalize the condition
8340 further. */
8342 rtx
8344 rtx insn;
8345 rtx cond;
8348 rtx want_reg;
8349 {
8352 rtx set;
8353 rtx tem;
8371 /* If we are comparing a register with zero, see if the register is set
8372 in the previous insn to a COMPARE or a comparison operation. Perform
8373 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8374 in cse.c */
8379 {
8380 /* Set non-zero when we find something of interest. */
8383 #ifdef HAVE_cc0
8384 /* If comparison with cc0, import actual comparison from compare
8385 insn. */
8387 {
8398 }
8399 #endif
8401 /* If this is a COMPARE, pick up the two things being compared. */
8403 {
8407 }
8411 /* Go back to the previous insn. Stop if it is not an INSN. We also
8412 stop if it isn't a single set or if it has a REG_INC note because
8413 we don't want to bother dealing with it. */
8427 /* If this is setting OP0, get what it sets it to if it looks
8428 relevant. */
8430 {
8433 /* ??? We may not combine comparisons done in a CCmode with
8434 comparisons not done in a CCmode. This is to aid targets
8435 like Alpha that have an IEEE compliant EQ instruction, and
8436 a non-IEEE compliant BEQ instruction. The use of CCmode is
8437 actually artificial, simply to prevent the combination, but
8438 should not affect other platforms.
8440 However, we must allow VOIDmode comparisons to match either
8441 CCmode or non-CCmode comparison, because some ports have
8442 modeless comparisons inside branch patterns.
8444 ??? This mode check should perhaps look more like the mode check
8445 in simplify_comparison in combine. */
8453 && (STORE_FLAG_VALUE
8456 #ifdef FLOAT_STORE_FLAG_VALUE
8459 && (REAL_VALUE_NEGATIVE
8461 #endif
8462 ))
8473 && (STORE_FLAG_VALUE
8476 #ifdef FLOAT_STORE_FLAG_VALUE
8479 && (REAL_VALUE_NEGATIVE
8481 #endif
8482 ))
8488 {
8491 }
8492 else
8494 }
8497 /* If this sets OP0, but not directly, we have to give up. */
8501 {
8505 {
8510 }
8515 }
8516 }
8518 /* If constant is first, put it last. */
8522 /* If OP0 is the result of a comparison, we weren't able to find what
8523 was really being compared, so fail. */
8527 /* Canonicalize any ordered comparison with integers involving equality
8528 if we can do computations in the relevant mode and we do not
8529 overflow. */
8534 {
8537 unsigned HOST_WIDE_INT max_val
8541 {
8547 /* When cross-compiling, const_val might be sign-extended from
8548 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
8568 }
8569 }
8571 #ifdef HAVE_cc0
8572 /* Never return CC0; return zero instead. */
8575 #endif
8578 }
8580 /* Given a jump insn JUMP, return the condition that will cause it to branch
8581 to its JUMP_LABEL. If the condition cannot be understood, or is an
8582 inequality floating-point comparison which needs to be reversed, 0 will
8583 be returned.
8585 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8586 insn used in locating the condition was found. If a replacement test
8587 of the condition is desired, it should be placed in front of that
8588 insn and we will be sure that the inputs are still valid. */
8590 rtx
8592 rtx jump;
8594 {
8595 rtx cond;
8597 rtx set;
8599 /* If this is not a standard conditional jump, we can't parse it. */
8607 /* If this branches to JUMP_LABEL when the condition is false, reverse
8608 the condition. */
8609 reverse
8614 }
8616 /* Similar to above routine, except that we also put an invariant last
8617 unless both operands are invariants. */
8619 rtx
8622 rtx x;
8623 {
8633 }
8635 /* Scan the function and determine whether it has indirect (computed) jumps.
8637 This is taken mostly from flow.c; similar code exists elsewhere
8638 in the compiler. It may be useful to put this into rtlanal.c. */
8639 static int
8641 rtx start;
8642 {
8643 rtx insn;
8650 }
8652 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
8653 documentation for LOOP_MEMS for the definition of `appropriate'.
