| blob | 066ef931111dc3a40eb6c70170da09ebf0f4a0be |
1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997,
3 1998, 1999, 2000 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
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. */
23 /* This is the loop optimization pass of the compiler.
24 It finds invariant computations within loops and moves them
25 to the beginning of the loop. Then it identifies basic and
26 general induction variables. Strength reduction is applied to the general
27 induction variables, and induction variable elimination is applied to
28 the basic induction variables.
30 It also finds cases where
31 a register is set within the loop by zero-extending a narrower value
32 and changes these to zero the entire register once before the loop
33 and merely copy the low part within the loop.
35 Most of the complexity is in heuristics to decide when it is worth
36 while to do these things. */
58 /* Vector mapping INSN_UIDs to luids.
59 The luids are like uids but increase monotonically always.
60 We use them to see whether a jump comes from outside a given loop. */
64 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
65 number the insn is contained in. */
69 /* 1 + largest uid of any insn. */
73 /* 1 + luid of last insn. */
77 /* Number of loops detected in current function. Used as index to the
78 next few tables. */
82 /* Indexed by register number, contains the number of times the reg
83 is set during the loop being scanned.
84 During code motion, a negative value indicates a reg that has been
85 made a candidate; in particular -2 means that it is an candidate that
86 we know is equal to a constant and -1 means that it is an candidate
87 not known equal to a constant.
88 After code motion, regs moved have 0 (which is accurate now)
89 while the failed candidates have the original number of times set.
91 Therefore, at all times, == 0 indicates an invariant register;
92 < 0 a conditionally invariant one. */
96 /* Original value of set_in_loop; same except that this value
97 is not set negative for a reg whose sets have been made candidates
98 and not set to 0 for a reg that is moved. */
102 /* Index by register number, 1 indicates that the register
103 cannot be moved or strength reduced. */
107 /* Contains the insn in which a register was used if it was used
108 exactly once; contains const0_rtx if it was used more than once. */
112 /* Nonzero means reg N has already been moved out of one loop.
113 This reduces the desire to move it out of another. */
117 /* List of MEMs that are stored in this loop. */
121 /* The insn where the first of these was found. */
130 /* Array of MEMs that are used (read or written) in this loop, but
131 cannot be aliased by anything in this loop, except perhaps
132 themselves. In other words, if loop_mems[i] is altered during the
133 loop, it is altered by an expression that is rtx_equal_p to it. */
137 /* The index of the next available slot in LOOP_MEMS. */
141 /* The number of elements allocated in LOOP_MEMs. */
145 /* Nonzero if we don't know what MEMs were changed in the current
146 loop. This happens if the loop contains a call (in which case
147 `loop_info->has_call' will also be set) or if we store into more
148 than NUM_STORES MEMs. */
152 /* The above doesn't count any readonly memory locations that are stored.
153 This does. */
157 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
160 /* Count of memory write instructions discovered in the loop. */
163 /* Bound on pseudo register number before loop optimization.
164 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
167 /* The value to pass to the next call of reg_scan_update. */
170 /* This obstack is used in product_cheap_p to allocate its rtl. It
171 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
172 If we used the same obstack that it did, we would be deallocating
173 that array. */
177 /* This is where the pointer to the obstack being used for RTL is stored. */
181 #define obstack_chunk_alloc xmalloc
182 #define obstack_chunk_free free
183 \f
184 /* During the analysis of a loop, a chain of `struct movable's
185 is made to record all the movable insns found.
186 Then the entire chain can be scanned to decide which to move. */
188 struct movable
189 {
194 of any registers used within the LIBCALL. */
196 that must be moved with this one. */
199 may be adjusted when matching movables
200 that load the same value are found. */
202 including other movables that force this
203 or match this one. */
207 /* If PARTIAL is 1, GLOBAL means something different:
208 that the reg is live outside the range from where it is set
209 to the following label. */
213 In particular, moving it does not make it
214 invariant. */
216 load SRC, rather than copying INSN. */
218 first insn of a consecutive sets group. */
221 that we should avoid changing when clearing
222 the rest of the reg. */
226 };
232 /* Forward declarations. */
250 #if 0
252 #endif
315 rtx r;
320 rtx r1;
321 rtx r2;
324 /* Nonzero iff INSN is between START and END, inclusive. */
325 #define INSN_IN_RANGE_P(INSN, START, END) \
326 (INSN_UID (INSN) < max_uid_for_loop \
327 && INSN_LUID (INSN) >= INSN_LUID (START) \
328 && INSN_LUID (INSN) <= INSN_LUID (END))
330 #ifdef HAVE_decrement_and_branch_on_count
331 /* Test whether BCT applicable and safe. */
334 /* Auxiliary function that inserts the BCT pattern into the loop. */
338 /* Indirect_jump_in_function is computed once per function. */
346 \f
347 /* Relative gain of eliminating various kinds of operations. */
349 #if 0
352 #endif
354 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
355 copy the value of the strength reduced giv to its original register. */
358 /* Cost of using a register, to normalize the benefits of a giv. */
362 void
364 {
372 /* We multiply by 2 to reconcile the difference in scale between
373 these two ways of computing costs. Otherwise the cost of a copy
374 will be far less than the cost of an add. */
378 /* Free the objects we just allocated. */
381 /* Initialize the obstack used for rtl in product_cheap_p. */
383 }
384 \f
385 /* Compute the mapping from uids to luids.
386 LUIDs are numbers assigned to insns, like uids,
387 except that luids increase monotonically through the code.
388 Start at insn START and stop just before END. Assign LUIDs
389 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
390 static int
394 {
396 rtx insn;
399 {
402 /* Don't assign luids to line-number NOTEs, so that the distance in
403 luids between two insns is not affected by -g. */
407 else
408 /* Give a line number note the same luid as preceding insn. */
410 }
412 }
413 \f
414 /* Entry point of this file. Perform loop optimization
415 on the current function. F is the first insn of the function
416 and DUMPFILE is a stream for output of a trace of actions taken
417 (or 0 if none should be output). */
419 void
421 /* f is the first instruction of a chain of insns for one function */
422 rtx f;
425 {
441 /* Count the number of loops. */
445 {
448 max_loop_num++;
449 }
451 /* Don't waste time if no loops. */
459 /* Get size to use for tables indexed by uids.
460 Leave some space for labels allocated by find_and_verify_loops. */
467 /* Allocate storage for array of loops. */
471 /* Find and process each loop.
472 First, find them, and record them in order of their beginnings. */
475 /* Allocate and initialize auxiliary loop information. */
480 /* Now find all register lifetimes. This must be done after
481 find_and_verify_loops, because it might reorder the insns in the
482 function. */
485 /* This must occur after reg_scan so that registers created by gcse
486 will have entries in the register tables.
488 We could have added a call to reg_scan after gcse_main in toplev.c,
489 but moving this call to init_alias_analysis is more efficient. */
492 /* See if we went too far. Note that get_max_uid already returns
493 one more that the maximum uid of all insn. */
496 /* Now reset it to the actual size we need. See above. */
499 /* find_and_verify_loops has already called compute_luids, but it
500 might have rearranged code afterwards, so we need to recompute
501 the luids now. */
504 /* Don't leave gaps in uid_luid for insns that have been
505 deleted. It is possible that the first or last insn
506 using some register has been deleted by cross-jumping.
507 Make sure that uid_luid for that former insn's uid
508 points to the general area where that insn used to be. */
510 {
514 }
519 /* Determine if the function has indirect jump. On some systems
520 this prevents low overhead loop instructions from being used. */
523 /* Now scan the loops, last ones first, since this means inner ones are done
524 before outer ones. */
526 {
531 }
533 /* If there were lexical blocks inside the loop, they have been
534 replicated. We will now have more than one NOTE_INSN_BLOCK_BEG
535 and NOTE_INSN_BLOCK_END for each such block. We must duplicate
536 the BLOCKs as well. */
542 /* Clean up. */
548 }
549 \f
550 /* Returns the next insn, in execution order, after INSN. START and
551 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
552 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
553 insn-stream; it is used with loops that are entered near the
554 bottom. */
556 static rtx
559 rtx insn;
560 {
564 {
566 /* Go to the top of the loop, and continue there. */
568 else
569 /* We're done. */
571 }
574 /* We're done. */
578 }
580 /* Optimize one loop described by LOOP. */
582 /* ??? Could also move memory writes out of loops if the destination address
583 is invariant, the source is invariant, the memory write is not volatile,
584 and if we can prove that no read inside the loop can read this address
585 before the write occurs. If there is a read of this address after the
586 write, then we can also mark the memory read as invariant. */
588 static void
592 {
596 /* Additional information about the current loop being processed
597 that is used to compute the number of loop iterations for loop
598 unrolling and doloop optimization. */
600 rtx p;
601 /* 1 if we are scanning insns that could be executed zero times. */
603 /* 1 if we are scanning insns that might never be executed
604 due to a subroutine call which might exit before they are reached. */
606 /* Jump insn that enters the loop, or 0 if control drops in. */
608 /* Number of insns in the loop. */
613 /* The SET from an insn, if it is the only SET in the insn. */
615 /* Chain describing insns movable in current loop. */
617 /* Last element in `movables' -- so we can add elements at the end. */
619 /* Ratio of extra register life span we can justify
620 for saving an instruction. More if loop doesn't call subroutines
621 since in that case saving an insn makes more difference
622 and more registers are available. */
624 /* Nonzero if we are scanning instructions in a sub-loop. */
630 /* Determine whether this loop starts with a jump down to a test at
631 the end. This will occur for a small number of loops with a test
632 that is too complex to duplicate in front of the loop.
634 We search for the first insn or label in the loop, skipping NOTEs.
635 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
636 (because we might have a loop executed only once that contains a
637 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
638 (in case we have a degenerate loop).
640 Note that if we mistakenly think that a loop is entered at the top
641 when, in fact, it is entered at the exit test, the only effect will be
642 slightly poorer optimization. Making the opposite error can generate
643 incorrect code. Since very few loops now start with a jump to the
644 exit test, the code here to detect that case is very conservative. */
647 p != loop_end
653 ;
657 /* Set up variables describing this loop. */
661 /* If loop has a jump before the first label,
662 the true entry is the target of that jump.
663 Start scan from there.
664 But record in LOOP->TOP the place where the end-test jumps
665 back to so we can scan that after the end of the loop. */
667 {
670 /* Loop entry must be unconditional jump (and not a RETURN) */
673 /* Check to see whether the jump actually
674 jumps out of the loop (meaning it's no loop).
675 This case can happen for things like
676 do {..} while (0). If this label was generated previously
677 by loop, we can't tell anything about it and have to reject
678 the loop. */
680 {
683 }
684 }
686 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
687 as required by loop_reg_used_before_p. So skip such loops. (This
688 test may never be true, but it's best to play it safe.)
690 Also, skip loops where we do not start scanning at a label. This
691 test also rejects loops starting with a JUMP_INSN that failed the
692 test above. */
696 {
701 }
703 /* Count number of times each reg is set during this loop.
704 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
705 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
707 /* Allocate extra space for REGS that might be created by
708 load_mems. We allocate a little extra slop as well, in the hopes
709 that even after the moving of movables creates some new registers
710 we won't have to reallocate these arrays. However, we do grow
711 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
722 {
725 }
727 #ifdef AVOID_CCMODE_COPIES
728 /* Don't try to move insns which set CC registers if we should not
729 create CCmode register copies. */
733 #endif
739 {
745 }
747 /* Scan through the loop finding insns that are safe to move.
748 Set set_in_loop negative for the reg being set, so that
749 this reg will be considered invariant for subsequent insns.
750 We consider whether subsequent insns use the reg
751 in deciding whether it is worth actually moving.
753 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
754 and therefore it is possible that the insns we are scanning
755 would never be executed. At such times, we must make sure
756 that it is safe to execute the insn once instead of zero times.
757 When MAYBE_NEVER is 0, all insns will be executed at least once
758 so that is not a problem. */
763 {
775 {
782 /* Figure out what to use as a source of this insn. If a REG_EQUIV
783 note is given or if a REG_EQUAL note with a constant operand is
784 specified, use it as the source and mark that we should move
785 this insn by calling emit_move_insn rather that duplicating the
786 insn.
788 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
789 is present. */
793 else
794 {
799 {
801 /* A libcall block can use regs that don't appear in
802 the equivalent expression. To move the libcall,
803 we must move those regs too. */
805 }
806 }
808 /* Don't try to optimize a register that was made
809 by loop-optimization for an inner loop.
810 We don't know its life-span, so we can't compute the benefit. */
812 ;
814 than the one where it is set (meaning that
815 something after this point in the loop might
816 depend on its value before the set). */
818 /* And the set is not guaranteed to be executed one
819 the loop starts, or the value before the set is
820 needed before the set occurs...
822 ??? Note we have quadratic behaviour here, mitigated
823 by the fact that the previous test will often fail for
824 large loops. Rather than re-scanning the entire loop
825 each time for register usage, we should build tables
826 of the register usage and use them here instead. */
827 && (maybe_never
829 /* It is unsafe to move the set.
831 This code used to consider it OK to move a set of a variable
832 which was not created by the user and not used in an exit test.
833 That behavior is incorrect and was removed. */
834 ;
840 || (tem1
841 = consec_sets_invariant_p
844 p)))
845 /* If the insn can cause a trap (such as divide by zero),
846 can't move it unless it's guaranteed to be executed
847 once loop is entered. Even a function call might
848 prevent the trap insn from being reached
849 (since it might exit!) */
852 {
856 /* A potential lossage is where we have a case where two insns
857 can be combined as long as they are both in the loop, but
858 we move one of them outside the loop. For large loops,
859 this can lose. The most common case of this is the address
860 of a function being called.
862 Therefore, if this register is marked as being used exactly
863 once if we are in a loop with calls (a "large loop"), see if
864 we can replace the usage of this register with the source
865 of this SET. If we can, delete this insn.
867 Don't do this if P has a REG_RETVAL note or if we have
868 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
879 && (! SMALL_REGISTER_CLASSES
882 /* This test is not redundant; SET_SRC (set) might be
883 a call-clobbered register and the life of REGNO
884 might span a call. */
886 VARRAY_RTX
890 VARRAY_RTX
892 {
893 /* Replace any usage in a REG_EQUAL note. Must copy the
894 new source, so that we don't get rtx sharing between the
895 SET_SOURCE and REG_NOTES of insn p. */
906 }
925 /* Set M->cond if either loop_invariant_p
926 or consec_sets_invariant_p returned 2
927 (only conditionally invariant). */
939 /* Add M to the end of the chain MOVABLES. */
942 else
947 {
948 /* It is possible for the first instruction to have a
949 REG_EQUAL note but a non-invariant SET_SRC, so we must
950 remember the status of the first instruction in case
951 the last instruction doesn't have a REG_EQUAL note. */
954 /* Skip this insn, not checking REG_LIBCALL notes. */
956 /* Skip the consecutive insns, if there are any. */
958 /* Back up to the last insn of the consecutive group. */
961 /* We must now reset m->move_insn, m->is_equiv, and possibly
962 m->set_src to correspond to the effects of all the
963 insns. */
967 else
968 {
972 else
975 }
977 }
978 }
979 /* If this register is always set within a STRICT_LOW_PART
980 or set to zero, then its high bytes are constant.
981 So clear them outside the loop and within the loop
982 just load the low bytes.
983 We must check that the machine has an instruction to do so.
984 Also, if the value loaded into the register
985 depends on the same register, this cannot be done. */
995 {
998 {
1012 /* If the insn may not be executed on some cycles,
1013 we can't clear the whole reg; clear just high part.
1014 Not even if the reg is used only within this loop.
1015 Consider this:
1016 while (1)
1017 while (s != t) {
1018 if (foo ()) x = *s;
1019 use (x);
1020 }
1021 Clearing x before the inner loop could clobber a value
1022 being saved from the last time around the outer loop.
1023 However, if the reg is not used outside this loop
1024 and all uses of the register are in the same
1025 basic block as the store, there is no problem.
1027 If this insn was made by loop, we don't know its
1028 INSN_LUID and hence must make a conservative
1029 assumption. */
1035 || (labels_in_range_p
1039 else
1048 /* Add M to the end of the chain MOVABLES. */
1051 else
1054 }
1055 }
1056 }
1057 /* Past a call insn, we get to insns which might not be executed
1058 because the call might exit. This matters for insns that trap.
1059 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1060 so they don't count. */
1063 /* Past a label or a jump, we get to insns for which we
1064 can't count on whether or how many times they will be
1065 executed during each iteration. Therefore, we can
1066 only move out sets of trivial variables
1067 (those not used after the loop). */
1068 /* Similar code appears twice in strength_reduce. */
1070 /* If we enter the loop in the middle, and scan around to the
1071 beginning, don't set maybe_never for that. This must be an
1072 unconditional jump, otherwise the code at the top of the
1073 loop might never be executed. Unconditional jumps are
1074 followed a by barrier then loop end. */
1080 {
1081 /* At the virtual top of a converted loop, insns are again known to
1082 be executed: logically, the loop begins here even though the exit
1083 code has been duplicated. */
1087 loop_depth++;
1089 loop_depth--;
1090 }
1091 }
1093 /* If one movable subsumes another, ignore that other. */
1097 /* For each movable insn, see if the reg that it loads
1098 leads when it dies right into another conditionally movable insn.
1099 If so, record that the second insn "forces" the first one,
1100 since the second can be moved only if the first is. */
1104 /* See if there are multiple movable insns that load the same value.
