Support TI mode and soft float on PA64
[official-gcc.git] / gcc / cse.c
blobc1c7d0ca27b73c4b944b4719f95fece74e0358d5
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2021 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 /* The sizes of these fields should match the sizes of the
252 code and mode fields of struct rtx_def (see rtl.h). */
253 ENUM_BITFIELD(rtx_code) comparison_code : 16;
254 ENUM_BITFIELD(machine_mode) mode : 8;
257 /* The table of all qtys, indexed by qty number. */
258 static struct qty_table_elem *qty_table;
260 /* Insn being scanned. */
262 static rtx_insn *this_insn;
263 static bool optimize_this_for_speed_p;
265 /* Index by register number, gives the number of the next (or
266 previous) register in the chain of registers sharing the same
267 value.
269 Or -1 if this register is at the end of the chain.
271 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
273 /* Per-register equivalence chain. */
274 struct reg_eqv_elem
276 int next, prev;
279 /* The table of all register equivalence chains. */
280 static struct reg_eqv_elem *reg_eqv_table;
282 struct cse_reg_info
284 /* The timestamp at which this register is initialized. */
285 unsigned int timestamp;
287 /* The quantity number of the register's current contents. */
288 int reg_qty;
290 /* The number of times the register has been altered in the current
291 basic block. */
292 int reg_tick;
294 /* The REG_TICK value at which rtx's containing this register are
295 valid in the hash table. If this does not equal the current
296 reg_tick value, such expressions existing in the hash table are
297 invalid. */
298 int reg_in_table;
300 /* The SUBREG that was set when REG_TICK was last incremented. Set
301 to -1 if the last store was to the whole register, not a subreg. */
302 unsigned int subreg_ticked;
305 /* A table of cse_reg_info indexed by register numbers. */
306 static struct cse_reg_info *cse_reg_info_table;
308 /* The size of the above table. */
309 static unsigned int cse_reg_info_table_size;
311 /* The index of the first entry that has not been initialized. */
312 static unsigned int cse_reg_info_table_first_uninitialized;
314 /* The timestamp at the beginning of the current run of
315 cse_extended_basic_block. We increment this variable at the beginning of
316 the current run of cse_extended_basic_block. The timestamp field of a
317 cse_reg_info entry matches the value of this variable if and only
318 if the entry has been initialized during the current run of
319 cse_extended_basic_block. */
320 static unsigned int cse_reg_info_timestamp;
322 /* A HARD_REG_SET containing all the hard registers for which there is
323 currently a REG expression in the hash table. Note the difference
324 from the above variables, which indicate if the REG is mentioned in some
325 expression in the table. */
327 static HARD_REG_SET hard_regs_in_table;
329 /* True if CSE has altered the CFG. */
330 static bool cse_cfg_altered;
332 /* True if CSE has altered conditional jump insns in such a way
333 that jump optimization should be redone. */
334 static bool cse_jumps_altered;
336 /* True if we put a LABEL_REF into the hash table for an INSN
337 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
338 to put in the note. */
339 static bool recorded_label_ref;
341 /* canon_hash stores 1 in do_not_record if it notices a reference to PC or
342 some other volatile subexpression. */
344 static int do_not_record;
346 /* canon_hash stores 1 in hash_arg_in_memory
347 if it notices a reference to memory within the expression being hashed. */
349 static int hash_arg_in_memory;
351 /* The hash table contains buckets which are chains of `struct table_elt's,
352 each recording one expression's information.
353 That expression is in the `exp' field.
355 The canon_exp field contains a canonical (from the point of view of
356 alias analysis) version of the `exp' field.
358 Those elements with the same hash code are chained in both directions
359 through the `next_same_hash' and `prev_same_hash' fields.
361 Each set of expressions with equivalent values
362 are on a two-way chain through the `next_same_value'
363 and `prev_same_value' fields, and all point with
364 the `first_same_value' field at the first element in
365 that chain. The chain is in order of increasing cost.
366 Each element's cost value is in its `cost' field.
368 The `in_memory' field is nonzero for elements that
369 involve any reference to memory. These elements are removed
370 whenever a write is done to an unidentified location in memory.
371 To be safe, we assume that a memory address is unidentified unless
372 the address is either a symbol constant or a constant plus
373 the frame pointer or argument pointer.
375 The `related_value' field is used to connect related expressions
376 (that differ by adding an integer).
377 The related expressions are chained in a circular fashion.
378 `related_value' is zero for expressions for which this
379 chain is not useful.
381 The `cost' field stores the cost of this element's expression.
382 The `regcost' field stores the value returned by approx_reg_cost for
383 this element's expression.
385 The `is_const' flag is set if the element is a constant (including
386 a fixed address).
388 The `flag' field is used as a temporary during some search routines.
390 The `mode' field is usually the same as GET_MODE (`exp'), but
391 if `exp' is a CONST_INT and has no machine mode then the `mode'
392 field is the mode it was being used as. Each constant is
393 recorded separately for each mode it is used with. */
395 struct table_elt
397 rtx exp;
398 rtx canon_exp;
399 struct table_elt *next_same_hash;
400 struct table_elt *prev_same_hash;
401 struct table_elt *next_same_value;
402 struct table_elt *prev_same_value;
403 struct table_elt *first_same_value;
404 struct table_elt *related_value;
405 int cost;
406 int regcost;
407 /* The size of this field should match the size
408 of the mode field of struct rtx_def (see rtl.h). */
409 ENUM_BITFIELD(machine_mode) mode : 8;
410 char in_memory;
411 char is_const;
412 char flag;
415 /* We don't want a lot of buckets, because we rarely have very many
416 things stored in the hash table, and a lot of buckets slows
417 down a lot of loops that happen frequently. */
418 #define HASH_SHIFT 5
419 #define HASH_SIZE (1 << HASH_SHIFT)
420 #define HASH_MASK (HASH_SIZE - 1)
422 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
423 register (hard registers may require `do_not_record' to be set). */
425 #define HASH(X, M) \
426 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
427 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
428 : canon_hash (X, M)) & HASH_MASK)
430 /* Like HASH, but without side-effects. */
431 #define SAFE_HASH(X, M) \
432 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
433 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
434 : safe_hash (X, M)) & HASH_MASK)
436 /* Determine whether register number N is considered a fixed register for the
437 purpose of approximating register costs.
438 It is desirable to replace other regs with fixed regs, to reduce need for
439 non-fixed hard regs.
440 A reg wins if it is either the frame pointer or designated as fixed. */
441 #define FIXED_REGNO_P(N) \
442 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
443 || fixed_regs[N] || global_regs[N])
445 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
446 hard registers and pointers into the frame are the cheapest with a cost
447 of 0. Next come pseudos with a cost of one and other hard registers with
448 a cost of 2. Aside from these special cases, call `rtx_cost'. */
450 #define CHEAP_REGNO(N) \
451 (REGNO_PTR_FRAME_P (N) \
452 || (HARD_REGISTER_NUM_P (N) \
453 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
455 #define COST(X, MODE) \
456 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
457 #define COST_IN(X, MODE, OUTER, OPNO) \
458 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
460 /* Get the number of times this register has been updated in this
461 basic block. */
463 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
465 /* Get the point at which REG was recorded in the table. */
467 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
469 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
470 SUBREG). */
472 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
474 /* Get the quantity number for REG. */
476 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
478 /* Determine if the quantity number for register X represents a valid index
479 into the qty_table. */
481 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
483 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
485 #define CHEAPER(X, Y) \
486 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
488 static struct table_elt *table[HASH_SIZE];
490 /* Chain of `struct table_elt's made so far for this function
491 but currently removed from the table. */
493 static struct table_elt *free_element_chain;
495 /* Trace a patch through the CFG. */
497 struct branch_path
499 /* The basic block for this path entry. */
500 basic_block bb;
503 /* This data describes a block that will be processed by
504 cse_extended_basic_block. */
506 struct cse_basic_block_data
508 /* Total number of SETs in block. */
509 int nsets;
510 /* Size of current branch path, if any. */
511 int path_size;
512 /* Current path, indicating which basic_blocks will be processed. */
513 struct branch_path *path;
517 /* Pointers to the live in/live out bitmaps for the boundaries of the
518 current EBB. */
519 static bitmap cse_ebb_live_in, cse_ebb_live_out;
521 /* A simple bitmap to track which basic blocks have been visited
522 already as part of an already processed extended basic block. */
523 static sbitmap cse_visited_basic_blocks;
525 static bool fixed_base_plus_p (rtx x);
526 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
527 static int preferable (int, int, int, int);
528 static void new_basic_block (void);
529 static void make_new_qty (unsigned int, machine_mode);
530 static void make_regs_eqv (unsigned int, unsigned int);
531 static void delete_reg_equiv (unsigned int);
532 static int mention_regs (rtx);
533 static int insert_regs (rtx, struct table_elt *, int);
534 static void remove_from_table (struct table_elt *, unsigned);
535 static void remove_pseudo_from_table (rtx, unsigned);
536 static struct table_elt *lookup (rtx, unsigned, machine_mode);
537 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
538 static rtx lookup_as_function (rtx, enum rtx_code);
539 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
540 machine_mode, int, int);
541 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
542 machine_mode);
543 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
544 static void invalidate (rtx, machine_mode);
545 static void remove_invalid_refs (unsigned int);
546 static void remove_invalid_subreg_refs (unsigned int, poly_uint64,
547 machine_mode);
548 static void rehash_using_reg (rtx);
549 static void invalidate_memory (void);
550 static rtx use_related_value (rtx, struct table_elt *);
552 static inline unsigned canon_hash (rtx, machine_mode);
553 static inline unsigned safe_hash (rtx, machine_mode);
554 static inline unsigned hash_rtx_string (const char *);
556 static rtx canon_reg (rtx, rtx_insn *);
557 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
558 machine_mode *,
559 machine_mode *);
560 static rtx fold_rtx (rtx, rtx_insn *);
561 static rtx equiv_constant (rtx);
562 static void record_jump_equiv (rtx_insn *, bool);
563 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
564 int);
565 static void cse_insn (rtx_insn *);
566 static void cse_prescan_path (struct cse_basic_block_data *);
567 static void invalidate_from_clobbers (rtx_insn *);
568 static void invalidate_from_sets_and_clobbers (rtx_insn *);
569 static void cse_extended_basic_block (struct cse_basic_block_data *);
570 extern void dump_class (struct table_elt*);
571 static void get_cse_reg_info_1 (unsigned int regno);
572 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
574 static void flush_hash_table (void);
575 static bool insn_live_p (rtx_insn *, int *);
576 static bool set_live_p (rtx, int *);
577 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
578 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
579 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
580 bool);
583 #undef RTL_HOOKS_GEN_LOWPART
584 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
586 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
588 /* Nonzero if X has the form (PLUS frame-pointer integer). */
590 static bool
591 fixed_base_plus_p (rtx x)
593 switch (GET_CODE (x))
595 case REG:
596 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
597 return true;
598 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
599 return true;
600 return false;
602 case PLUS:
603 if (!CONST_INT_P (XEXP (x, 1)))
604 return false;
605 return fixed_base_plus_p (XEXP (x, 0));
607 default:
608 return false;
612 /* Dump the expressions in the equivalence class indicated by CLASSP.
613 This function is used only for debugging. */
614 DEBUG_FUNCTION void
615 dump_class (struct table_elt *classp)
617 struct table_elt *elt;
619 fprintf (stderr, "Equivalence chain for ");
620 print_rtl (stderr, classp->exp);
621 fprintf (stderr, ": \n");
623 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
625 print_rtl (stderr, elt->exp);
626 fprintf (stderr, "\n");
630 /* Return an estimate of the cost of the registers used in an rtx.
631 This is mostly the number of different REG expressions in the rtx;
632 however for some exceptions like fixed registers we use a cost of
633 0. If any other hard register reference occurs, return MAX_COST. */
635 static int
636 approx_reg_cost (const_rtx x)
638 int cost = 0;
639 subrtx_iterator::array_type array;
640 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
642 const_rtx x = *iter;
643 if (REG_P (x))
645 unsigned int regno = REGNO (x);
646 if (!CHEAP_REGNO (regno))
648 if (regno < FIRST_PSEUDO_REGISTER)
650 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
651 return MAX_COST;
652 cost += 2;
654 else
655 cost += 1;
659 return cost;
662 /* Return a negative value if an rtx A, whose costs are given by COST_A
663 and REGCOST_A, is more desirable than an rtx B.
664 Return a positive value if A is less desirable, or 0 if the two are
665 equally good. */
666 static int
667 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
669 /* First, get rid of cases involving expressions that are entirely
670 unwanted. */
671 if (cost_a != cost_b)
673 if (cost_a == MAX_COST)
674 return 1;
675 if (cost_b == MAX_COST)
676 return -1;
679 /* Avoid extending lifetimes of hardregs. */
680 if (regcost_a != regcost_b)
682 if (regcost_a == MAX_COST)
683 return 1;
684 if (regcost_b == MAX_COST)
685 return -1;
688 /* Normal operation costs take precedence. */
689 if (cost_a != cost_b)
690 return cost_a - cost_b;
691 /* Only if these are identical consider effects on register pressure. */
692 if (regcost_a != regcost_b)
693 return regcost_a - regcost_b;
694 return 0;
697 /* Internal function, to compute cost when X is not a register; called
698 from COST macro to keep it simple. */
700 static int
701 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
703 scalar_int_mode int_mode, inner_mode;
704 return ((GET_CODE (x) == SUBREG
705 && REG_P (SUBREG_REG (x))
706 && is_int_mode (mode, &int_mode)
707 && is_int_mode (GET_MODE (SUBREG_REG (x)), &inner_mode)
708 && GET_MODE_SIZE (int_mode) < GET_MODE_SIZE (inner_mode)
709 && subreg_lowpart_p (x)
710 && TRULY_NOOP_TRUNCATION_MODES_P (int_mode, inner_mode))
712 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
716 /* Initialize CSE_REG_INFO_TABLE. */
718 static void
719 init_cse_reg_info (unsigned int nregs)
721 /* Do we need to grow the table? */
722 if (nregs > cse_reg_info_table_size)
724 unsigned int new_size;
726 if (cse_reg_info_table_size < 2048)
728 /* Compute a new size that is a power of 2 and no smaller
729 than the large of NREGS and 64. */
730 new_size = (cse_reg_info_table_size
731 ? cse_reg_info_table_size : 64);
733 while (new_size < nregs)
734 new_size *= 2;
736 else
738 /* If we need a big table, allocate just enough to hold
739 NREGS registers. */
740 new_size = nregs;
743 /* Reallocate the table with NEW_SIZE entries. */
744 free (cse_reg_info_table);
745 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
746 cse_reg_info_table_size = new_size;
747 cse_reg_info_table_first_uninitialized = 0;
750 /* Do we have all of the first NREGS entries initialized? */
751 if (cse_reg_info_table_first_uninitialized < nregs)
753 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
754 unsigned int i;
756 /* Put the old timestamp on newly allocated entries so that they
757 will all be considered out of date. We do not touch those
758 entries beyond the first NREGS entries to be nice to the
759 virtual memory. */
760 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
761 cse_reg_info_table[i].timestamp = old_timestamp;
763 cse_reg_info_table_first_uninitialized = nregs;
767 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
769 static void
770 get_cse_reg_info_1 (unsigned int regno)
772 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
773 entry will be considered to have been initialized. */
774 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
776 /* Initialize the rest of the entry. */
777 cse_reg_info_table[regno].reg_tick = 1;
778 cse_reg_info_table[regno].reg_in_table = -1;
779 cse_reg_info_table[regno].subreg_ticked = -1;
780 cse_reg_info_table[regno].reg_qty = -regno - 1;
783 /* Find a cse_reg_info entry for REGNO. */
785 static inline struct cse_reg_info *
786 get_cse_reg_info (unsigned int regno)
788 struct cse_reg_info *p = &cse_reg_info_table[regno];
790 /* If this entry has not been initialized, go ahead and initialize
791 it. */
792 if (p->timestamp != cse_reg_info_timestamp)
793 get_cse_reg_info_1 (regno);
795 return p;
798 /* Clear the hash table and initialize each register with its own quantity,
799 for a new basic block. */
801 static void
802 new_basic_block (void)
804 int i;
806 next_qty = 0;
808 /* Invalidate cse_reg_info_table. */
809 cse_reg_info_timestamp++;
811 /* Clear out hash table state for this pass. */
812 CLEAR_HARD_REG_SET (hard_regs_in_table);
814 /* The per-quantity values used to be initialized here, but it is
815 much faster to initialize each as it is made in `make_new_qty'. */
817 for (i = 0; i < HASH_SIZE; i++)
819 struct table_elt *first;
821 first = table[i];
822 if (first != NULL)
824 struct table_elt *last = first;
826 table[i] = NULL;
828 while (last->next_same_hash != NULL)
829 last = last->next_same_hash;
831 /* Now relink this hash entire chain into
832 the free element list. */
834 last->next_same_hash = free_element_chain;
835 free_element_chain = first;
840 /* Say that register REG contains a quantity in mode MODE not in any
841 register before and initialize that quantity. */
843 static void
844 make_new_qty (unsigned int reg, machine_mode mode)
846 int q;
847 struct qty_table_elem *ent;
848 struct reg_eqv_elem *eqv;
850 gcc_assert (next_qty < max_qty);
852 q = REG_QTY (reg) = next_qty++;
853 ent = &qty_table[q];
854 ent->first_reg = reg;
855 ent->last_reg = reg;
856 ent->mode = mode;
857 ent->const_rtx = ent->const_insn = NULL;
858 ent->comparison_code = UNKNOWN;
860 eqv = &reg_eqv_table[reg];
861 eqv->next = eqv->prev = -1;
864 /* Make reg NEW equivalent to reg OLD.
865 OLD is not changing; NEW is. */
867 static void
868 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
870 unsigned int lastr, firstr;
871 int q = REG_QTY (old_reg);
872 struct qty_table_elem *ent;
874 ent = &qty_table[q];
876 /* Nothing should become eqv until it has a "non-invalid" qty number. */
877 gcc_assert (REGNO_QTY_VALID_P (old_reg));
879 REG_QTY (new_reg) = q;
880 firstr = ent->first_reg;
881 lastr = ent->last_reg;
883 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
884 hard regs. Among pseudos, if NEW will live longer than any other reg
885 of the same qty, and that is beyond the current basic block,
886 make it the new canonical replacement for this qty. */
887 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
888 /* Certain fixed registers might be of the class NO_REGS. This means
889 that not only can they not be allocated by the compiler, but
890 they cannot be used in substitutions or canonicalizations
891 either. */
892 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
893 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
894 || (new_reg >= FIRST_PSEUDO_REGISTER
895 && (firstr < FIRST_PSEUDO_REGISTER
896 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
897 && !bitmap_bit_p (cse_ebb_live_out, firstr))
898 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
899 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
901 reg_eqv_table[firstr].prev = new_reg;
902 reg_eqv_table[new_reg].next = firstr;
903 reg_eqv_table[new_reg].prev = -1;
904 ent->first_reg = new_reg;
906 else
908 /* If NEW is a hard reg (known to be non-fixed), insert at end.
909 Otherwise, insert before any non-fixed hard regs that are at the
910 end. Registers of class NO_REGS cannot be used as an
911 equivalent for anything. */
912 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
913 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
914 && new_reg >= FIRST_PSEUDO_REGISTER)
915 lastr = reg_eqv_table[lastr].prev;
916 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
917 if (reg_eqv_table[lastr].next >= 0)
918 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
919 else
920 qty_table[q].last_reg = new_reg;
921 reg_eqv_table[lastr].next = new_reg;
922 reg_eqv_table[new_reg].prev = lastr;
926 /* Remove REG from its equivalence class. */
928 static void
929 delete_reg_equiv (unsigned int reg)
931 struct qty_table_elem *ent;
932 int q = REG_QTY (reg);
933 int p, n;
935 /* If invalid, do nothing. */
936 if (! REGNO_QTY_VALID_P (reg))
937 return;
939 ent = &qty_table[q];
941 p = reg_eqv_table[reg].prev;
942 n = reg_eqv_table[reg].next;
944 if (n != -1)
945 reg_eqv_table[n].prev = p;
946 else
947 ent->last_reg = p;
948 if (p != -1)
949 reg_eqv_table[p].next = n;
950 else
951 ent->first_reg = n;
953 REG_QTY (reg) = -reg - 1;
956 /* Remove any invalid expressions from the hash table
957 that refer to any of the registers contained in expression X.
959 Make sure that newly inserted references to those registers
960 as subexpressions will be considered valid.
962 mention_regs is not called when a register itself
963 is being stored in the table.
965 Return 1 if we have done something that may have changed the hash code
966 of X. */
968 static int
969 mention_regs (rtx x)
971 enum rtx_code code;
972 int i, j;
973 const char *fmt;
974 int changed = 0;
976 if (x == 0)
977 return 0;
979 code = GET_CODE (x);
980 if (code == REG)
982 unsigned int regno = REGNO (x);
983 unsigned int endregno = END_REGNO (x);
984 unsigned int i;
986 for (i = regno; i < endregno; i++)
988 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
989 remove_invalid_refs (i);
991 REG_IN_TABLE (i) = REG_TICK (i);
992 SUBREG_TICKED (i) = -1;
995 return 0;
998 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
999 pseudo if they don't use overlapping words. We handle only pseudos
1000 here for simplicity. */
1001 if (code == SUBREG && REG_P (SUBREG_REG (x))
1002 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1004 unsigned int i = REGNO (SUBREG_REG (x));
1006 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1008 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1009 the last store to this register really stored into this
1010 subreg, then remove the memory of this subreg.
1011 Otherwise, remove any memory of the entire register and
1012 all its subregs from the table. */
1013 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1014 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1015 remove_invalid_refs (i);
1016 else
1017 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1020 REG_IN_TABLE (i) = REG_TICK (i);
1021 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1022 return 0;
1025 /* If X is a comparison or a COMPARE and either operand is a register
1026 that does not have a quantity, give it one. This is so that a later
1027 call to record_jump_equiv won't cause X to be assigned a different
1028 hash code and not found in the table after that call.
1030 It is not necessary to do this here, since rehash_using_reg can
1031 fix up the table later, but doing this here eliminates the need to
1032 call that expensive function in the most common case where the only
1033 use of the register is in the comparison. */
1035 if (code == COMPARE || COMPARISON_P (x))
1037 if (REG_P (XEXP (x, 0))
1038 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1039 if (insert_regs (XEXP (x, 0), NULL, 0))
1041 rehash_using_reg (XEXP (x, 0));
1042 changed = 1;
1045 if (REG_P (XEXP (x, 1))
1046 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1047 if (insert_regs (XEXP (x, 1), NULL, 0))
1049 rehash_using_reg (XEXP (x, 1));
1050 changed = 1;
1054 fmt = GET_RTX_FORMAT (code);
1055 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1056 if (fmt[i] == 'e')
1057 changed |= mention_regs (XEXP (x, i));
1058 else if (fmt[i] == 'E')
1059 for (j = 0; j < XVECLEN (x, i); j++)
1060 changed |= mention_regs (XVECEXP (x, i, j));
1062 return changed;
1065 /* Update the register quantities for inserting X into the hash table
1066 with a value equivalent to CLASSP.
1067 (If the class does not contain a REG, it is irrelevant.)
1068 If MODIFIED is nonzero, X is a destination; it is being modified.
1069 Note that delete_reg_equiv should be called on a register
1070 before insert_regs is done on that register with MODIFIED != 0.
