re PR testsuite/60672 (FAIL: g++.dg/cpp1y/auto-fn25.C -std=gnu++1y (test for errors...
[official-gcc.git] / gcc / cse.c
blob852d13ebf5bcdc13a1a7ad6fcb2759abba371c0c
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2014 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 "tm.h"
24 #include "rtl.h"
25 #include "tm_p.h"
26 #include "hard-reg-set.h"
27 #include "regs.h"
28 #include "basic-block.h"
29 #include "flags.h"
30 #include "insn-config.h"
31 #include "recog.h"
32 #include "function.h"
33 #include "expr.h"
34 #include "diagnostic-core.h"
35 #include "toplev.h"
36 #include "ggc.h"
37 #include "except.h"
38 #include "target.h"
39 #include "params.h"
40 #include "rtlhooks-def.h"
41 #include "tree-pass.h"
42 #include "df.h"
43 #include "dbgcnt.h"
44 #include "pointer-set.h"
46 /* The basic idea of common subexpression elimination is to go
47 through the code, keeping a record of expressions that would
48 have the same value at the current scan point, and replacing
49 expressions encountered with the cheapest equivalent expression.
51 It is too complicated to keep track of the different possibilities
52 when control paths merge in this code; so, at each label, we forget all
53 that is known and start fresh. This can be described as processing each
54 extended basic block separately. We have a separate pass to perform
55 global CSE.
57 Note CSE can turn a conditional or computed jump into a nop or
58 an unconditional jump. When this occurs we arrange to run the jump
59 optimizer after CSE to delete the unreachable code.
61 We use two data structures to record the equivalent expressions:
62 a hash table for most expressions, and a vector of "quantity
63 numbers" to record equivalent (pseudo) registers.
65 The use of the special data structure for registers is desirable
66 because it is faster. It is possible because registers references
67 contain a fairly small number, the register number, taken from
68 a contiguously allocated series, and two register references are
69 identical if they have the same number. General expressions
70 do not have any such thing, so the only way to retrieve the
71 information recorded on an expression other than a register
72 is to keep it in a hash table.
74 Registers and "quantity numbers":
76 At the start of each basic block, all of the (hardware and pseudo)
77 registers used in the function are given distinct quantity
78 numbers to indicate their contents. During scan, when the code
79 copies one register into another, we copy the quantity number.
80 When a register is loaded in any other way, we allocate a new
81 quantity number to describe the value generated by this operation.
82 `REG_QTY (N)' records what quantity register N is currently thought
83 of as containing.
85 All real quantity numbers are greater than or equal to zero.
86 If register N has not been assigned a quantity, `REG_QTY (N)' will
87 equal -N - 1, which is always negative.
89 Quantity numbers below zero do not exist and none of the `qty_table'
90 entries should be referenced with a negative index.
92 We also maintain a bidirectional chain of registers for each
93 quantity number. The `qty_table` members `first_reg' and `last_reg',
94 and `reg_eqv_table' members `next' and `prev' hold these chains.
96 The first register in a chain is the one whose lifespan is least local.
97 Among equals, it is the one that was seen first.
98 We replace any equivalent register with that one.
100 If two registers have the same quantity number, it must be true that
101 REG expressions with qty_table `mode' must be in the hash table for both
102 registers and must be in the same class.
104 The converse is not true. Since hard registers may be referenced in
105 any mode, two REG expressions might be equivalent in the hash table
106 but not have the same quantity number if the quantity number of one
107 of the registers is not the same mode as those expressions.
109 Constants and quantity numbers
111 When a quantity has a known constant value, that value is stored
112 in the appropriate qty_table `const_rtx'. This is in addition to
113 putting the constant in the hash table as is usual for non-regs.
115 Whether a reg or a constant is preferred is determined by the configuration
116 macro CONST_COSTS and will often depend on the constant value. In any
117 event, expressions containing constants can be simplified, by fold_rtx.
119 When a quantity has a known nearly constant value (such as an address
120 of a stack slot), that value is stored in the appropriate qty_table
121 `const_rtx'.
123 Integer constants don't have a machine mode. However, cse
124 determines the intended machine mode from the destination
125 of the instruction that moves the constant. The machine mode
126 is recorded in the hash table along with the actual RTL
127 constant expression so that different modes are kept separate.
129 Other expressions:
131 To record known equivalences among expressions in general
132 we use a hash table called `table'. It has a fixed number of buckets
133 that contain chains of `struct table_elt' elements for expressions.
134 These chains connect the elements whose expressions have the same
135 hash codes.
137 Other chains through the same elements connect the elements which
138 currently have equivalent values.
140 Register references in an expression are canonicalized before hashing
141 the expression. This is done using `reg_qty' and qty_table `first_reg'.
142 The hash code of a register reference is computed using the quantity
143 number, not the register number.
145 When the value of an expression changes, it is necessary to remove from the
146 hash table not just that expression but all expressions whose values
147 could be different as a result.
149 1. If the value changing is in memory, except in special cases
150 ANYTHING referring to memory could be changed. That is because
151 nobody knows where a pointer does not point.
152 The function `invalidate_memory' removes what is necessary.
154 The special cases are when the address is constant or is
155 a constant plus a fixed register such as the frame pointer
156 or a static chain pointer. When such addresses are stored in,
157 we can tell exactly which other such addresses must be invalidated
158 due to overlap. `invalidate' does this.
159 All expressions that refer to non-constant
160 memory addresses are also invalidated. `invalidate_memory' does this.
162 2. If the value changing is a register, all expressions
163 containing references to that register, and only those,
164 must be removed.
166 Because searching the entire hash table for expressions that contain
167 a register is very slow, we try to figure out when it isn't necessary.
168 Precisely, this is necessary only when expressions have been
169 entered in the hash table using this register, and then the value has
170 changed, and then another expression wants to be added to refer to
171 the register's new value. This sequence of circumstances is rare
172 within any one basic block.
174 `REG_TICK' and `REG_IN_TABLE', accessors for members of
175 cse_reg_info, are used to detect this case. REG_TICK (i) is
176 incremented whenever a value is stored in register i.
177 REG_IN_TABLE (i) holds -1 if no references to register i have been
178 entered in the table; otherwise, it contains the value REG_TICK (i)
179 had when the references were entered. If we want to enter a
180 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
181 remove old references. Until we want to enter a new entry, the
182 mere fact that the two vectors don't match makes the entries be
183 ignored if anyone tries to match them.
185 Registers themselves are entered in the hash table as well as in
186 the equivalent-register chains. However, `REG_TICK' and
187 `REG_IN_TABLE' do not apply to expressions which are simple
188 register references. These expressions are removed from the table
189 immediately when they become invalid, and this can be done even if
190 we do not immediately search for all the expressions that refer to
191 the register.
193 A CLOBBER rtx in an instruction invalidates its operand for further
194 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
195 invalidates everything that resides in memory.
197 Related expressions:
199 Constant expressions that differ only by an additive integer
200 are called related. When a constant expression is put in
201 the table, the related expression with no constant term
202 is also entered. These are made to point at each other
203 so that it is possible to find out if there exists any
204 register equivalent to an expression related to a given expression. */
206 /* Length of qty_table vector. We know in advance we will not need
207 a quantity number this big. */
209 static int max_qty;
211 /* Next quantity number to be allocated.
212 This is 1 + the largest number needed so far. */
214 static int next_qty;
216 /* Per-qty information tracking.
218 `first_reg' and `last_reg' track the head and tail of the
219 chain of registers which currently contain this quantity.
221 `mode' contains the machine mode of this quantity.
223 `const_rtx' holds the rtx of the constant value of this
224 quantity, if known. A summations of the frame/arg pointer
225 and a constant can also be entered here. When this holds
226 a known value, `const_insn' is the insn which stored the
227 constant value.
229 `comparison_{code,const,qty}' are used to track when a
230 comparison between a quantity and some constant or register has
231 been passed. In such a case, we know the results of the comparison
232 in case we see it again. These members record a comparison that
233 is known to be true. `comparison_code' holds the rtx code of such
234 a comparison, else it is set to UNKNOWN and the other two
235 comparison members are undefined. `comparison_const' holds
236 the constant being compared against, or zero if the comparison
237 is not against a constant. `comparison_qty' holds the quantity
238 being compared against when the result is known. If the comparison
239 is not with a register, `comparison_qty' is -1. */
241 struct qty_table_elem
243 rtx const_rtx;
244 rtx const_insn;
245 rtx comparison_const;
246 int comparison_qty;
247 unsigned int first_reg, last_reg;
248 /* The sizes of these fields should match the sizes of the
249 code and mode fields of struct rtx_def (see rtl.h). */
250 ENUM_BITFIELD(rtx_code) comparison_code : 16;
251 ENUM_BITFIELD(machine_mode) mode : 8;
254 /* The table of all qtys, indexed by qty number. */
255 static struct qty_table_elem *qty_table;
257 /* Structure used to pass arguments via for_each_rtx to function
258 cse_change_cc_mode. */
259 struct change_cc_mode_args
261 rtx insn;
262 rtx newreg;
265 #ifdef HAVE_cc0
266 /* For machines that have a CC0, we do not record its value in the hash
267 table since its use is guaranteed to be the insn immediately following
268 its definition and any other insn is presumed to invalidate it.
270 Instead, we store below the current and last value assigned to CC0.
271 If it should happen to be a constant, it is stored in preference
272 to the actual assigned value. In case it is a constant, we store
273 the mode in which the constant should be interpreted. */
275 static rtx this_insn_cc0, prev_insn_cc0;
276 static enum machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
277 #endif
279 /* Insn being scanned. */
281 static rtx this_insn;
282 static bool optimize_this_for_speed_p;
284 /* Index by register number, gives the number of the next (or
285 previous) register in the chain of registers sharing the same
286 value.
288 Or -1 if this register is at the end of the chain.
290 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
292 /* Per-register equivalence chain. */
293 struct reg_eqv_elem
295 int next, prev;
298 /* The table of all register equivalence chains. */
299 static struct reg_eqv_elem *reg_eqv_table;
301 struct cse_reg_info
303 /* The timestamp at which this register is initialized. */
304 unsigned int timestamp;
306 /* The quantity number of the register's current contents. */
307 int reg_qty;
309 /* The number of times the register has been altered in the current
310 basic block. */
311 int reg_tick;
313 /* The REG_TICK value at which rtx's containing this register are
314 valid in the hash table. If this does not equal the current
315 reg_tick value, such expressions existing in the hash table are
316 invalid. */
317 int reg_in_table;
319 /* The SUBREG that was set when REG_TICK was last incremented. Set
320 to -1 if the last store was to the whole register, not a subreg. */
321 unsigned int subreg_ticked;
324 /* A table of cse_reg_info indexed by register numbers. */
325 static struct cse_reg_info *cse_reg_info_table;
327 /* The size of the above table. */
328 static unsigned int cse_reg_info_table_size;
330 /* The index of the first entry that has not been initialized. */
331 static unsigned int cse_reg_info_table_first_uninitialized;
333 /* The timestamp at the beginning of the current run of
334 cse_extended_basic_block. We increment this variable at the beginning of
335 the current run of cse_extended_basic_block. The timestamp field of a
336 cse_reg_info entry matches the value of this variable if and only
337 if the entry has been initialized during the current run of
338 cse_extended_basic_block. */
339 static unsigned int cse_reg_info_timestamp;
341 /* A HARD_REG_SET containing all the hard registers for which there is
342 currently a REG expression in the hash table. Note the difference
343 from the above variables, which indicate if the REG is mentioned in some
344 expression in the table. */
346 static HARD_REG_SET hard_regs_in_table;
348 /* True if CSE has altered the CFG. */
349 static bool cse_cfg_altered;
351 /* True if CSE has altered conditional jump insns in such a way
352 that jump optimization should be redone. */
353 static bool cse_jumps_altered;
355 /* True if we put a LABEL_REF into the hash table for an INSN
356 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
357 to put in the note. */
358 static bool recorded_label_ref;
360 /* canon_hash stores 1 in do_not_record
361 if it notices a reference to CC0, PC, or some other volatile
362 subexpression. */
364 static int do_not_record;
366 /* canon_hash stores 1 in hash_arg_in_memory
367 if it notices a reference to memory within the expression being hashed. */
369 static int hash_arg_in_memory;
371 /* The hash table contains buckets which are chains of `struct table_elt's,
372 each recording one expression's information.
373 That expression is in the `exp' field.
375 The canon_exp field contains a canonical (from the point of view of
376 alias analysis) version of the `exp' field.
378 Those elements with the same hash code are chained in both directions
379 through the `next_same_hash' and `prev_same_hash' fields.
381 Each set of expressions with equivalent values
382 are on a two-way chain through the `next_same_value'
383 and `prev_same_value' fields, and all point with
384 the `first_same_value' field at the first element in
385 that chain. The chain is in order of increasing cost.
386 Each element's cost value is in its `cost' field.
388 The `in_memory' field is nonzero for elements that
389 involve any reference to memory. These elements are removed
390 whenever a write is done to an unidentified location in memory.
391 To be safe, we assume that a memory address is unidentified unless
392 the address is either a symbol constant or a constant plus
393 the frame pointer or argument pointer.
395 The `related_value' field is used to connect related expressions
396 (that differ by adding an integer).
397 The related expressions are chained in a circular fashion.
398 `related_value' is zero for expressions for which this
399 chain is not useful.
401 The `cost' field stores the cost of this element's expression.
402 The `regcost' field stores the value returned by approx_reg_cost for
403 this element's expression.
405 The `is_const' flag is set if the element is a constant (including
406 a fixed address).
408 The `flag' field is used as a temporary during some search routines.
410 The `mode' field is usually the same as GET_MODE (`exp'), but
411 if `exp' is a CONST_INT and has no machine mode then the `mode'
412 field is the mode it was being used as. Each constant is
413 recorded separately for each mode it is used with. */
415 struct table_elt
417 rtx exp;
418 rtx canon_exp;
419 struct table_elt *next_same_hash;
420 struct table_elt *prev_same_hash;
421 struct table_elt *next_same_value;
422 struct table_elt *prev_same_value;
423 struct table_elt *first_same_value;
424 struct table_elt *related_value;
425 int cost;
426 int regcost;
427 /* The size of this field should match the size
428 of the mode field of struct rtx_def (see rtl.h). */
429 ENUM_BITFIELD(machine_mode) mode : 8;
430 char in_memory;
431 char is_const;
432 char flag;
435 /* We don't want a lot of buckets, because we rarely have very many
436 things stored in the hash table, and a lot of buckets slows
437 down a lot of loops that happen frequently. */
438 #define HASH_SHIFT 5
439 #define HASH_SIZE (1 << HASH_SHIFT)
440 #define HASH_MASK (HASH_SIZE - 1)
442 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
443 register (hard registers may require `do_not_record' to be set). */
445 #define HASH(X, M) \
446 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
447 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
448 : canon_hash (X, M)) & HASH_MASK)
450 /* Like HASH, but without side-effects. */
451 #define SAFE_HASH(X, M) \
452 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
453 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
454 : safe_hash (X, M)) & HASH_MASK)
456 /* Determine whether register number N is considered a fixed register for the
457 purpose of approximating register costs.
458 It is desirable to replace other regs with fixed regs, to reduce need for
459 non-fixed hard regs.