1105 If there are, make all but the first point at the first one
1106 through the `match' field, and add the priorities of them
1107 all together as the priority of the first. */
1111 /* Now consider each movable insn to decide whether it is worth moving.
1112 Store 0 in set_in_loop for each reg that is moved.
1114 Generally this increases code size, so do not move moveables when
1115 optimizing for code size. */
1120 /* Now candidates that still are negative are those not moved.
1121 Change set_in_loop to indicate that those are not actually invariant. */
1126 /* Now that we've moved some things out of the loop, we might be able to
1127 hoist even more memory references. */
1134 ;
1141 {
1147 }
1153 }
1154 \f
1155 /* Add elements to *OUTPUT to record all the pseudo-regs
1156 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1158 void
1162 {
1170 {
1188 }
1192 {
1196 {
1205 }
1206 }
1207 }
1208 \f
1209 /* Check what regs are referred to in the libcall block ending with INSN,
1210 aside from those mentioned in the equivalent value.
1211 If there are none, return 0.
1212 If there are one or more, return an EXPR_LIST containing all of them. */
1214 rtx
1217 {
1222 /* First, find all the regs used in the libcall block
1223 that are not mentioned as inputs to the result. */
1226 {
1231 }
1234 }
1235 \f
1236 /* Return 1 if all uses of REG
1237 are between INSN and the end of the basic block. */
1239 static int
1242 {
1244 rtx p;
1249 /* Search this basic block for the already recorded last use of the reg. */
1251 {
1253 {
1259 /* Ordinary insn: if this is the last use, we win. */
1265 /* Jump insn: if this is the last use, we win. */
1268 /* Otherwise, it's the end of the basic block, so we lose. */
1273 /* It's the end of the basic block, so we lose. */
1278 }
1279 }
1281 /* The "last use" that was recorded can't be found after the first
1282 use. This can happen when the last use was deleted while
1283 processing an inner loop, this inner loop was then completely
1284 unrolled, and the outer loop is always exited after the inner loop,
1285 so that everything after the first use becomes a single basic block. */
1287 }
1288 \f
1289 /* Compute the benefit of eliminating the insns in the block whose
1290 last insn is LAST. This may be a group of insns used to compute a
1291 value directly or can contain a library call. */
1293 static int
1295 rtx last;
1296 {
1297 rtx insn;
1302 {
1305 routine. */
1309 benefit++;
1310 }
1313 }
1314 \f
1315 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1317 static rtx
1319 rtx insn;
1321 {
1323 {
1324 rtx temp;
1326 /* If first insn of libcall sequence, skip to end. */
1327 /* Do this at start of loop, since INSN is guaranteed to
1328 be an insn here. */
1335 }
1338 }
1340 /* Ignore any movable whose insn falls within a libcall
1341 which is part of another movable.
1342 We make use of the fact that the movable for the libcall value
1343 was made later and so appears later on the chain. */
1345 static void
1348 {
1352 {
1353 /* Is this a movable for the value of a libcall? */
1356 {
1357 rtx insn;
1358 /* Check for earlier movables inside that range,
1359 and mark them invalid. We cannot use LUIDs here because
1360 insns created by loop.c for prior loops don't have LUIDs.
1361 Rather than reject all such insns from movables, we just
1362 explicitly check each insn in the libcall (since invariant
1363 libcalls aren't that common). */
1368 }
1369 }
1370 }
1372 /* For each movable insn, see if the reg that it loads
1373 leads when it dies right into another conditionally movable insn.
1374 If so, record that the second insn "forces" the first one,
1375 since the second can be moved only if the first is. */
1377 static void
1380 {
1383 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1385 {
1388 /* ??? Could this be a bug? What if CSE caused the
1389 register of M1 to be used after this insn?
1390 Since CSE does not update regno_last_uid,
1391 this insn M->insn might not be where it dies.
1392 But very likely this doesn't matter; what matters is
1393 that M's reg is computed from M1's reg. */
1398 /* If m->consec, m->set_src isn't valid. */
1402 /* Increase the priority of the moving the first insn
1403 since it permits the second to be moved as well. */
1405 {
1409 }
1410 }
1411 }
1412 \f
1413 /* Find invariant expressions that are equal and can be combined into
1414 one register. */
1416 static void
1420 {
1425 /* Regs that are set more than once are not allowed to match
1426 or be matched. I'm no longer sure why not. */
1427 /* Perhaps testing m->consec_sets would be more appropriate here? */
1431 {
1438 /* We want later insns to match the first one. Don't make the first
1439 one match any later ones. So start this loop at m->next. */
1442 /* A reg used outside the loop mustn't be eliminated. */
1444 /* A reg used for zero-extending mustn't be eliminated. */
1447 ||
1448 (
1449 /* Can combine regs with different modes loaded from the
1450 same constant only if the modes are the same or
1451 if both are integer modes with M wider or the same
1452 width as M1. The check for integer is redundant, but
1453 safe, since the only case of differing destination
1454 modes with equal sources is when both sources are
1455 VOIDmode, i.e., CONST_INT. */
1461 /* See if the source of M1 says it matches M. */
1465 movables))))
1468 {
1474 }
1475 }
1477 /* Now combine the regs used for zero-extension.
1478 This can be done for those not marked `global'
1479 provided their lives don't overlap. */
1483 {
1486 /* Combine all the registers for extension from mode MODE.
1487 Don't combine any that are used outside this loop. */
1491 {
1497 {
1498 /* First one: don't check for overlap, just record it. */
1501 }
1503 /* Make sure they extend to the same mode.
1504 (Almost always true.) */
1508 /* We already have one: check for overlap with those
1509 already combined together. */
1516 /* No overlap: we can combine this with the others. */
1522 overlap: ;
1523 }
1524 }
1526 /* Clean up. */
1528 }
1529 \f
1530 /* Return 1 if regs X and Y will become the same if moved. */
1532 static int
1536 {
1553 }
1555 /* Return 1 if X and Y are identical-looking rtx's.
1556 This is the Lisp function EQUAL for rtx arguments.
1558 If two registers are matching movables or a movable register and an
1559 equivalent constant, consider them equal. */
1561 static int
1565 {
1579 /* If we have a register and a constant, they may sometimes be
1580 equal. */
1583 {
1588 }
1591 {
1596 }
1598 /* Otherwise, rtx's of different codes cannot be equal. */
1602 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1603 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1608 /* These three types of rtx's can be compared nonrecursively. */
1617 /* Compare the elements. If any pair of corresponding elements
1618 fail to match, return 0 for the whole things. */
1622 {
1624 {
1636 /* Two vectors must have the same length. */
1640 /* And the corresponding elements must match. */
1657 /* These are just backpointers, so they don't matter. */
1663 /* It is believed that rtx's at this level will never
1664 contain anything but integers and other rtx's,
1665 except for within LABEL_REFs and SYMBOL_REFs. */
1668 }
1669 }
1671 }
1672 \f
1673 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1674 insns in INSNS which use the reference. */
1676 static void
1678 rtx x;
1679 rtx insns;
1680 {
1684 rtx insn;
1687 {
1688 /* This code used to ignore labels that referred to dispatch tables to
1689 avoid flow generating (slighly) worse code.
1691 We no longer ignore such label references (see LABEL_REF handling in
1692 mark_jump_label for additional information). */
1697 }
1701 {
1707 }
1708 }
1709 \f
1710 /* Scan MOVABLES, and move the insns that deserve to be moved.
1711 If two matching movables are combined, replace one reg with the
1712 other throughout. */
1714 static void
1721 {
1727 /* Map of pseudo-register replacements to handle combining
1728 when we move several insns that load the same value
1729 into different pseudo-registers. */
1736 {
1737 /* Describe this movable insn. */
1740 {
1761 }
1763 /* Count movables. Value used in heuristics in strength_reduce. */
1764 num_movables++;
1766 /* Ignore the insn if it's already done (it matched something else).
1767 Otherwise, see if it is now safe to move. */
1779 {
1784 /* We have an insn that is safe to move.
1785 Compute its desirability. */
1796 /* An insn MUST be moved if we already moved something else
1797 which is safe only if this one is moved too: that is,
1798 if already_moved[REGNO] is nonzero. */
1800 /* An insn is desirable to move if the new lifetime of the
1801 register is no more than THRESHOLD times the old lifetime.
1802 If it's not desirable, it means the loop is so big
1803 that moving won't speed things up much,
1804 and it is liable to make register usage worse. */
1806 /* It is also desirable to move if it can be moved at no
1807 extra cost because something else was already moved. */
1810 || flag_move_all_movables
1815 {
1820 /* Now move the insns that set the reg. */
1823 {
1826 /* Find the end of this chain of matching regs.
1827 Thus, we load each reg in the chain from that one reg.
1828 And that reg is loaded with 0 directly,
1829 since it has ->match == 0. */
1835 /* Mark the moved, invariant reg as being allowed to
1836 share a hard reg with the other matching invariant. */
1840 regs_may_share
1843 regs_may_share));
1851 }
1852 /* If we are to re-generate the item being moved with a
1853 new move insn, first delete what we have and then emit
1854 the move insn before the loop. */
1856 {
1860 {
1861 /* If this is the first insn of a library call sequence,
1862 skip to the end. */
1867 /* If this is the last insn of a libcall sequence, then
1868 delete every insn in the sequence except the last.
1869 The last insn is handled in the normal manner. */
1872 {
1876 }
1881 /* simplify_giv_expr expects that it can walk the insns
1882 at m->insn forwards and see this old sequence we are
1883 tossing here. delete_insn does preserve the next
1884 pointers, but when we skip over a NOTE we must fix
1885 it up. Otherwise that code walks into the non-deleted
1886 insn stream. */
1889 }
1907 /* The more regs we move, the less we like moving them. */
1909 }
1910 else
1911 {
1913 {
1916 /* If first insn of libcall sequence, skip to end. */
1917 /* Do this at start of loop, since p is guaranteed to
1918 be an insn here. */
1923 /* If last insn of libcall sequence, move all
1924 insns except the last before the loop. The last
1925 insn is handled in the normal manner. */
1928 {
1936 {
1937 rtx body;
1938 rtx n;
1939 rtx next;
1946 /* Find the next insn after TEMP,
1947 not counting USE or NOTE insns. */
1955 /* If that is the call, this may be the insn
1956 that loads the function address.
1958 Extract the function address from the insn
1959 that loads it into a register.
1960 If this insn was cse'd, we get incorrect code.
1962 So emit a new move insn that copies the
1963 function address into the register that the
1964 call insn will use. flow.c will delete any
1965 redundant stores that we have created. */
1970 NULL_RTX)))
1971 {
1977 }
1978 /* We have the call insn.
1979 If it uses the register we suspect it might,
1980 load it with the correct address directly. */
1985 fn_address),
1986 fn_address_insn);
1989 {
1991 /* Because the USAGE information potentially
1992 contains objects other than hard registers
1993 we need to copy it. */
1997 }
1998 else
2006 }
2009 }
2011 {
2012 /* P sets REG to zero; but we should clear only
2013 the bits that are not covered by the mode
2014 m->savemode. */
2016 rtx sequence;
2017 rtx tem;
2020 tem = expand_binop
2033 }
2035 {
2037 /* Because the USAGE information potentially
2038 contains objects other than hard registers
2039 we need to copy it. */
2043 }
2045 {
2046 /* The SET_SRC might not be invariant, so we must
2047 use the REG_EQUAL note. */
2061 }
2062 else
2066 {
2069 /* If there is a REG_EQUAL note present whose value
2070 is not loop invariant, then delete it, since it
2071 may cause problems with later optimization passes.
2072 It is possible for cse to create such notes
2073 like this as a result of record_jump_cond. */
2078 }
2087 /* If library call, now fix the REG_NOTES that contain
2088 insn pointers, namely REG_LIBCALL on FIRST
2089 and REG_RETVAL on I1. */
2091 {
2095 }
2101 /* simplify_giv_expr expects that it can walk the insns
2102 at m->insn forwards and see this old sequence we are
2103 tossing here. delete_insn does preserve the next
2104 pointers, but when we skip over a NOTE we must fix
2105 it up. Otherwise that code walks into the non-deleted
2106 insn stream. */
2109 }
2111 /* The more regs we move, the less we like moving them. */
2113 }
2115 /* Any other movable that loads the same register
2116 MUST be moved. */
2119 /* This reg has been moved out of one loop. */
2122 /* The reg set here is now invariant. */
2128 /* Change the length-of-life info for the register
2129 to say it lives at least the full length of this loop.
2130 This will help guide optimizations in outer loops. */
2133 /* This is the old insn before all the moved insns.
2134 We can't use the moved insn because it is out of range
2135 in uid_luid. Only the old insns have luids. */
2140 /* Combine with this moved insn any other matching movables. */
2145 {
2146 rtx temp;
2148 /* Schedule the reg loaded by M1
2149 for replacement so that shares the reg of M.
2150 If the modes differ (only possible in restricted
2151 circumstances, make a SUBREG.
2153 Note this assumes that the target dependent files
2154 treat REG and SUBREG equally, including within
2155 GO_IF_LEGITIMATE_ADDRESS and in all the
2156 predicates since we never verify that replacing the
2157 original register with a SUBREG results in a
2158 recognizable insn. */
2161 else
2166 /* Get rid of the matching insn
2167 and prevent further processing of it. */
2170 /* if library call, delete all insn except last, which
2171 is deleted below */
2173 NULL_RTX)))
2174 {
2178 }
2181 /* Any other movable that loads the same register
2182 MUST be moved. */
2185 /* The reg merged here is now invariant,
2186 if the reg it matches is invariant. */
2189 }
2190 }
2193 }
2199 }
2204 /* Go through all the instructions in the loop, making
2205 all the register substitutions scheduled in REG_MAP. */
2209 {
2213 }
2215 /* Clean up. */
2218 }
2219 \f
2220 #if 0
2221 /* Scan X and replace the address of any MEM in it with ADDR.
2222 REG is the address that MEM should have before the replacement. */
2224 static void
2227 {
2236 {
2248 /* Short cut for very common case. */
2253 /* Short cut for very common case. */
2258 /* If this MEM uses a reg other than the one we expected,
2259 something is wrong. */
2267 }
2271 {
2275 {
2279 }
2280 }
2281 }
2282 #endif
2283 \f
2284 /* Return the number of memory refs to addresses that vary
2285 in the rtx X. */
2287 static int
2290 rtx x;
2291 {
2302 {
2319 }
2324 {
2328 {
2332 }
2333 }
2335 }
2337 \f
2338 #if 0
2339 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2340 Replace it with an instruction to load just the low bytes
2341 if the machine supports such an instruction,
2342 and insert above LOOP_START an instruction to clear the register. */
2344 static void
2347 {
2351 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2352 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2354 new
2355 = gen_rtx_SET
2357 gen_rtx_STRICT_LOW_PART
2366 {
2369 /* Clear destination register before the loop. */
2372 loop_start);
2374 /* Inside the loop, just load the low part. */
2376 }
2377 }
2378 #endif
2379 \f
2380 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2381 `has_call', `has_volatile', and `has_tablejump' within LOOP.
2382 Set the global variables `unknown_address_altered',
2383 `unknown_constant_address_altered', and `num_mem_sets'. Also, fill
2384 in the array `loop_mems' and the list `loop_store_mems'. */
2386 static void
2389 {
2391 rtx insn;
2395 /* The label after END. Jumping here is just like falling off the
2396 end of the loop. We use next_nonnote_insn instead of next_label
2397 as a hedge against the (pathological) case where some actual insn
2398 might end up between the two. */
2419 {
2421 {
2423 {
2425 /* Count number of loops contained in this one. */
2427 }
2429 {
2432 {
2435 }
2436 }
2438 {
2441 }
2443 {
2444 /* If there is a NOTE_INSN_LOOP_VTOP, then this is a for
2445 or while style loop, with a loop exit test at the
2446 start. Thus, we can assume that the loop condition
2447 was true when the loop was entered. */
2450 }
2451 }
2453 {
2457 }
2459 {
2479 {
2481 {
2484 }
2485 else
2486 {
2488 }
2492 {
2494 {
2495 /* Something tricky. */
2498 }
2501 {
2502 /* A jump outside the current loop. */
2505 }
2506 }
2511 }
2512 }
2515 }
2517 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2519 ! unknown_address_altered
2520 /* An exception thrown by a called function might land us
2521 anywhere. */
2523 /* We don't want loads for MEMs moved to a location before the
2524 one at which their stack memory becomes allocated. (Note
2525 that this is not a problem for malloc, etc., since those
2526 require actual function calls. */
2527 && ! current_function_calls_alloca
2528 /* There are ways to leave the loop other than falling off the
2529 end. */
2534 }
2535 \f
2536 /* LOOP->CONT_DOMINATOR is now the last label between the loop start
2537 and the continue note that is a the destination of a (cond)jump after
2538 the continue note. If there is any (cond)jump between the loop start
2539 and what we have so far as LOOP->CONT_DOMINATOR that has a
2540 target between LOOP->DOMINATOR and the continue note, move
2541 LOOP->CONT_DOMINATOR forward to that label; if a jump's
2542 destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
2544 static void
2547 {
2548 rtx insn;
2551 /* This can happen for an empty loop, e.g. in
2552 gcc.c-torture/compile/920410-2.c */
2555 {
2558 }
2561 {
2564 {
2568 /* If it is not a jump we can easily understand or for
2569 which we do not have jump target information in the JUMP_LABEL
2570 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2571 LOOP->CONT_DOMINATOR. */
2574 {
2577 }
2581 && (label_luid
2584 }
2585 }
2586 }
2588 /* Scan the function looking for loops. Record the start and end of each loop.
