tree-ssa-strlen: Fix up handle_store [PR113603]
[official-gcc.git] / gcc / cse.cc
blobc53deecbe54765b4703acfd8cf246b6056b65d1e
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2024 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "cfghooks.h"
28 #include "df.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "insn-config.h"
32 #include "regs.h"
33 #include "emit-rtl.h"
34 #include "recog.h"
35 #include "cfgrtl.h"
36 #include "cfganal.h"
37 #include "cfgcleanup.h"
38 #include "alias.h"
39 #include "toplev.h"
40 #include "rtlhooks-def.h"
41 #include "tree-pass.h"
42 #include "dbgcnt.h"
43 #include "rtl-iter.h"
44 #include "regs.h"
45 #include "function-abi.h"
46 #include "rtlanal.h"
47 #include "expr.h"
49 /* The basic idea of common subexpression elimination is to go
50 through the code, keeping a record of expressions that would
51 have the same value at the current scan point, and replacing
52 expressions encountered with the cheapest equivalent expression.
54 It is too complicated to keep track of the different possibilities
55 when control paths merge in this code; so, at each label, we forget all
56 that is known and start fresh. This can be described as processing each
57 extended basic block separately. We have a separate pass to perform
58 global CSE.
60 Note CSE can turn a conditional or computed jump into a nop or
61 an unconditional jump. When this occurs we arrange to run the jump
62 optimizer after CSE to delete the unreachable code.
64 We use two data structures to record the equivalent expressions:
65 a hash table for most expressions, and a vector of "quantity
66 numbers" to record equivalent (pseudo) registers.
68 The use of the special data structure for registers is desirable
69 because it is faster. It is possible because registers references
70 contain a fairly small number, the register number, taken from
71 a contiguously allocated series, and two register references are
72 identical if they have the same number. General expressions
73 do not have any such thing, so the only way to retrieve the
74 information recorded on an expression other than a register
75 is to keep it in a hash table.
77 Registers and "quantity numbers":
79 At the start of each basic block, all of the (hardware and pseudo)
80 registers used in the function are given distinct quantity
81 numbers to indicate their contents. During scan, when the code
82 copies one register into another, we copy the quantity number.
83 When a register is loaded in any other way, we allocate a new
84 quantity number to describe the value generated by this operation.
85 `REG_QTY (N)' records what quantity register N is currently thought
86 of as containing.
88 All real quantity numbers are greater than or equal to zero.
89 If register N has not been assigned a quantity, `REG_QTY (N)' will
90 equal -N - 1, which is always negative.
92 Quantity numbers below zero do not exist and none of the `qty_table'
93 entries should be referenced with a negative index.
95 We also maintain a bidirectional chain of registers for each
96 quantity number. The `qty_table` members `first_reg' and `last_reg',
97 and `reg_eqv_table' members `next' and `prev' hold these chains.
99 The first register in a chain is the one whose lifespan is least local.
100 Among equals, it is the one that was seen first.
101 We replace any equivalent register with that one.
103 If two registers have the same quantity number, it must be true that
104 REG expressions with qty_table `mode' must be in the hash table for both
105 registers and must be in the same class.
107 The converse is not true. Since hard registers may be referenced in
108 any mode, two REG expressions might be equivalent in the hash table
109 but not have the same quantity number if the quantity number of one
110 of the registers is not the same mode as those expressions.
112 Constants and quantity numbers
114 When a quantity has a known constant value, that value is stored
115 in the appropriate qty_table `const_rtx'. This is in addition to
116 putting the constant in the hash table as is usual for non-regs.
118 Whether a reg or a constant is preferred is determined by the configuration
119 macro CONST_COSTS and will often depend on the constant value. In any
120 event, expressions containing constants can be simplified, by fold_rtx.
122 When a quantity has a known nearly constant value (such as an address
123 of a stack slot), that value is stored in the appropriate qty_table
124 `const_rtx'.
126 Integer constants don't have a machine mode. However, cse
127 determines the intended machine mode from the destination
128 of the instruction that moves the constant. The machine mode
129 is recorded in the hash table along with the actual RTL
130 constant expression so that different modes are kept separate.
132 Other expressions:
134 To record known equivalences among expressions in general
135 we use a hash table called `table'. It has a fixed number of buckets
136 that contain chains of `struct table_elt' elements for expressions.
137 These chains connect the elements whose expressions have the same
138 hash codes.
140 Other chains through the same elements connect the elements which
141 currently have equivalent values.
143 Register references in an expression are canonicalized before hashing
144 the expression. This is done using `reg_qty' and qty_table `first_reg'.
145 The hash code of a register reference is computed using the quantity
146 number, not the register number.
148 When the value of an expression changes, it is necessary to remove from the
149 hash table not just that expression but all expressions whose values
150 could be different as a result.
152 1. If the value changing is in memory, except in special cases
153 ANYTHING referring to memory could be changed. That is because
154 nobody knows where a pointer does not point.
155 The function `invalidate_memory' removes what is necessary.
157 The special cases are when the address is constant or is
158 a constant plus a fixed register such as the frame pointer
159 or a static chain pointer. When such addresses are stored in,
160 we can tell exactly which other such addresses must be invalidated
161 due to overlap. `invalidate' does this.
162 All expressions that refer to non-constant
163 memory addresses are also invalidated. `invalidate_memory' does this.
165 2. If the value changing is a register, all expressions
166 containing references to that register, and only those,
167 must be removed.
169 Because searching the entire hash table for expressions that contain
170 a register is very slow, we try to figure out when it isn't necessary.
171 Precisely, this is necessary only when expressions have been
172 entered in the hash table using this register, and then the value has
173 changed, and then another expression wants to be added to refer to
174 the register's new value. This sequence of circumstances is rare
175 within any one basic block.
177 `REG_TICK' and `REG_IN_TABLE', accessors for members of
178 cse_reg_info, are used to detect this case. REG_TICK (i) is
179 incremented whenever a value is stored in register i.
180 REG_IN_TABLE (i) holds -1 if no references to register i have been
181 entered in the table; otherwise, it contains the value REG_TICK (i)
182 had when the references were entered. If we want to enter a
183 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
184 remove old references. Until we want to enter a new entry, the
185 mere fact that the two vectors don't match makes the entries be
186 ignored if anyone tries to match them.
188 Registers themselves are entered in the hash table as well as in
189 the equivalent-register chains. However, `REG_TICK' and
190 `REG_IN_TABLE' do not apply to expressions which are simple
191 register references. These expressions are removed from the table
192 immediately when they become invalid, and this can be done even if
193 we do not immediately search for all the expressions that refer to
194 the register.
196 A CLOBBER rtx in an instruction invalidates its operand for further
197 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
198 invalidates everything that resides in memory.
200 Related expressions:
202 Constant expressions that differ only by an additive integer
203 are called related. When a constant expression is put in
204 the table, the related expression with no constant term
205 is also entered. These are made to point at each other
206 so that it is possible to find out if there exists any
207 register equivalent to an expression related to a given expression. */
209 /* Length of qty_table vector. We know in advance we will not need
210 a quantity number this big. */
212 static int max_qty;
214 /* Next quantity number to be allocated.
215 This is 1 + the largest number needed so far. */
217 static int next_qty;
219 /* Per-qty information tracking.
221 `first_reg' and `last_reg' track the head and tail of the
222 chain of registers which currently contain this quantity.
224 `mode' contains the machine mode of this quantity.
226 `const_rtx' holds the rtx of the constant value of this
227 quantity, if known. A summations of the frame/arg pointer
228 and a constant can also be entered here. When this holds
229 a known value, `const_insn' is the insn which stored the
230 constant value.
232 `comparison_{code,const,qty}' are used to track when a
233 comparison between a quantity and some constant or register has
234 been passed. In such a case, we know the results of the comparison
235 in case we see it again. These members record a comparison that
236 is known to be true. `comparison_code' holds the rtx code of such
237 a comparison, else it is set to UNKNOWN and the other two
238 comparison members are undefined. `comparison_const' holds
239 the constant being compared against, or zero if the comparison
240 is not against a constant. `comparison_qty' holds the quantity
241 being compared against when the result is known. If the comparison
242 is not with a register, `comparison_qty' is -1. */
244 struct qty_table_elem
246 rtx const_rtx;
247 rtx_insn *const_insn;
248 rtx comparison_const;
249 int comparison_qty;
250 unsigned int first_reg, last_reg;
251 ENUM_BITFIELD(machine_mode) mode : MACHINE_MODE_BITSIZE;
252 ENUM_BITFIELD(rtx_code) comparison_code : RTX_CODE_BITSIZE;
255 /* The table of all qtys, indexed by qty number. */
256 static struct qty_table_elem *qty_table;
258 /* Insn being scanned. */
260 static rtx_insn *this_insn;
261 static bool optimize_this_for_speed_p;
263 /* Index by register number, gives the number of the next (or
264 previous) register in the chain of registers sharing the same
265 value.
267 Or -1 if this register is at the end of the chain.
269 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
271 /* Per-register equivalence chain. */
272 struct reg_eqv_elem
274 int next, prev;
277 /* The table of all register equivalence chains. */
278 static struct reg_eqv_elem *reg_eqv_table;
280 struct cse_reg_info
282 /* The timestamp at which this register is initialized. */
283 unsigned int timestamp;
285 /* The quantity number of the register's current contents. */
286 int reg_qty;
288 /* The number of times the register has been altered in the current
289 basic block. */
290 int reg_tick;
292 /* The REG_TICK value at which rtx's containing this register are
293 valid in the hash table. If this does not equal the current
294 reg_tick value, such expressions existing in the hash table are
295 invalid. */
296 int reg_in_table;
298 /* The SUBREG that was set when REG_TICK was last incremented. Set
299 to -1 if the last store was to the whole register, not a subreg. */
300 unsigned int subreg_ticked;
303 /* A table of cse_reg_info indexed by register numbers. */
304 static struct cse_reg_info *cse_reg_info_table;
306 /* The size of the above table. */
307 static unsigned int cse_reg_info_table_size;
309 /* The index of the first entry that has not been initialized. */
310 static unsigned int cse_reg_info_table_first_uninitialized;
312 /* The timestamp at the beginning of the current run of
313 cse_extended_basic_block. We increment this variable at the beginning of
314 the current run of cse_extended_basic_block. The timestamp field of a
315 cse_reg_info entry matches the value of this variable if and only
316 if the entry has been initialized during the current run of
317 cse_extended_basic_block. */
318 static unsigned int cse_reg_info_timestamp;
320 /* A HARD_REG_SET containing all the hard registers for which there is
321 currently a REG expression in the hash table. Note the difference
322 from the above variables, which indicate if the REG is mentioned in some
323 expression in the table. */
325 static HARD_REG_SET hard_regs_in_table;
327 /* True if CSE has altered the CFG. */
328 static bool cse_cfg_altered;
330 /* True if CSE has altered conditional jump insns in such a way
331 that jump optimization should be redone. */
332 static bool cse_jumps_altered;
334 /* True if we put a LABEL_REF into the hash table for an INSN
335 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
336 to put in the note. */
337 static bool recorded_label_ref;
339 /* canon_hash stores 1 in do_not_record if it notices a reference to PC or
340 some other volatile subexpression. */
342 static int do_not_record;
344 /* canon_hash stores 1 in hash_arg_in_memory
345 if it notices a reference to memory within the expression being hashed. */
347 static int hash_arg_in_memory;
349 /* The hash table contains buckets which are chains of `struct table_elt's,
350 each recording one expression's information.
351 That expression is in the `exp' field.
353 The canon_exp field contains a canonical (from the point of view of
354 alias analysis) version of the `exp' field.
356 Those elements with the same hash code are chained in both directions
357 through the `next_same_hash' and `prev_same_hash' fields.
359 Each set of expressions with equivalent values
360 are on a two-way chain through the `next_same_value'
361 and `prev_same_value' fields, and all point with
362 the `first_same_value' field at the first element in
363 that chain. The chain is in order of increasing cost.
364 Each element's cost value is in its `cost' field.
366 The `in_memory' field is nonzero for elements that
367 involve any reference to memory. These elements are removed
368 whenever a write is done to an unidentified location in memory.
369 To be safe, we assume that a memory address is unidentified unless
370 the address is either a symbol constant or a constant plus
371 the frame pointer or argument pointer.
373 The `related_value' field is used to connect related expressions
374 (that differ by adding an integer).
375 The related expressions are chained in a circular fashion.
376 `related_value' is zero for expressions for which this
377 chain is not useful.
379 The `cost' field stores the cost of this element's expression.
380 The `regcost' field stores the value returned by approx_reg_cost for
381 this element's expression.
383 The `is_const' flag is set if the element is a constant (including
384 a fixed address).
386 The `flag' field is used as a temporary during some search routines.
388 The `mode' field is usually the same as GET_MODE (`exp'), but
389 if `exp' is a CONST_INT and has no machine mode then the `mode'
390 field is the mode it was being used as. Each constant is
391 recorded separately for each mode it is used with. */
393 struct table_elt
395 rtx exp;
396 rtx canon_exp;
397 struct table_elt *next_same_hash;
398 struct table_elt *prev_same_hash;
399 struct table_elt *next_same_value;
400 struct table_elt *prev_same_value;
401 struct table_elt *first_same_value;
402 struct table_elt *related_value;
403 int cost;
404 int regcost;
405 ENUM_BITFIELD(machine_mode) mode : MACHINE_MODE_BITSIZE;
406 char in_memory;
407 char is_const;
408 char flag;
411 /* We don't want a lot of buckets, because we rarely have very many
412 things stored in the hash table, and a lot of buckets slows
413 down a lot of loops that happen frequently. */
414 #define HASH_SHIFT 5
415 #define HASH_SIZE (1 << HASH_SHIFT)
416 #define HASH_MASK (HASH_SIZE - 1)
418 /* Determine whether register number N is considered a fixed register for the
419 purpose of approximating register costs.
420 It is desirable to replace other regs with fixed regs, to reduce need for
421 non-fixed hard regs.
422 A reg wins if it is either the frame pointer or designated as fixed. */
423 #define FIXED_REGNO_P(N) \
424 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
425 || fixed_regs[N] || global_regs[N])
427 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
428 hard registers and pointers into the frame are the cheapest with a cost
429 of 0. Next come pseudos with a cost of one and other hard registers with
430 a cost of 2. Aside from these special cases, call `rtx_cost'. */
432 #define CHEAP_REGNO(N) \
433 (REGNO_PTR_FRAME_P (N) \
434 || (HARD_REGISTER_NUM_P (N) \
435 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
437 #define COST(X, MODE) \
438 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
439 #define COST_IN(X, MODE, OUTER, OPNO) \
440 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
442 /* Get the number of times this register has been updated in this
443 basic block. */
445 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
447 /* Get the point at which REG was recorded in the table. */
449 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
451 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
452 SUBREG). */
454 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
456 /* Get the quantity number for REG. */
458 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
460 /* Determine if the quantity number for register X represents a valid index
461 into the qty_table. */
463 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
465 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
467 #define CHEAPER(X, Y) \
468 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
470 static struct table_elt *table[HASH_SIZE];
472 /* Chain of `struct table_elt's made so far for this function
473 but currently removed from the table. */
475 static struct table_elt *free_element_chain;
477 /* Trace a patch through the CFG. */
479 struct branch_path
481 /* The basic block for this path entry. */
482 basic_block bb;
485 /* This data describes a block that will be processed by
486 cse_extended_basic_block. */
488 struct cse_basic_block_data
490 /* Total number of SETs in block. */
491 int nsets;
492 /* Size of current branch path, if any. */
493 int path_size;
494 /* Current path, indicating which basic_blocks will be processed. */
495 struct branch_path *path;
499 /* Pointers to the live in/live out bitmaps for the boundaries of the
500 current EBB. */
501 static bitmap cse_ebb_live_in, cse_ebb_live_out;
503 /* A simple bitmap to track which basic blocks have been visited
504 already as part of an already processed extended basic block. */
505 static sbitmap cse_visited_basic_blocks;
507 static bool fixed_base_plus_p (rtx x);
508 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
509 static int preferable (int, int, int, int);
510 static void new_basic_block (void);
511 static void make_new_qty (unsigned int, machine_mode);
512 static void make_regs_eqv (unsigned int, unsigned int);
513 static void delete_reg_equiv (unsigned int);
514 static bool mention_regs (rtx);
515 static bool insert_regs (rtx, struct table_elt *, bool);
516 static void remove_from_table (struct table_elt *, unsigned);
517 static void remove_pseudo_from_table (rtx, unsigned);
518 static struct table_elt *lookup (rtx, unsigned, machine_mode);
519 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
520 static rtx lookup_as_function (rtx, enum rtx_code);
521 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
522 machine_mode, int, int);
523 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
524 machine_mode);
525 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
526 static void invalidate (rtx, machine_mode);
527 static void remove_invalid_refs (unsigned int);
528 static void remove_invalid_subreg_refs (unsigned int, poly_uint64,
529 machine_mode);
530 static void rehash_using_reg (rtx);
531 static void invalidate_memory (void);
532 static rtx use_related_value (rtx, struct table_elt *);
534 static inline unsigned canon_hash (rtx, machine_mode);
535 static inline unsigned safe_hash (rtx, machine_mode);
536 static inline unsigned hash_rtx_string (const char *);
538 static rtx canon_reg (rtx, rtx_insn *);
539 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
540 machine_mode *,
541 machine_mode *);
542 static rtx fold_rtx (rtx, rtx_insn *);
543 static rtx equiv_constant (rtx);
544 static void record_jump_equiv (rtx_insn *, bool);
545 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx);
546 static void cse_insn (rtx_insn *);
547 static void cse_prescan_path (struct cse_basic_block_data *);
548 static void invalidate_from_clobbers (rtx_insn *);
549 static void invalidate_from_sets_and_clobbers (rtx_insn *);
550 static void cse_extended_basic_block (struct cse_basic_block_data *);
551 extern void dump_class (struct table_elt*);
552 static void get_cse_reg_info_1 (unsigned int regno);
553 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
555 static void flush_hash_table (void);
556 static bool insn_live_p (rtx_insn *, int *);
557 static bool set_live_p (rtx, int *);
558 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
559 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
560 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
561 bool);
564 #undef RTL_HOOKS_GEN_LOWPART
565 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
567 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
569 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
570 register (hard registers may require `do_not_record' to be set). */
572 static inline unsigned
573 HASH (rtx x, machine_mode mode)
575 unsigned h = (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER
576 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (x)))
577 : canon_hash (x, mode));
578 return (h ^ (h >> HASH_SHIFT)) & HASH_MASK;
581 /* Like HASH, but without side-effects. */
583 static inline unsigned
584 SAFE_HASH (rtx x, machine_mode mode)
586 unsigned h = (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER
587 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (x)))
588 : safe_hash (x, mode));
589 return (h ^ (h >> HASH_SHIFT)) & HASH_MASK;
592 /* Nonzero if X has the form (PLUS frame-pointer integer). */
594 static bool
595 fixed_base_plus_p (rtx x)
597 switch (GET_CODE (x))
599 case REG:
600 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
601 return true;
602 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
603 return true;
604 return false;
606 case PLUS:
607 if (!CONST_INT_P (XEXP (x, 1)))
608 return false;
609 return fixed_base_plus_p (XEXP (x, 0));
611 default:
612 return false;
616 /* Dump the expressions in the equivalence class indicated by CLASSP.
617 This function is used only for debugging. */
618 DEBUG_FUNCTION void
619 dump_class (struct table_elt *classp)
621 struct table_elt *elt;
623 fprintf (stderr, "Equivalence chain for ");
624 print_rtl (stderr, classp->exp);
625 fprintf (stderr, ": \n");
627 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
629 print_rtl (stderr, elt->exp);
630 fprintf (stderr, "\n");
634 /* Return an estimate of the cost of the registers used in an rtx.
635 This is mostly the number of different REG expressions in the rtx;
636 however for some exceptions like fixed registers we use a cost of
637 0. If any other hard register reference occurs, return MAX_COST. */
639 static int
640 approx_reg_cost (const_rtx x)
642 int cost = 0;
643 subrtx_iterator::array_type array;
644 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
646 const_rtx x = *iter;
647 if (REG_P (x))
649 unsigned int regno = REGNO (x);
650 if (!CHEAP_REGNO (regno))
652 if (regno < FIRST_PSEUDO_REGISTER)
654 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
655 return MAX_COST;
656 cost += 2;
658 else
659 cost += 1;
663 return cost;
666 /* Return a negative value if an rtx A, whose costs are given by COST_A
667 and REGCOST_A, is more desirable than an rtx B.
668 Return a positive value if A is less desirable, or 0 if the two are
669 equally good. */
670 static int
671 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
673 /* First, get rid of cases involving expressions that are entirely
674 unwanted. */
675 if (cost_a != cost_b)
677 if (cost_a == MAX_COST)
678 return 1;
679 if (cost_b == MAX_COST)
680 return -1;
683 /* Avoid extending lifetimes of hardregs. */
684 if (regcost_a != regcost_b)
686 if (regcost_a == MAX_COST)
687 return 1;
688 if (regcost_b == MAX_COST)
689 return -1;
692 /* Normal operation costs take precedence. */
693 if (cost_a != cost_b)
694 return cost_a - cost_b;
695 /* Only if these are identical consider effects on register pressure. */
696 if (regcost_a != regcost_b)
697 return regcost_a - regcost_b;
698 return 0;
701 /* Internal function, to compute cost when X is not a register; called
702 from COST macro to keep it simple. */
704 static int
705 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
707 scalar_int_mode int_mode, inner_mode;
708 return ((GET_CODE (x) == SUBREG
709 && REG_P (SUBREG_REG (x))
710 && is_int_mode (mode, &int_mode)
711 && is_int_mode (GET_MODE (SUBREG_REG (x)), &inner_mode)
712 && GET_MODE_SIZE (int_mode) < GET_MODE_SIZE (inner_mode)
713 && subreg_lowpart_p (x)
714 && TRULY_NOOP_TRUNCATION_MODES_P (int_mode, inner_mode))
716 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
720 /* Initialize CSE_REG_INFO_TABLE. */
722 static void
723 init_cse_reg_info (unsigned int nregs)
725 /* Do we need to grow the table? */
726 if (nregs > cse_reg_info_table_size)
728 unsigned int new_size;
730 if (cse_reg_info_table_size < 2048)
732 /* Compute a new size that is a power of 2 and no smaller
733 than the large of NREGS and 64. */
734 new_size = (cse_reg_info_table_size
735 ? cse_reg_info_table_size : 64);
737 while (new_size < nregs)
738 new_size *= 2;
740 else
742 /* If we need a big table, allocate just enough to hold
743 NREGS registers. */
744 new_size = nregs;
747 /* Reallocate the table with NEW_SIZE entries. */
748 free (cse_reg_info_table);
749 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
750 cse_reg_info_table_size = new_size;
751 cse_reg_info_table_first_uninitialized = 0;
754 /* Do we have all of the first NREGS entries initialized? */
755 if (cse_reg_info_table_first_uninitialized < nregs)
757 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
758 unsigned int i;
760 /* Put the old timestamp on newly allocated entries so that they
761 will all be considered out of date. We do not touch those
762 entries beyond the first NREGS entries to be nice to the
763 virtual memory. */
764 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
765 cse_reg_info_table[i].timestamp = old_timestamp;
767 cse_reg_info_table_first_uninitialized = nregs;
771 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
773 static void
774 get_cse_reg_info_1 (unsigned int regno)
776 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
777 entry will be considered to have been initialized. */
778 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
780 /* Initialize the rest of the entry. */
781 cse_reg_info_table[regno].reg_tick = 1;
782 cse_reg_info_table[regno].reg_in_table = -1;
783 cse_reg_info_table[regno].subreg_ticked = -1;
784 cse_reg_info_table[regno].reg_qty = -regno - 1;
787 /* Find a cse_reg_info entry for REGNO. */
789 static inline struct cse_reg_info *
790 get_cse_reg_info (unsigned int regno)
792 struct cse_reg_info *p = &cse_reg_info_table[regno];
794 /* If this entry has not been initialized, go ahead and initialize
795 it. */
796 if (p->timestamp != cse_reg_info_timestamp)
797 get_cse_reg_info_1 (regno);
799 return p;
802 /* Clear the hash table and initialize each register with its own quantity,
803 for a new basic block. */
805 static void
806 new_basic_block (void)
808 int i;
810 next_qty = 0;
812 /* Invalidate cse_reg_info_table. */
813 cse_reg_info_timestamp++;
815 /* Clear out hash table state for this pass. */
816 CLEAR_HARD_REG_SET (hard_regs_in_table);
818 /* The per-quantity values used to be initialized here, but it is
819 much faster to initialize each as it is made in `make_new_qty'. */
821 for (i = 0; i < HASH_SIZE; i++)
823 struct table_elt *first;
825 first = table[i];
826 if (first != NULL)
828 struct table_elt *last = first;
830 table[i] = NULL;
832 while (last->next_same_hash != NULL)
833 last = last->next_same_hash;
835 /* Now relink this hash entire chain into
836 the free element list. */
838 last->next_same_hash = free_element_chain;
839 free_element_chain = first;
844 /* Say that register REG contains a quantity in mode MODE not in any
845 register before and initialize that quantity. */
847 static void
848 make_new_qty (unsigned int reg, machine_mode mode)
850 int q;
851 struct qty_table_elem *ent;
852 struct reg_eqv_elem *eqv;
854 gcc_assert (next_qty < max_qty);
856 q = REG_QTY (reg) = next_qty++;
857 ent = &qty_table[q];
858 ent->first_reg = reg;
859 ent->last_reg = reg;
860 ent->mode = mode;
861 ent->const_rtx = ent->const_insn = NULL;
862 ent->comparison_code = UNKNOWN;
864 eqv = &reg_eqv_table[reg];
865 eqv->next = eqv->prev = -1;
868 /* Make reg NEW equivalent to reg OLD.
869 OLD is not changing; NEW is. */
871 static void
872 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
874 unsigned int lastr, firstr;
875 int q = REG_QTY (old_reg);
876 struct qty_table_elem *ent;
878 ent = &qty_table[q];
880 /* Nothing should become eqv until it has a "non-invalid" qty number. */
881 gcc_assert (REGNO_QTY_VALID_P (old_reg));
883 REG_QTY (new_reg) = q;
884 firstr = ent->first_reg;
885 lastr = ent->last_reg;
887 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
888 hard regs. Among pseudos, if NEW will live longer than any other reg
889 of the same qty, and that is beyond the current basic block,
890 make it the new canonical replacement for this qty. */
891 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
892 /* Certain fixed registers might be of the class NO_REGS. This means
893 that not only can they not be allocated by the compiler, but
894 they cannot be used in substitutions or canonicalizations
895 either. */
896 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
897 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
898 || (new_reg >= FIRST_PSEUDO_REGISTER
899 && (firstr < FIRST_PSEUDO_REGISTER
900 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
901 && !bitmap_bit_p (cse_ebb_live_out, firstr))
902 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
903 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
905 reg_eqv_table[firstr].prev = new_reg;
906 reg_eqv_table[new_reg].next = firstr;
907 reg_eqv_table[new_reg].prev = -1;
908 ent->first_reg = new_reg;
910 else
912 /* If NEW is a hard reg (known to be non-fixed), insert at end.
913 Otherwise, insert before any non-fixed hard regs that are at the
914 end. Registers of class NO_REGS cannot be used as an
915 equivalent for anything. */
916 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
917 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
918 && new_reg >= FIRST_PSEUDO_REGISTER)
919 lastr = reg_eqv_table[lastr].prev;
920 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
921 if (reg_eqv_table[lastr].next >= 0)
922 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
923 else
924 qty_table[q].last_reg = new_reg;
925 reg_eqv_table[lastr].next = new_reg;
926 reg_eqv_table[new_reg].prev = lastr;
930 /* Remove REG from its equivalence class. */
932 static void
933 delete_reg_equiv (unsigned int reg)
935 struct qty_table_elem *ent;
936 int q = REG_QTY (reg);
937 int p, n;
939 /* If invalid, do nothing. */
940 if (! REGNO_QTY_VALID_P (reg))
941 return;
943 ent = &qty_table[q];
945 p = reg_eqv_table[reg].prev;
946 n = reg_eqv_table[reg].next;
948 if (n != -1)
949 reg_eqv_table[n].prev = p;
950 else
951 ent->last_reg = p;
952 if (p != -1)
953 reg_eqv_table[p].next = n;
954 else
955 ent->first_reg = n;
957 REG_QTY (reg) = -reg - 1;
960 /* Remove any invalid expressions from the hash table
961 that refer to any of the registers contained in expression X.