460 A reg wins if it is either the frame pointer or designated as fixed. */
461 #define FIXED_REGNO_P(N) \
462 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
463 || fixed_regs[N] || global_regs[N])
465 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
466 hard registers and pointers into the frame are the cheapest with a cost
467 of 0. Next come pseudos with a cost of one and other hard registers with
468 a cost of 2. Aside from these special cases, call `rtx_cost'. */
470 #define CHEAP_REGNO(N) \
471 (REGNO_PTR_FRAME_P (N) \
472 || (HARD_REGISTER_NUM_P (N) \
473 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
475 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET, 1))
476 #define COST_IN(X, OUTER, OPNO) (REG_P (X) ? 0 : notreg_cost (X, OUTER, OPNO))
478 /* Get the number of times this register has been updated in this
479 basic block. */
481 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
483 /* Get the point at which REG was recorded in the table. */
485 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
487 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
488 SUBREG). */
490 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
492 /* Get the quantity number for REG. */
494 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
496 /* Determine if the quantity number for register X represents a valid index
497 into the qty_table. */
499 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
501 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
503 #define CHEAPER(X, Y) \
504 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
506 static struct table_elt *table[HASH_SIZE];
508 /* Chain of `struct table_elt's made so far for this function
509 but currently removed from the table. */
511 static struct table_elt *free_element_chain;
513 /* Set to the cost of a constant pool reference if one was found for a
514 symbolic constant. If this was found, it means we should try to
515 convert constants into constant pool entries if they don't fit in
516 the insn. */
518 static int constant_pool_entries_cost;
519 static int constant_pool_entries_regcost;
521 /* Trace a patch through the CFG. */
523 struct branch_path
525 /* The basic block for this path entry. */
526 basic_block bb;
529 /* This data describes a block that will be processed by
530 cse_extended_basic_block. */
532 struct cse_basic_block_data
534 /* Total number of SETs in block. */
535 int nsets;
536 /* Size of current branch path, if any. */
537 int path_size;
538 /* Current path, indicating which basic_blocks will be processed. */
539 struct branch_path *path;
543 /* Pointers to the live in/live out bitmaps for the boundaries of the
544 current EBB. */
545 static bitmap cse_ebb_live_in, cse_ebb_live_out;
547 /* A simple bitmap to track which basic blocks have been visited
548 already as part of an already processed extended basic block. */
549 static sbitmap cse_visited_basic_blocks;
551 static bool fixed_base_plus_p (rtx x);
552 static int notreg_cost (rtx, enum rtx_code, int);
553 static int approx_reg_cost_1 (rtx *, void *);
554 static int approx_reg_cost (rtx);
555 static int preferable (int, int, int, int);
556 static void new_basic_block (void);
557 static void make_new_qty (unsigned int, enum machine_mode);
558 static void make_regs_eqv (unsigned int, unsigned int);
559 static void delete_reg_equiv (unsigned int);
560 static int mention_regs (rtx);
561 static int insert_regs (rtx, struct table_elt *, int);
562 static void remove_from_table (struct table_elt *, unsigned);
563 static void remove_pseudo_from_table (rtx, unsigned);
564 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
565 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
566 static rtx lookup_as_function (rtx, enum rtx_code);
567 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
568 enum machine_mode, int, int);
569 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
570 enum machine_mode);
571 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
572 static void invalidate (rtx, enum machine_mode);
573 static void remove_invalid_refs (unsigned int);
574 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
575 enum machine_mode);
576 static void rehash_using_reg (rtx);
577 static void invalidate_memory (void);
578 static void invalidate_for_call (void);
579 static rtx use_related_value (rtx, struct table_elt *);
581 static inline unsigned canon_hash (rtx, enum machine_mode);
582 static inline unsigned safe_hash (rtx, enum machine_mode);
583 static inline unsigned hash_rtx_string (const char *);
585 static rtx canon_reg (rtx, rtx);
586 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
587 enum machine_mode *,
588 enum machine_mode *);
589 static rtx fold_rtx (rtx, rtx);
590 static rtx equiv_constant (rtx);
591 static void record_jump_equiv (rtx, bool);
592 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
593 int);
594 static void cse_insn (rtx);
595 static void cse_prescan_path (struct cse_basic_block_data *);
596 static void invalidate_from_clobbers (rtx);
597 static void invalidate_from_sets_and_clobbers (rtx);
598 static rtx cse_process_notes (rtx, rtx, bool *);
599 static void cse_extended_basic_block (struct cse_basic_block_data *);
600 static int check_for_label_ref (rtx *, void *);
601 extern void dump_class (struct table_elt*);
602 static void get_cse_reg_info_1 (unsigned int regno);
603 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
604 static int check_dependence (rtx *, void *);
606 static void flush_hash_table (void);
607 static bool insn_live_p (rtx, int *);
608 static bool set_live_p (rtx, rtx, int *);
609 static int cse_change_cc_mode (rtx *, void *);
610 static void cse_change_cc_mode_insn (rtx, rtx);
611 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
612 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
613 bool);
616 #undef RTL_HOOKS_GEN_LOWPART
617 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
619 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
621 /* Nonzero if X has the form (PLUS frame-pointer integer). */
623 static bool
624 fixed_base_plus_p (rtx x)
626 switch (GET_CODE (x))
628 case REG:
629 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
630 return true;
631 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
632 return true;
633 return false;
635 case PLUS:
636 if (!CONST_INT_P (XEXP (x, 1)))
637 return false;
638 return fixed_base_plus_p (XEXP (x, 0));
640 default:
641 return false;
645 /* Dump the expressions in the equivalence class indicated by CLASSP.
646 This function is used only for debugging. */
647 DEBUG_FUNCTION void
648 dump_class (struct table_elt *classp)
650 struct table_elt *elt;
652 fprintf (stderr, "Equivalence chain for ");
653 print_rtl (stderr, classp->exp);
654 fprintf (stderr, ": \n");
656 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
658 print_rtl (stderr, elt->exp);
659 fprintf (stderr, "\n");
663 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
665 static int
666 approx_reg_cost_1 (rtx *xp, void *data)
668 rtx x = *xp;
669 int *cost_p = (int *) data;
671 if (x && REG_P (x))
673 unsigned int regno = REGNO (x);
675 if (! CHEAP_REGNO (regno))
677 if (regno < FIRST_PSEUDO_REGISTER)
679 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
680 return 1;
681 *cost_p += 2;
683 else
684 *cost_p += 1;
688 return 0;
691 /* Return an estimate of the cost of the registers used in an rtx.
692 This is mostly the number of different REG expressions in the rtx;
693 however for some exceptions like fixed registers we use a cost of
694 0. If any other hard register reference occurs, return MAX_COST. */
696 static int
697 approx_reg_cost (rtx x)
699 int cost = 0;
701 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
702 return MAX_COST;
704 return cost;
707 /* Return a negative value if an rtx A, whose costs are given by COST_A
708 and REGCOST_A, is more desirable than an rtx B.
709 Return a positive value if A is less desirable, or 0 if the two are
710 equally good. */
711 static int
712 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
714 /* First, get rid of cases involving expressions that are entirely
715 unwanted. */
716 if (cost_a != cost_b)
718 if (cost_a == MAX_COST)
719 return 1;
720 if (cost_b == MAX_COST)
721 return -1;
724 /* Avoid extending lifetimes of hardregs. */
725 if (regcost_a != regcost_b)
727 if (regcost_a == MAX_COST)
728 return 1;
729 if (regcost_b == MAX_COST)
730 return -1;
733 /* Normal operation costs take precedence. */
734 if (cost_a != cost_b)
735 return cost_a - cost_b;
736 /* Only if these are identical consider effects on register pressure. */
737 if (regcost_a != regcost_b)
738 return regcost_a - regcost_b;
739 return 0;
742 /* Internal function, to compute cost when X is not a register; called
743 from COST macro to keep it simple. */
745 static int
746 notreg_cost (rtx x, enum rtx_code outer, int opno)
748 return ((GET_CODE (x) == SUBREG
749 && REG_P (SUBREG_REG (x))
750 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
751 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
752 && (GET_MODE_SIZE (GET_MODE (x))
753 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
754 && subreg_lowpart_p (x)
755 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (x),
756 GET_MODE (SUBREG_REG (x))))
758 : rtx_cost (x, outer, opno, optimize_this_for_speed_p) * 2);
762 /* Initialize CSE_REG_INFO_TABLE. */
764 static void
765 init_cse_reg_info (unsigned int nregs)
767 /* Do we need to grow the table? */
768 if (nregs > cse_reg_info_table_size)
770 unsigned int new_size;
772 if (cse_reg_info_table_size < 2048)
774 /* Compute a new size that is a power of 2 and no smaller
775 than the large of NREGS and 64. */
776 new_size = (cse_reg_info_table_size
777 ? cse_reg_info_table_size : 64);
779 while (new_size < nregs)
780 new_size *= 2;
782 else
784 /* If we need a big table, allocate just enough to hold
785 NREGS registers. */
786 new_size = nregs;
789 /* Reallocate the table with NEW_SIZE entries. */
790 free (cse_reg_info_table);
791 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
792 cse_reg_info_table_size = new_size;
793 cse_reg_info_table_first_uninitialized = 0;
796 /* Do we have all of the first NREGS entries initialized? */
797 if (cse_reg_info_table_first_uninitialized < nregs)
799 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
800 unsigned int i;
802 /* Put the old timestamp on newly allocated entries so that they
803 will all be considered out of date. We do not touch those
804 entries beyond the first NREGS entries to be nice to the
805 virtual memory. */
806 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
807 cse_reg_info_table[i].timestamp = old_timestamp;
809 cse_reg_info_table_first_uninitialized = nregs;
813 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
815 static void
816 get_cse_reg_info_1 (unsigned int regno)
818 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
819 entry will be considered to have been initialized. */
820 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
822 /* Initialize the rest of the entry. */
823 cse_reg_info_table[regno].reg_tick = 1;
824 cse_reg_info_table[regno].reg_in_table = -1;
825 cse_reg_info_table[regno].subreg_ticked = -1;
826 cse_reg_info_table[regno].reg_qty = -regno - 1;
829 /* Find a cse_reg_info entry for REGNO. */
831 static inline struct cse_reg_info *
832 get_cse_reg_info (unsigned int regno)
834 struct cse_reg_info *p = &cse_reg_info_table[regno];
836 /* If this entry has not been initialized, go ahead and initialize
837 it. */
838 if (p->timestamp != cse_reg_info_timestamp)
839 get_cse_reg_info_1 (regno);
841 return p;
844 /* Clear the hash table and initialize each register with its own quantity,
845 for a new basic block. */
847 static void
848 new_basic_block (void)
850 int i;
852 next_qty = 0;
854 /* Invalidate cse_reg_info_table. */
855 cse_reg_info_timestamp++;
857 /* Clear out hash table state for this pass. */
858 CLEAR_HARD_REG_SET (hard_regs_in_table);
860 /* The per-quantity values used to be initialized here, but it is
861 much faster to initialize each as it is made in `make_new_qty'. */
863 for (i = 0; i < HASH_SIZE; i++)
865 struct table_elt *first;
867 first = table[i];
868 if (first != NULL)
870 struct table_elt *last = first;
872 table[i] = NULL;
874 while (last->next_same_hash != NULL)
875 last = last->next_same_hash;
877 /* Now relink this hash entire chain into
878 the free element list. */
880 last->next_same_hash = free_element_chain;
881 free_element_chain = first;
885 #ifdef HAVE_cc0
886 prev_insn_cc0 = 0;
887 #endif
890 /* Say that register REG contains a quantity in mode MODE not in any
891 register before and initialize that quantity. */
893 static void
894 make_new_qty (unsigned int reg, enum machine_mode mode)
896 int q;
897 struct qty_table_elem *ent;
898 struct reg_eqv_elem *eqv;
900 gcc_assert (next_qty < max_qty);
902 q = REG_QTY (reg) = next_qty++;
903 ent = &qty_table[q];
904 ent->first_reg = reg;
905 ent->last_reg = reg;
906 ent->mode = mode;
907 ent->const_rtx = ent->const_insn = NULL_RTX;
908 ent->comparison_code = UNKNOWN;
910 eqv = &reg_eqv_table[reg];
911 eqv->next = eqv->prev = -1;
914 /* Make reg NEW equivalent to reg OLD.
915 OLD is not changing; NEW is. */
917 static void
918 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
920 unsigned int lastr, firstr;
921 int q = REG_QTY (old_reg);
922 struct qty_table_elem *ent;
924 ent = &qty_table[q];
926 /* Nothing should become eqv until it has a "non-invalid" qty number. */
927 gcc_assert (REGNO_QTY_VALID_P (old_reg));
929 REG_QTY (new_reg) = q;
930 firstr = ent->first_reg;
931 lastr = ent->last_reg;
933 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
934 hard regs. Among pseudos, if NEW will live longer than any other reg
935 of the same qty, and that is beyond the current basic block,
936 make it the new canonical replacement for this qty. */
937 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
938 /* Certain fixed registers might be of the class NO_REGS. This means
939 that not only can they not be allocated by the compiler, but
940 they cannot be used in substitutions or canonicalizations
941 either. */
942 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
943 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
944 || (new_reg >= FIRST_PSEUDO_REGISTER
945 && (firstr < FIRST_PSEUDO_REGISTER
946 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
947 && !bitmap_bit_p (cse_ebb_live_out, firstr))
948 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
949 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
951 reg_eqv_table[firstr].prev = new_reg;
952 reg_eqv_table[new_reg].next = firstr;
953 reg_eqv_table[new_reg].prev = -1;
954 ent->first_reg = new_reg;
956 else
958 /* If NEW is a hard reg (known to be non-fixed), insert at end.
959 Otherwise, insert before any non-fixed hard regs that are at the
960 end. Registers of class NO_REGS cannot be used as an
961 equivalent for anything. */
962 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
963 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
964 && new_reg >= FIRST_PSEUDO_REGISTER)
965 lastr = reg_eqv_table[lastr].prev;
966 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
967 if (reg_eqv_table[lastr].next >= 0)
968 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
969 else
970 qty_table[q].last_reg = new_reg;
971 reg_eqv_table[lastr].next = new_reg;
972 reg_eqv_table[new_reg].prev = lastr;
976 /* Remove REG from its equivalence class. */
978 static void
979 delete_reg_equiv (unsigned int reg)
981 struct qty_table_elem *ent;
982 int q = REG_QTY (reg);
983 int p, n;
985 /* If invalid, do nothing. */
986 if (! REGNO_QTY_VALID_P (reg))
987 return;
989 ent = &qty_table[q];
991 p = reg_eqv_table[reg].prev;
992 n = reg_eqv_table[reg].next;
994 if (n != -1)
995 reg_eqv_table[n].prev = p;
996 else
997 ent->last_reg = p;
998 if (p != -1)
999 reg_eqv_table[p].next = n;
1000 else
1001 ent->first_reg = n;
1003 REG_QTY (reg) = -reg - 1;
1006 /* Remove any invalid expressions from the hash table
1007 that refer to any of the registers contained in expression X.
1009 Make sure that newly inserted references to those registers
1010 as subexpressions will be considered valid.
1012 mention_regs is not called when a register itself
1013 is being stored in the table.
1015 Return 1 if we have done something that may have changed the hash code
1016 of X. */
1018 static int
1019 mention_regs (rtx x)
1021 enum rtx_code code;
1022 int i, j;
1023 const char *fmt;
1024 int changed = 0;
1026 if (x == 0)
1027 return 0;
1029 code = GET_CODE (x);
1030 if (code == REG)
1032 unsigned int regno = REGNO (x);
1033 unsigned int endregno = END_REGNO (x);
1034 unsigned int i;
1036 for (i = regno; i < endregno; i++)
1038 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1039 remove_invalid_refs (i);
1041 REG_IN_TABLE (i) = REG_TICK (i);
1042 SUBREG_TICKED (i) = -1;
1045 return 0;
1048 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1049 pseudo if they don't use overlapping words. We handle only pseudos
1050 here for simplicity. */
1051 if (code == SUBREG && REG_P (SUBREG_REG (x))
1052 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1054 unsigned int i = REGNO (SUBREG_REG (x));
1056 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1058 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1059 the last store to this register really stored into this
1060 subreg, then remove the memory of this subreg.
1061 Otherwise, remove any memory of the entire register and
1062 all its subregs from the table. */
1063 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1064 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1065 remove_invalid_refs (i);
1066 else
1067 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1070 REG_IN_TABLE (i) = REG_TICK (i);
1071 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1072 return 0;
1075 /* If X is a comparison or a COMPARE and either operand is a register
1076 that does not have a quantity, give it one. This is so that a later
1077 call to record_jump_equiv won't cause X to be assigned a different
1078 hash code and not found in the table after that call.
1080 It is not necessary to do this here, since rehash_using_reg can
1081 fix up the table later, but doing this here eliminates the need to
1082 call that expensive function in the most common case where the only
1083 use of the register is in the comparison. */
1085 if (code == COMPARE || COMPARISON_P (x))
1087 if (REG_P (XEXP (x, 0))
1088 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1089 if (insert_regs (XEXP (x, 0), NULL, 0))
1091 rehash_using_reg (XEXP (x, 0));
1092 changed = 1;
1095 if (REG_P (XEXP (x, 1))
1096 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1097 if (insert_regs (XEXP (x, 1), NULL, 0))
1099 rehash_using_reg (XEXP (x, 1));
1100 changed = 1;
1104 fmt = GET_RTX_FORMAT (code);
1105 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1106 if (fmt[i] == 'e')
1107 changed |= mention_regs (XEXP (x, i));
1108 else if (fmt[i] == 'E')
1109 for (j = 0; j < XVECLEN (x, i); j++)
1110 changed |= mention_regs (XVECEXP (x, i, j));
1112 return changed;
1115 /* Update the register quantities for inserting X into the hash table
1116 with a value equivalent to CLASSP.
1117 (If the class does not contain a REG, it is irrelevant.)
1118 If MODIFIED is nonzero, X is a destination; it is being modified.
1119 Note that delete_reg_equiv should be called on a register
1120 before insert_regs is done on that register with MODIFIED != 0.