8654 This function is called from prescan_loop via for_each_rtx. */
8656 static int
8660 {
8669 {
8674 /* We're not interested in MEMs that are only clobbered. */
8678 /* We're not interested in the MEM associated with a
8679 CONST_DOUBLE, so there's no need to traverse into this. */
8683 /* We're not interested in any MEMs that only appear in notes. */
8687 /* This is not a MEM. */
8689 }
8691 /* See if we've already seen this MEM. */
8694 {
8696 /* The modes of the two memory accesses are different. If
8697 this happens, something tricky is going on, and we just
8698 don't optimize accesses to this MEM. */
8702 }
8704 /* Resize the array, if necessary. */
8706 {
8709 else
8715 }
8717 /* Actually insert the MEM. */
8719 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
8720 because we can't put it in a register. We still store it in the
8721 table, though, so that if we see the same address later, but in a
8722 non-BLK mode, we'll not think we can optimize it at that point. */
8728 }
8731 /* Allocate REGS->ARRAY or reallocate it if it is too small.
8733 Increment REGS->ARRAY[I].SET_IN_LOOP at the index I of each
8734 register that is modified by an insn between FROM and TO. If the
8735 value of an element of REGS->array[I].SET_IN_LOOP becomes 127 or
8736 more, stop incrementing it, to avoid overflow.
8738 Store in REGS->ARRAY[I].SINGLE_USAGE the single insn in which
8739 register I is used, if it is only used once. Otherwise, it is set
8740 to 0 (for no uses) or const0_rtx for more than one use. This
8741 parameter may be zero, in which case this processing is not done.
8743 Set REGS->ARRAY[I].MAY_NOT_OPTIMIZE nonzero if we should not
8744 optimize register I. */
8746 static void
8750 {
8753 /* last_set[n] is nonzero iff reg n has been set in the current
8754 basic block. In that case, it is the insn that last set reg n. */
8756 rtx insn;
8762 /* Grow the regs array if not allocated or too small. */
8764 {
8770 /* Zero the new elements. */
8773 }
8775 /* Clear previously scanned fields but do not clear n_times_set. */
8777 {
8781 }
8785 /* Scan the loop, recording register usage. */
8788 {
8790 {
8791 /* Record registers that have exactly one use. */
8794 /* Include uses in REG_EQUAL notes. */
8802 {
8806 last_set);
8807 }
8808 }
8812 }
8815 {
8818 }
8820 #ifdef AVOID_CCMODE_COPIES
8821 /* Don't try to move insns which set CC registers if we should not
8822 create CCmode register copies. */
8826 #endif
8828 /* Set regs->array[I].n_times_set for the new registers. */
8833 }
8835 /* Returns the number of real INSNs in the LOOP. */
8837 static int
8840 {
8842 rtx insn;
8850 }
8852 /* Move MEMs into registers for the duration of the loop. */
8854 static void
8857 {
8864 rtx end_label;
8865 /* Nonzero if the next instruction may never be executed. */
8872 /* We cannot use next_label here because it skips over normal insns. */
8877 /* Check to see if it's possible that some instructions in the loop are
8878 never executed. Also check if there is a goto out of the loop other
8879 than right after the end of the loop. */
8883 {
8887 /* If we enter the loop in the middle, and scan
8888 around to the beginning, don't set maybe_never
8889 for that. This must be an unconditional jump,
8890 otherwise the code at the top of the loop might
8891 never be executed. Unconditional jumps are
8892 followed a by barrier then loop end. */
8897 {
8898 /* If this is a jump outside of the loop but not right
8899 after the end of the loop, we would have to emit new fixup
8900 sequences for each such label. */
8902 JUMP_INSN is executed, we must be conservative. */
8911 /* Something complicated. */
8913 else
8914 /* If there are any more instructions in the loop, they