963 Make sure that newly inserted references to those registers
964 as subexpressions will be considered valid.
966 mention_regs is not called when a register itself
967 is being stored in the table.
969 Return true if we have done something that may have changed
970 the hash code of X. */
972 static bool
973 mention_regs (rtx x)
975 enum rtx_code code;
976 int i, j;
977 const char *fmt;
978 bool changed = false;
980 if (x == 0)
981 return false;
983 code = GET_CODE (x);
984 if (code == REG)
986 unsigned int regno = REGNO (x);
987 unsigned int endregno = END_REGNO (x);
988 unsigned int i;
990 for (i = regno; i < endregno; i++)
992 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
993 remove_invalid_refs (i);
995 REG_IN_TABLE (i) = REG_TICK (i);
996 SUBREG_TICKED (i) = -1;
999 return false;
1002 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1003 pseudo if they don't use overlapping words. We handle only pseudos
1004 here for simplicity. */
1005 if (code == SUBREG && REG_P (SUBREG_REG (x))
1006 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1008 unsigned int i = REGNO (SUBREG_REG (x));
1010 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1012 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1013 the last store to this register really stored into this
1014 subreg, then remove the memory of this subreg.
1015 Otherwise, remove any memory of the entire register and
1016 all its subregs from the table. */
1017 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1018 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1019 remove_invalid_refs (i);
1020 else
1021 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1024 REG_IN_TABLE (i) = REG_TICK (i);
1025 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1026 return false;
1029 /* If X is a comparison or a COMPARE and either operand is a register
1030 that does not have a quantity, give it one. This is so that a later
1031 call to record_jump_equiv won't cause X to be assigned a different
1032 hash code and not found in the table after that call.
1034 It is not necessary to do this here, since rehash_using_reg can
1035 fix up the table later, but doing this here eliminates the need to
1036 call that expensive function in the most common case where the only
1037 use of the register is in the comparison. */
1039 if (code == COMPARE || COMPARISON_P (x))
1041 if (REG_P (XEXP (x, 0))
1042 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1043 if (insert_regs (XEXP (x, 0), NULL, false))
1045 rehash_using_reg (XEXP (x, 0));
1046 changed = true;
1049 if (REG_P (XEXP (x, 1))
1050 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1051 if (insert_regs (XEXP (x, 1), NULL, false))
1053 rehash_using_reg (XEXP (x, 1));
1054 changed = true;
1058 fmt = GET_RTX_FORMAT (code);
1059 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1060 if (fmt[i] == 'e')
1062 if (mention_regs (XEXP (x, i)))
1063 changed = true;
1065 else if (fmt[i] == 'E')
1066 for (j = 0; j < XVECLEN (x, i); j++)
1067 if (mention_regs (XVECEXP (x, i, j)))
1068 changed = true;
1070 return changed;
1073 /* Update the register quantities for inserting X into the hash table
1074 with a value equivalent to CLASSP.
1075 (If the class does not contain a REG, it is irrelevant.)
1076 If MODIFIED is true, X is a destination; it is being modified.
1077 Note that delete_reg_equiv should be called on a register
1078 before insert_regs is done on that register with MODIFIED != 0.
1080 True value means that elements of reg_qty have changed
1081 so X's hash code may be different. */
1083 static bool
1084 insert_regs (rtx x, struct table_elt *classp, bool modified)
1086 if (REG_P (x))
1088 unsigned int regno = REGNO (x);
1089 int qty_valid;
1091 /* If REGNO is in the equivalence table already but is of the
1092 wrong mode for that equivalence, don't do anything here. */
1094 qty_valid = REGNO_QTY_VALID_P (regno);
1095 if (qty_valid)
1097 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1099 if (ent->mode != GET_MODE (x))
1100 return false;
1103 if (modified || ! qty_valid)
1105 if (classp)
1106 for (classp = classp->first_same_value;
1107 classp != 0;
1108 classp = classp->next_same_value)
1109 if (REG_P (classp->exp)
1110 && GET_MODE (classp->exp) == GET_MODE (x))
1112 unsigned c_regno = REGNO (classp->exp);
1114 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1116 /* Suppose that 5 is hard reg and 100 and 101 are
1117 pseudos. Consider
1119 (set (reg:si 100) (reg:si 5))
1120 (set (reg:si 5) (reg:si 100))
1121 (set (reg:di 101) (reg:di 5))
1123 We would now set REG_QTY (101) = REG_QTY (5), but the
1124 entry for 5 is in SImode. When we use this later in
1125 copy propagation, we get the register in wrong mode. */
1126 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1127 continue;
1129 make_regs_eqv (regno, c_regno);
1130 return true;
1133 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1134 than REG_IN_TABLE to find out if there was only a single preceding
1135 invalidation - for the SUBREG - or another one, which would be
1136 for the full register. However, if we find here that REG_TICK
1137 indicates that the register is invalid, it means that it has
1138 been invalidated in a separate operation. The SUBREG might be used
1139 now (then this is a recursive call), or we might use the full REG
1140 now and a SUBREG of it later. So bump up REG_TICK so that
1141 mention_regs will do the right thing. */
1142 if (! modified
1143 && REG_IN_TABLE (regno) >= 0
1144 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1145 REG_TICK (regno)++;
1146 make_new_qty (regno, GET_MODE (x));
1147 return true;
1150 return false;
1153 /* If X is a SUBREG, we will likely be inserting the inner register in the
1154 table. If that register doesn't have an assigned quantity number at
1155 this point but does later, the insertion that we will be doing now will
1156 not be accessible because its hash code will have changed. So assign
1157 a quantity number now. */
1159 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1160 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1162 insert_regs (SUBREG_REG (x), NULL, false);
1163 mention_regs (x);
1164 return true;
1166 else
1167 return mention_regs (x);
1171 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1172 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1173 CST is equal to an anchor. */
1175 static bool
1176 compute_const_anchors (rtx cst,
1177 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1178 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1180 unsigned HOST_WIDE_INT n = UINTVAL (cst);
1182 *lower_base = n & ~(targetm.const_anchor - 1);
1183 if ((unsigned HOST_WIDE_INT) *lower_base == n)
1184 return false;
1186 *upper_base = ((n + (targetm.const_anchor - 1))
1187 & ~(targetm.const_anchor - 1));
1188 *upper_offs = n - *upper_base;
1189 *lower_offs = n - *lower_base;
1190 return true;
1193 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1195 static void
1196 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1197 machine_mode mode)
1199 struct table_elt *elt;
1200 unsigned hash;
1201 rtx anchor_exp;
1202 rtx exp;
1204 anchor_exp = gen_int_mode (anchor, mode);
1205 hash = HASH (anchor_exp, mode);
1206 elt = lookup (anchor_exp, hash, mode);
1207 if (!elt)
1208 elt = insert (anchor_exp, NULL, hash, mode);
1210 exp = plus_constant (mode, reg, offs);
1211 /* REG has just been inserted and the hash codes recomputed. */
1212 mention_regs (exp);
1213 hash = HASH (exp, mode);
1215 /* Use the cost of the register rather than the whole expression. When
1216 looking up constant anchors we will further offset the corresponding
1217 expression therefore it does not make sense to prefer REGs over
1218 reg-immediate additions. Prefer instead the oldest expression. Also
1219 don't prefer pseudos over hard regs so that we derive constants in
1220 argument registers from other argument registers rather than from the
1221 original pseudo that was used to synthesize the constant. */
1222 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1225 /* The constant CST is equivalent to the register REG. Create
1226 equivalences between the two anchors of CST and the corresponding
1227 register-offset expressions using REG. */
1229 static void
1230 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1232 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1234 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1235 &upper_base, &upper_offs))
1236 return;
1238 /* Ignore anchors of value 0. Constants accessible from zero are
1239 simple. */
1240 if (lower_base != 0)
1241 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1243 if (upper_base != 0)
1244 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1247 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1248 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1249 valid expression. Return the cheapest and oldest of such expressions. In
1250 *OLD, return how old the resulting expression is compared to the other
1251 equivalent expressions. */
1253 static rtx
1254 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1255 unsigned *old)
1257 struct table_elt *elt;
1258 unsigned idx;
1259 struct table_elt *match_elt;
1260 rtx match;
1262 /* Find the cheapest and *oldest* expression to maximize the chance of
1263 reusing the same pseudo. */
1265 match_elt = NULL;
1266 match = NULL_RTX;
1267 for (elt = anchor_elt->first_same_value, idx = 0;
1268 elt;
1269 elt = elt->next_same_value, idx++)
1271 if (match_elt && CHEAPER (match_elt, elt))
1272 return match;
1274 if (REG_P (elt->exp)
1275 || (GET_CODE (elt->exp) == PLUS
1276 && REG_P (XEXP (elt->exp, 0))
1277 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1279 rtx x;
1281 /* Ignore expressions that are no longer valid. */
1282 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1283 continue;
1285 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1286 if (REG_P (x)
1287 || (GET_CODE (x) == PLUS
1288 && IN_RANGE (INTVAL (XEXP (x, 1)),
1289 -targetm.const_anchor,
1290 targetm.const_anchor - 1)))
1292 match = x;
1293 match_elt = elt;
1294 *old = idx;
1299 return match;
1302 /* Try to express the constant SRC_CONST using a register+offset expression
1303 derived from a constant anchor. Return it if successful or NULL_RTX,
1304 otherwise. */
1306 static rtx
1307 try_const_anchors (rtx src_const, machine_mode mode)
1309 struct table_elt *lower_elt, *upper_elt;
1310 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1311 rtx lower_anchor_rtx, upper_anchor_rtx;
1312 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1313 unsigned lower_old, upper_old;
1315 /* CONST_INT may be in various modes, avoid non-scalar-int mode. */
1316 if (!SCALAR_INT_MODE_P (mode))
1317 return NULL_RTX;
1319 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1320 &upper_base, &upper_offs))
1321 return NULL_RTX;
1323 lower_anchor_rtx = GEN_INT (lower_base);
1324 upper_anchor_rtx = GEN_INT (upper_base);
1325 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1326 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1328 if (lower_elt)
1329 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1330 if (upper_elt)
1331 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1333 if (!lower_exp)
1334 return upper_exp;
1335 if (!upper_exp)
1336 return lower_exp;
1338 /* Return the older expression. */
1339 return (upper_old > lower_old ? upper_exp : lower_exp);
1342 /* Look in or update the hash table. */
1344 /* Remove table element ELT from use in the table.
1345 HASH is its hash code, made using the HASH macro.
1346 It's an argument because often that is known in advance
1347 and we save much time not recomputing it. */
1349 static void
1350 remove_from_table (struct table_elt *elt, unsigned int hash)
1352 if (elt == 0)
1353 return;
1355 /* Mark this element as removed. See cse_insn. */
1356 elt->first_same_value = 0;
1358 /* Remove the table element from its equivalence class. */
1361 struct table_elt *prev = elt->prev_same_value;
1362 struct table_elt *next = elt->next_same_value;
1364 if (next)
1365 next->prev_same_value = prev;
1367 if (prev)
1368 prev->next_same_value = next;
1369 else
1371 struct table_elt *newfirst = next;
1372 while (next)
1374 next->first_same_value = newfirst;
1375 next = next->next_same_value;
1380 /* Remove the table element from its hash bucket. */
1383 struct table_elt *prev = elt->prev_same_hash;
1384 struct table_elt *next = elt->next_same_hash;
1386 if (next)
1387 next->prev_same_hash = prev;
1389 if (prev)
1390 prev->next_same_hash = next;
1391 else if (table[hash] == elt)
1392 table[hash] = next;
1393 else
1395 /* This entry is not in the proper hash bucket. This can happen
1396 when two classes were merged by `merge_equiv_classes'. Search
1397 for the hash bucket that it heads. This happens only very
1398 rarely, so the cost is acceptable. */
1399 for (hash = 0; hash < HASH_SIZE; hash++)
1400 if (table[hash] == elt)
1401 table[hash] = next;
1405 /* Remove the table element from its related-value circular chain. */
1407 if (elt->related_value != 0 && elt->related_value != elt)
1409 struct table_elt *p = elt->related_value;
1411 while (p->related_value != elt)
1412 p = p->related_value;
1413 p->related_value = elt->related_value;
1414 if (p->related_value == p)
1415 p->related_value = 0;
1418 /* Now add it to the free element chain. */
1419 elt->next_same_hash = free_element_chain;
1420 free_element_chain = elt;
1423 /* Same as above, but X is a pseudo-register. */
1425 static void
1426 remove_pseudo_from_table (rtx x, unsigned int hash)
1428 struct table_elt *elt;
1430 /* Because a pseudo-register can be referenced in more than one
1431 mode, we might have to remove more than one table entry. */
1432 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1433 remove_from_table (elt, hash);
1436 /* Look up X in the hash table and return its table element,
1437 or 0 if X is not in the table.
1439 MODE is the machine-mode of X, or if X is an integer constant
1440 with VOIDmode then MODE is the mode with which X will be used.
1442 Here we are satisfied to find an expression whose tree structure
1443 looks like X. */
1445 static struct table_elt *
1446 lookup (rtx x, unsigned int hash, machine_mode mode)
1448 struct table_elt *p;
1450 for (p = table[hash]; p; p = p->next_same_hash)
1451 if (mode == p->mode && ((x == p->exp && REG_P (x))
1452 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1453 return p;
1455 return 0;
1458 /* Like `lookup' but don't care whether the table element uses invalid regs.
1459 Also ignore discrepancies in the machine mode of a register. */
1461 static struct table_elt *
1462 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1464 struct table_elt *p;
1466 if (REG_P (x))
1468 unsigned int regno = REGNO (x);
1470 /* Don't check the machine mode when comparing registers;
1471 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1472 for (p = table[hash]; p; p = p->next_same_hash)
1473 if (REG_P (p->exp)
1474 && REGNO (p->exp) == regno)
1475 return p;
1477 else
1479 for (p = table[hash]; p; p = p->next_same_hash)
1480 if (mode == p->mode
1481 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1482 return p;
1485 return 0;
1488 /* Look for an expression equivalent to X and with code CODE.
1489 If one is found, return that expression. */
1491 static rtx
1492 lookup_as_function (rtx x, enum rtx_code code)
1494 struct table_elt *p
1495 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1497 if (p == 0)
1498 return 0;
1500 for (p = p->first_same_value; p; p = p->next_same_value)
1501 if (GET_CODE (p->exp) == code
1502 /* Make sure this is a valid entry in the table. */
1503 && exp_equiv_p (p->exp, p->exp, 1, false))
1504 return p->exp;
1506 return 0;
1509 /* Insert X in the hash table, assuming HASH is its hash code and
1510 CLASSP is an element of the class it should go in (or 0 if a new
1511 class should be made). COST is the code of X and reg_cost is the
1512 cost of registers in X. It is inserted at the proper position to
1513 keep the class in the order cheapest first.
1515 MODE is the machine-mode of X, or if X is an integer constant
1516 with VOIDmode then MODE is the mode with which X will be used.
1518 For elements of equal cheapness, the most recent one
1519 goes in front, except that the first element in the list
1520 remains first unless a cheaper element is added. The order of
1521 pseudo-registers does not matter, as canon_reg will be called to
1522 find the cheapest when a register is retrieved from the table.
1524 The in_memory field in the hash table element is set to 0.
1525 The caller must set it nonzero if appropriate.
1527 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1528 and if insert_regs returns a nonzero value
1529 you must then recompute its hash code before calling here.
1531 If necessary, update table showing constant values of quantities. */
1533 static struct table_elt *
1534 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1535 machine_mode mode, int cost, int reg_cost)
1537 struct table_elt *elt;
1539 /* If X is a register and we haven't made a quantity for it,
1540 something is wrong. */
1541 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1543 /* If X is a hard register, show it is being put in the table. */
1544 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1545 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1547 /* Put an element for X into the right hash bucket. */
1549 elt = free_element_chain;
1550 if (elt)
1551 free_element_chain = elt->next_same_hash;
1552 else
1553 elt = XNEW (struct table_elt);
1555 elt->exp = x;
1556 elt->canon_exp = NULL_RTX;
1557 elt->cost = cost;
1558 elt->regcost = reg_cost;
1559 elt->next_same_value = 0;
1560 elt->prev_same_value = 0;
1561 elt->next_same_hash = table[hash];
1562 elt->prev_same_hash = 0;
1563 elt->related_value = 0;
1564 elt->in_memory = 0;
1565 elt->mode = mode;
1566 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1568 if (table[hash])
1569 table[hash]->prev_same_hash = elt;
1570 table[hash] = elt;
1572 /* Put it into the proper value-class. */
1573 if (classp)
1575 classp = classp->first_same_value;
1576 if (CHEAPER (elt, classp))
1577 /* Insert at the head of the class. */
1579 struct table_elt *p;
1580 elt->next_same_value = classp;
1581 classp->prev_same_value = elt;
1582 elt->first_same_value = elt;
1584 for (p = classp; p; p = p->next_same_value)
1585 p->first_same_value = elt;
1587 else
1589 /* Insert not at head of the class. */
1590 /* Put it after the last element cheaper than X. */
1591 struct table_elt *p, *next;
1593 for (p = classp;
1594 (next = p->next_same_value) && CHEAPER (next, elt);
1595 p = next)
1598 /* Put it after P and before NEXT. */
1599 elt->next_same_value = next;
1600 if (next)
1601 next->prev_same_value = elt;
1603 elt->prev_same_value = p;
1604 p->next_same_value = elt;
1605 elt->first_same_value = classp;
1608 else
1609 elt->first_same_value = elt;
1611 /* If this is a constant being set equivalent to a register or a register
1612 being set equivalent to a constant, note the constant equivalence.
1614 If this is a constant, it cannot be equivalent to a different constant,
1615 and a constant is the only thing that can be cheaper than a register. So
1616 we know the register is the head of the class (before the constant was
1617 inserted).
1619 If this is a register that is not already known equivalent to a
1620 constant, we must check the entire class.
1622 If this is a register that is already known equivalent to an insn,
1623 update the qtys `const_insn' to show that `this_insn' is the latest
1624 insn making that quantity equivalent to the constant. */
1626 if (elt->is_const && classp && REG_P (classp->exp)
1627 && !REG_P (x))
1629 int exp_q = REG_QTY (REGNO (classp->exp));
1630 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1632 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1633 exp_ent->const_insn = this_insn;
1636 else if (REG_P (x)
1637 && classp
1638 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1639 && ! elt->is_const)
1641 struct table_elt *p;
1643 for (p = classp; p != 0; p = p->next_same_value)
1645 if (p->is_const && !REG_P (p->exp))
1647 int x_q = REG_QTY (REGNO (x));
1648 struct qty_table_elem *x_ent = &qty_table[x_q];
1650 x_ent->const_rtx
1651 = gen_lowpart (GET_MODE (x), p->exp);
1652 x_ent->const_insn = this_insn;
1653 break;
1658 else if (REG_P (x)
1659 && qty_table[REG_QTY (REGNO (x))].const_rtx
1660 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1661 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1663 /* If this is a constant with symbolic value,
1664 and it has a term with an explicit integer value,
1665 link it up with related expressions. */
1666 if (GET_CODE (x) == CONST)
1668 rtx subexp = get_related_value (x);
1669 unsigned subhash;
1670 struct table_elt *subelt, *subelt_prev;
1672 if (subexp != 0)
1674 /* Get the integer-free subexpression in the hash table. */
1675 subhash = SAFE_HASH (subexp, mode);
1676 subelt = lookup (subexp, subhash, mode);
1677 if (subelt == 0)
1678 subelt = insert (subexp, NULL, subhash, mode);
1679 /* Initialize SUBELT's circular chain if it has none. */
1680 if (subelt->related_value == 0)
1681 subelt->related_value = subelt;
1682 /* Find the element in the circular chain that precedes SUBELT. */
1683 subelt_prev = subelt;
1684 while (subelt_prev->related_value != subelt)
1685 subelt_prev = subelt_prev->related_value;
1686 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1687 This way the element that follows SUBELT is the oldest one. */
1688 elt->related_value = subelt_prev->related_value;
1689 subelt_prev->related_value = elt;
1693 return elt;
1696 /* Wrap insert_with_costs by passing the default costs. */
1698 static struct table_elt *
1699 insert (rtx x, struct table_elt *classp, unsigned int hash,
1700 machine_mode mode)
1702 return insert_with_costs (x, classp, hash, mode,
1703 COST (x, mode), approx_reg_cost (x));
1707 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1708 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1709 the two classes equivalent.
1711 CLASS1 will be the surviving class; CLASS2 should not be used after this
1712 call.
1714 Any invalid entries in CLASS2 will not be copied. */
1716 static void
1717 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1719 struct table_elt *elt, *next, *new_elt;
1721 /* Ensure we start with the head of the classes. */
1722 class1 = class1->first_same_value;
1723 class2 = class2->first_same_value;
1725 /* If they were already equal, forget it. */
1726 if (class1 == class2)
1727 return;
1729 for (elt = class2; elt; elt = next)
1731 unsigned int hash;
1732 rtx exp = elt->exp;
1733 machine_mode mode = elt->mode;
1735 next = elt->next_same_value;
1737 /* Remove old entry, make a new one in CLASS1's class.
1738 Don't do this for invalid entries as we cannot find their
1739 hash code (it also isn't necessary). */
1740 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1742 bool need_rehash = false;
1744 hash_arg_in_memory = 0;
1745 hash = HASH (exp, mode);
1747 if (REG_P (exp))
1749 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1750 delete_reg_equiv (REGNO (exp));
1753 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1754 remove_pseudo_from_table (exp, hash);
1755 else
1756 remove_from_table (elt, hash);
1758 if (insert_regs (exp, class1, false) || need_rehash)
1760 rehash_using_reg (exp);
1761 hash = HASH (exp, mode);
1763 new_elt = insert (exp, class1, hash, mode);
1764 new_elt->in_memory = hash_arg_in_memory;
1765 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1766 new_elt->cost = MAX_COST;
1771 /* Flush the entire hash table. */
1773 static void
1774 flush_hash_table (void)
1776 int i;
1777 struct table_elt *p;
1779 for (i = 0; i < HASH_SIZE; i++)
1780 for (p = table[i]; p; p = table[i])
1782 /* Note that invalidate can remove elements
1783 after P in the current hash chain. */
1784 if (REG_P (p->exp))
1785 invalidate (p->exp, VOIDmode);
1786 else
1787 remove_from_table (p, i);
1791 /* Check whether an anti dependence exists between X and EXP. MODE and
1792 ADDR are as for canon_anti_dependence. */
1794 static bool
1795 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1797 subrtx_iterator::array_type array;
1798 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1800 const_rtx x = *iter;
1801 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1802 return true;
1804 return false;
1807 /* Remove from the hash table, or mark as invalid, all expressions whose
1808 values could be altered by storing in register X. */
1810 static void
1811 invalidate_reg (rtx x)
1813 gcc_assert (GET_CODE (x) == REG);
1815 /* If X is a register, dependencies on its contents are recorded
1816 through the qty number mechanism. Just change the qty number of
1817 the register, mark it as invalid for expressions that refer to it,
1818 and remove it itself. */
1819 unsigned int regno = REGNO (x);
1820 unsigned int hash = HASH (x, GET_MODE (x));
1822 /* Remove REGNO from any quantity list it might be on and indicate
1823 that its value might have changed. If it is a pseudo, remove its
1824 entry from the hash table.
1826 For a hard register, we do the first two actions above for any
1827 additional hard registers corresponding to X. Then, if any of these
1828 registers are in the table, we must remove any REG entries that
1829 overlap these registers. */
1831 delete_reg_equiv (regno);
1832 REG_TICK (regno)++;
1833 SUBREG_TICKED (regno) = -1;
1835 if (regno >= FIRST_PSEUDO_REGISTER)
1836 remove_pseudo_from_table (x, hash);
1837 else
1839 HOST_WIDE_INT in_table = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1840 unsigned int endregno = END_REGNO (x);
1841 unsigned int rn;
1842 struct table_elt *p, *next;
1844 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1846 for (rn = regno + 1; rn < endregno; rn++)
1848 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1849 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1850 delete_reg_equiv (rn);
1851 REG_TICK (rn)++;
1852 SUBREG_TICKED (rn) = -1;
1855 if (in_table)
1856 for (hash = 0; hash < HASH_SIZE; hash++)
1857 for (p = table[hash]; p; p = next)
1859 next = p->next_same_hash;
1861 if (!REG_P (p->exp) || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1862 continue;
1864 unsigned int tregno = REGNO (p->exp);
1865 unsigned int tendregno = END_REGNO (p->exp);
1866 if (tendregno > regno && tregno < endregno)
1867 remove_from_table (p, hash);
1872 /* Remove from the hash table, or mark as invalid, all expressions whose
1873 values could be altered by storing in X. X is a register, a subreg, or
1874 a memory reference with nonvarying address (because, when a memory
1875 reference with a varying address is stored in, all memory references are
1876 removed by invalidate_memory so specific invalidation is superfluous).
1877 FULL_MODE, if not VOIDmode, indicates that this much should be
1878 invalidated instead of just the amount indicated by the mode of X. This
1879 is only used for bitfield stores into memory.
1881 A nonvarying address may be just a register or just a symbol reference,
1882 or it may be either of those plus a numeric offset. */
1884 static void
1885 invalidate (rtx x, machine_mode full_mode)
1887 int i;
1888 struct table_elt *p;
1889 rtx addr;
1891 switch (GET_CODE (x))
1893 case REG:
1894 invalidate_reg (x);
1895 return;
1897 case SUBREG:
1898 invalidate (SUBREG_REG (x), VOIDmode);
1899 return;
1901 case PARALLEL:
1902 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1903 invalidate (XVECEXP (x, 0, i), VOIDmode);
1904 return;
1906 case EXPR_LIST:
1907 /* This is part of a disjoint return value; extract the location in
1908 question ignoring the offset. */
1909 invalidate (XEXP (x, 0), VOIDmode);
1910 return;
1912 case MEM:
1913 addr = canon_rtx (get_addr (XEXP (x, 0)));
1914 /* Calculate the canonical version of X here so that
1915 true_dependence doesn't generate new RTL for X on each call. */
1916 x = canon_rtx (x);
1918 /* Remove all hash table elements that refer to overlapping pieces of
1919 memory. */
1920 if (full_mode == VOIDmode)
1921 full_mode = GET_MODE (x);
1923 for (i = 0; i < HASH_SIZE; i++)
1925 struct table_elt *next;
1927 for (p = table[i]; p; p = next)
1929 next = p->next_same_hash;
1930 if (p->in_memory)
1932 /* Just canonicalize the expression once;
1933 otherwise each time we call invalidate
1934 true_dependence will canonicalize the
1935 expression again. */
1936 if (!p->canon_exp)
1937 p->canon_exp = canon_rtx (p->exp);
1938 if (check_dependence (p->canon_exp, x, full_mode, addr))
1939 remove_from_table (p, i);
1943 return;
1945 default:
1946 gcc_unreachable ();
1950 /* Invalidate DEST. Used when DEST is not going to be added
1951 into the hash table for some reason, e.g. do_not_record
1952 flagged on it. */
1954 static void
1955 invalidate_dest (rtx dest)
1957 if (REG_P (dest)
1958 || GET_CODE (dest) == SUBREG
1959 || MEM_P (dest))
1960 invalidate (dest, VOIDmode);
1961 else if (GET_CODE (dest) == STRICT_LOW_PART
1962 || GET_CODE (dest) == ZERO_EXTRACT)
1963 invalidate (XEXP (dest, 0), GET_MODE (dest));
1966 /* Remove all expressions that refer to register REGNO,
1967 since they are already invalid, and we are about to
1968 mark that register valid again and don't want the old
1969 expressions to reappear as valid. */
1971 static void
1972 remove_invalid_refs (unsigned int regno)
1974 unsigned int i;
1975 struct table_elt *p, *next;
1977 for (i = 0; i < HASH_SIZE; i++)
1978 for (p = table[i]; p; p = next)
1980 next = p->next_same_hash;
1981 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1982 remove_from_table (p, i);
1986 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1987 and mode MODE. */
1988 static void
1989 remove_invalid_subreg_refs (unsigned int regno, poly_uint64 offset,
1990 machine_mode mode)
1992 unsigned int i;
1993 struct table_elt *p, *next;
1995 for (i = 0; i < HASH_SIZE; i++)
1996 for (p = table[i]; p; p = next)
1998 rtx exp = p->exp;
1999 next = p->next_same_hash;
2001 if (!REG_P (exp)
2002 && (GET_CODE (exp) != SUBREG
2003 || !REG_P (SUBREG_REG (exp))
2004 || REGNO (SUBREG_REG (exp)) != regno
2005 || ranges_maybe_overlap_p (SUBREG_BYTE (exp),
2006 GET_MODE_SIZE (GET_MODE (exp)),
2007 offset, GET_MODE_SIZE (mode)))
2008 && refers_to_regno_p (regno, p->exp))
2009 remove_from_table (p, i);
2013 /* Recompute the hash codes of any valid entries in the hash table that
2014 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2016 This is called when we make a jump equivalence. */
2018 static void
2019 rehash_using_reg (rtx x)
2021 unsigned int i;
2022 struct table_elt *p, *next;
2023 unsigned hash;
2025 if (GET_CODE (x) == SUBREG)
2026 x = SUBREG_REG (x);
2028 /* If X is not a register or if the register is known not to be in any
2029 valid entries in the table, we have no work to do. */
2031 if (!REG_P (x)
2032 || REG_IN_TABLE (REGNO (x)) < 0
2033 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2034 return;
2036 /* Scan all hash chains looking for valid entries that mention X.