1122 Nonzero value means that elements of reg_qty have changed
1123 so X's hash code may be different. */
1125 static int
1126 insert_regs (rtx x, struct table_elt *classp, int modified)
1128 if (REG_P (x))
1130 unsigned int regno = REGNO (x);
1131 int qty_valid;
1133 /* If REGNO is in the equivalence table already but is of the
1134 wrong mode for that equivalence, don't do anything here. */
1136 qty_valid = REGNO_QTY_VALID_P (regno);
1137 if (qty_valid)
1139 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1141 if (ent->mode != GET_MODE (x))
1142 return 0;
1145 if (modified || ! qty_valid)
1147 if (classp)
1148 for (classp = classp->first_same_value;
1149 classp != 0;
1150 classp = classp->next_same_value)
1151 if (REG_P (classp->exp)
1152 && GET_MODE (classp->exp) == GET_MODE (x))
1154 unsigned c_regno = REGNO (classp->exp);
1156 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1158 /* Suppose that 5 is hard reg and 100 and 101 are
1159 pseudos. Consider
1161 (set (reg:si 100) (reg:si 5))
1162 (set (reg:si 5) (reg:si 100))
1163 (set (reg:di 101) (reg:di 5))
1165 We would now set REG_QTY (101) = REG_QTY (5), but the
1166 entry for 5 is in SImode. When we use this later in
1167 copy propagation, we get the register in wrong mode. */
1168 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1169 continue;
1171 make_regs_eqv (regno, c_regno);
1172 return 1;
1175 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1176 than REG_IN_TABLE to find out if there was only a single preceding
1177 invalidation - for the SUBREG - or another one, which would be
1178 for the full register. However, if we find here that REG_TICK
1179 indicates that the register is invalid, it means that it has
1180 been invalidated in a separate operation. The SUBREG might be used
1181 now (then this is a recursive call), or we might use the full REG
1182 now and a SUBREG of it later. So bump up REG_TICK so that
1183 mention_regs will do the right thing. */
1184 if (! modified
1185 && REG_IN_TABLE (regno) >= 0
1186 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1187 REG_TICK (regno)++;
1188 make_new_qty (regno, GET_MODE (x));
1189 return 1;
1192 return 0;
1195 /* If X is a SUBREG, we will likely be inserting the inner register in the
1196 table. If that register doesn't have an assigned quantity number at
1197 this point but does later, the insertion that we will be doing now will
1198 not be accessible because its hash code will have changed. So assign
1199 a quantity number now. */
1201 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1202 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1204 insert_regs (SUBREG_REG (x), NULL, 0);
1205 mention_regs (x);
1206 return 1;
1208 else
1209 return mention_regs (x);
1213 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1214 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1215 CST is equal to an anchor. */
1217 static bool
1218 compute_const_anchors (rtx cst,
1219 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1220 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1222 HOST_WIDE_INT n = INTVAL (cst);
1224 *lower_base = n & ~(targetm.const_anchor - 1);
1225 if (*lower_base == n)
1226 return false;
1228 *upper_base =
1229 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1230 *upper_offs = n - *upper_base;
1231 *lower_offs = n - *lower_base;
1232 return true;
1235 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1237 static void
1238 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1239 enum machine_mode mode)
1241 struct table_elt *elt;
1242 unsigned hash;
1243 rtx anchor_exp;
1244 rtx exp;
1246 anchor_exp = GEN_INT (anchor);
1247 hash = HASH (anchor_exp, mode);
1248 elt = lookup (anchor_exp, hash, mode);
1249 if (!elt)
1250 elt = insert (anchor_exp, NULL, hash, mode);
1252 exp = plus_constant (mode, reg, offs);
1253 /* REG has just been inserted and the hash codes recomputed. */
1254 mention_regs (exp);
1255 hash = HASH (exp, mode);
1257 /* Use the cost of the register rather than the whole expression. When
1258 looking up constant anchors we will further offset the corresponding
1259 expression therefore it does not make sense to prefer REGs over
1260 reg-immediate additions. Prefer instead the oldest expression. Also
1261 don't prefer pseudos over hard regs so that we derive constants in
1262 argument registers from other argument registers rather than from the
1263 original pseudo that was used to synthesize the constant. */
1264 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1267 /* The constant CST is equivalent to the register REG. Create
1268 equivalences between the two anchors of CST and the corresponding
1269 register-offset expressions using REG. */
1271 static void
1272 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1274 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1276 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1277 &upper_base, &upper_offs))
1278 return;
1280 /* Ignore anchors of value 0. Constants accessible from zero are
1281 simple. */
1282 if (lower_base != 0)
1283 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1285 if (upper_base != 0)
1286 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1289 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1290 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1291 valid expression. Return the cheapest and oldest of such expressions. In
1292 *OLD, return how old the resulting expression is compared to the other
1293 equivalent expressions. */
1295 static rtx
1296 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1297 unsigned *old)
1299 struct table_elt *elt;
1300 unsigned idx;
1301 struct table_elt *match_elt;
1302 rtx match;
1304 /* Find the cheapest and *oldest* expression to maximize the chance of
1305 reusing the same pseudo. */
1307 match_elt = NULL;
1308 match = NULL_RTX;
1309 for (elt = anchor_elt->first_same_value, idx = 0;
1310 elt;
1311 elt = elt->next_same_value, idx++)
1313 if (match_elt && CHEAPER (match_elt, elt))
1314 return match;
1316 if (REG_P (elt->exp)
1317 || (GET_CODE (elt->exp) == PLUS
1318 && REG_P (XEXP (elt->exp, 0))
1319 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1321 rtx x;
1323 /* Ignore expressions that are no longer valid. */
1324 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1325 continue;
1327 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1328 if (REG_P (x)
1329 || (GET_CODE (x) == PLUS
1330 && IN_RANGE (INTVAL (XEXP (x, 1)),
1331 -targetm.const_anchor,
1332 targetm.const_anchor - 1)))
1334 match = x;
1335 match_elt = elt;
1336 *old = idx;
1341 return match;
1344 /* Try to express the constant SRC_CONST using a register+offset expression
1345 derived from a constant anchor. Return it if successful or NULL_RTX,
1346 otherwise. */
1348 static rtx
1349 try_const_anchors (rtx src_const, enum machine_mode mode)
1351 struct table_elt *lower_elt, *upper_elt;
1352 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1353 rtx lower_anchor_rtx, upper_anchor_rtx;
1354 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1355 unsigned lower_old, upper_old;
1357 /* CONST_INT is used for CC modes, but we should leave those alone. */
1358 if (GET_MODE_CLASS (mode) == MODE_CC)
1359 return NULL_RTX;
1361 gcc_assert (SCALAR_INT_MODE_P (mode));
1362 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1363 &upper_base, &upper_offs))
1364 return NULL_RTX;
1366 lower_anchor_rtx = GEN_INT (lower_base);
1367 upper_anchor_rtx = GEN_INT (upper_base);
1368 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1369 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1371 if (lower_elt)
1372 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1373 if (upper_elt)
1374 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1376 if (!lower_exp)
1377 return upper_exp;
1378 if (!upper_exp)
1379 return lower_exp;
1381 /* Return the older expression. */
1382 return (upper_old > lower_old ? upper_exp : lower_exp);
1385 /* Look in or update the hash table. */
1387 /* Remove table element ELT from use in the table.
1388 HASH is its hash code, made using the HASH macro.
1389 It's an argument because often that is known in advance
1390 and we save much time not recomputing it. */
1392 static void
1393 remove_from_table (struct table_elt *elt, unsigned int hash)
1395 if (elt == 0)
1396 return;
1398 /* Mark this element as removed. See cse_insn. */
1399 elt->first_same_value = 0;
1401 /* Remove the table element from its equivalence class. */
1404 struct table_elt *prev = elt->prev_same_value;
1405 struct table_elt *next = elt->next_same_value;
1407 if (next)
1408 next->prev_same_value = prev;
1410 if (prev)
1411 prev->next_same_value = next;
1412 else
1414 struct table_elt *newfirst = next;
1415 while (next)
1417 next->first_same_value = newfirst;
1418 next = next->next_same_value;
1423 /* Remove the table element from its hash bucket. */
1426 struct table_elt *prev = elt->prev_same_hash;
1427 struct table_elt *next = elt->next_same_hash;
1429 if (next)
1430 next->prev_same_hash = prev;
1432 if (prev)
1433 prev->next_same_hash = next;
1434 else if (table[hash] == elt)
1435 table[hash] = next;
1436 else
1438 /* This entry is not in the proper hash bucket. This can happen
1439 when two classes were merged by `merge_equiv_classes'. Search
1440 for the hash bucket that it heads. This happens only very
1441 rarely, so the cost is acceptable. */
1442 for (hash = 0; hash < HASH_SIZE; hash++)
1443 if (table[hash] == elt)
1444 table[hash] = next;
1448 /* Remove the table element from its related-value circular chain. */
1450 if (elt->related_value != 0 && elt->related_value != elt)
1452 struct table_elt *p = elt->related_value;
1454 while (p->related_value != elt)
1455 p = p->related_value;
1456 p->related_value = elt->related_value;
1457 if (p->related_value == p)
1458 p->related_value = 0;
1461 /* Now add it to the free element chain. */
1462 elt->next_same_hash = free_element_chain;
1463 free_element_chain = elt;
1466 /* Same as above, but X is a pseudo-register. */
1468 static void
1469 remove_pseudo_from_table (rtx x, unsigned int hash)
1471 struct table_elt *elt;
1473 /* Because a pseudo-register can be referenced in more than one
1474 mode, we might have to remove more than one table entry. */
1475 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1476 remove_from_table (elt, hash);
1479 /* Look up X in the hash table and return its table element,
1480 or 0 if X is not in the table.
1482 MODE is the machine-mode of X, or if X is an integer constant
1483 with VOIDmode then MODE is the mode with which X will be used.
1485 Here we are satisfied to find an expression whose tree structure
1486 looks like X. */
1488 static struct table_elt *
1489 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1491 struct table_elt *p;
1493 for (p = table[hash]; p; p = p->next_same_hash)
1494 if (mode == p->mode && ((x == p->exp && REG_P (x))
1495 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1496 return p;
1498 return 0;
1501 /* Like `lookup' but don't care whether the table element uses invalid regs.
1502 Also ignore discrepancies in the machine mode of a register. */
1504 static struct table_elt *
1505 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1507 struct table_elt *p;
1509 if (REG_P (x))
1511 unsigned int regno = REGNO (x);
1513 /* Don't check the machine mode when comparing registers;
1514 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1515 for (p = table[hash]; p; p = p->next_same_hash)
1516 if (REG_P (p->exp)
1517 && REGNO (p->exp) == regno)
1518 return p;
1520 else
1522 for (p = table[hash]; p; p = p->next_same_hash)
1523 if (mode == p->mode
1524 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1525 return p;
1528 return 0;
1531 /* Look for an expression equivalent to X and with code CODE.
1532 If one is found, return that expression. */
1534 static rtx
1535 lookup_as_function (rtx x, enum rtx_code code)
1537 struct table_elt *p
1538 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1540 if (p == 0)
1541 return 0;
1543 for (p = p->first_same_value; p; p = p->next_same_value)
1544 if (GET_CODE (p->exp) == code
1545 /* Make sure this is a valid entry in the table. */
1546 && exp_equiv_p (p->exp, p->exp, 1, false))
1547 return p->exp;
1549 return 0;
1552 /* Insert X in the hash table, assuming HASH is its hash code and
1553 CLASSP is an element of the class it should go in (or 0 if a new
1554 class should be made). COST is the code of X and reg_cost is the
1555 cost of registers in X. It is inserted at the proper position to
1556 keep the class in the order cheapest first.
1558 MODE is the machine-mode of X, or if X is an integer constant
1559 with VOIDmode then MODE is the mode with which X will be used.
1561 For elements of equal cheapness, the most recent one
1562 goes in front, except that the first element in the list
1563 remains first unless a cheaper element is added. The order of
1564 pseudo-registers does not matter, as canon_reg will be called to
1565 find the cheapest when a register is retrieved from the table.
1567 The in_memory field in the hash table element is set to 0.
1568 The caller must set it nonzero if appropriate.
1570 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1571 and if insert_regs returns a nonzero value
1572 you must then recompute its hash code before calling here.
1574 If necessary, update table showing constant values of quantities. */
1576 static struct table_elt *
1577 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1578 enum machine_mode mode, int cost, int reg_cost)
1580 struct table_elt *elt;
1582 /* If X is a register and we haven't made a quantity for it,
1583 something is wrong. */
1584 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1586 /* If X is a hard register, show it is being put in the table. */
1587 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1588 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1590 /* Put an element for X into the right hash bucket. */
1592 elt = free_element_chain;
1593 if (elt)
1594 free_element_chain = elt->next_same_hash;
1595 else
1596 elt = XNEW (struct table_elt);
1598 elt->exp = x;
1599 elt->canon_exp = NULL_RTX;
1600 elt->cost = cost;
1601 elt->regcost = reg_cost;
1602 elt->next_same_value = 0;
1603 elt->prev_same_value = 0;
1604 elt->next_same_hash = table[hash];
1605 elt->prev_same_hash = 0;
1606 elt->related_value = 0;
1607 elt->in_memory = 0;
1608 elt->mode = mode;
1609 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1611 if (table[hash])
1612 table[hash]->prev_same_hash = elt;
1613 table[hash] = elt;
1615 /* Put it into the proper value-class. */
1616 if (classp)
1618 classp = classp->first_same_value;
1619 if (CHEAPER (elt, classp))
1620 /* Insert at the head of the class. */
1622 struct table_elt *p;
1623 elt->next_same_value = classp;
1624 classp->prev_same_value = elt;
1625 elt->first_same_value = elt;
1627 for (p = classp; p; p = p->next_same_value)
1628 p->first_same_value = elt;
1630 else
1632 /* Insert not at head of the class. */
1633 /* Put it after the last element cheaper than X. */
1634 struct table_elt *p, *next;
1636 for (p = classp;
1637 (next = p->next_same_value) && CHEAPER (next, elt);
1638 p = next)
1641 /* Put it after P and before NEXT. */
1642 elt->next_same_value = next;
1643 if (next)
1644 next->prev_same_value = elt;
1646 elt->prev_same_value = p;
1647 p->next_same_value = elt;
1648 elt->first_same_value = classp;
1651 else
1652 elt->first_same_value = elt;
1654 /* If this is a constant being set equivalent to a register or a register
1655 being set equivalent to a constant, note the constant equivalence.
1657 If this is a constant, it cannot be equivalent to a different constant,
1658 and a constant is the only thing that can be cheaper than a register. So
1659 we know the register is the head of the class (before the constant was
1660 inserted).
1662 If this is a register that is not already known equivalent to a
1663 constant, we must check the entire class.
1665 If this is a register that is already known equivalent to an insn,
1666 update the qtys `const_insn' to show that `this_insn' is the latest
1667 insn making that quantity equivalent to the constant. */
1669 if (elt->is_const && classp && REG_P (classp->exp)
1670 && !REG_P (x))
1672 int exp_q = REG_QTY (REGNO (classp->exp));
1673 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1675 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1676 exp_ent->const_insn = this_insn;
1679 else if (REG_P (x)
1680 && classp
1681 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1682 && ! elt->is_const)
1684 struct table_elt *p;
1686 for (p = classp; p != 0; p = p->next_same_value)
1688 if (p->is_const && !REG_P (p->exp))
1690 int x_q = REG_QTY (REGNO (x));
1691 struct qty_table_elem *x_ent = &qty_table[x_q];
1693 x_ent->const_rtx
1694 = gen_lowpart (GET_MODE (x), p->exp);
1695 x_ent->const_insn = this_insn;
1696 break;
1701 else if (REG_P (x)
1702 && qty_table[REG_QTY (REGNO (x))].const_rtx
1703 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1704 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1706 /* If this is a constant with symbolic value,
1707 and it has a term with an explicit integer value,
1708 link it up with related expressions. */
1709 if (GET_CODE (x) == CONST)
1711 rtx subexp = get_related_value (x);
1712 unsigned subhash;
1713 struct table_elt *subelt, *subelt_prev;
1715 if (subexp != 0)
1717 /* Get the integer-free subexpression in the hash table. */
1718 subhash = SAFE_HASH (subexp, mode);
1719 subelt = lookup (subexp, subhash, mode);
1720 if (subelt == 0)
1721 subelt = insert (subexp, NULL, subhash, mode);
1722 /* Initialize SUBELT's circular chain if it has none. */
1723 if (subelt->related_value == 0)
1724 subelt->related_value = subelt;
1725 /* Find the element in the circular chain that precedes SUBELT. */
1726 subelt_prev = subelt;
1727 while (subelt_prev->related_value != subelt)
1728 subelt_prev = subelt_prev->related_value;
1729 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1730 This way the element that follows SUBELT is the oldest one. */
1731 elt->related_value = subelt_prev->related_value;
1732 subelt_prev->related_value = elt;
1736 return elt;
1739 /* Wrap insert_with_costs by passing the default costs. */
1741 static struct table_elt *
1742 insert (rtx x, struct table_elt *classp, unsigned int hash,
1743 enum machine_mode mode)
1745 return
1746 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1750 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1751 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1752 the two classes equivalent.