1072 Nonzero value means that elements of reg_qty have changed
1073 so X's hash code may be different. */
1075 static int
1076 insert_regs (rtx x, struct table_elt *classp, int modified)
1078 if (REG_P (x))
1080 unsigned int regno = REGNO (x);
1081 int qty_valid;
1083 /* If REGNO is in the equivalence table already but is of the
1084 wrong mode for that equivalence, don't do anything here. */
1086 qty_valid = REGNO_QTY_VALID_P (regno);
1087 if (qty_valid)
1089 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1091 if (ent->mode != GET_MODE (x))
1092 return 0;
1095 if (modified || ! qty_valid)
1097 if (classp)
1098 for (classp = classp->first_same_value;
1099 classp != 0;
1100 classp = classp->next_same_value)
1101 if (REG_P (classp->exp)
1102 && GET_MODE (classp->exp) == GET_MODE (x))
1104 unsigned c_regno = REGNO (classp->exp);
1106 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1108 /* Suppose that 5 is hard reg and 100 and 101 are
1109 pseudos. Consider
1111 (set (reg:si 100) (reg:si 5))
1112 (set (reg:si 5) (reg:si 100))
1113 (set (reg:di 101) (reg:di 5))
1115 We would now set REG_QTY (101) = REG_QTY (5), but the
1116 entry for 5 is in SImode. When we use this later in
1117 copy propagation, we get the register in wrong mode. */
1118 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1119 continue;
1121 make_regs_eqv (regno, c_regno);
1122 return 1;
1125 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1126 than REG_IN_TABLE to find out if there was only a single preceding
1127 invalidation - for the SUBREG - or another one, which would be
1128 for the full register. However, if we find here that REG_TICK
1129 indicates that the register is invalid, it means that it has
1130 been invalidated in a separate operation. The SUBREG might be used
1131 now (then this is a recursive call), or we might use the full REG
1132 now and a SUBREG of it later. So bump up REG_TICK so that
1133 mention_regs will do the right thing. */
1134 if (! modified
1135 && REG_IN_TABLE (regno) >= 0
1136 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1137 REG_TICK (regno)++;
1138 make_new_qty (regno, GET_MODE (x));
1139 return 1;
1142 return 0;
1145 /* If X is a SUBREG, we will likely be inserting the inner register in the
1146 table. If that register doesn't have an assigned quantity number at
1147 this point but does later, the insertion that we will be doing now will
1148 not be accessible because its hash code will have changed. So assign
1149 a quantity number now. */
1151 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1152 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1154 insert_regs (SUBREG_REG (x), NULL, 0);
1155 mention_regs (x);
1156 return 1;
1158 else
1159 return mention_regs (x);
1163 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1164 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1165 CST is equal to an anchor. */
1167 static bool
1168 compute_const_anchors (rtx cst,
1169 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1170 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1172 HOST_WIDE_INT n = INTVAL (cst);
1174 *lower_base = n & ~(targetm.const_anchor - 1);
1175 if (*lower_base == n)
1176 return false;
1178 *upper_base =
1179 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1180 *upper_offs = n - *upper_base;
1181 *lower_offs = n - *lower_base;
1182 return true;
1185 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1187 static void
1188 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1189 machine_mode mode)
1191 struct table_elt *elt;
1192 unsigned hash;
1193 rtx anchor_exp;
1194 rtx exp;
1196 anchor_exp = GEN_INT (anchor);
1197 hash = HASH (anchor_exp, mode);
1198 elt = lookup (anchor_exp, hash, mode);
1199 if (!elt)
1200 elt = insert (anchor_exp, NULL, hash, mode);
1202 exp = plus_constant (mode, reg, offs);
1203 /* REG has just been inserted and the hash codes recomputed. */
1204 mention_regs (exp);
1205 hash = HASH (exp, mode);
1207 /* Use the cost of the register rather than the whole expression. When
1208 looking up constant anchors we will further offset the corresponding
1209 expression therefore it does not make sense to prefer REGs over
1210 reg-immediate additions. Prefer instead the oldest expression. Also
1211 don't prefer pseudos over hard regs so that we derive constants in
1212 argument registers from other argument registers rather than from the
1213 original pseudo that was used to synthesize the constant. */
1214 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1217 /* The constant CST is equivalent to the register REG. Create
1218 equivalences between the two anchors of CST and the corresponding
1219 register-offset expressions using REG. */
1221 static void
1222 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1224 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1226 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1227 &upper_base, &upper_offs))
1228 return;
1230 /* Ignore anchors of value 0. Constants accessible from zero are
1231 simple. */
1232 if (lower_base != 0)
1233 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1235 if (upper_base != 0)
1236 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1239 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1240 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1241 valid expression. Return the cheapest and oldest of such expressions. In
1242 *OLD, return how old the resulting expression is compared to the other
1243 equivalent expressions. */
1245 static rtx
1246 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1247 unsigned *old)
1249 struct table_elt *elt;
1250 unsigned idx;
1251 struct table_elt *match_elt;
1252 rtx match;
1254 /* Find the cheapest and *oldest* expression to maximize the chance of
1255 reusing the same pseudo. */
1257 match_elt = NULL;
1258 match = NULL_RTX;
1259 for (elt = anchor_elt->first_same_value, idx = 0;
1260 elt;
1261 elt = elt->next_same_value, idx++)
1263 if (match_elt && CHEAPER (match_elt, elt))
1264 return match;
1266 if (REG_P (elt->exp)
1267 || (GET_CODE (elt->exp) == PLUS
1268 && REG_P (XEXP (elt->exp, 0))
1269 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1271 rtx x;
1273 /* Ignore expressions that are no longer valid. */
1274 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1275 continue;
1277 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1278 if (REG_P (x)
1279 || (GET_CODE (x) == PLUS
1280 && IN_RANGE (INTVAL (XEXP (x, 1)),
1281 -targetm.const_anchor,
1282 targetm.const_anchor - 1)))
1284 match = x;
1285 match_elt = elt;
1286 *old = idx;
1291 return match;
1294 /* Try to express the constant SRC_CONST using a register+offset expression
1295 derived from a constant anchor. Return it if successful or NULL_RTX,
1296 otherwise. */
1298 static rtx
1299 try_const_anchors (rtx src_const, machine_mode mode)
1301 struct table_elt *lower_elt, *upper_elt;
1302 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1303 rtx lower_anchor_rtx, upper_anchor_rtx;
1304 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1305 unsigned lower_old, upper_old;
1307 /* CONST_INT is used for CC modes, but we should leave those alone. */
1308 if (GET_MODE_CLASS (mode) == MODE_CC)
1309 return NULL_RTX;
1311 gcc_assert (SCALAR_INT_MODE_P (mode));
1312 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1313 &upper_base, &upper_offs))
1314 return NULL_RTX;
1316 lower_anchor_rtx = GEN_INT (lower_base);
1317 upper_anchor_rtx = GEN_INT (upper_base);
1318 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1319 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1321 if (lower_elt)
1322 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1323 if (upper_elt)
1324 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1326 if (!lower_exp)
1327 return upper_exp;
1328 if (!upper_exp)
1329 return lower_exp;
1331 /* Return the older expression. */
1332 return (upper_old > lower_old ? upper_exp : lower_exp);
1335 /* Look in or update the hash table. */
1337 /* Remove table element ELT from use in the table.
1338 HASH is its hash code, made using the HASH macro.
1339 It's an argument because often that is known in advance
1340 and we save much time not recomputing it. */
1342 static void
1343 remove_from_table (struct table_elt *elt, unsigned int hash)
1345 if (elt == 0)
1346 return;
1348 /* Mark this element as removed. See cse_insn. */
1349 elt->first_same_value = 0;
1351 /* Remove the table element from its equivalence class. */
1354 struct table_elt *prev = elt->prev_same_value;
1355 struct table_elt *next = elt->next_same_value;
1357 if (next)
1358 next->prev_same_value = prev;
1360 if (prev)
1361 prev->next_same_value = next;
1362 else
1364 struct table_elt *newfirst = next;
1365 while (next)
1367 next->first_same_value = newfirst;
1368 next = next->next_same_value;
1373 /* Remove the table element from its hash bucket. */
1376 struct table_elt *prev = elt->prev_same_hash;
1377 struct table_elt *next = elt->next_same_hash;
1379 if (next)
1380 next->prev_same_hash = prev;
1382 if (prev)
1383 prev->next_same_hash = next;
1384 else if (table[hash] == elt)
1385 table[hash] = next;
1386 else
1388 /* This entry is not in the proper hash bucket. This can happen
1389 when two classes were merged by `merge_equiv_classes'. Search
1390 for the hash bucket that it heads. This happens only very
1391 rarely, so the cost is acceptable. */
1392 for (hash = 0; hash < HASH_SIZE; hash++)
1393 if (table[hash] == elt)
1394 table[hash] = next;
1398 /* Remove the table element from its related-value circular chain. */
1400 if (elt->related_value != 0 && elt->related_value != elt)
1402 struct table_elt *p = elt->related_value;
1404 while (p->related_value != elt)
1405 p = p->related_value;
1406 p->related_value = elt->related_value;
1407 if (p->related_value == p)
1408 p->related_value = 0;
1411 /* Now add it to the free element chain. */
1412 elt->next_same_hash = free_element_chain;
1413 free_element_chain = elt;
1416 /* Same as above, but X is a pseudo-register. */
1418 static void
1419 remove_pseudo_from_table (rtx x, unsigned int hash)
1421 struct table_elt *elt;
1423 /* Because a pseudo-register can be referenced in more than one
1424 mode, we might have to remove more than one table entry. */
1425 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1426 remove_from_table (elt, hash);
1429 /* Look up X in the hash table and return its table element,
1430 or 0 if X is not in the table.
1432 MODE is the machine-mode of X, or if X is an integer constant
1433 with VOIDmode then MODE is the mode with which X will be used.
1435 Here we are satisfied to find an expression whose tree structure
1436 looks like X. */
1438 static struct table_elt *
1439 lookup (rtx x, unsigned int hash, machine_mode mode)
1441 struct table_elt *p;
1443 for (p = table[hash]; p; p = p->next_same_hash)
1444 if (mode == p->mode && ((x == p->exp && REG_P (x))
1445 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1446 return p;
1448 return 0;
1451 /* Like `lookup' but don't care whether the table element uses invalid regs.
1452 Also ignore discrepancies in the machine mode of a register. */
1454 static struct table_elt *
1455 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1457 struct table_elt *p;
1459 if (REG_P (x))
1461 unsigned int regno = REGNO (x);
1463 /* Don't check the machine mode when comparing registers;
1464 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1465 for (p = table[hash]; p; p = p->next_same_hash)
1466 if (REG_P (p->exp)
1467 && REGNO (p->exp) == regno)
1468 return p;
1470 else
1472 for (p = table[hash]; p; p = p->next_same_hash)
1473 if (mode == p->mode
1474 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1475 return p;
1478 return 0;
1481 /* Look for an expression equivalent to X and with code CODE.
1482 If one is found, return that expression. */
1484 static rtx
1485 lookup_as_function (rtx x, enum rtx_code code)
1487 struct table_elt *p
1488 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1490 if (p == 0)
1491 return 0;
1493 for (p = p->first_same_value; p; p = p->next_same_value)
1494 if (GET_CODE (p->exp) == code
1495 /* Make sure this is a valid entry in the table. */
1496 && exp_equiv_p (p->exp, p->exp, 1, false))
1497 return p->exp;
1499 return 0;
1502 /* Insert X in the hash table, assuming HASH is its hash code and
1503 CLASSP is an element of the class it should go in (or 0 if a new
1504 class should be made). COST is the code of X and reg_cost is the
1505 cost of registers in X. It is inserted at the proper position to
1506 keep the class in the order cheapest first.
1508 MODE is the machine-mode of X, or if X is an integer constant
1509 with VOIDmode then MODE is the mode with which X will be used.
1511 For elements of equal cheapness, the most recent one
1512 goes in front, except that the first element in the list
1513 remains first unless a cheaper element is added. The order of
1514 pseudo-registers does not matter, as canon_reg will be called to
1515 find the cheapest when a register is retrieved from the table.
1517 The in_memory field in the hash table element is set to 0.
1518 The caller must set it nonzero if appropriate.
1520 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1521 and if insert_regs returns a nonzero value
1522 you must then recompute its hash code before calling here.
1524 If necessary, update table showing constant values of quantities. */
1526 static struct table_elt *
1527 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1528 machine_mode mode, int cost, int reg_cost)
1530 struct table_elt *elt;
1532 /* If X is a register and we haven't made a quantity for it,
1533 something is wrong. */
1534 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1536 /* If X is a hard register, show it is being put in the table. */
1537 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1538 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1540 /* Put an element for X into the right hash bucket. */
1542 elt = free_element_chain;
1543 if (elt)
1544 free_element_chain = elt->next_same_hash;
1545 else
1546 elt = XNEW (struct table_elt);
1548 elt->exp = x;
1549 elt->canon_exp = NULL_RTX;
1550 elt->cost = cost;
1551 elt->regcost = reg_cost;
1552 elt->next_same_value = 0;
1553 elt->prev_same_value = 0;
1554 elt->next_same_hash = table[hash];
1555 elt->prev_same_hash = 0;
1556 elt->related_value = 0;
1557 elt->in_memory = 0;
1558 elt->mode = mode;
1559 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1561 if (table[hash])
1562 table[hash]->prev_same_hash = elt;
1563 table[hash] = elt;
1565 /* Put it into the proper value-class. */
1566 if (classp)
1568 classp = classp->first_same_value;
1569 if (CHEAPER (elt, classp))
1570 /* Insert at the head of the class. */
1572 struct table_elt *p;
1573 elt->next_same_value = classp;
1574 classp->prev_same_value = elt;
1575 elt->first_same_value = elt;
1577 for (p = classp; p; p = p->next_same_value)
1578 p->first_same_value = elt;
1580 else
1582 /* Insert not at head of the class. */
1583 /* Put it after the last element cheaper than X. */
1584 struct table_elt *p, *next;
1586 for (p = classp;
1587 (next = p->next_same_value) && CHEAPER (next, elt);
1588 p = next)
1591 /* Put it after P and before NEXT. */
1592 elt->next_same_value = next;
1593 if (next)
1594 next->prev_same_value = elt;
1596 elt->prev_same_value = p;
1597 p->next_same_value = elt;
1598 elt->first_same_value = classp;
1601 else
1602 elt->first_same_value = elt;
1604 /* If this is a constant being set equivalent to a register or a register
1605 being set equivalent to a constant, note the constant equivalence.
1607 If this is a constant, it cannot be equivalent to a different constant,
1608 and a constant is the only thing that can be cheaper than a register. So
1609 we know the register is the head of the class (before the constant was
1610 inserted).
1612 If this is a register that is not already known equivalent to a
1613 constant, we must check the entire class.
1615 If this is a register that is already known equivalent to an insn,
1616 update the qtys `const_insn' to show that `this_insn' is the latest
1617 insn making that quantity equivalent to the constant. */
1619 if (elt->is_const && classp && REG_P (classp->exp)
1620 && !REG_P (x))
1622 int exp_q = REG_QTY (REGNO (classp->exp));
1623 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1625 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1626 exp_ent->const_insn = this_insn;
1629 else if (REG_P (x)
1630 && classp
1631 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1632 && ! elt->is_const)
1634 struct table_elt *p;
1636 for (p = classp; p != 0; p = p->next_same_value)
1638 if (p->is_const && !REG_P (p->exp))
1640 int x_q = REG_QTY (REGNO (x));
1641 struct qty_table_elem *x_ent = &qty_table[x_q];
1643 x_ent->const_rtx
1644 = gen_lowpart (GET_MODE (x), p->exp);
1645 x_ent->const_insn = this_insn;
1646 break;
1651 else if (REG_P (x)
1652 && qty_table[REG_QTY (REGNO (x))].const_rtx
1653 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1654 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1656 /* If this is a constant with symbolic value,
1657 and it has a term with an explicit integer value,
1658 link it up with related expressions. */
1659 if (GET_CODE (x) == CONST)
1661 rtx subexp = get_related_value (x);
1662 unsigned subhash;
1663 struct table_elt *subelt, *subelt_prev;
1665 if (subexp != 0)
1667 /* Get the integer-free subexpression in the hash table. */
1668 subhash = SAFE_HASH (subexp, mode);
1669 subelt = lookup (subexp, subhash, mode);
1670 if (subelt == 0)
1671 subelt = insert (subexp, NULL, subhash, mode);
1672 /* Initialize SUBELT's circular chain if it has none. */
1673 if (subelt->related_value == 0)
1674 subelt->related_value = subelt;
1675 /* Find the element in the circular chain that precedes SUBELT. */
1676 subelt_prev = subelt;
1677 while (subelt_prev->related_value != subelt)
1678 subelt_prev = subelt_prev->related_value;
1679 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1680 This way the element that follows SUBELT is the oldest one. */
1681 elt->related_value = subelt_prev->related_value;
1682 subelt_prev->related_value = elt;
1686 return elt;
1689 /* Wrap insert_with_costs by passing the default costs. */
1691 static struct table_elt *
1692 insert (rtx x, struct table_elt *classp, unsigned int hash,
1693 machine_mode mode)
1695 return insert_with_costs (x, classp, hash, mode,
1696 COST (x, mode), approx_reg_cost (x));
1700 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1701 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1702 the two classes equivalent.
1704 CLASS1 will be the surviving class; CLASS2 should not be used after this
1705 call.
1707 Any invalid entries in CLASS2 will not be copied. */
1709 static void
1710 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1712 struct table_elt *elt, *next, *new_elt;
1714 /* Ensure we start with the head of the classes. */
1715 class1 = class1->first_same_value;
1716 class2 = class2->first_same_value;
1718 /* If they were already equal, forget it. */
1719 if (class1 == class2)
1720 return;
1722 for (elt = class2; elt; elt = next)
1724 unsigned int hash;
1725 rtx exp = elt->exp;
1726 machine_mode mode = elt->mode;
1728 next = elt->next_same_value;
1730 /* Remove old entry, make a new one in CLASS1's class.
1731 Don't do this for invalid entries as we cannot find their
1732 hash code (it also isn't necessary). */
1733 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1735 bool need_rehash = false;
1737 hash_arg_in_memory = 0;
1738 hash = HASH (exp, mode);
1740 if (REG_P (exp))
1742 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1743 delete_reg_equiv (REGNO (exp));
1746 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1747 remove_pseudo_from_table (exp, hash);
1748 else
1749 remove_from_table (elt, hash);
1751 if (insert_regs (exp, class1, 0) || need_rehash)
1753 rehash_using_reg (exp);
1754 hash = HASH (exp, mode);
1756 new_elt = insert (exp, class1, hash, mode);
1757 new_elt->in_memory = hash_arg_in_memory;
1758 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1759 new_elt->cost = MAX_COST;
1764 /* Flush the entire hash table. */
1766 static void
1767 flush_hash_table (void)
1769 int i;
1770 struct table_elt *p;
1772 for (i = 0; i < HASH_SIZE; i++)
1773 for (p = table[i]; p; p = table[i])
1775 /* Note that invalidate can remove elements
1776 after P in the current hash chain. */
1777 if (REG_P (p->exp))
1778 invalidate (p->exp, VOIDmode);
1779 else
1780 remove_from_table (p, i);
1784 /* Check whether an anti dependence exists between X and EXP. MODE and
1785 ADDR are as for canon_anti_dependence. */
1787 static bool
1788 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1790 subrtx_iterator::array_type array;
1791 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1793 const_rtx x = *iter;
1794 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1795 return true;
1797 return false;
1800 /* Remove from the hash table, or mark as invalid, all expressions whose
1801 values could be altered by storing in register X. */
1803 static void
1804 invalidate_reg (rtx x)
1806 gcc_assert (GET_CODE (x) == REG);
1808 /* If X is a register, dependencies on its contents are recorded
1809 through the qty number mechanism. Just change the qty number of
1810 the register, mark it as invalid for expressions that refer to it,
1811 and remove it itself. */
1812 unsigned int regno = REGNO (x);
1813 unsigned int hash = HASH (x, GET_MODE (x));
1815 /* Remove REGNO from any quantity list it might be on and indicate
1816 that its value might have changed. If it is a pseudo, remove its
1817 entry from the hash table.
1819 For a hard register, we do the first two actions above for any
1820 additional hard registers corresponding to X. Then, if any of these
1821 registers are in the table, we must remove any REG entries that
1822 overlap these registers. */
1824 delete_reg_equiv (regno);
1825 REG_TICK (regno)++;
1826 SUBREG_TICKED (regno) = -1;
1828 if (regno >= FIRST_PSEUDO_REGISTER)
1829 remove_pseudo_from_table (x, hash);
1830 else
1832 HOST_WIDE_INT in_table = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1833 unsigned int endregno = END_REGNO (x);
1834 unsigned int rn;
1835 struct table_elt *p, *next;
1837 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1839 for (rn = regno + 1; rn < endregno; rn++)
1841 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1842 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1843 delete_reg_equiv (rn);
1844 REG_TICK (rn)++;
1845 SUBREG_TICKED (rn) = -1;
1848 if (in_table)
1849 for (hash = 0; hash < HASH_SIZE; hash++)
1850 for (p = table[hash]; p; p = next)
1852 next = p->next_same_hash;
1854 if (!REG_P (p->exp) || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1855 continue;
1857 unsigned int tregno = REGNO (p->exp);
1858 unsigned int tendregno = END_REGNO (p->exp);
1859 if (tendregno > regno && tregno < endregno)
1860 remove_from_table (p, hash);
1865 /* Remove from the hash table, or mark as invalid, all expressions whose
1866 values could be altered by storing in X. X is a register, a subreg, or
1867 a memory reference with nonvarying address (because, when a memory
1868 reference with a varying address is stored in, all memory references are
1869 removed by invalidate_memory so specific invalidation is superfluous).
1870 FULL_MODE, if not VOIDmode, indicates that this much should be
1871 invalidated instead of just the amount indicated by the mode of X. This
1872 is only used for bitfield stores into memory.
1874 A nonvarying address may be just a register or just a symbol reference,
1875 or it may be either of those plus a numeric offset. */
1877 static void
1878 invalidate (rtx x, machine_mode full_mode)
1880 int i;
1881 struct table_elt *p;
1882 rtx addr;
1884 switch (GET_CODE (x))
1886 case REG:
1887 invalidate_reg (x);
1888 return;
1890 case SUBREG:
1891 invalidate (SUBREG_REG (x), VOIDmode);
1892 return;
1894 case PARALLEL:
1895 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1896 invalidate (XVECEXP (x, 0, i), VOIDmode);
1897 return;
1899 case EXPR_LIST:
1900 /* This is part of a disjoint return value; extract the location in
1901 question ignoring the offset. */
1902 invalidate (XEXP (x, 0), VOIDmode);
1903 return;
1905 case MEM:
1906 addr = canon_rtx (get_addr (XEXP (x, 0)));
1907 /* Calculate the canonical version of X here so that
1908 true_dependence doesn't generate new RTL for X on each call. */
1909 x = canon_rtx (x);
1911 /* Remove all hash table elements that refer to overlapping pieces of
1912 memory. */
1913 if (full_mode == VOIDmode)
1914 full_mode = GET_MODE (x);
1916 for (i = 0; i < HASH_SIZE; i++)
1918 struct table_elt *next;
1920 for (p = table[i]; p; p = next)
1922 next = p->next_same_hash;
1923 if (p->in_memory)
1925 /* Just canonicalize the expression once;
1926 otherwise each time we call invalidate
1927 true_dependence will canonicalize the
1928 expression again. */
1929 if (!p->canon_exp)
1930 p->canon_exp = canon_rtx (p->exp);
1931 if (check_dependence (p->canon_exp, x, full_mode, addr))
1932 remove_from_table (p, i);
1936 return;
1938 default:
1939 gcc_unreachable ();
1943 /* Invalidate DEST. Used when DEST is not going to be added
1944 into the hash table for some reason, e.g. do_not_record
1945 flagged on it. */
1947 static void
1948 invalidate_dest (rtx dest)
1950 if (REG_P (dest)
1951 || GET_CODE (dest) == SUBREG
1952 || MEM_P (dest))
1953 invalidate (dest, VOIDmode);
1954 else if (GET_CODE (dest) == STRICT_LOW_PART
1955 || GET_CODE (dest) == ZERO_EXTRACT)
1956 invalidate (XEXP (dest, 0), GET_MODE (dest));
1959 /* Remove all expressions that refer to register REGNO,
1960 since they are already invalid, and we are about to
1961 mark that register valid again and don't want the old
1962 expressions to reappear as valid. */
1964 static void
1965 remove_invalid_refs (unsigned int regno)
1967 unsigned int i;
1968 struct table_elt *p, *next;
1970 for (i = 0; i < HASH_SIZE; i++)
1971 for (p = table[i]; p; p = next)
1973 next = p->next_same_hash;
1974 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1975 remove_from_table (p, i);
1979 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1980 and mode MODE. */
1981 static void
1982 remove_invalid_subreg_refs (unsigned int regno, poly_uint64 offset,
1983 machine_mode mode)
1985 unsigned int i;
1986 struct table_elt *p, *next;
1988 for (i = 0; i < HASH_SIZE; i++)
1989 for (p = table[i]; p; p = next)
1991 rtx exp = p->exp;
1992 next = p->next_same_hash;
1994 if (!REG_P (exp)
1995 && (GET_CODE (exp) != SUBREG
1996 || !REG_P (SUBREG_REG (exp))
1997 || REGNO (SUBREG_REG (exp)) != regno
1998 || ranges_maybe_overlap_p (SUBREG_BYTE (exp),
1999 GET_MODE_SIZE (GET_MODE (exp)),
2000 offset, GET_MODE_SIZE (mode)))
2001 && refers_to_regno_p (regno, p->exp))
2002 remove_from_table (p, i);
2006 /* Recompute the hash codes of any valid entries in the hash table that
2007 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2009 This is called when we make a jump equivalence. */
2011 static void
2012 rehash_using_reg (rtx x)
2014 unsigned int i;
2015 struct table_elt *p, *next;
2016 unsigned hash;
2018 if (GET_CODE (x) == SUBREG)
2019 x = SUBREG_REG (x);
2021 /* If X is not a register or if the register is known not to be in any
2022 valid entries in the table, we have no work to do. */
2024 if (!REG_P (x)
2025 || REG_IN_TABLE (REGNO (x)) < 0
2026 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2027 return;
2029 /* Scan all hash chains looking for valid entries that mention X.
2030 If we find one and it is in the wrong hash chain, move it. */
2032 for (i = 0; i < HASH_SIZE; i++)
2033 for (p = table[i]; p; p = next)
2035 next = p->next_same_hash;
2036 if (reg_mentioned_p (x, p->exp)
2037 && exp_equiv_p (p->exp, p->exp, 1, false)
2038 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2040 if (p->next_same_hash)
2041 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2043 if (p->prev_same_hash)
2044 p->prev_same_hash->next_same_hash = p->next_same_hash;
2045 else
2046 table[i] = p->next_same_hash;
2048 p->next_same_hash = table[hash];
2049 p->prev_same_hash = 0;
2050 if (table[hash])
2051 table[hash]->prev_same_hash = p;
2052 table[hash] = p;
2057 /* Remove from the hash table any expression that is a call-clobbered
2058 register in INSN. Also update their TICK values. */
2060 static void
2061 invalidate_for_call (rtx_insn *insn)
2063 unsigned int regno;
2064 unsigned hash;
2065 struct table_elt *p, *next;
2066 int in_table = 0;
2067 hard_reg_set_iterator hrsi;
2069 /* Go through all the hard registers. For each that might be clobbered
2070 in call insn INSN, remove the register from quantity chains and update
2071 reg_tick if defined. Also see if any of these registers is currently
2072 in the table.
2074 ??? We could be more precise for partially-clobbered registers,
2075 and only invalidate values that actually occupy the clobbered part
2076 of the registers. It doesn't seem worth the effort though, since
2077 we shouldn't see this situation much before RA. Whatever choice
2078 we make here has to be consistent with the table walk below,
2079 so any change to this test will require a change there too. */
2080 HARD_REG_SET callee_clobbers
2081 = insn_callee_abi (insn).full_and_partial_reg_clobbers ();
2082 EXECUTE_IF_SET_IN_HARD_REG_SET (callee_clobbers, 0, regno, hrsi)
2084 delete_reg_equiv (regno);
2085 if (REG_TICK (regno) >= 0)
2087 REG_TICK (regno)++;
2088 SUBREG_TICKED (regno) = -1;
2090 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2093 /* In the case where we have no call-clobbered hard registers in the
2094 table, we are done. Otherwise, scan the table and remove any
2095 entry that overlaps a call-clobbered register. */
2097 if (in_table)
2098 for (hash = 0; hash < HASH_SIZE; hash++)
2099 for (p = table[hash]; p; p = next)
2101 next = p->next_same_hash;
2103 if (!REG_P (p->exp)
2104 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2105 continue;
2107 /* This must use the same test as above rather than the
2108 more accurate clobbers_reg_p. */
2109 if (overlaps_hard_reg_set_p (callee_clobbers, GET_MODE (p->exp),
2110 REGNO (p->exp)))
2111 remove_from_table (p, hash);
2115 /* Given an expression X of type CONST,
2116 and ELT which is its table entry (or 0 if it
2117 is not in the hash table),
2118 return an alternate expression for X as a register plus integer.
2119 If none can be found, return 0. */
2121 static rtx
2122 use_related_value (rtx x, struct table_elt *elt)
2124 struct table_elt *relt = 0;
2125 struct table_elt *p, *q;
2126 HOST_WIDE_INT offset;
2128 /* First, is there anything related known?
2129 If we have a table element, we can tell from that.
2130 Otherwise, must look it up. */
2132 if (elt != 0 && elt->related_value != 0)
2133 relt = elt;
2134 else if (elt == 0 && GET_CODE (x) == CONST)
2136 rtx subexp = get_related_value (x);
2137 if (subexp != 0)
2138 relt = lookup (subexp,
2139 SAFE_HASH (subexp, GET_MODE (subexp)),
2140 GET_MODE (subexp));
2143 if (relt == 0)
2144 return 0;
2146 /* Search all related table entries for one that has an
2147 equivalent register. */
2149 p = relt;
2150 while (1)
2152 /* This loop is strange in that it is executed in two different cases.