2037 If we find one and it is in the wrong hash chain, move it. */
2039 for (i = 0; i < HASH_SIZE; i++)
2040 for (p = table[i]; p; p = next)
2042 next = p->next_same_hash;
2043 if (reg_mentioned_p (x, p->exp)
2044 && exp_equiv_p (p->exp, p->exp, 1, false)
2045 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2047 if (p->next_same_hash)
2048 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2050 if (p->prev_same_hash)
2051 p->prev_same_hash->next_same_hash = p->next_same_hash;
2052 else
2053 table[i] = p->next_same_hash;
2055 p->next_same_hash = table[hash];
2056 p->prev_same_hash = 0;
2057 if (table[hash])
2058 table[hash]->prev_same_hash = p;
2059 table[hash] = p;
2064 /* Remove from the hash table any expression that is a call-clobbered
2065 register in INSN. Also update their TICK values. */
2067 static void
2068 invalidate_for_call (rtx_insn *insn)
2070 unsigned int regno;
2071 unsigned hash;
2072 struct table_elt *p, *next;
2073 int in_table = 0;
2074 hard_reg_set_iterator hrsi;
2076 /* Go through all the hard registers. For each that might be clobbered
2077 in call insn INSN, remove the register from quantity chains and update
2078 reg_tick if defined. Also see if any of these registers is currently
2079 in the table.
2081 ??? We could be more precise for partially-clobbered registers,
2082 and only invalidate values that actually occupy the clobbered part
2083 of the registers. It doesn't seem worth the effort though, since
2084 we shouldn't see this situation much before RA. Whatever choice
2085 we make here has to be consistent with the table walk below,
2086 so any change to this test will require a change there too. */
2087 HARD_REG_SET callee_clobbers
2088 = insn_callee_abi (insn).full_and_partial_reg_clobbers ();
2089 EXECUTE_IF_SET_IN_HARD_REG_SET (callee_clobbers, 0, regno, hrsi)
2091 delete_reg_equiv (regno);
2092 if (REG_TICK (regno) >= 0)
2094 REG_TICK (regno)++;
2095 SUBREG_TICKED (regno) = -1;
2097 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2100 /* In the case where we have no call-clobbered hard registers in the
2101 table, we are done. Otherwise, scan the table and remove any
2102 entry that overlaps a call-clobbered register. */
2104 if (in_table)
2105 for (hash = 0; hash < HASH_SIZE; hash++)
2106 for (p = table[hash]; p; p = next)
2108 next = p->next_same_hash;
2110 if (!REG_P (p->exp)
2111 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2112 continue;
2114 /* This must use the same test as above rather than the
2115 more accurate clobbers_reg_p. */
2116 if (overlaps_hard_reg_set_p (callee_clobbers, GET_MODE (p->exp),
2117 REGNO (p->exp)))
2118 remove_from_table (p, hash);
2122 /* Given an expression X of type CONST,
2123 and ELT which is its table entry (or 0 if it
2124 is not in the hash table),
2125 return an alternate expression for X as a register plus integer.
2126 If none can be found, return 0. */
2128 static rtx
2129 use_related_value (rtx x, struct table_elt *elt)
2131 struct table_elt *relt = 0;
2132 struct table_elt *p, *q;
2133 HOST_WIDE_INT offset;
2135 /* First, is there anything related known?
2136 If we have a table element, we can tell from that.
2137 Otherwise, must look it up. */
2139 if (elt != 0 && elt->related_value != 0)
2140 relt = elt;
2141 else if (elt == 0 && GET_CODE (x) == CONST)
2143 rtx subexp = get_related_value (x);
2144 if (subexp != 0)
2145 relt = lookup (subexp,
2146 SAFE_HASH (subexp, GET_MODE (subexp)),
2147 GET_MODE (subexp));
2150 if (relt == 0)
2151 return 0;
2153 /* Search all related table entries for one that has an
2154 equivalent register. */
2156 p = relt;
2157 while (1)
2159 /* This loop is strange in that it is executed in two different cases.
2160 The first is when X is already in the table. Then it is searching
2161 the RELATED_VALUE list of X's class (RELT). The second case is when
2162 X is not in the table. Then RELT points to a class for the related
2163 value.
2165 Ensure that, whatever case we are in, that we ignore classes that have
2166 the same value as X. */
2168 if (rtx_equal_p (x, p->exp))
2169 q = 0;
2170 else
2171 for (q = p->first_same_value; q; q = q->next_same_value)
2172 if (REG_P (q->exp))
2173 break;
2175 if (q)
2176 break;
2178 p = p->related_value;
2180 /* We went all the way around, so there is nothing to be found.
2181 Alternatively, perhaps RELT was in the table for some other reason
2182 and it has no related values recorded. */
2183 if (p == relt || p == 0)
2184 break;
2187 if (q == 0)
2188 return 0;
2190 offset = (get_integer_term (x) - get_integer_term (p->exp));
2191 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2192 return plus_constant (q->mode, q->exp, offset);
2196 /* Hash a string. Just add its bytes up. */
2197 static inline unsigned
2198 hash_rtx_string (const char *ps)
2200 unsigned hash = 0;
2201 const unsigned char *p = (const unsigned char *) ps;
2203 if (p)
2204 while (*p)
2205 hash += *p++;
2207 return hash;
2210 /* Hash an rtx. We are careful to make sure the value is never negative.
2211 Equivalent registers hash identically.
2212 MODE is used in hashing for CONST_INTs only;
2213 otherwise the mode of X is used.
2215 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2217 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2218 a MEM rtx which does not have the MEM_READONLY_P flag set.
2220 Note that cse_insn knows that the hash code of a MEM expression
2221 is just (int) MEM plus the hash code of the address.
2223 Call CB on each rtx if CB is not NULL.
2224 When the callback returns true, we continue with the new rtx. */
2226 unsigned
2227 hash_rtx (const_rtx x, machine_mode mode,
2228 int *do_not_record_p, int *hash_arg_in_memory_p,
2229 bool have_reg_qty, hash_rtx_callback_function cb)
2231 int i, j;
2232 unsigned hash = 0;
2233 enum rtx_code code;
2234 const char *fmt;
2235 machine_mode newmode;
2236 rtx newx;
2238 /* Used to turn recursion into iteration. We can't rely on GCC's
2239 tail-recursion elimination since we need to keep accumulating values
2240 in HASH. */
2241 repeat:
2242 if (x == 0)
2243 return hash;
2245 /* Invoke the callback first. */
2246 if (cb != NULL
2247 && ((*cb) (x, mode, &newx, &newmode)))
2249 hash += hash_rtx (newx, newmode, do_not_record_p,
2250 hash_arg_in_memory_p, have_reg_qty, cb);
2251 return hash;
2254 code = GET_CODE (x);
2255 switch (code)
2257 case REG:
2259 unsigned int regno = REGNO (x);
2261 if (do_not_record_p && !reload_completed)
2263 /* On some machines, we can't record any non-fixed hard register,
2264 because extending its life will cause reload problems. We
2265 consider ap, fp, sp, gp to be fixed for this purpose.
2267 We also consider CCmode registers to be fixed for this purpose;
2268 failure to do so leads to failure to simplify 0<100 type of
2269 conditionals.
2271 On all machines, we can't record any global registers.
2272 Nor should we record any register that is in a small
2273 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2274 bool record;
2276 if (regno >= FIRST_PSEUDO_REGISTER)
2277 record = true;
2278 else if (x == frame_pointer_rtx
2279 || x == hard_frame_pointer_rtx
2280 || x == arg_pointer_rtx
2281 || x == stack_pointer_rtx
2282 || x == pic_offset_table_rtx)
2283 record = true;
2284 else if (global_regs[regno])
2285 record = false;
2286 else if (fixed_regs[regno])
2287 record = true;
2288 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2289 record = true;
2290 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2291 record = false;
2292 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2293 record = false;
2294 else
2295 record = true;
2297 if (!record)
2299 *do_not_record_p = 1;
2300 return 0;
2304 hash += ((unsigned int) REG << 7);
2305 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2306 return hash;
2309 /* We handle SUBREG of a REG specially because the underlying
2310 reg changes its hash value with every value change; we don't
2311 want to have to forget unrelated subregs when one subreg changes. */
2312 case SUBREG:
2314 if (REG_P (SUBREG_REG (x)))
2316 hash += (((unsigned int) SUBREG << 7)
2317 + REGNO (SUBREG_REG (x))
2318 + (constant_lower_bound (SUBREG_BYTE (x))
2319 / UNITS_PER_WORD));
2320 return hash;
2322 break;
2325 case CONST_INT:
2326 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2327 + (unsigned int) INTVAL (x));
2328 return hash;
2330 case CONST_WIDE_INT:
2331 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2332 hash += CONST_WIDE_INT_ELT (x, i);
2333 return hash;
2335 case CONST_POLY_INT:
2337 inchash::hash h;
2338 h.add_int (hash);
2339 for (unsigned int i = 0; i < NUM_POLY_INT_COEFFS; ++i)
2340 h.add_wide_int (CONST_POLY_INT_COEFFS (x)[i]);
2341 return h.end ();
2344 case CONST_DOUBLE:
2345 /* This is like the general case, except that it only counts
2346 the integers representing the constant. */
2347 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2348 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2349 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2350 + (unsigned int) CONST_DOUBLE_HIGH (x));
2351 else
2352 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2353 return hash;
2355 case CONST_FIXED:
2356 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2357 hash += fixed_hash (CONST_FIXED_VALUE (x));
2358 return hash;
2360 case CONST_VECTOR:
2362 int units;
2363 rtx elt;
2365 units = const_vector_encoded_nelts (x);
2367 for (i = 0; i < units; ++i)
2369 elt = CONST_VECTOR_ENCODED_ELT (x, i);
2370 hash += hash_rtx (elt, GET_MODE (elt),
2371 do_not_record_p, hash_arg_in_memory_p,
2372 have_reg_qty, cb);
2375 return hash;
2378 /* Assume there is only one rtx object for any given label. */
2379 case LABEL_REF:
2380 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2381 differences and differences between each stage's debugging dumps. */
2382 hash += (((unsigned int) LABEL_REF << 7)
2383 + CODE_LABEL_NUMBER (label_ref_label (x)));
2384 return hash;
2386 case SYMBOL_REF:
2388 /* Don't hash on the symbol's address to avoid bootstrap differences.
2389 Different hash values may cause expressions to be recorded in
2390 different orders and thus different registers to be used in the
2391 final assembler. This also avoids differences in the dump files
2392 between various stages. */
2393 unsigned int h = 0;
2394 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2396 while (*p)
2397 h += (h << 7) + *p++; /* ??? revisit */
2399 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2400 return hash;
2403 case MEM:
2404 /* We don't record if marked volatile or if BLKmode since we don't
2405 know the size of the move. */
2406 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2408 *do_not_record_p = 1;
2409 return 0;
2411 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2412 *hash_arg_in_memory_p = 1;
2414 /* Now that we have already found this special case,
2415 might as well speed it up as much as possible. */
2416 hash += (unsigned) MEM;
2417 x = XEXP (x, 0);
2418 goto repeat;
2420 case USE:
2421 /* A USE that mentions non-volatile memory needs special
2422 handling since the MEM may be BLKmode which normally
2423 prevents an entry from being made. Pure calls are
2424 marked by a USE which mentions BLKmode memory.
2425 See calls.cc:emit_call_1. */
2426 if (MEM_P (XEXP (x, 0))
2427 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2429 hash += (unsigned) USE;
2430 x = XEXP (x, 0);
2432 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2433 *hash_arg_in_memory_p = 1;
2435 /* Now that we have already found this special case,
2436 might as well speed it up as much as possible. */
2437 hash += (unsigned) MEM;
2438 x = XEXP (x, 0);
2439 goto repeat;
2441 break;
2443 case PRE_DEC:
2444 case PRE_INC:
2445 case POST_DEC:
2446 case POST_INC:
2447 case PRE_MODIFY:
2448 case POST_MODIFY:
2449 case PC:
2450 case CALL:
2451 case UNSPEC_VOLATILE:
2452 if (do_not_record_p) {
2453 *do_not_record_p = 1;
2454 return 0;
2456 else
2457 return hash;
2458 break;
2460 case ASM_OPERANDS:
2461 if (do_not_record_p && MEM_VOLATILE_P (x))
2463 *do_not_record_p = 1;
2464 return 0;
2466 else
2468 /* We don't want to take the filename and line into account. */
2469 hash += (unsigned) code + (unsigned) GET_MODE (x)
2470 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2471 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2472 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2474 if (ASM_OPERANDS_INPUT_LENGTH (x))
2476 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2478 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2479 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2480 do_not_record_p, hash_arg_in_memory_p,
2481 have_reg_qty, cb)
2482 + hash_rtx_string
2483 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2486 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2487 x = ASM_OPERANDS_INPUT (x, 0);
2488 mode = GET_MODE (x);
2489 goto repeat;
2492 return hash;
2494 break;
2496 default:
2497 break;
2500 i = GET_RTX_LENGTH (code) - 1;
2501 hash += (unsigned) code + (unsigned) GET_MODE (x);
2502 fmt = GET_RTX_FORMAT (code);
2503 for (; i >= 0; i--)
2505 switch (fmt[i])
2507 case 'e':
2508 /* If we are about to do the last recursive call
2509 needed at this level, change it into iteration.
2510 This function is called enough to be worth it. */
2511 if (i == 0)
2513 x = XEXP (x, i);
2514 goto repeat;
2517 hash += hash_rtx (XEXP (x, i), VOIDmode, do_not_record_p,
2518 hash_arg_in_memory_p,
2519 have_reg_qty, cb);
2520 break;
2522 case 'E':
2523 for (j = 0; j < XVECLEN (x, i); j++)
2524 hash += hash_rtx (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2525 hash_arg_in_memory_p,
2526 have_reg_qty, cb);
2527 break;
2529 case 's':
2530 hash += hash_rtx_string (XSTR (x, i));
2531 break;
2533 case 'i':
2534 hash += (unsigned int) XINT (x, i);
2535 break;
2537 case 'p':
2538 hash += constant_lower_bound (SUBREG_BYTE (x));
2539 break;
2541 case '0': case 't':
2542 /* Unused. */
2543 break;
2545 default:
2546 gcc_unreachable ();
2550 return hash;
2553 /* Hash an rtx X for cse via hash_rtx.
2554 Stores 1 in do_not_record if any subexpression is volatile.
2555 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2556 does not have the MEM_READONLY_P flag set. */
2558 static inline unsigned
2559 canon_hash (rtx x, machine_mode mode)
2561 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2564 /* Like canon_hash but with no side effects, i.e. do_not_record
2565 and hash_arg_in_memory are not changed. */
2567 static inline unsigned
2568 safe_hash (rtx x, machine_mode mode)
2570 int dummy_do_not_record;
2571 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2574 /* Return true iff X and Y would canonicalize into the same thing,
2575 without actually constructing the canonicalization of either one.
2576 If VALIDATE is nonzero,
2577 we assume X is an expression being processed from the rtl
2578 and Y was found in the hash table. We check register refs
2579 in Y for being marked as valid.
2581 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2583 bool
2584 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2586 int i, j;
2587 enum rtx_code code;
2588 const char *fmt;
2590 /* Note: it is incorrect to assume an expression is equivalent to itself
2591 if VALIDATE is nonzero. */
2592 if (x == y && !validate)
2593 return true;
2595 if (x == 0 || y == 0)
2596 return x == y;
2598 code = GET_CODE (x);
2599 if (code != GET_CODE (y))
2600 return false;
2602 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2603 if (GET_MODE (x) != GET_MODE (y))
2604 return false;
2606 /* MEMs referring to different address space are not equivalent. */
2607 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2608 return false;
2610 switch (code)
2612 case PC:
2613 CASE_CONST_UNIQUE:
2614 return x == y;
2616 case CONST_VECTOR:
2617 if (!same_vector_encodings_p (x, y))
2618 return false;
2619 break;
2621 case LABEL_REF:
2622 return label_ref_label (x) == label_ref_label (y);
2624 case SYMBOL_REF:
2625 return XSTR (x, 0) == XSTR (y, 0);
2627 case REG:
2628 if (for_gcse)
2629 return REGNO (x) == REGNO (y);
2630 else
2632 unsigned int regno = REGNO (y);
2633 unsigned int i;
2634 unsigned int endregno = END_REGNO (y);
2636 /* If the quantities are not the same, the expressions are not
2637 equivalent. If there are and we are not to validate, they
2638 are equivalent. Otherwise, ensure all regs are up-to-date. */
2640 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2641 return false;
2643 if (! validate)
2644 return true;
2646 for (i = regno; i < endregno; i++)
2647 if (REG_IN_TABLE (i) != REG_TICK (i))
2648 return false;
2650 return true;
2653 case MEM:
2654 if (for_gcse)
2656 /* A volatile mem should not be considered equivalent to any
2657 other. */
2658 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2659 return false;
2661 /* Can't merge two expressions in different alias sets, since we
2662 can decide that the expression is transparent in a block when
2663 it isn't, due to it being set with the different alias set.
2665 Also, can't merge two expressions with different MEM_ATTRS.
2666 They could e.g. be two different entities allocated into the
2667 same space on the stack (see e.g. PR25130). In that case, the
2668 MEM addresses can be the same, even though the two MEMs are
2669 absolutely not equivalent.
2671 But because really all MEM attributes should be the same for
2672 equivalent MEMs, we just use the invariant that MEMs that have
2673 the same attributes share the same mem_attrs data structure. */
2674 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2675 return false;
2677 /* If we are handling exceptions, we cannot consider two expressions
2678 with different trapping status as equivalent, because simple_mem
2679 might accept one and reject the other. */
2680 if (cfun->can_throw_non_call_exceptions
2681 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2682 return false;
2684 break;
2686 /* For commutative operations, check both orders. */
2687 case PLUS:
2688 case MULT:
2689 case AND:
2690 case IOR:
2691 case XOR:
2692 case NE:
2693 case EQ:
2694 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2695 validate, for_gcse)
2696 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2697 validate, for_gcse))
2698 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2699 validate, for_gcse)
2700 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2701 validate, for_gcse)));
2703 case ASM_OPERANDS:
2704 /* We don't use the generic code below because we want to
2705 disregard filename and line numbers. */
2707 /* A volatile asm isn't equivalent to any other. */
2708 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2709 return false;
2711 if (GET_MODE (x) != GET_MODE (y)
2712 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2713 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2714 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2715 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2716 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2717 return false;
2719 if (ASM_OPERANDS_INPUT_LENGTH (x))
2721 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2722 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2723 ASM_OPERANDS_INPUT (y, i),
2724 validate, for_gcse)
2725 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2726 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2727 return false;
2730 return true;
2732 default:
2733 break;
2736 /* Compare the elements. If any pair of corresponding elements
2737 fail to match, return 0 for the whole thing. */
2739 fmt = GET_RTX_FORMAT (code);
2740 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2742 switch (fmt[i])
2744 case 'e':
2745 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2746 validate, for_gcse))
2747 return false;
2748 break;
2750 case 'E':
2751 if (XVECLEN (x, i) != XVECLEN (y, i))
2752 return 0;
2753 for (j = 0; j < XVECLEN (x, i); j++)
2754 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2755 validate, for_gcse))
2756 return false;
2757 break;
2759 case 's':
2760 if (strcmp (XSTR (x, i), XSTR (y, i)))
2761 return false;
2762 break;
2764 case 'i':
2765 if (XINT (x, i) != XINT (y, i))
2766 return false;
2767 break;
2769 case 'w':
2770 if (XWINT (x, i) != XWINT (y, i))
2771 return false;
2772 break;
2774 case 'p':
2775 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
2776 return false;
2777 break;
2779 case '0':
2780 case 't':
2781 break;
2783 default:
2784 gcc_unreachable ();
2788 return true;
2791 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2792 the result if necessary. INSN is as for canon_reg. */
2794 static void
2795 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2797 if (*xloc)
2799 rtx new_rtx = canon_reg (*xloc, insn);
2801 /* If replacing pseudo with hard reg or vice versa, ensure the
2802 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2803 gcc_assert (insn && new_rtx);
2804 validate_change (insn, xloc, new_rtx, 1);
2808 /* Canonicalize an expression:
2809 replace each register reference inside it
2810 with the "oldest" equivalent register.
2812 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2813 after we make our substitution. The calls are made with IN_GROUP nonzero
2814 so apply_change_group must be called upon the outermost return from this
2815 function (unless INSN is zero). The result of apply_change_group can
2816 generally be discarded since the changes we are making are optional. */
2818 static rtx
2819 canon_reg (rtx x, rtx_insn *insn)
2821 int i;
2822 enum rtx_code code;
2823 const char *fmt;
2825 if (x == 0)
2826 return x;
2828 code = GET_CODE (x);
2829 switch (code)
2831 case PC:
2832 case CONST:
2833 CASE_CONST_ANY:
2834 case SYMBOL_REF:
2835 case LABEL_REF:
2836 case ADDR_VEC:
2837 case ADDR_DIFF_VEC:
2838 return x;
2840 case REG:
2842 int first;
2843 int q;
2844 struct qty_table_elem *ent;
2846 /* Never replace a hard reg, because hard regs can appear
2847 in more than one machine mode, and we must preserve the mode
2848 of each occurrence. Also, some hard regs appear in
2849 MEMs that are shared and mustn't be altered. Don't try to
2850 replace any reg that maps to a reg of class NO_REGS. */
2851 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2852 || ! REGNO_QTY_VALID_P (REGNO (x)))
2853 return x;
2855 q = REG_QTY (REGNO (x));
2856 ent = &qty_table[q];
2857 first = ent->first_reg;
2858 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2859 : REGNO_REG_CLASS (first) == NO_REGS ? x
2860 : gen_rtx_REG (ent->mode, first));
2863 default:
2864 break;
2867 fmt = GET_RTX_FORMAT (code);
2868 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2870 int j;
2872 if (fmt[i] == 'e')
2873 validate_canon_reg (&XEXP (x, i), insn);
2874 else if (fmt[i] == 'E')
2875 for (j = 0; j < XVECLEN (x, i); j++)
2876 validate_canon_reg (&XVECEXP (x, i, j), insn);
2879 return x;
2882 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2883 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2884 what values are being compared.
2886 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2887 actually being compared. For example, if *PARG1 was (reg:CC CC_REG) and
2888 *PARG2 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that
2889 were compared to produce (reg:CC CC_REG).
2891 The return value is the comparison operator and is either the code of
2892 A or the code corresponding to the inverse of the comparison. */
2894 static enum rtx_code
2895 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2896 machine_mode *pmode1, machine_mode *pmode2)
2898 rtx arg1, arg2;
2899 hash_set<rtx> *visited = NULL;
2900 /* Set nonzero when we find something of interest. */
2901 rtx x = NULL;
2903 arg1 = *parg1, arg2 = *parg2;
2905 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2907 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2909 int reverse_code = 0;
2910 struct table_elt *p = 0;
2912 /* Remember state from previous iteration. */
2913 if (x)
2915 if (!visited)
2916 visited = new hash_set<rtx>;
2917 visited->add (x);
2918 x = 0;
2921 /* If arg1 is a COMPARE, extract the comparison arguments from it. */
2923 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2924 x = arg1;
2926 /* If ARG1 is a comparison operator and CODE is testing for
2927 STORE_FLAG_VALUE, get the inner arguments. */
2929 else if (COMPARISON_P (arg1))
2931 #ifdef FLOAT_STORE_FLAG_VALUE
2932 REAL_VALUE_TYPE fsfv;
2933 #endif
2935 if (code == NE
2936 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2937 && code == LT && STORE_FLAG_VALUE == -1)
2938 #ifdef FLOAT_STORE_FLAG_VALUE
2939 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2940 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2941 REAL_VALUE_NEGATIVE (fsfv)))
2942 #endif
2944 x = arg1;
2945 else if (code == EQ
2946 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2947 && code == GE && STORE_FLAG_VALUE == -1)
2948 #ifdef FLOAT_STORE_FLAG_VALUE
2949 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2950 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2951 REAL_VALUE_NEGATIVE (fsfv)))
2952 #endif
2954 x = arg1, reverse_code = 1;
2957 /* ??? We could also check for
2959 (ne (and (eq (...) (const_int 1))) (const_int 0))
2961 and related forms, but let's wait until we see them occurring. */
2963 if (x == 0)
2964 /* Look up ARG1 in the hash table and see if it has an equivalence
2965 that lets us see what is being compared. */
2966 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2967 if (p)
2969 p = p->first_same_value;
2971 /* If what we compare is already known to be constant, that is as
2972 good as it gets.
2973 We need to break the loop in this case, because otherwise we
2974 can have an infinite loop when looking at a reg that is known
2975 to be a constant which is the same as a comparison of a reg
2976 against zero which appears later in the insn stream, which in
2977 turn is constant and the same as the comparison of the first reg
2978 against zero... */
2979 if (p->is_const)
2980 break;
2983 for (; p; p = p->next_same_value)
2985 machine_mode inner_mode = GET_MODE (p->exp);
2986 #ifdef FLOAT_STORE_FLAG_VALUE
2987 REAL_VALUE_TYPE fsfv;
2988 #endif
2990 /* If the entry isn't valid, skip it. */
2991 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2992 continue;
2994 /* If it's a comparison we've used before, skip it. */
2995 if (visited && visited->contains (p->exp))
2996 continue;
2998 if (GET_CODE (p->exp) == COMPARE
2999 /* Another possibility is that this machine has a compare insn
3000 that includes the comparison code. In that case, ARG1 would
3001 be equivalent to a comparison operation that would set ARG1 to
3002 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3003 ORIG_CODE is the actual comparison being done; if it is an EQ,
3004 we must reverse ORIG_CODE. On machine with a negative value
3005 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3006 || ((code == NE
3007 || (code == LT
3008 && val_signbit_known_set_p (inner_mode,
3009 STORE_FLAG_VALUE))
3010 #ifdef FLOAT_STORE_FLAG_VALUE
3011 || (code == LT
3012 && SCALAR_FLOAT_MODE_P (inner_mode)
3013 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3014 REAL_VALUE_NEGATIVE (fsfv)))
3015 #endif
3017 && COMPARISON_P (p->exp)))
3019 x = p->exp;
3020 break;
3022 else if ((code == EQ
3023 || (code == GE
3024 && val_signbit_known_set_p (inner_mode,
3025 STORE_FLAG_VALUE))
3026 #ifdef FLOAT_STORE_FLAG_VALUE
3027 || (code == GE
3028 && SCALAR_FLOAT_MODE_P (inner_mode)
3029 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3030 REAL_VALUE_NEGATIVE (fsfv)))
3031 #endif
3033 && COMPARISON_P (p->exp))
3035 reverse_code = 1;
3036 x = p->exp;
3037 break;
3040 /* If this non-trapping address, e.g. fp + constant, the
3041 equivalent is a better operand since it may let us predict
3042 the value of the comparison. */
3043 else if (!rtx_addr_can_trap_p (p->exp))
3045 arg1 = p->exp;
3046 continue;
3050 /* If we didn't find a useful equivalence for ARG1, we are done.