2589 Also mark as invalid loops any loops that contain a setjmp or are branched
2590 to from outside the loop. */
2592 static void
2594 rtx f;
2596 {
2597 rtx insn;
2598 rtx label;
2608 /* If there are jumps to undefined labels,
2609 treat them as jumps out of any/all loops.
2610 This also avoids writing past end of tables when there are no loops. */
2613 /* Find boundaries of loops, mark which loops are contained within
2614 loops, and invalidate loops that have setjmp. */
2619 {
2622 {
2626 num_loops++;
2633 /* In this case, we must invalidate our current loop and any
2634 enclosing loop. */
2636 {
2642 }
2659 }
2660 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2661 and NOTE_INSN_LOOP_END notes, update loop->dominator. */
2665 {
2672 do
2673 {
2674 /* First see if we care about this loop. */
2676 {
2677 /* If the jump destination is not known, invalidate
2678 loop->const_dominator. */
2681 else
2682 /* Check if the destination is between loop start and
2683 cont. */
2688 /* And if there is no later destination already
2689 recorded. */
2694 }
2696 }
2698 }
2700 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2701 enclosing loop, but this doesn't matter. */
2703 }
2705 /* Any loop containing a label used in an initializer must be invalidated,
2706 because it can be jumped into from anywhere. */
2709 {
2713 }
2715 /* Any loop containing a label used for an exception handler must be
2716 invalidated, because it can be jumped into from anywhere. */
2719 {
2723 }
2725 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2726 loop that it is not contained within, that loop is marked invalid.
2727 If any INSN or CALL_INSN uses a label's address, then the loop containing
2728 that label is marked invalid, because it could be jumped into from
2729 anywhere.
2731 Also look for blocks of code ending in an unconditional branch that
2732 exits the loop. If such a block is surrounded by a conditional
2733 branch around the block, move the block elsewhere (see below) and
2734 invert the jump to point to the code block. This may eliminate a
2735 label in our loop and will simplify processing by both us and a
2736 possible second cse pass. */
2740 {
2744 {
2747 {
2751 }
2752 }
2759 /* See if this is an unconditional branch outside the loop. */
2767 {
2768 rtx p;
2774 /* Go backwards until we reach the start of the loop, a label,
2775 or a JUMP_INSN. */
2782 ;
2784 /* Check for the case where we have a jump to an inner nested
2785 loop, and do not perform the optimization in that case. */
2788 {
2791 {
2796 }
2797 }
2799 /* Make sure that the target of P is within the current loop. */
2805 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2806 we have a block of code to try to move.
2808 We look backward and then forward from the target of INSN
2809 to find a BARRIER at the same loop depth as the target.
2810 If we find such a BARRIER, we make a new label for the start
2811 of the block, invert the jump in P and point it to that label,
2812 and move the block of code to the spot we found. */
2817 /* Just ignore jumps to labels that were never emitted.
2818 These always indicate compilation errors. */
2822 /* If it's not safe to move the sequence, then we
2823 mustn't try. */
2826 {
2827 rtx target
2834 /* Don't move things inside a tablejump. */
2847 /* Don't move things inside a tablejump. */
2858 {
2862 /* Ensure our label doesn't go away. */
2865 /* Verify that uid_loop is large enough and that
2866 we can invert P. */
2868 {
2871 /* If no suitable BARRIER was found, create a suitable
2872 one before TARGET. Since TARGET is a fall through
2873 path, we'll need to insert an jump around our block
2874 and a add a BARRIER before TARGET.
2876 This creates an extra unconditional jump outside
2877 the loop. However, the benefits of removing rarely
2878 executed instructions from inside the loop usually
2879 outweighs the cost of the extra unconditional jump
2880 outside the loop. */
2882 {
2883 rtx temp;
2890 }
2892 /* Include the BARRIER after INSN and copy the
2893 block after LOC. */
2895 last_insn_to_move);
2898 /* All those insns are now in TARGET_LOOP. */
2904 /* The label jumped to by INSN is no longer a loop
2905 exit. Unless INSN does not have a label (e.g.,
2906 it is a RETURN insn), search loop->exit_labels
2907 to find its label_ref, and remove it. Also turn
2908 off LABEL_OUTSIDE_LOOP_P bit. */
2910 {
2915 {
2919 else
2922 }
2928 /* If we didn't find it, then something is
2929 wrong. */
2932 }
2934 /* P is now a jump outside the loop, so it must be put
2935 in loop->exit_labels, and marked as such.
2936 The easiest way to do this is to just call
2937 mark_loop_jump again for P. */
2940 /* If INSN now jumps to the insn after it,
2941 delete INSN. */
2946 }
2948 /* Continue the loop after where the conditional
2949 branch used to jump, since the only branch insn
2950 in the block (if it still remains) is an inter-loop
2951 branch and hence needs no processing. */
2957 /* This loop will be continued with NEXT_INSN (insn). */
2959 }
2960 }
2961 }
2962 }
2963 }
2965 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2966 loops it is contained in, mark the target loop invalid.
2968 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2970 static void
2972 rtx x;
2974 {
2980 {
2992 /* There could be a label reference in here. */
3004 /* This may refer to a LABEL_REF or SYMBOL_REF. */
3016 /* Link together all labels that branch outside the loop. This
3017 is used by final_[bg]iv_value and the loop unrolling code. Also
3018 mark this LABEL_REF so we know that this branch should predict
3019 false. */
3021 /* A check to make sure the label is not in an inner nested loop,
3022 since this does not count as a loop exit. */
3024 {
3029 }
3030 else
3034 {
3043 }
3045 /* If this is inside a loop, but not in the current loop or one enclosed
3046 by it, it invalidates at least one loop. */
3051 /* We must invalidate every nested loop containing the target of this
3052 label, except those that also contain the jump insn. */
3055 {
3056 /* Stop when we reach a loop that also contains the jump insn. */
3061 /* If we get here, we know we need to invalidate a loop. */
3068 }
3072 /* If this is not setting pc, ignore. */
3094 /* Strictly speaking this is not a jump into the loop, only a possible
3095 jump out of the loop. However, we have no way to link the destination
3096 of this jump onto the list of exit labels. To be safe we mark this
3097 loop and any containing loops as invalid. */
3099 {
3101 {
3107 }
3108 }
3110 }
3111 }
3112 \f
3113 /* Return nonzero if there is a label in the range from
3114 insn INSN to and including the insn whose luid is END
3115 INSN must have an assigned luid (i.e., it must not have
3116 been previously created by loop.c). */
3118 static int
3120 rtx insn;
3122 {
3124 {
3128 }
3131 }
3133 /* Record that a memory reference X is being set. */
3135 static void
3137 rtx x;
3138 rtx y ATTRIBUTE_UNUSED;
3140 {
3144 /* Count number of memory writes.
3145 This affects heuristics in strength_reduce. */
3146 num_mem_sets++;
3148 /* BLKmode MEM means all memory is clobbered. */
3150 {
3153 else
3157 }
3160 }
3162 /* X is a value modified by an INSN that references a biv inside a loop
3163 exit test (ie, X is somehow related to the value of the biv). If X
3164 is a pseudo that is used more than once, then the biv is (effectively)
3165 used more than once. DATA is really an `int *', and is set if the
3166 biv is used more than once. */
3168 static void
3170 rtx x;
3171 rtx y ATTRIBUTE_UNUSED;
3173 {
3186 /* If we do not have usage information, or if we know the register
3187 is used more than once, note that fact for check_dbra_loop. */
3192 }
3193 \f
3194 /* Return nonzero if the rtx X is invariant over the current loop.
3196 The value is 2 if we refer to something only conditionally invariant.
3198 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3199 Otherwise, a memory ref is invariant if it does not conflict with
3200 anything stored in `loop_store_mems'. */
3202 int
3206 {
3211 rtx mem_list_entry;
3217 {
3225 /* A LABEL_REF is normally invariant, however, if we are unrolling
3226 loops, and this label is inside the loop, then it isn't invariant.
3227 This is because each unrolled copy of the loop body will have
3228 a copy of this label. If this was invariant, then an insn loading
3229 the address of this label into a register might get moved outside
3230 the loop, and then each loop body would end up using the same label.
3232 We don't know the loop bounds here though, so just fail for all
3233 labels. */
3236 else
3245 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3246 since the reg might be set by initialization within the loop. */
3263 /* Volatile memory references must be rejected. Do this before
3264 checking for read-only items, so that volatile read-only items
3265 will be rejected also. */
3269 /* If we had a subroutine call, any location in memory could
3270 have been clobbered. We used to test here for volatile and
3271 readonly, but true_dependence knows how to do that better
3272 than we do. */
3277 /* See if there is any dependence between a store and this load. */
3280 {
3286 }
3288 /* It's not invalidated by a store in memory
3289 but we must still verify the address is invariant. */
3293 /* Don't mess with insns declared volatile. */
3300 }
3304 {
3306 {
3312 }
3314 {
3317 {
3323 }
3325 }
3326 }
3329 }
3331 \f
3332 /* Return nonzero if all the insns in the loop that set REG
3333 are INSN and the immediately following insns,
3334 and if each of those insns sets REG in an invariant way
3335 (not counting uses of REG in them).
3337 The value is 2 if some of these insns are only conditionally invariant.
3339 We assume that INSN itself is the first set of REG
3340 and that its source is invariant. */
3342 static int
3347 {
3350 rtx temp;
3351 /* Number of sets we have to insist on finding after INSN. */
3357 /* If N_SETS hit the limit, we can't rely on its value. */
3364 {
3366 rtx set;
3371 /* If library call, skip to end of it. */
3380 {
3385 {
3386 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3387 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3388 notes are OK. */
3394 }
3395 }
3397 count--;
3399 {
3402 }
3403 }
3406 /* If loop_invariant_p ever returned 2, we return 2. */
3408 }
3410 #if 0
3411 /* I don't think this condition is sufficient to allow INSN
3412 to be moved, so we no longer test it. */
3414 /* Return 1 if all insns in the basic block of INSN and following INSN
3415 that set REG are invariant according to TABLE. */
3417 static int
3421 {
3426 {
3435 {
3438 }
3439 }
3440 }
3442 \f
3443 /* Look at all uses (not sets) of registers in X. For each, if it is
3444 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3445 a different insn, set USAGE[REGNO] to const0_rtx. */
3447 static void
3449 rtx insn;
3450 rtx x;
3451 varray_type usage;
3452 {
3464 {
3465 /* Don't count SET_DEST if it is a REG; otherwise count things
3466 in SET_DEST because if a register is partially modified, it won't
3467 show up as a potential movable so we don't care how USAGE is set
3468 for it. */
3472 }
3473 else
3475 {
3481 }
3482 }
3483 \f
3484 /* Count and record any set in X which is contained in INSN. Update
3485 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3487 static void
3490 varray_type may_not_move;
3492 {
3494 /* Don't move a reg that has an explicit clobber.
3495 It's not worth the pain to try to do it correctly. */
3499 {
3507 {
3509 /* If this is the first setting of this reg
3510 in current basic block, and it was set before,
3511 it must be set in two basic blocks, so it cannot
3512 be moved out of the loop. */
3516 /* If this is not first setting in current basic block,
3517 see if reg was used in between previous one and this.
3518 If so, neither one can be moved. */
3525 }
3526 }
3527 }
3529 /* Increment SET_IN_LOOP at the index of each register
3530 that is modified by an insn between FROM and TO.
3531 If the value of an element of SET_IN_LOOP becomes 127 or more,
3532 stop incrementing it, to avoid overflow.
3534 Store in SINGLE_USAGE[I] the single insn in which register I is
3535 used, if it is only used once. Otherwise, it is set to 0 (for no
3536 uses) or const0_rtx for more than one use. This parameter may be zero,
3537 in which case this processing is not done.
3539 Store in *COUNT_PTR the number of actual instruction
3540 in the loop. We use this to decide what is worth moving out. */
3542 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3543 In that case, it is the insn that last set reg n. */
3545 static void
3548 varray_type may_not_move;
3549 varray_type single_usage;
3552 {
3558 {
3560 {
3563 /* Record registers that have exactly one use. */
3566 /* Include uses in REG_EQUAL notes. */
3574 {
3579 }
3580 }
3584 }
3587 /* Clean up. */
3589 }
3590 \f
3591 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3592 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3593 contained in insn INSN is used by any insn that precedes INSN in
3594 cyclic order starting from the loop entry point.
3596 We don't want to use INSN_LUID here because if we restrict INSN to those
3597 that have a valid INSN_LUID, it means we cannot move an invariant out
3598 from an inner loop past two loops. */
3600 static int
3604 {
3606 rtx p;
3608 /* Scan forward checking for register usage. If we hit INSN, we
3609 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3611 {
3618 }
3621 }
3622 \f
3623 /* A "basic induction variable" or biv is a pseudo reg that is set
3624 (within this loop) only by incrementing or decrementing it. */
3625 /* A "general induction variable" or giv is a pseudo reg whose
3626 value is a linear function of a biv. */
3628 /* Bivs are recognized by `basic_induction_var';
3629 Givs by `general_induction_var'. */
3631 /* Indexed by register number, indicates whether or not register is an
3632 induction variable, and if so what type. */
3634 varray_type reg_iv_type;
3636 /* Indexed by register number, contains pointer to `struct induction'
3637 if register is an induction variable. This holds general info for
3638 all induction variables. */
3640 varray_type reg_iv_info;
3642 /* Indexed by register number, contains pointer to `struct iv_class'
3643 if register is a basic induction variable. This holds info describing
3644 the class (a related group) of induction variables that the biv belongs
3645 to. */
3649 /* The head of a list which links together (via the next field)
3650 every iv class for the current loop. */
3654 /* Givs made from biv increments are always splittable for loop unrolling.
3655 Since there is no regscan info for them, we have to keep track of them
3656 separately. */
3659 /* Communication with routines called via `note_stores'. */
3663 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3667 /* ??? Unfinished optimizations, and possible future optimizations,
3668 for the strength reduction code. */
3670 /* ??? The interaction of biv elimination, and recognition of 'constant'
3671 bivs, may cause problems. */
3673 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3674 performance problems.
3676 Perhaps don't eliminate things that can be combined with an addressing
3677 mode. Find all givs that have the same biv, mult_val, and add_val;
3678 then for each giv, check to see if its only use dies in a following
3679 memory address. If so, generate a new memory address and check to see
3680 if it is valid. If it is valid, then store the modified memory address,
3681 otherwise, mark the giv as not done so that it will get its own iv. */
3683 /* ??? Could try to optimize branches when it is known that a biv is always
3684 positive. */
3686 /* ??? When replace a biv in a compare insn, we should replace with closest
3687 giv so that an optimized branch can still be recognized by the combiner,
3688 e.g. the VAX acb insn. */
3690 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3691 was rerun in loop_optimize whenever a register was added or moved.
3692 Also, some of the optimizations could be a little less conservative. */
3693 \f
3694 /* Scan the loop body and call FNCALL for each insn. In the addition to the
3695 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
3696 callback.
3698 NOT_EVERY_ITERATION if current insn is not executed at least once for every
3699 loop iteration except for the last one.
3701 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
3702 loop iteration.
3703 */
3704 void
3707 loop_insn_callback fncall;
3708 {
3709 /* This is 1 if current insn is not executed at least once for every loop
3710 iteration. */
3715 rtx p;
3717 /* If loop_scan_start points to the loop exit test, we have to be wary of
3718 subversive use of gotos inside expression statements. */
3722 /* Scan through loop to find all possible bivs. */
3727 {
3730 /* Past CODE_LABEL, we get to insns that may be executed multiple
3731 times. The only way we can be sure that they can't is if every
3732 jump insn between here and the end of the loop either
3733 returns, exits the loop, is a jump to a location that is still
3734 behind the label, or is a jump to the loop start. */
3737 {
3743 {
3748 {
3751 else
3755 }
3763 {
3766 }
3767 }
3768 }
3770 /* Past a jump, we get to insns for which we can't count
3771 on whether they will be executed during each iteration. */
3772 /* This code appears twice in strength_reduce. There is also similar
3773 code in scan_loop. */
3775 /* If we enter the loop in the middle, and scan around to the
3776 beginning, don't set not_every_iteration for that.
3777 This can be any kind of jump, since we want to know if insns
3778 will be executed if the loop is executed. */
3783 {
3786 /* If this is a jump outside the loop, then it also doesn't
3787 matter. Check to see if the target of this branch is on the
3788 loop->exits_labels list. */
3796 }
3799 {
3800 /* At the virtual top of a converted loop, insns are again known to
3801 be executed each iteration: logically, the loop begins here
3802 even though the exit code has been duplicated.
3804 Insns are also again known to be executed each iteration at
3805 the LOOP_CONT note. */
3811 loop_depth++;
3813 loop_depth--;
3814 }
3816 /* Note if we pass a loop latch. If we do, then we can not clear
3817 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3818 a loop since a jump before the last CODE_LABEL may have started
3819 a new loop iteration.
3821 Note that LOOP_TOP is only set for rotated loops and we need
3822 this check for all loops, so compare against the CODE_LABEL
3823 which immediately follows LOOP_START. */
3828 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3829 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3830 or not an insn is known to be executed each iteration of the
3831 loop, whether or not any iterations are known to occur.
