c++: fix explicit/copy problem [PR109247]
[official-gcc.git] / gcc / cse.cc
blob2bb63ac4105bde7d82a4768f994ae019e3242018
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2023 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "cfghooks.h"
28 #include "df.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "insn-config.h"
32 #include "regs.h"
33 #include "emit-rtl.h"
34 #include "recog.h"
35 #include "cfgrtl.h"
36 #include "cfganal.h"
37 #include "cfgcleanup.h"
38 #include "alias.h"
39 #include "toplev.h"
40 #include "rtlhooks-def.h"
41 #include "tree-pass.h"
42 #include "dbgcnt.h"
43 #include "rtl-iter.h"
44 #include "regs.h"
45 #include "function-abi.h"
46 #include "rtlanal.h"
47 #include "expr.h"
49 /* The basic idea of common subexpression elimination is to go
50 through the code, keeping a record of expressions that would
51 have the same value at the current scan point, and replacing
52 expressions encountered with the cheapest equivalent expression.
54 It is too complicated to keep track of the different possibilities
55 when control paths merge in this code; so, at each label, we forget all
56 that is known and start fresh. This can be described as processing each
57 extended basic block separately. We have a separate pass to perform
58 global CSE.
60 Note CSE can turn a conditional or computed jump into a nop or
61 an unconditional jump. When this occurs we arrange to run the jump
62 optimizer after CSE to delete the unreachable code.
64 We use two data structures to record the equivalent expressions:
65 a hash table for most expressions, and a vector of "quantity
66 numbers" to record equivalent (pseudo) registers.
68 The use of the special data structure for registers is desirable
69 because it is faster. It is possible because registers references
70 contain a fairly small number, the register number, taken from
71 a contiguously allocated series, and two register references are
72 identical if they have the same number. General expressions
73 do not have any such thing, so the only way to retrieve the
74 information recorded on an expression other than a register
75 is to keep it in a hash table.
77 Registers and "quantity numbers":
79 At the start of each basic block, all of the (hardware and pseudo)
80 registers used in the function are given distinct quantity
81 numbers to indicate their contents. During scan, when the code
82 copies one register into another, we copy the quantity number.
83 When a register is loaded in any other way, we allocate a new
84 quantity number to describe the value generated by this operation.
85 `REG_QTY (N)' records what quantity register N is currently thought
86 of as containing.
88 All real quantity numbers are greater than or equal to zero.
89 If register N has not been assigned a quantity, `REG_QTY (N)' will
90 equal -N - 1, which is always negative.
92 Quantity numbers below zero do not exist and none of the `qty_table'
93 entries should be referenced with a negative index.
95 We also maintain a bidirectional chain of registers for each
96 quantity number. The `qty_table` members `first_reg' and `last_reg',
97 and `reg_eqv_table' members `next' and `prev' hold these chains.
99 The first register in a chain is the one whose lifespan is least local.
100 Among equals, it is the one that was seen first.
101 We replace any equivalent register with that one.
103 If two registers have the same quantity number, it must be true that
104 REG expressions with qty_table `mode' must be in the hash table for both
105 registers and must be in the same class.
107 The converse is not true. Since hard registers may be referenced in
108 any mode, two REG expressions might be equivalent in the hash table
109 but not have the same quantity number if the quantity number of one
110 of the registers is not the same mode as those expressions.
112 Constants and quantity numbers
114 When a quantity has a known constant value, that value is stored
115 in the appropriate qty_table `const_rtx'. This is in addition to
116 putting the constant in the hash table as is usual for non-regs.
118 Whether a reg or a constant is preferred is determined by the configuration
119 macro CONST_COSTS and will often depend on the constant value. In any
120 event, expressions containing constants can be simplified, by fold_rtx.
122 When a quantity has a known nearly constant value (such as an address
123 of a stack slot), that value is stored in the appropriate qty_table
124 `const_rtx'.
126 Integer constants don't have a machine mode. However, cse
127 determines the intended machine mode from the destination
128 of the instruction that moves the constant. The machine mode
129 is recorded in the hash table along with the actual RTL
130 constant expression so that different modes are kept separate.
132 Other expressions:
134 To record known equivalences among expressions in general
135 we use a hash table called `table'. It has a fixed number of buckets
136 that contain chains of `struct table_elt' elements for expressions.
137 These chains connect the elements whose expressions have the same
138 hash codes.
140 Other chains through the same elements connect the elements which
141 currently have equivalent values.
143 Register references in an expression are canonicalized before hashing
144 the expression. This is done using `reg_qty' and qty_table `first_reg'.
145 The hash code of a register reference is computed using the quantity
146 number, not the register number.
148 When the value of an expression changes, it is necessary to remove from the
149 hash table not just that expression but all expressions whose values
150 could be different as a result.
152 1. If the value changing is in memory, except in special cases
153 ANYTHING referring to memory could be changed. That is because
154 nobody knows where a pointer does not point.
155 The function `invalidate_memory' removes what is necessary.
157 The special cases are when the address is constant or is
158 a constant plus a fixed register such as the frame pointer
159 or a static chain pointer. When such addresses are stored in,
160 we can tell exactly which other such addresses must be invalidated
161 due to overlap. `invalidate' does this.
162 All expressions that refer to non-constant
163 memory addresses are also invalidated. `invalidate_memory' does this.
165 2. If the value changing is a register, all expressions
166 containing references to that register, and only those,
167 must be removed.
169 Because searching the entire hash table for expressions that contain
170 a register is very slow, we try to figure out when it isn't necessary.
171 Precisely, this is necessary only when expressions have been
172 entered in the hash table using this register, and then the value has
173 changed, and then another expression wants to be added to refer to
174 the register's new value. This sequence of circumstances is rare
175 within any one basic block.
177 `REG_TICK' and `REG_IN_TABLE', accessors for members of
178 cse_reg_info, are used to detect this case. REG_TICK (i) is
179 incremented whenever a value is stored in register i.
180 REG_IN_TABLE (i) holds -1 if no references to register i have been
181 entered in the table; otherwise, it contains the value REG_TICK (i)
182 had when the references were entered. If we want to enter a
183 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
184 remove old references. Until we want to enter a new entry, the
185 mere fact that the two vectors don't match makes the entries be
186 ignored if anyone tries to match them.
188 Registers themselves are entered in the hash table as well as in
189 the equivalent-register chains. However, `REG_TICK' and
190 `REG_IN_TABLE' do not apply to expressions which are simple
191 register references. These expressions are removed from the table
192 immediately when they become invalid, and this can be done even if
193 we do not immediately search for all the expressions that refer to
194 the register.
196 A CLOBBER rtx in an instruction invalidates its operand for further
197 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
198 invalidates everything that resides in memory.
200 Related expressions:
202 Constant expressions that differ only by an additive integer
203 are called related. When a constant expression is put in
204 the table, the related expression with no constant term
205 is also entered. These are made to point at each other
206 so that it is possible to find out if there exists any
207 register equivalent to an expression related to a given expression. */
209 /* Length of qty_table vector. We know in advance we will not need
210 a quantity number this big. */
212 static int max_qty;
214 /* Next quantity number to be allocated.
215 This is 1 + the largest number needed so far. */
217 static int next_qty;
219 /* Per-qty information tracking.
221 `first_reg' and `last_reg' track the head and tail of the
222 chain of registers which currently contain this quantity.
224 `mode' contains the machine mode of this quantity.
226 `const_rtx' holds the rtx of the constant value of this
227 quantity, if known. A summations of the frame/arg pointer
228 and a constant can also be entered here. When this holds
229 a known value, `const_insn' is the insn which stored the
230 constant value.
232 `comparison_{code,const,qty}' are used to track when a
233 comparison between a quantity and some constant or register has
234 been passed. In such a case, we know the results of the comparison
235 in case we see it again. These members record a comparison that
236 is known to be true. `comparison_code' holds the rtx code of such
237 a comparison, else it is set to UNKNOWN and the other two
238 comparison members are undefined. `comparison_const' holds
239 the constant being compared against, or zero if the comparison
240 is not against a constant. `comparison_qty' holds the quantity
241 being compared against when the result is known. If the comparison
242 is not with a register, `comparison_qty' is -1. */
244 struct qty_table_elem
246 rtx const_rtx;
247 rtx_insn *const_insn;
248 rtx comparison_const;
249 int comparison_qty;
250 unsigned int first_reg, last_reg;
251 ENUM_BITFIELD(machine_mode) mode : MACHINE_MODE_BITSIZE;
252 ENUM_BITFIELD(rtx_code) comparison_code : RTX_CODE_BITSIZE;
255 /* The table of all qtys, indexed by qty number. */
256 static struct qty_table_elem *qty_table;
258 /* Insn being scanned. */
260 static rtx_insn *this_insn;
261 static bool optimize_this_for_speed_p;
263 /* Index by register number, gives the number of the next (or
264 previous) register in the chain of registers sharing the same
265 value.
267 Or -1 if this register is at the end of the chain.
269 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
271 /* Per-register equivalence chain. */
272 struct reg_eqv_elem
274 int next, prev;
277 /* The table of all register equivalence chains. */
278 static struct reg_eqv_elem *reg_eqv_table;
280 struct cse_reg_info
282 /* The timestamp at which this register is initialized. */
283 unsigned int timestamp;
285 /* The quantity number of the register's current contents. */
286 int reg_qty;
288 /* The number of times the register has been altered in the current
289 basic block. */
290 int reg_tick;
292 /* The REG_TICK value at which rtx's containing this register are
293 valid in the hash table. If this does not equal the current
294 reg_tick value, such expressions existing in the hash table are
295 invalid. */
296 int reg_in_table;
298 /* The SUBREG that was set when REG_TICK was last incremented. Set
299 to -1 if the last store was to the whole register, not a subreg. */
300 unsigned int subreg_ticked;
303 /* A table of cse_reg_info indexed by register numbers. */
304 static struct cse_reg_info *cse_reg_info_table;
306 /* The size of the above table. */
307 static unsigned int cse_reg_info_table_size;
309 /* The index of the first entry that has not been initialized. */
310 static unsigned int cse_reg_info_table_first_uninitialized;
312 /* The timestamp at the beginning of the current run of
313 cse_extended_basic_block. We increment this variable at the beginning of
314 the current run of cse_extended_basic_block. The timestamp field of a
315 cse_reg_info entry matches the value of this variable if and only
316 if the entry has been initialized during the current run of
317 cse_extended_basic_block. */
318 static unsigned int cse_reg_info_timestamp;
320 /* A HARD_REG_SET containing all the hard registers for which there is
321 currently a REG expression in the hash table. Note the difference
322 from the above variables, which indicate if the REG is mentioned in some
323 expression in the table. */
325 static HARD_REG_SET hard_regs_in_table;
327 /* True if CSE has altered the CFG. */
328 static bool cse_cfg_altered;
330 /* True if CSE has altered conditional jump insns in such a way
331 that jump optimization should be redone. */
332 static bool cse_jumps_altered;
334 /* True if we put a LABEL_REF into the hash table for an INSN
335 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
336 to put in the note. */
337 static bool recorded_label_ref;
339 /* canon_hash stores 1 in do_not_record if it notices a reference to PC or
340 some other volatile subexpression. */
342 static int do_not_record;
344 /* canon_hash stores 1 in hash_arg_in_memory
345 if it notices a reference to memory within the expression being hashed. */
347 static int hash_arg_in_memory;
349 /* The hash table contains buckets which are chains of `struct table_elt's,
350 each recording one expression's information.
351 That expression is in the `exp' field.
353 The canon_exp field contains a canonical (from the point of view of
354 alias analysis) version of the `exp' field.
356 Those elements with the same hash code are chained in both directions
357 through the `next_same_hash' and `prev_same_hash' fields.
359 Each set of expressions with equivalent values
360 are on a two-way chain through the `next_same_value'
361 and `prev_same_value' fields, and all point with
362 the `first_same_value' field at the first element in
363 that chain. The chain is in order of increasing cost.
364 Each element's cost value is in its `cost' field.
366 The `in_memory' field is nonzero for elements that
367 involve any reference to memory. These elements are removed
368 whenever a write is done to an unidentified location in memory.
369 To be safe, we assume that a memory address is unidentified unless
370 the address is either a symbol constant or a constant plus
371 the frame pointer or argument pointer.
373 The `related_value' field is used to connect related expressions
374 (that differ by adding an integer).
375 The related expressions are chained in a circular fashion.
376 `related_value' is zero for expressions for which this
377 chain is not useful.
379 The `cost' field stores the cost of this element's expression.
380 The `regcost' field stores the value returned by approx_reg_cost for
381 this element's expression.
383 The `is_const' flag is set if the element is a constant (including
384 a fixed address).
386 The `flag' field is used as a temporary during some search routines.
388 The `mode' field is usually the same as GET_MODE (`exp'), but
389 if `exp' is a CONST_INT and has no machine mode then the `mode'
390 field is the mode it was being used as. Each constant is
391 recorded separately for each mode it is used with. */
393 struct table_elt
395 rtx exp;
396 rtx canon_exp;
397 struct table_elt *next_same_hash;
398 struct table_elt *prev_same_hash;
399 struct table_elt *next_same_value;
400 struct table_elt *prev_same_value;
401 struct table_elt *first_same_value;
402 struct table_elt *related_value;
403 int cost;
404 int regcost;
405 ENUM_BITFIELD(machine_mode) mode : MACHINE_MODE_BITSIZE;
406 char in_memory;
407 char is_const;
408 char flag;
411 /* We don't want a lot of buckets, because we rarely have very many
412 things stored in the hash table, and a lot of buckets slows
413 down a lot of loops that happen frequently. */
414 #define HASH_SHIFT 5
415 #define HASH_SIZE (1 << HASH_SHIFT)
416 #define HASH_MASK (HASH_SIZE - 1)
418 /* Determine whether register number N is considered a fixed register for the
419 purpose of approximating register costs.
420 It is desirable to replace other regs with fixed regs, to reduce need for
421 non-fixed hard regs.
422 A reg wins if it is either the frame pointer or designated as fixed. */
423 #define FIXED_REGNO_P(N) \
424 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
425 || fixed_regs[N] || global_regs[N])
427 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
428 hard registers and pointers into the frame are the cheapest with a cost
429 of 0. Next come pseudos with a cost of one and other hard registers with
430 a cost of 2. Aside from these special cases, call `rtx_cost'. */
432 #define CHEAP_REGNO(N) \
433 (REGNO_PTR_FRAME_P (N) \
434 || (HARD_REGISTER_NUM_P (N) \
435 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
437 #define COST(X, MODE) \
438 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
439 #define COST_IN(X, MODE, OUTER, OPNO) \
440 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
442 /* Get the number of times this register has been updated in this
443 basic block. */
445 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
447 /* Get the point at which REG was recorded in the table. */
449 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
451 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
452 SUBREG). */
454 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
456 /* Get the quantity number for REG. */
458 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
460 /* Determine if the quantity number for register X represents a valid index
461 into the qty_table. */
463 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
465 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
467 #define CHEAPER(X, Y) \
468 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
470 static struct table_elt *table[HASH_SIZE];
472 /* Chain of `struct table_elt's made so far for this function
473 but currently removed from the table. */
475 static struct table_elt *free_element_chain;
477 /* Trace a patch through the CFG. */
479 struct branch_path
481 /* The basic block for this path entry. */
482 basic_block bb;
485 /* This data describes a block that will be processed by
486 cse_extended_basic_block. */
488 struct cse_basic_block_data
490 /* Total number of SETs in block. */
491 int nsets;
492 /* Size of current branch path, if any. */
493 int path_size;
494 /* Current path, indicating which basic_blocks will be processed. */
495 struct branch_path *path;
499 /* Pointers to the live in/live out bitmaps for the boundaries of the
500 current EBB. */
501 static bitmap cse_ebb_live_in, cse_ebb_live_out;
503 /* A simple bitmap to track which basic blocks have been visited
504 already as part of an already processed extended basic block. */
505 static sbitmap cse_visited_basic_blocks;
507 static bool fixed_base_plus_p (rtx x);
508 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
509 static int preferable (int, int, int, int);
510 static void new_basic_block (void);
511 static void make_new_qty (unsigned int, machine_mode);
512 static void make_regs_eqv (unsigned int, unsigned int);
513 static void delete_reg_equiv (unsigned int);
514 static bool mention_regs (rtx);
515 static bool insert_regs (rtx, struct table_elt *, bool);
516 static void remove_from_table (struct table_elt *, unsigned);
517 static void remove_pseudo_from_table (rtx, unsigned);
518 static struct table_elt *lookup (rtx, unsigned, machine_mode);
519 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
520 static rtx lookup_as_function (rtx, enum rtx_code);
521 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
522 machine_mode, int, int);
523 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
524 machine_mode);
525 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
526 static void invalidate (rtx, machine_mode);
527 static void remove_invalid_refs (unsigned int);
528 static void remove_invalid_subreg_refs (unsigned int, poly_uint64,
529 machine_mode);
530 static void rehash_using_reg (rtx);
531 static void invalidate_memory (void);
532 static rtx use_related_value (rtx, struct table_elt *);
534 static inline unsigned canon_hash (rtx, machine_mode);
535 static inline unsigned safe_hash (rtx, machine_mode);
536 static inline unsigned hash_rtx_string (const char *);
538 static rtx canon_reg (rtx, rtx_insn *);
539 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
540 machine_mode *,
541 machine_mode *);
542 static rtx fold_rtx (rtx, rtx_insn *);
543 static rtx equiv_constant (rtx);
544 static void record_jump_equiv (rtx_insn *, bool);
545 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx);
546 static void cse_insn (rtx_insn *);
547 static void cse_prescan_path (struct cse_basic_block_data *);
548 static void invalidate_from_clobbers (rtx_insn *);
549 static void invalidate_from_sets_and_clobbers (rtx_insn *);
550 static void cse_extended_basic_block (struct cse_basic_block_data *);
551 extern void dump_class (struct table_elt*);
552 static void get_cse_reg_info_1 (unsigned int regno);
553 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
555 static void flush_hash_table (void);
556 static bool insn_live_p (rtx_insn *, int *);
557 static bool set_live_p (rtx, int *);
558 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
559 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
560 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
561 bool);
564 #undef RTL_HOOKS_GEN_LOWPART
565 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
567 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
569 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
570 register (hard registers may require `do_not_record' to be set). */
572 static inline unsigned
573 HASH (rtx x, machine_mode mode)
575 unsigned h = (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER
576 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (x)))
577 : canon_hash (x, mode));
578 return (h ^ (h >> HASH_SHIFT)) & HASH_MASK;
581 /* Like HASH, but without side-effects. */
583 static inline unsigned
584 SAFE_HASH (rtx x, machine_mode mode)
586 unsigned h = (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER
587 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (x)))
588 : safe_hash (x, mode));
589 return (h ^ (h >> HASH_SHIFT)) & HASH_MASK;
592 /* Nonzero if X has the form (PLUS frame-pointer integer). */
594 static bool
595 fixed_base_plus_p (rtx x)
597 switch (GET_CODE (x))
599 case REG:
600 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
601 return true;
602 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
603 return true;
604 return false;
606 case PLUS:
607 if (!CONST_INT_P (XEXP (x, 1)))
608 return false;
609 return fixed_base_plus_p (XEXP (x, 0));
611 default:
612 return false;
616 /* Dump the expressions in the equivalence class indicated by CLASSP.
617 This function is used only for debugging. */
618 DEBUG_FUNCTION void
619 dump_class (struct table_elt *classp)
621 struct table_elt *elt;
623 fprintf (stderr, "Equivalence chain for ");
624 print_rtl (stderr, classp->exp);
625 fprintf (stderr, ": \n");
627 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
629 print_rtl (stderr, elt->exp);
630 fprintf (stderr, "\n");
634 /* Return an estimate of the cost of the registers used in an rtx.
635 This is mostly the number of different REG expressions in the rtx;
636 however for some exceptions like fixed registers we use a cost of
637 0. If any other hard register reference occurs, return MAX_COST. */
639 static int
640 approx_reg_cost (const_rtx x)
642 int cost = 0;
643 subrtx_iterator::array_type array;
644 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
646 const_rtx x = *iter;
647 if (REG_P (x))
649 unsigned int regno = REGNO (x);
650 if (!CHEAP_REGNO (regno))
652 if (regno < FIRST_PSEUDO_REGISTER)
654 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
655 return MAX_COST;
656 cost += 2;
658 else
659 cost += 1;
663 return cost;
666 /* Return a negative value if an rtx A, whose costs are given by COST_A
667 and REGCOST_A, is more desirable than an rtx B.
668 Return a positive value if A is less desirable, or 0 if the two are
669 equally good. */
670 static int
671 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
673 /* First, get rid of cases involving expressions that are entirely
674 unwanted. */
675 if (cost_a != cost_b)
677 if (cost_a == MAX_COST)
678 return 1;
679 if (cost_b == MAX_COST)
680 return -1;
683 /* Avoid extending lifetimes of hardregs. */
684 if (regcost_a != regcost_b)
686 if (regcost_a == MAX_COST)
687 return 1;
688 if (regcost_b == MAX_COST)
689 return -1;
692 /* Normal operation costs take precedence. */
693 if (cost_a != cost_b)
694 return cost_a - cost_b;
695 /* Only if these are identical consider effects on register pressure. */
696 if (regcost_a != regcost_b)
697 return regcost_a - regcost_b;
698 return 0;
701 /* Internal function, to compute cost when X is not a register; called
702 from COST macro to keep it simple. */
704 static int
705 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
707 scalar_int_mode int_mode, inner_mode;
708 return ((GET_CODE (x) == SUBREG
709 && REG_P (SUBREG_REG (x))
710 && is_int_mode (mode, &int_mode)
711 && is_int_mode (GET_MODE (SUBREG_REG (x)), &inner_mode)
712 && GET_MODE_SIZE (int_mode) < GET_MODE_SIZE (inner_mode)
713 && subreg_lowpart_p (x)
714 && TRULY_NOOP_TRUNCATION_MODES_P (int_mode, inner_mode))
716 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
720 /* Initialize CSE_REG_INFO_TABLE. */
722 static void
723 init_cse_reg_info (unsigned int nregs)
725 /* Do we need to grow the table? */
726 if (nregs > cse_reg_info_table_size)
728 unsigned int new_size;
730 if (cse_reg_info_table_size < 2048)
732 /* Compute a new size that is a power of 2 and no smaller
733 than the large of NREGS and 64. */
734 new_size = (cse_reg_info_table_size
735 ? cse_reg_info_table_size : 64);
737 while (new_size < nregs)
738 new_size *= 2;
740 else
742 /* If we need a big table, allocate just enough to hold
743 NREGS registers. */
744 new_size = nregs;
747 /* Reallocate the table with NEW_SIZE entries. */
748 free (cse_reg_info_table);
749 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
750 cse_reg_info_table_size = new_size;
751 cse_reg_info_table_first_uninitialized = 0;
754 /* Do we have all of the first NREGS entries initialized? */
755 if (cse_reg_info_table_first_uninitialized < nregs)
757 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
758 unsigned int i;
760 /* Put the old timestamp on newly allocated entries so that they
761 will all be considered out of date. We do not touch those
762 entries beyond the first NREGS entries to be nice to the
763 virtual memory. */
764 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
765 cse_reg_info_table[i].timestamp = old_timestamp;
767 cse_reg_info_table_first_uninitialized = nregs;
771 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
773 static void
774 get_cse_reg_info_1 (unsigned int regno)
776 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
777 entry will be considered to have been initialized. */
778 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
780 /* Initialize the rest of the entry. */
781 cse_reg_info_table[regno].reg_tick = 1;
782 cse_reg_info_table[regno].reg_in_table = -1;
783 cse_reg_info_table[regno].subreg_ticked = -1;
784 cse_reg_info_table[regno].reg_qty = -regno - 1;
787 /* Find a cse_reg_info entry for REGNO. */
789 static inline struct cse_reg_info *
790 get_cse_reg_info (unsigned int regno)
792 struct cse_reg_info *p = &cse_reg_info_table[regno];
794 /* If this entry has not been initialized, go ahead and initialize
795 it. */
796 if (p->timestamp != cse_reg_info_timestamp)
797 get_cse_reg_info_1 (regno);
799 return p;
802 /* Clear the hash table and initialize each register with its own quantity,
803 for a new basic block. */
805 static void
806 new_basic_block (void)
808 int i;
810 next_qty = 0;
812 /* Invalidate cse_reg_info_table. */
813 cse_reg_info_timestamp++;
815 /* Clear out hash table state for this pass. */
816 CLEAR_HARD_REG_SET (hard_regs_in_table);
818 /* The per-quantity values used to be initialized here, but it is
819 much faster to initialize each as it is made in `make_new_qty'. */
821 for (i = 0; i < HASH_SIZE; i++)
823 struct table_elt *first;
825 first = table[i];
826 if (first != NULL)
828 struct table_elt *last = first;
830 table[i] = NULL;
832 while (last->next_same_hash != NULL)
833 last = last->next_same_hash;
835 /* Now relink this hash entire chain into
836 the free element list. */
838 last->next_same_hash = free_element_chain;
839 free_element_chain = first;
844 /* Say that register REG contains a quantity in mode MODE not in any
845 register before and initialize that quantity. */
847 static void
848 make_new_qty (unsigned int reg, machine_mode mode)
850 int q;
851 struct qty_table_elem *ent;
852 struct reg_eqv_elem *eqv;
854 gcc_assert (next_qty < max_qty);
856 q = REG_QTY (reg) = next_qty++;
857 ent = &qty_table[q];
858 ent->first_reg = reg;
859 ent->last_reg = reg;
860 ent->mode = mode;
861 ent->const_rtx = ent->const_insn = NULL;
862 ent->comparison_code = UNKNOWN;
864 eqv = &reg_eqv_table[reg];
865 eqv->next = eqv->prev = -1;
868 /* Make reg NEW equivalent to reg OLD.
869 OLD is not changing; NEW is. */
871 static void
872 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
874 unsigned int lastr, firstr;
875 int q = REG_QTY (old_reg);
876 struct qty_table_elem *ent;
878 ent = &qty_table[q];
880 /* Nothing should become eqv until it has a "non-invalid" qty number. */
881 gcc_assert (REGNO_QTY_VALID_P (old_reg));
883 REG_QTY (new_reg) = q;
884 firstr = ent->first_reg;
885 lastr = ent->last_reg;
887 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
888 hard regs. Among pseudos, if NEW will live longer than any other reg
889 of the same qty, and that is beyond the current basic block,
890 make it the new canonical replacement for this qty. */
891 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
892 /* Certain fixed registers might be of the class NO_REGS. This means
893 that not only can they not be allocated by the compiler, but
894 they cannot be used in substitutions or canonicalizations
895 either. */
896 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
897 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
898 || (new_reg >= FIRST_PSEUDO_REGISTER
899 && (firstr < FIRST_PSEUDO_REGISTER
900 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
901 && !bitmap_bit_p (cse_ebb_live_out, firstr))
902 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
903 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
905 reg_eqv_table[firstr].prev = new_reg;
906 reg_eqv_table[new_reg].next = firstr;
907 reg_eqv_table[new_reg].prev = -1;
908 ent->first_reg = new_reg;
910 else
912 /* If NEW is a hard reg (known to be non-fixed), insert at end.
913 Otherwise, insert before any non-fixed hard regs that are at the
914 end. Registers of class NO_REGS cannot be used as an
915 equivalent for anything. */
916 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
917 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
918 && new_reg >= FIRST_PSEUDO_REGISTER)
919 lastr = reg_eqv_table[lastr].prev;
920 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
921 if (reg_eqv_table[lastr].next >= 0)
922 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
923 else
924 qty_table[q].last_reg = new_reg;
925 reg_eqv_table[lastr].next = new_reg;
926 reg_eqv_table[new_reg].prev = lastr;
930 /* Remove REG from its equivalence class. */
932 static void
933 delete_reg_equiv (unsigned int reg)
935 struct qty_table_elem *ent;
936 int q = REG_QTY (reg);
937 int p, n;
939 /* If invalid, do nothing. */
940 if (! REGNO_QTY_VALID_P (reg))
941 return;
943 ent = &qty_table[q];
945 p = reg_eqv_table[reg].prev;
946 n = reg_eqv_table[reg].next;
948 if (n != -1)
949 reg_eqv_table[n].prev = p;
950 else
951 ent->last_reg = p;
952 if (p != -1)
953 reg_eqv_table[p].next = n;
954 else
955 ent->first_reg = n;
957 REG_QTY (reg) = -reg - 1;
960 /* Remove any invalid expressions from the hash table
961 that refer to any of the registers contained in expression X.