2153 The first is when X is already in the table. Then it is searching
2154 the RELATED_VALUE list of X's class (RELT). The second case is when
2155 X is not in the table. Then RELT points to a class for the related
2156 value.
2158 Ensure that, whatever case we are in, that we ignore classes that have
2159 the same value as X. */
2161 if (rtx_equal_p (x, p->exp))
2162 q = 0;
2163 else
2164 for (q = p->first_same_value; q; q = q->next_same_value)
2165 if (REG_P (q->exp))
2166 break;
2168 if (q)
2169 break;
2171 p = p->related_value;
2173 /* We went all the way around, so there is nothing to be found.
2174 Alternatively, perhaps RELT was in the table for some other reason
2175 and it has no related values recorded. */
2176 if (p == relt || p == 0)
2177 break;
2180 if (q == 0)
2181 return 0;
2183 offset = (get_integer_term (x) - get_integer_term (p->exp));
2184 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2185 return plus_constant (q->mode, q->exp, offset);
2189 /* Hash a string. Just add its bytes up. */
2190 static inline unsigned
2191 hash_rtx_string (const char *ps)
2193 unsigned hash = 0;
2194 const unsigned char *p = (const unsigned char *) ps;
2196 if (p)
2197 while (*p)
2198 hash += *p++;
2200 return hash;
2203 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2204 When the callback returns true, we continue with the new rtx. */
2206 unsigned
2207 hash_rtx_cb (const_rtx x, machine_mode mode,
2208 int *do_not_record_p, int *hash_arg_in_memory_p,
2209 bool have_reg_qty, hash_rtx_callback_function cb)
2211 int i, j;
2212 unsigned hash = 0;
2213 enum rtx_code code;
2214 const char *fmt;
2215 machine_mode newmode;
2216 rtx newx;
2218 /* Used to turn recursion into iteration. We can't rely on GCC's
2219 tail-recursion elimination since we need to keep accumulating values
2220 in HASH. */
2221 repeat:
2222 if (x == 0)
2223 return hash;
2225 /* Invoke the callback first. */
2226 if (cb != NULL
2227 && ((*cb) (x, mode, &newx, &newmode)))
2229 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2230 hash_arg_in_memory_p, have_reg_qty, cb);
2231 return hash;
2234 code = GET_CODE (x);
2235 switch (code)
2237 case REG:
2239 unsigned int regno = REGNO (x);
2241 if (do_not_record_p && !reload_completed)
2243 /* On some machines, we can't record any non-fixed hard register,
2244 because extending its life will cause reload problems. We
2245 consider ap, fp, sp, gp to be fixed for this purpose.
2247 We also consider CCmode registers to be fixed for this purpose;
2248 failure to do so leads to failure to simplify 0<100 type of
2249 conditionals.
2251 On all machines, we can't record any global registers.
2252 Nor should we record any register that is in a small
2253 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2254 bool record;
2256 if (regno >= FIRST_PSEUDO_REGISTER)
2257 record = true;
2258 else if (x == frame_pointer_rtx
2259 || x == hard_frame_pointer_rtx
2260 || x == arg_pointer_rtx
2261 || x == stack_pointer_rtx
2262 || x == pic_offset_table_rtx)
2263 record = true;
2264 else if (global_regs[regno])
2265 record = false;
2266 else if (fixed_regs[regno])
2267 record = true;
2268 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2269 record = true;
2270 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2271 record = false;
2272 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2273 record = false;
2274 else
2275 record = true;
2277 if (!record)
2279 *do_not_record_p = 1;
2280 return 0;
2284 hash += ((unsigned int) REG << 7);
2285 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2286 return hash;
2289 /* We handle SUBREG of a REG specially because the underlying
2290 reg changes its hash value with every value change; we don't
2291 want to have to forget unrelated subregs when one subreg changes. */
2292 case SUBREG:
2294 if (REG_P (SUBREG_REG (x)))
2296 hash += (((unsigned int) SUBREG << 7)
2297 + REGNO (SUBREG_REG (x))
2298 + (constant_lower_bound (SUBREG_BYTE (x))
2299 / UNITS_PER_WORD));
2300 return hash;
2302 break;
2305 case CONST_INT:
2306 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2307 + (unsigned int) INTVAL (x));
2308 return hash;
2310 case CONST_WIDE_INT:
2311 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2312 hash += CONST_WIDE_INT_ELT (x, i);
2313 return hash;
2315 case CONST_POLY_INT:
2317 inchash::hash h;
2318 h.add_int (hash);
2319 for (unsigned int i = 0; i < NUM_POLY_INT_COEFFS; ++i)
2320 h.add_wide_int (CONST_POLY_INT_COEFFS (x)[i]);
2321 return h.end ();
2324 case CONST_DOUBLE:
2325 /* This is like the general case, except that it only counts
2326 the integers representing the constant. */
2327 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2328 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2329 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2330 + (unsigned int) CONST_DOUBLE_HIGH (x));
2331 else
2332 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2333 return hash;
2335 case CONST_FIXED:
2336 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2337 hash += fixed_hash (CONST_FIXED_VALUE (x));
2338 return hash;
2340 case CONST_VECTOR:
2342 int units;
2343 rtx elt;
2345 units = const_vector_encoded_nelts (x);
2347 for (i = 0; i < units; ++i)
2349 elt = CONST_VECTOR_ENCODED_ELT (x, i);
2350 hash += hash_rtx_cb (elt, GET_MODE (elt),
2351 do_not_record_p, hash_arg_in_memory_p,
2352 have_reg_qty, cb);
2355 return hash;
2358 /* Assume there is only one rtx object for any given label. */
2359 case LABEL_REF:
2360 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2361 differences and differences between each stage's debugging dumps. */
2362 hash += (((unsigned int) LABEL_REF << 7)
2363 + CODE_LABEL_NUMBER (label_ref_label (x)));
2364 return hash;
2366 case SYMBOL_REF:
2368 /* Don't hash on the symbol's address to avoid bootstrap differences.
2369 Different hash values may cause expressions to be recorded in
2370 different orders and thus different registers to be used in the
2371 final assembler. This also avoids differences in the dump files
2372 between various stages. */
2373 unsigned int h = 0;
2374 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2376 while (*p)
2377 h += (h << 7) + *p++; /* ??? revisit */
2379 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2380 return hash;
2383 case MEM:
2384 /* We don't record if marked volatile or if BLKmode since we don't
2385 know the size of the move. */
2386 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2388 *do_not_record_p = 1;
2389 return 0;
2391 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2392 *hash_arg_in_memory_p = 1;
2394 /* Now that we have already found this special case,
2395 might as well speed it up as much as possible. */
2396 hash += (unsigned) MEM;
2397 x = XEXP (x, 0);
2398 goto repeat;
2400 case USE:
2401 /* A USE that mentions non-volatile memory needs special
2402 handling since the MEM may be BLKmode which normally
2403 prevents an entry from being made. Pure calls are
2404 marked by a USE which mentions BLKmode memory.
2405 See calls.c:emit_call_1. */
2406 if (MEM_P (XEXP (x, 0))
2407 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2409 hash += (unsigned) USE;
2410 x = XEXP (x, 0);
2412 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2413 *hash_arg_in_memory_p = 1;
2415 /* Now that we have already found this special case,
2416 might as well speed it up as much as possible. */
2417 hash += (unsigned) MEM;
2418 x = XEXP (x, 0);
2419 goto repeat;
2421 break;
2423 case PRE_DEC:
2424 case PRE_INC:
2425 case POST_DEC:
2426 case POST_INC:
2427 case PRE_MODIFY:
2428 case POST_MODIFY:
2429 case PC:
2430 case CALL:
2431 case UNSPEC_VOLATILE:
2432 if (do_not_record_p) {
2433 *do_not_record_p = 1;
2434 return 0;
2436 else
2437 return hash;
2438 break;
2440 case ASM_OPERANDS:
2441 if (do_not_record_p && MEM_VOLATILE_P (x))
2443 *do_not_record_p = 1;
2444 return 0;
2446 else
2448 /* We don't want to take the filename and line into account. */
2449 hash += (unsigned) code + (unsigned) GET_MODE (x)
2450 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2451 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2452 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2454 if (ASM_OPERANDS_INPUT_LENGTH (x))
2456 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2458 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2459 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2460 do_not_record_p, hash_arg_in_memory_p,
2461 have_reg_qty, cb)
2462 + hash_rtx_string
2463 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2466 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2467 x = ASM_OPERANDS_INPUT (x, 0);
2468 mode = GET_MODE (x);
2469 goto repeat;
2472 return hash;
2474 break;
2476 default:
2477 break;
2480 i = GET_RTX_LENGTH (code) - 1;
2481 hash += (unsigned) code + (unsigned) GET_MODE (x);
2482 fmt = GET_RTX_FORMAT (code);
2483 for (; i >= 0; i--)
2485 switch (fmt[i])
2487 case 'e':
2488 /* If we are about to do the last recursive call
2489 needed at this level, change it into iteration.
2490 This function is called enough to be worth it. */
2491 if (i == 0)
2493 x = XEXP (x, i);
2494 goto repeat;
2497 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2498 hash_arg_in_memory_p,
2499 have_reg_qty, cb);
2500 break;
2502 case 'E':
2503 for (j = 0; j < XVECLEN (x, i); j++)
2504 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2505 hash_arg_in_memory_p,
2506 have_reg_qty, cb);
2507 break;
2509 case 's':
2510 hash += hash_rtx_string (XSTR (x, i));
2511 break;
2513 case 'i':
2514 hash += (unsigned int) XINT (x, i);
2515 break;
2517 case 'p':
2518 hash += constant_lower_bound (SUBREG_BYTE (x));
2519 break;
2521 case '0': case 't':
2522 /* Unused. */
2523 break;
2525 default:
2526 gcc_unreachable ();
2530 return hash;
2533 /* Hash an rtx. We are careful to make sure the value is never negative.
2534 Equivalent registers hash identically.
2535 MODE is used in hashing for CONST_INTs only;
2536 otherwise the mode of X is used.
2538 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2540 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2541 a MEM rtx which does not have the MEM_READONLY_P flag set.
2543 Note that cse_insn knows that the hash code of a MEM expression
2544 is just (int) MEM plus the hash code of the address. */
2546 unsigned
2547 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2548 int *hash_arg_in_memory_p, bool have_reg_qty)
2550 return hash_rtx_cb (x, mode, do_not_record_p,
2551 hash_arg_in_memory_p, have_reg_qty, NULL);
2554 /* Hash an rtx X for cse via hash_rtx.
2555 Stores 1 in do_not_record if any subexpression is volatile.
2556 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2557 does not have the MEM_READONLY_P flag set. */
2559 static inline unsigned
2560 canon_hash (rtx x, machine_mode mode)
2562 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2565 /* Like canon_hash but with no side effects, i.e. do_not_record
2566 and hash_arg_in_memory are not changed. */
2568 static inline unsigned
2569 safe_hash (rtx x, machine_mode mode)
2571 int dummy_do_not_record;
2572 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2575 /* Return 1 iff X and Y would canonicalize into the same thing,
2576 without actually constructing the canonicalization of either one.
2577 If VALIDATE is nonzero,
2578 we assume X is an expression being processed from the rtl
2579 and Y was found in the hash table. We check register refs
2580 in Y for being marked as valid.
2582 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2585 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2587 int i, j;
2588 enum rtx_code code;
2589 const char *fmt;
2591 /* Note: it is incorrect to assume an expression is equivalent to itself
2592 if VALIDATE is nonzero. */
2593 if (x == y && !validate)
2594 return 1;
2596 if (x == 0 || y == 0)
2597 return x == y;
2599 code = GET_CODE (x);
2600 if (code != GET_CODE (y))
2601 return 0;
2603 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2604 if (GET_MODE (x) != GET_MODE (y))
2605 return 0;
2607 /* MEMs referring to different address space are not equivalent. */
2608 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2609 return 0;
2611 switch (code)
2613 case PC:
2614 CASE_CONST_UNIQUE:
2615 return x == y;
2617 case CONST_VECTOR:
2618 if (!same_vector_encodings_p (x, y))
2619 return false;
2620 break;
2622 case LABEL_REF:
2623 return label_ref_label (x) == label_ref_label (y);
2625 case SYMBOL_REF:
2626 return XSTR (x, 0) == XSTR (y, 0);
2628 case REG:
2629 if (for_gcse)
2630 return REGNO (x) == REGNO (y);
2631 else
2633 unsigned int regno = REGNO (y);
2634 unsigned int i;
2635 unsigned int endregno = END_REGNO (y);
2637 /* If the quantities are not the same, the expressions are not
2638 equivalent. If there are and we are not to validate, they
2639 are equivalent. Otherwise, ensure all regs are up-to-date. */
2641 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2642 return 0;
2644 if (! validate)
2645 return 1;
2647 for (i = regno; i < endregno; i++)
2648 if (REG_IN_TABLE (i) != REG_TICK (i))
2649 return 0;
2651 return 1;
2654 case MEM:
2655 if (for_gcse)
2657 /* A volatile mem should not be considered equivalent to any
2658 other. */
2659 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2660 return 0;
2662 /* Can't merge two expressions in different alias sets, since we
2663 can decide that the expression is transparent in a block when
2664 it isn't, due to it being set with the different alias set.
2666 Also, can't merge two expressions with different MEM_ATTRS.
2667 They could e.g. be two different entities allocated into the
2668 same space on the stack (see e.g. PR25130). In that case, the
2669 MEM addresses can be the same, even though the two MEMs are
2670 absolutely not equivalent.
2672 But because really all MEM attributes should be the same for
2673 equivalent MEMs, we just use the invariant that MEMs that have
2674 the same attributes share the same mem_attrs data structure. */
2675 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2676 return 0;
2678 /* If we are handling exceptions, we cannot consider two expressions
2679 with different trapping status as equivalent, because simple_mem
2680 might accept one and reject the other. */
2681 if (cfun->can_throw_non_call_exceptions
2682 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2683 return 0;
2685 break;
2687 /* For commutative operations, check both orders. */
2688 case PLUS:
2689 case MULT:
2690 case AND:
2691 case IOR:
2692 case XOR:
2693 case NE:
2694 case EQ:
2695 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2696 validate, for_gcse)
2697 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2698 validate, for_gcse))
2699 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2700 validate, for_gcse)
2701 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2702 validate, for_gcse)));
2704 case ASM_OPERANDS:
2705 /* We don't use the generic code below because we want to
2706 disregard filename and line numbers. */
2708 /* A volatile asm isn't equivalent to any other. */
2709 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2710 return 0;
2712 if (GET_MODE (x) != GET_MODE (y)
2713 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2714 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2715 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2716 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2717 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2718 return 0;
2720 if (ASM_OPERANDS_INPUT_LENGTH (x))
2722 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2723 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2724 ASM_OPERANDS_INPUT (y, i),
2725 validate, for_gcse)
2726 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2727 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2728 return 0;
2731 return 1;
2733 default:
2734 break;
2737 /* Compare the elements. If any pair of corresponding elements
2738 fail to match, return 0 for the whole thing. */
2740 fmt = GET_RTX_FORMAT (code);
2741 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2743 switch (fmt[i])
2745 case 'e':
2746 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2747 validate, for_gcse))
2748 return 0;
2749 break;
2751 case 'E':
2752 if (XVECLEN (x, i) != XVECLEN (y, i))
2753 return 0;
2754 for (j = 0; j < XVECLEN (x, i); j++)
2755 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2756 validate, for_gcse))
2757 return 0;
2758 break;
2760 case 's':
2761 if (strcmp (XSTR (x, i), XSTR (y, i)))
2762 return 0;
2763 break;
2765 case 'i':
2766 if (XINT (x, i) != XINT (y, i))
2767 return 0;
2768 break;
2770 case 'w':
2771 if (XWINT (x, i) != XWINT (y, i))
2772 return 0;
2773 break;
2775 case 'p':
2776 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
2777 return 0;
2778 break;
2780 case '0':
2781 case 't':
2782 break;
2784 default:
2785 gcc_unreachable ();
2789 return 1;
2792 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2793 the result if necessary. INSN is as for canon_reg. */
2795 static void
2796 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2798 if (*xloc)
2800 rtx new_rtx = canon_reg (*xloc, insn);
2802 /* If replacing pseudo with hard reg or vice versa, ensure the
2803 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2804 gcc_assert (insn && new_rtx);
2805 validate_change (insn, xloc, new_rtx, 1);
2809 /* Canonicalize an expression:
2810 replace each register reference inside it
2811 with the "oldest" equivalent register.
2813 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2814 after we make our substitution. The calls are made with IN_GROUP nonzero
2815 so apply_change_group must be called upon the outermost return from this
2816 function (unless INSN is zero). The result of apply_change_group can
2817 generally be discarded since the changes we are making are optional. */
2819 static rtx
2820 canon_reg (rtx x, rtx_insn *insn)
2822 int i;
2823 enum rtx_code code;
2824 const char *fmt;
2826 if (x == 0)
2827 return x;
2829 code = GET_CODE (x);
2830 switch (code)
2832 case PC:
2833 case CONST:
2834 CASE_CONST_ANY:
2835 case SYMBOL_REF:
2836 case LABEL_REF:
2837 case ADDR_VEC:
2838 case ADDR_DIFF_VEC:
2839 return x;
2841 case REG:
2843 int first;
2844 int q;
2845 struct qty_table_elem *ent;
2847 /* Never replace a hard reg, because hard regs can appear
2848 in more than one machine mode, and we must preserve the mode
2849 of each occurrence. Also, some hard regs appear in
2850 MEMs that are shared and mustn't be altered. Don't try to
2851 replace any reg that maps to a reg of class NO_REGS. */
2852 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2853 || ! REGNO_QTY_VALID_P (REGNO (x)))
2854 return x;
2856 q = REG_QTY (REGNO (x));
2857 ent = &qty_table[q];
2858 first = ent->first_reg;
2859 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2860 : REGNO_REG_CLASS (first) == NO_REGS ? x
2861 : gen_rtx_REG (ent->mode, first));
2864 default:
2865 break;
2868 fmt = GET_RTX_FORMAT (code);
2869 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2871 int j;
2873 if (fmt[i] == 'e')
2874 validate_canon_reg (&XEXP (x, i), insn);
2875 else if (fmt[i] == 'E')
2876 for (j = 0; j < XVECLEN (x, i); j++)
2877 validate_canon_reg (&XVECEXP (x, i, j), insn);
2880 return x;
2883 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2884 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2885 what values are being compared.
2887 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2888 actually being compared. For example, if *PARG1 was (reg:CC CC_REG) and
2889 *PARG2 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that
2890 were compared to produce (reg:CC CC_REG).
2892 The return value is the comparison operator and is either the code of
2893 A or the code corresponding to the inverse of the comparison. */
2895 static enum rtx_code
2896 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2897 machine_mode *pmode1, machine_mode *pmode2)
2899 rtx arg1, arg2;
2900 hash_set<rtx> *visited = NULL;
2901 /* Set nonzero when we find something of interest. */
2902 rtx x = NULL;
2904 arg1 = *parg1, arg2 = *parg2;
2906 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2908 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2910 int reverse_code = 0;
2911 struct table_elt *p = 0;
2913 /* Remember state from previous iteration. */
2914 if (x)
2916 if (!visited)
2917 visited = new hash_set<rtx>;
2918 visited->add (x);
2919 x = 0;
2922 /* If arg1 is a COMPARE, extract the comparison arguments from it. */
2924 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2925 x = arg1;
2927 /* If ARG1 is a comparison operator and CODE is testing for
2928 STORE_FLAG_VALUE, get the inner arguments. */
2930 else if (COMPARISON_P (arg1))
2932 #ifdef FLOAT_STORE_FLAG_VALUE
2933 REAL_VALUE_TYPE fsfv;
2934 #endif
2936 if (code == NE
2937 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2938 && code == LT && STORE_FLAG_VALUE == -1)
2939 #ifdef FLOAT_STORE_FLAG_VALUE
2940 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2941 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2942 REAL_VALUE_NEGATIVE (fsfv)))
2943 #endif
2945 x = arg1;
2946 else if (code == EQ
2947 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2948 && code == GE && STORE_FLAG_VALUE == -1)
2949 #ifdef FLOAT_STORE_FLAG_VALUE
2950 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2951 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2952 REAL_VALUE_NEGATIVE (fsfv)))
2953 #endif
2955 x = arg1, reverse_code = 1;
2958 /* ??? We could also check for
2960 (ne (and (eq (...) (const_int 1))) (const_int 0))
2962 and related forms, but let's wait until we see them occurring. */
2964 if (x == 0)
2965 /* Look up ARG1 in the hash table and see if it has an equivalence
2966 that lets us see what is being compared. */
2967 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2968 if (p)
2970 p = p->first_same_value;
2972 /* If what we compare is already known to be constant, that is as
2973 good as it gets.
2974 We need to break the loop in this case, because otherwise we
2975 can have an infinite loop when looking at a reg that is known
2976 to be a constant which is the same as a comparison of a reg
2977 against zero which appears later in the insn stream, which in
2978 turn is constant and the same as the comparison of the first reg
2979 against zero... */
2980 if (p->is_const)
2981 break;
2984 for (; p; p = p->next_same_value)
2986 machine_mode inner_mode = GET_MODE (p->exp);
2987 #ifdef FLOAT_STORE_FLAG_VALUE
2988 REAL_VALUE_TYPE fsfv;
2989 #endif
2991 /* If the entry isn't valid, skip it. */
2992 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2993 continue;
2995 /* If it's a comparison we've used before, skip it. */
2996 if (visited && visited->contains (p->exp))
2997 continue;
2999 if (GET_CODE (p->exp) == COMPARE
3000 /* Another possibility is that this machine has a compare insn
3001 that includes the comparison code. In that case, ARG1 would
3002 be equivalent to a comparison operation that would set ARG1 to
3003 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3004 ORIG_CODE is the actual comparison being done; if it is an EQ,
3005 we must reverse ORIG_CODE. On machine with a negative value
3006 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3007 || ((code == NE
3008 || (code == LT
3009 && val_signbit_known_set_p (inner_mode,
3010 STORE_FLAG_VALUE))
3011 #ifdef FLOAT_STORE_FLAG_VALUE
3012 || (code == LT
3013 && SCALAR_FLOAT_MODE_P (inner_mode)
3014 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3015 REAL_VALUE_NEGATIVE (fsfv)))
3016 #endif
3018 && COMPARISON_P (p->exp)))
3020 x = p->exp;
3021 break;
3023 else if ((code == EQ
3024 || (code == GE
3025 && val_signbit_known_set_p (inner_mode,
3026 STORE_FLAG_VALUE))
3027 #ifdef FLOAT_STORE_FLAG_VALUE
3028 || (code == GE
3029 && SCALAR_FLOAT_MODE_P (inner_mode)
3030 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3031 REAL_VALUE_NEGATIVE (fsfv)))
3032 #endif
3034 && COMPARISON_P (p->exp))
3036 reverse_code = 1;
3037 x = p->exp;
3038 break;
3041 /* If this non-trapping address, e.g. fp + constant, the
3042 equivalent is a better operand since it may let us predict
3043 the value of the comparison. */
3044 else if (!rtx_addr_can_trap_p (p->exp))
3046 arg1 = p->exp;
3047 continue;
3051 /* If we didn't find a useful equivalence for ARG1, we are done.
3052 Otherwise, set up for the next iteration. */
3053 if (x == 0)
3054 break;
3056 /* If we need to reverse the comparison, make sure that is
3057 possible -- we can't necessarily infer the value of GE from LT
3058 with floating-point operands. */
3059 if (reverse_code)
3061 enum rtx_code reversed = reversed_comparison_code (x, NULL);
3062 if (reversed == UNKNOWN)
3063 break;
3064 else
3065 code = reversed;
3067 else if (COMPARISON_P (x))
3068 code = GET_CODE (x);
3069 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3072 /* Return our results. Return the modes from before fold_rtx
3073 because fold_rtx might produce const_int, and then it's too late. */
3074 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3075 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3077 if (visited)
3078 delete visited;
3079 return code;
3082 /* If X is a nontrivial arithmetic operation on an argument for which
3083 a constant value can be determined, return the result of operating
3084 on that value, as a constant. Otherwise, return X, possibly with
3085 one or more operands changed to a forward-propagated constant.
3087 If X is a register whose contents are known, we do NOT return
3088 those contents here; equiv_constant is called to perform that task.
3089 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3091 INSN is the insn that we may be modifying. If it is 0, make a copy
3092 of X before modifying it. */
3094 static rtx
3095 fold_rtx (rtx x, rtx_insn *insn)
3097 enum rtx_code code;
3098 machine_mode mode;
3099 const char *fmt;
3100 int i;
3101 rtx new_rtx = 0;
3102 int changed = 0;
3103 poly_int64 xval;
3105 /* Operands of X. */
3106 /* Workaround -Wmaybe-uninitialized false positive during
3107 profiledbootstrap by initializing them. */
3108 rtx folded_arg0 = NULL_RTX;
3109 rtx folded_arg1 = NULL_RTX;
3111 /* Constant equivalents of first three operands of X;
3112 0 when no such equivalent is known. */
3113 rtx const_arg0;
3114 rtx const_arg1;
3115 rtx const_arg2;
3117 /* The mode of the first operand of X. We need this for sign and zero
3118 extends. */
3119 machine_mode mode_arg0;
3121 if (x == 0)
3122 return x;
3124 /* Try to perform some initial simplifications on X. */
3125 code = GET_CODE (x);
3126 switch (code)
3128 case MEM:
3129 case SUBREG:
3130 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3131 than it would in other contexts. Basically its mode does not
3132 signify the size of the object read. That information is carried
3133 by size operand. If we happen to have a MEM of the appropriate
3134 mode in our tables with a constant value we could simplify the
3135 extraction incorrectly if we allowed substitution of that value
3136 for the MEM. */
3137 case ZERO_EXTRACT:
3138 case SIGN_EXTRACT:
3139 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3140 return new_rtx;
3141 return x;
3143 case CONST:
3144 CASE_CONST_ANY:
3145 case SYMBOL_REF:
3146 case LABEL_REF:
3147 case REG:
3148 case PC:
3149 /* No use simplifying an EXPR_LIST
3150 since they are used only for lists of args
3151 in a function call's REG_EQUAL note. */
3152 case EXPR_LIST:
3153 return x;
3155 case ASM_OPERANDS:
3156 if (insn)
3158 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3159 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3160 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3162 return x;
3164 case CALL:
3165 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3166 return x;
3167 break;
3168 case VEC_SELECT:
3170 rtx trueop0 = XEXP (x, 0);
3171 mode = GET_MODE (trueop0);
3172 rtx trueop1 = XEXP (x, 1);
3173 /* If we select a low-part subreg, return that. */
3174 if (vec_series_lowpart_p (GET_MODE (x), mode, trueop1))
3176 rtx new_rtx = lowpart_subreg (GET_MODE (x), trueop0, mode);
3177 if (new_rtx != NULL_RTX)
3178 return new_rtx;
3182 /* Anything else goes through the loop below. */
3183 default:
3184 break;
3187 mode = GET_MODE (x);
3188 const_arg0 = 0;
3189 const_arg1 = 0;
3190 const_arg2 = 0;
3191 mode_arg0 = VOIDmode;
3193 /* Try folding our operands.