3833 Therefore, if we have just passed a label and have no more labels
3834 between here and the test insn of the loop, and we have not passed
3835 a jump to the top of the loop, then we know these insns will be
3836 executed each iteration. */
3839 && !past_loop_latch
3844 }
3845 }
3846 \f
3847 /* Perform strength reduction and induction variable elimination.
3849 Pseudo registers created during this function will be beyond the last
3850 valid index in several tables including n_times_set and regno_last_uid.
3851 This does not cause a problem here, because the added registers cannot be
3852 givs outside of their loop, and hence will never be reconsidered.
3853 But scan_loop must check regnos to make sure they are in bounds. */
3855 static void
3860 {
3861 rtx p;
3862 /* Temporary list pointers for traversing loop_iv_list. */
3865 /* Ratio of extra register life span we can justify
3866 for saving an instruction. More if loop doesn't call subroutines
3867 since in that case saving an insn makes more difference
3868 and more registers are available. */
3869 /* ??? could set this to last value of threshold in move_movables */
3871 /* Map of pseudo-register replacements. */
3875 rtx test;
3876 rtx end_insert_before;
3891 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3892 must be put before this insn, so that they will appear in the right
3893 order (i.e. loop order).
3895 If loop_end is the end of the current function, then emit a
3896 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3897 dummy note insn. */
3900 else
3905 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3906 Make a sanity check against n_times_set. */
3908 {
3910 /* Above happens if register modified by subreg, etc. */
3911 /* Make sure it is not recognized as a basic induction var: */
3913 /* If never incremented, it is invariant that we decided not to
3914 move. So leave it alone. */
3916 {
3927 }
3928 else
3929 {
3934 }
3935 }
3937 /* Exit if there are no bivs. */
3939 {
3940 /* Can still unroll the loop anyways, but indicate that there is no
3941 strength reduction info available. */
3946 }
3948 /* Find initial value for each biv by searching backwards from loop_start,
3949 halting at first label. Also record any test condition. */
3953 {
3962 /* Record any test of a biv that branches around the loop if no store
3963 between it and the start of loop. We only care about tests with
3964 constants and registers and only certain of those. */
3974 {
3975 /* If an NE test, we have an initial value! */
3977 {
3981 }
3982 else
3984 }
3985 }
3987 /* Look at the each biv and see if we can say anything better about its
3988 initial value from any initializing insns set up above. (This is done
3989 in two passes to avoid missing SETs in a PARALLEL.) */
3991 {
3992 rtx src;
3993 rtx note;
3998 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3999 is a constant, use the value of that. */
4005 else
4016 {
4020 {
4022 {
4025 }
4026 else
4027 {
4030 }
4031 }
4032 }
4033 else
4034 {
4038 /* Biv initial value is not a simple move. If it is the sum of
4039 another biv and a constant, check if both bivs are incremented
4040 in lockstep. Then we are actually looking at a giv.
4041 For simplicity, we only handle the case where there is but a
4042 single increment, and the register is not used elsewhere. */
4050 {
4056 }
4058 /* Now, can we transform this biv into a giv? */
4062 /* init_insn is only set to insns that are before loop_start
4063 without any intervening labels. */
4066 /* The register from BL2 must be set before the register from
4067 BL is set, or we must be able to move the latter set after
4068 the former set. Currently there can't be any labels
4069 in-between when biv_total_increment returns nonzero both times
4070 but we test it here in case some day some real cfg analysis
4071 gets used to set always_computable. */
4080 {
4088 /* Let this giv be discovered by the generic code. */
4091 /* We can get better optimization if we can move the giv setting
4092 before the first giv use. */
4096 dominator)
4099 {
4100 rtx p;
4101 rtx next;
4104 {
4109 #ifdef HAVE_cc0
4112 #endif
4114 }
4118 /* Avoid problems with luids by actually moving the insn
4119 and adjusting all luids in the range. */
4124 /* If the only purpose of the init insn is to initialize
4125 this giv, delete it. */
4129 }
4131 {
4138 }
4139 /* Remove this biv from the chain. */
4141 }
4143 /* If we can't make it a giv,
4144 let biv keep initial value of "itself". */
4147 }
4148 }
4150 /* If a biv is unconditionally incremented several times in a row, convert
4151 all but the last increment into a giv. */
4153 /* Get an upper bound for the number of registers
4154 we might have after all bivs have been processed. */
4159 /* If the loop contains volatile memory references do not allow any
4160 replacements to take place, since this could loose the volatile
4161 markers. */
4163 {
4166 /* Reallocate reg_iv_type and reg_iv_info. */
4171 {
4175 /* The biv increments lists are in reverse order. Fix this
4176 first. */
4178 {
4182 }
4184 /* We must guard against the case that an early exit between v->insn
4185 and next->insn leaves the biv live after the loop, since that
4186 would mean that we'd be missing an increment for the final
4187 value. The following test to set biv_dead_after_loop is like
4188 the first part of the test to set bl->eliminable.
4189 We don't check here if we can calculate the final value, since
4190 this can't succeed if we already know that there is a jump
4191 between v->insn and next->insn, yet next->always_executed is
4192 set and next->maybe_multiple is cleared. Such a combination
4193 implies that the jump destination is outside the loop.
4194 If we want to make this check more sophisticated, we should
4195 check each branch between v->insn and next->insn individually
4196 to see if the biv is dead at its destination. */
4203 #ifdef HAVE_decrement_and_branch_until_zero
4205 #endif
4210 {
4211 HOST_WIDE_INT offset;
4214 rtx next_loc_insn;
4224 || ! (biv_dead_after_loop
4226 {
4229 }
4236 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4237 have been allocated with some slop space, so we may not
4238 actually need to reallocate them. If we do, the following
4239 if statement will be executed just once in this loop. */
4241 {
4242 /* Grow all the remaining arrays. */
4247 }
4249 /* Some bivs are incremented with a multi-insn sequence.
4250 The first insn contains the add. */
4260 {
4263 }
4265 /* Here we can try to eliminate the increment by combining
4266 it into the uses. */
4268 /* Set last_use_insn so that we can check against it. */
4273 {
4277 {
4279 }
4280 }
4282 /* If we can't get the LUIDs for the insns, we can't
4283 calculate the lifetime. This is likely from unrolling
4284 of an inner loop, so there is little point in making this
4285 a DEST_REG giv anyways. */
4289 {
4290 /* Change the increment at NEXT back to what it was. */
4296 }
4330 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4331 it must be replaced. */
4336 /* Remove the increment from the list of biv increments,
4337 and record it as a giv. */
4346 /* Now replace the biv with DEST_REG in all insns between
4347 the replaced increment and the next increment, and
4348 remember the last insn that needed a replacement. */
4352 {
4353 rtx note;
4358 {
4362 }
4364 {
4368 }
4369 }
4373 /* If the lifetime is zero, it means that this register is really
4374 a dead store. So mark this as a giv that can be ignored.
4375 This will not prevent the biv from being eliminated. */
4383 }
4384 }
4385 }
4388 /* Search the loop for general induction variables. */
4392 /* Try to calculate and save the number of loop iterations. This is
4393 set to zero if the actual number can not be calculated. This must
4394 be called after all giv's have been identified, since otherwise it may
4395 fail if the iteration variable is a giv. */
4399 /* Now for each giv for which we still don't know whether or not it is
4400 replaceable, check to see if it is replaceable because its final value
4401 can be calculated. This must be done after loop_iterations is called,
4402 so that final_giv_value will work correctly. */
4405 {
4411 }
4413 /* Try to prove that the loop counter variable (if any) is always
4414 nonnegative; if so, record that fact with a REG_NONNEG note
4415 so that "decrement and branch until zero" insn can be used. */
4418 /* Create reg_map to hold substitutions for replaceable giv regs.
4419 Some givs might have been made from biv increments, so look at
4420 reg_iv_type for a suitable size. */
4424 /* Examine each iv class for feasibility of strength reduction/induction
4425 variable elimination. */
4428 {
4435 /* Test whether it will be possible to eliminate this biv
4436 provided all givs are reduced. This is possible if either
4437 the reg is not used outside the loop, or we can compute
4438 what its final value will be.
4440 For architectures with a decrement_and_branch_until_zero insn,
4441 don't do this if we put a REG_NONNEG note on the endtest for
4442 this biv. */
4444 /* Compare against bl->init_insn rather than loop_start.
4445 We aren't concerned with any uses of the biv between
4446 init_insn and loop_start since these won't be affected
4447 by the value of the biv elsewhere in the function, so
4448 long as init_insn doesn't use the biv itself.
4449 March 14, 1989 -- self@bayes.arc.nasa.gov */
4455 #ifdef HAVE_decrement_and_branch_until_zero
4457 #endif
4460 #ifdef HAVE_decrement_and_branch_until_zero
4462 #endif
4463 ))
4465 insn_count);
4466 else
4467 {
4469 {
4477 }
4478 }
4480 /* Combine all giv's for this iv_class. */
4483 /* This will be true at the end, if all givs which depend on this
4484 biv have been strength reduced.
4485 We can't (currently) eliminate the biv unless this is so. */
4488 /* Check each giv in this class to see if we will benefit by reducing
4489 it. Skip giv's combined with others. */
4491 {
4499 /* Reduce benefit if not replaceable, since we will insert
4500 a move-insn to replace the insn that calculates this giv.
4501 Don't do this unless the giv is a user variable, since it
4502 will often be marked non-replaceable because of the duplication
4503 of the exit code outside the loop. In such a case, the copies
4504 we insert are dead and will be deleted. So they don't have
4505 a cost. Similar situations exist. */
4506 /* ??? The new final_[bg]iv_value code does a much better job
4507 of finding replaceable giv's, and hence this code may no longer
4508 be necessary. */
4513 /* Decrease the benefit to count the add-insns that we will
4514 insert to increment the reduced reg for the giv. */
4517 /* Decide whether to strength-reduce this giv or to leave the code
4518 unchanged (recompute it from the biv each time it is used).
4519 This decision can be made independently for each giv. */
4521 #ifdef AUTO_INC_DEC
4522 /* Attempt to guess whether autoincrement will handle some of the
4523 new add insns; if so, increase BENEFIT (undo the subtraction of
4524 add_cost that was done above). */
4527 {
4540 }
4541 #endif
4543 /* If an insn is not to be strength reduced, then set its ignore
4544 flag, and clear all_reduced. */
4546 /* A giv that depends on a reversed biv must be reduced if it is
4547 used after the loop exit, otherwise, it would have the wrong
4548 value after the loop exit. To make it simple, just reduce all
4549 of such giv's whether or not we know they are used after the loop
4550 exit. */
4554 {
4562 }
4563 else
4564 {
4565 /* Check that we can increment the reduced giv without a
4566 multiply insn. If not, reject it. */
4571 {
4579 }
4580 }
4581 }
4583 /* Check for givs whose first use is their definition and whose
4584 last use is the definition of another giv. If so, it is likely
4585 dead and should not be used to derive another giv nor to
4586 eliminate a biv. */
4588 {
4594 {
4600 }
4603 {
4609 }
4610 }
4612 /* Now that we know which givs will be reduced, try to rearrange the
4613 combinations to reduce register pressure.
4614 recombine_givs calls find_life_end, which needs reg_iv_type and
4615 reg_iv_info to be valid for all pseudos. We do the necessary
4616 reallocation here since it allows to check if there are still
4617 more bivs to process. */
4620 {
4621 /* If there are still more bivs to process, allocate some slack
4622 space so that we're not constantly reallocating these arrays. */
4625 /* Reallocate reg_iv_type and reg_iv_info. */
4628 }
4631 /* Reduce each giv that we decided to reduce. */
4634 {
4637 {
4640 /* If the code for derived givs immediately below has already
4641 allocated a new_reg, we must keep it. */
4646 {
4649 /* In case d->dest_reg is not replaceable, we have
4650 to replace it in v->insn now. */
4657 /* For each place where the biv is incremented, add an
4658 insn to set the new, reduced reg for the giv.
4659 We used to do this only for biv_count != 1, but
4660 this fails when there is a giv after a single biv
4661 increment, e.g. when the last giv was expressed as
4662 pre-decrement. */
4664 {
4665 /* We always emit reduced giv increments before the
4666 biv increment when bl->biv_count != 1. So by
4667 emitting the add insns for derived givs after the
4668 biv increment, they pick up the updated value of
4669 the reduced giv.
4670 If the reduced giv is processed with
4671 auto_inc_opt == 1, then it is incremented earlier
4672 than the biv, hence we'll still pick up the right
4673 value.
4674 If it's processed with auto_inc_opt == -1,
4675 that implies that the biv increment is before the
4676 first reduced giv's use. The derived giv's lifetime
4677 is after the reduced giv's lifetime, hence in this
4678 case, the biv increment doesn't matter. */
4680 }
4682 }
4684 #ifdef AUTO_INC_DEC
4685 /* If the target has auto-increment addressing modes, and
4686 this is an address giv, then try to put the increment
4687 immediately after its use, so that flow can create an
4688 auto-increment addressing mode. */
4691 /* We don't handle reversed biv's because bl->biv->insn
4692 does not have a valid INSN_LUID. */
4696 {
4697 /* If other giv's have been combined with this one, then
4698 this will work only if all uses of the other giv's occur
4699 before this giv's insn. This is difficult to check.
4701 We simplify this by looking for the common case where
4702 there is one DEST_REG giv, and this giv's insn is the
4703 last use of the dest_reg of that DEST_REG giv. If the
4704 increment occurs after the address giv, then we can
4705 perform the optimization. (Otherwise, the increment
4706 would have to go before other_giv, and we would not be
4707 able to combine it with the address giv to get an
4708 auto-inc address.) */
4710 {
4715 {
4718 else
4720 }
4727 }
4728 /* Check for case where increment is before the address
4729 giv. Do this test in "loop order". */
4738 else
4741 #ifdef HAVE_cc0
4742 {
4743 rtx prev;
4745 /* We can't put an insn immediately after one setting
4746 cc0, or immediately before one using cc0. */
4753 }
4754 #endif
4758 }
4759 #endif
4761 /* For each place where the biv is incremented, add an insn
4762 to increment the new, reduced reg for the giv. */
4764 {
4765 rtx insert_before;
4771 else
4778 /* A multiply is acceptable here
4779 since this is presumed to be seldom executed. */
4782 }
4784 /* Add code at loop start to initialize giv's reduced reg. */
4788 }
4789 }
4791 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4792 as not reduced.
4794 For each giv register that can be reduced now: if replaceable,
4795 substitute reduced reg wherever the old giv occurs;
4796 else add new move insn "giv_reg = reduced_reg". */
4799 {
4806 /* Update expression if this was combined, in case other giv was
4807 replaced. */
4813 /* Store reduced reg as the address in the memref where we found
4814 this giv. */
4817 {
4820 #if 0
4821 /* I can no longer duplicate the original problem. Perhaps
4822 this is unnecessary now? */
4824 /* Replaceable; it isn't strictly necessary to delete the old
4825 insn and emit a new one, because v->dest_reg is now dead.
4827 However, especially when unrolling loops, the special
4828 handling for (set REG0 REG1) in the second cse pass may
4829 make v->dest_reg live again. To avoid this problem, emit
4830 an insn to set the original giv reg from the reduced giv.
4831 We can not delete the original insn, since it may be part
4832 of a LIBCALL, and the code in flow that eliminates dead
4833 libcalls will fail if it is deleted. */
4836 #endif
4837 }
4838 else
4839 {
4840 /* Not replaceable; emit an insn to set the original giv reg from
4841 the reduced giv, same as above. */
4844 }
4846 /* When a loop is reversed, givs which depend on the reversed
4847 biv, and which are live outside the loop, must be set to their
4848 correct final value. This insn is only needed if the giv is
4849 not replaceable. The correct final value is the same as the
4850 value that the giv starts the reversed loop with. */
4855 {
4856 rtx insert_before;
4858 /* If the loop has multiple exits, emit the insn before the
4859 loop to ensure that it will always be executed no matter
4860 how the loop exits. Otherwise, emit the insn after the loop,
4861 since this is slightly more efficient. */
4864 else
4867 insert_before);
4869 #if 0
4870 /* If the insn to set the final value of the giv was emitted
4871 before the loop, then we must delete the insn inside the loop
4872 that sets it. If this is a LIBCALL, then we must delete
4873 every insn in the libcall. Note, however, that
4874 final_giv_value will only succeed when there are multiple
4875 exits if the giv is dead at each exit, hence it does not
4876 matter that the original insn remains because it is dead
4877 anyways. */
4878 /* Delete the insn inside the loop that sets the giv since
4879 the giv is now set before (or after) the loop. */
4881 #endif
4882 }
4885 {
4890 }
4891 }
4893 /* All the givs based on the biv bl have been reduced if they
4894 merit it. */
4896 /* For each giv not marked as maybe dead that has been combined with a
4897 second giv, clear any "maybe dead" mark on that second giv.
4898 v->new_reg will either be or refer to the register of the giv it
4899 combined with.
4901 Doing this clearing avoids problems in biv elimination where a
4902 giv's new_reg is a complex value that can't be put in the insn but
4903 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4904 Since the register will be used in either case, we'd prefer it be
4905 used from the simpler giv. */
4911 /* Try to eliminate the biv, if it is a candidate.
4912 This won't work if ! all_reduced,
4913 since the givs we planned to use might not have been reduced.