1754 CLASS1 will be the surviving class; CLASS2 should not be used after this
1755 call.
1757 Any invalid entries in CLASS2 will not be copied. */
1759 static void
1760 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1762 struct table_elt *elt, *next, *new_elt;
1764 /* Ensure we start with the head of the classes. */
1765 class1 = class1->first_same_value;
1766 class2 = class2->first_same_value;
1768 /* If they were already equal, forget it. */
1769 if (class1 == class2)
1770 return;
1772 for (elt = class2; elt; elt = next)
1774 unsigned int hash;
1775 rtx exp = elt->exp;
1776 enum machine_mode mode = elt->mode;
1778 next = elt->next_same_value;
1780 /* Remove old entry, make a new one in CLASS1's class.
1781 Don't do this for invalid entries as we cannot find their
1782 hash code (it also isn't necessary). */
1783 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1785 bool need_rehash = false;
1787 hash_arg_in_memory = 0;
1788 hash = HASH (exp, mode);
1790 if (REG_P (exp))
1792 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1793 delete_reg_equiv (REGNO (exp));
1796 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1797 remove_pseudo_from_table (exp, hash);
1798 else
1799 remove_from_table (elt, hash);
1801 if (insert_regs (exp, class1, 0) || need_rehash)
1803 rehash_using_reg (exp);
1804 hash = HASH (exp, mode);
1806 new_elt = insert (exp, class1, hash, mode);
1807 new_elt->in_memory = hash_arg_in_memory;
1812 /* Flush the entire hash table. */
1814 static void
1815 flush_hash_table (void)
1817 int i;
1818 struct table_elt *p;
1820 for (i = 0; i < HASH_SIZE; i++)
1821 for (p = table[i]; p; p = table[i])
1823 /* Note that invalidate can remove elements
1824 after P in the current hash chain. */
1825 if (REG_P (p->exp))
1826 invalidate (p->exp, VOIDmode);
1827 else
1828 remove_from_table (p, i);
1832 /* Function called for each rtx to check whether an anti dependence exist. */
1833 struct check_dependence_data
1835 enum machine_mode mode;
1836 rtx exp;
1837 rtx addr;
1840 static int
1841 check_dependence (rtx *x, void *data)
1843 struct check_dependence_data *d = (struct check_dependence_data *) data;
1844 if (*x && MEM_P (*x))
1845 return canon_anti_dependence (*x, true, d->exp, d->mode, d->addr);
1846 else
1847 return 0;
1850 /* Remove from the hash table, or mark as invalid, all expressions whose
1851 values could be altered by storing in X. X is a register, a subreg, or
1852 a memory reference with nonvarying address (because, when a memory
1853 reference with a varying address is stored in, all memory references are
1854 removed by invalidate_memory so specific invalidation is superfluous).
1855 FULL_MODE, if not VOIDmode, indicates that this much should be
1856 invalidated instead of just the amount indicated by the mode of X. This
1857 is only used for bitfield stores into memory.
1859 A nonvarying address may be just a register or just a symbol reference,
1860 or it may be either of those plus a numeric offset. */
1862 static void
1863 invalidate (rtx x, enum machine_mode full_mode)
1865 int i;
1866 struct table_elt *p;
1867 rtx addr;
1869 switch (GET_CODE (x))
1871 case REG:
1873 /* If X is a register, dependencies on its contents are recorded
1874 through the qty number mechanism. Just change the qty number of
1875 the register, mark it as invalid for expressions that refer to it,
1876 and remove it itself. */
1877 unsigned int regno = REGNO (x);
1878 unsigned int hash = HASH (x, GET_MODE (x));
1880 /* Remove REGNO from any quantity list it might be on and indicate
1881 that its value might have changed. If it is a pseudo, remove its
1882 entry from the hash table.
1884 For a hard register, we do the first two actions above for any
1885 additional hard registers corresponding to X. Then, if any of these
1886 registers are in the table, we must remove any REG entries that
1887 overlap these registers. */
1889 delete_reg_equiv (regno);
1890 REG_TICK (regno)++;
1891 SUBREG_TICKED (regno) = -1;
1893 if (regno >= FIRST_PSEUDO_REGISTER)
1894 remove_pseudo_from_table (x, hash);
1895 else
1897 HOST_WIDE_INT in_table
1898 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1899 unsigned int endregno = END_HARD_REGNO (x);
1900 unsigned int tregno, tendregno, rn;
1901 struct table_elt *p, *next;
1903 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1905 for (rn = regno + 1; rn < endregno; rn++)
1907 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1908 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1909 delete_reg_equiv (rn);
1910 REG_TICK (rn)++;
1911 SUBREG_TICKED (rn) = -1;
1914 if (in_table)
1915 for (hash = 0; hash < HASH_SIZE; hash++)
1916 for (p = table[hash]; p; p = next)
1918 next = p->next_same_hash;
1920 if (!REG_P (p->exp)
1921 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1922 continue;
1924 tregno = REGNO (p->exp);
1925 tendregno = END_HARD_REGNO (p->exp);
1926 if (tendregno > regno && tregno < endregno)
1927 remove_from_table (p, hash);
1931 return;
1933 case SUBREG:
1934 invalidate (SUBREG_REG (x), VOIDmode);
1935 return;
1937 case PARALLEL:
1938 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1939 invalidate (XVECEXP (x, 0, i), VOIDmode);
1940 return;
1942 case EXPR_LIST:
1943 /* This is part of a disjoint return value; extract the location in
1944 question ignoring the offset. */
1945 invalidate (XEXP (x, 0), VOIDmode);
1946 return;
1948 case MEM:
1949 addr = canon_rtx (get_addr (XEXP (x, 0)));
1950 /* Calculate the canonical version of X here so that
1951 true_dependence doesn't generate new RTL for X on each call. */
1952 x = canon_rtx (x);
1954 /* Remove all hash table elements that refer to overlapping pieces of
1955 memory. */
1956 if (full_mode == VOIDmode)
1957 full_mode = GET_MODE (x);
1959 for (i = 0; i < HASH_SIZE; i++)
1961 struct table_elt *next;
1963 for (p = table[i]; p; p = next)
1965 next = p->next_same_hash;
1966 if (p->in_memory)
1968 struct check_dependence_data d;
1970 /* Just canonicalize the expression once;
1971 otherwise each time we call invalidate
1972 true_dependence will canonicalize the
1973 expression again. */
1974 if (!p->canon_exp)
1975 p->canon_exp = canon_rtx (p->exp);
1976 d.exp = x;
1977 d.addr = addr;
1978 d.mode = full_mode;
1979 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1980 remove_from_table (p, i);
1984 return;
1986 default:
1987 gcc_unreachable ();
1991 /* Remove all expressions that refer to register REGNO,
1992 since they are already invalid, and we are about to
1993 mark that register valid again and don't want the old
1994 expressions to reappear as valid. */
1996 static void
1997 remove_invalid_refs (unsigned int regno)
1999 unsigned int i;
2000 struct table_elt *p, *next;
2002 for (i = 0; i < HASH_SIZE; i++)
2003 for (p = table[i]; p; p = next)
2005 next = p->next_same_hash;
2006 if (!REG_P (p->exp)
2007 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2008 remove_from_table (p, i);
2012 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2013 and mode MODE. */
2014 static void
2015 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2016 enum machine_mode mode)
2018 unsigned int i;
2019 struct table_elt *p, *next;
2020 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2022 for (i = 0; i < HASH_SIZE; i++)
2023 for (p = table[i]; p; p = next)
2025 rtx exp = p->exp;
2026 next = p->next_same_hash;
2028 if (!REG_P (exp)
2029 && (GET_CODE (exp) != SUBREG
2030 || !REG_P (SUBREG_REG (exp))
2031 || REGNO (SUBREG_REG (exp)) != regno
2032 || (((SUBREG_BYTE (exp)
2033 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2034 && SUBREG_BYTE (exp) <= end))
2035 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2036 remove_from_table (p, i);
2040 /* Recompute the hash codes of any valid entries in the hash table that
2041 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2043 This is called when we make a jump equivalence. */
2045 static void
2046 rehash_using_reg (rtx x)
2048 unsigned int i;
2049 struct table_elt *p, *next;
2050 unsigned hash;
2052 if (GET_CODE (x) == SUBREG)
2053 x = SUBREG_REG (x);
2055 /* If X is not a register or if the register is known not to be in any
2056 valid entries in the table, we have no work to do. */
2058 if (!REG_P (x)
2059 || REG_IN_TABLE (REGNO (x)) < 0
2060 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2061 return;
2063 /* Scan all hash chains looking for valid entries that mention X.
2064 If we find one and it is in the wrong hash chain, move it. */
2066 for (i = 0; i < HASH_SIZE; i++)
2067 for (p = table[i]; p; p = next)
2069 next = p->next_same_hash;
2070 if (reg_mentioned_p (x, p->exp)
2071 && exp_equiv_p (p->exp, p->exp, 1, false)
2072 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2074 if (p->next_same_hash)
2075 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2077 if (p->prev_same_hash)
2078 p->prev_same_hash->next_same_hash = p->next_same_hash;
2079 else
2080 table[i] = p->next_same_hash;
2082 p->next_same_hash = table[hash];
2083 p->prev_same_hash = 0;
2084 if (table[hash])
2085 table[hash]->prev_same_hash = p;
2086 table[hash] = p;
2091 /* Remove from the hash table any expression that is a call-clobbered
2092 register. Also update their TICK values. */
2094 static void
2095 invalidate_for_call (void)
2097 unsigned int regno, endregno;
2098 unsigned int i;
2099 unsigned hash;
2100 struct table_elt *p, *next;
2101 int in_table = 0;
2102 hard_reg_set_iterator hrsi;
2104 /* Go through all the hard registers. For each that is clobbered in
2105 a CALL_INSN, remove the register from quantity chains and update
2106 reg_tick if defined. Also see if any of these registers is currently
2107 in the table. */
2108 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2110 delete_reg_equiv (regno);
2111 if (REG_TICK (regno) >= 0)
2113 REG_TICK (regno)++;
2114 SUBREG_TICKED (regno) = -1;
2116 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2119 /* In the case where we have no call-clobbered hard registers in the
2120 table, we are done. Otherwise, scan the table and remove any
2121 entry that overlaps a call-clobbered register. */
2123 if (in_table)
2124 for (hash = 0; hash < HASH_SIZE; hash++)
2125 for (p = table[hash]; p; p = next)
2127 next = p->next_same_hash;
2129 if (!REG_P (p->exp)
2130 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2131 continue;
2133 regno = REGNO (p->exp);
2134 endregno = END_HARD_REGNO (p->exp);
2136 for (i = regno; i < endregno; i++)
2137 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2139 remove_from_table (p, hash);
2140 break;
2145 /* Given an expression X of type CONST,
2146 and ELT which is its table entry (or 0 if it
2147 is not in the hash table),
2148 return an alternate expression for X as a register plus integer.
2149 If none can be found, return 0. */
2151 static rtx
2152 use_related_value (rtx x, struct table_elt *elt)
2154 struct table_elt *relt = 0;
2155 struct table_elt *p, *q;
2156 HOST_WIDE_INT offset;
2158 /* First, is there anything related known?
2159 If we have a table element, we can tell from that.
2160 Otherwise, must look it up. */
2162 if (elt != 0 && elt->related_value != 0)
2163 relt = elt;
2164 else if (elt == 0 && GET_CODE (x) == CONST)
2166 rtx subexp = get_related_value (x);
2167 if (subexp != 0)
2168 relt = lookup (subexp,
2169 SAFE_HASH (subexp, GET_MODE (subexp)),
2170 GET_MODE (subexp));
2173 if (relt == 0)
2174 return 0;
2176 /* Search all related table entries for one that has an
2177 equivalent register. */
2179 p = relt;
2180 while (1)
2182 /* This loop is strange in that it is executed in two different cases.
2183 The first is when X is already in the table. Then it is searching
2184 the RELATED_VALUE list of X's class (RELT). The second case is when
2185 X is not in the table. Then RELT points to a class for the related
2186 value.
2188 Ensure that, whatever case we are in, that we ignore classes that have
2189 the same value as X. */
2191 if (rtx_equal_p (x, p->exp))
2192 q = 0;
2193 else
2194 for (q = p->first_same_value; q; q = q->next_same_value)
2195 if (REG_P (q->exp))
2196 break;
2198 if (q)
2199 break;
2201 p = p->related_value;
2203 /* We went all the way around, so there is nothing to be found.
2204 Alternatively, perhaps RELT was in the table for some other reason
2205 and it has no related values recorded. */
2206 if (p == relt || p == 0)
2207 break;
2210 if (q == 0)
2211 return 0;
2213 offset = (get_integer_term (x) - get_integer_term (p->exp));
2214 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2215 return plus_constant (q->mode, q->exp, offset);
2219 /* Hash a string. Just add its bytes up. */
2220 static inline unsigned
2221 hash_rtx_string (const char *ps)
2223 unsigned hash = 0;
2224 const unsigned char *p = (const unsigned char *) ps;
2226 if (p)
2227 while (*p)
2228 hash += *p++;
2230 return hash;
2233 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2234 When the callback returns true, we continue with the new rtx. */
2236 unsigned
2237 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2238 int *do_not_record_p, int *hash_arg_in_memory_p,
2239 bool have_reg_qty, hash_rtx_callback_function cb)
2241 int i, j;
2242 unsigned hash = 0;
2243 enum rtx_code code;
2244 const char *fmt;
2245 enum machine_mode newmode;
2246 rtx newx;
2248 /* Used to turn recursion into iteration. We can't rely on GCC's
2249 tail-recursion elimination since we need to keep accumulating values
2250 in HASH. */
2251 repeat:
2252 if (x == 0)
2253 return hash;
2255 /* Invoke the callback first. */
2256 if (cb != NULL
2257 && ((*cb) (x, mode, &newx, &newmode)))
2259 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2260 hash_arg_in_memory_p, have_reg_qty, cb);
2261 return hash;
2264 code = GET_CODE (x);
2265 switch (code)
2267 case REG:
2269 unsigned int regno = REGNO (x);
2271 if (do_not_record_p && !reload_completed)
2273 /* On some machines, we can't record any non-fixed hard register,
2274 because extending its life will cause reload problems. We
2275 consider ap, fp, sp, gp to be fixed for this purpose.
2277 We also consider CCmode registers to be fixed for this purpose;
2278 failure to do so leads to failure to simplify 0<100 type of
2279 conditionals.
2281 On all machines, we can't record any global registers.
2282 Nor should we record any register that is in a small
2283 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2284 bool record;
2286 if (regno >= FIRST_PSEUDO_REGISTER)
2287 record = true;
2288 else if (x == frame_pointer_rtx
2289 || x == hard_frame_pointer_rtx
2290 || x == arg_pointer_rtx
2291 || x == stack_pointer_rtx
2292 || x == pic_offset_table_rtx)
2293 record = true;
2294 else if (global_regs[regno])
2295 record = false;
2296 else if (fixed_regs[regno])
2297 record = true;
2298 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2299 record = true;
2300 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2301 record = false;
2302 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2303 record = false;
2304 else
2305 record = true;
2307 if (!record)
2309 *do_not_record_p = 1;
2310 return 0;
2314 hash += ((unsigned int) REG << 7);
2315 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2316 return hash;
2319 /* We handle SUBREG of a REG specially because the underlying
2320 reg changes its hash value with every value change; we don't
2321 want to have to forget unrelated subregs when one subreg changes. */
2322 case SUBREG:
2324 if (REG_P (SUBREG_REG (x)))
2326 hash += (((unsigned int) SUBREG << 7)
2327 + REGNO (SUBREG_REG (x))
2328 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2329 return hash;
2331 break;
2334 case CONST_INT:
2335 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2336 + (unsigned int) INTVAL (x));
2337 return hash;
2339 case CONST_DOUBLE:
2340 /* This is like the general case, except that it only counts
2341 the integers representing the constant. */
2342 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2343 if (GET_MODE (x) != VOIDmode)
2344 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2345 else
2346 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2347 + (unsigned int) CONST_DOUBLE_HIGH (x));
2348 return hash;
2350 case CONST_FIXED:
2351 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2352 hash += fixed_hash (CONST_FIXED_VALUE (x));
2353 return hash;
2355 case CONST_VECTOR:
2357 int units;
2358 rtx elt;
2360 units = CONST_VECTOR_NUNITS (x);
2362 for (i = 0; i < units; ++i)
2364 elt = CONST_VECTOR_ELT (x, i);
2365 hash += hash_rtx_cb (elt, GET_MODE (elt),
2366 do_not_record_p, hash_arg_in_memory_p,
2367 have_reg_qty, cb);
2370 return hash;
2373 /* Assume there is only one rtx object for any given label. */
2374 case LABEL_REF:
2375 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2376 differences and differences between each stage's debugging dumps. */
2377 hash += (((unsigned int) LABEL_REF << 7)
2378 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2379 return hash;
2381 case SYMBOL_REF:
2383 /* Don't hash on the symbol's address to avoid bootstrap differences.