3194 Then see which ones have constant values known. */
3196 fmt = GET_RTX_FORMAT (code);
3197 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3198 if (fmt[i] == 'e')
3200 rtx folded_arg = XEXP (x, i), const_arg;
3201 machine_mode mode_arg = GET_MODE (folded_arg);
3203 switch (GET_CODE (folded_arg))
3205 case MEM:
3206 case REG:
3207 case SUBREG:
3208 const_arg = equiv_constant (folded_arg);
3209 break;
3211 case CONST:
3212 CASE_CONST_ANY:
3213 case SYMBOL_REF:
3214 case LABEL_REF:
3215 const_arg = folded_arg;
3216 break;
3218 default:
3219 folded_arg = fold_rtx (folded_arg, insn);
3220 const_arg = equiv_constant (folded_arg);
3221 break;
3224 /* For the first three operands, see if the operand
3225 is constant or equivalent to a constant. */
3226 switch (i)
3228 case 0:
3229 folded_arg0 = folded_arg;
3230 const_arg0 = const_arg;
3231 mode_arg0 = mode_arg;
3232 break;
3233 case 1:
3234 folded_arg1 = folded_arg;
3235 const_arg1 = const_arg;
3236 break;
3237 case 2:
3238 const_arg2 = const_arg;
3239 break;
3242 /* Pick the least expensive of the argument and an equivalent constant
3243 argument. */
3244 if (const_arg != 0
3245 && const_arg != folded_arg
3246 && (COST_IN (const_arg, mode_arg, code, i)
3247 <= COST_IN (folded_arg, mode_arg, code, i))
3249 /* It's not safe to substitute the operand of a conversion
3250 operator with a constant, as the conversion's identity
3251 depends upon the mode of its operand. This optimization
3252 is handled by the call to simplify_unary_operation. */
3253 && (GET_RTX_CLASS (code) != RTX_UNARY
3254 || GET_MODE (const_arg) == mode_arg0
3255 || (code != ZERO_EXTEND
3256 && code != SIGN_EXTEND
3257 && code != TRUNCATE
3258 && code != FLOAT_TRUNCATE
3259 && code != FLOAT_EXTEND
3260 && code != FLOAT
3261 && code != FIX
3262 && code != UNSIGNED_FLOAT
3263 && code != UNSIGNED_FIX)))
3264 folded_arg = const_arg;
3266 if (folded_arg == XEXP (x, i))
3267 continue;
3269 if (insn == NULL_RTX && !changed)
3270 x = copy_rtx (x);
3271 changed = 1;
3272 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3275 if (changed)
3277 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3278 consistent with the order in X. */
3279 if (canonicalize_change_group (insn, x))
3281 std::swap (const_arg0, const_arg1);
3282 std::swap (folded_arg0, folded_arg1);
3285 apply_change_group ();
3288 /* If X is an arithmetic operation, see if we can simplify it. */
3290 switch (GET_RTX_CLASS (code))
3292 case RTX_UNARY:
3294 /* We can't simplify extension ops unless we know the
3295 original mode. */
3296 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3297 && mode_arg0 == VOIDmode)
3298 break;
3300 new_rtx = simplify_unary_operation (code, mode,
3301 const_arg0 ? const_arg0 : folded_arg0,
3302 mode_arg0);
3304 break;
3306 case RTX_COMPARE:
3307 case RTX_COMM_COMPARE:
3308 /* See what items are actually being compared and set FOLDED_ARG[01]
3309 to those values and CODE to the actual comparison code. If any are
3310 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3311 do anything if both operands are already known to be constant. */
3313 /* ??? Vector mode comparisons are not supported yet. */
3314 if (VECTOR_MODE_P (mode))
3315 break;
3317 if (const_arg0 == 0 || const_arg1 == 0)
3319 struct table_elt *p0, *p1;
3320 rtx true_rtx, false_rtx;
3321 machine_mode mode_arg1;
3323 if (SCALAR_FLOAT_MODE_P (mode))
3325 #ifdef FLOAT_STORE_FLAG_VALUE
3326 true_rtx = (const_double_from_real_value
3327 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3328 #else
3329 true_rtx = NULL_RTX;
3330 #endif
3331 false_rtx = CONST0_RTX (mode);
3333 else
3335 true_rtx = const_true_rtx;
3336 false_rtx = const0_rtx;
3339 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3340 &mode_arg0, &mode_arg1);
3342 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3343 what kinds of things are being compared, so we can't do
3344 anything with this comparison. */
3346 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3347 break;
3349 const_arg0 = equiv_constant (folded_arg0);
3350 const_arg1 = equiv_constant (folded_arg1);
3352 /* If we do not now have two constants being compared, see
3353 if we can nevertheless deduce some things about the
3354 comparison. */
3355 if (const_arg0 == 0 || const_arg1 == 0)
3357 if (const_arg1 != NULL)
3359 rtx cheapest_simplification;
3360 int cheapest_cost;
3361 rtx simp_result;
3362 struct table_elt *p;
3364 /* See if we can find an equivalent of folded_arg0
3365 that gets us a cheaper expression, possibly a
3366 constant through simplifications. */
3367 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3368 mode_arg0);
3370 if (p != NULL)
3372 cheapest_simplification = x;
3373 cheapest_cost = COST (x, mode);
3375 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3377 int cost;
3379 /* If the entry isn't valid, skip it. */
3380 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3381 continue;
3383 /* Try to simplify using this equivalence. */
3384 simp_result
3385 = simplify_relational_operation (code, mode,
3386 mode_arg0,
3387 p->exp,
3388 const_arg1);
3390 if (simp_result == NULL)
3391 continue;
3393 cost = COST (simp_result, mode);
3394 if (cost < cheapest_cost)
3396 cheapest_cost = cost;
3397 cheapest_simplification = simp_result;
3401 /* If we have a cheaper expression now, use that
3402 and try folding it further, from the top. */
3403 if (cheapest_simplification != x)
3404 return fold_rtx (copy_rtx (cheapest_simplification),
3405 insn);
3409 /* See if the two operands are the same. */
3411 if ((REG_P (folded_arg0)
3412 && REG_P (folded_arg1)
3413 && (REG_QTY (REGNO (folded_arg0))
3414 == REG_QTY (REGNO (folded_arg1))))
3415 || ((p0 = lookup (folded_arg0,
3416 SAFE_HASH (folded_arg0, mode_arg0),
3417 mode_arg0))
3418 && (p1 = lookup (folded_arg1,
3419 SAFE_HASH (folded_arg1, mode_arg0),
3420 mode_arg0))
3421 && p0->first_same_value == p1->first_same_value))
3422 folded_arg1 = folded_arg0;
3424 /* If FOLDED_ARG0 is a register, see if the comparison we are
3425 doing now is either the same as we did before or the reverse
3426 (we only check the reverse if not floating-point). */
3427 else if (REG_P (folded_arg0))
3429 int qty = REG_QTY (REGNO (folded_arg0));
3431 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3433 struct qty_table_elem *ent = &qty_table[qty];
3435 if ((comparison_dominates_p (ent->comparison_code, code)
3436 || (! FLOAT_MODE_P (mode_arg0)
3437 && comparison_dominates_p (ent->comparison_code,
3438 reverse_condition (code))))
3439 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3440 || (const_arg1
3441 && rtx_equal_p (ent->comparison_const,
3442 const_arg1))
3443 || (REG_P (folded_arg1)
3444 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3446 if (comparison_dominates_p (ent->comparison_code, code))
3448 if (true_rtx)
3449 return true_rtx;
3450 else
3451 break;
3453 else
3454 return false_rtx;
3461 /* If we are comparing against zero, see if the first operand is
3462 equivalent to an IOR with a constant. If so, we may be able to
3463 determine the result of this comparison. */
3464 if (const_arg1 == const0_rtx && !const_arg0)
3466 rtx y = lookup_as_function (folded_arg0, IOR);
3467 rtx inner_const;
3469 if (y != 0
3470 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3471 && CONST_INT_P (inner_const)
3472 && INTVAL (inner_const) != 0)
3473 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3477 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3478 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3479 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3480 op0, op1);
3482 break;
3484 case RTX_BIN_ARITH:
3485 case RTX_COMM_ARITH:
3486 switch (code)
3488 case PLUS:
3489 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3490 with that LABEL_REF as its second operand. If so, the result is
3491 the first operand of that MINUS. This handles switches with an
3492 ADDR_DIFF_VEC table. */
3493 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3495 rtx y
3496 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3497 : lookup_as_function (folded_arg0, MINUS);
3499 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3500 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg1))
3501 return XEXP (y, 0);
3503 /* Now try for a CONST of a MINUS like the above. */
3504 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3505 : lookup_as_function (folded_arg0, CONST))) != 0
3506 && GET_CODE (XEXP (y, 0)) == MINUS
3507 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3508 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg1))
3509 return XEXP (XEXP (y, 0), 0);
3512 /* Likewise if the operands are in the other order. */
3513 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3515 rtx y
3516 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3517 : lookup_as_function (folded_arg1, MINUS);
3519 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3520 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg0))
3521 return XEXP (y, 0);
3523 /* Now try for a CONST of a MINUS like the above. */
3524 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3525 : lookup_as_function (folded_arg1, CONST))) != 0
3526 && GET_CODE (XEXP (y, 0)) == MINUS
3527 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3528 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg0))
3529 return XEXP (XEXP (y, 0), 0);
3532 /* If second operand is a register equivalent to a negative
3533 CONST_INT, see if we can find a register equivalent to the
3534 positive constant. Make a MINUS if so. Don't do this for
3535 a non-negative constant since we might then alternate between
3536 choosing positive and negative constants. Having the positive
3537 constant previously-used is the more common case. Be sure
3538 the resulting constant is non-negative; if const_arg1 were
3539 the smallest negative number this would overflow: depending
3540 on the mode, this would either just be the same value (and
3541 hence not save anything) or be incorrect. */
3542 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3543 && INTVAL (const_arg1) < 0
3544 /* This used to test
3546 -INTVAL (const_arg1) >= 0
3548 But The Sun V5.0 compilers mis-compiled that test. So
3549 instead we test for the problematic value in a more direct
3550 manner and hope the Sun compilers get it correct. */
3551 && INTVAL (const_arg1) !=
3552 (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1))
3553 && REG_P (folded_arg1))
3555 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3556 struct table_elt *p
3557 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3559 if (p)
3560 for (p = p->first_same_value; p; p = p->next_same_value)
3561 if (REG_P (p->exp))
3562 return simplify_gen_binary (MINUS, mode, folded_arg0,
3563 canon_reg (p->exp, NULL));
3565 goto from_plus;
3567 case MINUS:
3568 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3569 If so, produce (PLUS Z C2-C). */
3570 if (const_arg1 != 0 && poly_int_rtx_p (const_arg1, &xval))
3572 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3573 if (y && poly_int_rtx_p (XEXP (y, 1)))
3574 return fold_rtx (plus_constant (mode, copy_rtx (y), -xval),
3575 NULL);
3578 /* Fall through. */
3580 from_plus:
3581 case SMIN: case SMAX: case UMIN: case UMAX:
3582 case IOR: case AND: case XOR:
3583 case MULT:
3584 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3585 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3586 is known to be of similar form, we may be able to replace the
3587 operation with a combined operation. This may eliminate the
3588 intermediate operation if every use is simplified in this way.
3589 Note that the similar optimization done by combine.c only works
3590 if the intermediate operation's result has only one reference. */
3592 if (REG_P (folded_arg0)
3593 && const_arg1 && CONST_INT_P (const_arg1))
3595 int is_shift
3596 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3597 rtx y, inner_const, new_const;
3598 rtx canon_const_arg1 = const_arg1;
3599 enum rtx_code associate_code;
3601 if (is_shift
3602 && (INTVAL (const_arg1) >= GET_MODE_UNIT_PRECISION (mode)
3603 || INTVAL (const_arg1) < 0))
3605 if (SHIFT_COUNT_TRUNCATED)
3606 canon_const_arg1 = gen_int_shift_amount
3607 (mode, (INTVAL (const_arg1)
3608 & (GET_MODE_UNIT_BITSIZE (mode) - 1)));
3609 else
3610 break;
3613 y = lookup_as_function (folded_arg0, code);
3614 if (y == 0)
3615 break;
3617 /* If we have compiled a statement like
3618 "if (x == (x & mask1))", and now are looking at
3619 "x & mask2", we will have a case where the first operand
3620 of Y is the same as our first operand. Unless we detect
3621 this case, an infinite loop will result. */
3622 if (XEXP (y, 0) == folded_arg0)
3623 break;
3625 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3626 if (!inner_const || !CONST_INT_P (inner_const))
3627 break;
3629 /* Don't associate these operations if they are a PLUS with the
3630 same constant and it is a power of two. These might be doable
3631 with a pre- or post-increment. Similarly for two subtracts of
3632 identical powers of two with post decrement. */
3634 if (code == PLUS && const_arg1 == inner_const
3635 && ((HAVE_PRE_INCREMENT
3636 && pow2p_hwi (INTVAL (const_arg1)))
3637 || (HAVE_POST_INCREMENT
3638 && pow2p_hwi (INTVAL (const_arg1)))
3639 || (HAVE_PRE_DECREMENT
3640 && pow2p_hwi (- INTVAL (const_arg1)))
3641 || (HAVE_POST_DECREMENT
3642 && pow2p_hwi (- INTVAL (const_arg1)))))
3643 break;
3645 /* ??? Vector mode shifts by scalar
3646 shift operand are not supported yet. */
3647 if (is_shift && VECTOR_MODE_P (mode))
3648 break;
3650 if (is_shift
3651 && (INTVAL (inner_const) >= GET_MODE_UNIT_PRECISION (mode)
3652 || INTVAL (inner_const) < 0))
3654 if (SHIFT_COUNT_TRUNCATED)
3655 inner_const = gen_int_shift_amount
3656 (mode, (INTVAL (inner_const)
3657 & (GET_MODE_UNIT_BITSIZE (mode) - 1)));
3658 else
3659 break;
3662 /* Compute the code used to compose the constants. For example,
3663 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3665 associate_code = (is_shift || code == MINUS ? PLUS : code);
3667 new_const = simplify_binary_operation (associate_code, mode,
3668 canon_const_arg1,
3669 inner_const);
3671 if (new_const == 0)
3672 break;
3674 /* If we are associating shift operations, don't let this
3675 produce a shift of the size of the object or larger.
3676 This could occur when we follow a sign-extend by a right
3677 shift on a machine that does a sign-extend as a pair
3678 of shifts. */
3680 if (is_shift
3681 && CONST_INT_P (new_const)
3682 && INTVAL (new_const) >= GET_MODE_UNIT_PRECISION (mode))
3684 /* As an exception, we can turn an ASHIFTRT of this
3685 form into a shift of the number of bits - 1. */
3686 if (code == ASHIFTRT)
3687 new_const = gen_int_shift_amount
3688 (mode, GET_MODE_UNIT_BITSIZE (mode) - 1);
3689 else if (!side_effects_p (XEXP (y, 0)))
3690 return CONST0_RTX (mode);
3691 else
3692 break;
3695 y = copy_rtx (XEXP (y, 0));
3697 /* If Y contains our first operand (the most common way this
3698 can happen is if Y is a MEM), we would do into an infinite
3699 loop if we tried to fold it. So don't in that case. */
3701 if (! reg_mentioned_p (folded_arg0, y))
3702 y = fold_rtx (y, insn);
3704 return simplify_gen_binary (code, mode, y, new_const);
3706 break;
3708 case DIV: case UDIV:
3709 /* ??? The associative optimization performed immediately above is
3710 also possible for DIV and UDIV using associate_code of MULT.
3711 However, we would need extra code to verify that the
3712 multiplication does not overflow, that is, there is no overflow
3713 in the calculation of new_const. */
3714 break;
3716 default:
3717 break;
3720 new_rtx = simplify_binary_operation (code, mode,
3721 const_arg0 ? const_arg0 : folded_arg0,
3722 const_arg1 ? const_arg1 : folded_arg1);
3723 break;
3725 case RTX_OBJ:
3726 /* (lo_sum (high X) X) is simply X. */
3727 if (code == LO_SUM && const_arg0 != 0
3728 && GET_CODE (const_arg0) == HIGH
3729 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3730 return const_arg1;
3731 break;
3733 case RTX_TERNARY:
3734 case RTX_BITFIELD_OPS:
3735 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3736 const_arg0 ? const_arg0 : folded_arg0,
3737 const_arg1 ? const_arg1 : folded_arg1,
3738 const_arg2 ? const_arg2 : XEXP (x, 2));
3739 break;
3741 default:
3742 break;
3745 return new_rtx ? new_rtx : x;
3748 /* Return a constant value currently equivalent to X.
3749 Return 0 if we don't know one. */
3751 static rtx
3752 equiv_constant (rtx x)
3754 if (REG_P (x)
3755 && REGNO_QTY_VALID_P (REGNO (x)))
3757 int x_q = REG_QTY (REGNO (x));
3758 struct qty_table_elem *x_ent = &qty_table[x_q];
3760 if (x_ent->const_rtx)
3761 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3764 if (x == 0 || CONSTANT_P (x))
3765 return x;
3767 if (GET_CODE (x) == SUBREG)
3769 machine_mode mode = GET_MODE (x);
3770 machine_mode imode = GET_MODE (SUBREG_REG (x));
3771 rtx new_rtx;
3773 /* See if we previously assigned a constant value to this SUBREG. */
3774 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3775 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3776 || (NUM_POLY_INT_COEFFS > 1
3777 && (new_rtx = lookup_as_function (x, CONST_POLY_INT)) != 0)
3778 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3779 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3780 return new_rtx;
3782 /* If we didn't and if doing so makes sense, see if we previously
3783 assigned a constant value to the enclosing word mode SUBREG. */
3784 if (known_lt (GET_MODE_SIZE (mode), UNITS_PER_WORD)
3785 && known_lt (UNITS_PER_WORD, GET_MODE_SIZE (imode)))
3787 poly_int64 byte = (SUBREG_BYTE (x)
3788 - subreg_lowpart_offset (mode, word_mode));
3789 if (known_ge (byte, 0) && multiple_p (byte, UNITS_PER_WORD))
3791 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3792 new_rtx = lookup_as_function (y, CONST_INT);
3793 if (new_rtx)
3794 return gen_lowpart (mode, new_rtx);
3798 /* Otherwise see if we already have a constant for the inner REG,
3799 and if that is enough to calculate an equivalent constant for
3800 the subreg. Note that the upper bits of paradoxical subregs
3801 are undefined, so they cannot be said to equal anything. */
3802 if (REG_P (SUBREG_REG (x))
3803 && !paradoxical_subreg_p (x)
3804 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3805 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3807 return 0;
3810 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3811 the hash table in case its value was seen before. */
3813 if (MEM_P (x))
3815 struct table_elt *elt;
3817 x = avoid_constant_pool_reference (x);
3818 if (CONSTANT_P (x))
3819 return x;
3821 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3822 if (elt == 0)
3823 return 0;
3825 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3826 if (elt->is_const && CONSTANT_P (elt->exp))
3827 return elt->exp;
3830 return 0;
3833 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3834 "taken" branch.
3836 In certain cases, this can cause us to add an equivalence. For example,
3837 if we are following the taken case of
3838 if (i == 2)
3839 we can add the fact that `i' and '2' are now equivalent.
3841 In any case, we can record that this comparison was passed. If the same
3842 comparison is seen later, we will know its value. */
3844 static void
3845 record_jump_equiv (rtx_insn *insn, bool taken)
3847 int cond_known_true;
3848 rtx op0, op1;
3849 rtx set;
3850 machine_mode mode, mode0, mode1;
3851 int reversed_nonequality = 0;
3852 enum rtx_code code;
3854 /* Ensure this is the right kind of insn. */
3855 gcc_assert (any_condjump_p (insn));
3857 set = pc_set (insn);
3859 /* See if this jump condition is known true or false. */
3860 if (taken)
3861 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3862 else
3863 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3865 /* Get the type of comparison being done and the operands being compared.
3866 If we had to reverse a non-equality condition, record that fact so we
3867 know that it isn't valid for floating-point. */
3868 code = GET_CODE (XEXP (SET_SRC (set), 0));
3869 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3870 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3872 /* If fold_rtx returns NULL_RTX, there's nothing to record. */
3873 if (op0 == NULL_RTX || op1 == NULL_RTX)
3874 return;
3876 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3877 if (! cond_known_true)
3879 code = reversed_comparison_code_parts (code, op0, op1, insn);
3881 /* Don't remember if we can't find the inverse. */
3882 if (code == UNKNOWN)
3883 return;
3886 /* The mode is the mode of the non-constant. */
3887 mode = mode0;
3888 if (mode1 != VOIDmode)
3889 mode = mode1;
3891 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3894 /* Yet another form of subreg creation. In this case, we want something in
3895 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3897 static rtx
3898 record_jump_cond_subreg (machine_mode mode, rtx op)
3900 machine_mode op_mode = GET_MODE (op);
3901 if (op_mode == mode || op_mode == VOIDmode)
3902 return op;
3903 return lowpart_subreg (mode, op, op_mode);
3906 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3907 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3908 Make any useful entries we can with that information. Called from
3909 above function and called recursively. */
3911 static void
3912 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3913 rtx op1, int reversed_nonequality)
3915 unsigned op0_hash, op1_hash;
3916 int op0_in_memory, op1_in_memory;
3917 struct table_elt *op0_elt, *op1_elt;
3919 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3920 we know that they are also equal in the smaller mode (this is also
3921 true for all smaller modes whether or not there is a SUBREG, but
3922 is not worth testing for with no SUBREG). */
3924 /* Note that GET_MODE (op0) may not equal MODE. */
3925 if (code == EQ && paradoxical_subreg_p (op0))
3927 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3928 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3929 if (tem)
3930 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3931 reversed_nonequality);
3934 if (code == EQ && paradoxical_subreg_p (op1))
3936 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3937 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3938 if (tem)
3939 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3940 reversed_nonequality);
3943 /* Similarly, if this is an NE comparison, and either is a SUBREG
3944 making a smaller mode, we know the whole thing is also NE. */
3946 /* Note that GET_MODE (op0) may not equal MODE;
3947 if we test MODE instead, we can get an infinite recursion
3948 alternating between two modes each wider than MODE. */
3950 if (code == NE
3951 && partial_subreg_p (op0)
3952 && subreg_lowpart_p (op0))
3954 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3955 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3956 if (tem)
3957 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3958 reversed_nonequality);
3961 if (code == NE
3962 && partial_subreg_p (op1)
3963 && subreg_lowpart_p (op1))
3965 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3966 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3967 if (tem)
3968 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3969 reversed_nonequality);
3972 /* Hash both operands. */
3974 do_not_record = 0;
3975 hash_arg_in_memory = 0;
3976 op0_hash = HASH (op0, mode);
3977 op0_in_memory = hash_arg_in_memory;
3979 if (do_not_record)
3980 return;
3982 do_not_record = 0;
3983 hash_arg_in_memory = 0;
3984 op1_hash = HASH (op1, mode);
3985 op1_in_memory = hash_arg_in_memory;
3987 if (do_not_record)
3988 return;
3990 /* Look up both operands. */
3991 op0_elt = lookup (op0, op0_hash, mode);
3992 op1_elt = lookup (op1, op1_hash, mode);
3994 /* If both operands are already equivalent or if they are not in the
3995 table but are identical, do nothing. */
3996 if ((op0_elt != 0 && op1_elt != 0
3997 && op0_elt->first_same_value == op1_elt->first_same_value)
3998 || op0 == op1 || rtx_equal_p (op0, op1))
3999 return;
4001 /* If we aren't setting two things equal all we can do is save this
4002 comparison. Similarly if this is floating-point. In the latter
4003 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4004 If we record the equality, we might inadvertently delete code
4005 whose intent was to change -0 to +0. */
4007 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4009 struct qty_table_elem *ent;
4010 int qty;
4012 /* If we reversed a floating-point comparison, if OP0 is not a
4013 register, or if OP1 is neither a register or constant, we can't
4014 do anything. */
4016 if (!REG_P (op1))
4017 op1 = equiv_constant (op1);
4019 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4020 || !REG_P (op0) || op1 == 0)
4021 return;
4023 /* Put OP0 in the hash table if it isn't already. This gives it a
4024 new quantity number. */
4025 if (op0_elt == 0)
4027 if (insert_regs (op0, NULL, 0))
4029 rehash_using_reg (op0);
4030 op0_hash = HASH (op0, mode);
4032 /* If OP0 is contained in OP1, this changes its hash code
4033 as well. Faster to rehash than to check, except
4034 for the simple case of a constant. */
4035 if (! CONSTANT_P (op1))
4036 op1_hash = HASH (op1,mode);
4039 op0_elt = insert (op0, NULL, op0_hash, mode);
4040 op0_elt->in_memory = op0_in_memory;
4043 qty = REG_QTY (REGNO (op0));
4044 ent = &qty_table[qty];
4046 ent->comparison_code = code;
4047 if (REG_P (op1))
4049 /* Look it up again--in case op0 and op1 are the same. */
4050 op1_elt = lookup (op1, op1_hash, mode);
4052 /* Put OP1 in the hash table so it gets a new quantity number. */
4053 if (op1_elt == 0)
4055 if (insert_regs (op1, NULL, 0))
4057 rehash_using_reg (op1);
4058 op1_hash = HASH (op1, mode);
4061 op1_elt = insert (op1, NULL, op1_hash, mode);
4062 op1_elt->in_memory = op1_in_memory;
4065 ent->comparison_const = NULL_RTX;
4066 ent->comparison_qty = REG_QTY (REGNO (op1));
4068 else
4070 ent->comparison_const = op1;
4071 ent->comparison_qty = -1;
4074 return;
4077 /* If either side is still missing an equivalence, make it now,
4078 then merge the equivalences. */
4080 if (op0_elt == 0)
4082 if (insert_regs (op0, NULL, 0))
4084 rehash_using_reg (op0);
4085 op0_hash = HASH (op0, mode);
4088 op0_elt = insert (op0, NULL, op0_hash, mode);
4089 op0_elt->in_memory = op0_in_memory;
4092 if (op1_elt == 0)
4094 if (insert_regs (op1, NULL, 0))
4096 rehash_using_reg (op1);
4097 op1_hash = HASH (op1, mode);
4100 op1_elt = insert (op1, NULL, op1_hash, mode);
4101 op1_elt->in_memory = op1_in_memory;
4104 merge_equiv_classes (op0_elt, op1_elt);
4107 /* CSE processing for one instruction.