4915 We have to be careful that we didn't initially think we could eliminate
4916 this biv because of a giv that we now think may be dead and shouldn't
4917 be used as a biv replacement.
4919 Also, there is the possibility that we may have a giv that looks
4920 like it can be used to eliminate a biv, but the resulting insn
4921 isn't valid. This can happen, for example, on the 88k, where a
4922 JUMP_INSN can compare a register only with zero. Attempts to
4923 replace it with a compare with a constant will fail.
4925 Note that in cases where this call fails, we may have replaced some
4926 of the occurrences of the biv with a giv, but no harm was done in
4927 doing so in the rare cases where it can occur. */
4931 {
4932 /* ?? If we created a new test to bypass the loop entirely,
4933 or otherwise drop straight in, based on this test, then
4934 we might want to rewrite it also. This way some later
4935 pass has more hope of removing the initialization of this
4936 biv entirely. */
4938 /* If final_value != 0, then the biv may be used after loop end
4939 and we must emit an insn to set it just in case.
4941 Reversed bivs already have an insn after the loop setting their
4942 value, so we don't need another one. We can't calculate the
4943 proper final value for such a biv here anyways. */
4945 {
4946 rtx insert_before;
4948 /* If the loop has multiple exits, emit the insn before the
4949 loop to ensure that it will always be executed no matter
4950 how the loop exits. Otherwise, emit the insn after the
4951 loop, since this is slightly more efficient. */
4954 else
4958 end_insert_before);
4959 }
4961 #if 0
4962 /* Delete all of the instructions inside the loop which set
4963 the biv, as they are all dead. If is safe to delete them,
4964 because an insn setting a biv will never be part of a libcall. */
4965 /* However, deleting them will invalidate the regno_last_uid info,
4966 so keeping them around is more convenient. Final_biv_value
4967 will only succeed when there are multiple exits if the biv
4968 is dead at each exit, hence it does not matter that the original
4969 insn remains, because it is dead anyways. */
4972 #endif
4977 }
4978 }
4980 /* Go through all the instructions in the loop, making all the
4981 register substitutions scheduled in REG_MAP. */
4986 {
4990 }
4993 {
4994 /* When we completely unroll a loop we will likely not need the increment
4995 of the loop BIV and we will not need the conditional branch at the
4996 end of the loop. */
4999 #ifdef HAVE_cc0
5000 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
5001 need the comparison before the conditional branch at the end of the
5002 loop. */
5004 #endif
5006 /* We'll need one copy for each loop iteration. */
5009 /* A little slop to account for the ability to remove initialization
5010 code, better CSE, and other secondary benefits of completely
5011 unrolling some loops. */
5014 /* Clamp the value. */
5017 }
5019 /* Unroll loops from within strength reduction so that we can use the
5020 induction variable information that strength_reduce has already
5021 collected. Always unroll loops that would be as small or smaller
5022 unrolled than when rolled. */
5028 #ifdef HAVE_decrement_and_branch_on_count
5029 /* Instrument the loop with BCT insn. */
5038 egress:
5044 }
5045 \f
5046 /*Record all basic induction variables calculated in the insn. */
5047 static rtx
5050 rtx p;
5053 {
5054 rtx set;
5055 rtx dest_reg;
5056 rtx inc_val;
5057 rtx mult_val;
5063 {
5068 {
5073 {
5074 /* It is a possible basic induction variable.
5075 Create and initialize an induction structure for it. */
5083 }
5086 }
5087 }
5089 }
5090 \f
5091 /* Record all givs calculated in the insn.
5092 A register is a giv if: it is only set once, it is a function of a
5093 biv and a constant (or invariant), and it is not a biv. */
5094 static rtx
5097 rtx p;
5100 {
5101 rtx set;
5102 /* Look for a general induction variable in a register. */
5107 {
5108 rtx src_reg;
5109 rtx dest_reg;
5110 rtx add_val;
5111 rtx mult_val;
5114 rtx last_consec_insn;
5123 /* Equivalent expression is a giv. */
5128 /* Don't try to handle any regs made by loop optimization.
5129 We have nothing on them in regno_first_uid, etc. */
5131 /* Don't recognize a BASIC_INDUCT_VAR here. */
5133 /* This must be the only place where the register is set. */
5135 /* or all sets must be consecutive and make a giv. */
5140 {
5144 /* If this is a library call, increase benefit. */
5148 /* Skip the consecutive insns, if there are any. */
5156 }
5157 }
5159 #ifndef DONT_REDUCE_ADDR
5160 /* Look for givs which are memory addresses. */
5161 /* This resulted in worse code on a VAX 8600. I wonder if it
5162 still does. */
5165 maybe_multiple);
5166 #endif
5168 /* Update the status of whether giv can derive other givs. This can
5169 change when we pass a label or an insn that updates a biv. */
5174 }
5175 \f
5176 /* Return 1 if X is a valid source for an initial value (or as value being
5177 compared against in an initial test).
5179 X must be either a register or constant and must not be clobbered between
5180 the current insn and the start of the loop.
5182 INSN is the insn containing X. */
5184 static int
5186 rtx x;
5187 rtx insn;
5189 rtx loop_start;
5190 {
5194 /* Only consider pseudos we know about initialized in insns whose luids
5195 we know. */
5200 /* Don't use call-clobbered registers across a call which clobbers it. On
5201 some machines, don't use any hard registers at all. */
5203 && (SMALL_REGISTER_CLASSES
5207 /* Don't use registers that have been clobbered before the start of the
5208 loop. */
5213 }
5214 \f
5215 /* Scan X for memory refs and check each memory address
5216 as a possible giv. INSN is the insn whose pattern X comes from.
5217 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5218 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5219 more thanonce in each loop iteration. */
5221 static void
5224 rtx x;
5225 rtx insn;
5227 {
5237 {
5253 {
5254 rtx src_reg;
5255 rtx add_val;
5256 rtx mult_val;
5259 /* This code used to disable creating GIVs with mult_val == 1 and
5260 add_val == 0. However, this leads to lost optimizations when
5261 it comes time to combine a set of related DEST_ADDR GIVs, since
5262 this one would not be seen. */
5266 {
5267 /* Found one; record it. */
5276 }
5277 }
5282 }
5284 /* Recursively scan the subexpressions for other mem refs. */
5290 maybe_multiple);
5294 maybe_multiple);
5295 }
5296 \f
5297 /* Fill in the data about one biv update.
5298 V is the `struct induction' in which we record the biv. (It is
5299 allocated by the caller, with alloca.)
5300 INSN is the insn that sets it.
5301 DEST_REG is the biv's reg.
5303 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5304 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5305 being set to INC_VAL.
5307 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5308 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5309 can be executed more than once per iteration. If MAYBE_MULTIPLE
5310 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5311 executed exactly once per iteration. */
5313 static void
5317 rtx insn;
5318 rtx dest_reg;
5319 rtx inc_val;
5320 rtx mult_val;
5324 {
5338 /* Add this to the reg's iv_class, creating a class
5339 if this is the first incrementation of the reg. */
5343 {
5344 /* Create and initialize new iv_class. */
5354 /* Set initial value to the reg itself. */
5356 /* We haven't seen the initializing insn yet */
5366 /* Add this class to loop_iv_list. */
5370 /* Put it in the array of biv register classes. */
5372 }
5374 /* Update IV_CLASS entry for this biv. */
5382 {
5387 {
5391 }
5392 else
5393 {
5397 }
5398 }
5399 }
5400 \f
5401 /* Fill in the data about one giv.
5402 V is the `struct induction' in which we record the giv. (It is
5403 allocated by the caller, with alloca.)
5404 INSN is the insn that sets it.
5405 BENEFIT estimates the savings from deleting this insn.
5406 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5407 into a register or is used as a memory address.
5409 SRC_REG is the biv reg which the giv is computed from.
5410 DEST_REG is the giv's reg (if the giv is stored in a reg).
5411 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5412 LOCATION points to the place where this giv's value appears in INSN. */
5414 static void
5419 rtx insn;
5420 rtx src_reg;
5421 rtx dest_reg;
5427 {
5431 rtx temp;
5433 /* Attempt to prove constantness of the values. */
5462 /* The v->always_computable field is used in update_giv_derive, to
5463 determine whether a giv can be used to derive another giv. For a
5464 DEST_REG giv, INSN computes a new value for the giv, so its value
5465 isn't computable if INSN insn't executed every iteration.
5466 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5467 it does not compute a new value. Hence the value is always computable
5468 regardless of whether INSN is executed each iteration. */
5472 else
5478 {
5481 }
5483 {
5489 /* If the lifetime is zero, it means that this register is
5490 really a dead store. So mark this as a giv that can be
5491 ignored. This will not prevent the biv from being eliminated. */
5497 }
5499 /* Add the giv to the class of givs computed from one biv. */
5503 {
5506 /* Don't count DEST_ADDR. This is supposed to count the number of
5507 insns that calculate givs. */
5511 }
5512 else
5513 /* Fatal error, biv missing for this giv? */
5518 else
5519 {
5520 /* The giv can be replaced outright by the reduced register only if all
5521 of the following conditions are true:
5522 - the insn that sets the giv is always executed on any iteration
5523 on which the giv is used at all
5524 (there are two ways to deduce this:
5525 either the insn is executed on every iteration,
5526 or all uses follow that insn in the same basic block),
5527 - the giv is not used outside the loop
5528 - no assignments to the biv occur during the giv's lifetime. */
5531 /* Previous line always fails if INSN was moved by loop opt. */
5534 && (! not_every_iteration
5536 {
5537 /* Now check that there are no assignments to the biv within the
5538 giv's lifetime. This requires two separate checks. */
5540 /* Check each biv update, and fail if any are between the first
5541 and last use of the giv.
5543 If this loop contains an inner loop that was unrolled, then
5544 the insn modifying the biv may have been emitted by the loop
5545 unrolling code, and hence does not have a valid luid. Just
5546 mark the biv as not replaceable in this case. It is not very
5547 useful as a biv, because it is used in two different loops.
5548 It is very unlikely that we would be able to optimize the giv
5549 using this biv anyways. */
5553 {
5559 {
5563 }
5564 }
5566 /* If there are any backwards branches that go from after the
5567 biv update to before it, then this giv is not replaceable. */
5571 {
5575 }
5576 }
5577 else
5578 {
5579 /* May still be replaceable, we don't have enough info here to
5580 decide. */
5583 }
5584 }
5586 /* Record whether the add_val contains a const_int, for later use by
5587 combine_givs. */
5588 {
5593 ;
5597 {
5599 {
5604 else
5606 }
5609 }
5610 }
5613 {
5617 else
5633 {
5636 }
5637 else
5638 {
5641 }
5644 {
5647 }
5648 else
5649 {
5652 }
5653 }
5658 }
5661 /* All this does is determine whether a giv can be made replaceable because
5662 its final value can be calculated. This code can not be part of record_giv
5663 above, because final_giv_value requires that the number of loop iterations
5664 be known, and that can not be accurately calculated until after all givs
5665 have been identified. */
5667 static void
5671 {
5677 /* DEST_ADDR givs will never reach here, because they are always marked
5678 replaceable above in record_giv. */
5680 /* The giv can be replaced outright by the reduced register only if all
5681 of the following conditions are true:
5682 - the insn that sets the giv is always executed on any iteration
5683 on which the giv is used at all
5684 (there are two ways to deduce this:
5685 either the insn is executed on every iteration,
5686 or all uses follow that insn in the same basic block),
5687 - its final value can be calculated (this condition is different
5688 than the one above in record_giv)
5689 - no assignments to the biv occur during the giv's lifetime. */
5691 #if 0
5692 /* This is only called now when replaceable is known to be false. */
5693 /* Clear replaceable, so that it won't confuse final_giv_value. */
5695 #endif
5699 {
5702 rtx last_giv_use;
5706 /* When trying to determine whether or not a biv increment occurs
5707 during the lifetime of the giv, we can ignore uses of the variable
5708 outside the loop because final_value is true. Hence we can not
5709 use regno_last_uid and regno_first_uid as above in record_giv. */
5711 /* Search the loop to determine whether any assignments to the
5712 biv occur during the giv's lifetime. Start with the insn
5713 that sets the giv, and search around the loop until we come
5714 back to that insn again.
5716 Also fail if there is a jump within the giv's lifetime that jumps
5717 to somewhere outside the lifetime but still within the loop. This
5718 catches spaghetti code where the execution order is not linear, and
5719 hence the above test fails. Here we assume that the giv lifetime
5720 does not extend from one iteration of the loop to the next, so as
5721 to make the test easier. Since the lifetime isn't known yet,
5722 this requires two loops. See also record_giv above. */
5727 {
5736 {
5738 {
5740 {
5744 }
5745 }
5750 }
5751 }
5753 /* Now that the lifetime of the giv is known, check for branches
5754 from within the lifetime to outside the lifetime if it is still
5755 replaceable. */
5758 {
5761 {
5774 {
5783 }
5784 }
5785 }
5787 /* If it is replaceable, then save the final value. */
5790 }
5795 }
5796 \f
5797 /* Update the status of whether a giv can derive other givs.
5799 We need to do something special if there is or may be an update to the biv
5800 between the time the giv is defined and the time it is used to derive
5801 another giv.
5803 In addition, a giv that is only conditionally set is not allowed to
5804 derive another giv once a label has been passed.
5806 The cases we look at are when a label or an update to a biv is passed. */
5808 static void
5811 rtx p;
5812 {
5815 rtx tem;
5818 /* Search all IV classes, then all bivs, and finally all givs.
5820 There are three cases we are concerned with. First we have the situation
5821 of a giv that is only updated conditionally. In that case, it may not
5822 derive any givs after a label is passed.
5824 The second case is when a biv update occurs, or may occur, after the
5825 definition of a giv. For certain biv updates (see below) that are
5826 known to occur between the giv definition and use, we can adjust the
5827 giv definition. For others, or when the biv update is conditional,
5828 we must prevent the giv from deriving any other givs. There are two
5829 sub-cases within this case.
5831 If this is a label, we are concerned with any biv update that is done
5832 conditionally, since it may be done after the giv is defined followed by
5833 a branch here (actually, we need to pass both a jump and a label, but
5834 this extra tracking doesn't seem worth it).
5836 If this is a jump, we are concerned about any biv update that may be
5837 executed multiple times. We are actually only concerned about
5838 backward jumps, but it is probably not worth performing the test
5839 on the jump again here.
5841 If this is a biv update, we must adjust the giv status to show that a
5842 subsequent biv update was performed. If this adjustment cannot be done,
5843 the giv cannot derive further givs. */
5849 {
5851 {
5852 /* If cant_derive is already true, there is no point in
5853 checking all of these conditions again. */
5857 /* If this giv is conditionally set and we have passed a label,
5858 it cannot derive anything. */
5862 /* Skip givs that have mult_val == 0, since
5863 they are really invariants. Also skip those that are
5864 replaceable, since we know their lifetime doesn't contain
5865 any biv update. */
5869 /* The only way we can allow this giv to derive another
5870 is if this is a biv increment and we can form the product
5871 of biv->add_val and giv->mult_val. In this case, we will
5872 be able to compute a compensation. */
5874 {
5885 tem = simplify_giv_expr
5892 else
5894 }
5898 }
5899 }
5900 }
5901 \f
5902 /* Check whether an insn is an increment legitimate for a basic induction var.
5903 X is the source of insn P, or a part of it.
5904 MODE is the mode in which X should be interpreted.
5906 DEST_REG is the putative biv, also the destination of the insn.
5907 We accept patterns of these forms:
5908 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5909 REG = INVARIANT + REG
5911 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5912 store the additive term into *INC_VAL, and store the place where
5913 we found the additive term into *LOCATION.
5915 If X is an assignment of an invariant into DEST_REG, we set
5916 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5918 We also want to detect a BIV when it corresponds to a variable
5919 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5920 of the variable may be a PLUS that adds a SUBREG of that variable to
5921 an invariant and then sign- or zero-extends the result of the PLUS
5922 into the variable.
5924 Most GIVs in such cases will be in the promoted mode, since that is the
5925 probably the natural computation mode (and almost certainly the mode
5926 used for addresses) on the machine. So we view the pseudo-reg containing
5927 the variable as the BIV, as if it were simply incremented.
5929 Note that treating the entire pseudo as a BIV will result in making
5930 simple increments to any GIVs based on it. However, if the variable
5931 overflows in its declared mode but not its promoted mode, the result will
5932 be incorrect. This is acceptable if the variable is signed, since
5933 overflows in such cases are undefined, but not if it is unsigned, since
5934 those overflows are defined. So we only check for SIGN_EXTEND and
5935 not ZERO_EXTEND.
5937 If we cannot find a biv, we return 0. */
5939 static int
5944 rtx dest_reg;
5945 rtx p;
5949 {
5957 {
5963 {
5965 }
5970 {
5972 }
5973 else
5986 /* If this is a SUBREG for a promoted variable, check the inner
5987 value. */
5995 /* If this register is assigned in a previous insn, look at its
5996 source, but don't go outside the loop or past a label. */
5998 /* If this sets a register to itself, we would repeat any previous
5999 biv increment if we applied this strategy blindly. */
6005 {
6031 }
6032 /* ... fall through ... */
6034 /* Can accept constant setting of biv only when inside inner most loop.