8915 might not be reached. */
8917 }
8920 }
8922 /* Find start of the extended basic block that enters the loop. */
8926 ;
8931 /* Build table of mems that get set to constant values before the
8932 loop. */
8936 /* Actually move the MEMs. */
8938 {
8939 regset_head load_copies;
8940 regset_head store_copies;
8942 rtx reg;
8944 rtx mem_list_entry;
8948 /* There's no telling whether or not MEM is modified. */
8951 /* Go through the MEMs written to in the loop to see if this
8952 one is aliased by one of them. */
8955 {
8960 {
8961 /* MEM is indeed aliased by this store. */
8964 }
8966 }
8972 /* If this MEM is written to, we must be sure that there
8973 are no reads from another MEM that aliases this one. */
8975 {
8979 {
8983 VOIDmode,
8985 rtx_varies_p))
8986 {
8987 /* It's not safe to hoist loop_info->mems[i] out of
8988 the loop because writes to it might not be
8989 seen by reads from loop_info->mems[j]. */
8992 }
8993 }
8994 }
8997 /* We can't access the MEM outside the loop; it might
8998 cause a trap that wouldn't have happened otherwise. */
9002 /* We thought we were going to lift this MEM out of the
9003 loop, but later discovered that we could not. */
9009 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9010 order to keep scan_loop from moving stores to this MEM
9011 out of the loop just because this REG is neither a
9012 user-variable nor used in the loop test. */
9017 /* Now, replace all references to the MEM with the
9018 corresponding pseudos. */
9023 {
9025 {
9026 rtx set;
9030 /* See if this copies the mem into a register that isn't
9031 modified afterwards. We'll try to do copy propagation
9032 a little further on. */
9034 /* @@@ This test is _way_ too conservative. */
9035 && ! maybe_never
9043 /* See if this copies the mem from a register that isn't
9044 modified afterwards. We'll try to remove the
9045 redundant copy later on by doing a little register
9046 renaming and copy propagation. This will help
9047 to untangle things for the BIV detection code. */
9049 && ! maybe_never
9057 /* Replace the memory reference with the shadow register. */
9060 }
9065 }
9068 /* We couldn't replace all occurrences of the MEM. */
9070 else
9071 {
9072 /* Load the memory immediately before LOOP->START, which is
9073 the NOTE_LOOP_BEG. */
9075 rtx set;
9081 {
9085 {
9089 /* Extending hard register lifetimes causes crash
9090 on SRC targets. Doing so on non-SRC is
9091 probably also not good idea, since we most
9092 probably have pseudoregister equivalence as
9093 well. */
9096 }
9097 /* Use the constant equivalence if that is cheap enough. */
9103 {
9106 }
9108 /* If best_equiv is nonzero, we know that MEM is set to a
9109 constant or register before the loop. We will use this
9110 knowledge to initialize the shadow register with that
9111 constant or reg rather than by loading from MEM. */
9114 }
9119 {
9122 {
9125 }
9126 }
9134 {
9136 {
9139 }
9141 /* Store the memory immediately after END, which is
9142 the NOTE_LOOP_END. */
9145 }
9148 {
9153 }
9155 /* Attempt a bit of copy propagation. This helps untangle the
9156 data flow, and enables {basic,general}_induction_var to find
9157 more bivs/givs. */
9158 EXECUTE_IF_SET_IN_REG_SET
9160 {
9162 });
9165 EXECUTE_IF_SET_IN_REG_SET
9167 {
9169 });
9171 }
9172 }
9175 {
9176 /* Now, we need to replace all references to the previous exit
9177 label with the new one. */
9178 rtx_pair rr;
9183 {
9186 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9187 field. This is not handled by for_each_rtx because it doesn't
9188 handle unprinted ('0') fields. We need to update JUMP_LABEL
9189 because the immediately following unroll pass will use it.