3051 Otherwise, set up for the next iteration. */
3052 if (x == 0)
3053 break;
3055 /* If we need to reverse the comparison, make sure that is
3056 possible -- we can't necessarily infer the value of GE from LT
3057 with floating-point operands. */
3058 if (reverse_code)
3060 enum rtx_code reversed = reversed_comparison_code (x, NULL);
3061 if (reversed == UNKNOWN)
3062 break;
3063 else
3064 code = reversed;
3066 else if (COMPARISON_P (x))
3067 code = GET_CODE (x);
3068 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3071 /* Return our results. Return the modes from before fold_rtx
3072 because fold_rtx might produce const_int, and then it's too late. */
3073 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3074 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3076 if (visited)
3077 delete visited;
3078 return code;
3081 /* If X is a nontrivial arithmetic operation on an argument for which
3082 a constant value can be determined, return the result of operating
3083 on that value, as a constant. Otherwise, return X, possibly with
3084 one or more operands changed to a forward-propagated constant.
3086 If X is a register whose contents are known, we do NOT return
3087 those contents here; equiv_constant is called to perform that task.
3088 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3090 INSN is the insn that we may be modifying. If it is 0, make a copy
3091 of X before modifying it. */
3093 static rtx
3094 fold_rtx (rtx x, rtx_insn *insn)
3096 enum rtx_code code;
3097 machine_mode mode;
3098 const char *fmt;
3099 int i;
3100 rtx new_rtx = 0;
3101 bool changed = false;
3102 poly_int64 xval;
3104 /* Operands of X. */
3105 /* Workaround -Wmaybe-uninitialized false positive during
3106 profiledbootstrap by initializing them. */
3107 rtx folded_arg0 = NULL_RTX;
3108 rtx folded_arg1 = NULL_RTX;
3110 /* Constant equivalents of first three operands of X;
3111 0 when no such equivalent is known. */
3112 rtx const_arg0;
3113 rtx const_arg1;
3114 rtx const_arg2;
3116 /* The mode of the first operand of X. We need this for sign and zero
3117 extends. */
3118 machine_mode mode_arg0;
3120 if (x == 0)
3121 return x;
3123 /* Try to perform some initial simplifications on X. */
3124 code = GET_CODE (x);
3125 switch (code)
3127 case MEM:
3128 case SUBREG:
3129 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3130 than it would in other contexts. Basically its mode does not
3131 signify the size of the object read. That information is carried
3132 by size operand. If we happen to have a MEM of the appropriate
3133 mode in our tables with a constant value we could simplify the
3134 extraction incorrectly if we allowed substitution of that value
3135 for the MEM. */
3136 case ZERO_EXTRACT:
3137 case SIGN_EXTRACT:
3138 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3139 return new_rtx;
3140 return x;
3142 case CONST:
3143 CASE_CONST_ANY:
3144 case SYMBOL_REF:
3145 case LABEL_REF:
3146 case REG:
3147 case PC:
3148 /* No use simplifying an EXPR_LIST
3149 since they are used only for lists of args
3150 in a function call's REG_EQUAL note. */
3151 case EXPR_LIST:
3152 return x;
3154 case ASM_OPERANDS:
3155 if (insn)
3157 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3158 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3159 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3161 return x;
3163 case CALL:
3164 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3165 return x;
3166 break;
3167 case VEC_SELECT:
3169 rtx trueop0 = XEXP (x, 0);
3170 mode = GET_MODE (trueop0);
3171 rtx trueop1 = XEXP (x, 1);
3172 /* If we select a low-part subreg, return that. */
3173 if (vec_series_lowpart_p (GET_MODE (x), mode, trueop1))
3175 rtx new_rtx = lowpart_subreg (GET_MODE (x), trueop0, mode);
3176 if (new_rtx != NULL_RTX)
3177 return new_rtx;
3181 /* Anything else goes through the loop below. */
3182 default:
3183 break;
3186 mode = GET_MODE (x);
3187 const_arg0 = 0;
3188 const_arg1 = 0;
3189 const_arg2 = 0;
3190 mode_arg0 = VOIDmode;
3192 /* Try folding our operands.
3193 Then see which ones have constant values known. */
3195 fmt = GET_RTX_FORMAT (code);
3196 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3197 if (fmt[i] == 'e')
3199 rtx folded_arg = XEXP (x, i), const_arg;
3200 machine_mode mode_arg = GET_MODE (folded_arg);
3202 switch (GET_CODE (folded_arg))
3204 case MEM:
3205 case REG:
3206 case SUBREG:
3207 const_arg = equiv_constant (folded_arg);
3208 break;
3210 case CONST:
3211 CASE_CONST_ANY:
3212 case SYMBOL_REF:
3213 case LABEL_REF:
3214 const_arg = folded_arg;
3215 break;
3217 default:
3218 folded_arg = fold_rtx (folded_arg, insn);
3219 const_arg = equiv_constant (folded_arg);
3220 break;
3223 /* For the first three operands, see if the operand
3224 is constant or equivalent to a constant. */
3225 switch (i)
3227 case 0:
3228 folded_arg0 = folded_arg;
3229 const_arg0 = const_arg;
3230 mode_arg0 = mode_arg;
3231 break;
3232 case 1:
3233 folded_arg1 = folded_arg;
3234 const_arg1 = const_arg;
3235 break;
3236 case 2:
3237 const_arg2 = const_arg;
3238 break;
3241 /* Pick the least expensive of the argument and an equivalent constant
3242 argument. */
3243 if (const_arg != 0
3244 && const_arg != folded_arg
3245 && (COST_IN (const_arg, mode_arg, code, i)
3246 <= COST_IN (folded_arg, mode_arg, code, i))
3248 /* It's not safe to substitute the operand of a conversion
3249 operator with a constant, as the conversion's identity
3250 depends upon the mode of its operand. This optimization
3251 is handled by the call to simplify_unary_operation. */
3252 && (GET_RTX_CLASS (code) != RTX_UNARY
3253 || GET_MODE (const_arg) == mode_arg0
3254 || (code != ZERO_EXTEND
3255 && code != SIGN_EXTEND
3256 && code != TRUNCATE
3257 && code != FLOAT_TRUNCATE
3258 && code != FLOAT_EXTEND
3259 && code != FLOAT
3260 && code != FIX
3261 && code != UNSIGNED_FLOAT
3262 && code != UNSIGNED_FIX)))
3263 folded_arg = const_arg;
3265 if (folded_arg == XEXP (x, i))
3266 continue;
3268 if (insn == NULL_RTX && !changed)
3269 x = copy_rtx (x);
3270 changed = true;
3271 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3274 if (changed)
3276 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3277 consistent with the order in X. */
3278 if (canonicalize_change_group (insn, x))
3280 std::swap (const_arg0, const_arg1);
3281 std::swap (folded_arg0, folded_arg1);
3284 apply_change_group ();
3287 /* If X is an arithmetic operation, see if we can simplify it. */
3289 switch (GET_RTX_CLASS (code))
3291 case RTX_UNARY:
3293 /* We can't simplify extension ops unless we know the
3294 original mode. */
3295 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3296 && mode_arg0 == VOIDmode)
3297 break;
3299 new_rtx = simplify_unary_operation (code, mode,
3300 const_arg0 ? const_arg0 : folded_arg0,
3301 mode_arg0);
3303 break;
3305 case RTX_COMPARE:
3306 case RTX_COMM_COMPARE:
3307 /* See what items are actually being compared and set FOLDED_ARG[01]
3308 to those values and CODE to the actual comparison code. If any are
3309 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3310 do anything if both operands are already known to be constant. */
3312 /* ??? Vector mode comparisons are not supported yet. */
3313 if (VECTOR_MODE_P (mode))
3314 break;
3316 if (const_arg0 == 0 || const_arg1 == 0)
3318 struct table_elt *p0, *p1;
3319 rtx true_rtx, false_rtx;
3320 machine_mode mode_arg1;
3322 if (SCALAR_FLOAT_MODE_P (mode))
3324 #ifdef FLOAT_STORE_FLAG_VALUE
3325 true_rtx = (const_double_from_real_value
3326 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3327 #else
3328 true_rtx = NULL_RTX;
3329 #endif
3330 false_rtx = CONST0_RTX (mode);
3332 else
3334 true_rtx = const_true_rtx;
3335 false_rtx = const0_rtx;
3338 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3339 &mode_arg0, &mode_arg1);
3341 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3342 what kinds of things are being compared, so we can't do
3343 anything with this comparison. */
3345 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3346 break;
3348 const_arg0 = equiv_constant (folded_arg0);
3349 const_arg1 = equiv_constant (folded_arg1);
3351 /* If we do not now have two constants being compared, see
3352 if we can nevertheless deduce some things about the
3353 comparison. */
3354 if (const_arg0 == 0 || const_arg1 == 0)
3356 if (const_arg1 != NULL)
3358 rtx cheapest_simplification;
3359 int cheapest_cost;
3360 rtx simp_result;
3361 struct table_elt *p;
3363 /* See if we can find an equivalent of folded_arg0
3364 that gets us a cheaper expression, possibly a
3365 constant through simplifications. */
3366 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3367 mode_arg0);
3369 if (p != NULL)
3371 cheapest_simplification = x;
3372 cheapest_cost = COST (x, mode);
3374 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3376 int cost;
3378 /* If the entry isn't valid, skip it. */
3379 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3380 continue;
3382 /* Try to simplify using this equivalence. */
3383 simp_result
3384 = simplify_relational_operation (code, mode,
3385 mode_arg0,
3386 p->exp,
3387 const_arg1);
3389 if (simp_result == NULL)
3390 continue;
3392 cost = COST (simp_result, mode);
3393 if (cost < cheapest_cost)
3395 cheapest_cost = cost;
3396 cheapest_simplification = simp_result;
3400 /* If we have a cheaper expression now, use that
3401 and try folding it further, from the top. */
3402 if (cheapest_simplification != x)
3403 return fold_rtx (copy_rtx (cheapest_simplification),
3404 insn);
3408 /* See if the two operands are the same. */
3410 if ((REG_P (folded_arg0)
3411 && REG_P (folded_arg1)
3412 && (REG_QTY (REGNO (folded_arg0))
3413 == REG_QTY (REGNO (folded_arg1))))
3414 || ((p0 = lookup (folded_arg0,
3415 SAFE_HASH (folded_arg0, mode_arg0),
3416 mode_arg0))
3417 && (p1 = lookup (folded_arg1,
3418 SAFE_HASH (folded_arg1, mode_arg0),
3419 mode_arg0))
3420 && p0->first_same_value == p1->first_same_value))
3421 folded_arg1 = folded_arg0;
3423 /* If FOLDED_ARG0 is a register, see if the comparison we are
3424 doing now is either the same as we did before or the reverse
3425 (we only check the reverse if not floating-point). */
3426 else if (REG_P (folded_arg0))
3428 int qty = REG_QTY (REGNO (folded_arg0));
3430 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3432 struct qty_table_elem *ent = &qty_table[qty];
3434 if ((comparison_dominates_p (ent->comparison_code, code)
3435 || (! FLOAT_MODE_P (mode_arg0)
3436 && comparison_dominates_p (ent->comparison_code,
3437 reverse_condition (code))))
3438 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3439 || (const_arg1
3440 && rtx_equal_p (ent->comparison_const,
3441 const_arg1))
3442 || (REG_P (folded_arg1)
3443 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3445 if (comparison_dominates_p (ent->comparison_code, code))
3447 if (true_rtx)
3448 return true_rtx;
3449 else
3450 break;
3452 else
3453 return false_rtx;
3460 /* If we are comparing against zero, see if the first operand is
3461 equivalent to an IOR with a constant. If so, we may be able to
3462 determine the result of this comparison. */
3463 if (const_arg1 == const0_rtx && !const_arg0)
3465 rtx y = lookup_as_function (folded_arg0, IOR);
3466 rtx inner_const;
3468 if (y != 0
3469 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3470 && CONST_INT_P (inner_const)
3471 && INTVAL (inner_const) != 0)
3472 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3476 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3477 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3478 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3479 op0, op1);
3481 break;
3483 case RTX_BIN_ARITH:
3484 case RTX_COMM_ARITH:
3485 switch (code)
3487 case PLUS:
3488 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3489 with that LABEL_REF as its second operand. If so, the result is
3490 the first operand of that MINUS. This handles switches with an
3491 ADDR_DIFF_VEC table. */
3492 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3494 rtx y
3495 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3496 : lookup_as_function (folded_arg0, MINUS);
3498 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3499 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg1))
3500 return XEXP (y, 0);
3502 /* Now try for a CONST of a MINUS like the above. */
3503 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3504 : lookup_as_function (folded_arg0, CONST))) != 0
3505 && GET_CODE (XEXP (y, 0)) == MINUS
3506 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3507 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg1))
3508 return XEXP (XEXP (y, 0), 0);
3511 /* Likewise if the operands are in the other order. */
3512 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3514 rtx y
3515 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3516 : lookup_as_function (folded_arg1, MINUS);
3518 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3519 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg0))
3520 return XEXP (y, 0);
3522 /* Now try for a CONST of a MINUS like the above. */
3523 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3524 : lookup_as_function (folded_arg1, CONST))) != 0
3525 && GET_CODE (XEXP (y, 0)) == MINUS
3526 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3527 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg0))
3528 return XEXP (XEXP (y, 0), 0);
3531 /* If second operand is a register equivalent to a negative
3532 CONST_INT, see if we can find a register equivalent to the
3533 positive constant. Make a MINUS if so. Don't do this for
3534 a non-negative constant since we might then alternate between
3535 choosing positive and negative constants. Having the positive
3536 constant previously-used is the more common case. Be sure
3537 the resulting constant is non-negative; if const_arg1 were
3538 the smallest negative number this would overflow: depending
3539 on the mode, this would either just be the same value (and
3540 hence not save anything) or be incorrect. */
3541 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3542 && INTVAL (const_arg1) < 0
3543 /* This used to test
3545 -INTVAL (const_arg1) >= 0
3547 But The Sun V5.0 compilers mis-compiled that test. So
3548 instead we test for the problematic value in a more direct
3549 manner and hope the Sun compilers get it correct. */
3550 && INTVAL (const_arg1) !=
3551 (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1))
3552 && REG_P (folded_arg1))
3554 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3555 struct table_elt *p
3556 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3558 if (p)
3559 for (p = p->first_same_value; p; p = p->next_same_value)
3560 if (REG_P (p->exp))
3561 return simplify_gen_binary (MINUS, mode, folded_arg0,
3562 canon_reg (p->exp, NULL));
3564 goto from_plus;
3566 case MINUS:
3567 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3568 If so, produce (PLUS Z C2-C). */
3569 if (const_arg1 != 0 && poly_int_rtx_p (const_arg1, &xval))
3571 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3572 if (y && poly_int_rtx_p (XEXP (y, 1)))
3573 return fold_rtx (plus_constant (mode, copy_rtx (y), -xval),
3574 NULL);
3577 /* Fall through. */
3579 from_plus:
3580 case SMIN: case SMAX: case UMIN: case UMAX:
3581 case IOR: case AND: case XOR:
3582 case MULT:
3583 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3584 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3585 is known to be of similar form, we may be able to replace the
3586 operation with a combined operation. This may eliminate the
3587 intermediate operation if every use is simplified in this way.
3588 Note that the similar optimization done by combine.cc only works
3589 if the intermediate operation's result has only one reference. */
3591 if (REG_P (folded_arg0)
3592 && const_arg1 && CONST_INT_P (const_arg1))
3594 int is_shift
3595 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3596 rtx y, inner_const, new_const;
3597 rtx canon_const_arg1 = const_arg1;
3598 enum rtx_code associate_code;
3600 if (is_shift
3601 && (INTVAL (const_arg1) >= GET_MODE_UNIT_PRECISION (mode)
3602 || INTVAL (const_arg1) < 0))
3604 if (SHIFT_COUNT_TRUNCATED)
3605 canon_const_arg1 = gen_int_shift_amount
3606 (mode, (INTVAL (const_arg1)
3607 & (GET_MODE_UNIT_BITSIZE (mode) - 1)));
3608 else
3609 break;
3612 y = lookup_as_function (folded_arg0, code);
3613 if (y == 0)
3614 break;
3616 /* If we have compiled a statement like
3617 "if (x == (x & mask1))", and now are looking at
3618 "x & mask2", we will have a case where the first operand
3619 of Y is the same as our first operand. Unless we detect
3620 this case, an infinite loop will result. */
3621 if (XEXP (y, 0) == folded_arg0)
3622 break;
3624 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3625 if (!inner_const || !CONST_INT_P (inner_const))
3626 break;
3628 /* Don't associate these operations if they are a PLUS with the
3629 same constant and it is a power of two. These might be doable
3630 with a pre- or post-increment. Similarly for two subtracts of
3631 identical powers of two with post decrement. */
3633 if (code == PLUS && const_arg1 == inner_const
3634 && ((HAVE_PRE_INCREMENT
3635 && pow2p_hwi (INTVAL (const_arg1)))
3636 || (HAVE_POST_INCREMENT
3637 && pow2p_hwi (INTVAL (const_arg1)))
3638 || (HAVE_PRE_DECREMENT
3639 && pow2p_hwi (- INTVAL (const_arg1)))
3640 || (HAVE_POST_DECREMENT
3641 && pow2p_hwi (- INTVAL (const_arg1)))))
3642 break;
3644 /* ??? Vector mode shifts by scalar
3645 shift operand are not supported yet. */
3646 if (is_shift && VECTOR_MODE_P (mode))
3647 break;
3649 if (is_shift
3650 && (INTVAL (inner_const) >= GET_MODE_UNIT_PRECISION (mode)
3651 || INTVAL (inner_const) < 0))
3653 if (SHIFT_COUNT_TRUNCATED)
3654 inner_const = gen_int_shift_amount
3655 (mode, (INTVAL (inner_const)
3656 & (GET_MODE_UNIT_BITSIZE (mode) - 1)));
3657 else
3658 break;
3661 /* Compute the code used to compose the constants. For example,
3662 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3664 associate_code = (is_shift || code == MINUS ? PLUS : code);
3666 new_const = simplify_binary_operation (associate_code, mode,
3667 canon_const_arg1,
3668 inner_const);
3670 if (new_const == 0)
3671 break;
3673 /* If we are associating shift operations, don't let this
3674 produce a shift of the size of the object or larger.
3675 This could occur when we follow a sign-extend by a right
3676 shift on a machine that does a sign-extend as a pair
3677 of shifts. */
3679 if (is_shift
3680 && CONST_INT_P (new_const)
3681 && INTVAL (new_const) >= GET_MODE_UNIT_PRECISION (mode))
3683 /* As an exception, we can turn an ASHIFTRT of this
3684 form into a shift of the number of bits - 1. */
3685 if (code == ASHIFTRT)
3686 new_const = gen_int_shift_amount
3687 (mode, GET_MODE_UNIT_BITSIZE (mode) - 1);
3688 else if (!side_effects_p (XEXP (y, 0)))
3689 return CONST0_RTX (mode);
3690 else
3691 break;
3694 y = copy_rtx (XEXP (y, 0));
3696 /* If Y contains our first operand (the most common way this
3697 can happen is if Y is a MEM), we would do into an infinite
3698 loop if we tried to fold it. So don't in that case. */
3700 if (! reg_mentioned_p (folded_arg0, y))
3701 y = fold_rtx (y, insn);
3703 return simplify_gen_binary (code, mode, y, new_const);
3705 break;
3707 case DIV: case UDIV:
3708 /* ??? The associative optimization performed immediately above is
3709 also possible for DIV and UDIV using associate_code of MULT.
3710 However, we would need extra code to verify that the
3711 multiplication does not overflow, that is, there is no overflow
3712 in the calculation of new_const. */
3713 break;
3715 default:
3716 break;
3719 new_rtx = simplify_binary_operation (code, mode,
3720 const_arg0 ? const_arg0 : folded_arg0,
3721 const_arg1 ? const_arg1 : folded_arg1);
3722 break;
3724 case RTX_OBJ:
3725 /* (lo_sum (high X) X) is simply X. */
3726 if (code == LO_SUM && const_arg0 != 0
3727 && GET_CODE (const_arg0) == HIGH
3728 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3729 return const_arg1;
3730 break;
3732 case RTX_TERNARY:
3733 case RTX_BITFIELD_OPS:
3734 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3735 const_arg0 ? const_arg0 : folded_arg0,
3736 const_arg1 ? const_arg1 : folded_arg1,
3737 const_arg2 ? const_arg2 : XEXP (x, 2));
3738 break;
3740 default:
3741 break;
3744 return new_rtx ? new_rtx : x;
3747 /* Return a constant value currently equivalent to X.
3748 Return 0 if we don't know one. */
3750 static rtx
3751 equiv_constant (rtx x)
3753 if (REG_P (x)
3754 && REGNO_QTY_VALID_P (REGNO (x)))
3756 int x_q = REG_QTY (REGNO (x));
3757 struct qty_table_elem *x_ent = &qty_table[x_q];
3759 if (x_ent->const_rtx)
3760 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3763 if (x == 0 || CONSTANT_P (x))
3764 return x;
3766 if (GET_CODE (x) == SUBREG)
3768 machine_mode mode = GET_MODE (x);
3769 machine_mode imode = GET_MODE (SUBREG_REG (x));
3770 rtx new_rtx;
3772 /* See if we previously assigned a constant value to this SUBREG. */
3773 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3774 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3775 || (NUM_POLY_INT_COEFFS > 1
3776 && (new_rtx = lookup_as_function (x, CONST_POLY_INT)) != 0)
3777 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3778 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3779 return new_rtx;
3781 /* If we didn't and if doing so makes sense, see if we previously
3782 assigned a constant value to the enclosing word mode SUBREG. */
3783 if (known_lt (GET_MODE_SIZE (mode), UNITS_PER_WORD)
3784 && known_lt (UNITS_PER_WORD, GET_MODE_SIZE (imode)))
3786 poly_int64 byte = (SUBREG_BYTE (x)
3787 - subreg_lowpart_offset (mode, word_mode));
3788 if (known_ge (byte, 0) && multiple_p (byte, UNITS_PER_WORD))
3790 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3791 new_rtx = lookup_as_function (y, CONST_INT);
3792 if (new_rtx)
3793 return gen_lowpart (mode, new_rtx);
3797 /* Otherwise see if we already have a constant for the inner REG,
3798 and if that is enough to calculate an equivalent constant for
3799 the subreg. Note that the upper bits of paradoxical subregs
3800 are undefined, so they cannot be said to equal anything. */
3801 if (REG_P (SUBREG_REG (x))
3802 && !paradoxical_subreg_p (x)
3803 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3804 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3806 return 0;
3809 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3810 the hash table in case its value was seen before. */
3812 if (MEM_P (x))
3814 struct table_elt *elt;
3816 x = avoid_constant_pool_reference (x);
3817 if (CONSTANT_P (x))
3818 return x;
3820 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3821 if (elt == 0)
3822 return 0;
3824 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3825 if (elt->is_const && CONSTANT_P (elt->exp))
3826 return elt->exp;
3829 return 0;
3832 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3833 "taken" branch.
3835 In certain cases, this can cause us to add an equivalence. For example,
3836 if we are following the taken case of
3837 if (i == 2)
3838 we can add the fact that `i' and '2' are now equivalent.
3840 In any case, we can record that this comparison was passed. If the same
3841 comparison is seen later, we will know its value. */
3843 static void
3844 record_jump_equiv (rtx_insn *insn, bool taken)
3846 int cond_known_true;
3847 rtx op0, op1;
3848 rtx set;
3849 machine_mode mode, mode0, mode1;
3850 enum rtx_code code;
3852 /* Ensure this is the right kind of insn. */
3853 gcc_assert (any_condjump_p (insn));
3855 set = pc_set (insn);
3857 /* See if this jump condition is known true or false. */
3858 if (taken)
3859 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3860 else
3861 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3863 /* Get the type of comparison being done and the operands being compared.