6035 Otherwise, a biv of an inner loop may be incorrectly recognized
6036 as a biv of the outer loop,
6037 causing code to be moved INTO the inner loop. */
6044 /* convert_modes aborts if we try to convert to or from CCmode, so just
6045 exclude that case. It is very unlikely that a condition code value
6046 would be a useful iterator anyways. */
6050 {
6051 /* Possible bug here? Perhaps we don't know the mode of X. */
6055 }
6056 else
6064 /* Similar, since this can be a sign extension. */
6069 ;
6083 location);
6088 }
6089 }
6090 \f
6091 /* A general induction variable (giv) is any quantity that is a linear
6092 function of a basic induction variable,
6093 i.e. giv = biv * mult_val + add_val.
6094 The coefficients can be any loop invariant quantity.
6095 A giv need not be computed directly from the biv;
6096 it can be computed by way of other givs. */
6098 /* Determine whether X computes a giv.
6099 If it does, return a nonzero value
6100 which is the benefit from eliminating the computation of X;
6101 set *SRC_REG to the register of the biv that it is computed from;
6102 set *ADD_VAL and *MULT_VAL to the coefficients,
6103 such that the value of X is biv * mult + add; */
6105 static int
6109 rtx x;
6116 {
6120 /* If this is an invariant, forget it, it isn't a giv. */
6124 /* See if the expression could be a giv and get its form.
6125 Mark our place on the obstack in case we don't find a giv. */
6130 {
6133 }
6136 {
6139 /* Since this is now an invariant and wasn't before, it must be a giv
6140 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6141 with. */
6148 /* This is equivalent to a BIV. */
6155 /* Either (plus (biv) (invar)) or
6156 (plus (mult (biv) (invar_1)) (invar_2)). */
6158 {
6161 }
6162 else
6163 {
6166 }
6171 /* ADD_VAL is zero. */
6179 }
6181 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6182 unless they are CONST_INT). */
6190 else
6193 /* Always return true if this is a giv so it will be detected as such,
6194 even if the benefit is zero or negative. This allows elimination
6195 of bivs that might otherwise not be eliminated. */
6197 }
6198 \f
6199 /* Given an expression, X, try to form it as a linear function of a biv.
6200 We will canonicalize it to be of the form
6201 (plus (mult (BIV) (invar_1))
6202 (invar_2))
6203 with possible degeneracies.
6205 The invariant expressions must each be of a form that can be used as a
6206 machine operand. We surround then with a USE rtx (a hack, but localized
6207 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6208 routine; it is the caller's responsibility to strip them.
6210 If no such canonicalization is possible (i.e., two biv's are used or an
6211 expression that is neither invariant nor a biv or giv), this routine
6212 returns 0.
6214 For a non-zero return, the result will have a code of CONST_INT, USE,
6215 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6217 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6224 static rtx
6227 rtx x;
6229 {
6232 rtx tem;
6234 /* If this is not an integer mode, or if we cannot do arithmetic in this
6235 mode, this can't be a giv. */
6242 {
6249 /* Put constant last, CONST_INT last if both constant. */
6257 /* Handle addition of zero, then addition of an invariant. */
6262 {
6265 /* Adding two invariants must result in an invariant, so enclose
6266 addition operation inside a USE and return it. */
6276 else
6285 /* biv + invar or mult + invar. Return sum. */
6289 /* (a + invar_1) + invar_2. Associate. */
6290 return
6296 arg1)),
6297 benefit);
6301 }
6303 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6304 MULT to reduce cases. */
6310 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6311 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6312 Recurse to associate the second PLUS. */
6317 return
6323 benefit);
6325 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6338 benefit);
6341 /* Handle "a - b" as "a + b * (-1)". */
6347 constm1_rtx)),
6348 benefit);
6356 /* Put constant last, CONST_INT last if both constant. */
6361 /* If second argument is not now constant, not giv. */
6365 /* Handle multiply by 0 or 1. */
6373 {
6375 /* biv * invar. Done. */
6379 /* Product of two constants. */
6383 /* invar * invar is a giv, but attempt to simplify it somehow. */
6389 {
6390 /* (invar_0 * invar_1) * invar_2. Associate. */
6397 arg1)),
6398 benefit);
6399 }
6400 /* Porpagate the MULT expressions to the intermost nodes. */
6402 {
6403 /* (invar_0 + invar_1) * invar_2. Distribute. */
6409 arg1),
6413 arg1)),
6414 benefit);
6415 }
6419 /* (a * invar_1) * invar_2. Associate. */
6425 arg1)),
6426 benefit);
6429 /* (a + invar_1) * invar_2. Distribute. */
6434 arg1),
6437 arg1)),
6438 benefit);
6442 }
6445 /* Shift by constant is multiply by power of two. */
6449 return
6455 benefit);
6458 /* "-a" is "a * (-1)" */
6461 benefit);
6464 /* "~a" is "-a - 1". Silly, but easy. */
6468 const1_rtx),
6469 benefit);
6472 /* Already in proper form for invariant. */
6476 /* If this is a new register, we can't deal with it. */
6480 /* Check for biv or giv. */
6482 {
6486 {
6489 /* Form expression from giv and add benefit. Ensure this giv
6490 can derive another and subtract any needed adjustment if so. */
6502 }
6505 /* If it isn't an induction variable, and it is invariant, we
6506 may be able to simplify things further by looking through
6507 the bits we just moved outside the loop. */
6509 {
6514 {
6515 /* Ok, we found a match. Substitute and simplify. */
6517 /* If we match another movable, we must use that, as
6518 this one is going away. */
6521 benefit);
6523 /* If consec is non-zero, this is a member of a group of
6524 instructions that were moved together. We handle this
6525 case only to the point of seeking to the last insn and
6526 looking for a REG_EQUAL. Fail if we don't find one. */
6528 {
6536 }
6537 else
6538 {
6542 }
6545 {
6546 /* What we are most interested in is pointer
6547 arithmetic on invariants -- only take
6548 patterns we may be able to do something with. */
6554 {
6558 }
6563 {
6565 benefit);
6568 }
6569 }
6571 }
6572 }
6574 }
6576 /* Fall through to general case. */
6578 /* If invariant, return as USE (unless CONST_INT).
6579 Otherwise, not giv. */
6584 {
6593 }
6594 else
6596 }
6597 }
6599 /* This routine folds invariants such that there is only ever one
6600 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6602 static rtx
6605 {
6611 {
6614 }
6617 {
6620 }
6621 else
6622 {
6625 }
6626 }
6628 static rtx
6632 {
6634 {
6638 else
6641 }
6644 else
6647 }
6648 \f
6649 /* Help detect a giv that is calculated by several consecutive insns;
6650 for example,
6651 giv = biv * M
6652 giv = giv + A
6653 The caller has already identified the first insn P as having a giv as dest;
6654 we check that all other insns that set the same register follow
6655 immediately after P, that they alter nothing else,
6656 and that the result of the last is still a giv.
6658 The value is 0 if the reg set in P is not really a giv.
6659 Otherwise, the value is the amount gained by eliminating
6660 all the consecutive insns that compute the value.
6662 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6663 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6665 The coefficients of the ultimate giv value are stored in
6666 *MULT_VAL and *ADD_VAL. */
6668 static int
6673 rtx p;
6674 rtx src_reg;
6675 rtx dest_reg;
6679 {
6683 rtx temp;
6684 rtx set;
6686 /* Indicate that this is a giv so that we can update the value produced in
6687 each insn of the multi-insn sequence.
6689 This induction structure will be used only by the call to
6690 general_induction_var below, so we can allocate it on our stack.
6691 If this is a giv, our caller will replace the induct var entry with
6692 a new induction structure. */
6708 {
6712 /* If libcall, skip to end of call sequence. */
6722 /* Giv created by equivalent expression. */
6726 VOIDmode)))
6728 {
6732 count--;
6736 }
6738 {
6739 /* Allow insns that set something other than this giv to a
6740 constant. Such insns are needed on machines which cannot
6741 include long constants and should not disqualify a giv. */
6750 }
6751 }
6755 }
6756 \f
6757 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6758 represented by G1. If no such expression can be found, or it is clear that
6759 it cannot possibly be a valid address, 0 is returned.
6761 To perform the computation, we note that
6762 G1 = x * v + a and
6763 G2 = y * v + b
6764 where `v' is the biv.
6766 So G2 = (y/b) * G1 + (b - a*y/x).
6768 Note that MULT = y/x.
6770 Update: A and B are now allowed to be additive expressions such that
6771 B contains all variables in A. That is, computing B-A will not require
6772 subtracting variables. */
6774 static rtx
6777 {
6778 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6783 /* If MULT is not 1, we cannot handle A with non-constants, since we
6784 would then be required to subtract multiples of the registers in A.
6785 This is theoretically possible, and may even apply to some Fortran
6786 constructs, but it is a lot of work and we do not attempt it here. */
6791 /* In general these structures are sorted top to bottom (down the PLUS
6792 chain), but not left to right across the PLUS. If B is a higher
6793 order giv than A, we can strip one level and recurse. If A is higher
6794 order, we'll eventually bail out, but won't know that until the end.
6795 If they are the same, we'll strip one level around this loop. */
6798 {
6810 /* We matched: remove one reg completely. */
6813 /* An alternate match. */
6816 /* An alternate match. */
6818 else
6819 {
6820 /* Indicates an extra register in B. Strip one level from B and
6821 recurse, hoping B was the higher order expression. */
6826 }
6827 }
6829 /* Here we are at the last level of A, go through the cases hoping to
6830 get rid of everything but a constant. */
6833 {
6846 }
6848 {
6850 }
6852 {
6854 }
6856 {
6861 else
6863 }
6868 }
6870 rtx
6873 {
6876 /* The value that G1 will be multiplied by must be a constant integer. Also,
6877 the only chance we have of getting a valid address is if b*c/a (see above
6878 for notation) is also an integer. */
6881 {
6886 }
6889 else
6890 {
6891 /* ??? Find out if the one is a multiple of the other? */
6893 }
6897 {
6898 /* Failed. If we've got a multiplication factor between G1 and G2,
6899 scale G1's addend and try again. */
6901 {
6905 {
6906 HOST_WIDE_INT m;
6910 }
6911 else
6912 {
6914 mult);
6915 }
6918 }
6919 }
6923 /* Form simplified final result. */
6928 else
6933 else
6934 {
6937 {
6941 }
6944 }
6946 }
6947 \f
6948 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6949 represented by G1. This indicates that G2 should be combined with G1 and
6950 that G2 can use (either directly or via an address expression) a register
6951 used to represent G1. */
6953 static rtx
6956 {
6959 /* If these givs are identical, they can be combined. We use the results
6960 of express_from because the addends are not in a canonical form, so
6961 rtx_equal_p is a weaker test. */
6962 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6963 combination to be the other way round. */
6966 {
6968 }
6970 /* If G2 can be expressed as a function of G1 and that function is valid
6971 as an address and no more expensive than using a register for G2,
6972 the expression of G2 in terms of G1 can be used. */
6976 /* ??? Looses, especially with -fforce-addr, where *g2->location
6977 will always be a register, and so anything more complicated
6978 gets discarded. */
6979 #if 0
6980 #ifdef ADDRESS_COST
6982 #else
6984 #endif
6985 #endif
6986 )
6987 {
6989 }
6992 }
6993 \f
6994 struct combine_givs_stats
6995 {
6998 };
7000 static int
7004 {
7011 /* Stabilize the sort. */
7015 }
7017 /* Check all pairs of givs for iv_class BL and see if any can be combined with
7018 any other. If so, point SAME to the giv combined with and set NEW_REG to
7019 be an expression (in terms of the other giv's DEST_REG) equivalent to the
7020 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
7022 static void
7025 {
7026 /* Additional benefit to add for being combined multiple times. */
7034 /* Count givs, because bl->giv_count is incorrect here. */
7038 giv_count++;
7040 giv_array
7051 {
7053 rtx single_use;
7058 /* If a DEST_REG GIV is used only once, do not allow it to combine
7059 with anything, for in doing so we will gain nothing that cannot
7060 be had by simply letting the GIV with which we would have combined
7061 to be reduced on its own. The losage shows up in particular with
7062 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7063 be seen elsewhere as well. */
7070 /* Add an additional weight for zero addends. */
7075 {
7076 rtx this_combine;
7081 {
7084 }
7085 }
7087 }
7089 /* Iterate, combining until we can't. */
7090 restart:
7094 {
7097 {
7103 }
7105 }
7108 {
7114 /* If it has already been combined, skip. */
7119 {
7122 /* If it has already been combined, skip. */
7124 {
7134 /* ??? The new final_[bg]iv_value code does a much better job
7135 of finding replaceable giv's, and hence this code may no
7136 longer be necessary. */
7140 /* To help optimize the next set of combinations, remove
7141 this giv from the benefits of other potential mates. */
7143 {
7147 }
7153 }
7154 }
7156 /* To help optimize the next set of combinations, remove
7157 this giv from the benefits of other potential mates. */
7159 {
7161 {
7165 }
7169 /* We've finished with this giv, and everything it touched.
7170 Restart the combination so that proper weights for the
7171 rest of the givs are properly taken into account. */
7172 /* ??? Ideally we would compact the arrays at this point, so
7173 as to not cover old ground. But sanely compacting
7174 can_combine is tricky. */
7176 }
7177 }
7179 /* Clean up. */
7182 }
7183 \f
7184 struct recombine_givs_stats
7185 {
7188 };
7190 /* Used below as comparison function for qsort. We want a ascending luid
7191 when scanning the array starting at the end, thus the arguments are
7192 used in reverse. */
7193 static int
7197 {
7204 /* Stabilize the sort. */
7208 }
7210 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7211 look for the start of life of a giv where the start has not been seen
7212 yet to unlock the search for the end of its life.
7213 Only consider givs that belong to BIV.
7214 Return the total number of lifetime ends that have been found. */
7215 static int
7219 {
7227 {
7229 {
7232 {
7240 {
7241 /* If we see a 0 here for end_luid, it means that we have
7242 scanned the entire loop without finding any use at all.
7243 We must not predicate this code on a start_luid match
7244 since that would make the test fail for givs that have
7245 been hoisted out of inner loops. */
7247 {
7250 }
7253 }
7255 }
7257 }
7259 {
7267 {
7272 }
7274 }
7282 }
7286 {
7293 }
7295 }
7297 /* For each giv that has been combined with another, look if
7298 we can combine it with the most recently used one instead.
7299 This tends to shorten giv lifetimes, and helps the next step:
7300 try to derive givs from other givs. */
7301 static void
7306 {
7314 {
7316 giv_count++;
7317 }
7318 giv_array
7322 /* Initialize stats and set up the ix field for each giv in stats to name
7323 the corresponding index into stats. */
7325 {
7326 rtx p;
7332 /* If this giv has been hoisted out of an inner loop, use the luid of
7333 the previous insn. */
7337 i++;
7338 }
7342 /* Set up the ix field for each giv in stats to name
7343 the corresponding index into stats, and
7344 do the actual most-recently-used recombination. */
7346 {
7350 {
7352 rtx new_combine;
7354 /* combine_givs_p actually says if we can make this transformation.
7355 The other tests are here only to avoid keeping a giv alive
7356 that could otherwise be eliminated. */
7362 {
7367 /* No need to update lifetimes / benefits here since we have
7368 already decided what to reduce. */
7371 {
7377 }
7379 }
7381 }
7389 }
7392 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7393 lifetime ends from regscan info. */
7395 {
7400 {
7401 /* Loop unrolling of an inner loop can even create new DEST_REG
7402 givs. */
7403 rtx p;
7409 }
7411 {
7413 {
7416 }
7418 {
7419 rtx p;
7420 /* This insn has been created by loop optimization on an inner
7421 loop. We don't have a proper start_luid that will match
7422 when we see the first set. But we do know that there will
7423 be no use before the set, so we can set end_luid to 0 so that
7424 we'll start looking for the last use right away. */
7429 ends_need_computing++;
7430 }
7431 else
7432 {
7437 /* Find the first insn that sets the giv, so that we can verify
7438 if this giv's lifetime wraps around the loop. We also need
7439 the luid of the first setting insn in order to detect the
7440 last use properly. */
7442 {
7445 count--;
7446 }
7450 {
7452 ends_need_computing++;
7453 }
7454 else
7455 {
7458 {
7460 ends_need_computing++;
7461 }
7462 }
7463 }
7464 }
7465 }
7467 /* If the regscan information was unconclusive for one or more DEST_REG
7468 givs, scan the all insn in the loop to find out lifetime ends. */
7470 {
7474 do
7475 {
7482 }
7484 }
7486 /* Set start_luid back to the last insn that sets the giv. This allows
7487 more combinations. */
7489 {
7495 }
7497 /* Now adjust lifetime ends by taking combined givs into account. */
7499 {
7507 {
7510 /* Use unsigned arithmetic to model loop wrap-around. */
7514 }
7515 }
7519 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7520 same mult_val and non-overlapping lifetime. This reduces register
7521 pressure.
7522 Once we find a DEST_REG giv that is suitable to derive others from,
7523 we set last_giv to this giv, and try to derive as many other DEST_REG
7524 givs from it without joining overlapping lifetimes. If we then
7525 encounter a DEST_REG giv that we can't derive, we set rescan to the
7526 index for this giv (unless rescan is already set).
7527 When we are finished with the current LAST_GIV (i.e. the inner loop
7528 terminates), we start again with rescan, which then becomes the new
7529 LAST_GIV. */
7531 {
7535 {
7536 rtx sum;
7542 {
7543 /* Don't use a giv that's likely to be dead to derive
7544 others - that would be likely to keep that giv alive. */
7546 {
7550 }
7552 }
7553 /* Use unsigned arithmetic to model loop wrap around. */
7558 /* Check that the giv insn we're about to use for deriving
7559 precedes all uses of that giv. Note that initializing the
7560 derived giv would defeat the purpose of reducing register
7561 pressure.