4109 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4110 but the few that "leak through" are cleaned up by cse_insn, and complex
4111 addressing modes are often formed here.
4113 The main function is cse_insn, and between here and that function
4114 a couple of helper functions is defined to keep the size of cse_insn
4115 within reasonable proportions.
4117 Data is shared between the main and helper functions via STRUCT SET,
4118 that contains all data related for every set in the instruction that
4119 is being processed.
4121 Note that cse_main processes all sets in the instruction. Most
4122 passes in GCC only process simple SET insns or single_set insns, but
4123 CSE processes insns with multiple sets as well. */
4125 /* Data on one SET contained in the instruction. */
4127 struct set
4129 /* The SET rtx itself. */
4130 rtx rtl;
4131 /* The SET_SRC of the rtx (the original value, if it is changing). */
4132 rtx src;
4133 /* The hash-table element for the SET_SRC of the SET. */
4134 struct table_elt *src_elt;
4135 /* Hash value for the SET_SRC. */
4136 unsigned src_hash;
4137 /* Hash value for the SET_DEST. */
4138 unsigned dest_hash;
4139 /* The SET_DEST, with SUBREG, etc., stripped. */
4140 rtx inner_dest;
4141 /* Nonzero if the SET_SRC is in memory. */
4142 char src_in_memory;
4143 /* Nonzero if the SET_SRC contains something
4144 whose value cannot be predicted and understood. */
4145 char src_volatile;
4146 /* Original machine mode, in case it becomes a CONST_INT.
4147 The size of this field should match the size of the mode
4148 field of struct rtx_def (see rtl.h). */
4149 ENUM_BITFIELD(machine_mode) mode : 8;
4150 /* Hash value of constant equivalent for SET_SRC. */
4151 unsigned src_const_hash;
4152 /* A constant equivalent for SET_SRC, if any. */
4153 rtx src_const;
4154 /* Table entry for constant equivalent for SET_SRC, if any. */
4155 struct table_elt *src_const_elt;
4156 /* Table entry for the destination address. */
4157 struct table_elt *dest_addr_elt;
4160 /* Special handling for (set REG0 REG1) where REG0 is the
4161 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4162 be used in the sequel, so (if easily done) change this insn to
4163 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4164 that computed their value. Then REG1 will become a dead store
4165 and won't cloud the situation for later optimizations.
4167 Do not make this change if REG1 is a hard register, because it will
4168 then be used in the sequel and we may be changing a two-operand insn
4169 into a three-operand insn.
4171 This is the last transformation that cse_insn will try to do. */
4173 static void
4174 try_back_substitute_reg (rtx set, rtx_insn *insn)
4176 rtx dest = SET_DEST (set);
4177 rtx src = SET_SRC (set);
4179 if (REG_P (dest)
4180 && REG_P (src) && ! HARD_REGISTER_P (src)
4181 && REGNO_QTY_VALID_P (REGNO (src)))
4183 int src_q = REG_QTY (REGNO (src));
4184 struct qty_table_elem *src_ent = &qty_table[src_q];
4186 if (src_ent->first_reg == REGNO (dest))
4188 /* Scan for the previous nonnote insn, but stop at a basic
4189 block boundary. */
4190 rtx_insn *prev = insn;
4191 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4194 prev = PREV_INSN (prev);
4196 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4198 /* Do not swap the registers around if the previous instruction
4199 attaches a REG_EQUIV note to REG1.
4201 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4202 from the pseudo that originally shadowed an incoming argument
4203 to another register. Some uses of REG_EQUIV might rely on it
4204 being attached to REG1 rather than REG2.
4206 This section previously turned the REG_EQUIV into a REG_EQUAL
4207 note. We cannot do that because REG_EQUIV may provide an
4208 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4209 if (NONJUMP_INSN_P (prev)
4210 && GET_CODE (PATTERN (prev)) == SET
4211 && SET_DEST (PATTERN (prev)) == src
4212 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4214 rtx note;
4216 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4217 validate_change (insn, &SET_DEST (set), src, 1);
4218 validate_change (insn, &SET_SRC (set), dest, 1);
4219 apply_change_group ();
4221 /* If INSN has a REG_EQUAL note, and this note mentions
4222 REG0, then we must delete it, because the value in
4223 REG0 has changed. If the note's value is REG1, we must
4224 also delete it because that is now this insn's dest. */
4225 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4226 if (note != 0
4227 && (reg_mentioned_p (dest, XEXP (note, 0))
4228 || rtx_equal_p (src, XEXP (note, 0))))
4229 remove_note (insn, note);
4231 /* If INSN has a REG_ARGS_SIZE note, move it to PREV. */
4232 note = find_reg_note (insn, REG_ARGS_SIZE, NULL_RTX);
4233 if (note != 0)
4235 remove_note (insn, note);
4236 gcc_assert (!find_reg_note (prev, REG_ARGS_SIZE, NULL_RTX));
4237 set_unique_reg_note (prev, REG_ARGS_SIZE, XEXP (note, 0));
4244 /* Add an entry containing RTL X into SETS. */
4245 static inline void
4246 add_to_set (vec<struct set> *sets, rtx x)
4248 struct set entry = {};
4249 entry.rtl = x;
4250 sets->safe_push (entry);
4253 /* Record all the SETs in this instruction into SETS_PTR,
4254 and return the number of recorded sets. */
4255 static int
4256 find_sets_in_insn (rtx_insn *insn, vec<struct set> *psets)
4258 rtx x = PATTERN (insn);
4260 if (GET_CODE (x) == SET)
4262 /* Ignore SETs that are unconditional jumps.
4263 They never need cse processing, so this does not hurt.
4264 The reason is not efficiency but rather
4265 so that we can test at the end for instructions
4266 that have been simplified to unconditional jumps
4267 and not be misled by unchanged instructions
4268 that were unconditional jumps to begin with. */
4269 if (SET_DEST (x) == pc_rtx
4270 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4272 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4273 The hard function value register is used only once, to copy to
4274 someplace else, so it isn't worth cse'ing. */
4275 else if (GET_CODE (SET_SRC (x)) == CALL)
4277 else if (GET_CODE (SET_SRC (x)) == CONST_VECTOR
4278 && GET_MODE_CLASS (GET_MODE (SET_SRC (x))) != MODE_VECTOR_BOOL)
4280 /* First register the vector itself. */
4281 add_to_set (psets, x);
4282 rtx src = SET_SRC (x);
4283 /* Go over the constants of the CONST_VECTOR in forward order, to
4284 put them in the same order in the SETS array. */
4285 for (unsigned i = 0; i < const_vector_encoded_nelts (src) ; i++)
4287 /* These are templates and don't actually get emitted but are
4288 used to tell CSE how to get to a particular constant. */
4289 rtx y = simplify_gen_vec_select (SET_DEST (x), i);
4290 gcc_assert (y);
4291 add_to_set (psets, gen_rtx_SET (y, CONST_VECTOR_ELT (src, i)));
4294 else
4295 add_to_set (psets, x);
4297 else if (GET_CODE (x) == PARALLEL)
4299 int i, lim = XVECLEN (x, 0);
4301 /* Go over the expressions of the PARALLEL in forward order, to
4302 put them in the same order in the SETS array. */
4303 for (i = 0; i < lim; i++)
4305 rtx y = XVECEXP (x, 0, i);
4306 if (GET_CODE (y) == SET)
4308 /* As above, we ignore unconditional jumps and call-insns and
4309 ignore the result of apply_change_group. */
4310 if (SET_DEST (y) == pc_rtx
4311 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4313 else if (GET_CODE (SET_SRC (y)) == CALL)
4315 else
4316 add_to_set (psets, y);
4321 return psets->length ();
4324 /* Subroutine of canonicalize_insn. X is an ASM_OPERANDS in INSN. */
4326 static void
4327 canon_asm_operands (rtx x, rtx_insn *insn)
4329 for (int i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4331 rtx input = ASM_OPERANDS_INPUT (x, i);
4332 if (!(REG_P (input) && HARD_REGISTER_P (input)))
4334 input = canon_reg (input, insn);
4335 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4340 /* Where possible, substitute every register reference in the N_SETS
4341 number of SETS in INSN with the canonical register.
4343 Register canonicalization propagatest the earliest register (i.e.
4344 one that is set before INSN) with the same value. This is a very
4345 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4346 to RTL. For instance, a CONST for an address is usually expanded
4347 multiple times to loads into different registers, thus creating many
4348 subexpressions of the form:
4350 (set (reg1) (some_const))
4351 (set (mem (... reg1 ...) (thing)))
4352 (set (reg2) (some_const))
4353 (set (mem (... reg2 ...) (thing)))
4355 After canonicalizing, the code takes the following form:
4357 (set (reg1) (some_const))
4358 (set (mem (... reg1 ...) (thing)))
4359 (set (reg2) (some_const))
4360 (set (mem (... reg1 ...) (thing)))
4362 The set to reg2 is now trivially dead, and the memory reference (or
4363 address, or whatever) may be a candidate for further CSEing.
4365 In this function, the result of apply_change_group can be ignored;
4366 see canon_reg. */
4368 static void
4369 canonicalize_insn (rtx_insn *insn, vec<struct set> *psets)
4371 vec<struct set> sets = *psets;
4372 int n_sets = sets.length ();
4373 rtx tem;
4374 rtx x = PATTERN (insn);
4375 int i;
4377 if (CALL_P (insn))
4379 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4380 if (GET_CODE (XEXP (tem, 0)) != SET)
4381 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4384 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4386 canon_reg (SET_SRC (x), insn);
4387 apply_change_group ();
4388 fold_rtx (SET_SRC (x), insn);
4390 else if (GET_CODE (x) == CLOBBER)
4392 /* If we clobber memory, canon the address.
4393 This does nothing when a register is clobbered
4394 because we have already invalidated the reg. */
4395 if (MEM_P (XEXP (x, 0)))
4396 canon_reg (XEXP (x, 0), insn);
4398 else if (GET_CODE (x) == USE
4399 && ! (REG_P (XEXP (x, 0))
4400 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4401 /* Canonicalize a USE of a pseudo register or memory location. */
4402 canon_reg (x, insn);
4403 else if (GET_CODE (x) == ASM_OPERANDS)
4404 canon_asm_operands (x, insn);
4405 else if (GET_CODE (x) == CALL)
4407 canon_reg (x, insn);
4408 apply_change_group ();
4409 fold_rtx (x, insn);
4411 else if (DEBUG_INSN_P (insn))
4412 canon_reg (PATTERN (insn), insn);
4413 else if (GET_CODE (x) == PARALLEL)
4415 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4417 rtx y = XVECEXP (x, 0, i);
4418 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4420 canon_reg (SET_SRC (y), insn);
4421 apply_change_group ();
4422 fold_rtx (SET_SRC (y), insn);
4424 else if (GET_CODE (y) == CLOBBER)
4426 if (MEM_P (XEXP (y, 0)))
4427 canon_reg (XEXP (y, 0), insn);
4429 else if (GET_CODE (y) == USE
4430 && ! (REG_P (XEXP (y, 0))
4431 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4432 canon_reg (y, insn);
4433 else if (GET_CODE (y) == ASM_OPERANDS)
4434 canon_asm_operands (y, insn);
4435 else if (GET_CODE (y) == CALL)
4437 canon_reg (y, insn);
4438 apply_change_group ();
4439 fold_rtx (y, insn);
4444 if (n_sets == 1 && REG_NOTES (insn) != 0
4445 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4447 /* We potentially will process this insn many times. Therefore,
4448 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4449 unique set in INSN.
4451 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4452 because cse_insn handles those specially. */
4453 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4454 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4455 remove_note (insn, tem);
4456 else
4458 canon_reg (XEXP (tem, 0), insn);
4459 apply_change_group ();
4460 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4461 df_notes_rescan (insn);
4465 /* Canonicalize sources and addresses of destinations.
4466 We do this in a separate pass to avoid problems when a MATCH_DUP is
4467 present in the insn pattern. In that case, we want to ensure that
4468 we don't break the duplicate nature of the pattern. So we will replace
4469 both operands at the same time. Otherwise, we would fail to find an
4470 equivalent substitution in the loop calling validate_change below.
4472 We used to suppress canonicalization of DEST if it appears in SRC,
4473 but we don't do this any more. */
4475 for (i = 0; i < n_sets; i++)
4477 rtx dest = SET_DEST (sets[i].rtl);
4478 rtx src = SET_SRC (sets[i].rtl);
4479 rtx new_rtx = canon_reg (src, insn);
4481 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4483 if (GET_CODE (dest) == ZERO_EXTRACT)
4485 validate_change (insn, &XEXP (dest, 1),
4486 canon_reg (XEXP (dest, 1), insn), 1);
4487 validate_change (insn, &XEXP (dest, 2),
4488 canon_reg (XEXP (dest, 2), insn), 1);
4491 while (GET_CODE (dest) == SUBREG
4492 || GET_CODE (dest) == ZERO_EXTRACT
4493 || GET_CODE (dest) == STRICT_LOW_PART)
4494 dest = XEXP (dest, 0);
4496 if (MEM_P (dest))
4497 canon_reg (dest, insn);
4500 /* Now that we have done all the replacements, we can apply the change
4501 group and see if they all work. Note that this will cause some
4502 canonicalizations that would have worked individually not to be applied
4503 because some other canonicalization didn't work, but this should not
4504 occur often.
4506 The result of apply_change_group can be ignored; see canon_reg. */
4508 apply_change_group ();
4511 /* Main function of CSE.
4512 First simplify sources and addresses of all assignments
4513 in the instruction, using previously-computed equivalents values.
4514 Then install the new sources and destinations in the table
4515 of available values. */
4517 static void
4518 cse_insn (rtx_insn *insn)
4520 rtx x = PATTERN (insn);
4521 int i;
4522 rtx tem;
4523 int n_sets = 0;
4525 rtx src_eqv = 0;
4526 struct table_elt *src_eqv_elt = 0;
4527 int src_eqv_volatile = 0;
4528 int src_eqv_in_memory = 0;
4529 unsigned src_eqv_hash = 0;
4531 this_insn = insn;
4533 /* Find all regs explicitly clobbered in this insn,
4534 to ensure they are not replaced with any other regs
4535 elsewhere in this insn. */
4536 invalidate_from_sets_and_clobbers (insn);
4538 /* Record all the SETs in this instruction. */
4539 auto_vec<struct set, 8> sets;
4540 n_sets = find_sets_in_insn (insn, (vec<struct set>*)&sets);
4542 /* Substitute the canonical register where possible. */
4543 canonicalize_insn (insn, (vec<struct set>*)&sets);
4545 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4546 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4547 latter condition is necessary because SRC_EQV is handled specially for
4548 this case, and if it isn't set, then there will be no equivalence
4549 for the destination. */
4550 if (n_sets == 1 && REG_NOTES (insn) != 0
4551 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4554 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4555 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4556 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4557 src_eqv = copy_rtx (XEXP (tem, 0));
4558 /* If DEST is of the form ZERO_EXTACT, as in:
4559 (set (zero_extract:SI (reg:SI 119)
4560 (const_int 16 [0x10])
4561 (const_int 16 [0x10]))
4562 (const_int 51154 [0xc7d2]))
4563 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4564 point. Note that this is different from SRC_EQV. We can however
4565 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4566 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4567 && CONST_INT_P (XEXP (tem, 0))
4568 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4569 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4571 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4572 /* This is the mode of XEXP (tem, 0) as well. */
4573 scalar_int_mode dest_mode
4574 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
4575 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4576 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4577 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4578 HOST_WIDE_INT mask;
4579 unsigned int shift;
4580 if (BITS_BIG_ENDIAN)
4581 shift = (GET_MODE_PRECISION (dest_mode)
4582 - INTVAL (pos) - INTVAL (width));
4583 else
4584 shift = INTVAL (pos);
4585 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4586 mask = HOST_WIDE_INT_M1;
4587 else
4588 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
4589 val = (val >> shift) & mask;
4590 src_eqv = GEN_INT (val);
4594 /* Set sets[i].src_elt to the class each source belongs to.
4595 Detect assignments from or to volatile things
4596 and set set[i] to zero so they will be ignored
4597 in the rest of this function.
4599 Nothing in this loop changes the hash table or the register chains. */
4601 for (i = 0; i < n_sets; i++)
4603 bool repeat = false;
4604 bool noop_insn = false;
4605 rtx src, dest;
4606 rtx src_folded;
4607 struct table_elt *elt = 0, *p;
4608 machine_mode mode;
4609 rtx src_eqv_here;
4610 rtx src_const = 0;
4611 rtx src_related = 0;
4612 bool src_related_is_const_anchor = false;
4613 struct table_elt *src_const_elt = 0;
4614 int src_cost = MAX_COST;
4615 int src_eqv_cost = MAX_COST;
4616 int src_folded_cost = MAX_COST;
4617 int src_related_cost = MAX_COST;
4618 int src_elt_cost = MAX_COST;
4619 int src_regcost = MAX_COST;
4620 int src_eqv_regcost = MAX_COST;
4621 int src_folded_regcost = MAX_COST;
4622 int src_related_regcost = MAX_COST;
4623 int src_elt_regcost = MAX_COST;
4624 scalar_int_mode int_mode;
4626 dest = SET_DEST (sets[i].rtl);
4627 src = SET_SRC (sets[i].rtl);
4629 /* If SRC is a constant that has no machine mode,
4630 hash it with the destination's machine mode.
4631 This way we can keep different modes separate. */
4633 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4634 sets[i].mode = mode;
4636 if (src_eqv)
4638 machine_mode eqvmode = mode;
4639 if (GET_CODE (dest) == STRICT_LOW_PART)
4640 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4641 do_not_record = 0;
4642 hash_arg_in_memory = 0;
4643 src_eqv_hash = HASH (src_eqv, eqvmode);
4645 /* Find the equivalence class for the equivalent expression. */
4647 if (!do_not_record)
4648 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4650 src_eqv_volatile = do_not_record;
4651 src_eqv_in_memory = hash_arg_in_memory;
4654 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4655 value of the INNER register, not the destination. So it is not
4656 a valid substitution for the source. But save it for later. */
4657 if (GET_CODE (dest) == STRICT_LOW_PART)
4658 src_eqv_here = 0;
4659 else
4660 src_eqv_here = src_eqv;
4662 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4663 simplified result, which may not necessarily be valid. */
4664 src_folded = fold_rtx (src, NULL);
4666 #if 0
4667 /* ??? This caused bad code to be generated for the m68k port with -O2.
4668 Suppose src is (CONST_INT -1), and that after truncation src_folded
4669 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4670 At the end we will add src and src_const to the same equivalence
4671 class. We now have 3 and -1 on the same equivalence class. This
4672 causes later instructions to be mis-optimized. */
4673 /* If storing a constant in a bitfield, pre-truncate the constant
4674 so we will be able to record it later. */
4675 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4677 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4679 if (CONST_INT_P (src)
4680 && CONST_INT_P (width)
4681 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4682 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4683 src_folded
4684 = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1
4685 << INTVAL (width)) - 1));
4687 #endif
4689 /* Compute SRC's hash code, and also notice if it
4690 should not be recorded at all. In that case,
4691 prevent any further processing of this assignment.
4693 We set DO_NOT_RECORD if the destination has a REG_UNUSED note.
4694 This avoids getting the source register into the tables, where it
4695 may be invalidated later (via REG_QTY), then trigger an ICE upon
4696 re-insertion.
4698 This is only a problem in multi-set insns. If it were a single
4699 set the dead copy would have been removed. If the RHS were anything
4700 but a simple REG, then we won't call insert_regs and thus there's
4701 no potential for triggering the ICE. */
4702 do_not_record = (REG_P (dest)
4703 && REG_P (src)
4704 && find_reg_note (insn, REG_UNUSED, dest));
4705 hash_arg_in_memory = 0;
4707 sets[i].src = src;
4708 sets[i].src_hash = HASH (src, mode);
4709 sets[i].src_volatile = do_not_record;
4710 sets[i].src_in_memory = hash_arg_in_memory;
4712 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4713 a pseudo, do not record SRC. Using SRC as a replacement for
4714 anything else will be incorrect in that situation. Note that
4715 this usually occurs only for stack slots, in which case all the
4716 RTL would be referring to SRC, so we don't lose any optimization
4717 opportunities by not having SRC in the hash table. */
4719 if (MEM_P (src)
4720 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4721 && REG_P (dest)
4722 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4723 sets[i].src_volatile = 1;
4725 else if (GET_CODE (src) == ASM_OPERANDS
4726 && GET_CODE (x) == PARALLEL)
4728 /* Do not record result of a non-volatile inline asm with
4729 more than one result. */
4730 if (n_sets > 1)
4731 sets[i].src_volatile = 1;
4733 int j, lim = XVECLEN (x, 0);
4734 for (j = 0; j < lim; j++)
4736 rtx y = XVECEXP (x, 0, j);
4737 /* And do not record result of a non-volatile inline asm
4738 with "memory" clobber. */
4739 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4741 sets[i].src_volatile = 1;
4742 break;
4747 #if 0
4748 /* It is no longer clear why we used to do this, but it doesn't
4749 appear to still be needed. So let's try without it since this
4750 code hurts cse'ing widened ops. */
4751 /* If source is a paradoxical subreg (such as QI treated as an SI),
4752 treat it as volatile. It may do the work of an SI in one context
4753 where the extra bits are not being used, but cannot replace an SI
4754 in general. */
4755 if (paradoxical_subreg_p (src))
4756 sets[i].src_volatile = 1;
4757 #endif
4759 /* Locate all possible equivalent forms for SRC. Try to replace
4760 SRC in the insn with each cheaper equivalent.
4762 We have the following types of equivalents: SRC itself, a folded
4763 version, a value given in a REG_EQUAL note, or a value related
4764 to a constant.
4766 Each of these equivalents may be part of an additional class
4767 of equivalents (if more than one is in the table, they must be in
4768 the same class; we check for this).
4770 If the source is volatile, we don't do any table lookups.
4772 We note any constant equivalent for possible later use in a
4773 REG_NOTE. */
4775 if (!sets[i].src_volatile)
4776 elt = lookup (src, sets[i].src_hash, mode);
4778 sets[i].src_elt = elt;
4780 if (elt && src_eqv_here && src_eqv_elt)
4782 if (elt->first_same_value != src_eqv_elt->first_same_value)
4784 /* The REG_EQUAL is indicating that two formerly distinct
4785 classes are now equivalent. So merge them. */
4786 merge_equiv_classes (elt, src_eqv_elt);
4787 src_eqv_hash = HASH (src_eqv, elt->mode);
4788 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4791 src_eqv_here = 0;
4794 else if (src_eqv_elt)
4795 elt = src_eqv_elt;
4797 /* Try to find a constant somewhere and record it in `src_const'.
4798 Record its table element, if any, in `src_const_elt'. Look in
4799 any known equivalences first. (If the constant is not in the
4800 table, also set `sets[i].src_const_hash'). */
4801 if (elt)
4802 for (p = elt->first_same_value; p; p = p->next_same_value)
4803 if (p->is_const)
4805 src_const = p->exp;
4806 src_const_elt = elt;
4807 break;
4810 if (src_const == 0
4811 && (CONSTANT_P (src_folded)
4812 /* Consider (minus (label_ref L1) (label_ref L2)) as
4813 "constant" here so we will record it. This allows us
4814 to fold switch statements when an ADDR_DIFF_VEC is used. */
4815 || (GET_CODE (src_folded) == MINUS
4816 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4817 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4818 src_const = src_folded, src_const_elt = elt;
4819 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4820 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4822 /* If we don't know if the constant is in the table, get its
4823 hash code and look it up. */
4824 if (src_const && src_const_elt == 0)
4826 sets[i].src_const_hash = HASH (src_const, mode);
4827 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4830 sets[i].src_const = src_const;
4831 sets[i].src_const_elt = src_const_elt;
4833 /* If the constant and our source are both in the table, mark them as
4834 equivalent. Otherwise, if a constant is in the table but the source
4835 isn't, set ELT to it. */
4836 if (src_const_elt && elt
4837 && src_const_elt->first_same_value != elt->first_same_value)
4838 merge_equiv_classes (elt, src_const_elt);
4839 else if (src_const_elt && elt == 0)
4840 elt = src_const_elt;
4842 /* See if there is a register linearly related to a constant
4843 equivalent of SRC. */
4844 if (src_const
4845 && (GET_CODE (src_const) == CONST
4846 || (src_const_elt && src_const_elt->related_value != 0)))
4848 src_related = use_related_value (src_const, src_const_elt);
4849 if (src_related)
4851 struct table_elt *src_related_elt
4852 = lookup (src_related, HASH (src_related, mode), mode);
4853 if (src_related_elt && elt)
4855 if (elt->first_same_value
4856 != src_related_elt->first_same_value)
4857 /* This can occur when we previously saw a CONST
4858 involving a SYMBOL_REF and then see the SYMBOL_REF
4859 twice. Merge the involved classes. */
4860 merge_equiv_classes (elt, src_related_elt);
4862 src_related = 0;
4863 src_related_elt = 0;
4865 else if (src_related_elt && elt == 0)
4866 elt = src_related_elt;
4870 /* See if we have a CONST_INT that is already in a register in a
4871 wider mode. */
4873 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4874 && is_int_mode (mode, &int_mode)
4875 && GET_MODE_PRECISION (int_mode) < BITS_PER_WORD)
4877 opt_scalar_int_mode wider_mode_iter;
4878 FOR_EACH_WIDER_MODE (wider_mode_iter, int_mode)
4880 scalar_int_mode wider_mode = wider_mode_iter.require ();
4881 if (GET_MODE_PRECISION (wider_mode) > BITS_PER_WORD)
4882 break;
4884 struct table_elt *const_elt
4885 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4887 if (const_elt == 0)
4888 continue;
4890 for (const_elt = const_elt->first_same_value;
4891 const_elt; const_elt = const_elt->next_same_value)
4892 if (REG_P (const_elt->exp))
4894 src_related = gen_lowpart (int_mode, const_elt->exp);
4895 break;
4898 if (src_related != 0)
4899 break;
4903 /* Another possibility is that we have an AND with a constant in
4904 a mode narrower than a word. If so, it might have been generated
4905 as part of an "if" which would narrow the AND. If we already
4906 have done the AND in a wider mode, we can use a SUBREG of that
4907 value. */
4909 if (flag_expensive_optimizations && ! src_related
4910 && is_a <scalar_int_mode> (mode, &int_mode)
4911 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4912 && GET_MODE_SIZE (int_mode) < UNITS_PER_WORD)
4914 opt_scalar_int_mode tmode_iter;
4915 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4917 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4919 scalar_int_mode tmode = tmode_iter.require ();
4920 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4921 break;
4923 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4924 struct table_elt *larger_elt;
4926 if (inner)
4928 PUT_MODE (new_and, tmode);
4929 XEXP (new_and, 0) = inner;
4930 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4931 if (larger_elt == 0)
4932 continue;
4934 for (larger_elt = larger_elt->first_same_value;
4935 larger_elt; larger_elt = larger_elt->next_same_value)
4936 if (REG_P (larger_elt->exp))
4938 src_related
4939 = gen_lowpart (int_mode, larger_elt->exp);
4940 break;
4943 if (src_related)
4944 break;
4949 /* See if a MEM has already been loaded with a widening operation;
4950 if it has, we can use a subreg of that. Many CISC machines
4951 also have such operations, but this is only likely to be
4952 beneficial on these machines. */
4954 rtx_code extend_op;
4955 if (flag_expensive_optimizations && src_related == 0
4956 && MEM_P (src) && ! do_not_record
4957 && is_a <scalar_int_mode> (mode, &int_mode)
4958 && (extend_op = load_extend_op (int_mode)) != UNKNOWN)
4960 struct rtx_def memory_extend_buf;
4961 rtx memory_extend_rtx = &memory_extend_buf;
4963 /* Set what we are trying to extend and the operation it might
4964 have been extended with. */
4965 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4966 PUT_CODE (memory_extend_rtx, extend_op);
4967 XEXP (memory_extend_rtx, 0) = src;
4969 opt_scalar_int_mode tmode_iter;
4970 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4972 struct table_elt *larger_elt;
4974 scalar_int_mode tmode = tmode_iter.require ();
4975 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4976 break;
4978 PUT_MODE (memory_extend_rtx, tmode);
4979 larger_elt = lookup (memory_extend_rtx,
4980 HASH (memory_extend_rtx, tmode), tmode);
4981 if (larger_elt == 0)
4982 continue;
4984 for (larger_elt = larger_elt->first_same_value;
4985 larger_elt; larger_elt = larger_elt->next_same_value)
4986 if (REG_P (larger_elt->exp))
4988 src_related = gen_lowpart (int_mode, larger_elt->exp);
4989 break;
4992 if (src_related)
4993 break;
4997 /* Try to express the constant using a register+offset expression
4998 derived from a constant anchor. */
5000 if (targetm.const_anchor
5001 && !src_related
5002 && src_const
5003 && GET_CODE (src_const) == CONST_INT)
5005 src_related = try_const_anchors (src_const, mode);
5006 src_related_is_const_anchor = src_related != NULL_RTX;
5009 /* Try to re-materialize a vec_dup with an existing constant. */
5010 rtx src_elt;
5011 if ((!src_eqv_here || CONSTANT_P (src_eqv_here))
5012 && const_vec_duplicate_p (src, &src_elt))
5014 machine_mode const_mode = GET_MODE_INNER (GET_MODE (src));
5015 struct table_elt *related_elt
5016 = lookup (src_elt, HASH (src_elt, const_mode), const_mode);
5017 if (related_elt)
5019 for (related_elt = related_elt->first_same_value;
5020 related_elt; related_elt = related_elt->next_same_value)
5021 if (REG_P (related_elt->exp))
5023 /* We don't need to compare costs with an existing (constant)
5024 src_eqv_here, since any such src_eqv_here should already be
5025 available in src_const. */
5026 src_eqv_here
5027 = gen_rtx_VEC_DUPLICATE (GET_MODE (src),
5028 related_elt->exp);
5029 break;
5034 if (src == src_folded)
5035 src_folded = 0;
5037 /* At this point, ELT, if nonzero, points to a class of expressions
5038 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5039 and SRC_RELATED, if nonzero, each contain additional equivalent
5040 expressions. Prune these latter expressions by deleting expressions
5041 already in the equivalence class.