2384 Different hash values may cause expressions to be recorded in
2385 different orders and thus different registers to be used in the
2386 final assembler. This also avoids differences in the dump files
2387 between various stages. */
2388 unsigned int h = 0;
2389 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2391 while (*p)
2392 h += (h << 7) + *p++; /* ??? revisit */
2394 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2395 return hash;
2398 case MEM:
2399 /* We don't record if marked volatile or if BLKmode since we don't
2400 know the size of the move. */
2401 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2403 *do_not_record_p = 1;
2404 return 0;
2406 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2407 *hash_arg_in_memory_p = 1;
2409 /* Now that we have already found this special case,
2410 might as well speed it up as much as possible. */
2411 hash += (unsigned) MEM;
2412 x = XEXP (x, 0);
2413 goto repeat;
2415 case USE:
2416 /* A USE that mentions non-volatile memory needs special
2417 handling since the MEM may be BLKmode which normally
2418 prevents an entry from being made. Pure calls are
2419 marked by a USE which mentions BLKmode memory.
2420 See calls.c:emit_call_1. */
2421 if (MEM_P (XEXP (x, 0))
2422 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2424 hash += (unsigned) USE;
2425 x = XEXP (x, 0);
2427 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2428 *hash_arg_in_memory_p = 1;
2430 /* Now that we have already found this special case,
2431 might as well speed it up as much as possible. */
2432 hash += (unsigned) MEM;
2433 x = XEXP (x, 0);
2434 goto repeat;
2436 break;
2438 case PRE_DEC:
2439 case PRE_INC:
2440 case POST_DEC:
2441 case POST_INC:
2442 case PRE_MODIFY:
2443 case POST_MODIFY:
2444 case PC:
2445 case CC0:
2446 case CALL:
2447 case UNSPEC_VOLATILE:
2448 if (do_not_record_p) {
2449 *do_not_record_p = 1;
2450 return 0;
2452 else
2453 return hash;
2454 break;
2456 case ASM_OPERANDS:
2457 if (do_not_record_p && MEM_VOLATILE_P (x))
2459 *do_not_record_p = 1;
2460 return 0;
2462 else
2464 /* We don't want to take the filename and line into account. */
2465 hash += (unsigned) code + (unsigned) GET_MODE (x)
2466 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2467 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2468 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2470 if (ASM_OPERANDS_INPUT_LENGTH (x))
2472 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2474 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2475 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2476 do_not_record_p, hash_arg_in_memory_p,
2477 have_reg_qty, cb)
2478 + hash_rtx_string
2479 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2482 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2483 x = ASM_OPERANDS_INPUT (x, 0);
2484 mode = GET_MODE (x);
2485 goto repeat;
2488 return hash;
2490 break;
2492 default:
2493 break;
2496 i = GET_RTX_LENGTH (code) - 1;
2497 hash += (unsigned) code + (unsigned) GET_MODE (x);
2498 fmt = GET_RTX_FORMAT (code);
2499 for (; i >= 0; i--)
2501 switch (fmt[i])
2503 case 'e':
2504 /* If we are about to do the last recursive call
2505 needed at this level, change it into iteration.
2506 This function is called enough to be worth it. */
2507 if (i == 0)
2509 x = XEXP (x, i);
2510 goto repeat;
2513 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2514 hash_arg_in_memory_p,
2515 have_reg_qty, cb);
2516 break;
2518 case 'E':
2519 for (j = 0; j < XVECLEN (x, i); j++)
2520 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2521 hash_arg_in_memory_p,
2522 have_reg_qty, cb);
2523 break;
2525 case 's':
2526 hash += hash_rtx_string (XSTR (x, i));
2527 break;
2529 case 'i':
2530 hash += (unsigned int) XINT (x, i);
2531 break;
2533 case '0': case 't':
2534 /* Unused. */
2535 break;
2537 default:
2538 gcc_unreachable ();
2542 return hash;
2545 /* Hash an rtx. We are careful to make sure the value is never negative.
2546 Equivalent registers hash identically.
2547 MODE is used in hashing for CONST_INTs only;
2548 otherwise the mode of X is used.
2550 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2552 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2553 a MEM rtx which does not have the MEM_READONLY_P flag set.
2555 Note that cse_insn knows that the hash code of a MEM expression
2556 is just (int) MEM plus the hash code of the address. */
2558 unsigned
2559 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2560 int *hash_arg_in_memory_p, bool have_reg_qty)
2562 return hash_rtx_cb (x, mode, do_not_record_p,
2563 hash_arg_in_memory_p, have_reg_qty, NULL);
2566 /* Hash an rtx X for cse via hash_rtx.
2567 Stores 1 in do_not_record if any subexpression is volatile.
2568 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2569 does not have the MEM_READONLY_P flag set. */
2571 static inline unsigned
2572 canon_hash (rtx x, enum machine_mode mode)
2574 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2577 /* Like canon_hash but with no side effects, i.e. do_not_record
2578 and hash_arg_in_memory are not changed. */
2580 static inline unsigned
2581 safe_hash (rtx x, enum machine_mode mode)
2583 int dummy_do_not_record;
2584 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2587 /* Return 1 iff X and Y would canonicalize into the same thing,
2588 without actually constructing the canonicalization of either one.
2589 If VALIDATE is nonzero,
2590 we assume X is an expression being processed from the rtl
2591 and Y was found in the hash table. We check register refs
2592 in Y for being marked as valid.
2594 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2597 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2599 int i, j;
2600 enum rtx_code code;
2601 const char *fmt;
2603 /* Note: it is incorrect to assume an expression is equivalent to itself
2604 if VALIDATE is nonzero. */
2605 if (x == y && !validate)
2606 return 1;
2608 if (x == 0 || y == 0)
2609 return x == y;
2611 code = GET_CODE (x);
2612 if (code != GET_CODE (y))
2613 return 0;
2615 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2616 if (GET_MODE (x) != GET_MODE (y))
2617 return 0;
2619 /* MEMs referring to different address space are not equivalent. */
2620 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2621 return 0;
2623 switch (code)
2625 case PC:
2626 case CC0:
2627 CASE_CONST_UNIQUE:
2628 return x == y;
2630 case LABEL_REF:
2631 return XEXP (x, 0) == XEXP (y, 0);
2633 case SYMBOL_REF:
2634 return XSTR (x, 0) == XSTR (y, 0);
2636 case REG:
2637 if (for_gcse)
2638 return REGNO (x) == REGNO (y);
2639 else
2641 unsigned int regno = REGNO (y);
2642 unsigned int i;
2643 unsigned int endregno = END_REGNO (y);
2645 /* If the quantities are not the same, the expressions are not
2646 equivalent. If there are and we are not to validate, they
2647 are equivalent. Otherwise, ensure all regs are up-to-date. */
2649 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2650 return 0;
2652 if (! validate)
2653 return 1;
2655 for (i = regno; i < endregno; i++)
2656 if (REG_IN_TABLE (i) != REG_TICK (i))
2657 return 0;
2659 return 1;
2662 case MEM:
2663 if (for_gcse)
2665 /* A volatile mem should not be considered equivalent to any
2666 other. */
2667 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2668 return 0;
2670 /* Can't merge two expressions in different alias sets, since we
2671 can decide that the expression is transparent in a block when
2672 it isn't, due to it being set with the different alias set.
2674 Also, can't merge two expressions with different MEM_ATTRS.
2675 They could e.g. be two different entities allocated into the
2676 same space on the stack (see e.g. PR25130). In that case, the
2677 MEM addresses can be the same, even though the two MEMs are
2678 absolutely not equivalent.
2680 But because really all MEM attributes should be the same for
2681 equivalent MEMs, we just use the invariant that MEMs that have
2682 the same attributes share the same mem_attrs data structure. */
2683 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2684 return 0;
2686 break;
2688 /* For commutative operations, check both orders. */
2689 case PLUS:
2690 case MULT:
2691 case AND:
2692 case IOR:
2693 case XOR:
2694 case NE:
2695 case EQ:
2696 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2697 validate, for_gcse)
2698 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2699 validate, for_gcse))
2700 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2701 validate, for_gcse)
2702 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2703 validate, for_gcse)));
2705 case ASM_OPERANDS:
2706 /* We don't use the generic code below because we want to
2707 disregard filename and line numbers. */
2709 /* A volatile asm isn't equivalent to any other. */
2710 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2711 return 0;
2713 if (GET_MODE (x) != GET_MODE (y)
2714 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2715 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2716 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2717 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2718 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2719 return 0;
2721 if (ASM_OPERANDS_INPUT_LENGTH (x))
2723 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2724 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2725 ASM_OPERANDS_INPUT (y, i),
2726 validate, for_gcse)
2727 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2728 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2729 return 0;
2732 return 1;
2734 default:
2735 break;
2738 /* Compare the elements. If any pair of corresponding elements
2739 fail to match, return 0 for the whole thing. */
2741 fmt = GET_RTX_FORMAT (code);
2742 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2744 switch (fmt[i])
2746 case 'e':
2747 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2748 validate, for_gcse))
2749 return 0;
2750 break;
2752 case 'E':
2753 if (XVECLEN (x, i) != XVECLEN (y, i))
2754 return 0;
2755 for (j = 0; j < XVECLEN (x, i); j++)
2756 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2757 validate, for_gcse))
2758 return 0;
2759 break;
2761 case 's':
2762 if (strcmp (XSTR (x, i), XSTR (y, i)))
2763 return 0;
2764 break;
2766 case 'i':
2767 if (XINT (x, i) != XINT (y, i))
2768 return 0;
2769 break;
2771 case 'w':
2772 if (XWINT (x, i) != XWINT (y, i))
2773 return 0;
2774 break;
2776 case '0':
2777 case 't':
2778 break;
2780 default:
2781 gcc_unreachable ();
2785 return 1;
2788 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2789 the result if necessary. INSN is as for canon_reg. */
2791 static void
2792 validate_canon_reg (rtx *xloc, rtx insn)
2794 if (*xloc)
2796 rtx new_rtx = canon_reg (*xloc, insn);
2798 /* If replacing pseudo with hard reg or vice versa, ensure the
2799 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2800 gcc_assert (insn && new_rtx);
2801 validate_change (insn, xloc, new_rtx, 1);
2805 /* Canonicalize an expression:
2806 replace each register reference inside it
2807 with the "oldest" equivalent register.
2809 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2810 after we make our substitution. The calls are made with IN_GROUP nonzero
2811 so apply_change_group must be called upon the outermost return from this
2812 function (unless INSN is zero). The result of apply_change_group can
2813 generally be discarded since the changes we are making are optional. */
2815 static rtx
2816 canon_reg (rtx x, rtx insn)
2818 int i;
2819 enum rtx_code code;
2820 const char *fmt;
2822 if (x == 0)
2823 return x;
2825 code = GET_CODE (x);
2826 switch (code)
2828 case PC:
2829 case CC0:
2830 case CONST:
2831 CASE_CONST_ANY:
2832 case SYMBOL_REF:
2833 case LABEL_REF:
2834 case ADDR_VEC:
2835 case ADDR_DIFF_VEC:
2836 return x;
2838 case REG:
2840 int first;
2841 int q;
2842 struct qty_table_elem *ent;
2844 /* Never replace a hard reg, because hard regs can appear
2845 in more than one machine mode, and we must preserve the mode
2846 of each occurrence. Also, some hard regs appear in
2847 MEMs that are shared and mustn't be altered. Don't try to
2848 replace any reg that maps to a reg of class NO_REGS. */
2849 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2850 || ! REGNO_QTY_VALID_P (REGNO (x)))
2851 return x;
2853 q = REG_QTY (REGNO (x));
2854 ent = &qty_table[q];
2855 first = ent->first_reg;
2856 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2857 : REGNO_REG_CLASS (first) == NO_REGS ? x
2858 : gen_rtx_REG (ent->mode, first));
2861 default:
2862 break;
2865 fmt = GET_RTX_FORMAT (code);
2866 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2868 int j;
2870 if (fmt[i] == 'e')
2871 validate_canon_reg (&XEXP (x, i), insn);
2872 else if (fmt[i] == 'E')
2873 for (j = 0; j < XVECLEN (x, i); j++)
2874 validate_canon_reg (&XVECEXP (x, i, j), insn);
2877 return x;
2880 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2881 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2882 what values are being compared.
2884 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2885 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2886 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2887 compared to produce cc0.
2889 The return value is the comparison operator and is either the code of
2890 A or the code corresponding to the inverse of the comparison. */
2892 static enum rtx_code
2893 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2894 enum machine_mode *pmode1, enum machine_mode *pmode2)
2896 rtx arg1, arg2;
2897 struct pointer_set_t *visited = NULL;
2898 /* Set nonzero when we find something of interest. */
2899 rtx x = NULL;
2901 arg1 = *parg1, arg2 = *parg2;
2903 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2905 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2907 int reverse_code = 0;
2908 struct table_elt *p = 0;
2910 /* Remember state from previous iteration. */
2911 if (x)
2913 if (!visited)
2914 visited = pointer_set_create ();
2915 pointer_set_insert (visited, x);
2916 x = 0;
2919 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2920 On machines with CC0, this is the only case that can occur, since
2921 fold_rtx will return the COMPARE or item being compared with zero
2922 when given CC0. */
2924 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2925 x = arg1;
2927 /* If ARG1 is a comparison operator and CODE is testing for
2928 STORE_FLAG_VALUE, get the inner arguments. */
2930 else if (COMPARISON_P (arg1))
2932 #ifdef FLOAT_STORE_FLAG_VALUE
2933 REAL_VALUE_TYPE fsfv;
2934 #endif
2936 if (code == NE
2937 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2938 && code == LT && STORE_FLAG_VALUE == -1)
2939 #ifdef FLOAT_STORE_FLAG_VALUE
2940 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2941 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2942 REAL_VALUE_NEGATIVE (fsfv)))
2943 #endif
2945 x = arg1;
2946 else if (code == EQ
2947 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2948 && code == GE && STORE_FLAG_VALUE == -1)
2949 #ifdef FLOAT_STORE_FLAG_VALUE
2950 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2951 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2952 REAL_VALUE_NEGATIVE (fsfv)))
2953 #endif
2955 x = arg1, reverse_code = 1;
2958 /* ??? We could also check for
2960 (ne (and (eq (...) (const_int 1))) (const_int 0))
2962 and related forms, but let's wait until we see them occurring. */
2964 if (x == 0)
2965 /* Look up ARG1 in the hash table and see if it has an equivalence
2966 that lets us see what is being compared. */
2967 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2968 if (p)
2970 p = p->first_same_value;
2972 /* If what we compare is already known to be constant, that is as
2973 good as it gets.
2974 We need to break the loop in this case, because otherwise we
2975 can have an infinite loop when looking at a reg that is known
2976 to be a constant which is the same as a comparison of a reg
2977 against zero which appears later in the insn stream, which in
2978 turn is constant and the same as the comparison of the first reg
2979 against zero... */
2980 if (p->is_const)
2981 break;
2984 for (; p; p = p->next_same_value)
2986 enum machine_mode inner_mode = GET_MODE (p->exp);
2987 #ifdef FLOAT_STORE_FLAG_VALUE
2988 REAL_VALUE_TYPE fsfv;
2989 #endif
2991 /* If the entry isn't valid, skip it. */
2992 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2993 continue;
2995 /* If it's a comparison we've used before, skip it. */
2996 if (visited && pointer_set_contains (visited, p->exp))
2997 continue;
2999 if (GET_CODE (p->exp) == COMPARE
3000 /* Another possibility is that this machine has a compare insn
3001 that includes the comparison code. In that case, ARG1 would
3002 be equivalent to a comparison operation that would set ARG1 to
3003 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3004 ORIG_CODE is the actual comparison being done; if it is an EQ,
3005 we must reverse ORIG_CODE. On machine with a negative value
3006 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3007 || ((code == NE
3008 || (code == LT
3009 && val_signbit_known_set_p (inner_mode,
3010 STORE_FLAG_VALUE))
3011 #ifdef FLOAT_STORE_FLAG_VALUE
3012 || (code == LT
3013 && SCALAR_FLOAT_MODE_P (inner_mode)
3014 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3015 REAL_VALUE_NEGATIVE (fsfv)))
3016 #endif
3018 && COMPARISON_P (p->exp)))
3020 x = p->exp;
3021 break;
3023 else if ((code == EQ
3024 || (code == GE
3025 && val_signbit_known_set_p (inner_mode,
3026 STORE_FLAG_VALUE))
3027 #ifdef FLOAT_STORE_FLAG_VALUE
3028 || (code == GE
3029 && SCALAR_FLOAT_MODE_P (inner_mode)
3030 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3031 REAL_VALUE_NEGATIVE (fsfv)))
3032 #endif
3034 && COMPARISON_P (p->exp))
3036 reverse_code = 1;
3037 x = p->exp;
3038 break;
3041 /* If this non-trapping address, e.g. fp + constant, the
3042 equivalent is a better operand since it may let us predict
3043 the value of the comparison. */
3044 else if (!rtx_addr_can_trap_p (p->exp))
3046 arg1 = p->exp;
3047 continue;
3051 /* If we didn't find a useful equivalence for ARG1, we are done.