3864 If we had to reverse a non-equality condition, record that fact so we
3865 know that it isn't valid for floating-point. */
3866 code = GET_CODE (XEXP (SET_SRC (set), 0));
3867 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3868 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3870 /* If fold_rtx returns NULL_RTX, there's nothing to record. */
3871 if (op0 == NULL_RTX || op1 == NULL_RTX)
3872 return;
3874 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3875 if (! cond_known_true)
3877 code = reversed_comparison_code_parts (code, op0, op1, insn);
3879 /* Don't remember if we can't find the inverse. */
3880 if (code == UNKNOWN)
3881 return;
3884 /* The mode is the mode of the non-constant. */
3885 mode = mode0;
3886 if (mode1 != VOIDmode)
3887 mode = mode1;
3889 record_jump_cond (code, mode, op0, op1);
3892 /* Yet another form of subreg creation. In this case, we want something in
3893 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3895 static rtx
3896 record_jump_cond_subreg (machine_mode mode, rtx op)
3898 machine_mode op_mode = GET_MODE (op);
3899 if (op_mode == mode || op_mode == VOIDmode)
3900 return op;
3901 return lowpart_subreg (mode, op, op_mode);
3904 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3905 Make any useful entries we can with that information. Called from
3906 above function and called recursively. */
3908 static void
3909 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0, rtx op1)
3911 unsigned op0_hash, op1_hash;
3912 int op0_in_memory, op1_in_memory;
3913 struct table_elt *op0_elt, *op1_elt;
3915 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3916 we know that they are also equal in the smaller mode (this is also
3917 true for all smaller modes whether or not there is a SUBREG, but
3918 is not worth testing for with no SUBREG). */
3920 /* Note that GET_MODE (op0) may not equal MODE. */
3921 if (code == EQ && paradoxical_subreg_p (op0))
3923 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3924 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3925 if (tem)
3926 record_jump_cond (code, mode, SUBREG_REG (op0), tem);
3929 if (code == EQ && paradoxical_subreg_p (op1))
3931 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3932 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3933 if (tem)
3934 record_jump_cond (code, mode, SUBREG_REG (op1), tem);
3937 /* Similarly, if this is an NE comparison, and either is a SUBREG
3938 making a smaller mode, we know the whole thing is also NE. */
3940 /* Note that GET_MODE (op0) may not equal MODE;
3941 if we test MODE instead, we can get an infinite recursion
3942 alternating between two modes each wider than MODE. */
3944 if (code == NE
3945 && partial_subreg_p (op0)
3946 && subreg_lowpart_p (op0))
3948 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3949 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3950 if (tem)
3951 record_jump_cond (code, mode, SUBREG_REG (op0), tem);
3954 if (code == NE
3955 && partial_subreg_p (op1)
3956 && subreg_lowpart_p (op1))
3958 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3959 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3960 if (tem)
3961 record_jump_cond (code, mode, SUBREG_REG (op1), tem);
3964 /* Hash both operands. */
3966 do_not_record = 0;
3967 hash_arg_in_memory = 0;
3968 op0_hash = HASH (op0, mode);
3969 op0_in_memory = hash_arg_in_memory;
3971 if (do_not_record)
3972 return;
3974 do_not_record = 0;
3975 hash_arg_in_memory = 0;
3976 op1_hash = HASH (op1, mode);
3977 op1_in_memory = hash_arg_in_memory;
3979 if (do_not_record)
3980 return;
3982 /* Look up both operands. */
3983 op0_elt = lookup (op0, op0_hash, mode);
3984 op1_elt = lookup (op1, op1_hash, mode);
3986 /* If both operands are already equivalent or if they are not in the
3987 table but are identical, do nothing. */
3988 if ((op0_elt != 0 && op1_elt != 0
3989 && op0_elt->first_same_value == op1_elt->first_same_value)
3990 || op0 == op1 || rtx_equal_p (op0, op1))
3991 return;
3993 /* If we aren't setting two things equal all we can do is save this
3994 comparison. Similarly if this is floating-point. In the latter
3995 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
3996 If we record the equality, we might inadvertently delete code
3997 whose intent was to change -0 to +0. */
3999 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4001 struct qty_table_elem *ent;
4002 int qty;
4004 /* If OP0 is not a register, or if OP1 is neither a register
4005 or constant, we can't do anything. */
4007 if (!REG_P (op1))
4008 op1 = equiv_constant (op1);
4010 if (!REG_P (op0) || op1 == 0)
4011 return;
4013 /* Put OP0 in the hash table if it isn't already. This gives it a
4014 new quantity number. */
4015 if (op0_elt == 0)
4017 if (insert_regs (op0, NULL, false))
4019 rehash_using_reg (op0);
4020 op0_hash = HASH (op0, mode);
4022 /* If OP0 is contained in OP1, this changes its hash code
4023 as well. Faster to rehash than to check, except
4024 for the simple case of a constant. */
4025 if (! CONSTANT_P (op1))
4026 op1_hash = HASH (op1,mode);
4029 op0_elt = insert (op0, NULL, op0_hash, mode);
4030 op0_elt->in_memory = op0_in_memory;
4033 qty = REG_QTY (REGNO (op0));
4034 ent = &qty_table[qty];
4036 ent->comparison_code = code;
4037 if (REG_P (op1))
4039 /* Look it up again--in case op0 and op1 are the same. */
4040 op1_elt = lookup (op1, op1_hash, mode);
4042 /* Put OP1 in the hash table so it gets a new quantity number. */
4043 if (op1_elt == 0)
4045 if (insert_regs (op1, NULL, false))
4047 rehash_using_reg (op1);
4048 op1_hash = HASH (op1, mode);
4051 op1_elt = insert (op1, NULL, op1_hash, mode);
4052 op1_elt->in_memory = op1_in_memory;
4055 ent->comparison_const = NULL_RTX;
4056 ent->comparison_qty = REG_QTY (REGNO (op1));
4058 else
4060 ent->comparison_const = op1;
4061 ent->comparison_qty = -1;
4064 return;
4067 /* If either side is still missing an equivalence, make it now,
4068 then merge the equivalences. */
4070 if (op0_elt == 0)
4072 if (insert_regs (op0, NULL, false))
4074 rehash_using_reg (op0);
4075 op0_hash = HASH (op0, mode);
4078 op0_elt = insert (op0, NULL, op0_hash, mode);
4079 op0_elt->in_memory = op0_in_memory;
4082 if (op1_elt == 0)
4084 if (insert_regs (op1, NULL, false))
4086 rehash_using_reg (op1);
4087 op1_hash = HASH (op1, mode);
4090 op1_elt = insert (op1, NULL, op1_hash, mode);
4091 op1_elt->in_memory = op1_in_memory;
4094 merge_equiv_classes (op0_elt, op1_elt);
4097 /* CSE processing for one instruction.
4099 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4100 but the few that "leak through" are cleaned up by cse_insn, and complex
4101 addressing modes are often formed here.
4103 The main function is cse_insn, and between here and that function
4104 a couple of helper functions is defined to keep the size of cse_insn
4105 within reasonable proportions.
4107 Data is shared between the main and helper functions via STRUCT SET,
4108 that contains all data related for every set in the instruction that
4109 is being processed.
4111 Note that cse_main processes all sets in the instruction. Most
4112 passes in GCC only process simple SET insns or single_set insns, but
4113 CSE processes insns with multiple sets as well. */
4115 /* Data on one SET contained in the instruction. */
4117 struct set
4119 /* The SET rtx itself. */
4120 rtx rtl;
4121 /* The SET_SRC of the rtx (the original value, if it is changing). */
4122 rtx src;
4123 /* The hash-table element for the SET_SRC of the SET. */
4124 struct table_elt *src_elt;
4125 /* Hash value for the SET_SRC. */
4126 unsigned src_hash;
4127 /* Hash value for the SET_DEST. */
4128 unsigned dest_hash;
4129 /* The SET_DEST, with SUBREG, etc., stripped. */
4130 rtx inner_dest;
4131 /* Original machine mode, in case it becomes a CONST_INT. */
4132 ENUM_BITFIELD(machine_mode) mode : MACHINE_MODE_BITSIZE;
4133 /* Nonzero if the SET_SRC is in memory. */
4134 unsigned int src_in_memory : 1;
4135 /* Nonzero if the SET_SRC contains something
4136 whose value cannot be predicted and understood. */
4137 unsigned int src_volatile : 1;
4138 /* Nonzero if RTL is an artifical set that has been created to describe
4139 part of an insn's effect. Zero means that RTL appears directly in
4140 the insn pattern. */
4141 unsigned int is_fake_set : 1;
4142 /* Hash value of constant equivalent for SET_SRC. */
4143 unsigned src_const_hash;
4144 /* A constant equivalent for SET_SRC, if any. */
4145 rtx src_const;
4146 /* Table entry for constant equivalent for SET_SRC, if any. */
4147 struct table_elt *src_const_elt;
4148 /* Table entry for the destination address. */
4149 struct table_elt *dest_addr_elt;
4152 /* Special handling for (set REG0 REG1) where REG0 is the
4153 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4154 be used in the sequel, so (if easily done) change this insn to
4155 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4156 that computed their value. Then REG1 will become a dead store
4157 and won't cloud the situation for later optimizations.
4159 Do not make this change if REG1 is a hard register, because it will
4160 then be used in the sequel and we may be changing a two-operand insn
4161 into a three-operand insn.
4163 This is the last transformation that cse_insn will try to do. */
4165 static void
4166 try_back_substitute_reg (rtx set, rtx_insn *insn)
4168 rtx dest = SET_DEST (set);
4169 rtx src = SET_SRC (set);
4171 if (REG_P (dest)
4172 && REG_P (src) && ! HARD_REGISTER_P (src)
4173 && REGNO_QTY_VALID_P (REGNO (src)))
4175 int src_q = REG_QTY (REGNO (src));
4176 struct qty_table_elem *src_ent = &qty_table[src_q];
4178 if (src_ent->first_reg == REGNO (dest))
4180 /* Scan for the previous nonnote insn, but stop at a basic
4181 block boundary. */
4182 rtx_insn *prev = insn;
4183 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4186 prev = PREV_INSN (prev);
4188 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4190 /* Do not swap the registers around if the previous instruction
4191 attaches a REG_EQUIV note to REG1.
4193 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4194 from the pseudo that originally shadowed an incoming argument
4195 to another register. Some uses of REG_EQUIV might rely on it
4196 being attached to REG1 rather than REG2.
4198 This section previously turned the REG_EQUIV into a REG_EQUAL
4199 note. We cannot do that because REG_EQUIV may provide an
4200 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4201 if (NONJUMP_INSN_P (prev)
4202 && GET_CODE (PATTERN (prev)) == SET
4203 && SET_DEST (PATTERN (prev)) == src
4204 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4206 rtx note;
4208 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4209 validate_change (insn, &SET_DEST (set), src, 1);
4210 validate_change (insn, &SET_SRC (set), dest, 1);
4211 apply_change_group ();
4213 /* If INSN has a REG_EQUAL note, and this note mentions
4214 REG0, then we must delete it, because the value in
4215 REG0 has changed. If the note's value is REG1, we must
4216 also delete it because that is now this insn's dest. */
4217 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4218 if (note != 0
4219 && (reg_mentioned_p (dest, XEXP (note, 0))
4220 || rtx_equal_p (src, XEXP (note, 0))))
4221 remove_note (insn, note);
4223 /* If INSN has a REG_ARGS_SIZE note, move it to PREV. */
4224 note = find_reg_note (insn, REG_ARGS_SIZE, NULL_RTX);
4225 if (note != 0)
4227 remove_note (insn, note);
4228 gcc_assert (!find_reg_note (prev, REG_ARGS_SIZE, NULL_RTX));
4229 set_unique_reg_note (prev, REG_ARGS_SIZE, XEXP (note, 0));
4236 /* Add an entry containing RTL X into SETS. IS_FAKE_SET is true if X is
4237 an artifical set that has been created to describe part of an insn's
4238 effect. */
4239 static inline void
4240 add_to_set (vec<struct set> *sets, rtx x, bool is_fake_set)
4242 struct set entry = {};
4243 entry.rtl = x;
4244 entry.is_fake_set = is_fake_set;
4245 sets->safe_push (entry);
4248 /* Record all the SETs in this instruction into SETS_PTR,
4249 and return the number of recorded sets. */
4250 static int
4251 find_sets_in_insn (rtx_insn *insn, vec<struct set> *psets)
4253 rtx x = PATTERN (insn);
4255 if (GET_CODE (x) == SET)
4257 /* Ignore SETs that are unconditional jumps.
4258 They never need cse processing, so this does not hurt.
4259 The reason is not efficiency but rather
4260 so that we can test at the end for instructions
4261 that have been simplified to unconditional jumps
4262 and not be misled by unchanged instructions
4263 that were unconditional jumps to begin with. */
4264 if (SET_DEST (x) == pc_rtx
4265 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4267 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4268 The hard function value register is used only once, to copy to
4269 someplace else, so it isn't worth cse'ing. */
4270 else if (GET_CODE (SET_SRC (x)) == CALL)
4272 else if (GET_CODE (SET_SRC (x)) == CONST_VECTOR
4273 && GET_MODE_CLASS (GET_MODE (SET_SRC (x))) != MODE_VECTOR_BOOL
4274 /* Prevent duplicates from being generated if the type is a V1
4275 type and a subreg. Folding this will result in the same
4276 element as folding x itself. */
4277 && !(SUBREG_P (SET_DEST (x))
4278 && known_eq (GET_MODE_NUNITS (GET_MODE (SET_SRC (x))), 1)))
4280 /* First register the vector itself. */
4281 add_to_set (psets, x, false);
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 rtx set = gen_rtx_SET (y, CONST_VECTOR_ELT (src, i));
4292 add_to_set (psets, set, true);
4295 else
4296 add_to_set (psets, x, false);
4298 else if (GET_CODE (x) == PARALLEL)
4300 int i, lim = XVECLEN (x, 0);
4302 /* Go over the expressions of the PARALLEL in forward order, to
4303 put them in the same order in the SETS array. */
4304 for (i = 0; i < lim; i++)
4306 rtx y = XVECEXP (x, 0, i);
4307 if (GET_CODE (y) == SET)
4309 /* As above, we ignore unconditional jumps and call-insns and
4310 ignore the result of apply_change_group. */
4311 if (SET_DEST (y) == pc_rtx
4312 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4314 else if (GET_CODE (SET_SRC (y)) == CALL)
4316 else
4317 add_to_set (psets, y, false);
4322 return psets->length ();
4325 /* Subroutine of canonicalize_insn. X is an ASM_OPERANDS in INSN. */
4327 static void
4328 canon_asm_operands (rtx x, rtx_insn *insn)
4330 for (int i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4332 rtx input = ASM_OPERANDS_INPUT (x, i);
4333 if (!(REG_P (input) && HARD_REGISTER_P (input)))
4335 input = canon_reg (input, insn);
4336 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4341 /* Where possible, substitute every register reference in the N_SETS
4342 number of SETS in INSN with the canonical register.
4344 Register canonicalization propagatest the earliest register (i.e.
4345 one that is set before INSN) with the same value. This is a very
4346 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4347 to RTL. For instance, a CONST for an address is usually expanded
4348 multiple times to loads into different registers, thus creating many
4349 subexpressions of the form:
4351 (set (reg1) (some_const))
4352 (set (mem (... reg1 ...) (thing)))
4353 (set (reg2) (some_const))
4354 (set (mem (... reg2 ...) (thing)))
4356 After canonicalizing, the code takes the following form:
4358 (set (reg1) (some_const))
4359 (set (mem (... reg1 ...) (thing)))
4360 (set (reg2) (some_const))
4361 (set (mem (... reg1 ...) (thing)))
4363 The set to reg2 is now trivially dead, and the memory reference (or
4364 address, or whatever) may be a candidate for further CSEing.
4366 In this function, the result of apply_change_group can be ignored;
4367 see canon_reg. */
4369 static void
4370 canonicalize_insn (rtx_insn *insn, vec<struct set> *psets)
4372 vec<struct set> sets = *psets;
4373 int n_sets = sets.length ();
4374 rtx tem;
4375 rtx x = PATTERN (insn);
4376 int i;
4378 if (CALL_P (insn))
4380 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4381 if (GET_CODE (XEXP (tem, 0)) != SET)
4382 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4385 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4387 canon_reg (SET_SRC (x), insn);
4388 apply_change_group ();
4389 fold_rtx (SET_SRC (x), insn);
4391 else if (GET_CODE (x) == CLOBBER)
4393 /* If we clobber memory, canon the address.
4394 This does nothing when a register is clobbered
4395 because we have already invalidated the reg. */
4396 if (MEM_P (XEXP (x, 0)))
4397 canon_reg (XEXP (x, 0), insn);
4399 else if (GET_CODE (x) == USE
4400 && ! (REG_P (XEXP (x, 0))
4401 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4402 /* Canonicalize a USE of a pseudo register or memory location. */
4403 canon_reg (x, insn);
4404 else if (GET_CODE (x) == ASM_OPERANDS)
4405 canon_asm_operands (x, insn);
4406 else if (GET_CODE (x) == CALL)
4408 canon_reg (x, insn);
4409 apply_change_group ();
4410 fold_rtx (x, insn);
4412 else if (DEBUG_INSN_P (insn))
4413 canon_reg (PATTERN (insn), insn);
4414 else if (GET_CODE (x) == PARALLEL)
4416 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4418 rtx y = XVECEXP (x, 0, i);
4419 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4421 canon_reg (SET_SRC (y), insn);
4422 apply_change_group ();
4423 fold_rtx (SET_SRC (y), insn);
4425 else if (GET_CODE (y) == CLOBBER)
4427 if (MEM_P (XEXP (y, 0)))
4428 canon_reg (XEXP (y, 0), insn);
4430 else if (GET_CODE (y) == USE
4431 && ! (REG_P (XEXP (y, 0))
4432 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4433 canon_reg (y, insn);
4434 else if (GET_CODE (y) == ASM_OPERANDS)
4435 canon_asm_operands (y, insn);
4436 else if (GET_CODE (y) == CALL)
4438 canon_reg (y, insn);
4439 apply_change_group ();
4440 fold_rtx (y, insn);
4445 if (n_sets == 1 && REG_NOTES (insn) != 0
4446 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4448 /* We potentially will process this insn many times. Therefore,
4449 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4450 unique set in INSN.
4452 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4453 because cse_insn handles those specially. */
4454 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4455 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4456 remove_note (insn, tem);
4457 else
4459 canon_reg (XEXP (tem, 0), insn);
4460 apply_change_group ();
4461 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4462 df_notes_rescan (insn);
4466 /* Canonicalize sources and addresses of destinations.
4467 We do this in a separate pass to avoid problems when a MATCH_DUP is
4468 present in the insn pattern. In that case, we want to ensure that
4469 we don't break the duplicate nature of the pattern. So we will replace
4470 both operands at the same time. Otherwise, we would fail to find an
4471 equivalent substitution in the loop calling validate_change below.
4473 We used to suppress canonicalization of DEST if it appears in SRC,
4474 but we don't do this any more. */
4476 for (i = 0; i < n_sets; i++)
4478 rtx dest = SET_DEST (sets[i].rtl);
4479 rtx src = SET_SRC (sets[i].rtl);
4480 rtx new_rtx = canon_reg (src, insn);
4482 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4484 if (GET_CODE (dest) == ZERO_EXTRACT)
4486 validate_change (insn, &XEXP (dest, 1),
4487 canon_reg (XEXP (dest, 1), insn), 1);
4488 validate_change (insn, &XEXP (dest, 2),
4489 canon_reg (XEXP (dest, 2), insn), 1);
4492 while (GET_CODE (dest) == SUBREG
4493 || GET_CODE (dest) == ZERO_EXTRACT
4494 || GET_CODE (dest) == STRICT_LOW_PART)
4495 dest = XEXP (dest, 0);
4497 if (MEM_P (dest))
4498 canon_reg (dest, insn);
4501 /* Now that we have done all the replacements, we can apply the change
4502 group and see if they all work. Note that this will cause some
4503 canonicalizations that would have worked individually not to be applied
4504 because some other canonicalization didn't work, but this should not
4505 occur often.
4507 The result of apply_change_group can be ignored; see canon_reg. */
4509 apply_change_group ();
4512 /* Main function of CSE.
4513 First simplify sources and addresses of all assignments
4514 in the instruction, using previously-computed equivalents values.
4515 Then install the new sources and destinations in the table
4516 of available values. */
4518 static void
4519 cse_insn (rtx_insn *insn)
4521 rtx x = PATTERN (insn);
4522 int i;
4523 rtx tem;
4524 int n_sets = 0;
4526 rtx src_eqv = 0;
4527 struct table_elt *src_eqv_elt = 0;
4528 int src_eqv_volatile = 0;
4529 int src_eqv_in_memory = 0;
4530 unsigned src_eqv_hash = 0;
4532 this_insn = insn;
4534 /* Find all regs explicitly clobbered in this insn,
4535 to ensure they are not replaced with any other regs
4536 elsewhere in this insn. */
4537 invalidate_from_sets_and_clobbers (insn);
4539 /* Record all the SETs in this instruction. */
4540 auto_vec<struct set, 8> sets;
4541 n_sets = find_sets_in_insn (insn, (vec<struct set>*)&sets);
4543 /* Substitute the canonical register where possible. */
4544 canonicalize_insn (insn, (vec<struct set>*)&sets);
4546 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4547 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4548 latter condition is necessary because SRC_EQV is handled specially for
4549 this case, and if it isn't set, then there will be no equivalence
4550 for the destination. */
4551 if (n_sets == 1 && REG_NOTES (insn) != 0
4552 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4555 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4556 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4557 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4558 src_eqv = copy_rtx (XEXP (tem, 0));
4559 /* If DEST is of the form ZERO_EXTACT, as in:
4560 (set (zero_extract:SI (reg:SI 119)
4561 (const_int 16 [0x10])
4562 (const_int 16 [0x10]))
4563 (const_int 51154 [0xc7d2]))
4564 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4565 point. Note that this is different from SRC_EQV. We can however
4566 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4567 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4568 && CONST_INT_P (XEXP (tem, 0))
4569 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4570 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4572 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4573 /* This is the mode of XEXP (tem, 0) as well. */
4574 scalar_int_mode dest_mode
4575 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
4576 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4577 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4578 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4579 HOST_WIDE_INT mask;
4580 unsigned int shift;
4581 if (BITS_BIG_ENDIAN)
4582 shift = (GET_MODE_PRECISION (dest_mode)
4583 - INTVAL (pos) - INTVAL (width));
4584 else
4585 shift = INTVAL (pos);
4586 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4587 mask = HOST_WIDE_INT_M1;
4588 else
4589 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
4590 val = (val >> shift) & mask;
4591 src_eqv = GEN_INT (val);
4595 /* Set sets[i].src_elt to the class each source belongs to.
4596 Detect assignments from or to volatile things
4597 and set set[i] to zero so they will be ignored
4598 in the rest of this function.
4600 Nothing in this loop changes the hash table or the register chains. */
4602 for (i = 0; i < n_sets; i++)
4604 bool repeat = false;
4605 bool noop_insn = false;
4606 rtx src, dest;
4607 rtx src_folded;
4608 struct table_elt *elt = 0, *p;
4609 machine_mode mode;
4610 rtx src_eqv_here;
4611 rtx src_const = 0;
4612 rtx src_related = 0;
4613 rtx dest_related = 0;
4614 bool src_related_is_const_anchor = false;
4615 struct table_elt *src_const_elt = 0;
4616 int src_cost = MAX_COST;
4617 int src_eqv_cost = MAX_COST;
4618 int src_folded_cost = MAX_COST;
4619 int src_related_cost = MAX_COST;
4620 int src_elt_cost = MAX_COST;
4621 int src_regcost = MAX_COST;
4622 int src_eqv_regcost = MAX_COST;
4623 int src_folded_regcost = MAX_COST;
4624 int src_related_regcost = MAX_COST;
4625 int src_elt_regcost = MAX_COST;
4626 scalar_int_mode int_mode;
4627 bool is_fake_set = sets[i].is_fake_set;
4629 dest = SET_DEST (sets[i].rtl);
4630 src = SET_SRC (sets[i].rtl);
4632 /* If SRC is a constant that has no machine mode,
4633 hash it with the destination's machine mode.
4634 This way we can keep different modes separate. */
4636 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4637 sets[i].mode = mode;
4639 if (!is_fake_set && src_eqv)
4641 machine_mode eqvmode = mode;
4642 if (GET_CODE (dest) == STRICT_LOW_PART)
4643 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4644 do_not_record = 0;
4645 hash_arg_in_memory = 0;
4646 src_eqv_hash = HASH (src_eqv, eqvmode);
4648 /* Find the equivalence class for the equivalent expression. */
4650 if (!do_not_record)
4651 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4653 src_eqv_volatile = do_not_record;
4654 src_eqv_in_memory = hash_arg_in_memory;
4657 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4658 value of the INNER register, not the destination. So it is not
4659 a valid substitution for the source. But save it for later. */
4660 if (is_fake_set || GET_CODE (dest) == STRICT_LOW_PART)
4661 src_eqv_here = 0;
4662 else
4663 src_eqv_here = src_eqv;
4665 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4666 simplified result, which may not necessarily be valid. */
4667 src_folded = fold_rtx (src, NULL);
4669 #if 0
4670 /* ??? This caused bad code to be generated for the m68k port with -O2.
4671 Suppose src is (CONST_INT -1), and that after truncation src_folded
4672 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4673 At the end we will add src and src_const to the same equivalence
4674 class. We now have 3 and -1 on the same equivalence class. This
4675 causes later instructions to be mis-optimized. */
4676 /* If storing a constant in a bitfield, pre-truncate the constant
4677 so we will be able to record it later. */
4678 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4680 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4682 if (CONST_INT_P (src)
4683 && CONST_INT_P (width)
4684 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4685 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4686 src_folded
4687 = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1
4688 << INTVAL (width)) - 1));
4690 #endif
4692 /* Compute SRC's hash code, and also notice if it
4693 should not be recorded at all. In that case,
4694 prevent any further processing of this assignment.
4696 We set DO_NOT_RECORD if the destination has a REG_UNUSED note.
4697 This avoids getting the source register into the tables, where it
4698 may be invalidated later (via REG_QTY), then trigger an ICE upon
4699 re-insertion.
4701 This is only a problem in multi-set insns. If it were a single
4702 set the dead copy would have been removed. If the RHS were anything
4703 but a simple REG, then we won't call insert_regs and thus there's
4704 no potential for triggering the ICE. */
4705 do_not_record = (REG_P (dest)
4706 && REG_P (src)
4707 && find_reg_note (insn, REG_UNUSED, dest));
4708 hash_arg_in_memory = 0;
4710 sets[i].src = src;
4711 sets[i].src_hash = HASH (src, mode);
4712 sets[i].src_volatile = do_not_record;
4713 sets[i].src_in_memory = hash_arg_in_memory;
4715 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4716 a pseudo, do not record SRC. Using SRC as a replacement for
4717 anything else will be incorrect in that situation. Note that
4718 this usually occurs only for stack slots, in which case all the
4719 RTL would be referring to SRC, so we don't lose any optimization
4720 opportunities by not having SRC in the hash table. */
4722 if (MEM_P (src)
4723 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4724 && REG_P (dest)
4725 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4726 sets[i].src_volatile = 1;
4728 else if (GET_CODE (src) == ASM_OPERANDS
4729 && GET_CODE (x) == PARALLEL)
4731 /* Do not record result of a non-volatile inline asm with
4732 more than one result. */
4733 if (n_sets > 1)
4734 sets[i].src_volatile = 1;
4736 int j, lim = XVECLEN (x, 0);
4737 for (j = 0; j < lim; j++)
4739 rtx y = XVECEXP (x, 0, j);
4740 /* And do not record result of a non-volatile inline asm
4741 with "memory" clobber. */
4742 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4744 sets[i].src_volatile = 1;
4745 break;
4750 #if 0
4751 /* It is no longer clear why we used to do this, but it doesn't
4752 appear to still be needed. So let's try without it since this
4753 code hurts cse'ing widened ops. */
4754 /* If source is a paradoxical subreg (such as QI treated as an SI),
4755 treat it as volatile. It may do the work of an SI in one context
4756 where the extra bits are not being used, but cannot replace an SI
4757 in general. */
4758 if (paradoxical_subreg_p (src))
4759 sets[i].src_volatile = 1;
4760 #endif
4762 /* Locate all possible equivalent forms for SRC. Try to replace
4763 SRC in the insn with each cheaper equivalent.
4765 We have the following types of equivalents: SRC itself, a folded
4766 version, a value given in a REG_EQUAL note, or a value related
4767 to a constant.
4769 Each of these equivalents may be part of an additional class
4770 of equivalents (if more than one is in the table, they must be in
4771 the same class; we check for this).
4773 If the source is volatile, we don't do any table lookups.
4775 We note any constant equivalent for possible later use in a
4776 REG_NOTE. */
4778 if (!sets[i].src_volatile)
4779 elt = lookup (src, sets[i].src_hash, mode);
4781 sets[i].src_elt = elt;
4783 if (elt && src_eqv_here && src_eqv_elt)
4785 if (elt->first_same_value != src_eqv_elt->first_same_value)
4787 /* The REG_EQUAL is indicating that two formerly distinct
4788 classes are now equivalent. So merge them. */
4789 merge_equiv_classes (elt, src_eqv_elt);
4790 src_eqv_hash = HASH (src_eqv, elt->mode);
4791 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4794 src_eqv_here = 0;
4797 else if (src_eqv_elt)
4798 elt = src_eqv_elt;
4800 /* Try to find a constant somewhere and record it in `src_const'.