963 Make sure that newly inserted references to those registers
964 as subexpressions will be considered valid.
966 mention_regs is not called when a register itself
967 is being stored in the table.
969 Return true if we have done something that may have changed
970 the hash code of X. */
972 static bool
973 mention_regs (rtx x)
975 enum rtx_code code;
976 int i, j;
977 const char *fmt;
978 bool changed = false;
980 if (x == 0)
981 return false;
983 code = GET_CODE (x);
984 if (code == REG)
986 unsigned int regno = REGNO (x);
987 unsigned int endregno = END_REGNO (x);
988 unsigned int i;
990 for (i = regno; i < endregno; i++)
992 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
993 remove_invalid_refs (i);
995 REG_IN_TABLE (i) = REG_TICK (i);
996 SUBREG_TICKED (i) = -1;
999 return false;
1002 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1003 pseudo if they don't use overlapping words. We handle only pseudos
1004 here for simplicity. */
1005 if (code == SUBREG && REG_P (SUBREG_REG (x))
1006 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1008 unsigned int i = REGNO (SUBREG_REG (x));
1010 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1012 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1013 the last store to this register really stored into this
1014 subreg, then remove the memory of this subreg.
1015 Otherwise, remove any memory of the entire register and
1016 all its subregs from the table. */
1017 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1018 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1019 remove_invalid_refs (i);
1020 else
1021 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1024 REG_IN_TABLE (i) = REG_TICK (i);
1025 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1026 return false;
1029 /* If X is a comparison or a COMPARE and either operand is a register
1030 that does not have a quantity, give it one. This is so that a later
1031 call to record_jump_equiv won't cause X to be assigned a different
1032 hash code and not found in the table after that call.
1034 It is not necessary to do this here, since rehash_using_reg can
1035 fix up the table later, but doing this here eliminates the need to
1036 call that expensive function in the most common case where the only
1037 use of the register is in the comparison. */
1039 if (code == COMPARE || COMPARISON_P (x))
1041 if (REG_P (XEXP (x, 0))
1042 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1043 if (insert_regs (XEXP (x, 0), NULL, false))
1045 rehash_using_reg (XEXP (x, 0));
1046 changed = true;
1049 if (REG_P (XEXP (x, 1))
1050 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1051 if (insert_regs (XEXP (x, 1), NULL, false))
1053 rehash_using_reg (XEXP (x, 1));
1054 changed = true;
1058 fmt = GET_RTX_FORMAT (code);
1059 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1060 if (fmt[i] == 'e')
1062 if (mention_regs (XEXP (x, i)))
1063 changed = true;
1065 else if (fmt[i] == 'E')
1066 for (j = 0; j < XVECLEN (x, i); j++)
1067 if (mention_regs (XVECEXP (x, i, j)))
1068 changed = true;
1070 return changed;
1073 /* Update the register quantities for inserting X into the hash table
1074 with a value equivalent to CLASSP.
1075 (If the class does not contain a REG, it is irrelevant.)
1076 If MODIFIED is true, X is a destination; it is being modified.
1077 Note that delete_reg_equiv should be called on a register
1078 before insert_regs is done on that register with MODIFIED != 0.
1080 True value means that elements of reg_qty have changed
1081 so X's hash code may be different. */
1083 static bool
1084 insert_regs (rtx x, struct table_elt *classp, bool modified)
1086 if (REG_P (x))
1088 unsigned int regno = REGNO (x);
1089 int qty_valid;
1091 /* If REGNO is in the equivalence table already but is of the
1092 wrong mode for that equivalence, don't do anything here. */
1094 qty_valid = REGNO_QTY_VALID_P (regno);
1095 if (qty_valid)
1097 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1099 if (ent->mode != GET_MODE (x))
1100 return false;
1103 if (modified || ! qty_valid)
1105 if (classp)
1106 for (classp = classp->first_same_value;
1107 classp != 0;
1108 classp = classp->next_same_value)
1109 if (REG_P (classp->exp)
1110 && GET_MODE (classp->exp) == GET_MODE (x))
1112 unsigned c_regno = REGNO (classp->exp);
1114 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1116 /* Suppose that 5 is hard reg and 100 and 101 are
1117 pseudos. Consider
1119 (set (reg:si 100) (reg:si 5))
1120 (set (reg:si 5) (reg:si 100))
1121 (set (reg:di 101) (reg:di 5))
1123 We would now set REG_QTY (101) = REG_QTY (5), but the
1124 entry for 5 is in SImode. When we use this later in
1125 copy propagation, we get the register in wrong mode. */
1126 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1127 continue;
1129 make_regs_eqv (regno, c_regno);
1130 return true;
1133 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1134 than REG_IN_TABLE to find out if there was only a single preceding
1135 invalidation - for the SUBREG - or another one, which would be
1136 for the full register. However, if we find here that REG_TICK
1137 indicates that the register is invalid, it means that it has
1138 been invalidated in a separate operation. The SUBREG might be used
1139 now (then this is a recursive call), or we might use the full REG
1140 now and a SUBREG of it later. So bump up REG_TICK so that
1141 mention_regs will do the right thing. */
1142 if (! modified
1143 && REG_IN_TABLE (regno) >= 0
1144 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1145 REG_TICK (regno)++;
1146 make_new_qty (regno, GET_MODE (x));
1147 return true;
1150 return false;
1153 /* If X is a SUBREG, we will likely be inserting the inner register in the
1154 table. If that register doesn't have an assigned quantity number at
1155 this point but does later, the insertion that we will be doing now will
1156 not be accessible because its hash code will have changed. So assign
1157 a quantity number now. */
1159 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1160 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1162 insert_regs (SUBREG_REG (x), NULL, false);
1163 mention_regs (x);
1164 return true;
1166 else
1167 return mention_regs (x);
1171 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1172 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1173 CST is equal to an anchor. */
1175 static bool
1176 compute_const_anchors (rtx cst,
1177 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1178 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1180 unsigned HOST_WIDE_INT n = UINTVAL (cst);
1182 *lower_base = n & ~(targetm.const_anchor - 1);
1183 if ((unsigned HOST_WIDE_INT) *lower_base == n)
1184 return false;
1186 *upper_base = ((n + (targetm.const_anchor - 1))
1187 & ~(targetm.const_anchor - 1));
1188 *upper_offs = n - *upper_base;
1189 *lower_offs = n - *lower_base;
1190 return true;
1193 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1195 static void
1196 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1197 machine_mode mode)
1199 struct table_elt *elt;
1200 unsigned hash;
1201 rtx anchor_exp;
1202 rtx exp;
1204 anchor_exp = gen_int_mode (anchor, mode);
1205 hash = HASH (anchor_exp, mode);
1206 elt = lookup (anchor_exp, hash, mode);
1207 if (!elt)
1208 elt = insert (anchor_exp, NULL, hash, mode);
1210 exp = plus_constant (mode, reg, offs);
1211 /* REG has just been inserted and the hash codes recomputed. */
1212 mention_regs (exp);
1213 hash = HASH (exp, mode);
1215 /* Use the cost of the register rather than the whole expression. When
1216 looking up constant anchors we will further offset the corresponding
1217 expression therefore it does not make sense to prefer REGs over
1218 reg-immediate additions. Prefer instead the oldest expression. Also
1219 don't prefer pseudos over hard regs so that we derive constants in
1220 argument registers from other argument registers rather than from the
1221 original pseudo that was used to synthesize the constant. */
1222 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1225 /* The constant CST is equivalent to the register REG. Create
1226 equivalences between the two anchors of CST and the corresponding
1227 register-offset expressions using REG. */
1229 static void
1230 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1232 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1234 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1235 &upper_base, &upper_offs))
1236 return;
1238 /* Ignore anchors of value 0. Constants accessible from zero are
1239 simple. */
1240 if (lower_base != 0)
1241 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1243 if (upper_base != 0)
1244 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1247 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1248 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1249 valid expression. Return the cheapest and oldest of such expressions. In
1250 *OLD, return how old the resulting expression is compared to the other
1251 equivalent expressions. */
1253 static rtx
1254 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1255 unsigned *old)
1257 struct table_elt *elt;
1258 unsigned idx;
1259 struct table_elt *match_elt;
1260 rtx match;
1262 /* Find the cheapest and *oldest* expression to maximize the chance of
1263 reusing the same pseudo. */
1265 match_elt = NULL;
1266 match = NULL_RTX;
1267 for (elt = anchor_elt->first_same_value, idx = 0;
1268 elt;
1269 elt = elt->next_same_value, idx++)
1271 if (match_elt && CHEAPER (match_elt, elt))
1272 return match;
1274 if (REG_P (elt->exp)
1275 || (GET_CODE (elt->exp) == PLUS
1276 && REG_P (XEXP (elt->exp, 0))
1277 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1279 rtx x;
1281 /* Ignore expressions that are no longer valid. */
1282 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1283 continue;
1285 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1286 if (REG_P (x)
1287 || (GET_CODE (x) == PLUS
1288 && IN_RANGE (INTVAL (XEXP (x, 1)),
1289 -targetm.const_anchor,
1290 targetm.const_anchor - 1)))
1292 match = x;
1293 match_elt = elt;
1294 *old = idx;
1299 return match;
1302 /* Try to express the constant SRC_CONST using a register+offset expression
1303 derived from a constant anchor. Return it if successful or NULL_RTX,
1304 otherwise. */
1306 static rtx
1307 try_const_anchors (rtx src_const, machine_mode mode)
1309 struct table_elt *lower_elt, *upper_elt;
1310 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1311 rtx lower_anchor_rtx, upper_anchor_rtx;
1312 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1313 unsigned lower_old, upper_old;
1315 /* CONST_INT is used for CC modes, but we should leave those alone. */
1316 if (GET_MODE_CLASS (mode) == MODE_CC)
1317 return NULL_RTX;
1319 gcc_assert (SCALAR_INT_MODE_P (mode));
1320 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1321 &upper_base, &upper_offs))
1322 return NULL_RTX;
1324 lower_anchor_rtx = GEN_INT (lower_base);
1325 upper_anchor_rtx = GEN_INT (upper_base);
1326 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1327 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1329 if (lower_elt)
1330 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1331 if (upper_elt)
1332 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1334 if (!lower_exp)
1335 return upper_exp;
1336 if (!upper_exp)
1337 return lower_exp;
1339 /* Return the older expression. */
1340 return (upper_old > lower_old ? upper_exp : lower_exp);
1343 /* Look in or update the hash table. */
1345 /* Remove table element ELT from use in the table.
1346 HASH is its hash code, made using the HASH macro.
1347 It's an argument because often that is known in advance
1348 and we save much time not recomputing it. */
1350 static void
1351 remove_from_table (struct table_elt *elt, unsigned int hash)
1353 if (elt == 0)
1354 return;
1356 /* Mark this element as removed. See cse_insn. */
1357 elt->first_same_value = 0;
1359 /* Remove the table element from its equivalence class. */
1362 struct table_elt *prev = elt->prev_same_value;
1363 struct table_elt *next = elt->next_same_value;
1365 if (next)
1366 next->prev_same_value = prev;
1368 if (prev)
1369 prev->next_same_value = next;
1370 else
1372 struct table_elt *newfirst = next;
1373 while (next)
1375 next->first_same_value = newfirst;
1376 next = next->next_same_value;
1381 /* Remove the table element from its hash bucket. */
1384 struct table_elt *prev = elt->prev_same_hash;
1385 struct table_elt *next = elt->next_same_hash;
1387 if (next)
1388 next->prev_same_hash = prev;
1390 if (prev)
1391 prev->next_same_hash = next;
1392 else if (table[hash] == elt)
1393 table[hash] = next;
1394 else
1396 /* This entry is not in the proper hash bucket. This can happen
1397 when two classes were merged by `merge_equiv_classes'. Search
1398 for the hash bucket that it heads. This happens only very
1399 rarely, so the cost is acceptable. */
1400 for (hash = 0; hash < HASH_SIZE; hash++)
1401 if (table[hash] == elt)
1402 table[hash] = next;
1406 /* Remove the table element from its related-value circular chain. */
1408 if (elt->related_value != 0 && elt->related_value != elt)
1410 struct table_elt *p = elt->related_value;
1412 while (p->related_value != elt)
1413 p = p->related_value;
1414 p->related_value = elt->related_value;
1415 if (p->related_value == p)
1416 p->related_value = 0;
1419 /* Now add it to the free element chain. */
1420 elt->next_same_hash = free_element_chain;
1421 free_element_chain = elt;
1424 /* Same as above, but X is a pseudo-register. */
1426 static void
1427 remove_pseudo_from_table (rtx x, unsigned int hash)
1429 struct table_elt *elt;
1431 /* Because a pseudo-register can be referenced in more than one
1432 mode, we might have to remove more than one table entry. */
1433 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1434 remove_from_table (elt, hash);
1437 /* Look up X in the hash table and return its table element,
1438 or 0 if X is not in the table.
1440 MODE is the machine-mode of X, or if X is an integer constant
1441 with VOIDmode then MODE is the mode with which X will be used.
1443 Here we are satisfied to find an expression whose tree structure
1444 looks like X. */
1446 static struct table_elt *
1447 lookup (rtx x, unsigned int hash, machine_mode mode)
1449 struct table_elt *p;
1451 for (p = table[hash]; p; p = p->next_same_hash)
1452 if (mode == p->mode && ((x == p->exp && REG_P (x))
1453 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1454 return p;
1456 return 0;
1459 /* Like `lookup' but don't care whether the table element uses invalid regs.
1460 Also ignore discrepancies in the machine mode of a register. */
1462 static struct table_elt *
1463 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1465 struct table_elt *p;
1467 if (REG_P (x))
1469 unsigned int regno = REGNO (x);
1471 /* Don't check the machine mode when comparing registers;
1472 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1473 for (p = table[hash]; p; p = p->next_same_hash)
1474 if (REG_P (p->exp)
1475 && REGNO (p->exp) == regno)
1476 return p;
1478 else
1480 for (p = table[hash]; p; p = p->next_same_hash)
1481 if (mode == p->mode
1482 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1483 return p;
1486 return 0;
1489 /* Look for an expression equivalent to X and with code CODE.
1490 If one is found, return that expression. */
1492 static rtx
1493 lookup_as_function (rtx x, enum rtx_code code)
1495 struct table_elt *p
1496 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1498 if (p == 0)
1499 return 0;
1501 for (p = p->first_same_value; p; p = p->next_same_value)
1502 if (GET_CODE (p->exp) == code
1503 /* Make sure this is a valid entry in the table. */
1504 && exp_equiv_p (p->exp, p->exp, 1, false))
1505 return p->exp;
1507 return 0;
1510 /* Insert X in the hash table, assuming HASH is its hash code and
1511 CLASSP is an element of the class it should go in (or 0 if a new
1512 class should be made). COST is the code of X and reg_cost is the
1513 cost of registers in X. It is inserted at the proper position to
1514 keep the class in the order cheapest first.
1516 MODE is the machine-mode of X, or if X is an integer constant
1517 with VOIDmode then MODE is the mode with which X will be used.
1519 For elements of equal cheapness, the most recent one
1520 goes in front, except that the first element in the list
1521 remains first unless a cheaper element is added. The order of
1522 pseudo-registers does not matter, as canon_reg will be called to
1523 find the cheapest when a register is retrieved from the table.
1525 The in_memory field in the hash table element is set to 0.
1526 The caller must set it nonzero if appropriate.
1528 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1529 and if insert_regs returns a nonzero value
1530 you must then recompute its hash code before calling here.
1532 If necessary, update table showing constant values of quantities. */
1534 static struct table_elt *
1535 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1536 machine_mode mode, int cost, int reg_cost)
1538 struct table_elt *elt;
1540 /* If X is a register and we haven't made a quantity for it,
1541 something is wrong. */
1542 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1544 /* If X is a hard register, show it is being put in the table. */
1545 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1546 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1548 /* Put an element for X into the right hash bucket. */
1550 elt = free_element_chain;
1551 if (elt)
1552 free_element_chain = elt->next_same_hash;
1553 else
1554 elt = XNEW (struct table_elt);
1556 elt->exp = x;
1557 elt->canon_exp = NULL_RTX;
1558 elt->cost = cost;
1559 elt->regcost = reg_cost;
1560 elt->next_same_value = 0;
1561 elt->prev_same_value = 0;
1562 elt->next_same_hash = table[hash];
1563 elt->prev_same_hash = 0;
1564 elt->related_value = 0;
1565 elt->in_memory = 0;
1566 elt->mode = mode;
1567 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1569 if (table[hash])
1570 table[hash]->prev_same_hash = elt;
1571 table[hash] = elt;
1573 /* Put it into the proper value-class. */
1574 if (classp)
1576 classp = classp->first_same_value;
1577 if (CHEAPER (elt, classp))
1578 /* Insert at the head of the class. */
1580 struct table_elt *p;
1581 elt->next_same_value = classp;
1582 classp->prev_same_value = elt;
1583 elt->first_same_value = elt;
1585 for (p = classp; p; p = p->next_same_value)
1586 p->first_same_value = elt;
1588 else
1590 /* Insert not at head of the class. */
1591 /* Put it after the last element cheaper than X. */
1592 struct table_elt *p, *next;
1594 for (p = classp;
1595 (next = p->next_same_value) && CHEAPER (next, elt);
1596 p = next)
1599 /* Put it after P and before NEXT. */
1600 elt->next_same_value = next;
1601 if (next)
1602 next->prev_same_value = elt;
1604 elt->prev_same_value = p;
1605 p->next_same_value = elt;
1606 elt->first_same_value = classp;
1609 else
1610 elt->first_same_value = elt;
1612 /* If this is a constant being set equivalent to a register or a register
1613 being set equivalent to a constant, note the constant equivalence.
1615 If this is a constant, it cannot be equivalent to a different constant,
1616 and a constant is the only thing that can be cheaper than a register. So
1617 we know the register is the head of the class (before the constant was
1618 inserted).
1620 If this is a register that is not already known equivalent to a
1621 constant, we must check the entire class.
1623 If this is a register that is already known equivalent to an insn,
1624 update the qtys `const_insn' to show that `this_insn' is the latest
1625 insn making that quantity equivalent to the constant. */
1627 if (elt->is_const && classp && REG_P (classp->exp)
1628 && !REG_P (x))
1630 int exp_q = REG_QTY (REGNO (classp->exp));
1631 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1633 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1634 exp_ent->const_insn = this_insn;
1637 else if (REG_P (x)
1638 && classp
1639 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1640 && ! elt->is_const)
1642 struct table_elt *p;
1644 for (p = classp; p != 0; p = p->next_same_value)
1646 if (p->is_const && !REG_P (p->exp))
1648 int x_q = REG_QTY (REGNO (x));
1649 struct qty_table_elem *x_ent = &qty_table[x_q];
1651 x_ent->const_rtx
1652 = gen_lowpart (GET_MODE (x), p->exp);
1653 x_ent->const_insn = this_insn;
1654 break;
1659 else if (REG_P (x)
1660 && qty_table[REG_QTY (REGNO (x))].const_rtx
1661 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1662 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1664 /* If this is a constant with symbolic value,
1665 and it has a term with an explicit integer value,
1666 link it up with related expressions. */
1667 if (GET_CODE (x) == CONST)
1669 rtx subexp = get_related_value (x);
1670 unsigned subhash;
1671 struct table_elt *subelt, *subelt_prev;
1673 if (subexp != 0)
1675 /* Get the integer-free subexpression in the hash table. */
1676 subhash = SAFE_HASH (subexp, mode);
1677 subelt = lookup (subexp, subhash, mode);
1678 if (subelt == 0)
1679 subelt = insert (subexp, NULL, subhash, mode);
1680 /* Initialize SUBELT's circular chain if it has none. */
1681 if (subelt->related_value == 0)
1682 subelt->related_value = subelt;
1683 /* Find the element in the circular chain that precedes SUBELT. */
1684 subelt_prev = subelt;
1685 while (subelt_prev->related_value != subelt)
1686 subelt_prev = subelt_prev->related_value;
1687 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1688 This way the element that follows SUBELT is the oldest one. */
1689 elt->related_value = subelt_prev->related_value;
1690 subelt_prev->related_value = elt;
1694 return elt;
1697 /* Wrap insert_with_costs by passing the default costs. */
1699 static struct table_elt *
1700 insert (rtx x, struct table_elt *classp, unsigned int hash,
1701 machine_mode mode)
1703 return insert_with_costs (x, classp, hash, mode,
1704 COST (x, mode), approx_reg_cost (x));
1708 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1709 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1710 the two classes equivalent.
1712 CLASS1 will be the surviving class; CLASS2 should not be used after this
1713 call.
1715 Any invalid entries in CLASS2 will not be copied. */
1717 static void
1718 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1720 struct table_elt *elt, *next, *new_elt;
1722 /* Ensure we start with the head of the classes. */
1723 class1 = class1->first_same_value;
1724 class2 = class2->first_same_value;
1726 /* If they were already equal, forget it. */
1727 if (class1 == class2)
1728 return;
1730 for (elt = class2; elt; elt = next)
1732 unsigned int hash;
1733 rtx exp = elt->exp;
1734 machine_mode mode = elt->mode;
1736 next = elt->next_same_value;
1738 /* Remove old entry, make a new one in CLASS1's class.
1739 Don't do this for invalid entries as we cannot find their
1740 hash code (it also isn't necessary). */
1741 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1743 bool need_rehash = false;
1745 hash_arg_in_memory = 0;
1746 hash = HASH (exp, mode);
1748 if (REG_P (exp))
1750 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1751 delete_reg_equiv (REGNO (exp));
1754 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1755 remove_pseudo_from_table (exp, hash);
1756 else
1757 remove_from_table (elt, hash);
1759 if (insert_regs (exp, class1, false) || need_rehash)
1761 rehash_using_reg (exp);
1762 hash = HASH (exp, mode);
1764 new_elt = insert (exp, class1, hash, mode);
1765 new_elt->in_memory = hash_arg_in_memory;
1766 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1767 new_elt->cost = MAX_COST;
1772 /* Flush the entire hash table. */
1774 static void
1775 flush_hash_table (void)
1777 int i;
1778 struct table_elt *p;
1780 for (i = 0; i < HASH_SIZE; i++)
1781 for (p = table[i]; p; p = table[i])
1783 /* Note that invalidate can remove elements
1784 after P in the current hash chain. */
1785 if (REG_P (p->exp))
1786 invalidate (p->exp, VOIDmode);
1787 else
1788 remove_from_table (p, i);
1792 /* Check whether an anti dependence exists between X and EXP. MODE and
1793 ADDR are as for canon_anti_dependence. */
1795 static bool
1796 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1798 subrtx_iterator::array_type array;
1799 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1801 const_rtx x = *iter;
1802 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1803 return true;
1805 return false;
1808 /* Remove from the hash table, or mark as invalid, all expressions whose
1809 values could be altered by storing in register X. */
1811 static void
1812 invalidate_reg (rtx x)
1814 gcc_assert (GET_CODE (x) == REG);
1816 /* If X is a register, dependencies on its contents are recorded
1817 through the qty number mechanism. Just change the qty number of
1818 the register, mark it as invalid for expressions that refer to it,
1819 and remove it itself. */
1820 unsigned int regno = REGNO (x);
1821 unsigned int hash = HASH (x, GET_MODE (x));
1823 /* Remove REGNO from any quantity list it might be on and indicate
1824 that its value might have changed. If it is a pseudo, remove its
1825 entry from the hash table.
1827 For a hard register, we do the first two actions above for any
1828 additional hard registers corresponding to X. Then, if any of these
1829 registers are in the table, we must remove any REG entries that
1830 overlap these registers. */
1832 delete_reg_equiv (regno);
1833 REG_TICK (regno)++;
1834 SUBREG_TICKED (regno) = -1;
1836 if (regno >= FIRST_PSEUDO_REGISTER)
1837 remove_pseudo_from_table (x, hash);
1838 else
1840 HOST_WIDE_INT in_table = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1841 unsigned int endregno = END_REGNO (x);
1842 unsigned int rn;
1843 struct table_elt *p, *next;
1845 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1847 for (rn = regno + 1; rn < endregno; rn++)
1849 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1850 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1851 delete_reg_equiv (rn);
1852 REG_TICK (rn)++;
1853 SUBREG_TICKED (rn) = -1;
1856 if (in_table)
1857 for (hash = 0; hash < HASH_SIZE; hash++)
1858 for (p = table[hash]; p; p = next)
1860 next = p->next_same_hash;
1862 if (!REG_P (p->exp) || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1863 continue;
1865 unsigned int tregno = REGNO (p->exp);
1866 unsigned int tendregno = END_REGNO (p->exp);
1867 if (tendregno > regno && tregno < endregno)
1868 remove_from_table (p, hash);
1873 /* Remove from the hash table, or mark as invalid, all expressions whose
1874 values could be altered by storing in X. X is a register, a subreg, or
1875 a memory reference with nonvarying address (because, when a memory
1876 reference with a varying address is stored in, all memory references are
1877 removed by invalidate_memory so specific invalidation is superfluous).
1878 FULL_MODE, if not VOIDmode, indicates that this much should be
1879 invalidated instead of just the amount indicated by the mode of X. This
1880 is only used for bitfield stores into memory.