7562 ??? We could arrange to move the insn. */
7566 /* ??? Could handle libcalls, but would need more logic. */
7568 /* We would really like to know if for any giv that v
7569 is combined with, v->insn or any intervening biv increment
7570 dominates that combined giv. However, we
7571 don't have this detailed control flow information.
7572 N.B. since last_giv will be reduced, it is valid
7573 anywhere in the loop, so we don't need to check the
7574 validity of last_giv.
7575 We rely here on the fact that v->always_executed implies that
7576 there is no jump to someplace else in the loop before the
7577 giv insn, and hence any insn that is executed before the
7578 giv insn in the loop will have a lower luid. */
7581 /* Make sure we don't make the add more expensive. ADD_COST
7582 doesn't take different costs of registers and constants into
7583 account, so compare the cost of the actual SET_SRCs. */
7586 /* ??? unroll can't understand anything but reg + const_int
7587 sums. It would be cleaner to fix unroll. */
7594 {
7599 {
7605 }
7606 }
7609 }
7610 }
7612 /* Clean up. */
7615 }
7616 \f
7617 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7619 void
7625 rtx insert_before;
7626 {
7627 rtx seq;
7628 rtx result;
7630 /* Prevent unexpected sharing of these rtx. */
7634 /* Increase the lifetime of any invariants moved further in code. */
7648 /* It is entirely possible that the expansion created lots of new
7649 registers. Iterate over the sequence we just created and
7650 record them all. */
7653 {
7656 {
7660 }
7661 }
7665 }
7666 \f
7667 /* Test whether A * B can be computed without
7668 an actual multiply insn. Value is 1 if so. */
7670 static int
7672 rtx a;
7673 rtx b;
7674 {
7676 rtx tmp;
7681 /* If only one is constant, make it B. */
7685 /* If first constant, both constant, so don't need multiply. */
7689 /* If second not constant, neither is constant, so would need multiply. */
7693 /* One operand is constant, so might not need multiply insn. Generate the
7694 code for the multiply and see if a call or multiply, or long sequence
7695 of insns is generated. */
7704 {
7709 else
7711 {
7720 {
7723 }
7724 }
7725 }
7734 /* Free any storage we obtained in generating this multiply and restore rtl
7735 allocation to its normal obstack. */
7740 }
7741 \f
7742 /* Check to see if loop can be terminated by a "decrement and branch until
7743 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7744 Also try reversing an increment loop to a decrement loop
7745 to see if the optimization can be performed.
7746 Value is nonzero if optimization was performed. */
7748 /* This is useful even if the architecture doesn't have such an insn,
7749 because it might change a loops which increments from 0 to n to a loop
7750 which decrements from n to 0. A loop that decrements to zero is usually
7751 faster than one that increments from zero. */
7753 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7754 such as approx_final_value, biv_total_increment, loop_iterations, and
7755 final_[bg]iv_value. */
7757 static int
7761 {
7763 rtx reg;
7764 rtx jump_label;
7765 rtx final_value;
7766 rtx start_value;
7767 rtx new_add_val;
7768 rtx comparison;
7769 rtx before_comparison;
7770 rtx p;
7771 rtx jump;
7772 rtx first_compare;
7778 /* If last insn is a conditional branch, and the insn before tests a
7779 register value, try to optimize it. Otherwise, we can't do anything. */
7788 /* Try to compute whether the compare/branch at the loop end is one or
7789 two instructions. */
7795 else
7798 /* Check all of the bivs to see if the compare uses one of them.
7799 Skip biv's set more than once because we can't guarantee that
7800 it will be zero on the last iteration. Also skip if the biv is
7801 used between its update and the test insn. */
7804 {
7809 first_compare))
7811 }
7816 /* Look for the case where the basic induction variable is always
7817 nonnegative, and equals zero on the last iteration.
7818 In this case, add a reg_note REG_NONNEG, which allows the
7819 m68k DBRA instruction to be used. */
7827 {
7828 /* Initial value must be greater than 0,
7829 init_val % -dec_value == 0 to ensure that it equals zero on
7830 the last iteration */
7836 {
7837 /* register always nonnegative, add REG_NOTE to branch */
7845 }
7847 /* If the decrement is 1 and the value was tested as >= 0 before
7848 the loop, then we can safely optimize. */
7850 {
7863 {
7871 }
7872 }
7873 }
7876 {
7877 /* Try to change inc to dec, so can apply above optimization. */
7878 /* Can do this if:
7879 all registers modified are induction variables or invariant,
7880 all memory references have non-overlapping addresses
7881 (obviously true if only one write)
7882 allow 2 insns for the compare/jump at the end of the loop. */
7883 /* Also, we must avoid any instructions which use both the reversed
7884 biv and another biv. Such instructions will fail if the loop is
7885 reversed. We meet this condition by requiring that either
7886 no_use_except_counting is true, or else that there is only
7887 one biv. */
7889 /* 1 if the iteration var is used only to count iterations. */
7891 /* 1 if the loop has no memory store, or it has a single memory store
7892 which is reversible. */
7896 {
7899 /* If there are no givs for this biv, and the only exit is the
7900 fall through at the end of the loop, then
7901 see if perhaps there are no uses except to count. */
7905 {
7910 /* An insn that sets the biv is okay. */
7911 ;
7915 {
7916 /* If either of these insns uses the biv and sets a pseudo
7917 that has more than one usage, then the biv has uses
7918 other than counting since it's used to derive a value
7919 that is used more than one time. */
7924 {
7927 }
7928 }
7930 {
7933 }
7934 }
7935 }
7940 {
7945 /* If the loop has a single store, and the destination address is
7946 invariant, then we can't reverse the loop, because this address
7947 might then have the wrong value at loop exit.
7948 This would work if the source was invariant also, however, in that
7949 case, the insn should have been moved out of the loop. */
7952 {
7955 reversible_mem_store
7956 = (! unknown_address_altered
7957 && ! unknown_constant_address_altered
7962 /* If the store depends on a register that is set after the
7963 store, it depends on the initial value, and is thus not
7964 reversible. */
7966 {
7972 }
7973 }
7974 }
7975 else
7978 /* This code only acts for innermost loops. Also it simplifies
7979 the memory address check by only reversing loops with
7980 zero or one memory access.
7981 Two memory accesses could involve parts of the same array,
7982 and that can't be reversed.
7983 If the biv is used only for counting, than we don't need to worry
7984 about all these things. */
7989 && reversible_mem_store
7994 {
7995 rtx tem;
7997 /* Loop can be reversed. */
8001 /* Now check other conditions:
8003 The increment must be a constant, as must the initial value,
8004 and the comparison code must be LT.
8006 This test can probably be improved since +/- 1 in the constant
8007 can be obtained by changing LT to LE and vice versa; this is
8008 confusing. */
8011 /* for constants, LE gets turned into LT */
8015 {
8026 comparison_const_width
8028 else
8029 comparison_const_width
8033 comparison_sign_mask
8036 /* If the comparison value is not a loop invariant, then we
8037 can not reverse this loop.
8039 ??? If the insns which initialize the comparison value as
8040 a whole compute an invariant result, then we could move
8041 them out of the loop and proceed with loop reversal. */
8049 /* Normalize the initial value if it is an integer and
8050 has no other use except as a counter. This will allow
8051 a few more loops to be reversed. */
8055 {
8057 /* The code below requires comparison_val to be a multiple
8058 of add_val in order to do the loop reversal, so
8059 round up comparison_val to a multiple of add_val.
8060 Since comparison_value is constant, we know that the
8061 current comparison code is LT. */
8063 comparison_val
8065 /* We postpone overflow checks for COMPARISON_VAL here;
8066 even if there is an overflow, we might still be able to
8067 reverse the loop, if converting the loop exit test to
8068 NE is possible. */
8070 }
8072 /* First check if we can do a vanilla loop reversal. */
8074 /* If we have a decrement_and_branch_on_count,
8075 prefer the NE test, since this will allow that
8076 instruction to be generated. Note that we must
8077 use a vanilla loop reversal if the biv is used to
8078 calculate a giv or has a non-counting use. */
8079 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8080 && defined (HAVE_decrement_and_branch_on_count)
8084 #endif
8086 /* Now do postponed overflow checks on COMPARISON_VAL. */
8089 {
8090 /* Register will always be nonnegative, with value
8091 0 on last iteration */
8095 }
8099 {
8102 }
8103 else
8109 /* If the initial value is not zero, or if the comparison
8110 value is not an exact multiple of the increment, then we
8111 can not reverse this loop. */
8114 {
8117 }
8118 else
8119 {
8122 }
8126 /* Reset these in case we normalized the initial value
8127 and comparison value above. */
8130 {
8132 final_value
8134 }
8137 /* Save some info needed to produce the new insns. */
8144 /* Set start_value; if this is not a CONST_INT, we need
8145 to generate a SUB.
8146 Initialize biv to start_value before loop start.
8147 The old initializing insn will be deleted as a
8148 dead store by flow.c. */
8151 {
8154 loop_start);
8155 }
8157 {
8160 enum insn_code icode
8169 start_value
8173 loop_start);
8177 }
8179 {
8181 enum insn_code icode
8189 start_value
8193 loop_start);
8194 }
8195 else
8196 /* We could handle the other cases too, but it'll be
8197 better to have a testcase first. */
8200 /* We may not have a single insn which can increment a reg, so
8201 create a sequence to hold all the insns from expand_inc. */
8210 /* Update biv info to reflect its new status. */
8215 /* Update loop info. */
8224 /* Inc LABEL_NUSES so that delete_insn will
8225 not delete the label. */
8228 /* Emit an insn after the end of the loop to set the biv's
8229 proper exit value if it is used anywhere outside the loop. */
8234 loop_end);
8236 /* Delete compare/branch at end of loop. */
8241 /* Add new compare/branch insn at end of loop. */
8253 ;
8259 {
8261 {
8262 /* Increment of LABEL_NUSES done above. */
8263 /* Register is now always nonnegative,
8264 so add REG_NONNEG note to the branch. */
8267 }
8269 }
8271 /* No insn may reference both the reversed and another biv or it
8272 will fail (see comment near the top of the loop reversal
8273 code).
8274 Earlier on, we have verified that the biv has no use except
8275 counting, or it is the only biv in this function.
8276 However, the code that computes no_use_except_counting does
8277 not verify reg notes. It's possible to have an insn that
8278 references another biv, and has a REG_EQUAL note with an
8279 expression based on the reversed biv. To avoid this case,
8280 remove all REG_EQUAL notes based on the reversed biv
8281 here. */
8284 {
8287 /* If this is a set of a GIV based on the reversed biv, any
8288 REG_EQUAL notes should still be correct. */
8295 {
8300 else
8302 }
8303 }
8305 /* Mark that this biv has been reversed. Each giv which depends
8306 on this biv, and which is also live past the end of the loop
8307 will have to be fixed up. */
8312 {
8316 else
8318 }
8321 }
8322 }
8323 }
8326 }
8327 \f
8328 /* Verify whether the biv BL appears to be eliminable,
8329 based on the insns in the loop that refer to it.
8331 If ELIMINATE_P is non-zero, actually do the elimination.
8333 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8334 determine whether invariant insns should be placed inside or at the
8335 start of the loop. */
8337 static int
8343 {
8347 rtx p;
8349 /* Scan all insns in the loop, stopping if we find one that uses the
8350 biv in a way that we cannot eliminate. */
8353 {
8357 /* If this is a libcall that sets a giv, skip ahead to its end. */
8359 {
8363 {
8368 {
8375 }
8376 }
8377 }
8382 {
8388 }
8389 }
8392 {
8397 }
8400 }
8401 \f
8402 /* INSN and REFERENCE are instructions in the same insn chain.
8403 Return non-zero if INSN is first. */
8405 int
8408 {
8412 {
8413 /* Start with test for not first so that INSN == REFERENCE yields not
8414 first. */
8420 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8421 previous insn, hence the <= comparison below does not work if
8422 P is a note. */
8433 }
8434 }
8436 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8437 the offset that we have to take into account due to auto-increment /
8438 div derivation is zero. */
8439 static int
8442 rtx insn;
8443 {
8444 /* If the giv V had the auto-inc address optimization applied
8445 to it, and INSN occurs between the giv insn and the biv
8446 insn, then we'd have to adjust the value used here.
8447 This is rare, so we don't bother to make this possible. */
8455 /* If the giv V was derived from another giv, and INSN does
8456 not occur between the giv insn and the biv insn, then we'd
8457 have to adjust the value used here. This is rare, so we don't
8458 bother to make this possible. */
8472 }
8474 /* If BL appears in X (part of the pattern of INSN), see if we can
8475 eliminate its use. If so, return 1. If not, return 0.
8477 If BIV does not appear in X, return 1.
8479 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8480 where extra insns should be added. Depending on how many items have been
8481 moved out of the loop, it will either be before INSN or at the start of
8482 the loop. */
8484 static int
8490 rtx where;
8491 {
8497 #ifdef HAVE_cc0
8499 #endif
8505 {
8507 /* If we haven't already been able to do something with this BIV,
8508 we can't eliminate it. */
8514 /* If this sets the BIV, it is not a problem. */
8518 /* If this is an insn that defines a giv, it is also ok because
8519 it will go away when the giv is reduced. */
8524 #ifdef HAVE_cc0
8526 {
8527 /* Can replace with any giv that was reduced and
8528 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8529 Require a constant for MULT_VAL, so we know it's nonzero.
8530 ??? We disable this optimization to avoid potential
8531 overflows. */
8539 {
8546 /* If the giv has the opposite direction of change,
8547 then reverse the comparison. */
8551 else
8554 /* We can probably test that giv's reduced reg. */
8557 }
8559 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8560 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8561 Require a constant for MULT_VAL, so we know it's nonzero.
8562 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8563 overflow problem. */
8574 {
8581 /* If the giv has the opposite direction of change,
8582 then reverse the comparison. */
8586 else
8590 /* Replace biv with the giv's reduced register. */
8595 /* Insn doesn't support that constant or invariant. Copy it
8596 into a register (it will be a loop invariant.) */
8600 where);
8602 /* Substitute the new register for its invariant value in
8603 the compare expression. */
8607 }
8608 }
8609 #endif
8616 /* See if either argument is the biv. */
8621 else
8625 {
8626 /* First try to replace with any giv that has constant positive
8627 mult_val and constant add_val. We might be able to support
8628 negative mult_val, but it seems complex to do it in general. */
8639 {
8646 /* Replace biv with the giv's reduced reg. */
8649 /* If all constants are actually constant integers and
8650 the derived constant can be directly placed in the COMPARE,
8651 do so. */
8655 {
8660 }
8661 else
8662 {
8663 /* Otherwise, load it into a register. */
8667 }
8670 }
8672 /* Look for giv with positive constant mult_val and nonconst add_val.
8673 Insert insns to calculate new compare value.
8674 ??? Turn this off due to possible overflow. */
8681 {
8682 rtx tem;
8692 /* Replace biv with giv's reduced register. */
8696 /* Compute value to compare against. */
8698 /* Use it in this insn. */
8702 }
8703 }
8705 {
8707 {
8708 /* Look for giv with constant positive mult_val and nonconst
8709 add_val. Insert insns to compute new compare value.
8710 ??? Turn this off due to possible overflow. */
8717 {
8718 rtx tem;
8728 /* Replace biv with giv's reduced register. */
8732 /* Compute value to compare against. */
8738 }
8739 }
8741 /* This code has problems. Basically, you can't know when
8742 seeing if we will eliminate BL, whether a particular giv
8743 of ARG will be reduced. If it isn't going to be reduced,
8744 we can't eliminate BL. We can try forcing it to be reduced,
8745 but that can generate poor code.
8747 The problem is that the benefit of reducing TV, below should
8748 be increased if BL can actually be eliminated, but this means
8749 we might have to do a topological sort of the order in which
8750 we try to process biv. It doesn't seem worthwhile to do
8751 this sort of thing now. */
8753 #if 0
8754 /* Otherwise the reg compared with had better be a biv. */
8759 /* Look for a pair of givs, one for each biv,
8760 with identical coefficients. */
8762 {
8773 {
8780 /* Replace biv with its giv's reduced reg. */
8782 /* Replace other operand with the other giv's
8783 reduced reg. */
8786 }
8787 }
8788 #endif
8789 }
8791 /* If we get here, the biv can't be eliminated. */
8795 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8796 biv is used in it, since it will be replaced. */
8804 }
8806 /* See if any subexpression fails elimination. */
8809 {
8811 {
8824 }
8825 }
8828 }
8829 \f
8830 /* Return nonzero if the last use of REG
8831 is in an insn following INSN in the same basic block. */
8833 static int
8835 rtx reg;
8836 rtx insn;
8837 {
8838 rtx n;
8842 {
8845 }
8847 }
8848 \f
8849 /* Called via `note_stores' to record the initial value of a biv. Here we
8850 just record the location of the set and process it later. */
8852 static void
8854 rtx dest;
8855 rtx set;
8857 {
8867 /* If this is the first set found, record it. */
8869 {
8872 }
8873 }
8874 \f
8875 /* If any of the registers in X are "old" and currently have a last use earlier
8876 than INSN, update them to have a last use of INSN. Their actual last use
8877 will be the previous insn but it will not have a valid uid_luid so we can't
8878 use it. */
8880 static void
8882 rtx x;
8883 rtx insn;
8884 {
8885 /* Check for the case where INSN does not have a valid luid. In this case,
8886 there is no need to modify the regno_last_uid, as this can only happen
8887 when code is inserted after the loop_end to set a pseudo's final value,
8888 and hence this insn will never be the last use of x. */
8893 else
8894 {
8898 {
8904 }
8905 }
8906 }
8907 \f
8908 /* Given an insn INSN and condition COND, return the condition in a
8909 canonical form to simplify testing by callers. Specifically:
8911 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8912 (2) Both operands will be machine operands; (cc0) will have been replaced.