5043 Check for an equivalent identical to the destination. If found,
5044 this is the preferred equivalent since it will likely lead to
5045 elimination of the insn. Indicate this by placing it in
5046 `src_related'. */
5048 if (elt)
5049 elt = elt->first_same_value;
5050 for (p = elt; p; p = p->next_same_value)
5052 enum rtx_code code = GET_CODE (p->exp);
5054 /* If the expression is not valid, ignore it. Then we do not
5055 have to check for validity below. In most cases, we can use
5056 `rtx_equal_p', since canonicalization has already been done. */
5057 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5058 continue;
5060 /* Also skip paradoxical subregs, unless that's what we're
5061 looking for. */
5062 if (paradoxical_subreg_p (p->exp)
5063 && ! (src != 0
5064 && GET_CODE (src) == SUBREG
5065 && GET_MODE (src) == GET_MODE (p->exp)
5066 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5067 GET_MODE (SUBREG_REG (p->exp)))))
5068 continue;
5070 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5071 src = 0;
5072 else if (src_folded && GET_CODE (src_folded) == code
5073 && rtx_equal_p (src_folded, p->exp))
5074 src_folded = 0;
5075 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5076 && rtx_equal_p (src_eqv_here, p->exp))
5077 src_eqv_here = 0;
5078 else if (src_related && GET_CODE (src_related) == code
5079 && rtx_equal_p (src_related, p->exp))
5080 src_related = 0;
5082 /* This is the same as the destination of the insns, we want
5083 to prefer it. Copy it to src_related. The code below will
5084 then give it a negative cost. */
5085 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5086 src_related = p->exp;
5089 /* Find the cheapest valid equivalent, trying all the available
5090 possibilities. Prefer items not in the hash table to ones
5091 that are when they are equal cost. Note that we can never
5092 worsen an insn as the current contents will also succeed.
5093 If we find an equivalent identical to the destination, use it as best,
5094 since this insn will probably be eliminated in that case. */
5095 if (src)
5097 if (rtx_equal_p (src, dest))
5098 src_cost = src_regcost = -1;
5099 else
5101 src_cost = COST (src, mode);
5102 src_regcost = approx_reg_cost (src);
5106 if (src_eqv_here)
5108 if (rtx_equal_p (src_eqv_here, dest))
5109 src_eqv_cost = src_eqv_regcost = -1;
5110 else
5112 src_eqv_cost = COST (src_eqv_here, mode);
5113 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5117 if (src_folded)
5119 if (rtx_equal_p (src_folded, dest))
5120 src_folded_cost = src_folded_regcost = -1;
5121 else
5123 src_folded_cost = COST (src_folded, mode);
5124 src_folded_regcost = approx_reg_cost (src_folded);
5128 if (src_related)
5130 if (rtx_equal_p (src_related, dest))
5131 src_related_cost = src_related_regcost = -1;
5132 else
5134 src_related_cost = COST (src_related, mode);
5135 src_related_regcost = approx_reg_cost (src_related);
5137 /* If a const-anchor is used to synthesize a constant that
5138 normally requires multiple instructions then slightly prefer
5139 it over the original sequence. These instructions are likely
5140 to become redundant now. We can't compare against the cost
5141 of src_eqv_here because, on MIPS for example, multi-insn
5142 constants have zero cost; they are assumed to be hoisted from
5143 loops. */
5144 if (src_related_is_const_anchor
5145 && src_related_cost == src_cost
5146 && src_eqv_here)
5147 src_related_cost--;
5151 /* If this was an indirect jump insn, a known label will really be
5152 cheaper even though it looks more expensive. */
5153 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5154 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5156 /* Terminate loop when replacement made. This must terminate since
5157 the current contents will be tested and will always be valid. */
5158 while (1)
5160 rtx trial;
5162 /* Skip invalid entries. */
5163 while (elt && !REG_P (elt->exp)
5164 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5165 elt = elt->next_same_value;
5167 /* A paradoxical subreg would be bad here: it'll be the right
5168 size, but later may be adjusted so that the upper bits aren't
5169 what we want. So reject it. */
5170 if (elt != 0
5171 && paradoxical_subreg_p (elt->exp)
5172 /* It is okay, though, if the rtx we're trying to match
5173 will ignore any of the bits we can't predict. */
5174 && ! (src != 0
5175 && GET_CODE (src) == SUBREG
5176 && GET_MODE (src) == GET_MODE (elt->exp)
5177 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5178 GET_MODE (SUBREG_REG (elt->exp)))))
5180 elt = elt->next_same_value;
5181 continue;
5184 if (elt)
5186 src_elt_cost = elt->cost;
5187 src_elt_regcost = elt->regcost;
5190 /* Find cheapest and skip it for the next time. For items
5191 of equal cost, use this order:
5192 src_folded, src, src_eqv, src_related and hash table entry. */
5193 if (src_folded
5194 && preferable (src_folded_cost, src_folded_regcost,
5195 src_cost, src_regcost) <= 0
5196 && preferable (src_folded_cost, src_folded_regcost,
5197 src_eqv_cost, src_eqv_regcost) <= 0
5198 && preferable (src_folded_cost, src_folded_regcost,
5199 src_related_cost, src_related_regcost) <= 0
5200 && preferable (src_folded_cost, src_folded_regcost,
5201 src_elt_cost, src_elt_regcost) <= 0)
5202 trial = src_folded, src_folded_cost = MAX_COST;
5203 else if (src
5204 && preferable (src_cost, src_regcost,
5205 src_eqv_cost, src_eqv_regcost) <= 0
5206 && preferable (src_cost, src_regcost,
5207 src_related_cost, src_related_regcost) <= 0
5208 && preferable (src_cost, src_regcost,
5209 src_elt_cost, src_elt_regcost) <= 0)
5210 trial = src, src_cost = MAX_COST;
5211 else if (src_eqv_here
5212 && preferable (src_eqv_cost, src_eqv_regcost,
5213 src_related_cost, src_related_regcost) <= 0
5214 && preferable (src_eqv_cost, src_eqv_regcost,
5215 src_elt_cost, src_elt_regcost) <= 0)
5216 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5217 else if (src_related
5218 && preferable (src_related_cost, src_related_regcost,
5219 src_elt_cost, src_elt_regcost) <= 0)
5220 trial = src_related, src_related_cost = MAX_COST;
5221 else
5223 trial = elt->exp;
5224 elt = elt->next_same_value;
5225 src_elt_cost = MAX_COST;
5228 /* Try to optimize
5229 (set (reg:M N) (const_int A))
5230 (set (reg:M2 O) (const_int B))
5231 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5232 (reg:M2 O)). */
5233 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5234 && CONST_INT_P (trial)
5235 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5236 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5237 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5238 && (known_ge
5239 (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl))),
5240 INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))))
5241 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5242 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5243 <= HOST_BITS_PER_WIDE_INT))
5245 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5246 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5247 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5248 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5249 struct table_elt *dest_elt
5250 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5251 rtx dest_cst = NULL;
5253 if (dest_elt)
5254 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5255 if (p->is_const && CONST_INT_P (p->exp))
5257 dest_cst = p->exp;
5258 break;
5260 if (dest_cst)
5262 HOST_WIDE_INT val = INTVAL (dest_cst);
5263 HOST_WIDE_INT mask;
5264 unsigned int shift;
5265 /* This is the mode of DEST_CST as well. */
5266 scalar_int_mode dest_mode
5267 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
5268 if (BITS_BIG_ENDIAN)
5269 shift = GET_MODE_PRECISION (dest_mode)
5270 - INTVAL (pos) - INTVAL (width);
5271 else
5272 shift = INTVAL (pos);
5273 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5274 mask = HOST_WIDE_INT_M1;
5275 else
5276 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
5277 val &= ~(mask << shift);
5278 val |= (INTVAL (trial) & mask) << shift;
5279 val = trunc_int_for_mode (val, dest_mode);
5280 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5281 dest_reg, 1);
5282 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5283 GEN_INT (val), 1);
5284 if (apply_change_group ())
5286 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5287 if (note)
5289 remove_note (insn, note);
5290 df_notes_rescan (insn);
5292 src_eqv = NULL_RTX;
5293 src_eqv_elt = NULL;
5294 src_eqv_volatile = 0;
5295 src_eqv_in_memory = 0;
5296 src_eqv_hash = 0;
5297 repeat = true;
5298 break;
5303 /* We don't normally have an insn matching (set (pc) (pc)), so
5304 check for this separately here. We will delete such an
5305 insn below.
5307 For other cases such as a table jump or conditional jump
5308 where we know the ultimate target, go ahead and replace the
5309 operand. While that may not make a valid insn, we will
5310 reemit the jump below (and also insert any necessary
5311 barriers). */
5312 if (n_sets == 1 && dest == pc_rtx
5313 && (trial == pc_rtx
5314 || (GET_CODE (trial) == LABEL_REF
5315 && ! condjump_p (insn))))
5317 /* Don't substitute non-local labels, this confuses CFG. */
5318 if (GET_CODE (trial) == LABEL_REF
5319 && LABEL_REF_NONLOCAL_P (trial))
5320 continue;
5322 SET_SRC (sets[i].rtl) = trial;
5323 cse_jumps_altered = true;
5324 break;
5327 /* Similarly, lots of targets don't allow no-op
5328 (set (mem x) (mem x)) moves. Even (set (reg x) (reg x))
5329 might be impossible for certain registers (like CC registers). */
5330 else if (n_sets == 1
5331 && !CALL_P (insn)
5332 && (MEM_P (trial) || REG_P (trial))
5333 && rtx_equal_p (trial, dest)
5334 && !side_effects_p (dest)
5335 && (cfun->can_delete_dead_exceptions
5336 || insn_nothrow_p (insn))
5337 /* We can only remove the later store if the earlier aliases
5338 at least all accesses the later one. */
5339 && (!MEM_P (trial)
5340 || ((MEM_ALIAS_SET (dest) == MEM_ALIAS_SET (trial)
5341 || alias_set_subset_of (MEM_ALIAS_SET (dest),
5342 MEM_ALIAS_SET (trial)))
5343 && (!MEM_EXPR (trial)
5344 || refs_same_for_tbaa_p (MEM_EXPR (trial),
5345 MEM_EXPR (dest))))))
5347 SET_SRC (sets[i].rtl) = trial;
5348 noop_insn = true;
5349 break;
5352 /* Reject certain invalid forms of CONST that we create. */
5353 else if (CONSTANT_P (trial)
5354 && GET_CODE (trial) == CONST
5355 /* Reject cases that will cause decode_rtx_const to
5356 die. On the alpha when simplifying a switch, we
5357 get (const (truncate (minus (label_ref)
5358 (label_ref)))). */
5359 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5360 /* Likewise on IA-64, except without the
5361 truncate. */
5362 || (GET_CODE (XEXP (trial, 0)) == MINUS
5363 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5364 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5365 /* Do nothing for this case. */
5368 /* Do not replace anything with a MEM, except the replacement
5369 is a no-op. This allows this loop to terminate. */
5370 else if (MEM_P (trial) && !rtx_equal_p (trial, SET_SRC(sets[i].rtl)))
5371 /* Do nothing for this case. */
5374 /* Look for a substitution that makes a valid insn. */
5375 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5376 trial, 0))
5378 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5380 /* The result of apply_change_group can be ignored; see
5381 canon_reg. */
5383 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5384 apply_change_group ();
5386 break;
5389 /* If the current function uses a constant pool and this is a
5390 constant, try making a pool entry. Put it in src_folded
5391 unless we already have done this since that is where it
5392 likely came from. */
5394 else if (crtl->uses_const_pool
5395 && CONSTANT_P (trial)
5396 && !CONST_INT_P (trial)
5397 && (src_folded == 0 || !MEM_P (src_folded))
5398 && GET_MODE_CLASS (mode) != MODE_CC
5399 && mode != VOIDmode)
5401 src_folded = force_const_mem (mode, trial);
5402 if (src_folded)
5404 src_folded_cost = COST (src_folded, mode);
5405 src_folded_regcost = approx_reg_cost (src_folded);
5410 /* If we changed the insn too much, handle this set from scratch. */
5411 if (repeat)
5413 i--;
5414 continue;
5417 src = SET_SRC (sets[i].rtl);
5419 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5420 However, there is an important exception: If both are registers
5421 that are not the head of their equivalence class, replace SET_SRC
5422 with the head of the class. If we do not do this, we will have
5423 both registers live over a portion of the basic block. This way,
5424 their lifetimes will likely abut instead of overlapping. */
5425 if (REG_P (dest)
5426 && REGNO_QTY_VALID_P (REGNO (dest)))
5428 int dest_q = REG_QTY (REGNO (dest));
5429 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5431 if (dest_ent->mode == GET_MODE (dest)
5432 && dest_ent->first_reg != REGNO (dest)
5433 && REG_P (src) && REGNO (src) == REGNO (dest)
5434 /* Don't do this if the original insn had a hard reg as
5435 SET_SRC or SET_DEST. */
5436 && (!REG_P (sets[i].src)
5437 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5438 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5439 /* We can't call canon_reg here because it won't do anything if
5440 SRC is a hard register. */
5442 int src_q = REG_QTY (REGNO (src));
5443 struct qty_table_elem *src_ent = &qty_table[src_q];
5444 int first = src_ent->first_reg;
5445 rtx new_src
5446 = (first >= FIRST_PSEUDO_REGISTER
5447 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5449 /* We must use validate-change even for this, because this
5450 might be a special no-op instruction, suitable only to
5451 tag notes onto. */
5452 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5454 src = new_src;
5455 /* If we had a constant that is cheaper than what we are now
5456 setting SRC to, use that constant. We ignored it when we
5457 thought we could make this into a no-op. */
5458 if (src_const && COST (src_const, mode) < COST (src, mode)
5459 && validate_change (insn, &SET_SRC (sets[i].rtl),
5460 src_const, 0))
5461 src = src_const;
5466 /* If we made a change, recompute SRC values. */
5467 if (src != sets[i].src)
5469 do_not_record = 0;
5470 hash_arg_in_memory = 0;
5471 sets[i].src = src;
5472 sets[i].src_hash = HASH (src, mode);
5473 sets[i].src_volatile = do_not_record;
5474 sets[i].src_in_memory = hash_arg_in_memory;
5475 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5478 /* If this is a single SET, we are setting a register, and we have an
5479 equivalent constant, we want to add a REG_EQUAL note if the constant
5480 is different from the source. We don't want to do it for a constant
5481 pseudo since verifying that this pseudo hasn't been eliminated is a
5482 pain; moreover such a note won't help anything.
5484 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5485 which can be created for a reference to a compile time computable
5486 entry in a jump table. */
5487 if (n_sets == 1
5488 && REG_P (dest)
5489 && src_const
5490 && !REG_P (src_const)
5491 && !(GET_CODE (src_const) == SUBREG
5492 && REG_P (SUBREG_REG (src_const)))
5493 && !(GET_CODE (src_const) == CONST
5494 && GET_CODE (XEXP (src_const, 0)) == MINUS
5495 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5496 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5497 && !rtx_equal_p (src, src_const))
5499 /* Make sure that the rtx is not shared. */
5500 src_const = copy_rtx (src_const);
5502 /* Record the actual constant value in a REG_EQUAL note,
5503 making a new one if one does not already exist. */
5504 set_unique_reg_note (insn, REG_EQUAL, src_const);
5505 df_notes_rescan (insn);
5508 /* Now deal with the destination. */
5509 do_not_record = 0;
5511 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5512 while (GET_CODE (dest) == SUBREG
5513 || GET_CODE (dest) == ZERO_EXTRACT
5514 || GET_CODE (dest) == STRICT_LOW_PART)
5515 dest = XEXP (dest, 0);
5517 sets[i].inner_dest = dest;
5519 if (MEM_P (dest))
5521 #ifdef PUSH_ROUNDING
5522 /* Stack pushes invalidate the stack pointer. */
5523 rtx addr = XEXP (dest, 0);
5524 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5525 && XEXP (addr, 0) == stack_pointer_rtx)
5526 invalidate (stack_pointer_rtx, VOIDmode);
5527 #endif
5528 dest = fold_rtx (dest, insn);
5531 /* Compute the hash code of the destination now,
5532 before the effects of this instruction are recorded,
5533 since the register values used in the address computation
5534 are those before this instruction. */
5535 sets[i].dest_hash = HASH (dest, mode);
5537 /* Don't enter a bit-field in the hash table
5538 because the value in it after the store
5539 may not equal what was stored, due to truncation. */
5541 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5543 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5545 if (src_const != 0 && CONST_INT_P (src_const)
5546 && CONST_INT_P (width)
5547 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5548 && ! (INTVAL (src_const)
5549 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5550 /* Exception: if the value is constant,
5551 and it won't be truncated, record it. */
5553 else
5555 /* This is chosen so that the destination will be invalidated
5556 but no new value will be recorded.
5557 We must invalidate because sometimes constant
5558 values can be recorded for bitfields. */
5559 sets[i].src_elt = 0;
5560 sets[i].src_volatile = 1;
5561 src_eqv = 0;
5562 src_eqv_elt = 0;
5566 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5567 the insn. */
5568 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5570 /* One less use of the label this insn used to jump to. */
5571 cse_cfg_altered |= delete_insn_and_edges (insn);
5572 cse_jumps_altered = true;
5573 /* No more processing for this set. */
5574 sets[i].rtl = 0;
5577 /* Similarly for no-op moves. */
5578 else if (noop_insn)
5580 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5581 cse_cfg_altered = true;
5582 cse_cfg_altered |= delete_insn_and_edges (insn);
5583 /* No more processing for this set. */
5584 sets[i].rtl = 0;
5587 /* If this SET is now setting PC to a label, we know it used to
5588 be a conditional or computed branch. */
5589 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5590 && !LABEL_REF_NONLOCAL_P (src))
5592 /* We reemit the jump in as many cases as possible just in
5593 case the form of an unconditional jump is significantly
5594 different than a computed jump or conditional jump.
5596 If this insn has multiple sets, then reemitting the
5597 jump is nontrivial. So instead we just force rerecognition
5598 and hope for the best. */
5599 if (n_sets == 1)
5601 rtx_jump_insn *new_rtx;
5602 rtx note;
5604 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5605 new_rtx = emit_jump_insn_before (seq, insn);
5606 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5607 LABEL_NUSES (XEXP (src, 0))++;
5609 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5610 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5611 if (note)
5613 XEXP (note, 1) = NULL_RTX;
5614 REG_NOTES (new_rtx) = note;
5617 cse_cfg_altered |= delete_insn_and_edges (insn);
5618 insn = new_rtx;
5620 else
5621 INSN_CODE (insn) = -1;
5623 /* Do not bother deleting any unreachable code, let jump do it. */
5624 cse_jumps_altered = true;
5625 sets[i].rtl = 0;
5628 /* If destination is volatile, invalidate it and then do no further
5629 processing for this assignment. */
5631 else if (do_not_record)
5633 invalidate_dest (dest);
5634 sets[i].rtl = 0;
5637 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5639 do_not_record = 0;
5640 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5641 if (do_not_record)
5643 invalidate_dest (SET_DEST (sets[i].rtl));
5644 sets[i].rtl = 0;
5649 /* Now enter all non-volatile source expressions in the hash table
5650 if they are not already present.
5651 Record their equivalence classes in src_elt.
5652 This way we can insert the corresponding destinations into
5653 the same classes even if the actual sources are no longer in them
5654 (having been invalidated). */
5656 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5657 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5659 struct table_elt *elt;
5660 struct table_elt *classp = sets[0].src_elt;
5661 rtx dest = SET_DEST (sets[0].rtl);
5662 machine_mode eqvmode = GET_MODE (dest);
5664 if (GET_CODE (dest) == STRICT_LOW_PART)
5666 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5667 classp = 0;
5669 if (insert_regs (src_eqv, classp, 0))
5671 rehash_using_reg (src_eqv);
5672 src_eqv_hash = HASH (src_eqv, eqvmode);
5674 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5675 elt->in_memory = src_eqv_in_memory;
5676 src_eqv_elt = elt;
5678 /* Check to see if src_eqv_elt is the same as a set source which
5679 does not yet have an elt, and if so set the elt of the set source
5680 to src_eqv_elt. */
5681 for (i = 0; i < n_sets; i++)
5682 if (sets[i].rtl && sets[i].src_elt == 0
5683 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5684 sets[i].src_elt = src_eqv_elt;
5687 for (i = 0; i < n_sets; i++)
5688 if (sets[i].rtl && ! sets[i].src_volatile
5689 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5691 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5693 /* REG_EQUAL in setting a STRICT_LOW_PART
5694 gives an equivalent for the entire destination register,
5695 not just for the subreg being stored in now.
5696 This is a more interesting equivalence, so we arrange later
5697 to treat the entire reg as the destination. */
5698 sets[i].src_elt = src_eqv_elt;
5699 sets[i].src_hash = src_eqv_hash;
5701 else
5703 /* Insert source and constant equivalent into hash table, if not
5704 already present. */
5705 struct table_elt *classp = src_eqv_elt;
5706 rtx src = sets[i].src;
5707 rtx dest = SET_DEST (sets[i].rtl);
5708 machine_mode mode
5709 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5711 /* It's possible that we have a source value known to be
5712 constant but don't have a REG_EQUAL note on the insn.
5713 Lack of a note will mean src_eqv_elt will be NULL. This
5714 can happen where we've generated a SUBREG to access a
5715 CONST_INT that is already in a register in a wider mode.
5716 Ensure that the source expression is put in the proper
5717 constant class. */
5718 if (!classp)
5719 classp = sets[i].src_const_elt;
5721 if (sets[i].src_elt == 0)
5723 struct table_elt *elt;
5725 /* Note that these insert_regs calls cannot remove
5726 any of the src_elt's, because they would have failed to
5727 match if not still valid. */
5728 if (insert_regs (src, classp, 0))
5730 rehash_using_reg (src);
5731 sets[i].src_hash = HASH (src, mode);
5733 elt = insert (src, classp, sets[i].src_hash, mode);
5734 elt->in_memory = sets[i].src_in_memory;
5735 /* If inline asm has any clobbers, ensure we only reuse
5736 existing inline asms and never try to put the ASM_OPERANDS
5737 into an insn that isn't inline asm. */
5738 if (GET_CODE (src) == ASM_OPERANDS
5739 && GET_CODE (x) == PARALLEL)
5740 elt->cost = MAX_COST;
5741 sets[i].src_elt = classp = elt;
5743 if (sets[i].src_const && sets[i].src_const_elt == 0
5744 && src != sets[i].src_const
5745 && ! rtx_equal_p (sets[i].src_const, src))
5746 sets[i].src_elt = insert (sets[i].src_const, classp,
5747 sets[i].src_const_hash, mode);
5750 else if (sets[i].src_elt == 0)
5751 /* If we did not insert the source into the hash table (e.g., it was
5752 volatile), note the equivalence class for the REG_EQUAL value, if any,
5753 so that the destination goes into that class. */
5754 sets[i].src_elt = src_eqv_elt;
5756 /* Record destination addresses in the hash table. This allows us to
5757 check if they are invalidated by other sets. */
5758 for (i = 0; i < n_sets; i++)
5760 if (sets[i].rtl)
5762 rtx x = sets[i].inner_dest;
5763 struct table_elt *elt;
5764 machine_mode mode;
5765 unsigned hash;
5767 if (MEM_P (x))
5769 x = XEXP (x, 0);
5770 mode = GET_MODE (x);
5771 hash = HASH (x, mode);
5772 elt = lookup (x, hash, mode);
5773 if (!elt)
5775 if (insert_regs (x, NULL, 0))
5777 rtx dest = SET_DEST (sets[i].rtl);
5779 rehash_using_reg (x);
5780 hash = HASH (x, mode);
5781 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5783 elt = insert (x, NULL, hash, mode);
5786 sets[i].dest_addr_elt = elt;
5788 else
5789 sets[i].dest_addr_elt = NULL;
5793 invalidate_from_clobbers (insn);
5795 /* Some registers are invalidated by subroutine calls. Memory is
5796 invalidated by non-constant calls. */
5798 if (CALL_P (insn))
5800 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5801 invalidate_memory ();
5802 else
5803 /* For const/pure calls, invalidate any argument slots, because
5804 those are owned by the callee. */
5805 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
5806 if (GET_CODE (XEXP (tem, 0)) == USE
5807 && MEM_P (XEXP (XEXP (tem, 0), 0)))
5808 invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode);
5809 invalidate_for_call (insn);
5812 /* Now invalidate everything set by this instruction.