3052 Otherwise, set up for the next iteration. */
3053 if (x == 0)
3054 break;
3056 /* If we need to reverse the comparison, make sure that that is
3057 possible -- we can't necessarily infer the value of GE from LT
3058 with floating-point operands. */
3059 if (reverse_code)
3061 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3062 if (reversed == UNKNOWN)
3063 break;
3064 else
3065 code = reversed;
3067 else if (COMPARISON_P (x))
3068 code = GET_CODE (x);
3069 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3072 /* Return our results. Return the modes from before fold_rtx
3073 because fold_rtx might produce const_int, and then it's too late. */
3074 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3075 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3077 if (visited)
3078 pointer_set_destroy (visited);
3079 return code;
3082 /* If X is a nontrivial arithmetic operation on an argument for which
3083 a constant value can be determined, return the result of operating
3084 on that value, as a constant. Otherwise, return X, possibly with
3085 one or more operands changed to a forward-propagated constant.
3087 If X is a register whose contents are known, we do NOT return
3088 those contents here; equiv_constant is called to perform that task.
3089 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3091 INSN is the insn that we may be modifying. If it is 0, make a copy
3092 of X before modifying it. */
3094 static rtx
3095 fold_rtx (rtx x, rtx insn)
3097 enum rtx_code code;
3098 enum machine_mode mode;
3099 const char *fmt;
3100 int i;
3101 rtx new_rtx = 0;
3102 int changed = 0;
3104 /* Operands of X. */
3105 rtx folded_arg0;
3106 rtx folded_arg1;
3108 /* Constant equivalents of first three operands of X;
3109 0 when no such equivalent is known. */
3110 rtx const_arg0;
3111 rtx const_arg1;
3112 rtx const_arg2;
3114 /* The mode of the first operand of X. We need this for sign and zero
3115 extends. */
3116 enum machine_mode mode_arg0;
3118 if (x == 0)
3119 return x;
3121 /* Try to perform some initial simplifications on X. */
3122 code = GET_CODE (x);
3123 switch (code)
3125 case MEM:
3126 case SUBREG:
3127 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3128 return new_rtx;
3129 return x;
3131 case CONST:
3132 CASE_CONST_ANY:
3133 case SYMBOL_REF:
3134 case LABEL_REF:
3135 case REG:
3136 case PC:
3137 /* No use simplifying an EXPR_LIST
3138 since they are used only for lists of args
3139 in a function call's REG_EQUAL note. */
3140 case EXPR_LIST:
3141 return x;
3143 #ifdef HAVE_cc0
3144 case CC0:
3145 return prev_insn_cc0;
3146 #endif
3148 case ASM_OPERANDS:
3149 if (insn)
3151 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3152 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3153 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3155 return x;
3157 #ifdef NO_FUNCTION_CSE
3158 case CALL:
3159 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3160 return x;
3161 break;
3162 #endif
3164 /* Anything else goes through the loop below. */
3165 default:
3166 break;
3169 mode = GET_MODE (x);
3170 const_arg0 = 0;
3171 const_arg1 = 0;
3172 const_arg2 = 0;
3173 mode_arg0 = VOIDmode;
3175 /* Try folding our operands.
3176 Then see which ones have constant values known. */
3178 fmt = GET_RTX_FORMAT (code);
3179 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3180 if (fmt[i] == 'e')
3182 rtx folded_arg = XEXP (x, i), const_arg;
3183 enum machine_mode mode_arg = GET_MODE (folded_arg);
3185 switch (GET_CODE (folded_arg))
3187 case MEM:
3188 case REG:
3189 case SUBREG:
3190 const_arg = equiv_constant (folded_arg);
3191 break;
3193 case CONST:
3194 CASE_CONST_ANY:
3195 case SYMBOL_REF:
3196 case LABEL_REF:
3197 const_arg = folded_arg;
3198 break;
3200 #ifdef HAVE_cc0
3201 case CC0:
3202 /* The cc0-user and cc0-setter may be in different blocks if
3203 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3204 will have been cleared as we exited the block with the
3205 setter.
3207 While we could potentially track cc0 in this case, it just
3208 doesn't seem to be worth it given that cc0 targets are not
3209 terribly common or important these days and trapping math
3210 is rarely used. The combination of those two conditions
3211 necessary to trip this situation is exceedingly rare in the
3212 real world. */
3213 if (!prev_insn_cc0)
3215 const_arg = NULL_RTX;
3217 else
3219 folded_arg = prev_insn_cc0;
3220 mode_arg = prev_insn_cc0_mode;
3221 const_arg = equiv_constant (folded_arg);
3223 break;
3224 #endif
3226 default:
3227 folded_arg = fold_rtx (folded_arg, insn);
3228 const_arg = equiv_constant (folded_arg);
3229 break;
3232 /* For the first three operands, see if the operand
3233 is constant or equivalent to a constant. */
3234 switch (i)
3236 case 0:
3237 folded_arg0 = folded_arg;
3238 const_arg0 = const_arg;
3239 mode_arg0 = mode_arg;
3240 break;
3241 case 1:
3242 folded_arg1 = folded_arg;
3243 const_arg1 = const_arg;
3244 break;
3245 case 2:
3246 const_arg2 = const_arg;
3247 break;
3250 /* Pick the least expensive of the argument and an equivalent constant
3251 argument. */
3252 if (const_arg != 0
3253 && const_arg != folded_arg
3254 && COST_IN (const_arg, code, i) <= COST_IN (folded_arg, code, i)
3256 /* It's not safe to substitute the operand of a conversion
3257 operator with a constant, as the conversion's identity
3258 depends upon the mode of its operand. This optimization
3259 is handled by the call to simplify_unary_operation. */
3260 && (GET_RTX_CLASS (code) != RTX_UNARY
3261 || GET_MODE (const_arg) == mode_arg0
3262 || (code != ZERO_EXTEND
3263 && code != SIGN_EXTEND
3264 && code != TRUNCATE
3265 && code != FLOAT_TRUNCATE
3266 && code != FLOAT_EXTEND
3267 && code != FLOAT
3268 && code != FIX
3269 && code != UNSIGNED_FLOAT
3270 && code != UNSIGNED_FIX)))
3271 folded_arg = const_arg;
3273 if (folded_arg == XEXP (x, i))
3274 continue;
3276 if (insn == NULL_RTX && !changed)
3277 x = copy_rtx (x);
3278 changed = 1;
3279 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3282 if (changed)
3284 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3285 consistent with the order in X. */
3286 if (canonicalize_change_group (insn, x))
3288 rtx tem;
3289 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3290 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3293 apply_change_group ();
3296 /* If X is an arithmetic operation, see if we can simplify it. */
3298 switch (GET_RTX_CLASS (code))
3300 case RTX_UNARY:
3302 /* We can't simplify extension ops unless we know the
3303 original mode. */
3304 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3305 && mode_arg0 == VOIDmode)
3306 break;
3308 new_rtx = simplify_unary_operation (code, mode,
3309 const_arg0 ? const_arg0 : folded_arg0,
3310 mode_arg0);
3312 break;
3314 case RTX_COMPARE:
3315 case RTX_COMM_COMPARE:
3316 /* See what items are actually being compared and set FOLDED_ARG[01]
3317 to those values and CODE to the actual comparison code. If any are
3318 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3319 do anything if both operands are already known to be constant. */
3321 /* ??? Vector mode comparisons are not supported yet. */
3322 if (VECTOR_MODE_P (mode))
3323 break;
3325 if (const_arg0 == 0 || const_arg1 == 0)
3327 struct table_elt *p0, *p1;
3328 rtx true_rtx, false_rtx;
3329 enum machine_mode mode_arg1;
3331 if (SCALAR_FLOAT_MODE_P (mode))
3333 #ifdef FLOAT_STORE_FLAG_VALUE
3334 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3335 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3336 #else
3337 true_rtx = NULL_RTX;
3338 #endif
3339 false_rtx = CONST0_RTX (mode);
3341 else
3343 true_rtx = const_true_rtx;
3344 false_rtx = const0_rtx;
3347 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3348 &mode_arg0, &mode_arg1);
3350 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3351 what kinds of things are being compared, so we can't do
3352 anything with this comparison. */
3354 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3355 break;
3357 const_arg0 = equiv_constant (folded_arg0);
3358 const_arg1 = equiv_constant (folded_arg1);
3360 /* If we do not now have two constants being compared, see
3361 if we can nevertheless deduce some things about the
3362 comparison. */
3363 if (const_arg0 == 0 || const_arg1 == 0)
3365 if (const_arg1 != NULL)
3367 rtx cheapest_simplification;
3368 int cheapest_cost;
3369 rtx simp_result;
3370 struct table_elt *p;
3372 /* See if we can find an equivalent of folded_arg0
3373 that gets us a cheaper expression, possibly a
3374 constant through simplifications. */
3375 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3376 mode_arg0);
3378 if (p != NULL)
3380 cheapest_simplification = x;
3381 cheapest_cost = COST (x);
3383 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3385 int cost;
3387 /* If the entry isn't valid, skip it. */
3388 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3389 continue;
3391 /* Try to simplify using this equivalence. */
3392 simp_result
3393 = simplify_relational_operation (code, mode,
3394 mode_arg0,
3395 p->exp,
3396 const_arg1);
3398 if (simp_result == NULL)
3399 continue;
3401 cost = COST (simp_result);
3402 if (cost < cheapest_cost)
3404 cheapest_cost = cost;
3405 cheapest_simplification = simp_result;
3409 /* If we have a cheaper expression now, use that
3410 and try folding it further, from the top. */
3411 if (cheapest_simplification != x)
3412 return fold_rtx (copy_rtx (cheapest_simplification),
3413 insn);
3417 /* See if the two operands are the same. */
3419 if ((REG_P (folded_arg0)
3420 && REG_P (folded_arg1)
3421 && (REG_QTY (REGNO (folded_arg0))
3422 == REG_QTY (REGNO (folded_arg1))))
3423 || ((p0 = lookup (folded_arg0,
3424 SAFE_HASH (folded_arg0, mode_arg0),
3425 mode_arg0))
3426 && (p1 = lookup (folded_arg1,
3427 SAFE_HASH (folded_arg1, mode_arg0),
3428 mode_arg0))
3429 && p0->first_same_value == p1->first_same_value))
3430 folded_arg1 = folded_arg0;
3432 /* If FOLDED_ARG0 is a register, see if the comparison we are
3433 doing now is either the same as we did before or the reverse
3434 (we only check the reverse if not floating-point). */
3435 else if (REG_P (folded_arg0))
3437 int qty = REG_QTY (REGNO (folded_arg0));
3439 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3441 struct qty_table_elem *ent = &qty_table[qty];
3443 if ((comparison_dominates_p (ent->comparison_code, code)
3444 || (! FLOAT_MODE_P (mode_arg0)
3445 && comparison_dominates_p (ent->comparison_code,
3446 reverse_condition (code))))
3447 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3448 || (const_arg1
3449 && rtx_equal_p (ent->comparison_const,
3450 const_arg1))
3451 || (REG_P (folded_arg1)
3452 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3454 if (comparison_dominates_p (ent->comparison_code, code))
3456 if (true_rtx)
3457 return true_rtx;
3458 else
3459 break;
3461 else
3462 return false_rtx;
3469 /* If we are comparing against zero, see if the first operand is
3470 equivalent to an IOR with a constant. If so, we may be able to
3471 determine the result of this comparison. */
3472 if (const_arg1 == const0_rtx && !const_arg0)
3474 rtx y = lookup_as_function (folded_arg0, IOR);
3475 rtx inner_const;
3477 if (y != 0
3478 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3479 && CONST_INT_P (inner_const)
3480 && INTVAL (inner_const) != 0)
3481 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3485 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3486 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3487 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3488 op0, op1);
3490 break;
3492 case RTX_BIN_ARITH:
3493 case RTX_COMM_ARITH:
3494 switch (code)
3496 case PLUS:
3497 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3498 with that LABEL_REF as its second operand. If so, the result is
3499 the first operand of that MINUS. This handles switches with an
3500 ADDR_DIFF_VEC table. */
3501 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3503 rtx y
3504 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3505 : lookup_as_function (folded_arg0, MINUS);
3507 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3508 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3509 return XEXP (y, 0);
3511 /* Now try for a CONST of a MINUS like the above. */
3512 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3513 : lookup_as_function (folded_arg0, CONST))) != 0
3514 && GET_CODE (XEXP (y, 0)) == MINUS
3515 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3516 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3517 return XEXP (XEXP (y, 0), 0);
3520 /* Likewise if the operands are in the other order. */
3521 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3523 rtx y
3524 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3525 : lookup_as_function (folded_arg1, MINUS);
3527 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3528 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3529 return XEXP (y, 0);
3531 /* Now try for a CONST of a MINUS like the above. */
3532 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3533 : lookup_as_function (folded_arg1, CONST))) != 0
3534 && GET_CODE (XEXP (y, 0)) == MINUS
3535 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3536 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3537 return XEXP (XEXP (y, 0), 0);
3540 /* If second operand is a register equivalent to a negative
3541 CONST_INT, see if we can find a register equivalent to the
3542 positive constant. Make a MINUS if so. Don't do this for
3543 a non-negative constant since we might then alternate between
3544 choosing positive and negative constants. Having the positive
3545 constant previously-used is the more common case. Be sure
3546 the resulting constant is non-negative; if const_arg1 were
3547 the smallest negative number this would overflow: depending
3548 on the mode, this would either just be the same value (and
3549 hence not save anything) or be incorrect. */
3550 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3551 && INTVAL (const_arg1) < 0
3552 /* This used to test
3554 -INTVAL (const_arg1) >= 0
3556 But The Sun V5.0 compilers mis-compiled that test. So
3557 instead we test for the problematic value in a more direct
3558 manner and hope the Sun compilers get it correct. */
3559 && INTVAL (const_arg1) !=
3560 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3561 && REG_P (folded_arg1))
3563 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3564 struct table_elt *p
3565 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3567 if (p)
3568 for (p = p->first_same_value; p; p = p->next_same_value)
3569 if (REG_P (p->exp))
3570 return simplify_gen_binary (MINUS, mode, folded_arg0,
3571 canon_reg (p->exp, NULL_RTX));
3573 goto from_plus;
3575 case MINUS:
3576 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3577 If so, produce (PLUS Z C2-C). */
3578 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3580 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3581 if (y && CONST_INT_P (XEXP (y, 1)))
3582 return fold_rtx (plus_constant (mode, copy_rtx (y),
3583 -INTVAL (const_arg1)),
3584 NULL_RTX);
3587 /* Fall through. */
3589 from_plus:
3590 case SMIN: case SMAX: case UMIN: case UMAX:
3591 case IOR: case AND: case XOR:
3592 case MULT:
3593 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3594 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3595 is known to be of similar form, we may be able to replace the
3596 operation with a combined operation. This may eliminate the
3597 intermediate operation if every use is simplified in this way.