1882 A nonvarying address may be just a register or just a symbol reference,
1883 or it may be either of those plus a numeric offset. */
1885 static void
1886 invalidate (rtx x, machine_mode full_mode)
1888 int i;
1889 struct table_elt *p;
1890 rtx addr;
1892 switch (GET_CODE (x))
1894 case REG:
1895 invalidate_reg (x);
1896 return;
1898 case SUBREG:
1899 invalidate (SUBREG_REG (x), VOIDmode);
1900 return;
1902 case PARALLEL:
1903 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1904 invalidate (XVECEXP (x, 0, i), VOIDmode);
1905 return;
1907 case EXPR_LIST:
1908 /* This is part of a disjoint return value; extract the location in
1909 question ignoring the offset. */
1910 invalidate (XEXP (x, 0), VOIDmode);
1911 return;
1913 case MEM:
1914 addr = canon_rtx (get_addr (XEXP (x, 0)));
1915 /* Calculate the canonical version of X here so that
1916 true_dependence doesn't generate new RTL for X on each call. */
1917 x = canon_rtx (x);
1919 /* Remove all hash table elements that refer to overlapping pieces of
1920 memory. */
1921 if (full_mode == VOIDmode)
1922 full_mode = GET_MODE (x);
1924 for (i = 0; i < HASH_SIZE; i++)
1926 struct table_elt *next;
1928 for (p = table[i]; p; p = next)
1930 next = p->next_same_hash;
1931 if (p->in_memory)
1933 /* Just canonicalize the expression once;
1934 otherwise each time we call invalidate
1935 true_dependence will canonicalize the
1936 expression again. */
1937 if (!p->canon_exp)
1938 p->canon_exp = canon_rtx (p->exp);
1939 if (check_dependence (p->canon_exp, x, full_mode, addr))
1940 remove_from_table (p, i);
1944 return;
1946 default:
1947 gcc_unreachable ();
1951 /* Invalidate DEST. Used when DEST is not going to be added
1952 into the hash table for some reason, e.g. do_not_record
1953 flagged on it. */
1955 static void
1956 invalidate_dest (rtx dest)
1958 if (REG_P (dest)
1959 || GET_CODE (dest) == SUBREG
1960 || MEM_P (dest))
1961 invalidate (dest, VOIDmode);
1962 else if (GET_CODE (dest) == STRICT_LOW_PART
1963 || GET_CODE (dest) == ZERO_EXTRACT)
1964 invalidate (XEXP (dest, 0), GET_MODE (dest));
1967 /* Remove all expressions that refer to register REGNO,
1968 since they are already invalid, and we are about to
1969 mark that register valid again and don't want the old
1970 expressions to reappear as valid. */
1972 static void
1973 remove_invalid_refs (unsigned int regno)
1975 unsigned int i;
1976 struct table_elt *p, *next;
1978 for (i = 0; i < HASH_SIZE; i++)
1979 for (p = table[i]; p; p = next)
1981 next = p->next_same_hash;
1982 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1983 remove_from_table (p, i);
1987 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1988 and mode MODE. */
1989 static void
1990 remove_invalid_subreg_refs (unsigned int regno, poly_uint64 offset,
1991 machine_mode mode)
1993 unsigned int i;
1994 struct table_elt *p, *next;
1996 for (i = 0; i < HASH_SIZE; i++)
1997 for (p = table[i]; p; p = next)
1999 rtx exp = p->exp;
2000 next = p->next_same_hash;
2002 if (!REG_P (exp)
2003 && (GET_CODE (exp) != SUBREG
2004 || !REG_P (SUBREG_REG (exp))
2005 || REGNO (SUBREG_REG (exp)) != regno
2006 || ranges_maybe_overlap_p (SUBREG_BYTE (exp),
2007 GET_MODE_SIZE (GET_MODE (exp)),
2008 offset, GET_MODE_SIZE (mode)))
2009 && refers_to_regno_p (regno, p->exp))
2010 remove_from_table (p, i);
2014 /* Recompute the hash codes of any valid entries in the hash table that
2015 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2017 This is called when we make a jump equivalence. */
2019 static void
2020 rehash_using_reg (rtx x)
2022 unsigned int i;
2023 struct table_elt *p, *next;
2024 unsigned hash;
2026 if (GET_CODE (x) == SUBREG)
2027 x = SUBREG_REG (x);
2029 /* If X is not a register or if the register is known not to be in any
2030 valid entries in the table, we have no work to do. */
2032 if (!REG_P (x)
2033 || REG_IN_TABLE (REGNO (x)) < 0
2034 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2035 return;
2037 /* Scan all hash chains looking for valid entries that mention X.
2038 If we find one and it is in the wrong hash chain, move it. */
2040 for (i = 0; i < HASH_SIZE; i++)
2041 for (p = table[i]; p; p = next)
2043 next = p->next_same_hash;
2044 if (reg_mentioned_p (x, p->exp)
2045 && exp_equiv_p (p->exp, p->exp, 1, false)
2046 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2048 if (p->next_same_hash)
2049 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2051 if (p->prev_same_hash)
2052 p->prev_same_hash->next_same_hash = p->next_same_hash;
2053 else
2054 table[i] = p->next_same_hash;
2056 p->next_same_hash = table[hash];
2057 p->prev_same_hash = 0;
2058 if (table[hash])
2059 table[hash]->prev_same_hash = p;
2060 table[hash] = p;
2065 /* Remove from the hash table any expression that is a call-clobbered
2066 register in INSN. Also update their TICK values. */
2068 static void
2069 invalidate_for_call (rtx_insn *insn)
2071 unsigned int regno;
2072 unsigned hash;
2073 struct table_elt *p, *next;
2074 int in_table = 0;
2075 hard_reg_set_iterator hrsi;
2077 /* Go through all the hard registers. For each that might be clobbered
2078 in call insn INSN, remove the register from quantity chains and update
2079 reg_tick if defined. Also see if any of these registers is currently
2080 in the table.
2082 ??? We could be more precise for partially-clobbered registers,
2083 and only invalidate values that actually occupy the clobbered part
2084 of the registers. It doesn't seem worth the effort though, since
2085 we shouldn't see this situation much before RA. Whatever choice
2086 we make here has to be consistent with the table walk below,
2087 so any change to this test will require a change there too. */
2088 HARD_REG_SET callee_clobbers
2089 = insn_callee_abi (insn).full_and_partial_reg_clobbers ();
2090 EXECUTE_IF_SET_IN_HARD_REG_SET (callee_clobbers, 0, regno, hrsi)
2092 delete_reg_equiv (regno);
2093 if (REG_TICK (regno) >= 0)
2095 REG_TICK (regno)++;
2096 SUBREG_TICKED (regno) = -1;
2098 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2101 /* In the case where we have no call-clobbered hard registers in the
2102 table, we are done. Otherwise, scan the table and remove any
2103 entry that overlaps a call-clobbered register. */
2105 if (in_table)
2106 for (hash = 0; hash < HASH_SIZE; hash++)
2107 for (p = table[hash]; p; p = next)
2109 next = p->next_same_hash;
2111 if (!REG_P (p->exp)
2112 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2113 continue;
2115 /* This must use the same test as above rather than the
2116 more accurate clobbers_reg_p. */
2117 if (overlaps_hard_reg_set_p (callee_clobbers, GET_MODE (p->exp),
2118 REGNO (p->exp)))
2119 remove_from_table (p, hash);
2123 /* Given an expression X of type CONST,
2124 and ELT which is its table entry (or 0 if it
2125 is not in the hash table),
2126 return an alternate expression for X as a register plus integer.
2127 If none can be found, return 0. */
2129 static rtx
2130 use_related_value (rtx x, struct table_elt *elt)
2132 struct table_elt *relt = 0;
2133 struct table_elt *p, *q;
2134 HOST_WIDE_INT offset;
2136 /* First, is there anything related known?
2137 If we have a table element, we can tell from that.
2138 Otherwise, must look it up. */
2140 if (elt != 0 && elt->related_value != 0)
2141 relt = elt;
2142 else if (elt == 0 && GET_CODE (x) == CONST)
2144 rtx subexp = get_related_value (x);
2145 if (subexp != 0)
2146 relt = lookup (subexp,
2147 SAFE_HASH (subexp, GET_MODE (subexp)),
2148 GET_MODE (subexp));
2151 if (relt == 0)
2152 return 0;
2154 /* Search all related table entries for one that has an
2155 equivalent register. */
2157 p = relt;
2158 while (1)
2160 /* This loop is strange in that it is executed in two different cases.
2161 The first is when X is already in the table. Then it is searching
2162 the RELATED_VALUE list of X's class (RELT). The second case is when
2163 X is not in the table. Then RELT points to a class for the related
2164 value.
2166 Ensure that, whatever case we are in, that we ignore classes that have
2167 the same value as X. */
2169 if (rtx_equal_p (x, p->exp))
2170 q = 0;
2171 else
2172 for (q = p->first_same_value; q; q = q->next_same_value)
2173 if (REG_P (q->exp))
2174 break;
2176 if (q)
2177 break;
2179 p = p->related_value;
2181 /* We went all the way around, so there is nothing to be found.
2182 Alternatively, perhaps RELT was in the table for some other reason
2183 and it has no related values recorded. */
2184 if (p == relt || p == 0)
2185 break;
2188 if (q == 0)
2189 return 0;
2191 offset = (get_integer_term (x) - get_integer_term (p->exp));
2192 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2193 return plus_constant (q->mode, q->exp, offset);
2197 /* Hash a string. Just add its bytes up. */
2198 static inline unsigned
2199 hash_rtx_string (const char *ps)
2201 unsigned hash = 0;
2202 const unsigned char *p = (const unsigned char *) ps;
2204 if (p)
2205 while (*p)
2206 hash += *p++;
2208 return hash;
2211 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2212 When the callback returns true, we continue with the new rtx. */
2214 unsigned
2215 hash_rtx_cb (const_rtx x, machine_mode mode,
2216 int *do_not_record_p, int *hash_arg_in_memory_p,
2217 bool have_reg_qty, hash_rtx_callback_function cb)
2219 int i, j;
2220 unsigned hash = 0;
2221 enum rtx_code code;
2222 const char *fmt;
2223 machine_mode newmode;
2224 rtx newx;
2226 /* Used to turn recursion into iteration. We can't rely on GCC's
2227 tail-recursion elimination since we need to keep accumulating values
2228 in HASH. */
2229 repeat:
2230 if (x == 0)
2231 return hash;
2233 /* Invoke the callback first. */
2234 if (cb != NULL
2235 && ((*cb) (x, mode, &newx, &newmode)))
2237 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2238 hash_arg_in_memory_p, have_reg_qty, cb);
2239 return hash;
2242 code = GET_CODE (x);
2243 switch (code)
2245 case REG:
2247 unsigned int regno = REGNO (x);
2249 if (do_not_record_p && !reload_completed)
2251 /* On some machines, we can't record any non-fixed hard register,
2252 because extending its life will cause reload problems. We
2253 consider ap, fp, sp, gp to be fixed for this purpose.
2255 We also consider CCmode registers to be fixed for this purpose;
2256 failure to do so leads to failure to simplify 0<100 type of
2257 conditionals.
2259 On all machines, we can't record any global registers.
2260 Nor should we record any register that is in a small
2261 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2262 bool record;
2264 if (regno >= FIRST_PSEUDO_REGISTER)
2265 record = true;
2266 else if (x == frame_pointer_rtx
2267 || x == hard_frame_pointer_rtx
2268 || x == arg_pointer_rtx
2269 || x == stack_pointer_rtx
2270 || x == pic_offset_table_rtx)
2271 record = true;
2272 else if (global_regs[regno])
2273 record = false;
2274 else if (fixed_regs[regno])
2275 record = true;
2276 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2277 record = true;
2278 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2279 record = false;
2280 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2281 record = false;
2282 else
2283 record = true;
2285 if (!record)
2287 *do_not_record_p = 1;
2288 return 0;
2292 hash += ((unsigned int) REG << 7);
2293 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2294 return hash;
2297 /* We handle SUBREG of a REG specially because the underlying
2298 reg changes its hash value with every value change; we don't
2299 want to have to forget unrelated subregs when one subreg changes. */
2300 case SUBREG:
2302 if (REG_P (SUBREG_REG (x)))
2304 hash += (((unsigned int) SUBREG << 7)
2305 + REGNO (SUBREG_REG (x))
2306 + (constant_lower_bound (SUBREG_BYTE (x))
2307 / UNITS_PER_WORD));
2308 return hash;
2310 break;
2313 case CONST_INT:
2314 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2315 + (unsigned int) INTVAL (x));
2316 return hash;
2318 case CONST_WIDE_INT:
2319 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2320 hash += CONST_WIDE_INT_ELT (x, i);
2321 return hash;
2323 case CONST_POLY_INT:
2325 inchash::hash h;
2326 h.add_int (hash);
2327 for (unsigned int i = 0; i < NUM_POLY_INT_COEFFS; ++i)
2328 h.add_wide_int (CONST_POLY_INT_COEFFS (x)[i]);
2329 return h.end ();
2332 case CONST_DOUBLE:
2333 /* This is like the general case, except that it only counts
2334 the integers representing the constant. */
2335 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2336 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2337 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2338 + (unsigned int) CONST_DOUBLE_HIGH (x));
2339 else
2340 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2341 return hash;
2343 case CONST_FIXED:
2344 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2345 hash += fixed_hash (CONST_FIXED_VALUE (x));
2346 return hash;
2348 case CONST_VECTOR:
2350 int units;
2351 rtx elt;
2353 units = const_vector_encoded_nelts (x);
2355 for (i = 0; i < units; ++i)
2357 elt = CONST_VECTOR_ENCODED_ELT (x, i);
2358 hash += hash_rtx_cb (elt, GET_MODE (elt),
2359 do_not_record_p, hash_arg_in_memory_p,
2360 have_reg_qty, cb);
2363 return hash;
2366 /* Assume there is only one rtx object for any given label. */
2367 case LABEL_REF:
2368 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2369 differences and differences between each stage's debugging dumps. */
2370 hash += (((unsigned int) LABEL_REF << 7)
2371 + CODE_LABEL_NUMBER (label_ref_label (x)));
2372 return hash;
2374 case SYMBOL_REF:
2376 /* Don't hash on the symbol's address to avoid bootstrap differences.
2377 Different hash values may cause expressions to be recorded in
2378 different orders and thus different registers to be used in the
2379 final assembler. This also avoids differences in the dump files
2380 between various stages. */
2381 unsigned int h = 0;
2382 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2384 while (*p)
2385 h += (h << 7) + *p++; /* ??? revisit */
2387 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2388 return hash;
2391 case MEM:
2392 /* We don't record if marked volatile or if BLKmode since we don't
2393 know the size of the move. */
2394 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2396 *do_not_record_p = 1;
2397 return 0;
2399 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2400 *hash_arg_in_memory_p = 1;
2402 /* Now that we have already found this special case,
2403 might as well speed it up as much as possible. */
2404 hash += (unsigned) MEM;
2405 x = XEXP (x, 0);
2406 goto repeat;
2408 case USE:
2409 /* A USE that mentions non-volatile memory needs special
2410 handling since the MEM may be BLKmode which normally
2411 prevents an entry from being made. Pure calls are
2412 marked by a USE which mentions BLKmode memory.
2413 See calls.cc:emit_call_1. */
2414 if (MEM_P (XEXP (x, 0))
2415 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2417 hash += (unsigned) USE;
2418 x = XEXP (x, 0);
2420 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2421 *hash_arg_in_memory_p = 1;
2423 /* Now that we have already found this special case,
2424 might as well speed it up as much as possible. */
2425 hash += (unsigned) MEM;
2426 x = XEXP (x, 0);
2427 goto repeat;
2429 break;
2431 case PRE_DEC:
2432 case PRE_INC:
2433 case POST_DEC:
2434 case POST_INC:
2435 case PRE_MODIFY:
2436 case POST_MODIFY:
2437 case PC:
2438 case CALL:
2439 case UNSPEC_VOLATILE:
2440 if (do_not_record_p) {
2441 *do_not_record_p = 1;
2442 return 0;
2444 else
2445 return hash;
2446 break;
2448 case ASM_OPERANDS:
2449 if (do_not_record_p && MEM_VOLATILE_P (x))
2451 *do_not_record_p = 1;
2452 return 0;
2454 else
2456 /* We don't want to take the filename and line into account. */
2457 hash += (unsigned) code + (unsigned) GET_MODE (x)
2458 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2459 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2460 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2462 if (ASM_OPERANDS_INPUT_LENGTH (x))
2464 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2466 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2467 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2468 do_not_record_p, hash_arg_in_memory_p,
2469 have_reg_qty, cb)
2470 + hash_rtx_string
2471 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2474 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2475 x = ASM_OPERANDS_INPUT (x, 0);
2476 mode = GET_MODE (x);
2477 goto repeat;
2480 return hash;
2482 break;
2484 default:
2485 break;
2488 i = GET_RTX_LENGTH (code) - 1;
2489 hash += (unsigned) code + (unsigned) GET_MODE (x);
2490 fmt = GET_RTX_FORMAT (code);
2491 for (; i >= 0; i--)
2493 switch (fmt[i])
2495 case 'e':
2496 /* If we are about to do the last recursive call
2497 needed at this level, change it into iteration.
2498 This function is called enough to be worth it. */
2499 if (i == 0)
2501 x = XEXP (x, i);
2502 goto repeat;
2505 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2506 hash_arg_in_memory_p,
2507 have_reg_qty, cb);
2508 break;
2510 case 'E':
2511 for (j = 0; j < XVECLEN (x, i); j++)
2512 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2513 hash_arg_in_memory_p,
2514 have_reg_qty, cb);
2515 break;
2517 case 's':
2518 hash += hash_rtx_string (XSTR (x, i));
2519 break;
2521 case 'i':
2522 hash += (unsigned int) XINT (x, i);
2523 break;
2525 case 'p':
2526 hash += constant_lower_bound (SUBREG_BYTE (x));
2527 break;
2529 case '0': case 't':
2530 /* Unused. */
2531 break;
2533 default:
2534 gcc_unreachable ();
2538 return hash;
2541 /* Hash an rtx. We are careful to make sure the value is never negative.
2542 Equivalent registers hash identically.
2543 MODE is used in hashing for CONST_INTs only;
2544 otherwise the mode of X is used.
2546 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2548 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2549 a MEM rtx which does not have the MEM_READONLY_P flag set.
2551 Note that cse_insn knows that the hash code of a MEM expression
2552 is just (int) MEM plus the hash code of the address. */
2554 unsigned
2555 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2556 int *hash_arg_in_memory_p, bool have_reg_qty)
2558 return hash_rtx_cb (x, mode, do_not_record_p,
2559 hash_arg_in_memory_p, have_reg_qty, NULL);
2562 /* Hash an rtx X for cse via hash_rtx.
2563 Stores 1 in do_not_record if any subexpression is volatile.
2564 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2565 does not have the MEM_READONLY_P flag set. */
2567 static inline unsigned
2568 canon_hash (rtx x, machine_mode mode)
2570 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2573 /* Like canon_hash but with no side effects, i.e. do_not_record
2574 and hash_arg_in_memory are not changed. */
2576 static inline unsigned
2577 safe_hash (rtx x, machine_mode mode)
2579 int dummy_do_not_record;
2580 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2583 /* Return true iff X and Y would canonicalize into the same thing,
2584 without actually constructing the canonicalization of either one.
2585 If VALIDATE is nonzero,
2586 we assume X is an expression being processed from the rtl
2587 and Y was found in the hash table. We check register refs
2588 in Y for being marked as valid.
2590 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2592 bool
2593 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2595 int i, j;
2596 enum rtx_code code;
2597 const char *fmt;
2599 /* Note: it is incorrect to assume an expression is equivalent to itself
2600 if VALIDATE is nonzero. */
2601 if (x == y && !validate)
2602 return true;
2604 if (x == 0 || y == 0)
2605 return x == y;
2607 code = GET_CODE (x);
2608 if (code != GET_CODE (y))
2609 return false;
2611 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2612 if (GET_MODE (x) != GET_MODE (y))
2613 return false;
2615 /* MEMs referring to different address space are not equivalent. */
2616 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2617 return false;
2619 switch (code)
2621 case PC:
2622 CASE_CONST_UNIQUE:
2623 return x == y;
2625 case CONST_VECTOR:
2626 if (!same_vector_encodings_p (x, y))
2627 return false;
2628 break;
2630 case LABEL_REF:
2631 return label_ref_label (x) == label_ref_label (y);
2633 case SYMBOL_REF:
2634 return XSTR (x, 0) == XSTR (y, 0);
2636 case REG:
2637 if (for_gcse)
2638 return REGNO (x) == REGNO (y);
2639 else
2641 unsigned int regno = REGNO (y);
2642 unsigned int i;
2643 unsigned int endregno = END_REGNO (y);
2645 /* If the quantities are not the same, the expressions are not
2646 equivalent. If there are and we are not to validate, they
2647 are equivalent. Otherwise, ensure all regs are up-to-date. */
2649 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2650 return false;
2652 if (! validate)
2653 return true;
2655 for (i = regno; i < endregno; i++)
2656 if (REG_IN_TABLE (i) != REG_TICK (i))
2657 return false;
2659 return true;
2662 case MEM:
2663 if (for_gcse)
2665 /* A volatile mem should not be considered equivalent to any
2666 other. */
2667 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2668 return false;
2670 /* Can't merge two expressions in different alias sets, since we
2671 can decide that the expression is transparent in a block when
2672 it isn't, due to it being set with the different alias set.
2674 Also, can't merge two expressions with different MEM_ATTRS.
2675 They could e.g. be two different entities allocated into the
2676 same space on the stack (see e.g. PR25130). In that case, the
2677 MEM addresses can be the same, even though the two MEMs are
2678 absolutely not equivalent.
2680 But because really all MEM attributes should be the same for
2681 equivalent MEMs, we just use the invariant that MEMs that have
2682 the same attributes share the same mem_attrs data structure. */
2683 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2684 return false;
2686 /* If we are handling exceptions, we cannot consider two expressions
2687 with different trapping status as equivalent, because simple_mem
2688 might accept one and reject the other. */
2689 if (cfun->can_throw_non_call_exceptions
2690 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2691 return false;
2693 break;
2695 /* For commutative operations, check both orders. */
2696 case PLUS:
2697 case MULT:
2698 case AND:
2699 case IOR:
2700 case XOR:
2701 case NE:
2702 case EQ:
2703 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2704 validate, for_gcse)
2705 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2706 validate, for_gcse))
2707 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2708 validate, for_gcse)
2709 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2710 validate, for_gcse)));
2712 case ASM_OPERANDS:
2713 /* We don't use the generic code below because we want to
2714 disregard filename and line numbers. */
2716 /* A volatile asm isn't equivalent to any other. */
2717 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2718 return false;
2720 if (GET_MODE (x) != GET_MODE (y)
2721 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2722 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2723 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2724 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2725 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2726 return false;
2728 if (ASM_OPERANDS_INPUT_LENGTH (x))
2730 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2731 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2732 ASM_OPERANDS_INPUT (y, i),
2733 validate, for_gcse)
2734 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2735 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2736 return false;
2739 return true;
2741 default:
2742 break;
2745 /* Compare the elements. If any pair of corresponding elements
2746 fail to match, return 0 for the whole thing. */
2748 fmt = GET_RTX_FORMAT (code);
2749 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2751 switch (fmt[i])
2753 case 'e':
2754 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2755 validate, for_gcse))
2756 return false;
2757 break;
2759 case 'E':
2760 if (XVECLEN (x, i) != XVECLEN (y, i))
2761 return 0;
2762 for (j = 0; j < XVECLEN (x, i); j++)
2763 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2764 validate, for_gcse))
2765 return false;
2766 break;
2768 case 's':
2769 if (strcmp (XSTR (x, i), XSTR (y, i)))
2770 return false;
2771 break;
2773 case 'i':
2774 if (XINT (x, i) != XINT (y, i))
2775 return false;
2776 break;
2778 case 'w':
2779 if (XWINT (x, i) != XWINT (y, i))
2780 return false;
2781 break;
2783 case 'p':
2784 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
2785 return false;
2786 break;
2788 case '0':
2789 case 't':
2790 break;
2792 default:
2793 gcc_unreachable ();
2797 return true;
2800 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2801 the result if necessary. INSN is as for canon_reg. */
2803 static void
2804 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2806 if (*xloc)
2808 rtx new_rtx = canon_reg (*xloc, insn);
2810 /* If replacing pseudo with hard reg or vice versa, ensure the
2811 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2812 gcc_assert (insn && new_rtx);
2813 validate_change (insn, xloc, new_rtx, 1);
2817 /* Canonicalize an expression:
2818 replace each register reference inside it
2819 with the "oldest" equivalent register.
2821 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2822 after we make our substitution. The calls are made with IN_GROUP nonzero
2823 so apply_change_group must be called upon the outermost return from this
2824 function (unless INSN is zero). The result of apply_change_group can
2825 generally be discarded since the changes we are making are optional. */
2827 static rtx
2828 canon_reg (rtx x, rtx_insn *insn)
2830 int i;
2831 enum rtx_code code;
2832 const char *fmt;
2834 if (x == 0)
2835 return x;
2837 code = GET_CODE (x);
2838 switch (code)
2840 case PC:
2841 case CONST:
2842 CASE_CONST_ANY:
2843 case SYMBOL_REF:
2844 case LABEL_REF:
2845 case ADDR_VEC:
2846 case ADDR_DIFF_VEC:
2847 return x;
2849 case REG:
2851 int first;
2852 int q;
2853 struct qty_table_elem *ent;
2855 /* Never replace a hard reg, because hard regs can appear
2856 in more than one machine mode, and we must preserve the mode
2857 of each occurrence. Also, some hard regs appear in
2858 MEMs that are shared and mustn't be altered. Don't try to
2859 replace any reg that maps to a reg of class NO_REGS. */
2860 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2861 || ! REGNO_QTY_VALID_P (REGNO (x)))
2862 return x;
2864 q = REG_QTY (REGNO (x));
2865 ent = &qty_table[q];
2866 first = ent->first_reg;
2867 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2868 : REGNO_REG_CLASS (first) == NO_REGS ? x
2869 : gen_rtx_REG (ent->mode, first));
2872 default:
2873 break;
2876 fmt = GET_RTX_FORMAT (code);
2877 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2879 int j;
2881 if (fmt[i] == 'e')
2882 validate_canon_reg (&XEXP (x, i), insn);
2883 else if (fmt[i] == 'E')
2884 for (j = 0; j < XVECLEN (x, i); j++)
2885 validate_canon_reg (&XVECEXP (x, i, j), insn);
2888 return x;
2891 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2892 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2893 what values are being compared.
2895 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2896 actually being compared. For example, if *PARG1 was (reg:CC CC_REG) and
2897 *PARG2 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that
2898 were compared to produce (reg:CC CC_REG).
2900 The return value is the comparison operator and is either the code of
2901 A or the code corresponding to the inverse of the comparison. */
2903 static enum rtx_code
2904 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2905 machine_mode *pmode1, machine_mode *pmode2)
2907 rtx arg1, arg2;
2908 hash_set<rtx> *visited = NULL;
2909 /* Set nonzero when we find something of interest. */
2910 rtx x = NULL;
2912 arg1 = *parg1, arg2 = *parg2;
2914 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2916 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2918 int reverse_code = 0;
2919 struct table_elt *p = 0;
2921 /* Remember state from previous iteration. */
2922 if (x)
2924 if (!visited)
2925 visited = new hash_set<rtx>;
2926 visited->add (x);
2927 x = 0;
2930 /* If arg1 is a COMPARE, extract the comparison arguments from it. */
2932 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2933 x = arg1;
2935 /* If ARG1 is a comparison operator and CODE is testing for
2936 STORE_FLAG_VALUE, get the inner arguments. */
2938 else if (COMPARISON_P (arg1))
2940 #ifdef FLOAT_STORE_FLAG_VALUE
2941 REAL_VALUE_TYPE fsfv;
2942 #endif
2944 if (code == NE
2945 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2946 && code == LT && STORE_FLAG_VALUE == -1)
2947 #ifdef FLOAT_STORE_FLAG_VALUE
2948 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2949 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2950 REAL_VALUE_NEGATIVE (fsfv)))
2951 #endif
2953 x = arg1;
2954 else if (code == EQ
2955 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2956 && code == GE && STORE_FLAG_VALUE == -1)
2957 #ifdef FLOAT_STORE_FLAG_VALUE
2958 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2959 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2960 REAL_VALUE_NEGATIVE (fsfv)))
2961 #endif
2963 x = arg1, reverse_code = 1;
2966 /* ??? We could also check for
2968 (ne (and (eq (...) (const_int 1))) (const_int 0))
2970 and related forms, but let's wait until we see them occurring. */
2972 if (x == 0)
2973 /* Look up ARG1 in the hash table and see if it has an equivalence
2974 that lets us see what is being compared. */
2975 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2976 if (p)
2978 p = p->first_same_value;
2980 /* If what we compare is already known to be constant, that is as
2981 good as it gets.