5813 If a SUBREG or other funny destination is being set,
5814 sets[i].rtl is still nonzero, so here we invalidate the reg
5815 a part of which is being set. */
5817 for (i = 0; i < n_sets; i++)
5818 if (sets[i].rtl)
5820 /* We can't use the inner dest, because the mode associated with
5821 a ZERO_EXTRACT is significant. */
5822 rtx dest = SET_DEST (sets[i].rtl);
5824 /* Needed for registers to remove the register from its
5825 previous quantity's chain.
5826 Needed for memory if this is a nonvarying address, unless
5827 we have just done an invalidate_memory that covers even those. */
5828 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5829 invalidate (dest, VOIDmode);
5830 else if (MEM_P (dest))
5831 invalidate (dest, VOIDmode);
5832 else if (GET_CODE (dest) == STRICT_LOW_PART
5833 || GET_CODE (dest) == ZERO_EXTRACT)
5834 invalidate (XEXP (dest, 0), GET_MODE (dest));
5837 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5838 the regs restored by the longjmp come from a later time
5839 than the setjmp. */
5840 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5842 flush_hash_table ();
5843 goto done;
5846 /* Make sure registers mentioned in destinations
5847 are safe for use in an expression to be inserted.
5848 This removes from the hash table
5849 any invalid entry that refers to one of these registers.
5851 We don't care about the return value from mention_regs because
5852 we are going to hash the SET_DEST values unconditionally. */
5854 for (i = 0; i < n_sets; i++)
5856 if (sets[i].rtl)
5858 rtx x = SET_DEST (sets[i].rtl);
5860 if (!REG_P (x))
5861 mention_regs (x);
5862 else
5864 /* We used to rely on all references to a register becoming
5865 inaccessible when a register changes to a new quantity,
5866 since that changes the hash code. However, that is not
5867 safe, since after HASH_SIZE new quantities we get a
5868 hash 'collision' of a register with its own invalid
5869 entries. And since SUBREGs have been changed not to
5870 change their hash code with the hash code of the register,
5871 it wouldn't work any longer at all. So we have to check
5872 for any invalid references lying around now.
5873 This code is similar to the REG case in mention_regs,
5874 but it knows that reg_tick has been incremented, and
5875 it leaves reg_in_table as -1 . */
5876 unsigned int regno = REGNO (x);
5877 unsigned int endregno = END_REGNO (x);
5878 unsigned int i;
5880 for (i = regno; i < endregno; i++)
5882 if (REG_IN_TABLE (i) >= 0)
5884 remove_invalid_refs (i);
5885 REG_IN_TABLE (i) = -1;
5892 /* We may have just removed some of the src_elt's from the hash table.
5893 So replace each one with the current head of the same class.
5894 Also check if destination addresses have been removed. */
5896 for (i = 0; i < n_sets; i++)
5897 if (sets[i].rtl)
5899 if (sets[i].dest_addr_elt
5900 && sets[i].dest_addr_elt->first_same_value == 0)
5902 /* The elt was removed, which means this destination is not
5903 valid after this instruction. */
5904 sets[i].rtl = NULL_RTX;
5906 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5907 /* If elt was removed, find current head of same class,
5908 or 0 if nothing remains of that class. */
5910 struct table_elt *elt = sets[i].src_elt;
5912 while (elt && elt->prev_same_value)
5913 elt = elt->prev_same_value;
5915 while (elt && elt->first_same_value == 0)
5916 elt = elt->next_same_value;
5917 sets[i].src_elt = elt ? elt->first_same_value : 0;
5921 /* Now insert the destinations into their equivalence classes. */
5923 for (i = 0; i < n_sets; i++)
5924 if (sets[i].rtl)
5926 rtx dest = SET_DEST (sets[i].rtl);
5927 struct table_elt *elt;
5929 /* Don't record value if we are not supposed to risk allocating
5930 floating-point values in registers that might be wider than
5931 memory. */
5932 if ((flag_float_store
5933 && MEM_P (dest)
5934 && FLOAT_MODE_P (GET_MODE (dest)))
5935 /* Don't record BLKmode values, because we don't know the
5936 size of it, and can't be sure that other BLKmode values
5937 have the same or smaller size. */
5938 || GET_MODE (dest) == BLKmode
5939 /* If we didn't put a REG_EQUAL value or a source into the hash
5940 table, there is no point is recording DEST. */
5941 || sets[i].src_elt == 0)
5942 continue;
5944 /* STRICT_LOW_PART isn't part of the value BEING set,
5945 and neither is the SUBREG inside it.
5946 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5947 if (GET_CODE (dest) == STRICT_LOW_PART)
5948 dest = SUBREG_REG (XEXP (dest, 0));
5950 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5951 /* Registers must also be inserted into chains for quantities. */
5952 if (insert_regs (dest, sets[i].src_elt, 1))
5954 /* If `insert_regs' changes something, the hash code must be
5955 recalculated. */
5956 rehash_using_reg (dest);
5957 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5960 /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
5961 outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
5962 if (paradoxical_subreg_p (dest))
5963 continue;
5965 elt = insert (dest, sets[i].src_elt,
5966 sets[i].dest_hash, GET_MODE (dest));
5968 /* If this is a constant, insert the constant anchors with the
5969 equivalent register-offset expressions using register DEST. */
5970 if (targetm.const_anchor
5971 && REG_P (dest)
5972 && SCALAR_INT_MODE_P (GET_MODE (dest))
5973 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5974 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5976 elt->in_memory = (MEM_P (sets[i].inner_dest)
5977 && !MEM_READONLY_P (sets[i].inner_dest));
5979 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5980 narrower than M2, and both M1 and M2 are the same number of words,
5981 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5982 make that equivalence as well.
5984 However, BAR may have equivalences for which gen_lowpart
5985 will produce a simpler value than gen_lowpart applied to
5986 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5987 BAR's equivalences. If we don't get a simplified form, make
5988 the SUBREG. It will not be used in an equivalence, but will
5989 cause two similar assignments to be detected.
5991 Note the loop below will find SUBREG_REG (DEST) since we have
5992 already entered SRC and DEST of the SET in the table. */
5994 if (GET_CODE (dest) == SUBREG
5995 && (known_equal_after_align_down
5996 (GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1,
5997 GET_MODE_SIZE (GET_MODE (dest)) - 1,
5998 UNITS_PER_WORD))
5999 && !partial_subreg_p (dest)
6000 && sets[i].src_elt != 0)
6002 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6003 struct table_elt *elt, *classp = 0;
6005 for (elt = sets[i].src_elt->first_same_value; elt;
6006 elt = elt->next_same_value)
6008 rtx new_src = 0;
6009 unsigned src_hash;
6010 struct table_elt *src_elt;
6012 /* Ignore invalid entries. */
6013 if (!REG_P (elt->exp)
6014 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6015 continue;
6017 /* We may have already been playing subreg games. If the
6018 mode is already correct for the destination, use it. */
6019 if (GET_MODE (elt->exp) == new_mode)
6020 new_src = elt->exp;
6021 else
6023 poly_uint64 byte
6024 = subreg_lowpart_offset (new_mode, GET_MODE (dest));
6025 new_src = simplify_gen_subreg (new_mode, elt->exp,
6026 GET_MODE (dest), byte);
6029 /* The call to simplify_gen_subreg fails if the value
6030 is VOIDmode, yet we can't do any simplification, e.g.
6031 for EXPR_LISTs denoting function call results.
6032 It is invalid to construct a SUBREG with a VOIDmode
6033 SUBREG_REG, hence a zero new_src means we can't do
6034 this substitution. */
6035 if (! new_src)
6036 continue;
6038 src_hash = HASH (new_src, new_mode);
6039 src_elt = lookup (new_src, src_hash, new_mode);
6041 /* Put the new source in the hash table is if isn't
6042 already. */
6043 if (src_elt == 0)
6045 if (insert_regs (new_src, classp, 0))
6047 rehash_using_reg (new_src);
6048 src_hash = HASH (new_src, new_mode);
6050 src_elt = insert (new_src, classp, src_hash, new_mode);
6051 src_elt->in_memory = elt->in_memory;
6052 if (GET_CODE (new_src) == ASM_OPERANDS
6053 && elt->cost == MAX_COST)
6054 src_elt->cost = MAX_COST;
6056 else if (classp && classp != src_elt->first_same_value)
6057 /* Show that two things that we've seen before are
6058 actually the same. */
6059 merge_equiv_classes (src_elt, classp);
6061 classp = src_elt->first_same_value;
6062 /* Ignore invalid entries. */
6063 while (classp
6064 && !REG_P (classp->exp)
6065 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6066 classp = classp->next_same_value;
6071 /* Special handling for (set REG0 REG1) where REG0 is the
6072 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6073 be used in the sequel, so (if easily done) change this insn to
6074 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6075 that computed their value. Then REG1 will become a dead store
6076 and won't cloud the situation for later optimizations.
6078 Do not make this change if REG1 is a hard register, because it will
6079 then be used in the sequel and we may be changing a two-operand insn
6080 into a three-operand insn.
6082 Also do not do this if we are operating on a copy of INSN. */
6084 if (n_sets == 1 && sets[0].rtl)
6085 try_back_substitute_reg (sets[0].rtl, insn);
6087 done:;
6090 /* Remove from the hash table all expressions that reference memory. */
6092 static void
6093 invalidate_memory (void)
6095 int i;
6096 struct table_elt *p, *next;
6098 for (i = 0; i < HASH_SIZE; i++)
6099 for (p = table[i]; p; p = next)
6101 next = p->next_same_hash;
6102 if (p->in_memory)
6103 remove_from_table (p, i);
6107 /* Perform invalidation on the basis of everything about INSN,
6108 except for invalidating the actual places that are SET in it.
6109 This includes the places CLOBBERed, and anything that might
6110 alias with something that is SET or CLOBBERed. */
6112 static void
6113 invalidate_from_clobbers (rtx_insn *insn)
6115 rtx x = PATTERN (insn);
6117 if (GET_CODE (x) == CLOBBER)
6119 rtx ref = XEXP (x, 0);
6120 if (ref)
6122 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6123 || MEM_P (ref))
6124 invalidate (ref, VOIDmode);
6125 else if (GET_CODE (ref) == STRICT_LOW_PART
6126 || GET_CODE (ref) == ZERO_EXTRACT)
6127 invalidate (XEXP (ref, 0), GET_MODE (ref));
6130 else if (GET_CODE (x) == PARALLEL)
6132 int i;
6133 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6135 rtx y = XVECEXP (x, 0, i);
6136 if (GET_CODE (y) == CLOBBER)
6138 rtx ref = XEXP (y, 0);
6139 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6140 || MEM_P (ref))
6141 invalidate (ref, VOIDmode);
6142 else if (GET_CODE (ref) == STRICT_LOW_PART
6143 || GET_CODE (ref) == ZERO_EXTRACT)
6144 invalidate (XEXP (ref, 0), GET_MODE (ref));
6150 /* Perform invalidation on the basis of everything about INSN.
6151 This includes the places CLOBBERed, and anything that might
6152 alias with something that is SET or CLOBBERed. */
6154 static void
6155 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6157 rtx tem;
6158 rtx x = PATTERN (insn);
6160 if (CALL_P (insn))
6162 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6164 rtx temx = XEXP (tem, 0);
6165 if (GET_CODE (temx) == CLOBBER)
6166 invalidate (SET_DEST (temx), VOIDmode);
6170 /* Ensure we invalidate the destination register of a CALL insn.
6171 This is necessary for machines where this register is a fixed_reg,
6172 because no other code would invalidate it. */
6173 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6174 invalidate (SET_DEST (x), VOIDmode);
6176 else if (GET_CODE (x) == PARALLEL)
6178 int i;
6180 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6182 rtx y = XVECEXP (x, 0, i);
6183 if (GET_CODE (y) == CLOBBER)
6185 rtx clobbered = XEXP (y, 0);
6187 if (REG_P (clobbered)
6188 || GET_CODE (clobbered) == SUBREG)
6189 invalidate (clobbered, VOIDmode);
6190 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6191 || GET_CODE (clobbered) == ZERO_EXTRACT)
6192 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6194 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6195 invalidate (SET_DEST (y), VOIDmode);
6200 static rtx cse_process_note (rtx);
6202 /* A simplify_replace_fn_rtx callback for cse_process_note. Process X,
6203 part of the REG_NOTES of an insn. Replace any registers with either
6204 an equivalent constant or the canonical form of the register.
6205 Only replace addresses if the containing MEM remains valid.
6207 Return the replacement for X, or null if it should be simplified
6208 recursively. */
6210 static rtx
6211 cse_process_note_1 (rtx x, const_rtx, void *)
6213 if (MEM_P (x))
6215 validate_change (x, &XEXP (x, 0), cse_process_note (XEXP (x, 0)), false);
6216 return x;
6219 if (REG_P (x))
6221 int i = REG_QTY (REGNO (x));
6223 /* Return a constant or a constant register. */
6224 if (REGNO_QTY_VALID_P (REGNO (x)))
6226 struct qty_table_elem *ent = &qty_table[i];
6228 if (ent->const_rtx != NULL_RTX
6229 && (CONSTANT_P (ent->const_rtx)
6230 || REG_P (ent->const_rtx)))
6232 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6233 if (new_rtx)
6234 return copy_rtx (new_rtx);
6238 /* Otherwise, canonicalize this register. */
6239 return canon_reg (x, NULL);
6242 return NULL_RTX;
6245 /* Process X, part of the REG_NOTES of an insn. Replace any registers in it
6246 with either an equivalent constant or the canonical form of the register.
6247 Only replace addresses if the containing MEM remains valid. */
6249 static rtx
6250 cse_process_note (rtx x)
6252 return simplify_replace_fn_rtx (x, NULL_RTX, cse_process_note_1, NULL);
6256 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6258 DATA is a pointer to a struct cse_basic_block_data, that is used to
6259 describe the path.
6260 It is filled with a queue of basic blocks, starting with FIRST_BB
6261 and following a trace through the CFG.
6263 If all paths starting at FIRST_BB have been followed, or no new path
6264 starting at FIRST_BB can be constructed, this function returns FALSE.
6265 Otherwise, DATA->path is filled and the function returns TRUE indicating
6266 that a path to follow was found.
6268 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6269 block in the path will be FIRST_BB. */
6271 static bool
6272 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6273 int follow_jumps)
6275 basic_block bb;
6276 edge e;
6277 int path_size;
6279 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6281 /* See if there is a previous path. */
6282 path_size = data->path_size;
6284 /* There is a previous path. Make sure it started with FIRST_BB. */
6285 if (path_size)
6286 gcc_assert (data->path[0].bb == first_bb);
6288 /* There was only one basic block in the last path. Clear the path and
6289 return, so that paths starting at another basic block can be tried. */
6290 if (path_size == 1)
6292 path_size = 0;
6293 goto done;
6296 /* If the path was empty from the beginning, construct a new path. */
6297 if (path_size == 0)
6298 data->path[path_size++].bb = first_bb;
6299 else
6301 /* Otherwise, path_size must be equal to or greater than 2, because
6302 a previous path exists that is at least two basic blocks long.
6304 Update the previous branch path, if any. If the last branch was
6305 previously along the branch edge, take the fallthrough edge now. */
6306 while (path_size >= 2)
6308 basic_block last_bb_in_path, previous_bb_in_path;
6309 edge e;
6311 --path_size;
6312 last_bb_in_path = data->path[path_size].bb;
6313 previous_bb_in_path = data->path[path_size - 1].bb;
6315 /* If we previously followed a path along the branch edge, try
6316 the fallthru edge now. */
6317 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6318 && any_condjump_p (BB_END (previous_bb_in_path))
6319 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6320 && e == BRANCH_EDGE (previous_bb_in_path))
6322 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6323 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6324 && single_pred_p (bb)
6325 /* We used to assert here that we would only see blocks
6326 that we have not visited yet. But we may end up
6327 visiting basic blocks twice if the CFG has changed
6328 in this run of cse_main, because when the CFG changes
6329 the topological sort of the CFG also changes. A basic
6330 blocks that previously had more than two predecessors
6331 may now have a single predecessor, and become part of
6332 a path that starts at another basic block.
6334 We still want to visit each basic block only once, so
6335 halt the path here if we have already visited BB. */
6336 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6338 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6339 data->path[path_size++].bb = bb;
6340 break;
6344 data->path[path_size].bb = NULL;
6347 /* If only one block remains in the path, bail. */
6348 if (path_size == 1)
6350 path_size = 0;
6351 goto done;
6355 /* Extend the path if possible. */
6356 if (follow_jumps)
6358 bb = data->path[path_size - 1].bb;
6359 while (bb && path_size < param_max_cse_path_length)
6361 if (single_succ_p (bb))
6362 e = single_succ_edge (bb);
6363 else if (EDGE_COUNT (bb->succs) == 2
6364 && any_condjump_p (BB_END (bb)))
6366 /* First try to follow the branch. If that doesn't lead
6367 to a useful path, follow the fallthru edge. */
6368 e = BRANCH_EDGE (bb);
6369 if (!single_pred_p (e->dest))
6370 e = FALLTHRU_EDGE (bb);
6372 else
6373 e = NULL;
6375 if (e
6376 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6377 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6378 && single_pred_p (e->dest)
6379 /* Avoid visiting basic blocks twice. The large comment
6380 above explains why this can happen. */
6381 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6383 basic_block bb2 = e->dest;
6384 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6385 data->path[path_size++].bb = bb2;
6386 bb = bb2;
6388 else
6389 bb = NULL;
6393 done:
6394 data->path_size = path_size;
6395 return path_size != 0;
6398 /* Dump the path in DATA to file F. NSETS is the number of sets
6399 in the path. */
6401 static void
6402 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6404 int path_entry;
6406 fprintf (f, ";; Following path with %d sets: ", nsets);
6407 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6408 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6409 fputc ('\n', f);
6410 fflush (f);
6414 /* Return true if BB has exception handling successor edges. */
6416 static bool
6417 have_eh_succ_edges (basic_block bb)
6419 edge e;
6420 edge_iterator ei;
6422 FOR_EACH_EDGE (e, ei, bb->succs)
6423 if (e->flags & EDGE_EH)
6424 return true;
6426 return false;
6430 /* Scan to the end of the path described by DATA. Return an estimate of
6431 the total number of SETs of all insns in the path. */
6433 static void
6434 cse_prescan_path (struct cse_basic_block_data *data)
6436 int nsets = 0;
6437 int path_size = data->path_size;
6438 int path_entry;
6440 /* Scan to end of each basic block in the path. */
6441 for (path_entry = 0; path_entry < path_size; path_entry++)
6443 basic_block bb;
6444 rtx_insn *insn;
6446 bb = data->path[path_entry].bb;
6448 FOR_BB_INSNS (bb, insn)
6450 if (!INSN_P (insn))
6451 continue;
6453 /* A PARALLEL can have lots of SETs in it,
6454 especially if it is really an ASM_OPERANDS. */
6455 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6456 nsets += XVECLEN (PATTERN (insn), 0);
6457 else
6458 nsets += 1;
6462 data->nsets = nsets;
6465 /* Return true if the pattern of INSN uses a LABEL_REF for which
6466 there isn't a REG_LABEL_OPERAND note. */
6468 static bool
6469 check_for_label_ref (rtx_insn *insn)
6471 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6472 note for it, we must rerun jump since it needs to place the note. If
6473 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6474 don't do this since no REG_LABEL_OPERAND will be added. */
6475 subrtx_iterator::array_type array;
6476 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6478 const_rtx x = *iter;
6479 if (GET_CODE (x) == LABEL_REF
6480 && !LABEL_REF_NONLOCAL_P (x)
6481 && (!JUMP_P (insn)
6482 || !label_is_jump_target_p (label_ref_label (x), insn))
6483 && LABEL_P (label_ref_label (x))
6484 && INSN_UID (label_ref_label (x)) != 0
6485 && !find_reg_note (insn, REG_LABEL_OPERAND, label_ref_label (x)))
6486 return true;
6488 return false;
6491 /* Process a single extended basic block described by EBB_DATA. */
6493 static void
6494 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6496 int path_size = ebb_data->path_size;
6497 int path_entry;
6498 int num_insns = 0;
6500 /* Allocate the space needed by qty_table. */
6501 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6503 new_basic_block ();
6504 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6505 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6506 for (path_entry = 0; path_entry < path_size; path_entry++)
6508 basic_block bb;
6509 rtx_insn *insn;
6511 bb = ebb_data->path[path_entry].bb;
6513 /* Invalidate recorded information for eh regs if there is an EH
6514 edge pointing to that bb. */
6515 if (bb_has_eh_pred (bb))
6517 df_ref def;
6519 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6520 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6521 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6524 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6525 FOR_BB_INSNS (bb, insn)
6527 /* If we have processed 1,000 insns, flush the hash table to
6528 avoid extreme quadratic behavior. We must not include NOTEs
6529 in the count since there may be more of them when generating
6530 debugging information. If we clear the table at different
6531 times, code generated with -g -O might be different than code
6532 generated with -O but not -g.
6534 FIXME: This is a real kludge and needs to be done some other
6535 way. */
6536 if (NONDEBUG_INSN_P (insn)
6537 && num_insns++ > param_max_cse_insns)
6539 flush_hash_table ();
6540 num_insns = 0;
6543 if (INSN_P (insn))
6545 /* Process notes first so we have all notes in canonical forms
6546 when looking for duplicate operations. */
6547 bool changed = false;
6548 for (rtx note = REG_NOTES (insn); note; note = XEXP (note, 1))
6549 if (REG_NOTE_KIND (note) == REG_EQUAL)
6551 rtx newval = cse_process_note (XEXP (note, 0));
6552 if (newval != XEXP (note, 0))
6554 XEXP (note, 0) = newval;
6555 changed = true;
6558 if (changed)
6559 df_notes_rescan (insn);
6561 cse_insn (insn);
6563 /* If we haven't already found an insn where we added a LABEL_REF,
6564 check this one. */
6565 if (INSN_P (insn) && !recorded_label_ref
6566 && check_for_label_ref (insn))
6567 recorded_label_ref = true;
6571 /* With non-call exceptions, we are not always able to update
6572 the CFG properly inside cse_insn. So clean up possibly
6573 redundant EH edges here. */
6574 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6575 cse_cfg_altered |= purge_dead_edges (bb);
6577 /* If we changed a conditional jump, we may have terminated
6578 the path we are following. Check that by verifying that
6579 the edge we would take still exists. If the edge does
6580 not exist anymore, purge the remainder of the path.
6581 Note that this will cause us to return to the caller. */
6582 if (path_entry < path_size - 1)
6584 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6585 if (!find_edge (bb, next_bb))
6589 path_size--;
6591 /* If we truncate the path, we must also reset the
6592 visited bit on the remaining blocks in the path,
6593 or we will never visit them at all. */
6594 bitmap_clear_bit (cse_visited_basic_blocks,
6595 ebb_data->path[path_size].bb->index);
6596 ebb_data->path[path_size].bb = NULL;
6598 while (path_size - 1 != path_entry);
6599 ebb_data->path_size = path_size;
6603 /* If this is a conditional jump insn, record any known
6604 equivalences due to the condition being tested. */
6605 insn = BB_END (bb);
6606 if (path_entry < path_size - 1
6607 && EDGE_COUNT (bb->succs) == 2
6608 && JUMP_P (insn)
6609 && single_set (insn)
6610 && any_condjump_p (insn))
6612 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6613 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6614 record_jump_equiv (insn, taken);
6618 gcc_assert (next_qty <= max_qty);
6620 free (qty_table);
6624 /* Perform cse on the instructions of a function.
6625 F is the first instruction.
6626 NREGS is one plus the highest pseudo-reg number used in the instruction.
6628 Return 2 if jump optimizations should be redone due to simplifications
6629 in conditional jump instructions.
6630 Return 1 if the CFG should be cleaned up because it has been modified.
6631 Return 0 otherwise. */
6633 static int
6634 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6636 struct cse_basic_block_data ebb_data;
6637 basic_block bb;
6638 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6639 int i, n_blocks;
6641 /* CSE doesn't use dominane info but can invalidate it in different ways.
6642 For simplicity free dominance info here. */
6643 free_dominance_info (CDI_DOMINATORS);
6645 df_set_flags (DF_LR_RUN_DCE);
6646 df_note_add_problem ();
6647 df_analyze ();
6648 df_set_flags (DF_DEFER_INSN_RESCAN);
6650 reg_scan (get_insns (), max_reg_num ());
6651 init_cse_reg_info (nregs);
6653 ebb_data.path = XNEWVEC (struct branch_path,
6654 param_max_cse_path_length);
6656 cse_cfg_altered = false;
6657 cse_jumps_altered = false;
6658 recorded_label_ref = false;
6659 ebb_data.path_size = 0;
6660 ebb_data.nsets = 0;
6661 rtl_hooks = cse_rtl_hooks;
6663 init_recog ();
6664 init_alias_analysis ();
6666 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6668 /* Set up the table of already visited basic blocks. */
6669 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6670 bitmap_clear (cse_visited_basic_blocks);
6672 /* Loop over basic blocks in reverse completion order (RPO),
6673 excluding the ENTRY and EXIT blocks. */
6674 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6675 i = 0;
6676 while (i < n_blocks)
6678 /* Find the first block in the RPO queue that we have not yet
6679 processed before. */
6682 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6684 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6685 && i < n_blocks);
6687 /* Find all paths starting with BB, and process them. */
6688 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6690 /* Pre-scan the path. */
6691 cse_prescan_path (&ebb_data);
6693 /* If this basic block has no sets, skip it. */
6694 if (ebb_data.nsets == 0)
6695 continue;
6697 /* Get a reasonable estimate for the maximum number of qty's
6698 needed for this path. For this, we take the number of sets
6699 and multiply that by MAX_RECOG_OPERANDS. */
6700 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6702 /* Dump the path we're about to process. */
6703 if (dump_file)
6704 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6706 cse_extended_basic_block (&ebb_data);
6710 /* Clean up. */
6711 end_alias_analysis ();
6712 free (reg_eqv_table);
6713 free (ebb_data.path);
6714 sbitmap_free (cse_visited_basic_blocks);
6715 free (rc_order);
6716 rtl_hooks = general_rtl_hooks;
6718 if (cse_jumps_altered || recorded_label_ref)
6719 return 2;
6720 else if (cse_cfg_altered)
6721 return 1;
6722 else
6723 return 0;
6726 /* Count the number of times registers are used (not set) in X.