4801 Record its table element, if any, in `src_const_elt'. Look in
4802 any known equivalences first. (If the constant is not in the
4803 table, also set `sets[i].src_const_hash'). */
4804 if (elt)
4805 for (p = elt->first_same_value; p; p = p->next_same_value)
4806 if (p->is_const)
4808 src_const = p->exp;
4809 src_const_elt = elt;
4810 break;
4813 if (src_const == 0
4814 && (CONSTANT_P (src_folded)
4815 /* Consider (minus (label_ref L1) (label_ref L2)) as
4816 "constant" here so we will record it. This allows us
4817 to fold switch statements when an ADDR_DIFF_VEC is used. */
4818 || (GET_CODE (src_folded) == MINUS
4819 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4820 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4821 src_const = src_folded, src_const_elt = elt;
4822 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4823 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4825 /* If we don't know if the constant is in the table, get its
4826 hash code and look it up. */
4827 if (src_const && src_const_elt == 0)
4829 sets[i].src_const_hash = HASH (src_const, mode);
4830 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4833 sets[i].src_const = src_const;
4834 sets[i].src_const_elt = src_const_elt;
4836 /* If the constant and our source are both in the table, mark them as
4837 equivalent. Otherwise, if a constant is in the table but the source
4838 isn't, set ELT to it. */
4839 if (src_const_elt && elt
4840 && src_const_elt->first_same_value != elt->first_same_value)
4841 merge_equiv_classes (elt, src_const_elt);
4842 else if (src_const_elt && elt == 0)
4843 elt = src_const_elt;
4845 /* See if there is a register linearly related to a constant
4846 equivalent of SRC. */
4847 if (src_const
4848 && (GET_CODE (src_const) == CONST
4849 || (src_const_elt && src_const_elt->related_value != 0)))
4851 src_related = use_related_value (src_const, src_const_elt);
4852 if (src_related)
4854 struct table_elt *src_related_elt
4855 = lookup (src_related, HASH (src_related, mode), mode);
4856 if (src_related_elt && elt)
4858 if (elt->first_same_value
4859 != src_related_elt->first_same_value)
4860 /* This can occur when we previously saw a CONST
4861 involving a SYMBOL_REF and then see the SYMBOL_REF
4862 twice. Merge the involved classes. */
4863 merge_equiv_classes (elt, src_related_elt);
4865 src_related = 0;
4866 src_related_elt = 0;
4868 else if (src_related_elt && elt == 0)
4869 elt = src_related_elt;
4873 /* See if we have a CONST_INT that is already in a register in a
4874 wider mode. */
4876 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4877 && is_int_mode (mode, &int_mode)
4878 && GET_MODE_PRECISION (int_mode) < BITS_PER_WORD)
4880 opt_scalar_int_mode wider_mode_iter;
4881 FOR_EACH_WIDER_MODE (wider_mode_iter, int_mode)
4883 scalar_int_mode wider_mode = wider_mode_iter.require ();
4884 if (GET_MODE_PRECISION (wider_mode) > BITS_PER_WORD)
4885 break;
4887 struct table_elt *const_elt
4888 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4890 if (const_elt == 0)
4891 continue;
4893 for (const_elt = const_elt->first_same_value;
4894 const_elt; const_elt = const_elt->next_same_value)
4895 if (REG_P (const_elt->exp))
4897 src_related = gen_lowpart (int_mode, const_elt->exp);
4898 break;
4901 if (src_related != 0)
4902 break;
4906 /* Another possibility is that we have an AND with a constant in
4907 a mode narrower than a word. If so, it might have been generated
4908 as part of an "if" which would narrow the AND. If we already
4909 have done the AND in a wider mode, we can use a SUBREG of that
4910 value. */
4912 if (flag_expensive_optimizations && ! src_related
4913 && is_a <scalar_int_mode> (mode, &int_mode)
4914 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4915 && GET_MODE_SIZE (int_mode) < UNITS_PER_WORD)
4917 opt_scalar_int_mode tmode_iter;
4918 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4920 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4922 scalar_int_mode tmode = tmode_iter.require ();
4923 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4924 break;
4926 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4927 struct table_elt *larger_elt;
4929 if (inner)
4931 PUT_MODE (new_and, tmode);
4932 XEXP (new_and, 0) = inner;
4933 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4934 if (larger_elt == 0)
4935 continue;
4937 for (larger_elt = larger_elt->first_same_value;
4938 larger_elt; larger_elt = larger_elt->next_same_value)
4939 if (REG_P (larger_elt->exp))
4941 src_related
4942 = gen_lowpart (int_mode, larger_elt->exp);
4943 break;
4946 if (src_related)
4947 break;
4952 /* See if a MEM has already been loaded with a widening operation;
4953 if it has, we can use a subreg of that. Many CISC machines
4954 also have such operations, but this is only likely to be
4955 beneficial on these machines. */
4957 rtx_code extend_op;
4958 if (flag_expensive_optimizations && src_related == 0
4959 && MEM_P (src) && ! do_not_record
4960 && is_a <scalar_int_mode> (mode, &int_mode)
4961 && (extend_op = load_extend_op (int_mode)) != UNKNOWN)
4963 #if GCC_VERSION >= 5000
4964 struct rtx_def memory_extend_buf;
4965 rtx memory_extend_rtx = &memory_extend_buf;
4966 #else
4967 /* Workaround GCC < 5 bug, fixed in r5-3834 as part of PR63362
4968 fix. */
4969 alignas (rtx_def) unsigned char memory_extended_buf[sizeof (rtx_def)];
4970 rtx memory_extend_rtx = (rtx) &memory_extended_buf[0];
4971 #endif
4973 /* Set what we are trying to extend and the operation it might
4974 have been extended with. */
4975 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4976 PUT_CODE (memory_extend_rtx, extend_op);
4977 XEXP (memory_extend_rtx, 0) = src;
4979 opt_scalar_int_mode tmode_iter;
4980 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4982 struct table_elt *larger_elt;
4984 scalar_int_mode tmode = tmode_iter.require ();
4985 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4986 break;
4988 PUT_MODE (memory_extend_rtx, tmode);
4989 larger_elt = lookup (memory_extend_rtx,
4990 HASH (memory_extend_rtx, tmode), tmode);
4991 if (larger_elt == 0)
4992 continue;
4994 for (larger_elt = larger_elt->first_same_value;
4995 larger_elt; larger_elt = larger_elt->next_same_value)
4996 if (REG_P (larger_elt->exp))
4998 src_related = gen_lowpart (int_mode, larger_elt->exp);
4999 break;
5002 if (src_related)
5003 break;
5007 /* Try to express the constant using a register+offset expression
5008 derived from a constant anchor. */
5010 if (targetm.const_anchor
5011 && !src_related
5012 && src_const
5013 && GET_CODE (src_const) == CONST_INT)
5015 src_related = try_const_anchors (src_const, mode);
5016 src_related_is_const_anchor = src_related != NULL_RTX;
5019 /* Try to re-materialize a vec_dup with an existing constant. */
5020 rtx src_elt;
5021 if ((!src_eqv_here || CONSTANT_P (src_eqv_here))
5022 && const_vec_duplicate_p (src, &src_elt))
5024 machine_mode const_mode = GET_MODE_INNER (GET_MODE (src));
5025 struct table_elt *related_elt
5026 = lookup (src_elt, HASH (src_elt, const_mode), const_mode);
5027 if (related_elt)
5029 for (related_elt = related_elt->first_same_value;
5030 related_elt; related_elt = related_elt->next_same_value)
5031 if (REG_P (related_elt->exp))
5033 /* We don't need to compare costs with an existing (constant)
5034 src_eqv_here, since any such src_eqv_here should already be
5035 available in src_const. */
5036 src_eqv_here
5037 = gen_rtx_VEC_DUPLICATE (GET_MODE (src),
5038 related_elt->exp);
5039 break;
5044 if (src == src_folded)
5045 src_folded = 0;
5047 /* At this point, ELT, if nonzero, points to a class of expressions
5048 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5049 and SRC_RELATED, if nonzero, each contain additional equivalent
5050 expressions. Prune these latter expressions by deleting expressions
5051 already in the equivalence class.
5053 Check for an equivalent identical to the destination. If found,
5054 this is the preferred equivalent since it will likely lead to
5055 elimination of the insn. Indicate this by placing it in
5056 `src_related'. */
5058 if (elt)
5059 elt = elt->first_same_value;
5060 for (p = elt; p; p = p->next_same_value)
5062 enum rtx_code code = GET_CODE (p->exp);
5064 /* If the expression is not valid, ignore it. Then we do not
5065 have to check for validity below. In most cases, we can use
5066 `rtx_equal_p', since canonicalization has already been done. */
5067 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5068 continue;
5070 /* Also skip paradoxical subregs, unless that's what we're
5071 looking for. */
5072 if (paradoxical_subreg_p (p->exp)
5073 && ! (src != 0
5074 && GET_CODE (src) == SUBREG
5075 && GET_MODE (src) == GET_MODE (p->exp)
5076 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5077 GET_MODE (SUBREG_REG (p->exp)))))
5078 continue;
5080 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5081 src = 0;
5082 else if (src_folded && GET_CODE (src_folded) == code
5083 && rtx_equal_p (src_folded, p->exp))
5084 src_folded = 0;
5085 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5086 && rtx_equal_p (src_eqv_here, p->exp))
5087 src_eqv_here = 0;
5088 else if (src_related && GET_CODE (src_related) == code
5089 && rtx_equal_p (src_related, p->exp))
5090 src_related = 0;
5092 /* This is the same as the destination of the insns, we want
5093 to prefer it. The code below will then give it a negative
5094 cost. */
5095 if (!dest_related
5096 && GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5097 dest_related = p->exp;
5100 /* Find the cheapest valid equivalent, trying all the available
5101 possibilities. Prefer items not in the hash table to ones
5102 that are when they are equal cost. Note that we can never
5103 worsen an insn as the current contents will also succeed.
5104 If we find an equivalent identical to the destination, use it as best,
5105 since this insn will probably be eliminated in that case. */
5106 if (src)
5108 if (rtx_equal_p (src, dest))
5109 src_cost = src_regcost = -1;
5110 else
5112 src_cost = COST (src, mode);
5113 src_regcost = approx_reg_cost (src);
5117 if (src_eqv_here)
5119 if (rtx_equal_p (src_eqv_here, dest))
5120 src_eqv_cost = src_eqv_regcost = -1;
5121 else
5123 src_eqv_cost = COST (src_eqv_here, mode);
5124 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5128 if (src_folded)
5130 if (rtx_equal_p (src_folded, dest))
5131 src_folded_cost = src_folded_regcost = -1;
5132 else
5134 src_folded_cost = COST (src_folded, mode);
5135 src_folded_regcost = approx_reg_cost (src_folded);
5139 if (dest_related)
5141 src_related_cost = src_related_regcost = -1;
5142 /* Handle it as src_related. */
5143 src_related = dest_related;
5145 else if (src_related)
5147 src_related_cost = COST (src_related, mode);
5148 src_related_regcost = approx_reg_cost (src_related);
5150 /* If a const-anchor is used to synthesize a constant that
5151 normally requires multiple instructions then slightly prefer
5152 it over the original sequence. These instructions are likely
5153 to become redundant now. We can't compare against the cost
5154 of src_eqv_here because, on MIPS for example, multi-insn
5155 constants have zero cost; they are assumed to be hoisted from
5156 loops. */
5157 if (src_related_is_const_anchor
5158 && src_related_cost == src_cost
5159 && src_eqv_here)
5160 src_related_cost--;
5163 /* If this was an indirect jump insn, a known label will really be
5164 cheaper even though it looks more expensive. */
5165 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5166 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5168 /* Terminate loop when replacement made. This must terminate since
5169 the current contents will be tested and will always be valid. */
5170 while (!is_fake_set)
5172 rtx trial;
5174 /* Skip invalid entries. */
5175 while (elt && !REG_P (elt->exp)
5176 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5177 elt = elt->next_same_value;
5179 /* A paradoxical subreg would be bad here: it'll be the right
5180 size, but later may be adjusted so that the upper bits aren't
5181 what we want. So reject it. */
5182 if (elt != 0
5183 && paradoxical_subreg_p (elt->exp)
5184 /* It is okay, though, if the rtx we're trying to match
5185 will ignore any of the bits we can't predict. */
5186 && ! (src != 0
5187 && GET_CODE (src) == SUBREG
5188 && GET_MODE (src) == GET_MODE (elt->exp)
5189 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5190 GET_MODE (SUBREG_REG (elt->exp)))))
5192 elt = elt->next_same_value;
5193 continue;
5196 if (elt)
5198 src_elt_cost = elt->cost;
5199 src_elt_regcost = elt->regcost;
5202 /* Find cheapest and skip it for the next time. For items
5203 of equal cost, use this order:
5204 src_folded, src, src_eqv, src_related and hash table entry. */
5205 if (src_folded
5206 && preferable (src_folded_cost, src_folded_regcost,
5207 src_cost, src_regcost) <= 0
5208 && preferable (src_folded_cost, src_folded_regcost,
5209 src_eqv_cost, src_eqv_regcost) <= 0
5210 && preferable (src_folded_cost, src_folded_regcost,
5211 src_related_cost, src_related_regcost) <= 0
5212 && preferable (src_folded_cost, src_folded_regcost,
5213 src_elt_cost, src_elt_regcost) <= 0)
5214 trial = src_folded, src_folded_cost = MAX_COST;
5215 else if (src
5216 && preferable (src_cost, src_regcost,
5217 src_eqv_cost, src_eqv_regcost) <= 0
5218 && preferable (src_cost, src_regcost,
5219 src_related_cost, src_related_regcost) <= 0
5220 && preferable (src_cost, src_regcost,
5221 src_elt_cost, src_elt_regcost) <= 0)
5222 trial = src, src_cost = MAX_COST;
5223 else if (src_eqv_here
5224 && preferable (src_eqv_cost, src_eqv_regcost,
5225 src_related_cost, src_related_regcost) <= 0
5226 && preferable (src_eqv_cost, src_eqv_regcost,
5227 src_elt_cost, src_elt_regcost) <= 0)
5228 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5229 else if (src_related
5230 && preferable (src_related_cost, src_related_regcost,
5231 src_elt_cost, src_elt_regcost) <= 0)
5232 trial = src_related, src_related_cost = MAX_COST;
5233 else
5235 trial = elt->exp;
5236 elt = elt->next_same_value;
5237 src_elt_cost = MAX_COST;
5240 /* Try to optimize
5241 (set (reg:M N) (const_int A))
5242 (set (reg:M2 O) (const_int B))
5243 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5244 (reg:M2 O)). */
5245 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5246 && CONST_INT_P (trial)
5247 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5248 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5249 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5250 && (known_ge
5251 (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl))),
5252 INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))))
5253 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5254 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5255 <= HOST_BITS_PER_WIDE_INT))
5257 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5258 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5259 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5260 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5261 struct table_elt *dest_elt
5262 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5263 rtx dest_cst = NULL;
5265 if (dest_elt)
5266 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5267 if (p->is_const && CONST_INT_P (p->exp))
5269 dest_cst = p->exp;
5270 break;
5272 if (dest_cst)
5274 HOST_WIDE_INT val = INTVAL (dest_cst);
5275 HOST_WIDE_INT mask;
5276 unsigned int shift;
5277 /* This is the mode of DEST_CST as well. */
5278 scalar_int_mode dest_mode
5279 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
5280 if (BITS_BIG_ENDIAN)
5281 shift = GET_MODE_PRECISION (dest_mode)
5282 - INTVAL (pos) - INTVAL (width);
5283 else
5284 shift = INTVAL (pos);
5285 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5286 mask = HOST_WIDE_INT_M1;
5287 else
5288 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
5289 val &= ~(mask << shift);
5290 val |= (INTVAL (trial) & mask) << shift;
5291 val = trunc_int_for_mode (val, dest_mode);
5292 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5293 dest_reg, 1);
5294 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5295 GEN_INT (val), 1);
5296 if (apply_change_group ())
5298 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5299 if (note)
5301 remove_note (insn, note);
5302 df_notes_rescan (insn);
5304 src_eqv = NULL_RTX;
5305 src_eqv_elt = NULL;
5306 src_eqv_volatile = 0;
5307 src_eqv_in_memory = 0;
5308 src_eqv_hash = 0;
5309 repeat = true;
5310 break;
5315 /* We don't normally have an insn matching (set (pc) (pc)), so
5316 check for this separately here. We will delete such an
5317 insn below.
5319 For other cases such as a table jump or conditional jump
5320 where we know the ultimate target, go ahead and replace the
5321 operand. While that may not make a valid insn, we will
5322 reemit the jump below (and also insert any necessary
5323 barriers). */
5324 if (n_sets == 1 && dest == pc_rtx
5325 && (trial == pc_rtx
5326 || (GET_CODE (trial) == LABEL_REF
5327 && ! condjump_p (insn))))
5329 /* Don't substitute non-local labels, this confuses CFG. */
5330 if (GET_CODE (trial) == LABEL_REF
5331 && LABEL_REF_NONLOCAL_P (trial))
5332 continue;
5334 SET_SRC (sets[i].rtl) = trial;
5335 cse_jumps_altered = true;
5336 break;
5339 /* Similarly, lots of targets don't allow no-op
5340 (set (mem x) (mem x)) moves. Even (set (reg x) (reg x))
5341 might be impossible for certain registers (like CC registers). */
5342 else if (n_sets == 1
5343 && !CALL_P (insn)
5344 && (MEM_P (trial) || REG_P (trial))
5345 && rtx_equal_p (trial, dest)
5346 && !side_effects_p (dest)
5347 && (cfun->can_delete_dead_exceptions
5348 || insn_nothrow_p (insn))
5349 /* We can only remove the later store if the earlier aliases
5350 at least all accesses the later one. */
5351 && (!MEM_P (trial)
5352 || ((MEM_ALIAS_SET (dest) == MEM_ALIAS_SET (trial)
5353 || alias_set_subset_of (MEM_ALIAS_SET (dest),
5354 MEM_ALIAS_SET (trial)))
5355 && (!MEM_EXPR (trial)
5356 || refs_same_for_tbaa_p (MEM_EXPR (trial),
5357 MEM_EXPR (dest))))))
5359 SET_SRC (sets[i].rtl) = trial;
5360 noop_insn = true;
5361 break;
5364 /* Reject certain invalid forms of CONST that we create. */
5365 else if (CONSTANT_P (trial)
5366 && GET_CODE (trial) == CONST
5367 /* Reject cases that will cause decode_rtx_const to
5368 die. On the alpha when simplifying a switch, we
5369 get (const (truncate (minus (label_ref)
5370 (label_ref)))). */
5371 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5372 /* Likewise on IA-64, except without the
5373 truncate. */
5374 || (GET_CODE (XEXP (trial, 0)) == MINUS
5375 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5376 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5377 /* Do nothing for this case. */
5380 /* Do not replace anything with a MEM, except the replacement
5381 is a no-op. This allows this loop to terminate. */
5382 else if (MEM_P (trial) && !rtx_equal_p (trial, SET_SRC(sets[i].rtl)))
5383 /* Do nothing for this case. */
5386 /* Look for a substitution that makes a valid insn. */
5387 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5388 trial, 0))
5390 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5392 /* The result of apply_change_group can be ignored; see
5393 canon_reg. */
5395 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5396 apply_change_group ();
5398 break;
5401 /* If the current function uses a constant pool and this is a
5402 constant, try making a pool entry. Put it in src_folded
5403 unless we already have done this since that is where it
5404 likely came from. */
5406 else if (crtl->uses_const_pool
5407 && CONSTANT_P (trial)
5408 && !CONST_INT_P (trial)
5409 && (src_folded == 0 || !MEM_P (src_folded))
5410 && GET_MODE_CLASS (mode) != MODE_CC
5411 && mode != VOIDmode)
5413 src_folded = force_const_mem (mode, trial);
5414 if (src_folded)
5416 src_folded_cost = COST (src_folded, mode);
5417 src_folded_regcost = approx_reg_cost (src_folded);
5422 /* If we changed the insn too much, handle this set from scratch. */
5423 if (repeat)
5425 i--;
5426 continue;
5429 src = SET_SRC (sets[i].rtl);
5431 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5432 However, there is an important exception: If both are registers
5433 that are not the head of their equivalence class, replace SET_SRC
5434 with the head of the class. If we do not do this, we will have
5435 both registers live over a portion of the basic block. This way,
5436 their lifetimes will likely abut instead of overlapping. */
5437 if (!is_fake_set
5438 && REG_P (dest)
5439 && REGNO_QTY_VALID_P (REGNO (dest)))
5441 int dest_q = REG_QTY (REGNO (dest));
5442 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5444 if (dest_ent->mode == GET_MODE (dest)
5445 && dest_ent->first_reg != REGNO (dest)
5446 && REG_P (src) && REGNO (src) == REGNO (dest)
5447 /* Don't do this if the original insn had a hard reg as
5448 SET_SRC or SET_DEST. */
5449 && (!REG_P (sets[i].src)
5450 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5451 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5452 /* We can't call canon_reg here because it won't do anything if
5453 SRC is a hard register. */
5455 int src_q = REG_QTY (REGNO (src));
5456 struct qty_table_elem *src_ent = &qty_table[src_q];
5457 int first = src_ent->first_reg;
5458 rtx new_src
5459 = (first >= FIRST_PSEUDO_REGISTER
5460 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5462 /* We must use validate-change even for this, because this
5463 might be a special no-op instruction, suitable only to
5464 tag notes onto. */
5465 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5467 src = new_src;
5468 /* If we had a constant that is cheaper than what we are now
5469 setting SRC to, use that constant. We ignored it when we
5470 thought we could make this into a no-op. */
5471 if (src_const && COST (src_const, mode) < COST (src, mode)
5472 && validate_change (insn, &SET_SRC (sets[i].rtl),
5473 src_const, 0))
5474 src = src_const;
5479 /* If we made a change, recompute SRC values. */
5480 if (src != sets[i].src)
5482 do_not_record = 0;
5483 hash_arg_in_memory = 0;
5484 sets[i].src = src;
5485 sets[i].src_hash = HASH (src, mode);
5486 sets[i].src_volatile = do_not_record;
5487 sets[i].src_in_memory = hash_arg_in_memory;
5488 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5491 /* If this is a single SET, we are setting a register, and we have an
5492 equivalent constant, we want to add a REG_EQUAL note if the constant
5493 is different from the source. We don't want to do it for a constant
5494 pseudo since verifying that this pseudo hasn't been eliminated is a
5495 pain; moreover such a note won't help anything.
5497 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5498 which can be created for a reference to a compile time computable
5499 entry in a jump table. */
5500 if (n_sets == 1
5501 && REG_P (dest)
5502 && src_const
5503 && !REG_P (src_const)
5504 && !(GET_CODE (src_const) == SUBREG
5505 && REG_P (SUBREG_REG (src_const)))
5506 && !(GET_CODE (src_const) == CONST
5507 && GET_CODE (XEXP (src_const, 0)) == MINUS
5508 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5509 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5510 && !rtx_equal_p (src, src_const))
5512 /* Make sure that the rtx is not shared. */
5513 src_const = copy_rtx (src_const);
5515 /* Record the actual constant value in a REG_EQUAL note,
5516 making a new one if one does not already exist. */
5517 set_unique_reg_note (insn, REG_EQUAL, src_const);
5518 df_notes_rescan (insn);
5521 /* Now deal with the destination. */
5522 do_not_record = 0;
5524 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5525 while (GET_CODE (dest) == SUBREG
5526 || GET_CODE (dest) == ZERO_EXTRACT
5527 || GET_CODE (dest) == STRICT_LOW_PART)
5528 dest = XEXP (dest, 0);
5530 sets[i].inner_dest = dest;
5532 if (MEM_P (dest))
5534 #ifdef PUSH_ROUNDING
5535 /* Stack pushes invalidate the stack pointer. */
5536 rtx addr = XEXP (dest, 0);
5537 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5538 && XEXP (addr, 0) == stack_pointer_rtx)
5539 invalidate (stack_pointer_rtx, VOIDmode);
5540 #endif
5541 dest = fold_rtx (dest, insn);
5544 /* Compute the hash code of the destination now,
5545 before the effects of this instruction are recorded,
5546 since the register values used in the address computation
5547 are those before this instruction. */
5548 sets[i].dest_hash = HASH (dest, mode);
5550 /* Don't enter a bit-field in the hash table
5551 because the value in it after the store
5552 may not equal what was stored, due to truncation. */
5554 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5556 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5558 if (src_const != 0 && CONST_INT_P (src_const)
5559 && CONST_INT_P (width)
5560 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5561 && ! (INTVAL (src_const)
5562 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5563 /* Exception: if the value is constant,
5564 and it won't be truncated, record it. */
5566 else
5568 /* This is chosen so that the destination will be invalidated
5569 but no new value will be recorded.
5570 We must invalidate because sometimes constant
5571 values can be recorded for bitfields. */
5572 sets[i].src_elt = 0;
5573 sets[i].src_volatile = 1;
5574 src_eqv = 0;
5575 src_eqv_elt = 0;
5579 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5580 the insn. */
5581 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5583 /* One less use of the label this insn used to jump to. */
5584 cse_cfg_altered |= delete_insn_and_edges (insn);
5585 cse_jumps_altered = true;
5586 /* No more processing for this set. */
5587 sets[i].rtl = 0;
5590 /* Similarly for no-op moves. */
5591 else if (noop_insn)
5593 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5594 cse_cfg_altered = true;
5595 cse_cfg_altered |= delete_insn_and_edges (insn);
5596 /* No more processing for this set. */
5597 sets[i].rtl = 0;
5600 /* If this SET is now setting PC to a label, we know it used to
5601 be a conditional or computed branch. */
5602 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5603 && !LABEL_REF_NONLOCAL_P (src))
5605 /* We reemit the jump in as many cases as possible just in
5606 case the form of an unconditional jump is significantly
5607 different than a computed jump or conditional jump.
5609 If this insn has multiple sets, then reemitting the
5610 jump is nontrivial. So instead we just force rerecognition
5611 and hope for the best. */
5612 if (n_sets == 1)
5614 rtx_jump_insn *new_rtx;
5615 rtx note;
5617 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5618 new_rtx = emit_jump_insn_before (seq, insn);
5619 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5620 LABEL_NUSES (XEXP (src, 0))++;
5622 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5623 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5624 if (note)
5626 XEXP (note, 1) = NULL_RTX;
5627 REG_NOTES (new_rtx) = note;
5630 cse_cfg_altered |= delete_insn_and_edges (insn);
5631 insn = new_rtx;
5633 else
5634 INSN_CODE (insn) = -1;
5636 /* Do not bother deleting any unreachable code, let jump do it. */
5637 cse_jumps_altered = true;
5638 sets[i].rtl = 0;
5641 /* If destination is volatile, invalidate it and then do no further
5642 processing for this assignment. */
5644 else if (do_not_record)
5646 invalidate_dest (dest);
5647 sets[i].rtl = 0;
5650 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5652 do_not_record = 0;
5653 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5654 if (do_not_record)
5656 invalidate_dest (SET_DEST (sets[i].rtl));
5657 sets[i].rtl = 0;
5662 /* Now enter all non-volatile source expressions in the hash table
5663 if they are not already present.
5664 Record their equivalence classes in src_elt.
5665 This way we can insert the corresponding destinations into
5666 the same classes even if the actual sources are no longer in them
5667 (having been invalidated). */
5669 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5670 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5672 struct table_elt *elt;
5673 struct table_elt *classp = sets[0].src_elt;
5674 rtx dest = SET_DEST (sets[0].rtl);
5675 machine_mode eqvmode = GET_MODE (dest);
5677 if (GET_CODE (dest) == STRICT_LOW_PART)
5679 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5680 classp = 0;
5682 if (insert_regs (src_eqv, classp, false))
5684 rehash_using_reg (src_eqv);
5685 src_eqv_hash = HASH (src_eqv, eqvmode);
5687 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5688 elt->in_memory = src_eqv_in_memory;
5689 src_eqv_elt = elt;
5691 /* Check to see if src_eqv_elt is the same as a set source which
5692 does not yet have an elt, and if so set the elt of the set source
5693 to src_eqv_elt. */
5694 for (i = 0; i < n_sets; i++)
5695 if (sets[i].rtl && sets[i].src_elt == 0
5696 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5697 sets[i].src_elt = src_eqv_elt;
5700 for (i = 0; i < n_sets; i++)
5701 if (sets[i].rtl && ! sets[i].src_volatile
5702 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5704 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5706 /* REG_EQUAL in setting a STRICT_LOW_PART
5707 gives an equivalent for the entire destination register,
5708 not just for the subreg being stored in now.
5709 This is a more interesting equivalence, so we arrange later
5710 to treat the entire reg as the destination. */
5711 sets[i].src_elt = src_eqv_elt;
5712 sets[i].src_hash = src_eqv_hash;
5714 else
5716 /* Insert source and constant equivalent into hash table, if not
5717 already present. */
5718 struct table_elt *classp = src_eqv_elt;
5719 rtx src = sets[i].src;
5720 rtx dest = SET_DEST (sets[i].rtl);
5721 machine_mode mode
5722 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5724 /* It's possible that we have a source value known to be
5725 constant but don't have a REG_EQUAL note on the insn.