2982 We need to break the loop in this case, because otherwise we
2983 can have an infinite loop when looking at a reg that is known
2984 to be a constant which is the same as a comparison of a reg
2985 against zero which appears later in the insn stream, which in
2986 turn is constant and the same as the comparison of the first reg
2987 against zero... */
2988 if (p->is_const)
2989 break;
2992 for (; p; p = p->next_same_value)
2994 machine_mode inner_mode = GET_MODE (p->exp);
2995 #ifdef FLOAT_STORE_FLAG_VALUE
2996 REAL_VALUE_TYPE fsfv;
2997 #endif
2999 /* If the entry isn't valid, skip it. */
3000 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3001 continue;
3003 /* If it's a comparison we've used before, skip it. */
3004 if (visited && visited->contains (p->exp))
3005 continue;
3007 if (GET_CODE (p->exp) == COMPARE
3008 /* Another possibility is that this machine has a compare insn
3009 that includes the comparison code. In that case, ARG1 would
3010 be equivalent to a comparison operation that would set ARG1 to
3011 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3012 ORIG_CODE is the actual comparison being done; if it is an EQ,
3013 we must reverse ORIG_CODE. On machine with a negative value
3014 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3015 || ((code == NE
3016 || (code == LT
3017 && val_signbit_known_set_p (inner_mode,
3018 STORE_FLAG_VALUE))
3019 #ifdef FLOAT_STORE_FLAG_VALUE
3020 || (code == LT
3021 && SCALAR_FLOAT_MODE_P (inner_mode)
3022 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3023 REAL_VALUE_NEGATIVE (fsfv)))
3024 #endif
3026 && COMPARISON_P (p->exp)))
3028 x = p->exp;
3029 break;
3031 else if ((code == EQ
3032 || (code == GE
3033 && val_signbit_known_set_p (inner_mode,
3034 STORE_FLAG_VALUE))
3035 #ifdef FLOAT_STORE_FLAG_VALUE
3036 || (code == GE
3037 && SCALAR_FLOAT_MODE_P (inner_mode)
3038 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3039 REAL_VALUE_NEGATIVE (fsfv)))
3040 #endif
3042 && COMPARISON_P (p->exp))
3044 reverse_code = 1;
3045 x = p->exp;
3046 break;
3049 /* If this non-trapping address, e.g. fp + constant, the
3050 equivalent is a better operand since it may let us predict
3051 the value of the comparison. */
3052 else if (!rtx_addr_can_trap_p (p->exp))
3054 arg1 = p->exp;
3055 continue;
3059 /* If we didn't find a useful equivalence for ARG1, we are done.
3060 Otherwise, set up for the next iteration. */
3061 if (x == 0)
3062 break;
3064 /* If we need to reverse the comparison, make sure that is
3065 possible -- we can't necessarily infer the value of GE from LT
3066 with floating-point operands. */
3067 if (reverse_code)
3069 enum rtx_code reversed = reversed_comparison_code (x, NULL);
3070 if (reversed == UNKNOWN)
3071 break;
3072 else
3073 code = reversed;
3075 else if (COMPARISON_P (x))
3076 code = GET_CODE (x);
3077 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3080 /* Return our results. Return the modes from before fold_rtx
3081 because fold_rtx might produce const_int, and then it's too late. */
3082 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3083 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3085 if (visited)
3086 delete visited;
3087 return code;
3090 /* If X is a nontrivial arithmetic operation on an argument for which
3091 a constant value can be determined, return the result of operating
3092 on that value, as a constant. Otherwise, return X, possibly with
3093 one or more operands changed to a forward-propagated constant.
3095 If X is a register whose contents are known, we do NOT return
3096 those contents here; equiv_constant is called to perform that task.
3097 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3099 INSN is the insn that we may be modifying. If it is 0, make a copy
3100 of X before modifying it. */
3102 static rtx
3103 fold_rtx (rtx x, rtx_insn *insn)
3105 enum rtx_code code;
3106 machine_mode mode;
3107 const char *fmt;
3108 int i;
3109 rtx new_rtx = 0;
3110 bool changed = false;
3111 poly_int64 xval;
3113 /* Operands of X. */
3114 /* Workaround -Wmaybe-uninitialized false positive during
3115 profiledbootstrap by initializing them. */
3116 rtx folded_arg0 = NULL_RTX;
3117 rtx folded_arg1 = NULL_RTX;
3119 /* Constant equivalents of first three operands of X;
3120 0 when no such equivalent is known. */
3121 rtx const_arg0;
3122 rtx const_arg1;
3123 rtx const_arg2;
3125 /* The mode of the first operand of X. We need this for sign and zero
3126 extends. */
3127 machine_mode mode_arg0;
3129 if (x == 0)
3130 return x;
3132 /* Try to perform some initial simplifications on X. */
3133 code = GET_CODE (x);
3134 switch (code)
3136 case MEM:
3137 case SUBREG:
3138 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3139 than it would in other contexts. Basically its mode does not
3140 signify the size of the object read. That information is carried
3141 by size operand. If we happen to have a MEM of the appropriate
3142 mode in our tables with a constant value we could simplify the
3143 extraction incorrectly if we allowed substitution of that value
3144 for the MEM. */
3145 case ZERO_EXTRACT:
3146 case SIGN_EXTRACT:
3147 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3148 return new_rtx;
3149 return x;
3151 case CONST:
3152 CASE_CONST_ANY:
3153 case SYMBOL_REF:
3154 case LABEL_REF:
3155 case REG:
3156 case PC:
3157 /* No use simplifying an EXPR_LIST
3158 since they are used only for lists of args
3159 in a function call's REG_EQUAL note. */
3160 case EXPR_LIST:
3161 return x;
3163 case ASM_OPERANDS:
3164 if (insn)
3166 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3167 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3168 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3170 return x;
3172 case CALL:
3173 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3174 return x;
3175 break;
3176 case VEC_SELECT:
3178 rtx trueop0 = XEXP (x, 0);
3179 mode = GET_MODE (trueop0);
3180 rtx trueop1 = XEXP (x, 1);
3181 /* If we select a low-part subreg, return that. */
3182 if (vec_series_lowpart_p (GET_MODE (x), mode, trueop1))
3184 rtx new_rtx = lowpart_subreg (GET_MODE (x), trueop0, mode);
3185 if (new_rtx != NULL_RTX)
3186 return new_rtx;
3190 /* Anything else goes through the loop below. */
3191 default:
3192 break;
3195 mode = GET_MODE (x);
3196 const_arg0 = 0;
3197 const_arg1 = 0;
3198 const_arg2 = 0;
3199 mode_arg0 = VOIDmode;
3201 /* Try folding our operands.
3202 Then see which ones have constant values known. */
3204 fmt = GET_RTX_FORMAT (code);
3205 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3206 if (fmt[i] == 'e')
3208 rtx folded_arg = XEXP (x, i), const_arg;
3209 machine_mode mode_arg = GET_MODE (folded_arg);
3211 switch (GET_CODE (folded_arg))
3213 case MEM:
3214 case REG:
3215 case SUBREG:
3216 const_arg = equiv_constant (folded_arg);
3217 break;
3219 case CONST:
3220 CASE_CONST_ANY:
3221 case SYMBOL_REF:
3222 case LABEL_REF:
3223 const_arg = folded_arg;
3224 break;
3226 default:
3227 folded_arg = fold_rtx (folded_arg, insn);
3228 const_arg = equiv_constant (folded_arg);
3229 break;
3232 /* For the first three operands, see if the operand
3233 is constant or equivalent to a constant. */
3234 switch (i)
3236 case 0:
3237 folded_arg0 = folded_arg;
3238 const_arg0 = const_arg;
3239 mode_arg0 = mode_arg;
3240 break;
3241 case 1:
3242 folded_arg1 = folded_arg;
3243 const_arg1 = const_arg;
3244 break;
3245 case 2:
3246 const_arg2 = const_arg;
3247 break;
3250 /* Pick the least expensive of the argument and an equivalent constant
3251 argument. */
3252 if (const_arg != 0
3253 && const_arg != folded_arg
3254 && (COST_IN (const_arg, mode_arg, code, i)
3255 <= COST_IN (folded_arg, mode_arg, code, i))
3257 /* It's not safe to substitute the operand of a conversion
3258 operator with a constant, as the conversion's identity
3259 depends upon the mode of its operand. This optimization
3260 is handled by the call to simplify_unary_operation. */
3261 && (GET_RTX_CLASS (code) != RTX_UNARY
3262 || GET_MODE (const_arg) == mode_arg0
3263 || (code != ZERO_EXTEND
3264 && code != SIGN_EXTEND
3265 && code != TRUNCATE
3266 && code != FLOAT_TRUNCATE
3267 && code != FLOAT_EXTEND
3268 && code != FLOAT
3269 && code != FIX
3270 && code != UNSIGNED_FLOAT
3271 && code != UNSIGNED_FIX)))
3272 folded_arg = const_arg;
3274 if (folded_arg == XEXP (x, i))
3275 continue;
3277 if (insn == NULL_RTX && !changed)
3278 x = copy_rtx (x);
3279 changed = true;
3280 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3283 if (changed)
3285 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3286 consistent with the order in X. */
3287 if (canonicalize_change_group (insn, x))
3289 std::swap (const_arg0, const_arg1);
3290 std::swap (folded_arg0, folded_arg1);
3293 apply_change_group ();
3296 /* If X is an arithmetic operation, see if we can simplify it. */
3298 switch (GET_RTX_CLASS (code))
3300 case RTX_UNARY:
3302 /* We can't simplify extension ops unless we know the
3303 original mode. */
3304 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3305 && mode_arg0 == VOIDmode)
3306 break;
3308 new_rtx = simplify_unary_operation (code, mode,
3309 const_arg0 ? const_arg0 : folded_arg0,
3310 mode_arg0);
3312 break;
3314 case RTX_COMPARE:
3315 case RTX_COMM_COMPARE:
3316 /* See what items are actually being compared and set FOLDED_ARG[01]
3317 to those values and CODE to the actual comparison code. If any are
3318 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3319 do anything if both operands are already known to be constant. */
3321 /* ??? Vector mode comparisons are not supported yet. */
3322 if (VECTOR_MODE_P (mode))
3323 break;
3325 if (const_arg0 == 0 || const_arg1 == 0)
3327 struct table_elt *p0, *p1;
3328 rtx true_rtx, false_rtx;
3329 machine_mode mode_arg1;
3331 if (SCALAR_FLOAT_MODE_P (mode))
3333 #ifdef FLOAT_STORE_FLAG_VALUE
3334 true_rtx = (const_double_from_real_value
3335 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3336 #else
3337 true_rtx = NULL_RTX;
3338 #endif
3339 false_rtx = CONST0_RTX (mode);
3341 else
3343 true_rtx = const_true_rtx;
3344 false_rtx = const0_rtx;
3347 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3348 &mode_arg0, &mode_arg1);
3350 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3351 what kinds of things are being compared, so we can't do
3352 anything with this comparison. */
3354 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3355 break;
3357 const_arg0 = equiv_constant (folded_arg0);
3358 const_arg1 = equiv_constant (folded_arg1);
3360 /* If we do not now have two constants being compared, see
3361 if we can nevertheless deduce some things about the
3362 comparison. */
3363 if (const_arg0 == 0 || const_arg1 == 0)
3365 if (const_arg1 != NULL)
3367 rtx cheapest_simplification;
3368 int cheapest_cost;
3369 rtx simp_result;
3370 struct table_elt *p;
3372 /* See if we can find an equivalent of folded_arg0
3373 that gets us a cheaper expression, possibly a
3374 constant through simplifications. */
3375 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3376 mode_arg0);
3378 if (p != NULL)
3380 cheapest_simplification = x;
3381 cheapest_cost = COST (x, mode);
3383 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3385 int cost;
3387 /* If the entry isn't valid, skip it. */
3388 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3389 continue;
3391 /* Try to simplify using this equivalence. */
3392 simp_result
3393 = simplify_relational_operation (code, mode,
3394 mode_arg0,
3395 p->exp,
3396 const_arg1);
3398 if (simp_result == NULL)
3399 continue;
3401 cost = COST (simp_result, mode);
3402 if (cost < cheapest_cost)
3404 cheapest_cost = cost;
3405 cheapest_simplification = simp_result;
3409 /* If we have a cheaper expression now, use that
3410 and try folding it further, from the top. */
3411 if (cheapest_simplification != x)
3412 return fold_rtx (copy_rtx (cheapest_simplification),
3413 insn);
3417 /* See if the two operands are the same. */
3419 if ((REG_P (folded_arg0)
3420 && REG_P (folded_arg1)
3421 && (REG_QTY (REGNO (folded_arg0))
3422 == REG_QTY (REGNO (folded_arg1))))
3423 || ((p0 = lookup (folded_arg0,
3424 SAFE_HASH (folded_arg0, mode_arg0),
3425 mode_arg0))
3426 && (p1 = lookup (folded_arg1,
3427 SAFE_HASH (folded_arg1, mode_arg0),
3428 mode_arg0))
3429 && p0->first_same_value == p1->first_same_value))
3430 folded_arg1 = folded_arg0;
3432 /* If FOLDED_ARG0 is a register, see if the comparison we are
3433 doing now is either the same as we did before or the reverse
3434 (we only check the reverse if not floating-point). */
3435 else if (REG_P (folded_arg0))
3437 int qty = REG_QTY (REGNO (folded_arg0));
3439 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3441 struct qty_table_elem *ent = &qty_table[qty];
3443 if ((comparison_dominates_p (ent->comparison_code, code)
3444 || (! FLOAT_MODE_P (mode_arg0)
3445 && comparison_dominates_p (ent->comparison_code,
3446 reverse_condition (code))))
3447 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3448 || (const_arg1
3449 && rtx_equal_p (ent->comparison_const,
3450 const_arg1))
3451 || (REG_P (folded_arg1)
3452 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3454 if (comparison_dominates_p (ent->comparison_code, code))
3456 if (true_rtx)
3457 return true_rtx;
3458 else
3459 break;
3461 else
3462 return false_rtx;
3469 /* If we are comparing against zero, see if the first operand is
3470 equivalent to an IOR with a constant. If so, we may be able to
3471 determine the result of this comparison. */
3472 if (const_arg1 == const0_rtx && !const_arg0)
3474 rtx y = lookup_as_function (folded_arg0, IOR);
3475 rtx inner_const;
3477 if (y != 0
3478 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3479 && CONST_INT_P (inner_const)
3480 && INTVAL (inner_const) != 0)
3481 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3485 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3486 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3487 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3488 op0, op1);
3490 break;
3492 case RTX_BIN_ARITH:
3493 case RTX_COMM_ARITH:
3494 switch (code)
3496 case PLUS:
3497 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3498 with that LABEL_REF as its second operand. If so, the result is
3499 the first operand of that MINUS. This handles switches with an
3500 ADDR_DIFF_VEC table. */
3501 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3503 rtx y
3504 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3505 : lookup_as_function (folded_arg0, MINUS);
3507 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3508 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg1))
3509 return XEXP (y, 0);
3511 /* Now try for a CONST of a MINUS like the above. */
3512 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3513 : lookup_as_function (folded_arg0, CONST))) != 0
3514 && GET_CODE (XEXP (y, 0)) == MINUS
3515 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3516 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg1))
3517 return XEXP (XEXP (y, 0), 0);
3520 /* Likewise if the operands are in the other order. */
3521 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3523 rtx y
3524 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3525 : lookup_as_function (folded_arg1, MINUS);
3527 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3528 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg0))
3529 return XEXP (y, 0);
3531 /* Now try for a CONST of a MINUS like the above. */
3532 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3533 : lookup_as_function (folded_arg1, CONST))) != 0
3534 && GET_CODE (XEXP (y, 0)) == MINUS
3535 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3536 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg0))
3537 return XEXP (XEXP (y, 0), 0);
3540 /* If second operand is a register equivalent to a negative
3541 CONST_INT, see if we can find a register equivalent to the
3542 positive constant. Make a MINUS if so. Don't do this for
3543 a non-negative constant since we might then alternate between
3544 choosing positive and negative constants. Having the positive
3545 constant previously-used is the more common case. Be sure
3546 the resulting constant is non-negative; if const_arg1 were
3547 the smallest negative number this would overflow: depending
3548 on the mode, this would either just be the same value (and
3549 hence not save anything) or be incorrect. */
3550 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3551 && INTVAL (const_arg1) < 0
3552 /* This used to test
3554 -INTVAL (const_arg1) >= 0
3556 But The Sun V5.0 compilers mis-compiled that test. So
3557 instead we test for the problematic value in a more direct
3558 manner and hope the Sun compilers get it correct. */
3559 && INTVAL (const_arg1) !=
3560 (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1))
3561 && REG_P (folded_arg1))
3563 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3564 struct table_elt *p
3565 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3567 if (p)
3568 for (p = p->first_same_value; p; p = p->next_same_value)
3569 if (REG_P (p->exp))
3570 return simplify_gen_binary (MINUS, mode, folded_arg0,
3571 canon_reg (p->exp, NULL));
3573 goto from_plus;
3575 case MINUS:
3576 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3577 If so, produce (PLUS Z C2-C). */
3578 if (const_arg1 != 0 && poly_int_rtx_p (const_arg1, &xval))
3580 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3581 if (y && poly_int_rtx_p (XEXP (y, 1)))
3582 return fold_rtx (plus_constant (mode, copy_rtx (y), -xval),
3583 NULL);
3586 /* Fall through. */
3588 from_plus:
3589 case SMIN: case SMAX: case UMIN: case UMAX:
3590 case IOR: case AND: case XOR:
3591 case MULT:
3592 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3593 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3594 is known to be of similar form, we may be able to replace the
3595 operation with a combined operation. This may eliminate the
3596 intermediate operation if every use is simplified in this way.
3597 Note that the similar optimization done by combine.cc only works
3598 if the intermediate operation's result has only one reference. */
3600 if (REG_P (folded_arg0)
3601 && const_arg1 && CONST_INT_P (const_arg1))
3603 int is_shift
3604 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3605 rtx y, inner_const, new_const;
3606 rtx canon_const_arg1 = const_arg1;
3607 enum rtx_code associate_code;
3609 if (is_shift
3610 && (INTVAL (const_arg1) >= GET_MODE_UNIT_PRECISION (mode)
3611 || INTVAL (const_arg1) < 0))
3613 if (SHIFT_COUNT_TRUNCATED)
3614 canon_const_arg1 = gen_int_shift_amount
3615 (mode, (INTVAL (const_arg1)
3616 & (GET_MODE_UNIT_BITSIZE (mode) - 1)));
3617 else
3618 break;
3621 y = lookup_as_function (folded_arg0, code);
3622 if (y == 0)
3623 break;
3625 /* If we have compiled a statement like
3626 "if (x == (x & mask1))", and now are looking at
3627 "x & mask2", we will have a case where the first operand
3628 of Y is the same as our first operand. Unless we detect
3629 this case, an infinite loop will result. */
3630 if (XEXP (y, 0) == folded_arg0)
3631 break;
3633 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3634 if (!inner_const || !CONST_INT_P (inner_const))
3635 break;
3637 /* Don't associate these operations if they are a PLUS with the
3638 same constant and it is a power of two. These might be doable
3639 with a pre- or post-increment. Similarly for two subtracts of
3640 identical powers of two with post decrement. */
3642 if (code == PLUS && const_arg1 == inner_const
3643 && ((HAVE_PRE_INCREMENT
3644 && pow2p_hwi (INTVAL (const_arg1)))
3645 || (HAVE_POST_INCREMENT
3646 && pow2p_hwi (INTVAL (const_arg1)))
3647 || (HAVE_PRE_DECREMENT
3648 && pow2p_hwi (- INTVAL (const_arg1)))
3649 || (HAVE_POST_DECREMENT
3650 && pow2p_hwi (- INTVAL (const_arg1)))))
3651 break;
3653 /* ??? Vector mode shifts by scalar
3654 shift operand are not supported yet. */
3655 if (is_shift && VECTOR_MODE_P (mode))
3656 break;
3658 if (is_shift
3659 && (INTVAL (inner_const) >= GET_MODE_UNIT_PRECISION (mode)
3660 || INTVAL (inner_const) < 0))
3662 if (SHIFT_COUNT_TRUNCATED)
3663 inner_const = gen_int_shift_amount
3664 (mode, (INTVAL (inner_const)
3665 & (GET_MODE_UNIT_BITSIZE (mode) - 1)));
3666 else
3667 break;
3670 /* Compute the code used to compose the constants. For example,
3671 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3673 associate_code = (is_shift || code == MINUS ? PLUS : code);
3675 new_const = simplify_binary_operation (associate_code, mode,
3676 canon_const_arg1,
3677 inner_const);
3679 if (new_const == 0)
3680 break;
3682 /* If we are associating shift operations, don't let this
3683 produce a shift of the size of the object or larger.
3684 This could occur when we follow a sign-extend by a right
3685 shift on a machine that does a sign-extend as a pair
3686 of shifts. */
3688 if (is_shift
3689 && CONST_INT_P (new_const)
3690 && INTVAL (new_const) >= GET_MODE_UNIT_PRECISION (mode))
3692 /* As an exception, we can turn an ASHIFTRT of this
3693 form into a shift of the number of bits - 1. */
3694 if (code == ASHIFTRT)
3695 new_const = gen_int_shift_amount
3696 (mode, GET_MODE_UNIT_BITSIZE (mode) - 1);
3697 else if (!side_effects_p (XEXP (y, 0)))
3698 return CONST0_RTX (mode);
3699 else
3700 break;
3703 y = copy_rtx (XEXP (y, 0));
3705 /* If Y contains our first operand (the most common way this
3706 can happen is if Y is a MEM), we would do into an infinite
3707 loop if we tried to fold it. So don't in that case. */
3709 if (! reg_mentioned_p (folded_arg0, y))
3710 y = fold_rtx (y, insn);
3712 return simplify_gen_binary (code, mode, y, new_const);
3714 break;
3716 case DIV: case UDIV:
3717 /* ??? The associative optimization performed immediately above is
3718 also possible for DIV and UDIV using associate_code of MULT.
3719 However, we would need extra code to verify that the
3720 multiplication does not overflow, that is, there is no overflow
3721 in the calculation of new_const. */
3722 break;
3724 default:
3725 break;
3728 new_rtx = simplify_binary_operation (code, mode,
3729 const_arg0 ? const_arg0 : folded_arg0,
3730 const_arg1 ? const_arg1 : folded_arg1);
3731 break;
3733 case RTX_OBJ:
3734 /* (lo_sum (high X) X) is simply X. */
3735 if (code == LO_SUM && const_arg0 != 0
3736 && GET_CODE (const_arg0) == HIGH
3737 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3738 return const_arg1;
3739 break;
3741 case RTX_TERNARY:
3742 case RTX_BITFIELD_OPS:
3743 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3744 const_arg0 ? const_arg0 : folded_arg0,
3745 const_arg1 ? const_arg1 : folded_arg1,
3746 const_arg2 ? const_arg2 : XEXP (x, 2));
3747 break;
3749 default:
3750 break;
3753 return new_rtx ? new_rtx : x;
3756 /* Return a constant value currently equivalent to X.
3757 Return 0 if we don't know one. */
3759 static rtx
3760 equiv_constant (rtx x)
3762 if (REG_P (x)
3763 && REGNO_QTY_VALID_P (REGNO (x)))
3765 int x_q = REG_QTY (REGNO (x));
3766 struct qty_table_elem *x_ent = &qty_table[x_q];
3768 if (x_ent->const_rtx)
3769 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3772 if (x == 0 || CONSTANT_P (x))
3773 return x;
3775 if (GET_CODE (x) == SUBREG)
3777 machine_mode mode = GET_MODE (x);
3778 machine_mode imode = GET_MODE (SUBREG_REG (x));
3779 rtx new_rtx;
3781 /* See if we previously assigned a constant value to this SUBREG. */
3782 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3783 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3784 || (NUM_POLY_INT_COEFFS > 1
3785 && (new_rtx = lookup_as_function (x, CONST_POLY_INT)) != 0)
3786 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3787 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3788 return new_rtx;
3790 /* If we didn't and if doing so makes sense, see if we previously
3791 assigned a constant value to the enclosing word mode SUBREG. */
3792 if (known_lt (GET_MODE_SIZE (mode), UNITS_PER_WORD)
3793 && known_lt (UNITS_PER_WORD, GET_MODE_SIZE (imode)))
3795 poly_int64 byte = (SUBREG_BYTE (x)
3796 - subreg_lowpart_offset (mode, word_mode));
3797 if (known_ge (byte, 0) && multiple_p (byte, UNITS_PER_WORD))
3799 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3800 new_rtx = lookup_as_function (y, CONST_INT);
3801 if (new_rtx)
3802 return gen_lowpart (mode, new_rtx);
3806 /* Otherwise see if we already have a constant for the inner REG,
3807 and if that is enough to calculate an equivalent constant for
3808 the subreg. Note that the upper bits of paradoxical subregs
3809 are undefined, so they cannot be said to equal anything. */
3810 if (REG_P (SUBREG_REG (x))
3811 && !paradoxical_subreg_p (x)
3812 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3813 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3815 return 0;
3818 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3819 the hash table in case its value was seen before. */
3821 if (MEM_P (x))
3823 struct table_elt *elt;
3825 x = avoid_constant_pool_reference (x);
3826 if (CONSTANT_P (x))
3827 return x;
3829 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3830 if (elt == 0)
3831 return 0;
3833 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3834 if (elt->is_const && CONSTANT_P (elt->exp))
3835 return elt->exp;
3838 return 0;
3841 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3842 "taken" branch.
3844 In certain cases, this can cause us to add an equivalence. For example,
3845 if we are following the taken case of
3846 if (i == 2)
3847 we can add the fact that `i' and '2' are now equivalent.
3849 In any case, we can record that this comparison was passed. If the same
3850 comparison is seen later, we will know its value. */
3852 static void
3853 record_jump_equiv (rtx_insn *insn, bool taken)
3855 int cond_known_true;
3856 rtx op0, op1;
3857 rtx set;
3858 machine_mode mode, mode0, mode1;
3859 enum rtx_code code;
3861 /* Ensure this is the right kind of insn. */
3862 gcc_assert (any_condjump_p (insn));
3864 set = pc_set (insn);
3866 /* See if this jump condition is known true or false. */
3867 if (taken)
3868 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3869 else
3870 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3872 /* Get the type of comparison being done and the operands being compared.