9190 replace_label would not work anyways, because that only handles
9191 LABEL_REFs. */
9194 }
9195 }
9198 }
9200 /* For communication between note_reg_stored and its caller. */
9201 struct note_reg_stored_arg
9202 {
9204 rtx reg;
9205 };
9207 /* Called via note_stores, record in SET_SEEN whether X, which is written,
9208 is equal to ARG. */
9209 static void
9213 {
9217 }
9219 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
9220 There must be exactly one insn that sets this pseudo; it will be
9221 deleted if all replacements succeed and we can prove that the register
9222 is not used after the loop. */
9224 static void
9227 rtx replacement;
9229 {
9230 /* This is the reg that we are copying from. */
9233 rtx insn;
9234 /* These help keep track of whether we replaced all uses of the reg. */
9241 {
9242 rtx set;
9244 /* Only substitute within one extended basic block from the initializing
9245 insn. */
9252 /* Is this the initializing insn? */
9257 {
9264 }
9266 /* Only substitute after seeing the initializing insn. */
9268 {
9275 /* Stop replacing when REPLACEMENT is modified. */
9280 {
9283 /* It is possible that we've turned previously valid REG_EQUAL to
9284 invalid, as we change the REGNO to REPLACEMENT and unlike REGNO,
9285 REPLACEMENT is modified, we get different meaning. */
9289 }
9290 }
9291 }
9295 {
9299 {
9300 rtx first;
9301 rtx retval_note;
9303 /* Assume we're just deleting INIT_INSN. */
9305 /* Look for REG_RETVAL note. If we're deleting the end of
9306 the libcall sequence, the whole sequence can go. */
9308 /* If we found a REG_RETVAL note, find the first instruction
9309 in the sequence. */
9313 /* Delete the instructions. */
9315 }
9318 }
9319 }
9321 /* Replace all the instructions from FIRST up to and including LAST
9322 with NOTE_INSN_DELETED notes. */
9324 static void
9326 rtx first;
9327 rtx last;
9328 {
9330 {
9337 /* If this was the LAST instructions we're supposed to delete,
9338 we're done. */
9343 }
9344 }
9346 /* Try to replace occurrences of pseudo REGNO with REPLACEMENT within
9347 loop LOOP if the order of the sets of these registers can be
9348 swapped. There must be exactly one insn within the loop that sets
9349 this pseudo followed immediately by a move insn that sets
9350 REPLACEMENT with REGNO. */
9351 static void
9354 rtx replacement;
9356 {
9357 rtx insn;
9366 {
9367 /* Search for the insn that copies REGNO to NEW_REGNO? */
9375 }
9378 {
9379 rtx prev_insn;
9380 rtx prev_set;
9382 /* Some DEF-USE info would come in handy here to make this
9383 function more general. For now, just check the previous insn
9384 which is the most likely candidate for setting REGNO. */
9392 {
9393 /* We have:
9394 (set (reg regno) (expr))
9395 (set (reg new_regno) (reg regno))
9397 so try converting this to:
9398 (set (reg new_regno) (expr))
9399 (set (reg regno) (reg new_regno))
9401 The former construct is often generated when a global
9402 variable used for an induction variable is shadowed by a
9403 register (NEW_REGNO). The latter construct improves the
9404 chances of GIV replacement and BIV elimination. */
9414 {
9421 /* Update first use of REGNO. */
9425 /* Now perform copy propagation to hopefully
9426 remove all uses of REGNO within the loop. */
9428 }
9429 }
9430 }
9431 }
9433 /* Replace MEM with its associated pseudo register. This function is
9434 called from load_mems via for_each_rtx. DATA is actually a pointer
9435 to a structure describing the instruction currently being scanned
9436 and the MEM we are currently replacing. */
9438 static int
9442 {
9450 {
9455 /* We're not interested in the MEM associated with a