5726 Lack of a note will mean src_eqv_elt will be NULL. This
5727 can happen where we've generated a SUBREG to access a
5728 CONST_INT that is already in a register in a wider mode.
5729 Ensure that the source expression is put in the proper
5730 constant class. */
5731 if (!classp)
5732 classp = sets[i].src_const_elt;
5734 if (sets[i].src_elt == 0)
5736 struct table_elt *elt;
5738 /* Note that these insert_regs calls cannot remove
5739 any of the src_elt's, because they would have failed to
5740 match if not still valid. */
5741 if (insert_regs (src, classp, false))
5743 rehash_using_reg (src);
5744 sets[i].src_hash = HASH (src, mode);
5746 elt = insert (src, classp, sets[i].src_hash, mode);
5747 elt->in_memory = sets[i].src_in_memory;
5748 /* If inline asm has any clobbers, ensure we only reuse
5749 existing inline asms and never try to put the ASM_OPERANDS
5750 into an insn that isn't inline asm. */
5751 if (GET_CODE (src) == ASM_OPERANDS
5752 && GET_CODE (x) == PARALLEL)
5753 elt->cost = MAX_COST;
5754 sets[i].src_elt = classp = elt;
5756 if (sets[i].src_const && sets[i].src_const_elt == 0
5757 && src != sets[i].src_const
5758 && ! rtx_equal_p (sets[i].src_const, src))
5759 sets[i].src_elt = insert (sets[i].src_const, classp,
5760 sets[i].src_const_hash, mode);
5763 else if (sets[i].src_elt == 0)
5764 /* If we did not insert the source into the hash table (e.g., it was
5765 volatile), note the equivalence class for the REG_EQUAL value, if any,
5766 so that the destination goes into that class. */
5767 sets[i].src_elt = src_eqv_elt;
5769 /* Record destination addresses in the hash table. This allows us to
5770 check if they are invalidated by other sets. */
5771 for (i = 0; i < n_sets; i++)
5773 if (sets[i].rtl)
5775 rtx x = sets[i].inner_dest;
5776 struct table_elt *elt;
5777 machine_mode mode;
5778 unsigned hash;
5780 if (MEM_P (x))
5782 x = XEXP (x, 0);
5783 mode = GET_MODE (x);
5784 hash = HASH (x, mode);
5785 elt = lookup (x, hash, mode);
5786 if (!elt)
5788 if (insert_regs (x, NULL, false))
5790 rtx dest = SET_DEST (sets[i].rtl);
5792 rehash_using_reg (x);
5793 hash = HASH (x, mode);
5794 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5796 elt = insert (x, NULL, hash, mode);
5799 sets[i].dest_addr_elt = elt;
5801 else
5802 sets[i].dest_addr_elt = NULL;
5806 invalidate_from_clobbers (insn);
5808 /* Some registers are invalidated by subroutine calls. Memory is
5809 invalidated by non-constant calls. */
5811 if (CALL_P (insn))
5813 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5814 invalidate_memory ();
5815 else
5816 /* For const/pure calls, invalidate any argument slots, because
5817 those are owned by the callee. */
5818 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
5819 if (GET_CODE (XEXP (tem, 0)) == USE
5820 && MEM_P (XEXP (XEXP (tem, 0), 0)))
5821 invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode);
5822 invalidate_for_call (insn);
5825 /* Now invalidate everything set by this instruction.
5826 If a SUBREG or other funny destination is being set,
5827 sets[i].rtl is still nonzero, so here we invalidate the reg
5828 a part of which is being set. */
5830 for (i = 0; i < n_sets; i++)
5831 if (sets[i].rtl)
5833 /* We can't use the inner dest, because the mode associated with
5834 a ZERO_EXTRACT is significant. */
5835 rtx dest = SET_DEST (sets[i].rtl);
5837 /* Needed for registers to remove the register from its
5838 previous quantity's chain.
5839 Needed for memory if this is a nonvarying address, unless
5840 we have just done an invalidate_memory that covers even those. */
5841 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5842 invalidate (dest, VOIDmode);
5843 else if (MEM_P (dest))
5844 invalidate (dest, VOIDmode);
5845 else if (GET_CODE (dest) == STRICT_LOW_PART
5846 || GET_CODE (dest) == ZERO_EXTRACT)
5847 invalidate (XEXP (dest, 0), GET_MODE (dest));
5850 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5851 the regs restored by the longjmp come from a later time
5852 than the setjmp. */
5853 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5855 flush_hash_table ();
5856 goto done;
5859 /* Make sure registers mentioned in destinations
5860 are safe for use in an expression to be inserted.
5861 This removes from the hash table
5862 any invalid entry that refers to one of these registers.
5864 We don't care about the return value from mention_regs because
5865 we are going to hash the SET_DEST values unconditionally. */
5867 for (i = 0; i < n_sets; i++)
5869 if (sets[i].rtl)
5871 rtx x = SET_DEST (sets[i].rtl);
5873 if (!REG_P (x))
5874 mention_regs (x);
5875 else
5877 /* We used to rely on all references to a register becoming
5878 inaccessible when a register changes to a new quantity,
5879 since that changes the hash code. However, that is not
5880 safe, since after HASH_SIZE new quantities we get a
5881 hash 'collision' of a register with its own invalid
5882 entries. And since SUBREGs have been changed not to
5883 change their hash code with the hash code of the register,
5884 it wouldn't work any longer at all. So we have to check
5885 for any invalid references lying around now.
5886 This code is similar to the REG case in mention_regs,
5887 but it knows that reg_tick has been incremented, and
5888 it leaves reg_in_table as -1 . */
5889 unsigned int regno = REGNO (x);
5890 unsigned int endregno = END_REGNO (x);
5891 unsigned int i;
5893 for (i = regno; i < endregno; i++)
5895 if (REG_IN_TABLE (i) >= 0)
5897 remove_invalid_refs (i);
5898 REG_IN_TABLE (i) = -1;
5905 /* We may have just removed some of the src_elt's from the hash table.
5906 So replace each one with the current head of the same class.
5907 Also check if destination addresses have been removed. */
5909 for (i = 0; i < n_sets; i++)
5910 if (sets[i].rtl)
5912 if (sets[i].dest_addr_elt
5913 && sets[i].dest_addr_elt->first_same_value == 0)
5915 /* The elt was removed, which means this destination is not
5916 valid after this instruction. */
5917 sets[i].rtl = NULL_RTX;
5919 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5920 /* If elt was removed, find current head of same class,
5921 or 0 if nothing remains of that class. */
5923 struct table_elt *elt = sets[i].src_elt;
5925 while (elt && elt->prev_same_value)
5926 elt = elt->prev_same_value;
5928 while (elt && elt->first_same_value == 0)
5929 elt = elt->next_same_value;
5930 sets[i].src_elt = elt ? elt->first_same_value : 0;
5934 /* Now insert the destinations into their equivalence classes. */
5936 for (i = 0; i < n_sets; i++)
5937 if (sets[i].rtl)
5939 rtx dest = SET_DEST (sets[i].rtl);
5940 struct table_elt *elt;
5942 /* Don't record value if we are not supposed to risk allocating
5943 floating-point values in registers that might be wider than
5944 memory. */
5945 if ((flag_float_store
5946 && MEM_P (dest)
5947 && FLOAT_MODE_P (GET_MODE (dest)))
5948 /* Don't record BLKmode values, because we don't know the
5949 size of it, and can't be sure that other BLKmode values
5950 have the same or smaller size. */
5951 || GET_MODE (dest) == BLKmode
5952 /* If we didn't put a REG_EQUAL value or a source into the hash
5953 table, there is no point is recording DEST. */
5954 || sets[i].src_elt == 0)
5955 continue;
5957 /* STRICT_LOW_PART isn't part of the value BEING set,
5958 and neither is the SUBREG inside it.
5959 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5960 if (GET_CODE (dest) == STRICT_LOW_PART)
5961 dest = SUBREG_REG (XEXP (dest, 0));
5963 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5964 /* Registers must also be inserted into chains for quantities. */
5965 if (insert_regs (dest, sets[i].src_elt, true))
5967 /* If `insert_regs' changes something, the hash code must be
5968 recalculated. */
5969 rehash_using_reg (dest);
5970 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5973 /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
5974 outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
5975 if (paradoxical_subreg_p (dest))
5976 continue;
5978 elt = insert (dest, sets[i].src_elt,
5979 sets[i].dest_hash, GET_MODE (dest));
5981 /* If this is a constant, insert the constant anchors with the
5982 equivalent register-offset expressions using register DEST. */
5983 if (targetm.const_anchor
5984 && REG_P (dest)
5985 && SCALAR_INT_MODE_P (GET_MODE (dest))
5986 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5987 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5989 elt->in_memory = (MEM_P (sets[i].inner_dest)
5990 && !MEM_READONLY_P (sets[i].inner_dest));
5992 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5993 narrower than M2, and both M1 and M2 are the same number of words,
5994 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5995 make that equivalence as well.
5997 However, BAR may have equivalences for which gen_lowpart
5998 will produce a simpler value than gen_lowpart applied to
5999 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6000 BAR's equivalences. If we don't get a simplified form, make
6001 the SUBREG. It will not be used in an equivalence, but will
6002 cause two similar assignments to be detected.
6004 Note the loop below will find SUBREG_REG (DEST) since we have
6005 already entered SRC and DEST of the SET in the table. */
6007 if (GET_CODE (dest) == SUBREG
6008 && (known_equal_after_align_down
6009 (GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1,
6010 GET_MODE_SIZE (GET_MODE (dest)) - 1,
6011 UNITS_PER_WORD))
6012 && !partial_subreg_p (dest)
6013 && sets[i].src_elt != 0)
6015 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6016 struct table_elt *elt, *classp = 0;
6018 for (elt = sets[i].src_elt->first_same_value; elt;
6019 elt = elt->next_same_value)
6021 rtx new_src = 0;
6022 unsigned src_hash;
6023 struct table_elt *src_elt;
6025 /* Ignore invalid entries. */
6026 if (!REG_P (elt->exp)
6027 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6028 continue;
6030 /* We may have already been playing subreg games. If the
6031 mode is already correct for the destination, use it. */
6032 if (GET_MODE (elt->exp) == new_mode)
6033 new_src = elt->exp;
6034 else
6036 poly_uint64 byte
6037 = subreg_lowpart_offset (new_mode, GET_MODE (dest));
6038 new_src = simplify_gen_subreg (new_mode, elt->exp,
6039 GET_MODE (dest), byte);
6042 /* The call to simplify_gen_subreg fails if the value
6043 is VOIDmode, yet we can't do any simplification, e.g.
6044 for EXPR_LISTs denoting function call results.
6045 It is invalid to construct a SUBREG with a VOIDmode
6046 SUBREG_REG, hence a zero new_src means we can't do
6047 this substitution. */
6048 if (! new_src)
6049 continue;
6051 src_hash = HASH (new_src, new_mode);
6052 src_elt = lookup (new_src, src_hash, new_mode);
6054 /* Put the new source in the hash table is if isn't
6055 already. */
6056 if (src_elt == 0)
6058 if (insert_regs (new_src, classp, false))
6060 rehash_using_reg (new_src);
6061 src_hash = HASH (new_src, new_mode);
6063 src_elt = insert (new_src, classp, src_hash, new_mode);
6064 src_elt->in_memory = elt->in_memory;
6065 if (GET_CODE (new_src) == ASM_OPERANDS
6066 && elt->cost == MAX_COST)
6067 src_elt->cost = MAX_COST;
6069 else if (classp && classp != src_elt->first_same_value)
6070 /* Show that two things that we've seen before are
6071 actually the same. */
6072 merge_equiv_classes (src_elt, classp);
6074 classp = src_elt->first_same_value;
6075 /* Ignore invalid entries. */
6076 while (classp
6077 && !REG_P (classp->exp)
6078 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6079 classp = classp->next_same_value;
6084 /* Special handling for (set REG0 REG1) where REG0 is the
6085 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6086 be used in the sequel, so (if easily done) change this insn to
6087 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6088 that computed their value. Then REG1 will become a dead store
6089 and won't cloud the situation for later optimizations.
6091 Do not make this change if REG1 is a hard register, because it will
6092 then be used in the sequel and we may be changing a two-operand insn
6093 into a three-operand insn.
6095 Also do not do this if we are operating on a copy of INSN. */
6097 if (n_sets == 1 && sets[0].rtl)
6098 try_back_substitute_reg (sets[0].rtl, insn);
6100 done:;
6103 /* Remove from the hash table all expressions that reference memory. */
6105 static void
6106 invalidate_memory (void)
6108 int i;
6109 struct table_elt *p, *next;
6111 for (i = 0; i < HASH_SIZE; i++)
6112 for (p = table[i]; p; p = next)
6114 next = p->next_same_hash;
6115 if (p->in_memory)
6116 remove_from_table (p, i);
6120 /* Perform invalidation on the basis of everything about INSN,
6121 except for invalidating the actual places that are SET in it.
6122 This includes the places CLOBBERed, and anything that might
6123 alias with something that is SET or CLOBBERed. */
6125 static void
6126 invalidate_from_clobbers (rtx_insn *insn)
6128 rtx x = PATTERN (insn);
6130 if (GET_CODE (x) == CLOBBER)
6132 rtx ref = XEXP (x, 0);
6133 if (ref)
6135 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6136 || MEM_P (ref))
6137 invalidate (ref, VOIDmode);
6138 else if (GET_CODE (ref) == STRICT_LOW_PART
6139 || GET_CODE (ref) == ZERO_EXTRACT)
6140 invalidate (XEXP (ref, 0), GET_MODE (ref));
6143 else if (GET_CODE (x) == PARALLEL)
6145 int i;
6146 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6148 rtx y = XVECEXP (x, 0, i);
6149 if (GET_CODE (y) == CLOBBER)
6151 rtx ref = XEXP (y, 0);
6152 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6153 || MEM_P (ref))
6154 invalidate (ref, VOIDmode);
6155 else if (GET_CODE (ref) == STRICT_LOW_PART
6156 || GET_CODE (ref) == ZERO_EXTRACT)
6157 invalidate (XEXP (ref, 0), GET_MODE (ref));
6163 /* Perform invalidation on the basis of everything about INSN.
6164 This includes the places CLOBBERed, and anything that might
6165 alias with something that is SET or CLOBBERed. */
6167 static void
6168 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6170 rtx tem;
6171 rtx x = PATTERN (insn);
6173 if (CALL_P (insn))
6175 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6177 rtx temx = XEXP (tem, 0);
6178 if (GET_CODE (temx) == CLOBBER)
6179 invalidate (SET_DEST (temx), VOIDmode);
6183 /* Ensure we invalidate the destination register of a CALL insn.
6184 This is necessary for machines where this register is a fixed_reg,
6185 because no other code would invalidate it. */
6186 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6187 invalidate (SET_DEST (x), VOIDmode);
6189 else if (GET_CODE (x) == PARALLEL)
6191 int i;
6193 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6195 rtx y = XVECEXP (x, 0, i);
6196 if (GET_CODE (y) == CLOBBER)
6198 rtx clobbered = XEXP (y, 0);
6200 if (REG_P (clobbered)
6201 || GET_CODE (clobbered) == SUBREG)
6202 invalidate (clobbered, VOIDmode);
6203 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6204 || GET_CODE (clobbered) == ZERO_EXTRACT)
6205 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6207 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6208 invalidate (SET_DEST (y), VOIDmode);
6213 static rtx cse_process_note (rtx);
6215 /* A simplify_replace_fn_rtx callback for cse_process_note. Process X,
6216 part of the REG_NOTES of an insn. Replace any registers with either
6217 an equivalent constant or the canonical form of the register.
6218 Only replace addresses if the containing MEM remains valid.
6220 Return the replacement for X, or null if it should be simplified
6221 recursively. */
6223 static rtx
6224 cse_process_note_1 (rtx x, const_rtx, void *)
6226 if (MEM_P (x))
6228 validate_change (x, &XEXP (x, 0), cse_process_note (XEXP (x, 0)), false);
6229 return x;
6232 if (REG_P (x))
6234 int i = REG_QTY (REGNO (x));
6236 /* Return a constant or a constant register. */
6237 if (REGNO_QTY_VALID_P (REGNO (x)))
6239 struct qty_table_elem *ent = &qty_table[i];
6241 if (ent->const_rtx != NULL_RTX
6242 && (CONSTANT_P (ent->const_rtx)
6243 || REG_P (ent->const_rtx)))
6245 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6246 if (new_rtx)
6247 return copy_rtx (new_rtx);
6251 /* Otherwise, canonicalize this register. */
6252 return canon_reg (x, NULL);
6255 return NULL_RTX;
6258 /* Process X, part of the REG_NOTES of an insn. Replace any registers in it
6259 with either an equivalent constant or the canonical form of the register.
6260 Only replace addresses if the containing MEM remains valid. */
6262 static rtx
6263 cse_process_note (rtx x)
6265 return simplify_replace_fn_rtx (x, NULL_RTX, cse_process_note_1, NULL);
6269 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6271 DATA is a pointer to a struct cse_basic_block_data, that is used to
6272 describe the path.
6273 It is filled with a queue of basic blocks, starting with FIRST_BB
6274 and following a trace through the CFG.
6276 If all paths starting at FIRST_BB have been followed, or no new path
6277 starting at FIRST_BB can be constructed, this function returns FALSE.
6278 Otherwise, DATA->path is filled and the function returns TRUE indicating
6279 that a path to follow was found.
6281 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6282 block in the path will be FIRST_BB. */
6284 static bool
6285 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6286 int follow_jumps)
6288 basic_block bb;
6289 edge e;
6290 int path_size;
6292 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6294 /* See if there is a previous path. */
6295 path_size = data->path_size;
6297 /* There is a previous path. Make sure it started with FIRST_BB. */
6298 if (path_size)
6299 gcc_assert (data->path[0].bb == first_bb);
6301 /* There was only one basic block in the last path. Clear the path and
6302 return, so that paths starting at another basic block can be tried. */
6303 if (path_size == 1)
6305 path_size = 0;
6306 goto done;
6309 /* If the path was empty from the beginning, construct a new path. */
6310 if (path_size == 0)
6311 data->path[path_size++].bb = first_bb;
6312 else
6314 /* Otherwise, path_size must be equal to or greater than 2, because
6315 a previous path exists that is at least two basic blocks long.
6317 Update the previous branch path, if any. If the last branch was
6318 previously along the branch edge, take the fallthrough edge now. */
6319 while (path_size >= 2)
6321 basic_block last_bb_in_path, previous_bb_in_path;
6322 edge e;
6324 --path_size;
6325 last_bb_in_path = data->path[path_size].bb;
6326 previous_bb_in_path = data->path[path_size - 1].bb;
6328 /* If we previously followed a path along the branch edge, try
6329 the fallthru edge now. */
6330 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6331 && any_condjump_p (BB_END (previous_bb_in_path))
6332 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6333 && e == BRANCH_EDGE (previous_bb_in_path))
6335 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6336 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6337 && single_pred_p (bb)
6338 /* We used to assert here that we would only see blocks
6339 that we have not visited yet. But we may end up
6340 visiting basic blocks twice if the CFG has changed
6341 in this run of cse_main, because when the CFG changes
6342 the topological sort of the CFG also changes. A basic
6343 blocks that previously had more than two predecessors
6344 may now have a single predecessor, and become part of
6345 a path that starts at another basic block.
6347 We still want to visit each basic block only once, so
6348 halt the path here if we have already visited BB. */
6349 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6351 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6352 data->path[path_size++].bb = bb;
6353 break;
6357 data->path[path_size].bb = NULL;
6360 /* If only one block remains in the path, bail. */
6361 if (path_size == 1)
6363 path_size = 0;
6364 goto done;
6368 /* Extend the path if possible. */
6369 if (follow_jumps)
6371 bb = data->path[path_size - 1].bb;
6372 while (bb && path_size < param_max_cse_path_length)
6374 if (single_succ_p (bb))
6375 e = single_succ_edge (bb);
6376 else if (EDGE_COUNT (bb->succs) == 2
6377 && any_condjump_p (BB_END (bb)))
6379 /* First try to follow the branch. If that doesn't lead
6380 to a useful path, follow the fallthru edge. */
6381 e = BRANCH_EDGE (bb);
6382 if (!single_pred_p (e->dest))
6383 e = FALLTHRU_EDGE (bb);
6385 else
6386 e = NULL;
6388 if (e
6389 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6390 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6391 && single_pred_p (e->dest)
6392 /* Avoid visiting basic blocks twice. The large comment
6393 above explains why this can happen. */
6394 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6396 basic_block bb2 = e->dest;
6397 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6398 data->path[path_size++].bb = bb2;
6399 bb = bb2;
6401 else
6402 bb = NULL;
6406 done:
6407 data->path_size = path_size;
6408 return path_size != 0;
6411 /* Dump the path in DATA to file F. NSETS is the number of sets
6412 in the path. */
6414 static void
6415 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6417 int path_entry;
6419 fprintf (f, ";; Following path with %d sets: ", nsets);
6420 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6421 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6422 fputc ('\n', f);
6423 fflush (f);
6427 /* Return true if BB has exception handling successor edges. */
6429 static bool
6430 have_eh_succ_edges (basic_block bb)
6432 edge e;
6433 edge_iterator ei;
6435 FOR_EACH_EDGE (e, ei, bb->succs)
6436 if (e->flags & EDGE_EH)
6437 return true;
6439 return false;
6443 /* Scan to the end of the path described by DATA. Return an estimate of
6444 the total number of SETs of all insns in the path. */
6446 static void
6447 cse_prescan_path (struct cse_basic_block_data *data)
6449 int nsets = 0;
6450 int path_size = data->path_size;
6451 int path_entry;
6453 /* Scan to end of each basic block in the path. */
6454 for (path_entry = 0; path_entry < path_size; path_entry++)
6456 basic_block bb;
6457 rtx_insn *insn;
6459 bb = data->path[path_entry].bb;
6461 FOR_BB_INSNS (bb, insn)
6463 if (!INSN_P (insn))
6464 continue;
6466 /* A PARALLEL can have lots of SETs in it,
6467 especially if it is really an ASM_OPERANDS. */
6468 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6469 nsets += XVECLEN (PATTERN (insn), 0);
6470 else
6471 nsets += 1;
6475 data->nsets = nsets;
6478 /* Return true if the pattern of INSN uses a LABEL_REF for which
6479 there isn't a REG_LABEL_OPERAND note. */
6481 static bool
6482 check_for_label_ref (rtx_insn *insn)
6484 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6485 note for it, we must rerun jump since it needs to place the note. If
6486 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6487 don't do this since no REG_LABEL_OPERAND will be added. */
6488 subrtx_iterator::array_type array;
6489 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6491 const_rtx x = *iter;
6492 if (GET_CODE (x) == LABEL_REF
6493 && !LABEL_REF_NONLOCAL_P (x)
6494 && (!JUMP_P (insn)
6495 || !label_is_jump_target_p (label_ref_label (x), insn))
6496 && LABEL_P (label_ref_label (x))
6497 && INSN_UID (label_ref_label (x)) != 0
6498 && !find_reg_note (insn, REG_LABEL_OPERAND, label_ref_label (x)))
6499 return true;
6501 return false;
6504 /* Process a single extended basic block described by EBB_DATA. */
6506 static void
6507 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6509 int path_size = ebb_data->path_size;
6510 int path_entry;
6511 int num_insns = 0;
6513 /* Allocate the space needed by qty_table. */
6514 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6516 new_basic_block ();
6517 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6518 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6519 for (path_entry = 0; path_entry < path_size; path_entry++)
6521 basic_block bb;
6522 rtx_insn *insn;
6524 bb = ebb_data->path[path_entry].bb;
6526 /* Invalidate recorded information for eh regs if there is an EH
6527 edge pointing to that bb. */
6528 if (bb_has_eh_pred (bb))
6530 df_ref def;
6532 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6533 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6534 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6537 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6538 FOR_BB_INSNS (bb, insn)
6540 /* If we have processed 1,000 insns, flush the hash table to
6541 avoid extreme quadratic behavior. We must not include NOTEs
6542 in the count since there may be more of them when generating
6543 debugging information. If we clear the table at different
6544 times, code generated with -g -O might be different than code
6545 generated with -O but not -g.
6547 FIXME: This is a real kludge and needs to be done some other
6548 way. */
6549 if (NONDEBUG_INSN_P (insn)
6550 && num_insns++ > param_max_cse_insns)
6552 flush_hash_table ();
6553 num_insns = 0;
6556 if (INSN_P (insn))
6558 /* Process notes first so we have all notes in canonical forms
6559 when looking for duplicate operations. */
6560 bool changed = false;
6561 for (rtx note = REG_NOTES (insn); note; note = XEXP (note, 1))
6562 if (REG_NOTE_KIND (note) == REG_EQUAL)
6564 rtx newval = cse_process_note (XEXP (note, 0));
6565 if (newval != XEXP (note, 0))
6567 XEXP (note, 0) = newval;
6568 changed = true;
6571 if (changed)
6572 df_notes_rescan (insn);
6574 cse_insn (insn);
6576 /* If we haven't already found an insn where we added a LABEL_REF,
6577 check this one. */
6578 if (INSN_P (insn) && !recorded_label_ref
6579 && check_for_label_ref (insn))
6580 recorded_label_ref = true;
6584 /* With non-call exceptions, we are not always able to update
6585 the CFG properly inside cse_insn. So clean up possibly
6586 redundant EH edges here. */
6587 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6588 cse_cfg_altered |= purge_dead_edges (bb);
6590 /* If we changed a conditional jump, we may have terminated
6591 the path we are following. Check that by verifying that
6592 the edge we would take still exists. If the edge does
6593 not exist anymore, purge the remainder of the path.
6594 Note that this will cause us to return to the caller. */
6595 if (path_entry < path_size - 1)
6597 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6598 if (!find_edge (bb, next_bb))
6602 path_size--;
6604 /* If we truncate the path, we must also reset the
6605 visited bit on the remaining blocks in the path,
6606 or we will never visit them at all. */
6607 bitmap_clear_bit (cse_visited_basic_blocks,
6608 ebb_data->path[path_size].bb->index);
6609 ebb_data->path[path_size].bb = NULL;
6611 while (path_size - 1 != path_entry);
6612 ebb_data->path_size = path_size;
6616 /* If this is a conditional jump insn, record any known
6617 equivalences due to the condition being tested. */
6618 insn = BB_END (bb);
6619 if (path_entry < path_size - 1
6620 && EDGE_COUNT (bb->succs) == 2
6621 && JUMP_P (insn)
6622 && single_set (insn)
6623 && any_condjump_p (insn))
6625 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6626 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6627 record_jump_equiv (insn, taken);
6631 gcc_assert (next_qty <= max_qty);
6633 free (qty_table);
6637 /* Perform cse on the instructions of a function.
6638 F is the first instruction.
6639 NREGS is one plus the highest pseudo-reg number used in the instruction.
6641 Return 2 if jump optimizations should be redone due to simplifications
6642 in conditional jump instructions.
6643 Return 1 if the CFG should be cleaned up because it has been modified.
6644 Return 0 otherwise. */
6646 static int
6647 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6649 struct cse_basic_block_data ebb_data;
6650 basic_block bb;
6651 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6652 int i, n_blocks;
6654 /* CSE doesn't use dominane info but can invalidate it in different ways.