3873 If we had to reverse a non-equality condition, record that fact so we
3874 know that it isn't valid for floating-point. */
3875 code = GET_CODE (XEXP (SET_SRC (set), 0));
3876 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3877 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3879 /* If fold_rtx returns NULL_RTX, there's nothing to record. */
3880 if (op0 == NULL_RTX || op1 == NULL_RTX)
3881 return;
3883 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3884 if (! cond_known_true)
3886 code = reversed_comparison_code_parts (code, op0, op1, insn);
3888 /* Don't remember if we can't find the inverse. */
3889 if (code == UNKNOWN)
3890 return;
3893 /* The mode is the mode of the non-constant. */
3894 mode = mode0;
3895 if (mode1 != VOIDmode)
3896 mode = mode1;
3898 record_jump_cond (code, mode, op0, op1);
3901 /* Yet another form of subreg creation. In this case, we want something in
3902 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3904 static rtx
3905 record_jump_cond_subreg (machine_mode mode, rtx op)
3907 machine_mode op_mode = GET_MODE (op);
3908 if (op_mode == mode || op_mode == VOIDmode)
3909 return op;
3910 return lowpart_subreg (mode, op, op_mode);
3913 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3914 Make any useful entries we can with that information. Called from
3915 above function and called recursively. */
3917 static void
3918 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0, rtx op1)
3920 unsigned op0_hash, op1_hash;
3921 int op0_in_memory, op1_in_memory;
3922 struct table_elt *op0_elt, *op1_elt;
3924 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3925 we know that they are also equal in the smaller mode (this is also
3926 true for all smaller modes whether or not there is a SUBREG, but
3927 is not worth testing for with no SUBREG). */
3929 /* Note that GET_MODE (op0) may not equal MODE. */
3930 if (code == EQ && paradoxical_subreg_p (op0))
3932 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3933 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3934 if (tem)
3935 record_jump_cond (code, mode, SUBREG_REG (op0), tem);
3938 if (code == EQ && paradoxical_subreg_p (op1))
3940 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3941 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3942 if (tem)
3943 record_jump_cond (code, mode, SUBREG_REG (op1), tem);
3946 /* Similarly, if this is an NE comparison, and either is a SUBREG
3947 making a smaller mode, we know the whole thing is also NE. */
3949 /* Note that GET_MODE (op0) may not equal MODE;
3950 if we test MODE instead, we can get an infinite recursion
3951 alternating between two modes each wider than MODE. */
3953 if (code == NE
3954 && partial_subreg_p (op0)
3955 && subreg_lowpart_p (op0))
3957 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3958 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3959 if (tem)
3960 record_jump_cond (code, mode, SUBREG_REG (op0), tem);
3963 if (code == NE
3964 && partial_subreg_p (op1)
3965 && subreg_lowpart_p (op1))
3967 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3968 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3969 if (tem)
3970 record_jump_cond (code, mode, SUBREG_REG (op1), tem);
3973 /* Hash both operands. */
3975 do_not_record = 0;
3976 hash_arg_in_memory = 0;
3977 op0_hash = HASH (op0, mode);
3978 op0_in_memory = hash_arg_in_memory;
3980 if (do_not_record)
3981 return;
3983 do_not_record = 0;
3984 hash_arg_in_memory = 0;
3985 op1_hash = HASH (op1, mode);
3986 op1_in_memory = hash_arg_in_memory;
3988 if (do_not_record)
3989 return;
3991 /* Look up both operands. */
3992 op0_elt = lookup (op0, op0_hash, mode);
3993 op1_elt = lookup (op1, op1_hash, mode);
3995 /* If both operands are already equivalent or if they are not in the
3996 table but are identical, do nothing. */
3997 if ((op0_elt != 0 && op1_elt != 0
3998 && op0_elt->first_same_value == op1_elt->first_same_value)
3999 || op0 == op1 || rtx_equal_p (op0, op1))
4000 return;
4002 /* If we aren't setting two things equal all we can do is save this
4003 comparison. Similarly if this is floating-point. In the latter
4004 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4005 If we record the equality, we might inadvertently delete code
4006 whose intent was to change -0 to +0. */
4008 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4010 struct qty_table_elem *ent;
4011 int qty;
4013 /* If OP0 is not a register, or if OP1 is neither a register
4014 or constant, we can't do anything. */
4016 if (!REG_P (op1))
4017 op1 = equiv_constant (op1);
4019 if (!REG_P (op0) || op1 == 0)
4020 return;
4022 /* Put OP0 in the hash table if it isn't already. This gives it a
4023 new quantity number. */
4024 if (op0_elt == 0)
4026 if (insert_regs (op0, NULL, false))
4028 rehash_using_reg (op0);
4029 op0_hash = HASH (op0, mode);
4031 /* If OP0 is contained in OP1, this changes its hash code
4032 as well. Faster to rehash than to check, except
4033 for the simple case of a constant. */
4034 if (! CONSTANT_P (op1))
4035 op1_hash = HASH (op1,mode);
4038 op0_elt = insert (op0, NULL, op0_hash, mode);
4039 op0_elt->in_memory = op0_in_memory;
4042 qty = REG_QTY (REGNO (op0));
4043 ent = &qty_table[qty];
4045 ent->comparison_code = code;
4046 if (REG_P (op1))
4048 /* Look it up again--in case op0 and op1 are the same. */
4049 op1_elt = lookup (op1, op1_hash, mode);
4051 /* Put OP1 in the hash table so it gets a new quantity number. */
4052 if (op1_elt == 0)
4054 if (insert_regs (op1, NULL, false))
4056 rehash_using_reg (op1);
4057 op1_hash = HASH (op1, mode);
4060 op1_elt = insert (op1, NULL, op1_hash, mode);
4061 op1_elt->in_memory = op1_in_memory;
4064 ent->comparison_const = NULL_RTX;
4065 ent->comparison_qty = REG_QTY (REGNO (op1));
4067 else
4069 ent->comparison_const = op1;
4070 ent->comparison_qty = -1;
4073 return;
4076 /* If either side is still missing an equivalence, make it now,
4077 then merge the equivalences. */
4079 if (op0_elt == 0)
4081 if (insert_regs (op0, NULL, false))
4083 rehash_using_reg (op0);
4084 op0_hash = HASH (op0, mode);
4087 op0_elt = insert (op0, NULL, op0_hash, mode);
4088 op0_elt->in_memory = op0_in_memory;
4091 if (op1_elt == 0)
4093 if (insert_regs (op1, NULL, false))
4095 rehash_using_reg (op1);
4096 op1_hash = HASH (op1, mode);
4099 op1_elt = insert (op1, NULL, op1_hash, mode);
4100 op1_elt->in_memory = op1_in_memory;
4103 merge_equiv_classes (op0_elt, op1_elt);
4106 /* CSE processing for one instruction.
4108 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4109 but the few that "leak through" are cleaned up by cse_insn, and complex
4110 addressing modes are often formed here.
4112 The main function is cse_insn, and between here and that function
4113 a couple of helper functions is defined to keep the size of cse_insn
4114 within reasonable proportions.
4116 Data is shared between the main and helper functions via STRUCT SET,
4117 that contains all data related for every set in the instruction that
4118 is being processed.
4120 Note that cse_main processes all sets in the instruction. Most
4121 passes in GCC only process simple SET insns or single_set insns, but
4122 CSE processes insns with multiple sets as well. */
4124 /* Data on one SET contained in the instruction. */
4126 struct set
4128 /* The SET rtx itself. */
4129 rtx rtl;
4130 /* The SET_SRC of the rtx (the original value, if it is changing). */
4131 rtx src;
4132 /* The hash-table element for the SET_SRC of the SET. */
4133 struct table_elt *src_elt;
4134 /* Hash value for the SET_SRC. */
4135 unsigned src_hash;
4136 /* Hash value for the SET_DEST. */
4137 unsigned dest_hash;
4138 /* The SET_DEST, with SUBREG, etc., stripped. */
4139 rtx inner_dest;
4140 /* Nonzero if the SET_SRC is in memory. */
4141 char src_in_memory;
4142 /* Nonzero if the SET_SRC contains something
4143 whose value cannot be predicted and understood. */
4144 char src_volatile;
4145 /* Original machine mode, in case it becomes a CONST_INT. */
4146 ENUM_BITFIELD(machine_mode) mode : MACHINE_MODE_BITSIZE;
4147 /* Hash value of constant equivalent for SET_SRC. */
4148 unsigned src_const_hash;
4149 /* A constant equivalent for SET_SRC, if any. */
4150 rtx src_const;
4151 /* Table entry for constant equivalent for SET_SRC, if any. */
4152 struct table_elt *src_const_elt;
4153 /* Table entry for the destination address. */
4154 struct table_elt *dest_addr_elt;
4157 /* Special handling for (set REG0 REG1) where REG0 is the
4158 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4159 be used in the sequel, so (if easily done) change this insn to
4160 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4161 that computed their value. Then REG1 will become a dead store
4162 and won't cloud the situation for later optimizations.
4164 Do not make this change if REG1 is a hard register, because it will
4165 then be used in the sequel and we may be changing a two-operand insn
4166 into a three-operand insn.
4168 This is the last transformation that cse_insn will try to do. */
4170 static void
4171 try_back_substitute_reg (rtx set, rtx_insn *insn)
4173 rtx dest = SET_DEST (set);
4174 rtx src = SET_SRC (set);
4176 if (REG_P (dest)
4177 && REG_P (src) && ! HARD_REGISTER_P (src)
4178 && REGNO_QTY_VALID_P (REGNO (src)))
4180 int src_q = REG_QTY (REGNO (src));
4181 struct qty_table_elem *src_ent = &qty_table[src_q];
4183 if (src_ent->first_reg == REGNO (dest))
4185 /* Scan for the previous nonnote insn, but stop at a basic
4186 block boundary. */
4187 rtx_insn *prev = insn;
4188 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4191 prev = PREV_INSN (prev);
4193 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4195 /* Do not swap the registers around if the previous instruction
4196 attaches a REG_EQUIV note to REG1.
4198 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4199 from the pseudo that originally shadowed an incoming argument
4200 to another register. Some uses of REG_EQUIV might rely on it
4201 being attached to REG1 rather than REG2.
4203 This section previously turned the REG_EQUIV into a REG_EQUAL
4204 note. We cannot do that because REG_EQUIV may provide an
4205 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4206 if (NONJUMP_INSN_P (prev)
4207 && GET_CODE (PATTERN (prev)) == SET
4208 && SET_DEST (PATTERN (prev)) == src
4209 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4211 rtx note;
4213 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4214 validate_change (insn, &SET_DEST (set), src, 1);
4215 validate_change (insn, &SET_SRC (set), dest, 1);
4216 apply_change_group ();
4218 /* If INSN has a REG_EQUAL note, and this note mentions
4219 REG0, then we must delete it, because the value in
4220 REG0 has changed. If the note's value is REG1, we must
4221 also delete it because that is now this insn's dest. */
4222 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4223 if (note != 0
4224 && (reg_mentioned_p (dest, XEXP (note, 0))
4225 || rtx_equal_p (src, XEXP (note, 0))))
4226 remove_note (insn, note);
4228 /* If INSN has a REG_ARGS_SIZE note, move it to PREV. */
4229 note = find_reg_note (insn, REG_ARGS_SIZE, NULL_RTX);
4230 if (note != 0)
4232 remove_note (insn, note);
4233 gcc_assert (!find_reg_note (prev, REG_ARGS_SIZE, NULL_RTX));
4234 set_unique_reg_note (prev, REG_ARGS_SIZE, XEXP (note, 0));
4241 /* Add an entry containing RTL X into SETS. */
4242 static inline void
4243 add_to_set (vec<struct set> *sets, rtx x)
4245 struct set entry = {};
4246 entry.rtl = x;
4247 sets->safe_push (entry);
4250 /* Record all the SETs in this instruction into SETS_PTR,
4251 and return the number of recorded sets. */
4252 static int
4253 find_sets_in_insn (rtx_insn *insn, vec<struct set> *psets)
4255 rtx x = PATTERN (insn);
4257 if (GET_CODE (x) == SET)
4259 /* Ignore SETs that are unconditional jumps.
4260 They never need cse processing, so this does not hurt.
4261 The reason is not efficiency but rather
4262 so that we can test at the end for instructions
4263 that have been simplified to unconditional jumps
4264 and not be misled by unchanged instructions
4265 that were unconditional jumps to begin with. */
4266 if (SET_DEST (x) == pc_rtx
4267 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4269 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4270 The hard function value register is used only once, to copy to
4271 someplace else, so it isn't worth cse'ing. */
4272 else if (GET_CODE (SET_SRC (x)) == CALL)
4274 else if (GET_CODE (SET_SRC (x)) == CONST_VECTOR
4275 && GET_MODE_CLASS (GET_MODE (SET_SRC (x))) != MODE_VECTOR_BOOL
4276 /* Prevent duplicates from being generated if the type is a V1
4277 type and a subreg. Folding this will result in the same
4278 element as folding x itself. */
4279 && !(SUBREG_P (SET_DEST (x))
4280 && known_eq (GET_MODE_NUNITS (GET_MODE (SET_SRC (x))), 1)))
4282 /* First register the vector itself. */
4283 add_to_set (psets, x);
4284 rtx src = SET_SRC (x);
4285 /* Go over the constants of the CONST_VECTOR in forward order, to
4286 put them in the same order in the SETS array. */
4287 for (unsigned i = 0; i < const_vector_encoded_nelts (src) ; i++)
4289 /* These are templates and don't actually get emitted but are
4290 used to tell CSE how to get to a particular constant. */
4291 rtx y = simplify_gen_vec_select (SET_DEST (x), i);
4292 gcc_assert (y);
4293 add_to_set (psets, gen_rtx_SET (y, CONST_VECTOR_ELT (src, i)));
4296 else
4297 add_to_set (psets, x);
4299 else if (GET_CODE (x) == PARALLEL)
4301 int i, lim = XVECLEN (x, 0);
4303 /* Go over the expressions of the PARALLEL in forward order, to
4304 put them in the same order in the SETS array. */
4305 for (i = 0; i < lim; i++)
4307 rtx y = XVECEXP (x, 0, i);
4308 if (GET_CODE (y) == SET)
4310 /* As above, we ignore unconditional jumps and call-insns and
4311 ignore the result of apply_change_group. */
4312 if (SET_DEST (y) == pc_rtx
4313 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4315 else if (GET_CODE (SET_SRC (y)) == CALL)
4317 else
4318 add_to_set (psets, y);
4323 return psets->length ();
4326 /* Subroutine of canonicalize_insn. X is an ASM_OPERANDS in INSN. */
4328 static void
4329 canon_asm_operands (rtx x, rtx_insn *insn)
4331 for (int i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4333 rtx input = ASM_OPERANDS_INPUT (x, i);
4334 if (!(REG_P (input) && HARD_REGISTER_P (input)))
4336 input = canon_reg (input, insn);
4337 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4342 /* Where possible, substitute every register reference in the N_SETS
4343 number of SETS in INSN with the canonical register.
4345 Register canonicalization propagatest the earliest register (i.e.
4346 one that is set before INSN) with the same value. This is a very
4347 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4348 to RTL. For instance, a CONST for an address is usually expanded
4349 multiple times to loads into different registers, thus creating many
4350 subexpressions of the form:
4352 (set (reg1) (some_const))
4353 (set (mem (... reg1 ...) (thing)))
4354 (set (reg2) (some_const))
4355 (set (mem (... reg2 ...) (thing)))
4357 After canonicalizing, the code takes the following form:
4359 (set (reg1) (some_const))
4360 (set (mem (... reg1 ...) (thing)))
4361 (set (reg2) (some_const))
4362 (set (mem (... reg1 ...) (thing)))
4364 The set to reg2 is now trivially dead, and the memory reference (or
4365 address, or whatever) may be a candidate for further CSEing.
4367 In this function, the result of apply_change_group can be ignored;
4368 see canon_reg. */
4370 static void
4371 canonicalize_insn (rtx_insn *insn, vec<struct set> *psets)
4373 vec<struct set> sets = *psets;
4374 int n_sets = sets.length ();
4375 rtx tem;
4376 rtx x = PATTERN (insn);
4377 int i;
4379 if (CALL_P (insn))
4381 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4382 if (GET_CODE (XEXP (tem, 0)) != SET)
4383 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4386 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4388 canon_reg (SET_SRC (x), insn);
4389 apply_change_group ();
4390 fold_rtx (SET_SRC (x), insn);
4392 else if (GET_CODE (x) == CLOBBER)
4394 /* If we clobber memory, canon the address.
4395 This does nothing when a register is clobbered
4396 because we have already invalidated the reg. */
4397 if (MEM_P (XEXP (x, 0)))
4398 canon_reg (XEXP (x, 0), insn);
4400 else if (GET_CODE (x) == USE
4401 && ! (REG_P (XEXP (x, 0))
4402 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4403 /* Canonicalize a USE of a pseudo register or memory location. */
4404 canon_reg (x, insn);
4405 else if (GET_CODE (x) == ASM_OPERANDS)
4406 canon_asm_operands (x, insn);
4407 else if (GET_CODE (x) == CALL)
4409 canon_reg (x, insn);
4410 apply_change_group ();
4411 fold_rtx (x, insn);
4413 else if (DEBUG_INSN_P (insn))
4414 canon_reg (PATTERN (insn), insn);
4415 else if (GET_CODE (x) == PARALLEL)
4417 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4419 rtx y = XVECEXP (x, 0, i);
4420 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4422 canon_reg (SET_SRC (y), insn);
4423 apply_change_group ();
4424 fold_rtx (SET_SRC (y), insn);
4426 else if (GET_CODE (y) == CLOBBER)
4428 if (MEM_P (XEXP (y, 0)))
4429 canon_reg (XEXP (y, 0), insn);
4431 else if (GET_CODE (y) == USE
4432 && ! (REG_P (XEXP (y, 0))
4433 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4434 canon_reg (y, insn);
4435 else if (GET_CODE (y) == ASM_OPERANDS)
4436 canon_asm_operands (y, insn);
4437 else if (GET_CODE (y) == CALL)
4439 canon_reg (y, insn);
4440 apply_change_group ();
4441 fold_rtx (y, insn);
4446 if (n_sets == 1 && REG_NOTES (insn) != 0
4447 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4449 /* We potentially will process this insn many times. Therefore,
4450 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4451 unique set in INSN.
4453 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4454 because cse_insn handles those specially. */
4455 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4456 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4457 remove_note (insn, tem);
4458 else
4460 canon_reg (XEXP (tem, 0), insn);
4461 apply_change_group ();
4462 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4463 df_notes_rescan (insn);
4467 /* Canonicalize sources and addresses of destinations.
4468 We do this in a separate pass to avoid problems when a MATCH_DUP is
4469 present in the insn pattern. In that case, we want to ensure that
4470 we don't break the duplicate nature of the pattern. So we will replace
4471 both operands at the same time. Otherwise, we would fail to find an
4472 equivalent substitution in the loop calling validate_change below.
4474 We used to suppress canonicalization of DEST if it appears in SRC,
4475 but we don't do this any more. */
4477 for (i = 0; i < n_sets; i++)
4479 rtx dest = SET_DEST (sets[i].rtl);
4480 rtx src = SET_SRC (sets[i].rtl);
4481 rtx new_rtx = canon_reg (src, insn);
4483 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4485 if (GET_CODE (dest) == ZERO_EXTRACT)
4487 validate_change (insn, &XEXP (dest, 1),
4488 canon_reg (XEXP (dest, 1), insn), 1);
4489 validate_change (insn, &XEXP (dest, 2),
4490 canon_reg (XEXP (dest, 2), insn), 1);
4493 while (GET_CODE (dest) == SUBREG
4494 || GET_CODE (dest) == ZERO_EXTRACT
4495 || GET_CODE (dest) == STRICT_LOW_PART)
4496 dest = XEXP (dest, 0);
4498 if (MEM_P (dest))
4499 canon_reg (dest, insn);
4502 /* Now that we have done all the replacements, we can apply the change
4503 group and see if they all work. Note that this will cause some
4504 canonicalizations that would have worked individually not to be applied
4505 because some other canonicalization didn't work, but this should not
4506 occur often.
4508 The result of apply_change_group can be ignored; see canon_reg. */
4510 apply_change_group ();
4513 /* Main function of CSE.
4514 First simplify sources and addresses of all assignments
4515 in the instruction, using previously-computed equivalents values.
4516 Then install the new sources and destinations in the table
4517 of available values. */
4519 static void
4520 cse_insn (rtx_insn *insn)
4522 rtx x = PATTERN (insn);
4523 int i;
4524 rtx tem;
4525 int n_sets = 0;
4527 rtx src_eqv = 0;
4528 struct table_elt *src_eqv_elt = 0;
4529 int src_eqv_volatile = 0;
4530 int src_eqv_in_memory = 0;
4531 unsigned src_eqv_hash = 0;
4533 this_insn = insn;
4535 /* Find all regs explicitly clobbered in this insn,
4536 to ensure they are not replaced with any other regs
4537 elsewhere in this insn. */
4538 invalidate_from_sets_and_clobbers (insn);
4540 /* Record all the SETs in this instruction. */
4541 auto_vec<struct set, 8> sets;
4542 n_sets = find_sets_in_insn (insn, (vec<struct set>*)&sets);
4544 /* Substitute the canonical register where possible. */
4545 canonicalize_insn (insn, (vec<struct set>*)&sets);
4547 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4548 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4549 latter condition is necessary because SRC_EQV is handled specially for
4550 this case, and if it isn't set, then there will be no equivalence
4551 for the destination. */
4552 if (n_sets == 1 && REG_NOTES (insn) != 0
4553 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4556 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4557 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4558 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4559 src_eqv = copy_rtx (XEXP (tem, 0));
4560 /* If DEST is of the form ZERO_EXTACT, as in:
4561 (set (zero_extract:SI (reg:SI 119)
4562 (const_int 16 [0x10])
4563 (const_int 16 [0x10]))
4564 (const_int 51154 [0xc7d2]))
4565 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4566 point. Note that this is different from SRC_EQV. We can however
4567 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4568 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4569 && CONST_INT_P (XEXP (tem, 0))
4570 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4571 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4573 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4574 /* This is the mode of XEXP (tem, 0) as well. */
4575 scalar_int_mode dest_mode
4576 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
4577 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4578 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4579 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4580 HOST_WIDE_INT mask;
4581 unsigned int shift;
4582 if (BITS_BIG_ENDIAN)
4583 shift = (GET_MODE_PRECISION (dest_mode)
4584 - INTVAL (pos) - INTVAL (width));
4585 else
4586 shift = INTVAL (pos);
4587 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4588 mask = HOST_WIDE_INT_M1;
4589 else
4590 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
4591 val = (val >> shift) & mask;
4592 src_eqv = GEN_INT (val);
4596 /* Set sets[i].src_elt to the class each source belongs to.
4597 Detect assignments from or to volatile things
4598 and set set[i] to zero so they will be ignored
4599 in the rest of this function.
4601 Nothing in this loop changes the hash table or the register chains. */
4603 for (i = 0; i < n_sets; i++)
4605 bool repeat = false;
4606 bool noop_insn = false;
4607 rtx src, dest;
4608 rtx src_folded;
4609 struct table_elt *elt = 0, *p;
4610 machine_mode mode;
4611 rtx src_eqv_here;
4612 rtx src_const = 0;
4613 rtx src_related = 0;
4614 rtx dest_related = 0;
4615 bool src_related_is_const_anchor = false;
4616 struct table_elt *src_const_elt = 0;
4617 int src_cost = MAX_COST;
4618 int src_eqv_cost = MAX_COST;
4619 int src_folded_cost = MAX_COST;
4620 int src_related_cost = MAX_COST;
4621 int src_elt_cost = MAX_COST;
4622 int src_regcost = MAX_COST;
4623 int src_eqv_regcost = MAX_COST;
4624 int src_folded_regcost = MAX_COST;
4625 int src_related_regcost = MAX_COST;
4626 int src_elt_regcost = MAX_COST;
4627 scalar_int_mode int_mode;
4629 dest = SET_DEST (sets[i].rtl);
4630 src = SET_SRC (sets[i].rtl);
4632 /* If SRC is a constant that has no machine mode,
4633 hash it with the destination's machine mode.
4634 This way we can keep different modes separate. */
4636 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4637 sets[i].mode = mode;
4639 if (src_eqv)
4641 machine_mode eqvmode = mode;
4642 if (GET_CODE (dest) == STRICT_LOW_PART)
4643 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4644 do_not_record = 0;
4645 hash_arg_in_memory = 0;
4646 src_eqv_hash = HASH (src_eqv, eqvmode);
4648 /* Find the equivalence class for the equivalent expression. */
4650 if (!do_not_record)
4651 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4653 src_eqv_volatile = do_not_record;
4654 src_eqv_in_memory = hash_arg_in_memory;
4657 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4658 value of the INNER register, not the destination. So it is not
4659 a valid substitution for the source. But save it for later. */
4660 if (GET_CODE (dest) == STRICT_LOW_PART)
4661 src_eqv_here = 0;
4662 else
4663 src_eqv_here = src_eqv;
4665 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4666 simplified result, which may not necessarily be valid. */
4667 src_folded = fold_rtx (src, NULL);
4669 #if 0
4670 /* ??? This caused bad code to be generated for the m68k port with -O2.
4671 Suppose src is (CONST_INT -1), and that after truncation src_folded
4672 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4673 At the end we will add src and src_const to the same equivalence
4674 class. We now have 3 and -1 on the same equivalence class. This
4675 causes later instructions to be mis-optimized. */
4676 /* If storing a constant in a bitfield, pre-truncate the constant
4677 so we will be able to record it later. */
4678 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4680 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4682 if (CONST_INT_P (src)
4683 && CONST_INT_P (width)
4684 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4685 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4686 src_folded
4687 = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1
4688 << INTVAL (width)) - 1));
4690 #endif
4692 /* Compute SRC's hash code, and also notice if it
4693 should not be recorded at all. In that case,
4694 prevent any further processing of this assignment.
4696 We set DO_NOT_RECORD if the destination has a REG_UNUSED note.
4697 This avoids getting the source register into the tables, where it
4698 may be invalidated later (via REG_QTY), then trigger an ICE upon
4699 re-insertion.
4701 This is only a problem in multi-set insns. If it were a single
4702 set the dead copy would have been removed. If the RHS were anything
4703 but a simple REG, then we won't call insert_regs and thus there's
4704 no potential for triggering the ICE. */
4705 do_not_record = (REG_P (dest)
4706 && REG_P (src)
4707 && find_reg_note (insn, REG_UNUSED, dest));
4708 hash_arg_in_memory = 0;
4710 sets[i].src = src;
4711 sets[i].src_hash = HASH (src, mode);
4712 sets[i].src_volatile = do_not_record;
4713 sets[i].src_in_memory = hash_arg_in_memory;
4715 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4716 a pseudo, do not record SRC. Using SRC as a replacement for
4717 anything else will be incorrect in that situation. Note that
4718 this usually occurs only for stack slots, in which case all the
4719 RTL would be referring to SRC, so we don't lose any optimization
4720 opportunities by not having SRC in the hash table. */
4722 if (MEM_P (src)
4723 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4724 && REG_P (dest)
4725 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4726 sets[i].src_volatile = 1;
4728 else if (GET_CODE (src) == ASM_OPERANDS
4729 && GET_CODE (x) == PARALLEL)
4731 /* Do not record result of a non-volatile inline asm with
4732 more than one result. */
4733 if (n_sets > 1)
4734 sets[i].src_volatile = 1;
4736 int j, lim = XVECLEN (x, 0);
4737 for (j = 0; j < lim; j++)
4739 rtx y = XVECEXP (x, 0, j);
4740 /* And do not record result of a non-volatile inline asm
4741 with "memory" clobber. */
4742 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4744 sets[i].src_volatile = 1;
4745 break;
4750 #if 0
4751 /* It is no longer clear why we used to do this, but it doesn't
4752 appear to still be needed. So let's try without it since this
4753 code hurts cse'ing widened ops. */
4754 /* If source is a paradoxical subreg (such as QI treated as an SI),
4755 treat it as volatile. It may do the work of an SI in one context
4756 where the extra bits are not being used, but cannot replace an SI
4757 in general. */
4758 if (paradoxical_subreg_p (src))
4759 sets[i].src_volatile = 1;
4760 #endif
4762 /* Locate all possible equivalent forms for SRC. Try to replace
4763 SRC in the insn with each cheaper equivalent.
4765 We have the following types of equivalents: SRC itself, a folded
4766 version, a value given in a REG_EQUAL note, or a value related
4767 to a constant.
4769 Each of these equivalents may be part of an additional class
4770 of equivalents (if more than one is in the table, they must be in
4771 the same class; we check for this).
4773 If the source is volatile, we don't do any table lookups.
4775 We note any constant equivalent for possible later use in a
4776 REG_NOTE. */
4778 if (!sets[i].src_volatile)
4779 elt = lookup (src, sets[i].src_hash, mode);
4781 sets[i].src_elt = elt;
4783 if (elt && src_eqv_here && src_eqv_elt)
4785 if (elt->first_same_value != src_eqv_elt->first_same_value)
4787 /* The REG_EQUAL is indicating that two formerly distinct
4788 classes are now equivalent. So merge them. */
4789 merge_equiv_classes (elt, src_eqv_elt);
4790 src_eqv_hash = HASH (src_eqv, elt->mode);
4791 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4794 src_eqv_here = 0;
4797 else if (src_eqv_elt)
4798 elt = src_eqv_elt;
4800 /* Try to find a constant somewhere and record it in `src_const'.