6727 COUNTS is an array in which we accumulate the count, INCR is how much
6728 we count each register usage.
6730 Don't count a usage of DEST, which is the SET_DEST of a SET which
6731 contains X in its SET_SRC. This is because such a SET does not
6732 modify the liveness of DEST.
6733 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6734 We must then count uses of a SET_DEST regardless, because the insn can't be
6735 deleted here. */
6737 static void
6738 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6740 enum rtx_code code;
6741 rtx note;
6742 const char *fmt;
6743 int i, j;
6745 if (x == 0)
6746 return;
6748 switch (code = GET_CODE (x))
6750 case REG:
6751 if (x != dest)
6752 counts[REGNO (x)] += incr;
6753 return;
6755 case PC:
6756 case CONST:
6757 CASE_CONST_ANY:
6758 case SYMBOL_REF:
6759 case LABEL_REF:
6760 return;
6762 case CLOBBER:
6763 /* If we are clobbering a MEM, mark any registers inside the address
6764 as being used. */
6765 if (MEM_P (XEXP (x, 0)))
6766 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6767 return;
6769 case SET:
6770 /* Unless we are setting a REG, count everything in SET_DEST. */
6771 if (!REG_P (SET_DEST (x)))
6772 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6773 count_reg_usage (SET_SRC (x), counts,
6774 dest ? dest : SET_DEST (x),
6775 incr);
6776 return;
6778 case DEBUG_INSN:
6779 return;
6781 case CALL_INSN:
6782 case INSN:
6783 case JUMP_INSN:
6784 /* We expect dest to be NULL_RTX here. If the insn may throw,
6785 or if it cannot be deleted due to side-effects, mark this fact
6786 by setting DEST to pc_rtx. */
6787 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6788 || side_effects_p (PATTERN (x)))
6789 dest = pc_rtx;
6790 if (code == CALL_INSN)
6791 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6792 count_reg_usage (PATTERN (x), counts, dest, incr);
6794 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6795 use them. */
6797 note = find_reg_equal_equiv_note (x);
6798 if (note)
6800 rtx eqv = XEXP (note, 0);
6802 if (GET_CODE (eqv) == EXPR_LIST)
6803 /* This REG_EQUAL note describes the result of a function call.
6804 Process all the arguments. */
6807 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6808 eqv = XEXP (eqv, 1);
6810 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6811 else
6812 count_reg_usage (eqv, counts, dest, incr);
6814 return;
6816 case EXPR_LIST:
6817 if (REG_NOTE_KIND (x) == REG_EQUAL
6818 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6819 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6820 involving registers in the address. */
6821 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6822 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6824 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6825 return;
6827 case ASM_OPERANDS:
6828 /* Iterate over just the inputs, not the constraints as well. */
6829 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6830 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6831 return;
6833 case INSN_LIST:
6834 case INT_LIST:
6835 gcc_unreachable ();
6837 default:
6838 break;
6841 fmt = GET_RTX_FORMAT (code);
6842 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6844 if (fmt[i] == 'e')
6845 count_reg_usage (XEXP (x, i), counts, dest, incr);
6846 else if (fmt[i] == 'E')
6847 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6848 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6852 /* Return true if X is a dead register. */
6854 static inline int
6855 is_dead_reg (const_rtx x, int *counts)
6857 return (REG_P (x)
6858 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6859 && counts[REGNO (x)] == 0);
6862 /* Return true if set is live. */
6863 static bool
6864 set_live_p (rtx set, int *counts)
6866 if (set_noop_p (set))
6867 return false;
6869 if (!is_dead_reg (SET_DEST (set), counts)
6870 || side_effects_p (SET_SRC (set)))
6871 return true;
6873 return false;
6876 /* Return true if insn is live. */
6878 static bool
6879 insn_live_p (rtx_insn *insn, int *counts)
6881 int i;
6882 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6883 return true;
6884 else if (GET_CODE (PATTERN (insn)) == SET)
6885 return set_live_p (PATTERN (insn), counts);
6886 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6888 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6890 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6892 if (GET_CODE (elt) == SET)
6894 if (set_live_p (elt, counts))
6895 return true;
6897 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6898 return true;
6900 return false;
6902 else if (DEBUG_INSN_P (insn))
6904 rtx_insn *next;
6906 if (DEBUG_MARKER_INSN_P (insn))
6907 return true;
6909 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6910 if (NOTE_P (next))
6911 continue;
6912 else if (!DEBUG_INSN_P (next))
6913 return true;
6914 /* If we find an inspection point, such as a debug begin stmt,
6915 we want to keep the earlier debug insn. */
6916 else if (DEBUG_MARKER_INSN_P (next))
6917 return true;
6918 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6919 return false;
6921 return true;
6923 else
6924 return true;
6927 /* Count the number of stores into pseudo. Callback for note_stores. */
6929 static void
6930 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6932 int *counts = (int *) data;
6933 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6934 counts[REGNO (x)]++;
6937 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6938 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6939 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6940 Set *SEEN_REPL to true if we see a dead register that does have
6941 a replacement. */
6943 static bool
6944 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6945 bool *seen_repl)
6947 subrtx_iterator::array_type array;
6948 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6950 const_rtx x = *iter;
6951 if (is_dead_reg (x, counts))
6953 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6954 *seen_repl = true;
6955 else
6956 return true;
6959 return false;
6962 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6963 Callback for simplify_replace_fn_rtx. */
6965 static rtx
6966 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6968 rtx *replacements = (rtx *) data;
6970 if (REG_P (x)
6971 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6972 && replacements[REGNO (x)] != NULL_RTX)
6974 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6975 return replacements[REGNO (x)];
6976 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6977 GET_MODE (replacements[REGNO (x)]));
6979 return NULL_RTX;
6982 /* Scan all the insns and delete any that are dead; i.e., they store a register
6983 that is never used or they copy a register to itself.
6985 This is used to remove insns made obviously dead by cse, loop or other
6986 optimizations. It improves the heuristics in loop since it won't try to
6987 move dead invariants out of loops or make givs for dead quantities. The
6988 remaining passes of the compilation are also sped up. */
6991 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
6993 int *counts;
6994 rtx_insn *insn, *prev;
6995 rtx *replacements = NULL;
6996 int ndead = 0;
6998 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6999 /* First count the number of times each register is used. */
7000 if (MAY_HAVE_DEBUG_BIND_INSNS)
7002 counts = XCNEWVEC (int, nreg * 3);
7003 for (insn = insns; insn; insn = NEXT_INSN (insn))
7004 if (DEBUG_BIND_INSN_P (insn))
7005 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7006 NULL_RTX, 1);
7007 else if (INSN_P (insn))
7009 count_reg_usage (insn, counts, NULL_RTX, 1);
7010 note_stores (insn, count_stores, counts + nreg * 2);
7012 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7013 First one counts how many times each pseudo is used outside
7014 of debug insns, second counts how many times each pseudo is
7015 used in debug insns and third counts how many times a pseudo
7016 is stored. */
7018 else
7020 counts = XCNEWVEC (int, nreg);
7021 for (insn = insns; insn; insn = NEXT_INSN (insn))
7022 if (INSN_P (insn))
7023 count_reg_usage (insn, counts, NULL_RTX, 1);
7024 /* If no debug insns can be present, COUNTS is just an array
7025 which counts how many times each pseudo is used. */
7027 /* Pseudo PIC register should be considered as used due to possible
7028 new usages generated. */
7029 if (!reload_completed
7030 && pic_offset_table_rtx
7031 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7032 counts[REGNO (pic_offset_table_rtx)]++;
7033 /* Go from the last insn to the first and delete insns that only set unused
7034 registers or copy a register to itself. As we delete an insn, remove
7035 usage counts for registers it uses.
7037 The first jump optimization pass may leave a real insn as the last
7038 insn in the function. We must not skip that insn or we may end
7039 up deleting code that is not really dead.
7041 If some otherwise unused register is only used in DEBUG_INSNs,
7042 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7043 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7044 has been created for the unused register, replace it with
7045 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7046 for (insn = get_last_insn (); insn; insn = prev)
7048 int live_insn = 0;
7050 prev = PREV_INSN (insn);
7051 if (!INSN_P (insn))
7052 continue;
7054 live_insn = insn_live_p (insn, counts);
7056 /* If this is a dead insn, delete it and show registers in it aren't
7057 being used. */
7059 if (! live_insn && dbg_cnt (delete_trivial_dead))
7061 if (DEBUG_INSN_P (insn))
7063 if (DEBUG_BIND_INSN_P (insn))
7064 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7065 NULL_RTX, -1);
7067 else
7069 rtx set;
7070 if (MAY_HAVE_DEBUG_BIND_INSNS
7071 && (set = single_set (insn)) != NULL_RTX
7072 && is_dead_reg (SET_DEST (set), counts)
7073 /* Used at least once in some DEBUG_INSN. */
7074 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7075 /* And set exactly once. */
7076 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7077 && !side_effects_p (SET_SRC (set))
7078 && asm_noperands (PATTERN (insn)) < 0)
7080 rtx dval, bind_var_loc;
7081 rtx_insn *bind;
7083 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7084 dval = make_debug_expr_from_rtl (SET_DEST (set));
7086 /* Emit a debug bind insn before the insn in which
7087 reg dies. */
7088 bind_var_loc =
7089 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7090 DEBUG_EXPR_TREE_DECL (dval),
7091 SET_SRC (set),
7092 VAR_INIT_STATUS_INITIALIZED);
7093 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7095 bind = emit_debug_insn_before (bind_var_loc, insn);
7096 df_insn_rescan (bind);
7098 if (replacements == NULL)
7099 replacements = XCNEWVEC (rtx, nreg);
7100 replacements[REGNO (SET_DEST (set))] = dval;
7103 count_reg_usage (insn, counts, NULL_RTX, -1);
7104 ndead++;
7106 cse_cfg_altered |= delete_insn_and_edges (insn);
7110 if (MAY_HAVE_DEBUG_BIND_INSNS)
7112 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7113 if (DEBUG_BIND_INSN_P (insn))
7115 /* If this debug insn references a dead register that wasn't replaced
7116 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7117 bool seen_repl = false;
7118 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7119 counts, replacements, &seen_repl))
7121 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7122 df_insn_rescan (insn);
7124 else if (seen_repl)
7126 INSN_VAR_LOCATION_LOC (insn)
7127 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7128 NULL_RTX, replace_dead_reg,
7129 replacements);
7130 df_insn_rescan (insn);
7133 free (replacements);
7136 if (dump_file && ndead)
7137 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7138 ndead);
7139 /* Clean up. */
7140 free (counts);
7141 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7142 return ndead;
7145 /* If LOC contains references to NEWREG in a different mode, change them
7146 to use NEWREG instead. */
7148 static void
7149 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7150 rtx *loc, rtx_insn *insn, rtx newreg)
7152 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7154 rtx *loc = *iter;
7155 rtx x = *loc;
7156 if (x
7157 && REG_P (x)
7158 && REGNO (x) == REGNO (newreg)
7159 && GET_MODE (x) != GET_MODE (newreg))
7161 validate_change (insn, loc, newreg, 1);
7162 iter.skip_subrtxes ();
7167 /* Change the mode of any reference to the register REGNO (NEWREG) to
7168 GET_MODE (NEWREG) in INSN. */
7170 static void
7171 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7173 int success;
7175 if (!INSN_P (insn))
7176 return;
7178 subrtx_ptr_iterator::array_type array;
7179 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7180 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7182 /* If the following assertion was triggered, there is most probably
7183 something wrong with the cc_modes_compatible back end function.
7184 CC modes only can be considered compatible if the insn - with the mode
7185 replaced by any of the compatible modes - can still be recognized. */
7186 success = apply_change_group ();
7187 gcc_assert (success);
7190 /* Change the mode of any reference to the register REGNO (NEWREG) to
7191 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7192 any instruction which modifies NEWREG. */
7194 static void
7195 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7197 rtx_insn *insn;
7199 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7201 if (! INSN_P (insn))
7202 continue;
7204 if (reg_set_p (newreg, insn))
7205 return;
7207 cse_change_cc_mode_insn (insn, newreg);
7211 /* BB is a basic block which finishes with CC_REG as a condition code
7212 register which is set to CC_SRC. Look through the successors of BB
7213 to find blocks which have a single predecessor (i.e., this one),
7214 and look through those blocks for an assignment to CC_REG which is
7215 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7216 permitted to change the mode of CC_SRC to a compatible mode. This
7217 returns VOIDmode if no equivalent assignments were found.
7218 Otherwise it returns the mode which CC_SRC should wind up with.
7219 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7220 but is passed unmodified down to recursive calls in order to prevent
7221 endless recursion.
7223 The main complexity in this function is handling the mode issues.
7224 We may have more than one duplicate which we can eliminate, and we
7225 try to find a mode which will work for multiple duplicates. */
7227 static machine_mode
7228 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7229 bool can_change_mode)
7231 bool found_equiv;
7232 machine_mode mode;
7233 unsigned int insn_count;
7234 edge e;
7235 rtx_insn *insns[2];
7236 machine_mode modes[2];
7237 rtx_insn *last_insns[2];
7238 unsigned int i;
7239 rtx newreg;
7240 edge_iterator ei;
7242 /* We expect to have two successors. Look at both before picking
7243 the final mode for the comparison. If we have more successors
7244 (i.e., some sort of table jump, although that seems unlikely),
7245 then we require all beyond the first two to use the same
7246 mode. */
7248 found_equiv = false;
7249 mode = GET_MODE (cc_src);
7250 insn_count = 0;
7251 FOR_EACH_EDGE (e, ei, bb->succs)
7253 rtx_insn *insn;
7254 rtx_insn *end;
7256 if (e->flags & EDGE_COMPLEX)
7257 continue;
7259 if (EDGE_COUNT (e->dest->preds) != 1
7260 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7261 /* Avoid endless recursion on unreachable blocks. */
7262 || e->dest == orig_bb)
7263 continue;
7265 end = NEXT_INSN (BB_END (e->dest));
7266 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7268 rtx set;
7270 if (! INSN_P (insn))
7271 continue;
7273 /* If CC_SRC is modified, we have to stop looking for
7274 something which uses it. */
7275 if (modified_in_p (cc_src, insn))
7276 break;
7278 /* Check whether INSN sets CC_REG to CC_SRC. */
7279 set = single_set (insn);
7280 if (set
7281 && REG_P (SET_DEST (set))
7282 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7284 bool found;
7285 machine_mode set_mode;
7286 machine_mode comp_mode;
7288 found = false;
7289 set_mode = GET_MODE (SET_SRC (set));
7290 comp_mode = set_mode;
7291 if (rtx_equal_p (cc_src, SET_SRC (set)))
7292 found = true;
7293 else if (GET_CODE (cc_src) == COMPARE
7294 && GET_CODE (SET_SRC (set)) == COMPARE
7295 && mode != set_mode
7296 && rtx_equal_p (XEXP (cc_src, 0),
7297 XEXP (SET_SRC (set), 0))
7298 && rtx_equal_p (XEXP (cc_src, 1),
7299 XEXP (SET_SRC (set), 1)))
7302 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7303 if (comp_mode != VOIDmode
7304 && (can_change_mode || comp_mode == mode))
7305 found = true;
7308 if (found)
7310 found_equiv = true;
7311 if (insn_count < ARRAY_SIZE (insns))
7313 insns[insn_count] = insn;
7314 modes[insn_count] = set_mode;
7315 last_insns[insn_count] = end;
7316 ++insn_count;
7318 if (mode != comp_mode)
7320 gcc_assert (can_change_mode);
7321 mode = comp_mode;
7323 /* The modified insn will be re-recognized later. */
7324 PUT_MODE (cc_src, mode);
7327 else
7329 if (set_mode != mode)
7331 /* We found a matching expression in the
7332 wrong mode, but we don't have room to
7333 store it in the array. Punt. This case
7334 should be rare. */
7335 break;
7337 /* INSN sets CC_REG to a value equal to CC_SRC
7338 with the right mode. We can simply delete
7339 it. */
7340 delete_insn (insn);
7343 /* We found an instruction to delete. Keep looking,
7344 in the hopes of finding a three-way jump. */
7345 continue;
7348 /* We found an instruction which sets the condition
7349 code, so don't look any farther. */
7350 break;
7353 /* If INSN sets CC_REG in some other way, don't look any
7354 farther. */
7355 if (reg_set_p (cc_reg, insn))
7356 break;
7359 /* If we fell off the bottom of the block, we can keep looking
7360 through successors. We pass CAN_CHANGE_MODE as false because
7361 we aren't prepared to handle compatibility between the
7362 further blocks and this block. */
7363 if (insn == end)
7365 machine_mode submode;
7367 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7368 if (submode != VOIDmode)
7370 gcc_assert (submode == mode);
7371 found_equiv = true;
7372 can_change_mode = false;
7377 if (! found_equiv)
7378 return VOIDmode;
7380 /* Now INSN_COUNT is the number of instructions we found which set
7381 CC_REG to a value equivalent to CC_SRC. The instructions are in
7382 INSNS. The modes used by those instructions are in MODES. */
7384 newreg = NULL_RTX;
7385 for (i = 0; i < insn_count; ++i)
7387 if (modes[i] != mode)
7389 /* We need to change the mode of CC_REG in INSNS[i] and
7390 subsequent instructions. */
7391 if (! newreg)
7393 if (GET_MODE (cc_reg) == mode)
7394 newreg = cc_reg;
7395 else
7396 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7398 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7399 newreg);
7402 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7405 return mode;
7408 /* If we have a fixed condition code register (or two), walk through
7409 the instructions and try to eliminate duplicate assignments. */
7411 static void
7412 cse_condition_code_reg (void)
7414 unsigned int cc_regno_1;
7415 unsigned int cc_regno_2;
7416 rtx cc_reg_1;
7417 rtx cc_reg_2;
7418 basic_block bb;
7420 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7421 return;
7423 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7424 if (cc_regno_2 != INVALID_REGNUM)
7425 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7426 else
7427 cc_reg_2 = NULL_RTX;
7429 FOR_EACH_BB_FN (bb, cfun)
7431 rtx_insn *last_insn;
7432 rtx cc_reg;
7433 rtx_insn *insn;
7434 rtx_insn *cc_src_insn;
7435 rtx cc_src;
7436 machine_mode mode;
7437 machine_mode orig_mode;
7439 /* Look for blocks which end with a conditional jump based on a
7440 condition code register. Then look for the instruction which
7441 sets the condition code register. Then look through the
7442 successor blocks for instructions which set the condition
7443 code register to the same value. There are other possible
7444 uses of the condition code register, but these are by far the
7445 most common and the ones which we are most likely to be able
7446 to optimize. */
7448 last_insn = BB_END (bb);
7449 if (!JUMP_P (last_insn))
7450 continue;
7452 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7453 cc_reg = cc_reg_1;
7454 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7455 cc_reg = cc_reg_2;
7456 else
7457 continue;
7459 cc_src_insn = NULL;
7460 cc_src = NULL_RTX;
7461 for (insn = PREV_INSN (last_insn);
7462 insn && insn != PREV_INSN (BB_HEAD (bb));
7463 insn = PREV_INSN (insn))
7465 rtx set;
7467 if (! INSN_P (insn))
7468 continue;
7469 set = single_set (insn);
7470 if (set
7471 && REG_P (SET_DEST (set))
7472 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7474 cc_src_insn = insn;
7475 cc_src = SET_SRC (set);
7476 break;
7478 else if (reg_set_p (cc_reg, insn))
7479 break;
7482 if (! cc_src_insn)
7483 continue;
7485 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7486 continue;
7488 /* Now CC_REG is a condition code register used for a
7489 conditional jump at the end of the block, and CC_SRC, in
7490 CC_SRC_INSN, is the value to which that condition code
7491 register is set, and CC_SRC is still meaningful at the end of
7492 the basic block. */
7494 orig_mode = GET_MODE (cc_src);
7495 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7496 if (mode != VOIDmode)
7498 gcc_assert (mode == GET_MODE (cc_src));
7499 if (mode != orig_mode)
7501 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7503 cse_change_cc_mode_insn (cc_src_insn, newreg);
7505 /* Do the same in the following insns that use the
7506 current value of CC_REG within BB. */
7507 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7508 NEXT_INSN (last_insn),
7509 newreg);
7516 /* Perform common subexpression elimination. Nonzero value from
7517 `cse_main' means that jumps were simplified and some code may now
7518 be unreachable, so do jump optimization again. */
7519 static unsigned int
7520 rest_of_handle_cse (void)
7522 int tem;
7524 if (dump_file)
7525 dump_flow_info (dump_file, dump_flags);
7527 tem = cse_main (get_insns (), max_reg_num ());
7529 /* If we are not running more CSE passes, then we are no longer
7530 expecting CSE to be run. But always rerun it in a cheap mode. */
7531 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7533 if (tem == 2)
7535 timevar_push (TV_JUMP);
7536 rebuild_jump_labels (get_insns ());
7537 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7538 timevar_pop (TV_JUMP);
7540 else if (tem == 1 || optimize > 1)
7541 cse_cfg_altered |= cleanup_cfg (0);
7543 return 0;
7546 namespace {
7548 const pass_data pass_data_cse =
7550 RTL_PASS, /* type */
7551 "cse1", /* name */
7552 OPTGROUP_NONE, /* optinfo_flags */
7553 TV_CSE, /* tv_id */
7554 0, /* properties_required */
7555 0, /* properties_provided */
7556 0, /* properties_destroyed */
7557 0, /* todo_flags_start */
7558 TODO_df_finish, /* todo_flags_finish */
7561 class pass_cse : public rtl_opt_pass
7563 public:
7564 pass_cse (gcc::context *ctxt)
7565 : rtl_opt_pass (pass_data_cse, ctxt)
7568 /* opt_pass methods: */
7569 virtual bool gate (function *) { return optimize > 0; }
7570 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7572 }; // class pass_cse
7574 } // anon namespace
7576 rtl_opt_pass *
7577 make_pass_cse (gcc::context *ctxt)
7579 return new pass_cse (ctxt);
7583 /* Run second CSE pass after loop optimizations. */
7584 static unsigned int
7585 rest_of_handle_cse2 (void)
7587 int tem;
7589 if (dump_file)
7590 dump_flow_info (dump_file, dump_flags);
7592 tem = cse_main (get_insns (), max_reg_num ());
7594 /* Run a pass to eliminate duplicated assignments to condition code
7595 registers. We have to run this after bypass_jumps, because it
7596 makes it harder for that pass to determine whether a jump can be
7597 bypassed safely. */
7598 cse_condition_code_reg ();
7600 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7602 if (tem == 2)
7604 timevar_push (TV_JUMP);
7605 rebuild_jump_labels (get_insns ());
7606 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7607 timevar_pop (TV_JUMP);
7609 else if (tem == 1 || cse_cfg_altered)
7610 cse_cfg_altered |= cleanup_cfg (0);
7612 cse_not_expected = 1;
7613 return 0;
7617 namespace {
7619 const pass_data pass_data_cse2 =
7621 RTL_PASS, /* type */
7622 "cse2", /* name */
7623 OPTGROUP_NONE, /* optinfo_flags */
7624 TV_CSE2, /* tv_id */
7625 0, /* properties_required */
7626 0, /* properties_provided */
7627 0, /* properties_destroyed */
7628 0, /* todo_flags_start */
7629 TODO_df_finish, /* todo_flags_finish */
7632 class pass_cse2 : public rtl_opt_pass
7634 public:
7635 pass_cse2 (gcc::context *ctxt)
7636 : rtl_opt_pass (pass_data_cse2, ctxt)
7639 /* opt_pass methods: */
7640 virtual bool gate (function *)
7642 return optimize > 0 && flag_rerun_cse_after_loop;
7645 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7647 }; // class pass_cse2
7649 } // anon namespace
7651 rtl_opt_pass *
7652 make_pass_cse2 (gcc::context *ctxt)
7654 return new pass_cse2 (ctxt);
7657 /* Run second CSE pass after loop optimizations. */
7658 static unsigned int
7659 rest_of_handle_cse_after_global_opts (void)
7661 int save_cfj;
7662 int tem;
7664 /* We only want to do local CSE, so don't follow jumps. */
7665 save_cfj = flag_cse_follow_jumps;
7666 flag_cse_follow_jumps = 0;
7668 rebuild_jump_labels (get_insns ());
7669 tem = cse_main (get_insns (), max_reg_num ());
7670 cse_cfg_altered |= purge_all_dead_edges ();
7671 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7673 cse_not_expected = !flag_rerun_cse_after_loop;
7675 /* If cse altered any jumps, rerun jump opts to clean things up. */
7676 if (tem == 2)
7678 timevar_push (TV_JUMP);
7679 rebuild_jump_labels (get_insns ());
7680 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7681 timevar_pop (TV_JUMP);
7683 else if (tem == 1 || cse_cfg_altered)
7684 cse_cfg_altered |= cleanup_cfg (0);
7686 flag_cse_follow_jumps = save_cfj;
7687 return 0;
7690 namespace {
7692 const pass_data pass_data_cse_after_global_opts =
7694 RTL_PASS, /* type */
7695 "cse_local", /* name */
7696 OPTGROUP_NONE, /* optinfo_flags */
7697 TV_CSE, /* tv_id */
7698 0, /* properties_required */
7699 0, /* properties_provided */
7700 0, /* properties_destroyed */
7701 0, /* todo_flags_start */
7702 TODO_df_finish, /* todo_flags_finish */
7705 class pass_cse_after_global_opts : public rtl_opt_pass
7707 public:
7708 pass_cse_after_global_opts (gcc::context *ctxt)
7709 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7712 /* opt_pass methods: */
7713 virtual bool gate (function *)
7715 return optimize > 0 && flag_rerun_cse_after_global_opts;
7718 virtual unsigned int execute (function *)
7720 return rest_of_handle_cse_after_global_opts ();
7723 }; // class pass_cse_after_global_opts
7725 } // anon namespace
7727 rtl_opt_pass *
7728 make_pass_cse_after_global_opts (gcc::context *ctxt)
7730 return new pass_cse_after_global_opts (ctxt);