3598 Note that the similar optimization done by combine.c only works
3599 if the intermediate operation's result has only one reference. */
3601 if (REG_P (folded_arg0)
3602 && const_arg1 && CONST_INT_P (const_arg1))
3604 int is_shift
3605 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3606 rtx y, inner_const, new_const;
3607 rtx canon_const_arg1 = const_arg1;
3608 enum rtx_code associate_code;
3610 if (is_shift
3611 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3612 || INTVAL (const_arg1) < 0))
3614 if (SHIFT_COUNT_TRUNCATED)
3615 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3616 & (GET_MODE_BITSIZE (mode)
3617 - 1));
3618 else
3619 break;
3622 y = lookup_as_function (folded_arg0, code);
3623 if (y == 0)
3624 break;
3626 /* If we have compiled a statement like
3627 "if (x == (x & mask1))", and now are looking at
3628 "x & mask2", we will have a case where the first operand
3629 of Y is the same as our first operand. Unless we detect
3630 this case, an infinite loop will result. */
3631 if (XEXP (y, 0) == folded_arg0)
3632 break;
3634 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3635 if (!inner_const || !CONST_INT_P (inner_const))
3636 break;
3638 /* Don't associate these operations if they are a PLUS with the
3639 same constant and it is a power of two. These might be doable
3640 with a pre- or post-increment. Similarly for two subtracts of
3641 identical powers of two with post decrement. */
3643 if (code == PLUS && const_arg1 == inner_const
3644 && ((HAVE_PRE_INCREMENT
3645 && exact_log2 (INTVAL (const_arg1)) >= 0)
3646 || (HAVE_POST_INCREMENT
3647 && exact_log2 (INTVAL (const_arg1)) >= 0)
3648 || (HAVE_PRE_DECREMENT
3649 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3650 || (HAVE_POST_DECREMENT
3651 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3652 break;
3654 /* ??? Vector mode shifts by scalar
3655 shift operand are not supported yet. */
3656 if (is_shift && VECTOR_MODE_P (mode))
3657 break;
3659 if (is_shift
3660 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3661 || INTVAL (inner_const) < 0))
3663 if (SHIFT_COUNT_TRUNCATED)
3664 inner_const = GEN_INT (INTVAL (inner_const)
3665 & (GET_MODE_BITSIZE (mode) - 1));
3666 else
3667 break;
3670 /* Compute the code used to compose the constants. For example,
3671 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3673 associate_code = (is_shift || code == MINUS ? PLUS : code);
3675 new_const = simplify_binary_operation (associate_code, mode,
3676 canon_const_arg1,
3677 inner_const);
3679 if (new_const == 0)
3680 break;
3682 /* If we are associating shift operations, don't let this
3683 produce a shift of the size of the object or larger.
3684 This could occur when we follow a sign-extend by a right
3685 shift on a machine that does a sign-extend as a pair
3686 of shifts. */
3688 if (is_shift
3689 && CONST_INT_P (new_const)
3690 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3692 /* As an exception, we can turn an ASHIFTRT of this
3693 form into a shift of the number of bits - 1. */
3694 if (code == ASHIFTRT)
3695 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3696 else if (!side_effects_p (XEXP (y, 0)))
3697 return CONST0_RTX (mode);
3698 else
3699 break;
3702 y = copy_rtx (XEXP (y, 0));
3704 /* If Y contains our first operand (the most common way this
3705 can happen is if Y is a MEM), we would do into an infinite
3706 loop if we tried to fold it. So don't in that case. */
3708 if (! reg_mentioned_p (folded_arg0, y))
3709 y = fold_rtx (y, insn);
3711 return simplify_gen_binary (code, mode, y, new_const);
3713 break;
3715 case DIV: case UDIV:
3716 /* ??? The associative optimization performed immediately above is
3717 also possible for DIV and UDIV using associate_code of MULT.
3718 However, we would need extra code to verify that the
3719 multiplication does not overflow, that is, there is no overflow
3720 in the calculation of new_const. */
3721 break;
3723 default:
3724 break;
3727 new_rtx = simplify_binary_operation (code, mode,
3728 const_arg0 ? const_arg0 : folded_arg0,
3729 const_arg1 ? const_arg1 : folded_arg1);
3730 break;
3732 case RTX_OBJ:
3733 /* (lo_sum (high X) X) is simply X. */
3734 if (code == LO_SUM && const_arg0 != 0
3735 && GET_CODE (const_arg0) == HIGH
3736 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3737 return const_arg1;
3738 break;
3740 case RTX_TERNARY:
3741 case RTX_BITFIELD_OPS:
3742 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3743 const_arg0 ? const_arg0 : folded_arg0,
3744 const_arg1 ? const_arg1 : folded_arg1,
3745 const_arg2 ? const_arg2 : XEXP (x, 2));
3746 break;
3748 default:
3749 break;
3752 return new_rtx ? new_rtx : x;
3755 /* Return a constant value currently equivalent to X.
3756 Return 0 if we don't know one. */
3758 static rtx
3759 equiv_constant (rtx x)
3761 if (REG_P (x)
3762 && REGNO_QTY_VALID_P (REGNO (x)))
3764 int x_q = REG_QTY (REGNO (x));
3765 struct qty_table_elem *x_ent = &qty_table[x_q];
3767 if (x_ent->const_rtx)
3768 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3771 if (x == 0 || CONSTANT_P (x))
3772 return x;
3774 if (GET_CODE (x) == SUBREG)
3776 enum machine_mode mode = GET_MODE (x);
3777 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3778 rtx new_rtx;
3780 /* See if we previously assigned a constant value to this SUBREG. */
3781 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3782 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3783 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3784 return new_rtx;
3786 /* If we didn't and if doing so makes sense, see if we previously
3787 assigned a constant value to the enclosing word mode SUBREG. */
3788 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3789 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3791 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3792 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3794 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3795 new_rtx = lookup_as_function (y, CONST_INT);
3796 if (new_rtx)
3797 return gen_lowpart (mode, new_rtx);
3801 /* Otherwise see if we already have a constant for the inner REG,
3802 and if that is enough to calculate an equivalent constant for
3803 the subreg. Note that the upper bits of paradoxical subregs
3804 are undefined, so they cannot be said to equal anything. */
3805 if (REG_P (SUBREG_REG (x))
3806 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3807 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3808 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3810 return 0;
3813 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3814 the hash table in case its value was seen before. */
3816 if (MEM_P (x))
3818 struct table_elt *elt;
3820 x = avoid_constant_pool_reference (x);
3821 if (CONSTANT_P (x))
3822 return x;
3824 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3825 if (elt == 0)
3826 return 0;
3828 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3829 if (elt->is_const && CONSTANT_P (elt->exp))
3830 return elt->exp;
3833 return 0;
3836 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3837 "taken" branch.
3839 In certain cases, this can cause us to add an equivalence. For example,
3840 if we are following the taken case of
3841 if (i == 2)
3842 we can add the fact that `i' and '2' are now equivalent.
3844 In any case, we can record that this comparison was passed. If the same
3845 comparison is seen later, we will know its value. */
3847 static void
3848 record_jump_equiv (rtx insn, bool taken)
3850 int cond_known_true;
3851 rtx op0, op1;
3852 rtx set;
3853 enum machine_mode mode, mode0, mode1;
3854 int reversed_nonequality = 0;
3855 enum rtx_code code;
3857 /* Ensure this is the right kind of insn. */
3858 gcc_assert (any_condjump_p (insn));
3860 set = pc_set (insn);
3862 /* See if this jump condition is known true or false. */
3863 if (taken)
3864 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3865 else
3866 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3868 /* Get the type of comparison being done and the operands being compared.
3869 If we had to reverse a non-equality condition, record that fact so we
3870 know that it isn't valid for floating-point. */
3871 code = GET_CODE (XEXP (SET_SRC (set), 0));
3872 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3873 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3875 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3876 if (! cond_known_true)
3878 code = reversed_comparison_code_parts (code, op0, op1, insn);
3880 /* Don't remember if we can't find the inverse. */
3881 if (code == UNKNOWN)
3882 return;
3885 /* The mode is the mode of the non-constant. */
3886 mode = mode0;
3887 if (mode1 != VOIDmode)
3888 mode = mode1;
3890 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3893 /* Yet another form of subreg creation. In this case, we want something in
3894 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3896 static rtx
3897 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3899 enum machine_mode op_mode = GET_MODE (op);
3900 if (op_mode == mode || op_mode == VOIDmode)
3901 return op;
3902 return lowpart_subreg (mode, op, op_mode);
3905 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3906 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3907 Make any useful entries we can with that information. Called from
3908 above function and called recursively. */
3910 static void
3911 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3912 rtx op1, int reversed_nonequality)
3914 unsigned op0_hash, op1_hash;
3915 int op0_in_memory, op1_in_memory;
3916 struct table_elt *op0_elt, *op1_elt;
3918 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3919 we know that they are also equal in the smaller mode (this is also
3920 true for all smaller modes whether or not there is a SUBREG, but
3921 is not worth testing for with no SUBREG). */
3923 /* Note that GET_MODE (op0) may not equal MODE. */
3924 if (code == EQ && paradoxical_subreg_p (op0))
3926 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3927 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3928 if (tem)
3929 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3930 reversed_nonequality);
3933 if (code == EQ && paradoxical_subreg_p (op1))
3935 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3936 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3937 if (tem)
3938 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3939 reversed_nonequality);
3942 /* Similarly, if this is an NE comparison, and either is a SUBREG
3943 making a smaller mode, we know the whole thing is also NE. */
3945 /* Note that GET_MODE (op0) may not equal MODE;
3946 if we test MODE instead, we can get an infinite recursion
3947 alternating between two modes each wider than MODE. */
3949 if (code == NE && GET_CODE (op0) == SUBREG
3950 && subreg_lowpart_p (op0)
3951 && (GET_MODE_SIZE (GET_MODE (op0))
3952 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3954 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3955 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3956 if (tem)
3957 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3958 reversed_nonequality);
3961 if (code == NE && GET_CODE (op1) == SUBREG
3962 && subreg_lowpart_p (op1)
3963 && (GET_MODE_SIZE (GET_MODE (op1))
3964 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3966 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3967 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3968 if (tem)
3969 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3970 reversed_nonequality);
3973 /* Hash both operands. */
3975 do_not_record = 0;
3976 hash_arg_in_memory = 0;
3977 op0_hash = HASH (op0, mode);
3978 op0_in_memory = hash_arg_in_memory;
3980 if (do_not_record)
3981 return;
3983 do_not_record = 0;
3984 hash_arg_in_memory = 0;
3985 op1_hash = HASH (op1, mode);
3986 op1_in_memory = hash_arg_in_memory;
3988 if (do_not_record)
3989 return;
3991 /* Look up both operands. */
3992 op0_elt = lookup (op0, op0_hash, mode);
3993 op1_elt = lookup (op1, op1_hash, mode);
3995 /* If both operands are already equivalent or if they are not in the
3996 table but are identical, do nothing. */
3997 if ((op0_elt != 0 && op1_elt != 0
3998 && op0_elt->first_same_value == op1_elt->first_same_value)
3999 || op0 == op1 || rtx_equal_p (op0, op1))
4000 return;
4002 /* If we aren't setting two things equal all we can do is save this
4003 comparison. Similarly if this is floating-point. In the latter
4004 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4005 If we record the equality, we might inadvertently delete code
4006 whose intent was to change -0 to +0. */
4008 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4010 struct qty_table_elem *ent;
4011 int qty;
4013 /* If we reversed a floating-point comparison, if OP0 is not a
4014 register, or if OP1 is neither a register or constant, we can't
4015 do anything. */
4017 if (!REG_P (op1))
4018 op1 = equiv_constant (op1);
4020 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4021 || !REG_P (op0) || op1 == 0)
4022 return;
4024 /* Put OP0 in the hash table if it isn't already. This gives it a
4025 new quantity number. */
4026 if (op0_elt == 0)
4028 if (insert_regs (op0, NULL, 0))
4030 rehash_using_reg (op0);
4031 op0_hash = HASH (op0, mode);
4033 /* If OP0 is contained in OP1, this changes its hash code
4034 as well. Faster to rehash than to check, except
4035 for the simple case of a constant. */
4036 if (! CONSTANT_P (op1))
4037 op1_hash = HASH (op1,mode);
4040 op0_elt = insert (op0, NULL, op0_hash, mode);
4041 op0_elt->in_memory = op0_in_memory;
4044 qty = REG_QTY (REGNO (op0));
4045 ent = &qty_table[qty];
4047 ent->comparison_code = code;
4048 if (REG_P (op1))
4050 /* Look it up again--in case op0 and op1 are the same. */
4051 op1_elt = lookup (op1, op1_hash, mode);
4053 /* Put OP1 in the hash table so it gets a new quantity number. */
4054 if (op1_elt == 0)
4056 if (insert_regs (op1, NULL, 0))
4058 rehash_using_reg (op1);
4059 op1_hash = HASH (op1, mode);
4062 op1_elt = insert (op1, NULL, op1_hash, mode);
4063 op1_elt->in_memory = op1_in_memory;
4066 ent->comparison_const = NULL_RTX;
4067 ent->comparison_qty = REG_QTY (REGNO (op1));
4069 else
4071 ent->comparison_const = op1;
4072 ent->comparison_qty = -1;
4075 return;
4078 /* If either side is still missing an equivalence, make it now,
4079 then merge the equivalences. */
4081 if (op0_elt == 0)
4083 if (insert_regs (op0, NULL, 0))
4085 rehash_using_reg (op0);
4086 op0_hash = HASH (op0, mode);
4089 op0_elt = insert (op0, NULL, op0_hash, mode);
4090 op0_elt->in_memory = op0_in_memory;
4093 if (op1_elt == 0)
4095 if (insert_regs (op1, NULL, 0))
4097 rehash_using_reg (op1);
4098 op1_hash = HASH (op1, mode);
4101 op1_elt = insert (op1, NULL, op1_hash, mode);
4102 op1_elt->in_memory = op1_in_memory;
4105 merge_equiv_classes (op0_elt, op1_elt);
4108 /* CSE processing for one instruction.
4110 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4111 but the few that "leak through" are cleaned up by cse_insn, and complex
4112 addressing modes are often formed here.
4114 The main function is cse_insn, and between here and that function
4115 a couple of helper functions is defined to keep the size of cse_insn
4116 within reasonable proportions.
4118 Data is shared between the main and helper functions via STRUCT SET,
4119 that contains all data related for every set in the instruction that
4120 is being processed.
4122 Note that cse_main processes all sets in the instruction. Most
4123 passes in GCC only process simple SET insns or single_set insns, but
4124 CSE processes insns with multiple sets as well. */
4126 /* Data on one SET contained in the instruction. */
4128 struct set
4130 /* The SET rtx itself. */
4131 rtx rtl;
4132 /* The SET_SRC of the rtx (the original value, if it is changing). */
4133 rtx src;
4134 /* The hash-table element for the SET_SRC of the SET. */
4135 struct table_elt *src_elt;
4136 /* Hash value for the SET_SRC. */
4137 unsigned src_hash;
4138 /* Hash value for the SET_DEST. */
4139 unsigned dest_hash;
4140 /* The SET_DEST, with SUBREG, etc., stripped. */
4141 rtx inner_dest;
4142 /* Nonzero if the SET_SRC is in memory. */
4143 char src_in_memory;
4144 /* Nonzero if the SET_SRC contains something
4145 whose value cannot be predicted and understood. */
4146 char src_volatile;
4147 /* Original machine mode, in case it becomes a CONST_INT.
4148 The size of this field should match the size of the mode
4149 field of struct rtx_def (see rtl.h). */
4150 ENUM_BITFIELD(machine_mode) mode : 8;
4151 /* A constant equivalent for SET_SRC, if any. */
4152 rtx src_const;
4153 /* Hash value of constant equivalent for SET_SRC. */
4154 unsigned src_const_hash;
4155 /* Table entry for constant equivalent for SET_SRC, if any. */
4156 struct table_elt *src_const_elt;
4157 /* Table entry for the destination address. */
4158 struct table_elt *dest_addr_elt;
4161 /* Special handling for (set REG0 REG1) where REG0 is the
4162 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4163 be used in the sequel, so (if easily done) change this insn to
4164 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4165 that computed their value. Then REG1 will become a dead store
4166 and won't cloud the situation for later optimizations.
4168 Do not make this change if REG1 is a hard register, because it will
4169 then be used in the sequel and we may be changing a two-operand insn
4170 into a three-operand insn.
4172 This is the last transformation that cse_insn will try to do. */
4174 static void
4175 try_back_substitute_reg (rtx set, rtx insn)
4177 rtx dest = SET_DEST (set);
4178 rtx src = SET_SRC (set);
4180 if (REG_P (dest)
4181 && REG_P (src) && ! HARD_REGISTER_P (src)
4182 && REGNO_QTY_VALID_P (REGNO (src)))
4184 int src_q = REG_QTY (REGNO (src));
4185 struct qty_table_elem *src_ent = &qty_table[src_q];
4187 if (src_ent->first_reg == REGNO (dest))
4189 /* Scan for the previous nonnote insn, but stop at a basic
4190 block boundary. */
4191 rtx prev = insn;
4192 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4195 prev = PREV_INSN (prev);
4197 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4199 /* Do not swap the registers around if the previous instruction
4200 attaches a REG_EQUIV note to REG1.
4202 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4203 from the pseudo that originally shadowed an incoming argument
4204 to another register. Some uses of REG_EQUIV might rely on it
4205 being attached to REG1 rather than REG2.