4801 Record its table element, if any, in `src_const_elt'. Look in
4802 any known equivalences first. (If the constant is not in the
4803 table, also set `sets[i].src_const_hash'). */
4804 if (elt)
4805 for (p = elt->first_same_value; p; p = p->next_same_value)
4806 if (p->is_const)
4808 src_const = p->exp;
4809 src_const_elt = elt;
4810 break;
4813 if (src_const == 0
4814 && (CONSTANT_P (src_folded)
4815 /* Consider (minus (label_ref L1) (label_ref L2)) as
4816 "constant" here so we will record it. This allows us
4817 to fold switch statements when an ADDR_DIFF_VEC is used. */
4818 || (GET_CODE (src_folded) == MINUS
4819 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4820 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4821 src_const = src_folded, src_const_elt = elt;
4822 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4823 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4825 /* If we don't know if the constant is in the table, get its
4826 hash code and look it up. */
4827 if (src_const && src_const_elt == 0)
4829 sets[i].src_const_hash = HASH (src_const, mode);
4830 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4833 sets[i].src_const = src_const;
4834 sets[i].src_const_elt = src_const_elt;
4836 /* If the constant and our source are both in the table, mark them as
4837 equivalent. Otherwise, if a constant is in the table but the source
4838 isn't, set ELT to it. */
4839 if (src_const_elt && elt
4840 && src_const_elt->first_same_value != elt->first_same_value)
4841 merge_equiv_classes (elt, src_const_elt);
4842 else if (src_const_elt && elt == 0)
4843 elt = src_const_elt;
4845 /* See if there is a register linearly related to a constant
4846 equivalent of SRC. */
4847 if (src_const
4848 && (GET_CODE (src_const) == CONST
4849 || (src_const_elt && src_const_elt->related_value != 0)))
4851 src_related = use_related_value (src_const, src_const_elt);
4852 if (src_related)
4854 struct table_elt *src_related_elt
4855 = lookup (src_related, HASH (src_related, mode), mode);
4856 if (src_related_elt && elt)
4858 if (elt->first_same_value
4859 != src_related_elt->first_same_value)
4860 /* This can occur when we previously saw a CONST
4861 involving a SYMBOL_REF and then see the SYMBOL_REF
4862 twice. Merge the involved classes. */
4863 merge_equiv_classes (elt, src_related_elt);
4865 src_related = 0;
4866 src_related_elt = 0;
4868 else if (src_related_elt && elt == 0)
4869 elt = src_related_elt;
4873 /* See if we have a CONST_INT that is already in a register in a
4874 wider mode. */
4876 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4877 && is_int_mode (mode, &int_mode)
4878 && GET_MODE_PRECISION (int_mode) < BITS_PER_WORD)
4880 opt_scalar_int_mode wider_mode_iter;
4881 FOR_EACH_WIDER_MODE (wider_mode_iter, int_mode)
4883 scalar_int_mode wider_mode = wider_mode_iter.require ();
4884 if (GET_MODE_PRECISION (wider_mode) > BITS_PER_WORD)
4885 break;
4887 struct table_elt *const_elt
4888 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4890 if (const_elt == 0)
4891 continue;
4893 for (const_elt = const_elt->first_same_value;
4894 const_elt; const_elt = const_elt->next_same_value)
4895 if (REG_P (const_elt->exp))
4897 src_related = gen_lowpart (int_mode, const_elt->exp);
4898 break;
4901 if (src_related != 0)
4902 break;
4906 /* Another possibility is that we have an AND with a constant in
4907 a mode narrower than a word. If so, it might have been generated
4908 as part of an "if" which would narrow the AND. If we already
4909 have done the AND in a wider mode, we can use a SUBREG of that
4910 value. */
4912 if (flag_expensive_optimizations && ! src_related
4913 && is_a <scalar_int_mode> (mode, &int_mode)
4914 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4915 && GET_MODE_SIZE (int_mode) < UNITS_PER_WORD)
4917 opt_scalar_int_mode tmode_iter;
4918 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4920 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4922 scalar_int_mode tmode = tmode_iter.require ();
4923 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4924 break;
4926 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4927 struct table_elt *larger_elt;
4929 if (inner)
4931 PUT_MODE (new_and, tmode);
4932 XEXP (new_and, 0) = inner;
4933 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4934 if (larger_elt == 0)
4935 continue;
4937 for (larger_elt = larger_elt->first_same_value;
4938 larger_elt; larger_elt = larger_elt->next_same_value)
4939 if (REG_P (larger_elt->exp))
4941 src_related
4942 = gen_lowpart (int_mode, larger_elt->exp);
4943 break;
4946 if (src_related)
4947 break;
4952 /* See if a MEM has already been loaded with a widening operation;
4953 if it has, we can use a subreg of that. Many CISC machines
4954 also have such operations, but this is only likely to be
4955 beneficial on these machines. */
4957 rtx_code extend_op;
4958 if (flag_expensive_optimizations && src_related == 0
4959 && MEM_P (src) && ! do_not_record
4960 && is_a <scalar_int_mode> (mode, &int_mode)
4961 && (extend_op = load_extend_op (int_mode)) != UNKNOWN)
4963 struct rtx_def memory_extend_buf;
4964 rtx memory_extend_rtx = &memory_extend_buf;
4966 /* Set what we are trying to extend and the operation it might
4967 have been extended with. */
4968 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4969 PUT_CODE (memory_extend_rtx, extend_op);
4970 XEXP (memory_extend_rtx, 0) = src;
4972 opt_scalar_int_mode tmode_iter;
4973 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4975 struct table_elt *larger_elt;
4977 scalar_int_mode tmode = tmode_iter.require ();
4978 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4979 break;
4981 PUT_MODE (memory_extend_rtx, tmode);
4982 larger_elt = lookup (memory_extend_rtx,
4983 HASH (memory_extend_rtx, tmode), tmode);
4984 if (larger_elt == 0)
4985 continue;
4987 for (larger_elt = larger_elt->first_same_value;
4988 larger_elt; larger_elt = larger_elt->next_same_value)
4989 if (REG_P (larger_elt->exp))
4991 src_related = gen_lowpart (int_mode, larger_elt->exp);
4992 break;
4995 if (src_related)
4996 break;
5000 /* Try to express the constant using a register+offset expression
5001 derived from a constant anchor. */
5003 if (targetm.const_anchor
5004 && !src_related
5005 && src_const
5006 && GET_CODE (src_const) == CONST_INT)
5008 src_related = try_const_anchors (src_const, mode);
5009 src_related_is_const_anchor = src_related != NULL_RTX;
5012 /* Try to re-materialize a vec_dup with an existing constant. */
5013 rtx src_elt;
5014 if ((!src_eqv_here || CONSTANT_P (src_eqv_here))
5015 && const_vec_duplicate_p (src, &src_elt))
5017 machine_mode const_mode = GET_MODE_INNER (GET_MODE (src));
5018 struct table_elt *related_elt
5019 = lookup (src_elt, HASH (src_elt, const_mode), const_mode);
5020 if (related_elt)
5022 for (related_elt = related_elt->first_same_value;
5023 related_elt; related_elt = related_elt->next_same_value)
5024 if (REG_P (related_elt->exp))
5026 /* We don't need to compare costs with an existing (constant)
5027 src_eqv_here, since any such src_eqv_here should already be
5028 available in src_const. */
5029 src_eqv_here
5030 = gen_rtx_VEC_DUPLICATE (GET_MODE (src),
5031 related_elt->exp);
5032 break;
5037 if (src == src_folded)
5038 src_folded = 0;
5040 /* At this point, ELT, if nonzero, points to a class of expressions
5041 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5042 and SRC_RELATED, if nonzero, each contain additional equivalent
5043 expressions. Prune these latter expressions by deleting expressions
5044 already in the equivalence class.
5046 Check for an equivalent identical to the destination. If found,
5047 this is the preferred equivalent since it will likely lead to
5048 elimination of the insn. Indicate this by placing it in
5049 `src_related'. */
5051 if (elt)
5052 elt = elt->first_same_value;
5053 for (p = elt; p; p = p->next_same_value)
5055 enum rtx_code code = GET_CODE (p->exp);
5057 /* If the expression is not valid, ignore it. Then we do not
5058 have to check for validity below. In most cases, we can use
5059 `rtx_equal_p', since canonicalization has already been done. */
5060 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5061 continue;
5063 /* Also skip paradoxical subregs, unless that's what we're
5064 looking for. */
5065 if (paradoxical_subreg_p (p->exp)
5066 && ! (src != 0
5067 && GET_CODE (src) == SUBREG
5068 && GET_MODE (src) == GET_MODE (p->exp)
5069 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5070 GET_MODE (SUBREG_REG (p->exp)))))
5071 continue;
5073 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5074 src = 0;
5075 else if (src_folded && GET_CODE (src_folded) == code
5076 && rtx_equal_p (src_folded, p->exp))
5077 src_folded = 0;
5078 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5079 && rtx_equal_p (src_eqv_here, p->exp))
5080 src_eqv_here = 0;
5081 else if (src_related && GET_CODE (src_related) == code
5082 && rtx_equal_p (src_related, p->exp))
5083 src_related = 0;
5085 /* This is the same as the destination of the insns, we want
5086 to prefer it. The code below will then give it a negative
5087 cost. */
5088 if (!dest_related
5089 && GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5090 dest_related = p->exp;
5093 /* Find the cheapest valid equivalent, trying all the available
5094 possibilities. Prefer items not in the hash table to ones
5095 that are when they are equal cost. Note that we can never
5096 worsen an insn as the current contents will also succeed.
5097 If we find an equivalent identical to the destination, use it as best,
5098 since this insn will probably be eliminated in that case. */
5099 if (src)
5101 if (rtx_equal_p (src, dest))
5102 src_cost = src_regcost = -1;
5103 else
5105 src_cost = COST (src, mode);
5106 src_regcost = approx_reg_cost (src);
5110 if (src_eqv_here)
5112 if (rtx_equal_p (src_eqv_here, dest))
5113 src_eqv_cost = src_eqv_regcost = -1;
5114 else
5116 src_eqv_cost = COST (src_eqv_here, mode);
5117 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5121 if (src_folded)
5123 if (rtx_equal_p (src_folded, dest))
5124 src_folded_cost = src_folded_regcost = -1;
5125 else
5127 src_folded_cost = COST (src_folded, mode);
5128 src_folded_regcost = approx_reg_cost (src_folded);
5132 if (dest_related)
5134 src_related_cost = src_related_regcost = -1;
5135 /* Handle it as src_related. */
5136 src_related = dest_related;
5138 else if (src_related)
5140 src_related_cost = COST (src_related, mode);
5141 src_related_regcost = approx_reg_cost (src_related);
5143 /* If a const-anchor is used to synthesize a constant that
5144 normally requires multiple instructions then slightly prefer
5145 it over the original sequence. These instructions are likely
5146 to become redundant now. We can't compare against the cost
5147 of src_eqv_here because, on MIPS for example, multi-insn
5148 constants have zero cost; they are assumed to be hoisted from
5149 loops. */
5150 if (src_related_is_const_anchor
5151 && src_related_cost == src_cost
5152 && src_eqv_here)
5153 src_related_cost--;
5156 /* If this was an indirect jump insn, a known label will really be
5157 cheaper even though it looks more expensive. */
5158 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5159 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5161 /* Terminate loop when replacement made. This must terminate since
5162 the current contents will be tested and will always be valid. */
5163 while (1)
5165 rtx trial;
5167 /* Skip invalid entries. */
5168 while (elt && !REG_P (elt->exp)
5169 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5170 elt = elt->next_same_value;
5172 /* A paradoxical subreg would be bad here: it'll be the right
5173 size, but later may be adjusted so that the upper bits aren't
5174 what we want. So reject it. */
5175 if (elt != 0
5176 && paradoxical_subreg_p (elt->exp)
5177 /* It is okay, though, if the rtx we're trying to match
5178 will ignore any of the bits we can't predict. */
5179 && ! (src != 0
5180 && GET_CODE (src) == SUBREG
5181 && GET_MODE (src) == GET_MODE (elt->exp)
5182 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5183 GET_MODE (SUBREG_REG (elt->exp)))))
5185 elt = elt->next_same_value;
5186 continue;
5189 if (elt)
5191 src_elt_cost = elt->cost;
5192 src_elt_regcost = elt->regcost;
5195 /* Find cheapest and skip it for the next time. For items
5196 of equal cost, use this order:
5197 src_folded, src, src_eqv, src_related and hash table entry. */
5198 if (src_folded
5199 && preferable (src_folded_cost, src_folded_regcost,
5200 src_cost, src_regcost) <= 0
5201 && preferable (src_folded_cost, src_folded_regcost,
5202 src_eqv_cost, src_eqv_regcost) <= 0
5203 && preferable (src_folded_cost, src_folded_regcost,
5204 src_related_cost, src_related_regcost) <= 0
5205 && preferable (src_folded_cost, src_folded_regcost,
5206 src_elt_cost, src_elt_regcost) <= 0)
5207 trial = src_folded, src_folded_cost = MAX_COST;
5208 else if (src
5209 && preferable (src_cost, src_regcost,
5210 src_eqv_cost, src_eqv_regcost) <= 0
5211 && preferable (src_cost, src_regcost,
5212 src_related_cost, src_related_regcost) <= 0
5213 && preferable (src_cost, src_regcost,
5214 src_elt_cost, src_elt_regcost) <= 0)
5215 trial = src, src_cost = MAX_COST;
5216 else if (src_eqv_here
5217 && preferable (src_eqv_cost, src_eqv_regcost,
5218 src_related_cost, src_related_regcost) <= 0
5219 && preferable (src_eqv_cost, src_eqv_regcost,
5220 src_elt_cost, src_elt_regcost) <= 0)
5221 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5222 else if (src_related
5223 && preferable (src_related_cost, src_related_regcost,
5224 src_elt_cost, src_elt_regcost) <= 0)
5225 trial = src_related, src_related_cost = MAX_COST;
5226 else
5228 trial = elt->exp;
5229 elt = elt->next_same_value;
5230 src_elt_cost = MAX_COST;
5233 /* Try to optimize
5234 (set (reg:M N) (const_int A))
5235 (set (reg:M2 O) (const_int B))
5236 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5237 (reg:M2 O)). */
5238 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5239 && CONST_INT_P (trial)
5240 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5241 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5242 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5243 && (known_ge
5244 (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl))),
5245 INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))))
5246 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5247 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5248 <= HOST_BITS_PER_WIDE_INT))
5250 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5251 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5252 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5253 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5254 struct table_elt *dest_elt
5255 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5256 rtx dest_cst = NULL;
5258 if (dest_elt)
5259 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5260 if (p->is_const && CONST_INT_P (p->exp))
5262 dest_cst = p->exp;
5263 break;
5265 if (dest_cst)
5267 HOST_WIDE_INT val = INTVAL (dest_cst);
5268 HOST_WIDE_INT mask;
5269 unsigned int shift;
5270 /* This is the mode of DEST_CST as well. */
5271 scalar_int_mode dest_mode
5272 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
5273 if (BITS_BIG_ENDIAN)
5274 shift = GET_MODE_PRECISION (dest_mode)
5275 - INTVAL (pos) - INTVAL (width);
5276 else
5277 shift = INTVAL (pos);
5278 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5279 mask = HOST_WIDE_INT_M1;
5280 else
5281 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
5282 val &= ~(mask << shift);
5283 val |= (INTVAL (trial) & mask) << shift;
5284 val = trunc_int_for_mode (val, dest_mode);
5285 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5286 dest_reg, 1);
5287 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5288 GEN_INT (val), 1);
5289 if (apply_change_group ())
5291 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5292 if (note)
5294 remove_note (insn, note);
5295 df_notes_rescan (insn);
5297 src_eqv = NULL_RTX;
5298 src_eqv_elt = NULL;
5299 src_eqv_volatile = 0;
5300 src_eqv_in_memory = 0;
5301 src_eqv_hash = 0;
5302 repeat = true;
5303 break;
5308 /* We don't normally have an insn matching (set (pc) (pc)), so
5309 check for this separately here. We will delete such an
5310 insn below.
5312 For other cases such as a table jump or conditional jump
5313 where we know the ultimate target, go ahead and replace the
5314 operand. While that may not make a valid insn, we will
5315 reemit the jump below (and also insert any necessary
5316 barriers). */
5317 if (n_sets == 1 && dest == pc_rtx
5318 && (trial == pc_rtx
5319 || (GET_CODE (trial) == LABEL_REF
5320 && ! condjump_p (insn))))
5322 /* Don't substitute non-local labels, this confuses CFG. */
5323 if (GET_CODE (trial) == LABEL_REF
5324 && LABEL_REF_NONLOCAL_P (trial))
5325 continue;
5327 SET_SRC (sets[i].rtl) = trial;
5328 cse_jumps_altered = true;
5329 break;
5332 /* Similarly, lots of targets don't allow no-op
5333 (set (mem x) (mem x)) moves. Even (set (reg x) (reg x))
5334 might be impossible for certain registers (like CC registers). */
5335 else if (n_sets == 1
5336 && !CALL_P (insn)
5337 && (MEM_P (trial) || REG_P (trial))
5338 && rtx_equal_p (trial, dest)
5339 && !side_effects_p (dest)
5340 && (cfun->can_delete_dead_exceptions
5341 || insn_nothrow_p (insn))
5342 /* We can only remove the later store if the earlier aliases
5343 at least all accesses the later one. */
5344 && (!MEM_P (trial)
5345 || ((MEM_ALIAS_SET (dest) == MEM_ALIAS_SET (trial)
5346 || alias_set_subset_of (MEM_ALIAS_SET (dest),
5347 MEM_ALIAS_SET (trial)))
5348 && (!MEM_EXPR (trial)
5349 || refs_same_for_tbaa_p (MEM_EXPR (trial),
5350 MEM_EXPR (dest))))))
5352 SET_SRC (sets[i].rtl) = trial;
5353 noop_insn = true;
5354 break;
5357 /* Reject certain invalid forms of CONST that we create. */
5358 else if (CONSTANT_P (trial)
5359 && GET_CODE (trial) == CONST
5360 /* Reject cases that will cause decode_rtx_const to
5361 die. On the alpha when simplifying a switch, we
5362 get (const (truncate (minus (label_ref)
5363 (label_ref)))). */
5364 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5365 /* Likewise on IA-64, except without the
5366 truncate. */
5367 || (GET_CODE (XEXP (trial, 0)) == MINUS
5368 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5369 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5370 /* Do nothing for this case. */
5373 /* Do not replace anything with a MEM, except the replacement
5374 is a no-op. This allows this loop to terminate. */
5375 else if (MEM_P (trial) && !rtx_equal_p (trial, SET_SRC(sets[i].rtl)))
5376 /* Do nothing for this case. */
5379 /* Look for a substitution that makes a valid insn. */
5380 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5381 trial, 0))
5383 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5385 /* The result of apply_change_group can be ignored; see
5386 canon_reg. */
5388 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5389 apply_change_group ();
5391 break;
5394 /* If the current function uses a constant pool and this is a
5395 constant, try making a pool entry. Put it in src_folded
5396 unless we already have done this since that is where it
5397 likely came from. */
5399 else if (crtl->uses_const_pool
5400 && CONSTANT_P (trial)
5401 && !CONST_INT_P (trial)
5402 && (src_folded == 0 || !MEM_P (src_folded))
5403 && GET_MODE_CLASS (mode) != MODE_CC
5404 && mode != VOIDmode)
5406 src_folded = force_const_mem (mode, trial);
5407 if (src_folded)
5409 src_folded_cost = COST (src_folded, mode);
5410 src_folded_regcost = approx_reg_cost (src_folded);
5415 /* If we changed the insn too much, handle this set from scratch. */
5416 if (repeat)
5418 i--;
5419 continue;
5422 src = SET_SRC (sets[i].rtl);
5424 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5425 However, there is an important exception: If both are registers
5426 that are not the head of their equivalence class, replace SET_SRC
5427 with the head of the class. If we do not do this, we will have
5428 both registers live over a portion of the basic block. This way,
5429 their lifetimes will likely abut instead of overlapping. */
5430 if (REG_P (dest)
5431 && REGNO_QTY_VALID_P (REGNO (dest)))
5433 int dest_q = REG_QTY (REGNO (dest));
5434 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5436 if (dest_ent->mode == GET_MODE (dest)
5437 && dest_ent->first_reg != REGNO (dest)
5438 && REG_P (src) && REGNO (src) == REGNO (dest)
5439 /* Don't do this if the original insn had a hard reg as
5440 SET_SRC or SET_DEST. */
5441 && (!REG_P (sets[i].src)
5442 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5443 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5444 /* We can't call canon_reg here because it won't do anything if
5445 SRC is a hard register. */
5447 int src_q = REG_QTY (REGNO (src));
5448 struct qty_table_elem *src_ent = &qty_table[src_q];
5449 int first = src_ent->first_reg;
5450 rtx new_src
5451 = (first >= FIRST_PSEUDO_REGISTER
5452 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5454 /* We must use validate-change even for this, because this
5455 might be a special no-op instruction, suitable only to
5456 tag notes onto. */
5457 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5459 src = new_src;
5460 /* If we had a constant that is cheaper than what we are now
5461 setting SRC to, use that constant. We ignored it when we
5462 thought we could make this into a no-op. */
5463 if (src_const && COST (src_const, mode) < COST (src, mode)
5464 && validate_change (insn, &SET_SRC (sets[i].rtl),
5465 src_const, 0))
5466 src = src_const;
5471 /* If we made a change, recompute SRC values. */
5472 if (src != sets[i].src)
5474 do_not_record = 0;
5475 hash_arg_in_memory = 0;
5476 sets[i].src = src;
5477 sets[i].src_hash = HASH (src, mode);
5478 sets[i].src_volatile = do_not_record;
5479 sets[i].src_in_memory = hash_arg_in_memory;
5480 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5483 /* If this is a single SET, we are setting a register, and we have an
5484 equivalent constant, we want to add a REG_EQUAL note if the constant
5485 is different from the source. We don't want to do it for a constant
5486 pseudo since verifying that this pseudo hasn't been eliminated is a
5487 pain; moreover such a note won't help anything.
5489 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5490 which can be created for a reference to a compile time computable
5491 entry in a jump table. */
5492 if (n_sets == 1
5493 && REG_P (dest)
5494 && src_const
5495 && !REG_P (src_const)
5496 && !(GET_CODE (src_const) == SUBREG
5497 && REG_P (SUBREG_REG (src_const)))
5498 && !(GET_CODE (src_const) == CONST
5499 && GET_CODE (XEXP (src_const, 0)) == MINUS
5500 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5501 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5502 && !rtx_equal_p (src, src_const))
5504 /* Make sure that the rtx is not shared. */
5505 src_const = copy_rtx (src_const);
5507 /* Record the actual constant value in a REG_EQUAL note,
5508 making a new one if one does not already exist. */
5509 set_unique_reg_note (insn, REG_EQUAL, src_const);
5510 df_notes_rescan (insn);
5513 /* Now deal with the destination. */
5514 do_not_record = 0;
5516 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5517 while (GET_CODE (dest) == SUBREG
5518 || GET_CODE (dest) == ZERO_EXTRACT
5519 || GET_CODE (dest) == STRICT_LOW_PART)
5520 dest = XEXP (dest, 0);
5522 sets[i].inner_dest = dest;
5524 if (MEM_P (dest))
5526 #ifdef PUSH_ROUNDING
5527 /* Stack pushes invalidate the stack pointer. */
5528 rtx addr = XEXP (dest, 0);
5529 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5530 && XEXP (addr, 0) == stack_pointer_rtx)
5531 invalidate (stack_pointer_rtx, VOIDmode);
5532 #endif
5533 dest = fold_rtx (dest, insn);
5536 /* Compute the hash code of the destination now,
5537 before the effects of this instruction are recorded,
5538 since the register values used in the address computation
5539 are those before this instruction. */
5540 sets[i].dest_hash = HASH (dest, mode);
5542 /* Don't enter a bit-field in the hash table
5543 because the value in it after the store
5544 may not equal what was stored, due to truncation. */
5546 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5548 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5550 if (src_const != 0 && CONST_INT_P (src_const)
5551 && CONST_INT_P (width)
5552 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5553 && ! (INTVAL (src_const)
5554 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5555 /* Exception: if the value is constant,
5556 and it won't be truncated, record it. */
5558 else
5560 /* This is chosen so that the destination will be invalidated
5561 but no new value will be recorded.
5562 We must invalidate because sometimes constant
5563 values can be recorded for bitfields. */
5564 sets[i].src_elt = 0;
5565 sets[i].src_volatile = 1;
5566 src_eqv = 0;
5567 src_eqv_elt = 0;
5571 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5572 the insn. */
5573 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5575 /* One less use of the label this insn used to jump to. */
5576 cse_cfg_altered |= delete_insn_and_edges (insn);
5577 cse_jumps_altered = true;
5578 /* No more processing for this set. */
5579 sets[i].rtl = 0;
5582 /* Similarly for no-op moves. */
5583 else if (noop_insn)
5585 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5586 cse_cfg_altered = true;
5587 cse_cfg_altered |= delete_insn_and_edges (insn);
5588 /* No more processing for this set. */
5589 sets[i].rtl = 0;
5592 /* If this SET is now setting PC to a label, we know it used to
5593 be a conditional or computed branch. */
5594 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5595 && !LABEL_REF_NONLOCAL_P (src))
5597 /* We reemit the jump in as many cases as possible just in
5598 case the form of an unconditional jump is significantly
5599 different than a computed jump or conditional jump.
5601 If this insn has multiple sets, then reemitting the
5602 jump is nontrivial. So instead we just force rerecognition
5603 and hope for the best. */
5604 if (n_sets == 1)
5606 rtx_jump_insn *new_rtx;
5607 rtx note;
5609 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5610 new_rtx = emit_jump_insn_before (seq, insn);
5611 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5612 LABEL_NUSES (XEXP (src, 0))++;
5614 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5615 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5616 if (note)
5618 XEXP (note, 1) = NULL_RTX;
5619 REG_NOTES (new_rtx) = note;
5622 cse_cfg_altered |= delete_insn_and_edges (insn);
5623 insn = new_rtx;
5625 else
5626 INSN_CODE (insn) = -1;
5628 /* Do not bother deleting any unreachable code, let jump do it. */
5629 cse_jumps_altered = true;
5630 sets[i].rtl = 0;
5633 /* If destination is volatile, invalidate it and then do no further
5634 processing for this assignment. */
5636 else if (do_not_record)
5638 invalidate_dest (dest);
5639 sets[i].rtl = 0;
5642 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5644 do_not_record = 0;
5645 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5646 if (do_not_record)
5648 invalidate_dest (SET_DEST (sets[i].rtl));
5649 sets[i].rtl = 0;
5654 /* Now enter all non-volatile source expressions in the hash table
5655 if they are not already present.
5656 Record their equivalence classes in src_elt.
5657 This way we can insert the corresponding destinations into
5658 the same classes even if the actual sources are no longer in them
5659 (having been invalidated). */
5661 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5662 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5664 struct table_elt *elt;
5665 struct table_elt *classp = sets[0].src_elt;
5666 rtx dest = SET_DEST (sets[0].rtl);
5667 machine_mode eqvmode = GET_MODE (dest);
5669 if (GET_CODE (dest) == STRICT_LOW_PART)
5671 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5672 classp = 0;
5674 if (insert_regs (src_eqv, classp, false))
5676 rehash_using_reg (src_eqv);
5677 src_eqv_hash = HASH (src_eqv, eqvmode);
5679 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5680 elt->in_memory = src_eqv_in_memory;
5681 src_eqv_elt = elt;
5683 /* Check to see if src_eqv_elt is the same as a set source which
5684 does not yet have an elt, and if so set the elt of the set source
5685 to src_eqv_elt. */
5686 for (i = 0; i < n_sets; i++)
5687 if (sets[i].rtl && sets[i].src_elt == 0
5688 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5689 sets[i].src_elt = src_eqv_elt;
5692 for (i = 0; i < n_sets; i++)
5693 if (sets[i].rtl && ! sets[i].src_volatile
5694 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5696 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5698 /* REG_EQUAL in setting a STRICT_LOW_PART
5699 gives an equivalent for the entire destination register,
5700 not just for the subreg being stored in now.
5701 This is a more interesting equivalence, so we arrange later
5702 to treat the entire reg as the destination. */
5703 sets[i].src_elt = src_eqv_elt;
5704 sets[i].src_hash = src_eqv_hash;
5706 else
5708 /* Insert source and constant equivalent into hash table, if not
5709 already present. */
5710 struct table_elt *classp = src_eqv_elt;
5711 rtx src = sets[i].src;
5712 rtx dest = SET_DEST (sets[i].rtl);
5713 machine_mode mode
5714 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5716 /* It's possible that we have a source value known to be
5717 constant but don't have a REG_EQUAL note on the insn.