6655 For simplicity free dominance info here. */
6656 free_dominance_info (CDI_DOMINATORS);
6658 df_set_flags (DF_LR_RUN_DCE);
6659 df_note_add_problem ();
6660 df_analyze ();
6661 df_set_flags (DF_DEFER_INSN_RESCAN);
6663 reg_scan (get_insns (), max_reg_num ());
6664 init_cse_reg_info (nregs);
6666 ebb_data.path = XNEWVEC (struct branch_path,
6667 param_max_cse_path_length);
6669 cse_cfg_altered = false;
6670 cse_jumps_altered = false;
6671 recorded_label_ref = false;
6672 ebb_data.path_size = 0;
6673 ebb_data.nsets = 0;
6674 rtl_hooks = cse_rtl_hooks;
6676 init_recog ();
6677 init_alias_analysis ();
6679 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6681 /* Set up the table of already visited basic blocks. */
6682 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6683 bitmap_clear (cse_visited_basic_blocks);
6685 /* Loop over basic blocks in reverse completion order (RPO),
6686 excluding the ENTRY and EXIT blocks. */
6687 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6688 i = 0;
6689 while (i < n_blocks)
6691 /* Find the first block in the RPO queue that we have not yet
6692 processed before. */
6695 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6697 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6698 && i < n_blocks);
6700 /* Find all paths starting with BB, and process them. */
6701 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6703 /* Pre-scan the path. */
6704 cse_prescan_path (&ebb_data);
6706 /* If this basic block has no sets, skip it. */
6707 if (ebb_data.nsets == 0)
6708 continue;
6710 /* Get a reasonable estimate for the maximum number of qty's
6711 needed for this path. For this, we take the number of sets
6712 and multiply that by MAX_RECOG_OPERANDS. */
6713 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6715 /* Dump the path we're about to process. */
6716 if (dump_file)
6717 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6719 cse_extended_basic_block (&ebb_data);
6723 /* Clean up. */
6724 end_alias_analysis ();
6725 free (reg_eqv_table);
6726 free (ebb_data.path);
6727 sbitmap_free (cse_visited_basic_blocks);
6728 free (rc_order);
6729 rtl_hooks = general_rtl_hooks;
6731 if (cse_jumps_altered || recorded_label_ref)
6732 return 2;
6733 else if (cse_cfg_altered)
6734 return 1;
6735 else
6736 return 0;
6739 /* Count the number of times registers are used (not set) in X.
6740 COUNTS is an array in which we accumulate the count, INCR is how much
6741 we count each register usage.
6743 Don't count a usage of DEST, which is the SET_DEST of a SET which
6744 contains X in its SET_SRC. This is because such a SET does not
6745 modify the liveness of DEST.
6746 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6747 We must then count uses of a SET_DEST regardless, because the insn can't be
6748 deleted here. */
6750 static void
6751 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6753 enum rtx_code code;
6754 rtx note;
6755 const char *fmt;
6756 int i, j;
6758 if (x == 0)
6759 return;
6761 switch (code = GET_CODE (x))
6763 case REG:
6764 if (x != dest)
6765 counts[REGNO (x)] += incr;
6766 return;
6768 case PC:
6769 case CONST:
6770 CASE_CONST_ANY:
6771 case SYMBOL_REF:
6772 case LABEL_REF:
6773 return;
6775 case CLOBBER:
6776 /* If we are clobbering a MEM, mark any registers inside the address
6777 as being used. */
6778 if (MEM_P (XEXP (x, 0)))
6779 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6780 return;
6782 case SET:
6783 /* Unless we are setting a REG, count everything in SET_DEST. */
6784 if (!REG_P (SET_DEST (x)))
6785 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6786 count_reg_usage (SET_SRC (x), counts,
6787 dest ? dest : SET_DEST (x),
6788 incr);
6789 return;
6791 case DEBUG_INSN:
6792 return;
6794 case CALL_INSN:
6795 case INSN:
6796 case JUMP_INSN:
6797 /* We expect dest to be NULL_RTX here. If the insn may throw,
6798 or if it cannot be deleted due to side-effects, mark this fact
6799 by setting DEST to pc_rtx. */
6800 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6801 || side_effects_p (PATTERN (x)))
6802 dest = pc_rtx;
6803 if (code == CALL_INSN)
6804 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6805 count_reg_usage (PATTERN (x), counts, dest, incr);
6807 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6808 use them. */
6810 note = find_reg_equal_equiv_note (x);
6811 if (note)
6813 rtx eqv = XEXP (note, 0);
6815 if (GET_CODE (eqv) == EXPR_LIST)
6816 /* This REG_EQUAL note describes the result of a function call.
6817 Process all the arguments. */
6820 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6821 eqv = XEXP (eqv, 1);
6823 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6824 else
6825 count_reg_usage (eqv, counts, dest, incr);
6827 return;
6829 case EXPR_LIST:
6830 if (REG_NOTE_KIND (x) == REG_EQUAL
6831 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6832 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6833 involving registers in the address. */
6834 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6835 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6837 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6838 return;
6840 case ASM_OPERANDS:
6841 /* Iterate over just the inputs, not the constraints as well. */
6842 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6843 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6844 return;
6846 case INSN_LIST:
6847 case INT_LIST:
6848 gcc_unreachable ();
6850 default:
6851 break;
6854 fmt = GET_RTX_FORMAT (code);
6855 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6857 if (fmt[i] == 'e')
6858 count_reg_usage (XEXP (x, i), counts, dest, incr);
6859 else if (fmt[i] == 'E')
6860 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6861 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6865 /* Return true if X is a dead register. */
6867 static inline bool
6868 is_dead_reg (const_rtx x, int *counts)
6870 return (REG_P (x)
6871 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6872 && counts[REGNO (x)] == 0);
6875 /* Return true if set is live. */
6876 static bool
6877 set_live_p (rtx set, int *counts)
6879 if (set_noop_p (set))
6880 return false;
6882 if (!is_dead_reg (SET_DEST (set), counts)
6883 || side_effects_p (SET_SRC (set)))
6884 return true;
6886 return false;
6889 /* Return true if insn is live. */
6891 static bool
6892 insn_live_p (rtx_insn *insn, int *counts)
6894 int i;
6895 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6896 return true;
6897 else if (GET_CODE (PATTERN (insn)) == SET)
6898 return set_live_p (PATTERN (insn), counts);
6899 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6901 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6903 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6905 if (GET_CODE (elt) == SET)
6907 if (set_live_p (elt, counts))
6908 return true;
6910 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6911 return true;
6913 return false;
6915 else if (DEBUG_INSN_P (insn))
6917 if (DEBUG_MARKER_INSN_P (insn))
6918 return true;
6920 if (DEBUG_BIND_INSN_P (insn)
6921 && TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)))
6922 return false;
6924 return true;
6926 else
6927 return true;
6930 /* Count the number of stores into pseudo. Callback for note_stores. */
6932 static void
6933 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6935 int *counts = (int *) data;
6936 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6937 counts[REGNO (x)]++;
6940 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6941 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6942 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6943 Set *SEEN_REPL to true if we see a dead register that does have
6944 a replacement. */
6946 static bool
6947 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6948 bool *seen_repl)
6950 subrtx_iterator::array_type array;
6951 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6953 const_rtx x = *iter;
6954 if (is_dead_reg (x, counts))
6956 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6957 *seen_repl = true;
6958 else
6959 return true;
6962 return false;
6965 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6966 Callback for simplify_replace_fn_rtx. */
6968 static rtx
6969 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6971 rtx *replacements = (rtx *) data;
6973 if (REG_P (x)
6974 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6975 && replacements[REGNO (x)] != NULL_RTX)
6977 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6978 return replacements[REGNO (x)];
6979 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6980 GET_MODE (replacements[REGNO (x)]));
6982 return NULL_RTX;
6985 /* Scan all the insns and delete any that are dead; i.e., they store a register
6986 that is never used or they copy a register to itself.
6988 This is used to remove insns made obviously dead by cse, loop or other
6989 optimizations. It improves the heuristics in loop since it won't try to
6990 move dead invariants out of loops or make givs for dead quantities. The
6991 remaining passes of the compilation are also sped up. */
6994 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
6996 int *counts;
6997 rtx_insn *insn, *prev;
6998 rtx *replacements = NULL;
6999 int ndead = 0;
7001 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7002 /* First count the number of times each register is used. */
7003 if (MAY_HAVE_DEBUG_BIND_INSNS)
7005 counts = XCNEWVEC (int, nreg * 3);
7006 for (insn = insns; insn; insn = NEXT_INSN (insn))
7007 if (DEBUG_BIND_INSN_P (insn))
7009 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7010 NULL_RTX, 1);
7011 TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)) = 0;
7013 else if (INSN_P (insn))
7015 count_reg_usage (insn, counts, NULL_RTX, 1);
7016 note_stores (insn, count_stores, counts + nreg * 2);
7018 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7019 First one counts how many times each pseudo is used outside
7020 of debug insns, second counts how many times each pseudo is
7021 used in debug insns and third counts how many times a pseudo
7022 is stored. */
7024 else
7026 counts = XCNEWVEC (int, nreg);
7027 for (insn = insns; insn; insn = NEXT_INSN (insn))
7028 if (INSN_P (insn))
7029 count_reg_usage (insn, counts, NULL_RTX, 1);
7030 /* If no debug insns can be present, COUNTS is just an array
7031 which counts how many times each pseudo is used. */
7033 /* Pseudo PIC register should be considered as used due to possible
7034 new usages generated. */
7035 if (!reload_completed
7036 && pic_offset_table_rtx
7037 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7038 counts[REGNO (pic_offset_table_rtx)]++;
7039 /* Go from the last insn to the first and delete insns that only set unused
7040 registers or copy a register to itself. As we delete an insn, remove
7041 usage counts for registers it uses.
7043 The first jump optimization pass may leave a real insn as the last
7044 insn in the function. We must not skip that insn or we may end
7045 up deleting code that is not really dead.
7047 If some otherwise unused register is only used in DEBUG_INSNs,
7048 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7049 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7050 has been created for the unused register, replace it with
7051 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7052 auto_vec<tree, 32> later_debug_set_vars;
7053 for (insn = get_last_insn (); insn; insn = prev)
7055 int live_insn = 0;
7057 prev = PREV_INSN (insn);
7058 if (!INSN_P (insn))
7059 continue;
7061 live_insn = insn_live_p (insn, counts);
7063 /* If this is a dead insn, delete it and show registers in it aren't
7064 being used. */
7066 if (! live_insn && dbg_cnt (delete_trivial_dead))
7068 if (DEBUG_INSN_P (insn))
7070 if (DEBUG_BIND_INSN_P (insn))
7071 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7072 NULL_RTX, -1);
7074 else
7076 rtx set;
7077 if (MAY_HAVE_DEBUG_BIND_INSNS
7078 && (set = single_set (insn)) != NULL_RTX
7079 && is_dead_reg (SET_DEST (set), counts)
7080 /* Used at least once in some DEBUG_INSN. */
7081 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7082 /* And set exactly once. */
7083 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7084 && !side_effects_p (SET_SRC (set))
7085 && asm_noperands (PATTERN (insn)) < 0)
7087 rtx dval, bind_var_loc;
7088 rtx_insn *bind;
7090 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7091 dval = make_debug_expr_from_rtl (SET_DEST (set));
7093 /* Emit a debug bind insn before the insn in which
7094 reg dies. */
7095 bind_var_loc =
7096 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7097 DEBUG_EXPR_TREE_DECL (dval),
7098 SET_SRC (set),
7099 VAR_INIT_STATUS_INITIALIZED);
7100 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7102 bind = emit_debug_insn_before (bind_var_loc, insn);
7103 df_insn_rescan (bind);
7105 if (replacements == NULL)
7106 replacements = XCNEWVEC (rtx, nreg);
7107 replacements[REGNO (SET_DEST (set))] = dval;
7110 count_reg_usage (insn, counts, NULL_RTX, -1);
7111 ndead++;
7113 cse_cfg_altered |= delete_insn_and_edges (insn);
7115 else
7117 if (!DEBUG_INSN_P (insn) || DEBUG_MARKER_INSN_P (insn))
7119 for (tree var : later_debug_set_vars)
7120 TREE_VISITED (var) = 0;
7121 later_debug_set_vars.truncate (0);
7123 else if (DEBUG_BIND_INSN_P (insn)
7124 && !TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)))
7126 later_debug_set_vars.safe_push (INSN_VAR_LOCATION_DECL (insn));
7127 TREE_VISITED (INSN_VAR_LOCATION_DECL (insn)) = 1;
7132 if (MAY_HAVE_DEBUG_BIND_INSNS)
7134 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7135 if (DEBUG_BIND_INSN_P (insn))
7137 /* If this debug insn references a dead register that wasn't replaced
7138 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7139 bool seen_repl = false;
7140 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7141 counts, replacements, &seen_repl))
7143 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7144 df_insn_rescan (insn);
7146 else if (seen_repl)
7148 INSN_VAR_LOCATION_LOC (insn)
7149 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7150 NULL_RTX, replace_dead_reg,
7151 replacements);
7152 df_insn_rescan (insn);
7155 free (replacements);
7158 if (dump_file && ndead)
7159 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7160 ndead);
7161 /* Clean up. */
7162 free (counts);
7163 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7164 return ndead;
7167 /* If LOC contains references to NEWREG in a different mode, change them
7168 to use NEWREG instead. */
7170 static void
7171 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7172 rtx *loc, rtx_insn *insn, rtx newreg)
7174 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7176 rtx *loc = *iter;
7177 rtx x = *loc;
7178 if (x
7179 && REG_P (x)
7180 && REGNO (x) == REGNO (newreg)
7181 && GET_MODE (x) != GET_MODE (newreg))
7183 validate_change (insn, loc, newreg, 1);
7184 iter.skip_subrtxes ();
7189 /* Change the mode of any reference to the register REGNO (NEWREG) to
7190 GET_MODE (NEWREG) in INSN. */
7192 static void
7193 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7195 int success;
7197 if (!INSN_P (insn))
7198 return;
7200 subrtx_ptr_iterator::array_type array;
7201 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7202 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7204 /* If the following assertion was triggered, there is most probably
7205 something wrong with the cc_modes_compatible back end function.
7206 CC modes only can be considered compatible if the insn - with the mode
7207 replaced by any of the compatible modes - can still be recognized. */
7208 success = apply_change_group ();
7209 gcc_assert (success);
7212 /* Change the mode of any reference to the register REGNO (NEWREG) to
7213 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7214 any instruction which modifies NEWREG. */
7216 static void
7217 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7219 rtx_insn *insn;
7221 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7223 if (! INSN_P (insn))
7224 continue;
7226 if (reg_set_p (newreg, insn))
7227 return;
7229 cse_change_cc_mode_insn (insn, newreg);
7233 /* BB is a basic block which finishes with CC_REG as a condition code
7234 register which is set to CC_SRC. Look through the successors of BB
7235 to find blocks which have a single predecessor (i.e., this one),
7236 and look through those blocks for an assignment to CC_REG which is
7237 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7238 permitted to change the mode of CC_SRC to a compatible mode. This
7239 returns VOIDmode if no equivalent assignments were found.
7240 Otherwise it returns the mode which CC_SRC should wind up with.
7241 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7242 but is passed unmodified down to recursive calls in order to prevent
7243 endless recursion.
7245 The main complexity in this function is handling the mode issues.
7246 We may have more than one duplicate which we can eliminate, and we
7247 try to find a mode which will work for multiple duplicates. */
7249 static machine_mode
7250 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7251 bool can_change_mode)
7253 bool found_equiv;
7254 machine_mode mode;
7255 unsigned int insn_count;
7256 edge e;
7257 rtx_insn *insns[2];
7258 machine_mode modes[2];
7259 rtx_insn *last_insns[2];
7260 unsigned int i;
7261 rtx newreg;
7262 edge_iterator ei;
7264 /* We expect to have two successors. Look at both before picking
7265 the final mode for the comparison. If we have more successors
7266 (i.e., some sort of table jump, although that seems unlikely),
7267 then we require all beyond the first two to use the same
7268 mode. */
7270 found_equiv = false;
7271 mode = GET_MODE (cc_src);
7272 insn_count = 0;
7273 FOR_EACH_EDGE (e, ei, bb->succs)
7275 rtx_insn *insn;
7276 rtx_insn *end;
7278 if (e->flags & EDGE_COMPLEX)
7279 continue;
7281 if (EDGE_COUNT (e->dest->preds) != 1
7282 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7283 /* Avoid endless recursion on unreachable blocks. */
7284 || e->dest == orig_bb)
7285 continue;
7287 end = NEXT_INSN (BB_END (e->dest));
7288 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7290 rtx set;
7292 if (! INSN_P (insn))
7293 continue;
7295 /* If CC_SRC is modified, we have to stop looking for
7296 something which uses it. */
7297 if (modified_in_p (cc_src, insn))
7298 break;
7300 /* Check whether INSN sets CC_REG to CC_SRC. */
7301 set = single_set (insn);
7302 if (set
7303 && REG_P (SET_DEST (set))
7304 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7306 bool found;
7307 machine_mode set_mode;
7308 machine_mode comp_mode;
7310 found = false;
7311 set_mode = GET_MODE (SET_SRC (set));
7312 comp_mode = set_mode;
7313 if (rtx_equal_p (cc_src, SET_SRC (set)))
7314 found = true;
7315 else if (GET_CODE (cc_src) == COMPARE
7316 && GET_CODE (SET_SRC (set)) == COMPARE
7317 && mode != set_mode
7318 && rtx_equal_p (XEXP (cc_src, 0),
7319 XEXP (SET_SRC (set), 0))
7320 && rtx_equal_p (XEXP (cc_src, 1),
7321 XEXP (SET_SRC (set), 1)))
7324 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7325 if (comp_mode != VOIDmode
7326 && (can_change_mode || comp_mode == mode))
7327 found = true;
7330 if (found)
7332 found_equiv = true;
7333 if (insn_count < ARRAY_SIZE (insns))
7335 insns[insn_count] = insn;
7336 modes[insn_count] = set_mode;
7337 last_insns[insn_count] = end;
7338 ++insn_count;
7340 if (mode != comp_mode)
7342 gcc_assert (can_change_mode);
7343 mode = comp_mode;
7345 /* The modified insn will be re-recognized later. */
7346 PUT_MODE (cc_src, mode);
7349 else
7351 if (set_mode != mode)
7353 /* We found a matching expression in the
7354 wrong mode, but we don't have room to
7355 store it in the array. Punt. This case
7356 should be rare. */
7357 break;
7359 /* INSN sets CC_REG to a value equal to CC_SRC
7360 with the right mode. We can simply delete
7361 it. */
7362 delete_insn (insn);
7365 /* We found an instruction to delete. Keep looking,
7366 in the hopes of finding a three-way jump. */
7367 continue;
7370 /* We found an instruction which sets the condition
7371 code, so don't look any farther. */
7372 break;
7375 /* If INSN sets CC_REG in some other way, don't look any
7376 farther. */
7377 if (reg_set_p (cc_reg, insn))
7378 break;
7381 /* If we fell off the bottom of the block, we can keep looking
7382 through successors. We pass CAN_CHANGE_MODE as false because
7383 we aren't prepared to handle compatibility between the
7384 further blocks and this block. */
7385 if (insn == end)
7387 machine_mode submode;
7389 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7390 if (submode != VOIDmode)
7392 gcc_assert (submode == mode);
7393 found_equiv = true;
7394 can_change_mode = false;
7399 if (! found_equiv)
7400 return VOIDmode;
7402 /* Now INSN_COUNT is the number of instructions we found which set
7403 CC_REG to a value equivalent to CC_SRC. The instructions are in
7404 INSNS. The modes used by those instructions are in MODES. */
7406 newreg = NULL_RTX;
7407 for (i = 0; i < insn_count; ++i)
7409 if (modes[i] != mode)
7411 /* We need to change the mode of CC_REG in INSNS[i] and
7412 subsequent instructions. */
7413 if (! newreg)
7415 if (GET_MODE (cc_reg) == mode)
7416 newreg = cc_reg;
7417 else
7418 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7420 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7421 newreg);
7424 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7427 return mode;
7430 /* If we have a fixed condition code register (or two), walk through
7431 the instructions and try to eliminate duplicate assignments. */
7433 static void
7434 cse_condition_code_reg (void)
7436 unsigned int cc_regno_1;
7437 unsigned int cc_regno_2;
7438 rtx cc_reg_1;
7439 rtx cc_reg_2;
7440 basic_block bb;
7442 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7443 return;
7445 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7446 if (cc_regno_2 != INVALID_REGNUM)
7447 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7448 else
7449 cc_reg_2 = NULL_RTX;
7451 FOR_EACH_BB_FN (bb, cfun)
7453 rtx_insn *last_insn;
7454 rtx cc_reg;
7455 rtx_insn *insn;
7456 rtx_insn *cc_src_insn;
7457 rtx cc_src;
7458 machine_mode mode;
7459 machine_mode orig_mode;
7461 /* Look for blocks which end with a conditional jump based on a
7462 condition code register. Then look for the instruction which
7463 sets the condition code register. Then look through the
7464 successor blocks for instructions which set the condition
7465 code register to the same value. There are other possible
7466 uses of the condition code register, but these are by far the
7467 most common and the ones which we are most likely to be able
7468 to optimize. */
7470 last_insn = BB_END (bb);
7471 if (!JUMP_P (last_insn))
7472 continue;
7474 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7475 cc_reg = cc_reg_1;
7476 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7477 cc_reg = cc_reg_2;
7478 else
7479 continue;
7481 cc_src_insn = NULL;
7482 cc_src = NULL_RTX;
7483 for (insn = PREV_INSN (last_insn);
7484 insn && insn != PREV_INSN (BB_HEAD (bb));
7485 insn = PREV_INSN (insn))
7487 rtx set;
7489 if (! INSN_P (insn))
7490 continue;
7491 set = single_set (insn);
7492 if (set
7493 && REG_P (SET_DEST (set))
7494 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7496 cc_src_insn = insn;
7497 cc_src = SET_SRC (set);
7498 break;
7500 else if (reg_set_p (cc_reg, insn))
7501 break;
7504 if (! cc_src_insn)
7505 continue;
7507 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7508 continue;
7510 /* Now CC_REG is a condition code register used for a
7511 conditional jump at the end of the block, and CC_SRC, in
7512 CC_SRC_INSN, is the value to which that condition code
7513 register is set, and CC_SRC is still meaningful at the end of
7514 the basic block. */
7516 orig_mode = GET_MODE (cc_src);
7517 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7518 if (mode != VOIDmode)
7520 gcc_assert (mode == GET_MODE (cc_src));
7521 if (mode != orig_mode)
7523 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7525 cse_change_cc_mode_insn (cc_src_insn, newreg);
7527 /* Do the same in the following insns that use the
7528 current value of CC_REG within BB. */
7529 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7530 NEXT_INSN (last_insn),
7531 newreg);
7538 /* Perform common subexpression elimination. Nonzero value from
7539 `cse_main' means that jumps were simplified and some code may now
7540 be unreachable, so do jump optimization again. */
7541 static unsigned int
7542 rest_of_handle_cse (void)
7544 int tem;
7546 if (dump_file)
7547 dump_flow_info (dump_file, dump_flags);
7549 tem = cse_main (get_insns (), max_reg_num ());
7551 /* If we are not running more CSE passes, then we are no longer
7552 expecting CSE to be run. But always rerun it in a cheap mode. */
7553 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7555 if (tem == 2)
7557 timevar_push (TV_JUMP);
7558 rebuild_jump_labels (get_insns ());
7559 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7560 timevar_pop (TV_JUMP);
7562 else if (tem == 1 || optimize > 1)
7563 cse_cfg_altered |= cleanup_cfg (0);
7565 return 0;
7568 namespace {
7570 const pass_data pass_data_cse =
7572 RTL_PASS, /* type */
7573 "cse1", /* name */
7574 OPTGROUP_NONE, /* optinfo_flags */
7575 TV_CSE, /* tv_id */
7576 0, /* properties_required */
7577 0, /* properties_provided */
7578 0, /* properties_destroyed */
7579 0, /* todo_flags_start */
7580 TODO_df_finish, /* todo_flags_finish */
7583 class pass_cse : public rtl_opt_pass
7585 public:
7586 pass_cse (gcc::context *ctxt)
7587 : rtl_opt_pass (pass_data_cse, ctxt)
7590 /* opt_pass methods: */
7591 bool gate (function *) final override { return optimize > 0; }
7592 unsigned int execute (function *) final override
7594 return rest_of_handle_cse ();
7597 }; // class pass_cse
7599 } // anon namespace
7601 rtl_opt_pass *
7602 make_pass_cse (gcc::context *ctxt)
7604 return new pass_cse (ctxt);
7608 /* Run second CSE pass after loop optimizations. */
7609 static unsigned int
7610 rest_of_handle_cse2 (void)
7612 int tem;
7614 if (dump_file)
7615 dump_flow_info (dump_file, dump_flags);
7617 tem = cse_main (get_insns (), max_reg_num ());
7619 /* Run a pass to eliminate duplicated assignments to condition code
7620 registers. We have to run this after bypass_jumps, because it
7621 makes it harder for that pass to determine whether a jump can be
7622 bypassed safely. */
7623 cse_condition_code_reg ();
7625 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7627 if (tem == 2)
7629 timevar_push (TV_JUMP);
7630 rebuild_jump_labels (get_insns ());
7631 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7632 timevar_pop (TV_JUMP);
7634 else if (tem == 1 || cse_cfg_altered)
7635 cse_cfg_altered |= cleanup_cfg (0);
7637 cse_not_expected = 1;
7638 return 0;
7642 namespace {
7644 const pass_data pass_data_cse2 =
7646 RTL_PASS, /* type */
7647 "cse2", /* name */
7648 OPTGROUP_NONE, /* optinfo_flags */
7649 TV_CSE2, /* tv_id */
7650 0, /* properties_required */
7651 0, /* properties_provided */
7652 0, /* properties_destroyed */
7653 0, /* todo_flags_start */
7654 TODO_df_finish, /* todo_flags_finish */
7657 class pass_cse2 : public rtl_opt_pass
7659 public:
7660 pass_cse2 (gcc::context *ctxt)
7661 : rtl_opt_pass (pass_data_cse2, ctxt)
7664 /* opt_pass methods: */
7665 bool gate (function *) final override
7667 return optimize > 0 && flag_rerun_cse_after_loop;
7670 unsigned int execute (function *) final override
7672 return rest_of_handle_cse2 ();
7675 }; // class pass_cse2
7677 } // anon namespace
7679 rtl_opt_pass *
7680 make_pass_cse2 (gcc::context *ctxt)
7682 return new pass_cse2 (ctxt);
7685 /* Run second CSE pass after loop optimizations. */
7686 static unsigned int
7687 rest_of_handle_cse_after_global_opts (void)
7689 int save_cfj;
7690 int tem;
7692 /* We only want to do local CSE, so don't follow jumps. */
7693 save_cfj = flag_cse_follow_jumps;
7694 flag_cse_follow_jumps = 0;
7696 rebuild_jump_labels (get_insns ());
7697 tem = cse_main (get_insns (), max_reg_num ());
7698 cse_cfg_altered |= purge_all_dead_edges ();
7699 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7701 cse_not_expected = !flag_rerun_cse_after_loop;
7703 /* If cse altered any jumps, rerun jump opts to clean things up. */
7704 if (tem == 2)
7706 timevar_push (TV_JUMP);
7707 rebuild_jump_labels (get_insns ());
7708 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7709 timevar_pop (TV_JUMP);
7711 else if (tem == 1 || cse_cfg_altered)
7712 cse_cfg_altered |= cleanup_cfg (0);
7714 flag_cse_follow_jumps = save_cfj;
7715 return 0;
7718 namespace {
7720 const pass_data pass_data_cse_after_global_opts =
7722 RTL_PASS, /* type */
7723 "cse_local", /* name */
7724 OPTGROUP_NONE, /* optinfo_flags */
7725 TV_CSE, /* tv_id */
7726 0, /* properties_required */
7727 0, /* properties_provided */
7728 0, /* properties_destroyed */
7729 0, /* todo_flags_start */
7730 TODO_df_finish, /* todo_flags_finish */
7733 class pass_cse_after_global_opts : public rtl_opt_pass
7735 public:
7736 pass_cse_after_global_opts (gcc::context *ctxt)
7737 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7740 /* opt_pass methods: */
7741 bool gate (function *) final override
7743 return optimize > 0 && flag_rerun_cse_after_global_opts;
7746 unsigned int execute (function *) final override
7748 return rest_of_handle_cse_after_global_opts ();
7751 }; // class pass_cse_after_global_opts
7753 } // anon namespace
7755 rtl_opt_pass *
7756 make_pass_cse_after_global_opts (gcc::context *ctxt)
7758 return new pass_cse_after_global_opts (ctxt);