9456 CONST_DOUBLE, so there's no need to traverse into one. */
9460 /* This is not a MEM. */
9462 }
9465 /* This is not the MEM we are currently replacing. */
9468 /* Actually replace the MEM. */
9472 }
9474 static void
9476 rtx insn;
9477 rtx mem;
9478 rtx reg;
9479 {
9480 loop_replace_args args;
9487 }
9489 /* Replace one register with another. Called through for_each_rtx; PX points
9490 to the rtx being scanned. DATA is actually a pointer to
9491 a structure of arguments. */
9493 static int
9497 {
9508 }
9510 static void
9512 rtx insn;
9513 rtx reg;
9514 rtx replacement;
9515 {
9516 loop_replace_args args;
9523 }
9525 /* Replace occurrences of the old exit label for the loop with the new
9526 one. DATA is an rtx_pair containing the old and new labels,
9527 respectively. */
9529 static int
9533 {
9552 }
9553 \f
9554 /* Emit insn for PATTERN after WHERE_INSN in basic block WHERE_BB
9555 (ignored in the interim). */
9557 static rtx
9560 basic_block where_bb ATTRIBUTE_UNUSED;
9561 rtx where_insn;
9562 rtx pattern;
9563 {
9565 }
9568 /* If WHERE_INSN is non-zero emit insn for PATTERN before WHERE_INSN
9569 in basic block WHERE_BB (ignored in the interim) within the loop
9570 otherwise hoist PATTERN into the loop pre-header. */
9572 rtx
9575 basic_block where_bb ATTRIBUTE_UNUSED;
9576 rtx where_insn;
9577 rtx pattern;
9578 {
9582 }
9585 /* Emit call insn for PATTERN before WHERE_INSN in basic block
9586 WHERE_BB (ignored in the interim) within the loop. */
9588 static rtx
9591 basic_block where_bb ATTRIBUTE_UNUSED;
9592 rtx where_insn;
9593 rtx pattern;
9594 {
9596 }
9599 /* Hoist insn for PATTERN into the loop pre-header. */
9601 rtx
9604 rtx pattern;
9605 {
9607 }
9610 /* Hoist call insn for PATTERN into the loop pre-header. */
9612 static rtx
9615 rtx pattern;
9616 {
9618 }
9621 /* Sink insn for PATTERN after the loop end. */
9623 rtx
9626 rtx pattern;
9627 {
9629 }
9632 /* If the loop has multiple exits, emit insn for PATTERN before the
9633 loop to ensure that it will always be executed no matter how the
9634 loop exits. Otherwise, emit the insn for PATTERN after the loop,
9635 since this is slightly more efficient. */
9637 static rtx
9640 rtx pattern;
9641 {
9644 else
9646 }
9647 \f
9648 static void
9653 {
9661 iv_num++;
9666 {
9669 }
9670 }
9673 static void
9678 {
9680 rtx incr;
9691 {
9694 }
9696 {
9699 }
9703 {
9707 }
9710 {
9714 }
9716 /* List the increments. */
9718 {
9722 }
9724 /* List the givs. */
9726 {
9731 else
9734 }
9735 }
9738 static void
9743 {
9754 {
9758 }
9761 }
9764 static void
9769 {
9776 else
9792 {
9794 {
9806 }
9807 }
9818 {
9822 }
9825 }
9828 void
9831 {
9833 }
9836 void
9839 {
9841 }
9844 void
9847 {
9849 }
9852 void
9855 {
9857 }
9860 #define LOOP_BLOCK_NUM_1(INSN) \
9861 ((INSN) ? (BLOCK_FOR_INSN (INSN) ? BLOCK_NUM (INSN) : - 1) : -1)
9863 /* The notes do not have an assigned block, so look at the next insn. */
9864 #define LOOP_BLOCK_NUM(INSN) \
9865 ((INSN) ? (GET_CODE (INSN) == NOTE \
9866 ? LOOP_BLOCK_NUM_1 (next_nonnote_insn (INSN)) \
9867 : LOOP_BLOCK_NUM_1 (INSN)) \
9868 : -1)
9870 #define LOOP_INSN_UID(INSN) ((INSN) ? INSN_UID (INSN) : -1)
9872 static void
9877 {
9878 rtx label;
9883 /* Print diagnostics to compare our concept of a loop with
9884 what the loop notes say. */
9899 {
9919 {
9922 {
9925 }
9926 }
9929 /* This can happen when a marked loop appears as two nested loops,
9930 say from while (a || b) {}. The inner loop won't match
9931 the loop markers but the outer one will. */
9934 }
9935 }
9937 /* Call this function from the debugger to dump LOOP. */
9939 void
9942 {
9944 }
9946 /* Call this function from the debugger to dump LOOPS. */
9948 void
9951 {
9953 }