5718 Lack of a note will mean src_eqv_elt will be NULL. This
5719 can happen where we've generated a SUBREG to access a
5720 CONST_INT that is already in a register in a wider mode.
5721 Ensure that the source expression is put in the proper
5722 constant class. */
5723 if (!classp)
5724 classp = sets[i].src_const_elt;
5726 if (sets[i].src_elt == 0)
5728 struct table_elt *elt;
5730 /* Note that these insert_regs calls cannot remove
5731 any of the src_elt's, because they would have failed to
5732 match if not still valid. */
5733 if (insert_regs (src, classp, false))
5735 rehash_using_reg (src);
5736 sets[i].src_hash = HASH (src, mode);
5738 elt = insert (src, classp, sets[i].src_hash, mode);
5739 elt->in_memory = sets[i].src_in_memory;
5740 /* If inline asm has any clobbers, ensure we only reuse
5741 existing inline asms and never try to put the ASM_OPERANDS
5742 into an insn that isn't inline asm. */
5743 if (GET_CODE (src) == ASM_OPERANDS
5744 && GET_CODE (x) == PARALLEL)
5745 elt->cost = MAX_COST;
5746 sets[i].src_elt = classp = elt;
5748 if (sets[i].src_const && sets[i].src_const_elt == 0
5749 && src != sets[i].src_const
5750 && ! rtx_equal_p (sets[i].src_const, src))
5751 sets[i].src_elt = insert (sets[i].src_const, classp,
5752 sets[i].src_const_hash, mode);
5755 else if (sets[i].src_elt == 0)
5756 /* If we did not insert the source into the hash table (e.g., it was
5757 volatile), note the equivalence class for the REG_EQUAL value, if any,
5758 so that the destination goes into that class. */
5759 sets[i].src_elt = src_eqv_elt;
5761 /* Record destination addresses in the hash table. This allows us to
5762 check if they are invalidated by other sets. */
5763 for (i = 0; i < n_sets; i++)
5765 if (sets[i].rtl)
5767 rtx x = sets[i].inner_dest;
5768 struct table_elt *elt;
5769 machine_mode mode;
5770 unsigned hash;
5772 if (MEM_P (x))
5774 x = XEXP (x, 0);
5775 mode = GET_MODE (x);
5776 hash = HASH (x, mode);
5777 elt = lookup (x, hash, mode);
5778 if (!elt)
5780 if (insert_regs (x, NULL, false))
5782 rtx dest = SET_DEST (sets[i].rtl);
5784 rehash_using_reg (x);
5785 hash = HASH (x, mode);
5786 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5788 elt = insert (x, NULL, hash, mode);
5791 sets[i].dest_addr_elt = elt;
5793 else
5794 sets[i].dest_addr_elt = NULL;
5798 invalidate_from_clobbers (insn);
5800 /* Some registers are invalidated by subroutine calls. Memory is
5801 invalidated by non-constant calls. */
5803 if (CALL_P (insn))
5805 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5806 invalidate_memory ();
5807 else
5808 /* For const/pure calls, invalidate any argument slots, because
5809 those are owned by the callee. */
5810 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
5811 if (GET_CODE (XEXP (tem, 0)) == USE
5812 && MEM_P (XEXP (XEXP (tem, 0), 0)))
5813 invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode);
5814 invalidate_for_call (insn);
5817 /* Now invalidate everything set by this instruction.
5818 If a SUBREG or other funny destination is being set,
5819 sets[i].rtl is still nonzero, so here we invalidate the reg
5820 a part of which is being set. */
5822 for (i = 0; i < n_sets; i++)
5823 if (sets[i].rtl)
5825 /* We can't use the inner dest, because the mode associated with
5826 a ZERO_EXTRACT is significant. */
5827 rtx dest = SET_DEST (sets[i].rtl);
5829 /* Needed for registers to remove the register from its
5830 previous quantity's chain.
5831 Needed for memory if this is a nonvarying address, unless
5832 we have just done an invalidate_memory that covers even those. */
5833 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5834 invalidate (dest, VOIDmode);
5835 else if (MEM_P (dest))
5836 invalidate (dest, VOIDmode);
5837 else if (GET_CODE (dest) == STRICT_LOW_PART
5838 || GET_CODE (dest) == ZERO_EXTRACT)
5839 invalidate (XEXP (dest, 0), GET_MODE (dest));
5842 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5843 the regs restored by the longjmp come from a later time
5844 than the setjmp. */
5845 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5847 flush_hash_table ();
5848 goto done;
5851 /* Make sure registers mentioned in destinations
5852 are safe for use in an expression to be inserted.
5853 This removes from the hash table
5854 any invalid entry that refers to one of these registers.
5856 We don't care about the return value from mention_regs because
5857 we are going to hash the SET_DEST values unconditionally. */
5859 for (i = 0; i < n_sets; i++)
5861 if (sets[i].rtl)
5863 rtx x = SET_DEST (sets[i].rtl);
5865 if (!REG_P (x))
5866 mention_regs (x);
5867 else
5869 /* We used to rely on all references to a register becoming
5870 inaccessible when a register changes to a new quantity,
5871 since that changes the hash code. However, that is not
5872 safe, since after HASH_SIZE new quantities we get a
5873 hash 'collision' of a register with its own invalid
5874 entries. And since SUBREGs have been changed not to
5875 change their hash code with the hash code of the register,
5876 it wouldn't work any longer at all. So we have to check
5877 for any invalid references lying around now.
5878 This code is similar to the REG case in mention_regs,
5879 but it knows that reg_tick has been incremented, and
5880 it leaves reg_in_table as -1 . */
5881 unsigned int regno = REGNO (x);
5882 unsigned int endregno = END_REGNO (x);
5883 unsigned int i;
5885 for (i = regno; i < endregno; i++)
5887 if (REG_IN_TABLE (i) >= 0)
5889 remove_invalid_refs (i);
5890 REG_IN_TABLE (i) = -1;
5897 /* We may have just removed some of the src_elt's from the hash table.
5898 So replace each one with the current head of the same class.
5899 Also check if destination addresses have been removed. */
5901 for (i = 0; i < n_sets; i++)
5902 if (sets[i].rtl)
5904 if (sets[i].dest_addr_elt
5905 && sets[i].dest_addr_elt->first_same_value == 0)
5907 /* The elt was removed, which means this destination is not
5908 valid after this instruction. */
5909 sets[i].rtl = NULL_RTX;
5911 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5912 /* If elt was removed, find current head of same class,
5913 or 0 if nothing remains of that class. */
5915 struct table_elt *elt = sets[i].src_elt;
5917 while (elt && elt->prev_same_value)
5918 elt = elt->prev_same_value;
5920 while (elt && elt->first_same_value == 0)
5921 elt = elt->next_same_value;
5922 sets[i].src_elt = elt ? elt->first_same_value : 0;
5926 /* Now insert the destinations into their equivalence classes. */
5928 for (i = 0; i < n_sets; i++)
5929 if (sets[i].rtl)
5931 rtx dest = SET_DEST (sets[i].rtl);
5932 struct table_elt *elt;
5934 /* Don't record value if we are not supposed to risk allocating
5935 floating-point values in registers that might be wider than
5936 memory. */
5937 if ((flag_float_store
5938 && MEM_P (dest)
5939 && FLOAT_MODE_P (GET_MODE (dest)))
5940 /* Don't record BLKmode values, because we don't know the
5941 size of it, and can't be sure that other BLKmode values
5942 have the same or smaller size. */
5943 || GET_MODE (dest) == BLKmode
5944 /* If we didn't put a REG_EQUAL value or a source into the hash
5945 table, there is no point is recording DEST. */
5946 || sets[i].src_elt == 0)
5947 continue;
5949 /* STRICT_LOW_PART isn't part of the value BEING set,
5950 and neither is the SUBREG inside it.
5951 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5952 if (GET_CODE (dest) == STRICT_LOW_PART)
5953 dest = SUBREG_REG (XEXP (dest, 0));
5955 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5956 /* Registers must also be inserted into chains for quantities. */
5957 if (insert_regs (dest, sets[i].src_elt, true))
5959 /* If `insert_regs' changes something, the hash code must be
5960 recalculated. */
5961 rehash_using_reg (dest);
5962 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5965 /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
5966 outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
5967 if (paradoxical_subreg_p (dest))
5968 continue;
5970 elt = insert (dest, sets[i].src_elt,
5971 sets[i].dest_hash, GET_MODE (dest));
5973 /* If this is a constant, insert the constant anchors with the
5974 equivalent register-offset expressions using register DEST. */
5975 if (targetm.const_anchor
5976 && REG_P (dest)
5977 && SCALAR_INT_MODE_P (GET_MODE (dest))
5978 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5979 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5981 elt->in_memory = (MEM_P (sets[i].inner_dest)
5982 && !MEM_READONLY_P (sets[i].inner_dest));
5984 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5985 narrower than M2, and both M1 and M2 are the same number of words,
5986 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5987 make that equivalence as well.
5989 However, BAR may have equivalences for which gen_lowpart
5990 will produce a simpler value than gen_lowpart applied to
5991 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5992 BAR's equivalences. If we don't get a simplified form, make
5993 the SUBREG. It will not be used in an equivalence, but will
5994 cause two similar assignments to be detected.
5996 Note the loop below will find SUBREG_REG (DEST) since we have
5997 already entered SRC and DEST of the SET in the table. */
5999 if (GET_CODE (dest) == SUBREG
6000 && (known_equal_after_align_down
6001 (GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1,
6002 GET_MODE_SIZE (GET_MODE (dest)) - 1,
6003 UNITS_PER_WORD))
6004 && !partial_subreg_p (dest)
6005 && sets[i].src_elt != 0)
6007 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6008 struct table_elt *elt, *classp = 0;
6010 for (elt = sets[i].src_elt->first_same_value; elt;
6011 elt = elt->next_same_value)
6013 rtx new_src = 0;
6014 unsigned src_hash;
6015 struct table_elt *src_elt;
6017 /* Ignore invalid entries. */
6018 if (!REG_P (elt->exp)
6019 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6020 continue;
6022 /* We may have already been playing subreg games. If the
6023 mode is already correct for the destination, use it. */
6024 if (GET_MODE (elt->exp) == new_mode)
6025 new_src = elt->exp;
6026 else
6028 poly_uint64 byte
6029 = subreg_lowpart_offset (new_mode, GET_MODE (dest));
6030 new_src = simplify_gen_subreg (new_mode, elt->exp,
6031 GET_MODE (dest), byte);
6034 /* The call to simplify_gen_subreg fails if the value
6035 is VOIDmode, yet we can't do any simplification, e.g.
6036 for EXPR_LISTs denoting function call results.
6037 It is invalid to construct a SUBREG with a VOIDmode
6038 SUBREG_REG, hence a zero new_src means we can't do
6039 this substitution. */
6040 if (! new_src)
6041 continue;
6043 src_hash = HASH (new_src, new_mode);
6044 src_elt = lookup (new_src, src_hash, new_mode);
6046 /* Put the new source in the hash table is if isn't
6047 already. */
6048 if (src_elt == 0)
6050 if (insert_regs (new_src, classp, false))
6052 rehash_using_reg (new_src);
6053 src_hash = HASH (new_src, new_mode);
6055 src_elt = insert (new_src, classp, src_hash, new_mode);
6056 src_elt->in_memory = elt->in_memory;
6057 if (GET_CODE (new_src) == ASM_OPERANDS
6058 && elt->cost == MAX_COST)
6059 src_elt->cost = MAX_COST;
6061 else if (classp && classp != src_elt->first_same_value)
6062 /* Show that two things that we've seen before are
6063 actually the same. */
6064 merge_equiv_classes (src_elt, classp);
6066 classp = src_elt->first_same_value;
6067 /* Ignore invalid entries. */
6068 while (classp
6069 && !REG_P (classp->exp)
6070 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6071 classp = classp->next_same_value;
6076 /* Special handling for (set REG0 REG1) where REG0 is the
6077 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6078 be used in the sequel, so (if easily done) change this insn to
6079 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6080 that computed their value. Then REG1 will become a dead store
6081 and won't cloud the situation for later optimizations.
6083 Do not make this change if REG1 is a hard register, because it will
6084 then be used in the sequel and we may be changing a two-operand insn
6085 into a three-operand insn.
6087 Also do not do this if we are operating on a copy of INSN. */
6089 if (n_sets == 1 && sets[0].rtl)
6090 try_back_substitute_reg (sets[0].rtl, insn);
6092 done:;
6095 /* Remove from the hash table all expressions that reference memory. */
6097 static void
6098 invalidate_memory (void)
6100 int i;
6101 struct table_elt *p, *next;
6103 for (i = 0; i < HASH_SIZE; i++)
6104 for (p = table[i]; p; p = next)
6106 next = p->next_same_hash;
6107 if (p->in_memory)
6108 remove_from_table (p, i);
6112 /* Perform invalidation on the basis of everything about INSN,
6113 except for invalidating the actual places that are SET in it.
6114 This includes the places CLOBBERed, and anything that might
6115 alias with something that is SET or CLOBBERed. */
6117 static void
6118 invalidate_from_clobbers (rtx_insn *insn)
6120 rtx x = PATTERN (insn);
6122 if (GET_CODE (x) == CLOBBER)
6124 rtx ref = XEXP (x, 0);
6125 if (ref)
6127 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6128 || MEM_P (ref))
6129 invalidate (ref, VOIDmode);
6130 else if (GET_CODE (ref) == STRICT_LOW_PART
6131 || GET_CODE (ref) == ZERO_EXTRACT)
6132 invalidate (XEXP (ref, 0), GET_MODE (ref));
6135 else if (GET_CODE (x) == PARALLEL)
6137 int i;
6138 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6140 rtx y = XVECEXP (x, 0, i);
6141 if (GET_CODE (y) == CLOBBER)
6143 rtx ref = XEXP (y, 0);
6144 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6145 || MEM_P (ref))
6146 invalidate (ref, VOIDmode);
6147 else if (GET_CODE (ref) == STRICT_LOW_PART
6148 || GET_CODE (ref) == ZERO_EXTRACT)
6149 invalidate (XEXP (ref, 0), GET_MODE (ref));
6155 /* Perform invalidation on the basis of everything about INSN.
6156 This includes the places CLOBBERed, and anything that might
6157 alias with something that is SET or CLOBBERed. */
6159 static void
6160 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6162 rtx tem;
6163 rtx x = PATTERN (insn);
6165 if (CALL_P (insn))
6167 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6169 rtx temx = XEXP (tem, 0);
6170 if (GET_CODE (temx) == CLOBBER)
6171 invalidate (SET_DEST (temx), VOIDmode);
6175 /* Ensure we invalidate the destination register of a CALL insn.
6176 This is necessary for machines where this register is a fixed_reg,
6177 because no other code would invalidate it. */
6178 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6179 invalidate (SET_DEST (x), VOIDmode);
6181 else if (GET_CODE (x) == PARALLEL)
6183 int i;
6185 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6187 rtx y = XVECEXP (x, 0, i);
6188 if (GET_CODE (y) == CLOBBER)
6190 rtx clobbered = XEXP (y, 0);
6192 if (REG_P (clobbered)
6193 || GET_CODE (clobbered) == SUBREG)
6194 invalidate (clobbered, VOIDmode);
6195 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6196 || GET_CODE (clobbered) == ZERO_EXTRACT)
6197 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6199 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6200 invalidate (SET_DEST (y), VOIDmode);
6205 static rtx cse_process_note (rtx);
6207 /* A simplify_replace_fn_rtx callback for cse_process_note. Process X,
6208 part of the REG_NOTES of an insn. Replace any registers with either
6209 an equivalent constant or the canonical form of the register.
6210 Only replace addresses if the containing MEM remains valid.
6212 Return the replacement for X, or null if it should be simplified
6213 recursively. */
6215 static rtx
6216 cse_process_note_1 (rtx x, const_rtx, void *)
6218 if (MEM_P (x))
6220 validate_change (x, &XEXP (x, 0), cse_process_note (XEXP (x, 0)), false);
6221 return x;
6224 if (REG_P (x))
6226 int i = REG_QTY (REGNO (x));
6228 /* Return a constant or a constant register. */
6229 if (REGNO_QTY_VALID_P (REGNO (x)))
6231 struct qty_table_elem *ent = &qty_table[i];
6233 if (ent->const_rtx != NULL_RTX
6234 && (CONSTANT_P (ent->const_rtx)
6235 || REG_P (ent->const_rtx)))
6237 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6238 if (new_rtx)
6239 return copy_rtx (new_rtx);
6243 /* Otherwise, canonicalize this register. */
6244 return canon_reg (x, NULL);
6247 return NULL_RTX;
6250 /* Process X, part of the REG_NOTES of an insn. Replace any registers in it
6251 with either an equivalent constant or the canonical form of the register.
6252 Only replace addresses if the containing MEM remains valid. */
6254 static rtx
6255 cse_process_note (rtx x)
6257 return simplify_replace_fn_rtx (x, NULL_RTX, cse_process_note_1, NULL);
6261 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6263 DATA is a pointer to a struct cse_basic_block_data, that is used to
6264 describe the path.
6265 It is filled with a queue of basic blocks, starting with FIRST_BB
6266 and following a trace through the CFG.
6268 If all paths starting at FIRST_BB have been followed, or no new path
6269 starting at FIRST_BB can be constructed, this function returns FALSE.
6270 Otherwise, DATA->path is filled and the function returns TRUE indicating
6271 that a path to follow was found.
6273 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6274 block in the path will be FIRST_BB. */
6276 static bool
6277 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6278 int follow_jumps)
6280 basic_block bb;
6281 edge e;
6282 int path_size;
6284 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6286 /* See if there is a previous path. */
6287 path_size = data->path_size;
6289 /* There is a previous path. Make sure it started with FIRST_BB. */
6290 if (path_size)
6291 gcc_assert (data->path[0].bb == first_bb);
6293 /* There was only one basic block in the last path. Clear the path and
6294 return, so that paths starting at another basic block can be tried. */
6295 if (path_size == 1)
6297 path_size = 0;
6298 goto done;
6301 /* If the path was empty from the beginning, construct a new path. */
6302 if (path_size == 0)
6303 data->path[path_size++].bb = first_bb;
6304 else
6306 /* Otherwise, path_size must be equal to or greater than 2, because
6307 a previous path exists that is at least two basic blocks long.
6309 Update the previous branch path, if any. If the last branch was
6310 previously along the branch edge, take the fallthrough edge now. */
6311 while (path_size >= 2)
6313 basic_block last_bb_in_path, previous_bb_in_path;
6314 edge e;
6316 --path_size;
6317 last_bb_in_path = data->path[path_size].bb;
6318 previous_bb_in_path = data->path[path_size - 1].bb;
6320 /* If we previously followed a path along the branch edge, try
6321 the fallthru edge now. */
6322 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6323 && any_condjump_p (BB_END (previous_bb_in_path))
6324 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6325 && e == BRANCH_EDGE (previous_bb_in_path))
6327 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6328 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6329 && single_pred_p (bb)
6330 /* We used to assert here that we would only see blocks
6331 that we have not visited yet. But we may end up
6332 visiting basic blocks twice if the CFG has changed
6333 in this run of cse_main, because when the CFG changes
6334 the topological sort of the CFG also changes. A basic
6335 blocks that previously had more than two predecessors
6336 may now have a single predecessor, and become part of
6337 a path that starts at another basic block.
6339 We still want to visit each basic block only once, so
6340 halt the path here if we have already visited BB. */
6341 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6343 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6344 data->path[path_size++].bb = bb;
6345 break;
6349 data->path[path_size].bb = NULL;
6352 /* If only one block remains in the path, bail. */
6353 if (path_size == 1)
6355 path_size = 0;
6356 goto done;
6360 /* Extend the path if possible. */
6361 if (follow_jumps)
6363 bb = data->path[path_size - 1].bb;
6364 while (bb && path_size < param_max_cse_path_length)
6366 if (single_succ_p (bb))
6367 e = single_succ_edge (bb);
6368 else if (EDGE_COUNT (bb->succs) == 2
6369 && any_condjump_p (BB_END (bb)))
6371 /* First try to follow the branch. If that doesn't lead
6372 to a useful path, follow the fallthru edge. */
6373 e = BRANCH_EDGE (bb);
6374 if (!single_pred_p (e->dest))
6375 e = FALLTHRU_EDGE (bb);
6377 else
6378 e = NULL;
6380 if (e
6381 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6382 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6383 && single_pred_p (e->dest)
6384 /* Avoid visiting basic blocks twice. The large comment
6385 above explains why this can happen. */
6386 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6388 basic_block bb2 = e->dest;
6389 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6390 data->path[path_size++].bb = bb2;
6391 bb = bb2;
6393 else
6394 bb = NULL;
6398 done:
6399 data->path_size = path_size;
6400 return path_size != 0;
6403 /* Dump the path in DATA to file F. NSETS is the number of sets
6404 in the path. */
6406 static void
6407 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6409 int path_entry;
6411 fprintf (f, ";; Following path with %d sets: ", nsets);
6412 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6413 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6414 fputc ('\n', f);
6415 fflush (f);
6419 /* Return true if BB has exception handling successor edges. */
6421 static bool
6422 have_eh_succ_edges (basic_block bb)
6424 edge e;
6425 edge_iterator ei;
6427 FOR_EACH_EDGE (e, ei, bb->succs)
6428 if (e->flags & EDGE_EH)
6429 return true;
6431 return false;
6435 /* Scan to the end of the path described by DATA. Return an estimate of
6436 the total number of SETs of all insns in the path. */
6438 static void
6439 cse_prescan_path (struct cse_basic_block_data *data)
6441 int nsets = 0;
6442 int path_size = data->path_size;
6443 int path_entry;
6445 /* Scan to end of each basic block in the path. */
6446 for (path_entry = 0; path_entry < path_size; path_entry++)
6448 basic_block bb;
6449 rtx_insn *insn;
6451 bb = data->path[path_entry].bb;
6453 FOR_BB_INSNS (bb, insn)
6455 if (!INSN_P (insn))
6456 continue;
6458 /* A PARALLEL can have lots of SETs in it,
6459 especially if it is really an ASM_OPERANDS. */
6460 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6461 nsets += XVECLEN (PATTERN (insn), 0);
6462 else
6463 nsets += 1;
6467 data->nsets = nsets;
6470 /* Return true if the pattern of INSN uses a LABEL_REF for which
6471 there isn't a REG_LABEL_OPERAND note. */
6473 static bool
6474 check_for_label_ref (rtx_insn *insn)
6476 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6477 note for it, we must rerun jump since it needs to place the note. If
6478 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6479 don't do this since no REG_LABEL_OPERAND will be added. */
6480 subrtx_iterator::array_type array;
6481 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6483 const_rtx x = *iter;
6484 if (GET_CODE (x) == LABEL_REF
6485 && !LABEL_REF_NONLOCAL_P (x)
6486 && (!JUMP_P (insn)
6487 || !label_is_jump_target_p (label_ref_label (x), insn))
6488 && LABEL_P (label_ref_label (x))
6489 && INSN_UID (label_ref_label (x)) != 0
6490 && !find_reg_note (insn, REG_LABEL_OPERAND, label_ref_label (x)))
6491 return true;
6493 return false;
6496 /* Process a single extended basic block described by EBB_DATA. */
6498 static void
6499 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6501 int path_size = ebb_data->path_size;
6502 int path_entry;
6503 int num_insns = 0;
6505 /* Allocate the space needed by qty_table. */
6506 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6508 new_basic_block ();
6509 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6510 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6511 for (path_entry = 0; path_entry < path_size; path_entry++)
6513 basic_block bb;
6514 rtx_insn *insn;
6516 bb = ebb_data->path[path_entry].bb;
6518 /* Invalidate recorded information for eh regs if there is an EH
6519 edge pointing to that bb. */
6520 if (bb_has_eh_pred (bb))
6522 df_ref def;
6524 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6525 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6526 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6529 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6530 FOR_BB_INSNS (bb, insn)
6532 /* If we have processed 1,000 insns, flush the hash table to
6533 avoid extreme quadratic behavior. We must not include NOTEs
6534 in the count since there may be more of them when generating
6535 debugging information. If we clear the table at different
6536 times, code generated with -g -O might be different than code
6537 generated with -O but not -g.
6539 FIXME: This is a real kludge and needs to be done some other
6540 way. */
6541 if (NONDEBUG_INSN_P (insn)
6542 && num_insns++ > param_max_cse_insns)
6544 flush_hash_table ();
6545 num_insns = 0;
6548 if (INSN_P (insn))
6550 /* Process notes first so we have all notes in canonical forms
6551 when looking for duplicate operations. */
6552 bool changed = false;
6553 for (rtx note = REG_NOTES (insn); note; note = XEXP (note, 1))
6554 if (REG_NOTE_KIND (note) == REG_EQUAL)
6556 rtx newval = cse_process_note (XEXP (note, 0));
6557 if (newval != XEXP (note, 0))
6559 XEXP (note, 0) = newval;
6560 changed = true;
6563 if (changed)
6564 df_notes_rescan (insn);
6566 cse_insn (insn);
6568 /* If we haven't already found an insn where we added a LABEL_REF,
6569 check this one. */
6570 if (INSN_P (insn) && !recorded_label_ref
6571 && check_for_label_ref (insn))
6572 recorded_label_ref = true;
6576 /* With non-call exceptions, we are not always able to update
6577 the CFG properly inside cse_insn. So clean up possibly
6578 redundant EH edges here. */
6579 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6580 cse_cfg_altered |= purge_dead_edges (bb);
6582 /* If we changed a conditional jump, we may have terminated
6583 the path we are following. Check that by verifying that
6584 the edge we would take still exists. If the edge does
6585 not exist anymore, purge the remainder of the path.
6586 Note that this will cause us to return to the caller. */
6587 if (path_entry < path_size - 1)
6589 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6590 if (!find_edge (bb, next_bb))
6594 path_size--;
6596 /* If we truncate the path, we must also reset the
6597 visited bit on the remaining blocks in the path,
6598 or we will never visit them at all. */
6599 bitmap_clear_bit (cse_visited_basic_blocks,
6600 ebb_data->path[path_size].bb->index);
6601 ebb_data->path[path_size].bb = NULL;
6603 while (path_size - 1 != path_entry);
6604 ebb_data->path_size = path_size;
6608 /* If this is a conditional jump insn, record any known
6609 equivalences due to the condition being tested. */
6610 insn = BB_END (bb);
6611 if (path_entry < path_size - 1
6612 && EDGE_COUNT (bb->succs) == 2
6613 && JUMP_P (insn)
6614 && single_set (insn)
6615 && any_condjump_p (insn))
6617 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6618 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6619 record_jump_equiv (insn, taken);
6623 gcc_assert (next_qty <= max_qty);
6625 free (qty_table);
6629 /* Perform cse on the instructions of a function.
6630 F is the first instruction.
6631 NREGS is one plus the highest pseudo-reg number used in the instruction.
6633 Return 2 if jump optimizations should be redone due to simplifications
6634 in conditional jump instructions.
6635 Return 1 if the CFG should be cleaned up because it has been modified.
6636 Return 0 otherwise. */
6638 static int
6639 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6641 struct cse_basic_block_data ebb_data;
6642 basic_block bb;
6643 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6644 int i, n_blocks;
6646 /* CSE doesn't use dominane info but can invalidate it in different ways.