8913 (3) If an operand is a constant, it will be the second operand.
8914 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8915 for GE, GEU, and LEU.
8917 If the condition cannot be understood, or is an inequality floating-point
8918 comparison which needs to be reversed, 0 will be returned.
8920 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
8922 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8923 insn used in locating the condition was found. If a replacement test
8924 of the condition is desired, it should be placed in front of that
8925 insn and we will be sure that the inputs are still valid.
8927 If WANT_REG is non-zero, we wish the condition to be relative to that
8928 register, if possible. Therefore, do not canonicalize the condition
8929 further. */
8931 rtx
8933 rtx insn;
8934 rtx cond;
8937 rtx want_reg;
8938 {
8941 rtx set;
8942 rtx tem;
8954 {
8957 }
8962 /* If we are comparing a register with zero, see if the register is set
8963 in the previous insn to a COMPARE or a comparison operation. Perform
8964 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8965 in cse.c */
8970 {
8971 /* Set non-zero when we find something of interest. */
8974 #ifdef HAVE_cc0
8975 /* If comparison with cc0, import actual comparison from compare
8976 insn. */
8978 {
8989 }
8990 #endif
8992 /* If this is a COMPARE, pick up the two things being compared. */
8994 {
8998 }
9002 /* Go back to the previous insn. Stop if it is not an INSN. We also
9003 stop if it isn't a single set or if it has a REG_INC note because
9004 we don't want to bother dealing with it. */
9012 /* If this is setting OP0, get what it sets it to if it looks
9013 relevant. */
9015 {
9018 /* ??? We may not combine comparisons done in a CCmode with
9019 comparisons not done in a CCmode. This is to aid targets
9020 like Alpha that have an IEEE compliant EQ instruction, and
9021 a non-IEEE compliant BEQ instruction. The use of CCmode is
9022 actually artificial, simply to prevent the combination, but
9023 should not affect other platforms.
9025 However, we must allow VOIDmode comparisons to match either
9026 CCmode or non-CCmode comparison, because some ports have
9027 modeless comparisons inside branch patterns.
9029 ??? This mode check should perhaps look more like the mode check
9030 in simplify_comparison in combine. */
9038 && (STORE_FLAG_VALUE
9041 #ifdef FLOAT_STORE_FLAG_VALUE
9044 && (REAL_VALUE_NEGATIVE
9046 #endif
9047 ))
9058 && (STORE_FLAG_VALUE
9061 #ifdef FLOAT_STORE_FLAG_VALUE
9064 && (REAL_VALUE_NEGATIVE
9066 #endif
9067 ))
9073 {
9074 /* We might have reversed a LT to get a GE here. But this wasn't
9075 actually the comparison of data, so we don't flag that we
9076 have had to reverse the condition. */
9080 }
9081 else
9083 }
9086 /* If this sets OP0, but not directly, we have to give up. */
9090 {
9094 {
9100 }
9105 }
9106 }
9108 /* If constant is first, put it last. */
9112 /* If OP0 is the result of a comparison, we weren't able to find what
9113 was really being compared, so fail. */
9117 /* Canonicalize any ordered comparison with integers involving equality
9118 if we can do computations in the relevant mode and we do not
9119 overflow. */
9124 {
9127 unsigned HOST_WIDE_INT max_val
9131 {
9137 /* When cross-compiling, const_val might be sign-extended from
9138 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9158 }
9159 }
9161 /* If this was floating-point and we reversed anything other than an
9162 EQ or NE or (UN)ORDERED, return zero. */
9164 && did_reverse_condition
9166 && ! flag_fast_math
9170 #ifdef HAVE_cc0
9171 /* Never return CC0; return zero instead. */
9174 #endif
9177 }
9179 /* Given a jump insn JUMP, return the condition that will cause it to branch
9180 to its JUMP_LABEL. If the condition cannot be understood, or is an
9181 inequality floating-point comparison which needs to be reversed, 0 will
9182 be returned.
9184 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9185 insn used in locating the condition was found. If a replacement test
9186 of the condition is desired, it should be placed in front of that
9187 insn and we will be sure that the inputs are still valid. */
9189 rtx
9191 rtx jump;
9193 {
9194 rtx cond;
9196 rtx set;
9198 /* If this is not a standard conditional jump, we can't parse it. */
9206 /* If this branches to JUMP_LABEL when the condition is false, reverse
9207 the condition. */
9208 reverse
9213 }
9215 /* Similar to above routine, except that we also put an invariant last
9216 unless both operands are invariants. */
9218 rtx
9221 rtx x;
9222 {
9232 }
9234 #ifdef HAVE_decrement_and_branch_on_count
9235 /* Instrument loop for insertion of bct instruction. We distinguish between
9236 loops with compile-time bounds and those with run-time bounds.
9237 Information from loop_iterations() is used to compute compile-time bounds.
9238 Run-time bounds should use loop preconditioning, but currently ignored.
9239 */
9241 static void
9244 {
9251 #if 0
9253 /* If the loop condition is <= or >=, the number of iteration
9254 is 1 more than the range of the bounds of the loop. */
9257 #endif
9259 /* It's impossible to instrument a competely unrolled loop. */
9263 /* Make sure that the count register is not in use. */
9265 {
9269 loop_num);
9271 }
9273 /* Make sure that the function has no indirect jumps. */
9275 {
9279 loop_num);
9281 }
9283 /* Make sure that the last loop insn is a conditional jump. */
9287 {
9291 loop_num);
9293 }
9295 /* Make sure that the loop does not contain a function call
9296 (the count register might be altered by the called function). */
9298 {
9302 loop_num);
9304 }
9306 /* Make sure that the loop does not jump via a table.
9307 (the count register might be used to perform the branch on table). */
9309 {
9313 loop_num);
9315 }
9317 /* Account for loop unrolling in instrumented iteration count. */
9320 else
9324 {
9325 /* Allow an enclosing outer loop to benefit if possible. */
9329 loop_num);
9331 }
9333 /* Try to instrument the loop. */
9335 /* Handle the simpler case, where the bounds are known at compile time. */
9337 {
9341 /* Mark all enclosing loops that they cannot use count register. */
9343 {
9346 }
9349 }
9351 /* Handle the more complex case, that the bounds are NOT known
9352 at compile time. In this case we generate run_time calculation
9353 of the number of iterations. */
9356 {
9360 loop_num);
9362 }
9366 {
9370 loop_num);
9372 }
9374 /* With runtime bounds, if the compare is of the form '!=' we give up */
9376 {
9380 loop_num);
9382 }
9383 /* Use common loop preconditioning code instead. */
9384 #if 0
9385 else
9386 {
9387 /* We rely on the existence of run-time guard to ensure that the
9388 loop executes at least once. */
9389 rtx sequence;
9390 rtx iterations_num_reg;
9392 unsigned HOST_WIDE_INT increment_value_abs
9395 /* make sure that the increment is a power of two, otherwise (an
9396 expensive) divide is needed. */
9398 {
9403 }
9405 /* compute the number of iterations */
9407 {
9408 rtx temp_reg;
9410 /* Again, the number of iterations is calculated by:
9411 ;
9412 ; compare-val - initial-val + (increment -1) + additional-iteration
9413 ; num_iterations = -----------------------------------------------------------------
9414 ; increment
9415 */
9416 /* ??? Do we have to call copy_rtx here before passing rtx to
9417 expand_binop? */
9419 {
9420 /* <, <= :the loop variable is increasing */
9424 }
9425 else
9426 {
9430 }
9440 temp_reg,
9443 else
9445 }
9450 }
9454 }
9456 /* Instrument loop by inserting a bct in it as follows:
9457 1. A new counter register is created.
9458 2. In the head of the loop the new variable is initialized to the value
9459 passed in the loop_num_iterations parameter.
9460 3. At the end of the loop, comparison of the register with 0 is generated.
9461 The created comparison follows the pattern defined for the
9462 decrement_and_branch_on_count insn, so this insn will be generated.
9463 4. The branch on the old variable are deleted. The compare must remain
9464 because it might be used elsewhere. If the loop-variable or condition
9465 register are used elsewhere, they will be eliminated by flow. */
9467 static void
9470 rtx loop_num_iterations;
9471 {
9472 rtx counter_reg;
9473 rtx start_label;
9474 rtx sequence;
9477 {
9479 {
9484 else
9487 }
9489 /* Discard original jump to continue loop. Original compare result
9490 may still be live, so it cannot be discarded explicitly. */
9493 /* Insert the label which will delimit the start of the loop. */
9497 /* Insert initialization of the count register into the loop header. */
9505 /* Insert new comparison on the count register instead of the
9506 old one, generating the needed BCT pattern (that will be
9507 later recognized by assembly generation phase). */
9510 loop_end);
9516 }
9517 }
9520 /* Scan the function and determine whether it has indirect (computed) jumps.
9522 This is taken mostly from flow.c; similar code exists elsewhere
9523 in the compiler. It may be useful to put this into rtlanal.c. */
9524 static int
9526 rtx start;
9527 {
9528 rtx insn;
9535 }
9537 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9538 documentation for LOOP_MEMS for the definition of `appropriate'.
9539 This function is called from prescan_loop via for_each_rtx. */
9541 static int
9545 {
9553 {
9558 /* We're not interested in MEMs that are only clobbered. */
9562 /* We're not interested in the MEM associated with a
9563 CONST_DOUBLE, so there's no need to traverse into this. */
9567 /* We're not interested in any MEMs that only appear in notes. */
9571 /* This is not a MEM. */
9573 }
9575 /* See if we've already seen this MEM. */
9578 {
9580 /* The modes of the two memory accesses are different. If
9581 this happens, something tricky is going on, and we just
9582 don't optimize accesses to this MEM. */
9586 }
9588 /* Resize the array, if necessary. */
9590 {
9593 else
9599 }
9601 /* Actually insert the MEM. */
9603 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9604 because we can't put it in a register. We still store it in the
9605 table, though, so that if we see the same address later, but in a
9606 non-BLK mode, we'll not think we can optimize it at that point. */
9612 }
9614 /* Like load_mems, but also ensures that SET_IN_LOOP,
9615 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9616 values after load_mems. */
9618 static void
9622 {
9627 /* Recalculate set_in_loop and friends since load_mems may have
9628 created new registers. */
9630 {
9638 {
9639 /* Grow all the arrays. */
9644 }
9645 /* Clear the arrays */
9655 {
9658 }
9660 #ifdef AVOID_CCMODE_COPIES
9661 /* Don't try to move insns which set CC registers if we should not
9662 create CCmode register copies. */
9666 #endif
9668 /* Set n_times_set for the new registers. */
9672 }
9673 }
9675 /* Move MEMs into registers for the duration of the loop. */
9677 static void
9680 {
9683 rtx p;
9686 /* Nonzero if the next instruction may never be executed. */
9693 /* Find start of the extended basic block that enters the loop. */
9697 ;
9701 /* Build table of mems that get set to constant values before the
9702 loop. */
9706 /* Check to see if it's possible that some instructions in the
9707 loop are never executed. */
9711 {
9715 /* If we enter the loop in the middle, and scan
9716 around to the beginning, don't set maybe_never
9717 for that. This must be an unconditional jump,
9718 otherwise the code at the top of the loop might
9719 never be executed. Unconditional jumps are
9720 followed a by barrier then loop end. */
9725 {
9727 /* Something complicated. */
9729 else
9730 /* If there are any more instructions in the loop, they
9731 might not be reached. */
9733 }
9736 }
9738 /* Actually move the MEMs. */
9740 {
9741 regset_head copies;
9743 rtx reg;
9745 rtx mem_list_entry;
9749 /* There's no telling whether or not MEM is modified. */
9752 /* Go through the MEMs written to in the loop to see if this
9753 one is aliased by one of them. */
9756 {
9761 {
9762 /* MEM is indeed aliased by this store. */
9765 }
9767 }
9773 /* If this MEM is written to, we must be sure that there
9774 are no reads from another MEM that aliases this one. */
9776 {
9780 {
9784 VOIDmode,
9786 rtx_varies_p))
9787 {
9788 /* It's not safe to hoist loop_mems[i] out of
9789 the loop because writes to it might not be
9790 seen by reads from loop_mems[j]. */
9793 }
9794 }
9795 }
9798 /* We can't access the MEM outside the loop; it might
9799 cause a trap that wouldn't have happened otherwise. */
9803 /* We thought we were going to lift this MEM out of the
9804 loop, but later discovered that we could not. */
9809 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9810 order to keep scan_loop from moving stores to this MEM
9811 out of the loop just because this REG is neither a
9812 user-variable nor used in the loop test. */
9817 /* Now, replace all references to the MEM with the
9818 corresponding pesudos. */
9823 {
9824 rtx_and_int ri;
9825 rtx set;
9828 {
9829 /* See if this copies the mem into a register that isn't
9830 modified afterwards. We'll try to do copy propagation
9831 a little further on. */
9834 /* @@@ This test is _way_ too conservative. */
9835 && ! maybe_never
9845 }
9850 }
9853 /* We couldn't replace all occurrences of the MEM. */
9855 else
9856 {
9857 /* Load the memory immediately before LOOP->START, which is
9858 the NOTE_LOOP_BEG. */
9860 rtx set;
9866 {
9870 {
9874 /* Extending hard register lifetimes cuases crash
9875 on SRC targets. Doing so on non-SRC is
9876 probably also not good idea, since we most
9877 probably have pseudoregister equivalence as
9878 well. */
9881 }
9882 /* Use the constant equivalence if that is cheap enough. */
9888 {
9891 }
9893 /* If best_equiv is nonzero, we know that MEM is set to a
9894 constant or register before the loop. We will use this
9895 knowledge to initialize the shadow register with that
9896 constant or reg rather than by loading from MEM. */
9899 }
9908 {
9910 {
9911 /* We must compute the former
9912 right-after-the-end label before we insert
9913 the new one. */
9917 }
9919 /* Store the memory immediately after END, which is
9920 the NOTE_LOOP_END. */
9923 }
9926 {
9931 }
9933 /* Attempt a bit of copy propagation. This helps untangle the
9934 data flow, and enables {basic,general}_induction_var to find
9935 more bivs/givs. */
9936 EXECUTE_IF_SET_IN_REG_SET
9938 {
9940 });
9942 }
9943 }
9946 {
9947 /* Now, we need to replace all references to the previous exit
9948 label with the new one. */
9949 rtx_pair rr;
9954 {
9957 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9958 field. This is not handled by for_each_rtx because it doesn't
9959 handle unprinted ('0') fields. We need to update JUMP_LABEL
9960 because the immediately following unroll pass will use it.
9961 replace_label would not work anyways, because that only handles
9962 LABEL_REFs. */
9965 }
9966 }
9969 }
9971 /* For communication between note_reg_stored and its caller. */
9972 struct note_reg_stored_arg
9973 {
9975 rtx reg;
9976 };
9978 /* Called via note_stores, record in SET_SEEN whether X, which is written,
9979 is equal to ARG. */
9980 static void
9984 {
9988 }
9990 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
9991 There must be exactly one insn that sets this pseudo; it will be
9992 deleted if all replacements succeed and we can prove that the register
9993 is not used after the loop. */
9995 static void
9998 rtx replacement;
10000 {
10001 /* This is the reg that we are copying from. */
10004 rtx insn;
10005 /* These help keep track of whether we replaced all uses of the reg. */
10012 {
10013 rtx set;
10015 /* Only substitute within one extended basic block from the initializing
10016 insn. */
10023 /* Is this the initializing insn? */
10028 {
10035 }
10037 /* Only substitute after seeing the initializing insn. */
10039 {
10050 /* Stop replacing when REPLACEMENT is modified. */
10056 }
10057 }
10061 {
10065 {
10071 }
10074 }
10075 }
10077 /* Replace MEM with its associated pseudo register. This function is
10078 called from load_mems via for_each_rtx. DATA is actually an
10079 rtx_and_int * describing the instruction currently being scanned
10080 and the MEM we are currently replacing. */
10082 static int
10086 {
10088 rtx insn;
10096 {
10101 /* We're not interested in the MEM associated with a
10102 CONST_DOUBLE, so there's no need to traverse into one. */
10106 /* This is not a MEM. */
10108 }
10114 /* This is not the MEM we are currently replacing. */
10119 /* Actually replace the MEM. */
10123 }
10125 /* Replace one register with another. Called through for_each_rtx; PX points
10126 to the rtx being scanned. DATA is actually an array of three rtx's; the
10127 first one is the one to be replaced, and the second one the replacement.
10128 The third one is the current insn. */
10130 static int
10134 {
10145 }
10147 /* Replace occurrences of the old exit label for the loop with the new
10148 one. DATA is an rtx_pair containing the old and new labels,
10149 respectively. */
10151 static int
10155 {
10174 }