6647 For simplicity free dominance info here. */
6648 free_dominance_info (CDI_DOMINATORS);
6650 df_set_flags (DF_LR_RUN_DCE);
6651 df_note_add_problem ();
6652 df_analyze ();
6653 df_set_flags (DF_DEFER_INSN_RESCAN);
6655 reg_scan (get_insns (), max_reg_num ());
6656 init_cse_reg_info (nregs);
6658 ebb_data.path = XNEWVEC (struct branch_path,
6659 param_max_cse_path_length);
6661 cse_cfg_altered = false;
6662 cse_jumps_altered = false;
6663 recorded_label_ref = false;
6664 ebb_data.path_size = 0;
6665 ebb_data.nsets = 0;
6666 rtl_hooks = cse_rtl_hooks;
6668 init_recog ();
6669 init_alias_analysis ();
6671 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6673 /* Set up the table of already visited basic blocks. */
6674 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6675 bitmap_clear (cse_visited_basic_blocks);
6677 /* Loop over basic blocks in reverse completion order (RPO),
6678 excluding the ENTRY and EXIT blocks. */
6679 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6680 i = 0;
6681 while (i < n_blocks)
6683 /* Find the first block in the RPO queue that we have not yet
6684 processed before. */
6687 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6689 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6690 && i < n_blocks);
6692 /* Find all paths starting with BB, and process them. */
6693 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6695 /* Pre-scan the path. */
6696 cse_prescan_path (&ebb_data);
6698 /* If this basic block has no sets, skip it. */
6699 if (ebb_data.nsets == 0)
6700 continue;
6702 /* Get a reasonable estimate for the maximum number of qty's
6703 needed for this path. For this, we take the number of sets
6704 and multiply that by MAX_RECOG_OPERANDS. */
6705 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6707 /* Dump the path we're about to process. */
6708 if (dump_file)
6709 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6711 cse_extended_basic_block (&ebb_data);
6715 /* Clean up. */
6716 end_alias_analysis ();
6717 free (reg_eqv_table);
6718 free (ebb_data.path);
6719 sbitmap_free (cse_visited_basic_blocks);
6720 free (rc_order);
6721 rtl_hooks = general_rtl_hooks;
6723 if (cse_jumps_altered || recorded_label_ref)
6724 return 2;
6725 else if (cse_cfg_altered)
6726 return 1;
6727 else
6728 return 0;
6731 /* Count the number of times registers are used (not set) in X.
6732 COUNTS is an array in which we accumulate the count, INCR is how much
6733 we count each register usage.
6735 Don't count a usage of DEST, which is the SET_DEST of a SET which
6736 contains X in its SET_SRC. This is because such a SET does not
6737 modify the liveness of DEST.
6738 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6739 We must then count uses of a SET_DEST regardless, because the insn can't be
6740 deleted here. */
6742 static void
6743 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6745 enum rtx_code code;
6746 rtx note;
6747 const char *fmt;
6748 int i, j;
6750 if (x == 0)
6751 return;
6753 switch (code = GET_CODE (x))
6755 case REG:
6756 if (x != dest)
6757 counts[REGNO (x)] += incr;
6758 return;
6760 case PC:
6761 case CONST:
6762 CASE_CONST_ANY:
6763 case SYMBOL_REF:
6764 case LABEL_REF:
6765 return;
6767 case CLOBBER:
6768 /* If we are clobbering a MEM, mark any registers inside the address
6769 as being used. */
6770 if (MEM_P (XEXP (x, 0)))
6771 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6772 return;
6774 case SET:
6775 /* Unless we are setting a REG, count everything in SET_DEST. */
6776 if (!REG_P (SET_DEST (x)))
6777 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6778 count_reg_usage (SET_SRC (x), counts,
6779 dest ? dest : SET_DEST (x),
6780 incr);
6781 return;
6783 case DEBUG_INSN:
6784 return;
6786 case CALL_INSN:
6787 case INSN:
6788 case JUMP_INSN:
6789 /* We expect dest to be NULL_RTX here. If the insn may throw,
6790 or if it cannot be deleted due to side-effects, mark this fact
6791 by setting DEST to pc_rtx. */
6792 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6793 || side_effects_p (PATTERN (x)))
6794 dest = pc_rtx;
6795 if (code == CALL_INSN)
6796 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6797 count_reg_usage (PATTERN (x), counts, dest, incr);
6799 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6800 use them. */
6802 note = find_reg_equal_equiv_note (x);
6803 if (note)
6805 rtx eqv = XEXP (note, 0);
6807 if (GET_CODE (eqv) == EXPR_LIST)
6808 /* This REG_EQUAL note describes the result of a function call.
6809 Process all the arguments. */
6812 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6813 eqv = XEXP (eqv, 1);
6815 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6816 else
6817 count_reg_usage (eqv, counts, dest, incr);
6819 return;
6821 case EXPR_LIST:
6822 if (REG_NOTE_KIND (x) == REG_EQUAL
6823 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6824 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6825 involving registers in the address. */
6826 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6827 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6829 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6830 return;
6832 case ASM_OPERANDS:
6833 /* Iterate over just the inputs, not the constraints as well. */
6834 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6835 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6836 return;
6838 case INSN_LIST:
6839 case INT_LIST:
6840 gcc_unreachable ();
6842 default:
6843 break;
6846 fmt = GET_RTX_FORMAT (code);
6847 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6849 if (fmt[i] == 'e')
6850 count_reg_usage (XEXP (x, i), counts, dest, incr);
6851 else if (fmt[i] == 'E')
6852 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6853 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6857 /* Return true if X is a dead register. */
6859 static inline bool
6860 is_dead_reg (const_rtx x, int *counts)
6862 return (REG_P (x)
6863 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6864 && counts[REGNO (x)] == 0);
6867 /* Return true if set is live. */
6868 static bool
6869 set_live_p (rtx set, int *counts)
6871 if (set_noop_p (set))
6872 return false;
6874 if (!is_dead_reg (SET_DEST (set), counts)
6875 || side_effects_p (SET_SRC (set)))
6876 return true;
6878 return false;
6881 /* Return true if insn is live. */
6883 static bool
6884 insn_live_p (rtx_insn *insn, int *counts)
6886 int i;
6887 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6888 return true;
6889 else if (GET_CODE (PATTERN (insn)) == SET)
6890 return set_live_p (PATTERN (insn), counts);
6891 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6893 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6895 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6897 if (GET_CODE (elt) == SET)
6899 if (set_live_p (elt, counts))
6900 return true;
6902 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6903 return true;
6905 return false;
6907 else if (DEBUG_INSN_P (insn))
6909 if (DEBUG_MARKER_INSN_P (insn))
6910 return true;
6912 if (DEBUG_BIND_INSN_P (insn)
6913 && TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)))
6914 return false;
6916 return true;
6918 else
6919 return true;
6922 /* Count the number of stores into pseudo. Callback for note_stores. */
6924 static void
6925 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6927 int *counts = (int *) data;
6928 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6929 counts[REGNO (x)]++;
6932 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6933 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6934 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6935 Set *SEEN_REPL to true if we see a dead register that does have
6936 a replacement. */
6938 static bool
6939 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6940 bool *seen_repl)
6942 subrtx_iterator::array_type array;
6943 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6945 const_rtx x = *iter;
6946 if (is_dead_reg (x, counts))
6948 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6949 *seen_repl = true;
6950 else
6951 return true;
6954 return false;
6957 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6958 Callback for simplify_replace_fn_rtx. */
6960 static rtx
6961 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6963 rtx *replacements = (rtx *) data;
6965 if (REG_P (x)
6966 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6967 && replacements[REGNO (x)] != NULL_RTX)
6969 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6970 return replacements[REGNO (x)];
6971 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6972 GET_MODE (replacements[REGNO (x)]));
6974 return NULL_RTX;
6977 /* Scan all the insns and delete any that are dead; i.e., they store a register
6978 that is never used or they copy a register to itself.
6980 This is used to remove insns made obviously dead by cse, loop or other
6981 optimizations. It improves the heuristics in loop since it won't try to
6982 move dead invariants out of loops or make givs for dead quantities. The
6983 remaining passes of the compilation are also sped up. */
6986 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
6988 int *counts;
6989 rtx_insn *insn, *prev;
6990 rtx *replacements = NULL;
6991 int ndead = 0;
6993 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6994 /* First count the number of times each register is used. */
6995 if (MAY_HAVE_DEBUG_BIND_INSNS)
6997 counts = XCNEWVEC (int, nreg * 3);
6998 for (insn = insns; insn; insn = NEXT_INSN (insn))
6999 if (DEBUG_BIND_INSN_P (insn))
7001 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7002 NULL_RTX, 1);
7003 TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)) = 0;
7005 else if (INSN_P (insn))
7007 count_reg_usage (insn, counts, NULL_RTX, 1);
7008 note_stores (insn, count_stores, counts + nreg * 2);
7010 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7011 First one counts how many times each pseudo is used outside
7012 of debug insns, second counts how many times each pseudo is
7013 used in debug insns and third counts how many times a pseudo
7014 is stored. */
7016 else
7018 counts = XCNEWVEC (int, nreg);
7019 for (insn = insns; insn; insn = NEXT_INSN (insn))
7020 if (INSN_P (insn))
7021 count_reg_usage (insn, counts, NULL_RTX, 1);
7022 /* If no debug insns can be present, COUNTS is just an array
7023 which counts how many times each pseudo is used. */
7025 /* Pseudo PIC register should be considered as used due to possible
7026 new usages generated. */
7027 if (!reload_completed
7028 && pic_offset_table_rtx
7029 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7030 counts[REGNO (pic_offset_table_rtx)]++;
7031 /* Go from the last insn to the first and delete insns that only set unused
7032 registers or copy a register to itself. As we delete an insn, remove
7033 usage counts for registers it uses.
7035 The first jump optimization pass may leave a real insn as the last
7036 insn in the function. We must not skip that insn or we may end
7037 up deleting code that is not really dead.
7039 If some otherwise unused register is only used in DEBUG_INSNs,
7040 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7041 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7042 has been created for the unused register, replace it with
7043 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7044 auto_vec<tree, 32> later_debug_set_vars;
7045 for (insn = get_last_insn (); insn; insn = prev)
7047 int live_insn = 0;
7049 prev = PREV_INSN (insn);
7050 if (!INSN_P (insn))
7051 continue;
7053 live_insn = insn_live_p (insn, counts);
7055 /* If this is a dead insn, delete it and show registers in it aren't
7056 being used. */
7058 if (! live_insn && dbg_cnt (delete_trivial_dead))
7060 if (DEBUG_INSN_P (insn))
7062 if (DEBUG_BIND_INSN_P (insn))
7063 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7064 NULL_RTX, -1);
7066 else
7068 rtx set;
7069 if (MAY_HAVE_DEBUG_BIND_INSNS
7070 && (set = single_set (insn)) != NULL_RTX
7071 && is_dead_reg (SET_DEST (set), counts)
7072 /* Used at least once in some DEBUG_INSN. */
7073 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7074 /* And set exactly once. */
7075 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7076 && !side_effects_p (SET_SRC (set))
7077 && asm_noperands (PATTERN (insn)) < 0)
7079 rtx dval, bind_var_loc;
7080 rtx_insn *bind;
7082 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7083 dval = make_debug_expr_from_rtl (SET_DEST (set));
7085 /* Emit a debug bind insn before the insn in which
7086 reg dies. */
7087 bind_var_loc =
7088 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7089 DEBUG_EXPR_TREE_DECL (dval),
7090 SET_SRC (set),
7091 VAR_INIT_STATUS_INITIALIZED);
7092 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7094 bind = emit_debug_insn_before (bind_var_loc, insn);
7095 df_insn_rescan (bind);
7097 if (replacements == NULL)
7098 replacements = XCNEWVEC (rtx, nreg);
7099 replacements[REGNO (SET_DEST (set))] = dval;
7102 count_reg_usage (insn, counts, NULL_RTX, -1);
7103 ndead++;
7105 cse_cfg_altered |= delete_insn_and_edges (insn);
7107 else
7109 if (!DEBUG_INSN_P (insn) || DEBUG_MARKER_INSN_P (insn))
7111 for (tree var : later_debug_set_vars)
7112 TREE_VISITED (var) = 0;
7113 later_debug_set_vars.truncate (0);
7115 else if (DEBUG_BIND_INSN_P (insn)
7116 && !TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)))
7118 later_debug_set_vars.safe_push (INSN_VAR_LOCATION_DECL (insn));
7119 TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)) = 1;
7124 if (MAY_HAVE_DEBUG_BIND_INSNS)
7126 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7127 if (DEBUG_BIND_INSN_P (insn))
7129 /* If this debug insn references a dead register that wasn't replaced
7130 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7131 bool seen_repl = false;
7132 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7133 counts, replacements, &seen_repl))
7135 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7136 df_insn_rescan (insn);
7138 else if (seen_repl)
7140 INSN_VAR_LOCATION_LOC (insn)
7141 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7142 NULL_RTX, replace_dead_reg,
7143 replacements);
7144 df_insn_rescan (insn);
7147 free (replacements);
7150 if (dump_file && ndead)
7151 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7152 ndead);
7153 /* Clean up. */
7154 free (counts);
7155 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7156 return ndead;
7159 /* If LOC contains references to NEWREG in a different mode, change them
7160 to use NEWREG instead. */
7162 static void
7163 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7164 rtx *loc, rtx_insn *insn, rtx newreg)
7166 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7168 rtx *loc = *iter;
7169 rtx x = *loc;
7170 if (x
7171 && REG_P (x)
7172 && REGNO (x) == REGNO (newreg)
7173 && GET_MODE (x) != GET_MODE (newreg))
7175 validate_change (insn, loc, newreg, 1);
7176 iter.skip_subrtxes ();
7181 /* Change the mode of any reference to the register REGNO (NEWREG) to
7182 GET_MODE (NEWREG) in INSN. */
7184 static void
7185 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7187 int success;
7189 if (!INSN_P (insn))
7190 return;
7192 subrtx_ptr_iterator::array_type array;
7193 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7194 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7196 /* If the following assertion was triggered, there is most probably
7197 something wrong with the cc_modes_compatible back end function.
7198 CC modes only can be considered compatible if the insn - with the mode
7199 replaced by any of the compatible modes - can still be recognized. */
7200 success = apply_change_group ();
7201 gcc_assert (success);
7204 /* Change the mode of any reference to the register REGNO (NEWREG) to
7205 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7206 any instruction which modifies NEWREG. */
7208 static void
7209 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7211 rtx_insn *insn;
7213 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7215 if (! INSN_P (insn))
7216 continue;
7218 if (reg_set_p (newreg, insn))
7219 return;
7221 cse_change_cc_mode_insn (insn, newreg);
7225 /* BB is a basic block which finishes with CC_REG as a condition code
7226 register which is set to CC_SRC. Look through the successors of BB
7227 to find blocks which have a single predecessor (i.e., this one),
7228 and look through those blocks for an assignment to CC_REG which is
7229 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7230 permitted to change the mode of CC_SRC to a compatible mode. This
7231 returns VOIDmode if no equivalent assignments were found.
7232 Otherwise it returns the mode which CC_SRC should wind up with.
7233 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7234 but is passed unmodified down to recursive calls in order to prevent
7235 endless recursion.
7237 The main complexity in this function is handling the mode issues.
7238 We may have more than one duplicate which we can eliminate, and we
7239 try to find a mode which will work for multiple duplicates. */
7241 static machine_mode
7242 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7243 bool can_change_mode)
7245 bool found_equiv;
7246 machine_mode mode;
7247 unsigned int insn_count;
7248 edge e;
7249 rtx_insn *insns[2];
7250 machine_mode modes[2];
7251 rtx_insn *last_insns[2];
7252 unsigned int i;
7253 rtx newreg;
7254 edge_iterator ei;
7256 /* We expect to have two successors. Look at both before picking
7257 the final mode for the comparison. If we have more successors
7258 (i.e., some sort of table jump, although that seems unlikely),
7259 then we require all beyond the first two to use the same
7260 mode. */
7262 found_equiv = false;
7263 mode = GET_MODE (cc_src);
7264 insn_count = 0;
7265 FOR_EACH_EDGE (e, ei, bb->succs)
7267 rtx_insn *insn;
7268 rtx_insn *end;
7270 if (e->flags & EDGE_COMPLEX)
7271 continue;
7273 if (EDGE_COUNT (e->dest->preds) != 1
7274 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7275 /* Avoid endless recursion on unreachable blocks. */
7276 || e->dest == orig_bb)
7277 continue;
7279 end = NEXT_INSN (BB_END (e->dest));
7280 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7282 rtx set;
7284 if (! INSN_P (insn))
7285 continue;
7287 /* If CC_SRC is modified, we have to stop looking for
7288 something which uses it. */
7289 if (modified_in_p (cc_src, insn))
7290 break;
7292 /* Check whether INSN sets CC_REG to CC_SRC. */
7293 set = single_set (insn);
7294 if (set
7295 && REG_P (SET_DEST (set))
7296 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7298 bool found;
7299 machine_mode set_mode;
7300 machine_mode comp_mode;
7302 found = false;
7303 set_mode = GET_MODE (SET_SRC (set));
7304 comp_mode = set_mode;
7305 if (rtx_equal_p (cc_src, SET_SRC (set)))
7306 found = true;
7307 else if (GET_CODE (cc_src) == COMPARE
7308 && GET_CODE (SET_SRC (set)) == COMPARE
7309 && mode != set_mode
7310 && rtx_equal_p (XEXP (cc_src, 0),
7311 XEXP (SET_SRC (set), 0))
7312 && rtx_equal_p (XEXP (cc_src, 1),
7313 XEXP (SET_SRC (set), 1)))
7316 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7317 if (comp_mode != VOIDmode
7318 && (can_change_mode || comp_mode == mode))
7319 found = true;
7322 if (found)
7324 found_equiv = true;
7325 if (insn_count < ARRAY_SIZE (insns))
7327 insns[insn_count] = insn;
7328 modes[insn_count] = set_mode;
7329 last_insns[insn_count] = end;
7330 ++insn_count;
7332 if (mode != comp_mode)
7334 gcc_assert (can_change_mode);
7335 mode = comp_mode;
7337 /* The modified insn will be re-recognized later. */
7338 PUT_MODE (cc_src, mode);
7341 else
7343 if (set_mode != mode)
7345 /* We found a matching expression in the
7346 wrong mode, but we don't have room to
7347 store it in the array. Punt. This case
7348 should be rare. */
7349 break;
7351 /* INSN sets CC_REG to a value equal to CC_SRC
7352 with the right mode. We can simply delete
7353 it. */
7354 delete_insn (insn);
7357 /* We found an instruction to delete. Keep looking,
7358 in the hopes of finding a three-way jump. */
7359 continue;
7362 /* We found an instruction which sets the condition
7363 code, so don't look any farther. */
7364 break;
7367 /* If INSN sets CC_REG in some other way, don't look any
7368 farther. */
7369 if (reg_set_p (cc_reg, insn))
7370 break;
7373 /* If we fell off the bottom of the block, we can keep looking
7374 through successors. We pass CAN_CHANGE_MODE as false because
7375 we aren't prepared to handle compatibility between the
7376 further blocks and this block. */
7377 if (insn == end)
7379 machine_mode submode;
7381 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7382 if (submode != VOIDmode)
7384 gcc_assert (submode == mode);
7385 found_equiv = true;
7386 can_change_mode = false;
7391 if (! found_equiv)
7392 return VOIDmode;
7394 /* Now INSN_COUNT is the number of instructions we found which set
7395 CC_REG to a value equivalent to CC_SRC. The instructions are in
7396 INSNS. The modes used by those instructions are in MODES. */
7398 newreg = NULL_RTX;
7399 for (i = 0; i < insn_count; ++i)
7401 if (modes[i] != mode)
7403 /* We need to change the mode of CC_REG in INSNS[i] and
7404 subsequent instructions. */
7405 if (! newreg)
7407 if (GET_MODE (cc_reg) == mode)
7408 newreg = cc_reg;
7409 else
7410 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7412 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7413 newreg);
7416 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7419 return mode;
7422 /* If we have a fixed condition code register (or two), walk through
7423 the instructions and try to eliminate duplicate assignments. */
7425 static void
7426 cse_condition_code_reg (void)
7428 unsigned int cc_regno_1;
7429 unsigned int cc_regno_2;
7430 rtx cc_reg_1;
7431 rtx cc_reg_2;
7432 basic_block bb;
7434 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7435 return;
7437 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7438 if (cc_regno_2 != INVALID_REGNUM)
7439 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7440 else
7441 cc_reg_2 = NULL_RTX;
7443 FOR_EACH_BB_FN (bb, cfun)
7445 rtx_insn *last_insn;
7446 rtx cc_reg;
7447 rtx_insn *insn;
7448 rtx_insn *cc_src_insn;
7449 rtx cc_src;
7450 machine_mode mode;
7451 machine_mode orig_mode;
7453 /* Look for blocks which end with a conditional jump based on a
7454 condition code register. Then look for the instruction which
7455 sets the condition code register. Then look through the
7456 successor blocks for instructions which set the condition
7457 code register to the same value. There are other possible
7458 uses of the condition code register, but these are by far the
7459 most common and the ones which we are most likely to be able
7460 to optimize. */
7462 last_insn = BB_END (bb);
7463 if (!JUMP_P (last_insn))
7464 continue;
7466 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7467 cc_reg = cc_reg_1;
7468 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7469 cc_reg = cc_reg_2;
7470 else
7471 continue;
7473 cc_src_insn = NULL;
7474 cc_src = NULL_RTX;
7475 for (insn = PREV_INSN (last_insn);
7476 insn && insn != PREV_INSN (BB_HEAD (bb));
7477 insn = PREV_INSN (insn))
7479 rtx set;
7481 if (! INSN_P (insn))
7482 continue;
7483 set = single_set (insn);
7484 if (set
7485 && REG_P (SET_DEST (set))
7486 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7488 cc_src_insn = insn;
7489 cc_src = SET_SRC (set);
7490 break;
7492 else if (reg_set_p (cc_reg, insn))
7493 break;
7496 if (! cc_src_insn)
7497 continue;
7499 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7500 continue;
7502 /* Now CC_REG is a condition code register used for a
7503 conditional jump at the end of the block, and CC_SRC, in
7504 CC_SRC_INSN, is the value to which that condition code
7505 register is set, and CC_SRC is still meaningful at the end of
7506 the basic block. */
7508 orig_mode = GET_MODE (cc_src);
7509 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7510 if (mode != VOIDmode)
7512 gcc_assert (mode == GET_MODE (cc_src));
7513 if (mode != orig_mode)
7515 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7517 cse_change_cc_mode_insn (cc_src_insn, newreg);
7519 /* Do the same in the following insns that use the
7520 current value of CC_REG within BB. */
7521 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7522 NEXT_INSN (last_insn),
7523 newreg);
7530 /* Perform common subexpression elimination. Nonzero value from
7531 `cse_main' means that jumps were simplified and some code may now
7532 be unreachable, so do jump optimization again. */
7533 static unsigned int
7534 rest_of_handle_cse (void)
7536 int tem;
7538 if (dump_file)
7539 dump_flow_info (dump_file, dump_flags);
7541 tem = cse_main (get_insns (), max_reg_num ());
7543 /* If we are not running more CSE passes, then we are no longer
7544 expecting CSE to be run. But always rerun it in a cheap mode. */
7545 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7547 if (tem == 2)
7549 timevar_push (TV_JUMP);
7550 rebuild_jump_labels (get_insns ());
7551 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7552 timevar_pop (TV_JUMP);
7554 else if (tem == 1 || optimize > 1)
7555 cse_cfg_altered |= cleanup_cfg (0);
7557 return 0;
7560 namespace {
7562 const pass_data pass_data_cse =
7564 RTL_PASS, /* type */
7565 "cse1", /* name */
7566 OPTGROUP_NONE, /* optinfo_flags */
7567 TV_CSE, /* tv_id */
7568 0, /* properties_required */
7569 0, /* properties_provided */
7570 0, /* properties_destroyed */
7571 0, /* todo_flags_start */
7572 TODO_df_finish, /* todo_flags_finish */
7575 class pass_cse : public rtl_opt_pass
7577 public:
7578 pass_cse (gcc::context *ctxt)
7579 : rtl_opt_pass (pass_data_cse, ctxt)
7582 /* opt_pass methods: */
7583 bool gate (function *) final override { return optimize > 0; }
7584 unsigned int execute (function *) final override
7586 return rest_of_handle_cse ();
7589 }; // class pass_cse
7591 } // anon namespace
7593 rtl_opt_pass *
7594 make_pass_cse (gcc::context *ctxt)
7596 return new pass_cse (ctxt);
7600 /* Run second CSE pass after loop optimizations. */
7601 static unsigned int
7602 rest_of_handle_cse2 (void)
7604 int tem;
7606 if (dump_file)
7607 dump_flow_info (dump_file, dump_flags);
7609 tem = cse_main (get_insns (), max_reg_num ());
7611 /* Run a pass to eliminate duplicated assignments to condition code
7612 registers. We have to run this after bypass_jumps, because it
7613 makes it harder for that pass to determine whether a jump can be
7614 bypassed safely. */
7615 cse_condition_code_reg ();
7617 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7619 if (tem == 2)
7621 timevar_push (TV_JUMP);
7622 rebuild_jump_labels (get_insns ());
7623 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7624 timevar_pop (TV_JUMP);
7626 else if (tem == 1 || cse_cfg_altered)
7627 cse_cfg_altered |= cleanup_cfg (0);
7629 cse_not_expected = 1;
7630 return 0;
7634 namespace {
7636 const pass_data pass_data_cse2 =
7638 RTL_PASS, /* type */
7639 "cse2", /* name */
7640 OPTGROUP_NONE, /* optinfo_flags */
7641 TV_CSE2, /* tv_id */
7642 0, /* properties_required */
7643 0, /* properties_provided */
7644 0, /* properties_destroyed */
7645 0, /* todo_flags_start */
7646 TODO_df_finish, /* todo_flags_finish */
7649 class pass_cse2 : public rtl_opt_pass
7651 public:
7652 pass_cse2 (gcc::context *ctxt)
7653 : rtl_opt_pass (pass_data_cse2, ctxt)
7656 /* opt_pass methods: */
7657 bool gate (function *) final override
7659 return optimize > 0 && flag_rerun_cse_after_loop;
7662 unsigned int execute (function *) final override
7664 return rest_of_handle_cse2 ();
7667 }; // class pass_cse2
7669 } // anon namespace
7671 rtl_opt_pass *
7672 make_pass_cse2 (gcc::context *ctxt)
7674 return new pass_cse2 (ctxt);
7677 /* Run second CSE pass after loop optimizations. */
7678 static unsigned int
7679 rest_of_handle_cse_after_global_opts (void)
7681 int save_cfj;
7682 int tem;
7684 /* We only want to do local CSE, so don't follow jumps. */
7685 save_cfj = flag_cse_follow_jumps;
7686 flag_cse_follow_jumps = 0;
7688 rebuild_jump_labels (get_insns ());
7689 tem = cse_main (get_insns (), max_reg_num ());
7690 cse_cfg_altered |= purge_all_dead_edges ();
7691 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7693 cse_not_expected = !flag_rerun_cse_after_loop;
7695 /* If cse altered any jumps, rerun jump opts to clean things up. */
7696 if (tem == 2)
7698 timevar_push (TV_JUMP);
7699 rebuild_jump_labels (get_insns ());
7700 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7701 timevar_pop (TV_JUMP);
7703 else if (tem == 1 || cse_cfg_altered)
7704 cse_cfg_altered |= cleanup_cfg (0);
7706 flag_cse_follow_jumps = save_cfj;
7707 return 0;
7710 namespace {
7712 const pass_data pass_data_cse_after_global_opts =
7714 RTL_PASS, /* type */
7715 "cse_local", /* name */
7716 OPTGROUP_NONE, /* optinfo_flags */
7717 TV_CSE, /* tv_id */
7718 0, /* properties_required */
7719 0, /* properties_provided */
7720 0, /* properties_destroyed */
7721 0, /* todo_flags_start */
7722 TODO_df_finish, /* todo_flags_finish */
7725 class pass_cse_after_global_opts : public rtl_opt_pass
7727 public:
7728 pass_cse_after_global_opts (gcc::context *ctxt)
7729 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7732 /* opt_pass methods: */
7733 bool gate (function *) final override
7735 return optimize > 0 && flag_rerun_cse_after_global_opts;
7738 unsigned int execute (function *) final override
7740 return rest_of_handle_cse_after_global_opts ();
7743 }; // class pass_cse_after_global_opts
7745 } // anon namespace
7747 rtl_opt_pass *
7748 make_pass_cse_after_global_opts (gcc::context *ctxt)
7750 return new pass_cse_after_global_opts (ctxt);