4207 This section previously turned the REG_EQUIV into a REG_EQUAL
4208 note. We cannot do that because REG_EQUIV may provide an
4209 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4210 if (NONJUMP_INSN_P (prev)
4211 && GET_CODE (PATTERN (prev)) == SET
4212 && SET_DEST (PATTERN (prev)) == src
4213 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4215 rtx note;
4217 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4218 validate_change (insn, &SET_DEST (set), src, 1);
4219 validate_change (insn, &SET_SRC (set), dest, 1);
4220 apply_change_group ();
4222 /* If INSN has a REG_EQUAL note, and this note mentions
4223 REG0, then we must delete it, because the value in
4224 REG0 has changed. If the note's value is REG1, we must
4225 also delete it because that is now this insn's dest. */
4226 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4227 if (note != 0
4228 && (reg_mentioned_p (dest, XEXP (note, 0))
4229 || rtx_equal_p (src, XEXP (note, 0))))
4230 remove_note (insn, note);
4236 /* Record all the SETs in this instruction into SETS_PTR,
4237 and return the number of recorded sets. */
4238 static int
4239 find_sets_in_insn (rtx insn, struct set **psets)
4241 struct set *sets = *psets;
4242 int n_sets = 0;
4243 rtx x = PATTERN (insn);
4245 if (GET_CODE (x) == SET)
4247 /* Ignore SETs that are unconditional jumps.
4248 They never need cse processing, so this does not hurt.
4249 The reason is not efficiency but rather
4250 so that we can test at the end for instructions
4251 that have been simplified to unconditional jumps
4252 and not be misled by unchanged instructions
4253 that were unconditional jumps to begin with. */
4254 if (SET_DEST (x) == pc_rtx
4255 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4257 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4258 The hard function value register is used only once, to copy to
4259 someplace else, so it isn't worth cse'ing. */
4260 else if (GET_CODE (SET_SRC (x)) == CALL)
4262 else
4263 sets[n_sets++].rtl = x;
4265 else if (GET_CODE (x) == PARALLEL)
4267 int i, lim = XVECLEN (x, 0);
4269 /* Go over the epressions of the PARALLEL in forward order, to
4270 put them in the same order in the SETS array. */
4271 for (i = 0; i < lim; i++)
4273 rtx y = XVECEXP (x, 0, i);
4274 if (GET_CODE (y) == SET)
4276 /* As above, we ignore unconditional jumps and call-insns and
4277 ignore the result of apply_change_group. */
4278 if (SET_DEST (y) == pc_rtx
4279 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4281 else if (GET_CODE (SET_SRC (y)) == CALL)
4283 else
4284 sets[n_sets++].rtl = y;
4289 return n_sets;
4292 /* Where possible, substitute every register reference in the N_SETS
4293 number of SETS in INSN with the the canonical register.
4295 Register canonicalization propagatest the earliest register (i.e.
4296 one that is set before INSN) with the same value. This is a very
4297 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4298 to RTL. For instance, a CONST for an address is usually expanded
4299 multiple times to loads into different registers, thus creating many
4300 subexpressions of the form:
4302 (set (reg1) (some_const))
4303 (set (mem (... reg1 ...) (thing)))
4304 (set (reg2) (some_const))
4305 (set (mem (... reg2 ...) (thing)))
4307 After canonicalizing, the code takes the following form:
4309 (set (reg1) (some_const))
4310 (set (mem (... reg1 ...) (thing)))
4311 (set (reg2) (some_const))
4312 (set (mem (... reg1 ...) (thing)))
4314 The set to reg2 is now trivially dead, and the memory reference (or
4315 address, or whatever) may be a candidate for further CSEing.
4317 In this function, the result of apply_change_group can be ignored;
4318 see canon_reg. */
4320 static void
4321 canonicalize_insn (rtx insn, struct set **psets, int n_sets)
4323 struct set *sets = *psets;
4324 rtx tem;
4325 rtx x = PATTERN (insn);
4326 int i;
4328 if (CALL_P (insn))
4330 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4331 if (GET_CODE (XEXP (tem, 0)) != SET)
4332 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4335 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4337 canon_reg (SET_SRC (x), insn);
4338 apply_change_group ();
4339 fold_rtx (SET_SRC (x), insn);
4341 else if (GET_CODE (x) == CLOBBER)
4343 /* If we clobber memory, canon the address.
4344 This does nothing when a register is clobbered
4345 because we have already invalidated the reg. */
4346 if (MEM_P (XEXP (x, 0)))
4347 canon_reg (XEXP (x, 0), insn);
4349 else if (GET_CODE (x) == USE
4350 && ! (REG_P (XEXP (x, 0))
4351 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4352 /* Canonicalize a USE of a pseudo register or memory location. */
4353 canon_reg (x, insn);
4354 else if (GET_CODE (x) == ASM_OPERANDS)
4356 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4358 rtx input = ASM_OPERANDS_INPUT (x, i);
4359 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4361 input = canon_reg (input, insn);
4362 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4366 else if (GET_CODE (x) == CALL)
4368 canon_reg (x, insn);
4369 apply_change_group ();
4370 fold_rtx (x, insn);
4372 else if (DEBUG_INSN_P (insn))
4373 canon_reg (PATTERN (insn), insn);
4374 else if (GET_CODE (x) == PARALLEL)
4376 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4378 rtx y = XVECEXP (x, 0, i);
4379 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4381 canon_reg (SET_SRC (y), insn);
4382 apply_change_group ();
4383 fold_rtx (SET_SRC (y), insn);
4385 else if (GET_CODE (y) == CLOBBER)
4387 if (MEM_P (XEXP (y, 0)))
4388 canon_reg (XEXP (y, 0), insn);
4390 else if (GET_CODE (y) == USE
4391 && ! (REG_P (XEXP (y, 0))
4392 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4393 canon_reg (y, insn);
4394 else if (GET_CODE (y) == CALL)
4396 canon_reg (y, insn);
4397 apply_change_group ();
4398 fold_rtx (y, insn);
4403 if (n_sets == 1 && REG_NOTES (insn) != 0
4404 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4406 /* We potentially will process this insn many times. Therefore,
4407 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4408 unique set in INSN.
4410 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4411 because cse_insn handles those specially. */
4412 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4413 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4414 remove_note (insn, tem);
4415 else
4417 canon_reg (XEXP (tem, 0), insn);
4418 apply_change_group ();
4419 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4420 df_notes_rescan (insn);
4424 /* Canonicalize sources and addresses of destinations.
4425 We do this in a separate pass to avoid problems when a MATCH_DUP is
4426 present in the insn pattern. In that case, we want to ensure that
4427 we don't break the duplicate nature of the pattern. So we will replace
4428 both operands at the same time. Otherwise, we would fail to find an
4429 equivalent substitution in the loop calling validate_change below.
4431 We used to suppress canonicalization of DEST if it appears in SRC,
4432 but we don't do this any more. */
4434 for (i = 0; i < n_sets; i++)
4436 rtx dest = SET_DEST (sets[i].rtl);
4437 rtx src = SET_SRC (sets[i].rtl);
4438 rtx new_rtx = canon_reg (src, insn);
4440 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4442 if (GET_CODE (dest) == ZERO_EXTRACT)
4444 validate_change (insn, &XEXP (dest, 1),
4445 canon_reg (XEXP (dest, 1), insn), 1);
4446 validate_change (insn, &XEXP (dest, 2),
4447 canon_reg (XEXP (dest, 2), insn), 1);
4450 while (GET_CODE (dest) == SUBREG
4451 || GET_CODE (dest) == ZERO_EXTRACT
4452 || GET_CODE (dest) == STRICT_LOW_PART)
4453 dest = XEXP (dest, 0);
4455 if (MEM_P (dest))
4456 canon_reg (dest, insn);
4459 /* Now that we have done all the replacements, we can apply the change
4460 group and see if they all work. Note that this will cause some
4461 canonicalizations that would have worked individually not to be applied
4462 because some other canonicalization didn't work, but this should not
4463 occur often.
4465 The result of apply_change_group can be ignored; see canon_reg. */
4467 apply_change_group ();
4470 /* Main function of CSE.
4471 First simplify sources and addresses of all assignments
4472 in the instruction, using previously-computed equivalents values.
4473 Then install the new sources and destinations in the table
4474 of available values. */
4476 static void
4477 cse_insn (rtx insn)
4479 rtx x = PATTERN (insn);
4480 int i;
4481 rtx tem;
4482 int n_sets = 0;
4484 rtx src_eqv = 0;
4485 struct table_elt *src_eqv_elt = 0;
4486 int src_eqv_volatile = 0;
4487 int src_eqv_in_memory = 0;
4488 unsigned src_eqv_hash = 0;
4490 struct set *sets = (struct set *) 0;
4492 if (GET_CODE (x) == SET)
4493 sets = XALLOCA (struct set);
4494 else if (GET_CODE (x) == PARALLEL)
4495 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4497 this_insn = insn;
4498 #ifdef HAVE_cc0
4499 /* Records what this insn does to set CC0. */
4500 this_insn_cc0 = 0;
4501 this_insn_cc0_mode = VOIDmode;
4502 #endif
4504 /* Find all regs explicitly clobbered in this insn,
4505 to ensure they are not replaced with any other regs
4506 elsewhere in this insn. */
4507 invalidate_from_sets_and_clobbers (insn);
4509 /* Record all the SETs in this instruction. */
4510 n_sets = find_sets_in_insn (insn, &sets);
4512 /* Substitute the canonical register where possible. */
4513 canonicalize_insn (insn, &sets, n_sets);
4515 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4516 if different, or if the DEST is a STRICT_LOW_PART. The latter condition
4517 is necessary because SRC_EQV is handled specially for this case, and if
4518 it isn't set, then there will be no equivalence for the destination. */
4519 if (n_sets == 1 && REG_NOTES (insn) != 0
4520 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4521 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4522 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4523 src_eqv = copy_rtx (XEXP (tem, 0));
4525 /* Set sets[i].src_elt to the class each source belongs to.
4526 Detect assignments from or to volatile things
4527 and set set[i] to zero so they will be ignored
4528 in the rest of this function.
4530 Nothing in this loop changes the hash table or the register chains. */
4532 for (i = 0; i < n_sets; i++)
4534 bool repeat = false;
4535 rtx src, dest;
4536 rtx src_folded;
4537 struct table_elt *elt = 0, *p;
4538 enum machine_mode mode;
4539 rtx src_eqv_here;
4540 rtx src_const = 0;
4541 rtx src_related = 0;
4542 bool src_related_is_const_anchor = false;
4543 struct table_elt *src_const_elt = 0;
4544 int src_cost = MAX_COST;
4545 int src_eqv_cost = MAX_COST;
4546 int src_folded_cost = MAX_COST;
4547 int src_related_cost = MAX_COST;
4548 int src_elt_cost = MAX_COST;
4549 int src_regcost = MAX_COST;
4550 int src_eqv_regcost = MAX_COST;
4551 int src_folded_regcost = MAX_COST;
4552 int src_related_regcost = MAX_COST;
4553 int src_elt_regcost = MAX_COST;
4554 /* Set nonzero if we need to call force_const_mem on with the
4555 contents of src_folded before using it. */
4556 int src_folded_force_flag = 0;
4558 dest = SET_DEST (sets[i].rtl);
4559 src = SET_SRC (sets[i].rtl);
4561 /* If SRC is a constant that has no machine mode,
4562 hash it with the destination's machine mode.
4563 This way we can keep different modes separate. */
4565 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4566 sets[i].mode = mode;
4568 if (src_eqv)
4570 enum machine_mode eqvmode = mode;
4571 if (GET_CODE (dest) == STRICT_LOW_PART)
4572 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4573 do_not_record = 0;
4574 hash_arg_in_memory = 0;
4575 src_eqv_hash = HASH (src_eqv, eqvmode);
4577 /* Find the equivalence class for the equivalent expression. */
4579 if (!do_not_record)
4580 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4582 src_eqv_volatile = do_not_record;
4583 src_eqv_in_memory = hash_arg_in_memory;
4586 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4587 value of the INNER register, not the destination. So it is not
4588 a valid substitution for the source. But save it for later. */
4589 if (GET_CODE (dest) == STRICT_LOW_PART)
4590 src_eqv_here = 0;
4591 else
4592 src_eqv_here = src_eqv;
4594 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4595 simplified result, which may not necessarily be valid. */
4596 src_folded = fold_rtx (src, insn);
4598 #if 0
4599 /* ??? This caused bad code to be generated for the m68k port with -O2.
4600 Suppose src is (CONST_INT -1), and that after truncation src_folded
4601 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4602 At the end we will add src and src_const to the same equivalence
4603 class. We now have 3 and -1 on the same equivalence class. This
4604 causes later instructions to be mis-optimized. */
4605 /* If storing a constant in a bitfield, pre-truncate the constant
4606 so we will be able to record it later. */
4607 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4609 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4611 if (CONST_INT_P (src)
4612 && CONST_INT_P (width)
4613 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4614 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4615 src_folded
4616 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4617 << INTVAL (width)) - 1));
4619 #endif
4621 /* Compute SRC's hash code, and also notice if it
4622 should not be recorded at all. In that case,
4623 prevent any further processing of this assignment. */
4624 do_not_record = 0;
4625 hash_arg_in_memory = 0;
4627 sets[i].src = src;
4628 sets[i].src_hash = HASH (src, mode);
4629 sets[i].src_volatile = do_not_record;
4630 sets[i].src_in_memory = hash_arg_in_memory;
4632 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4633 a pseudo, do not record SRC. Using SRC as a replacement for
4634 anything else will be incorrect in that situation. Note that
4635 this usually occurs only for stack slots, in which case all the
4636 RTL would be referring to SRC, so we don't lose any optimization
4637 opportunities by not having SRC in the hash table. */
4639 if (MEM_P (src)
4640 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4641 && REG_P (dest)
4642 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4643 sets[i].src_volatile = 1;
4645 #if 0
4646 /* It is no longer clear why we used to do this, but it doesn't
4647 appear to still be needed. So let's try without it since this
4648 code hurts cse'ing widened ops. */
4649 /* If source is a paradoxical subreg (such as QI treated as an SI),
4650 treat it as volatile. It may do the work of an SI in one context
4651 where the extra bits are not being used, but cannot replace an SI
4652 in general. */
4653 if (paradoxical_subreg_p (src))
4654 sets[i].src_volatile = 1;
4655 #endif
4657 /* Locate all possible equivalent forms for SRC. Try to replace
4658 SRC in the insn with each cheaper equivalent.
4660 We have the following types of equivalents: SRC itself, a folded
4661 version, a value given in a REG_EQUAL note, or a value related
4662 to a constant.
4664 Each of these equivalents may be part of an additional class
4665 of equivalents (if more than one is in the table, they must be in
4666 the same class; we check for this).
4668 If the source is volatile, we don't do any table lookups.
4670 We note any constant equivalent for possible later use in a
4671 REG_NOTE. */
4673 if (!sets[i].src_volatile)
4674 elt = lookup (src, sets[i].src_hash, mode);
4676 sets[i].src_elt = elt;
4678 if (elt && src_eqv_here && src_eqv_elt)
4680 if (elt->first_same_value != src_eqv_elt->first_same_value)
4682 /* The REG_EQUAL is indicating that two formerly distinct
4683 classes are now equivalent. So merge them. */
4684 merge_equiv_classes (elt, src_eqv_elt);
4685 src_eqv_hash = HASH (src_eqv, elt->mode);
4686 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4689 src_eqv_here = 0;
4692 else if (src_eqv_elt)
4693 elt = src_eqv_elt;
4695 /* Try to find a constant somewhere and record it in `src_const'.
4696 Record its table element, if any, in `src_const_elt'. Look in
4697 any known equivalences first. (If the constant is not in the
4698 table, also set `sets[i].src_const_hash'). */
4699 if (elt)
4700 for (p = elt->first_same_value; p; p = p->next_same_value)
4701 if (p->is_const)
4703 src_const = p->exp;
4704 src_const_elt = elt;
4705 break;
4708 if (src_const == 0
4709 && (CONSTANT_P (src_folded)
4710 /* Consider (minus (label_ref L1) (label_ref L2)) as
4711 "constant" here so we will record it. This allows us
4712 to fold switch statements when an ADDR_DIFF_VEC is used. */
4713 || (GET_CODE (src_folded) == MINUS
4714 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4715 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4716 src_const = src_folded, src_const_elt = elt;
4717 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4718 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4720 /* If we don't know if the constant is in the table, get its
4721 hash code and look it up. */
4722 if (src_const && src_const_elt == 0)
4724 sets[i].src_const_hash = HASH (src_const, mode);
4725 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4728 sets[i].src_const = src_const;
4729 sets[i].src_const_elt = src_const_elt;
4731 /* If the constant and our source are both in the table, mark them as
4732 equivalent. Otherwise, if a constant is in the table but the source
4733 isn't, set ELT to it. */
4734 if (src_const_elt && elt
4735 && src_const_elt->first_same_value != elt->first_same_value)
4736 merge_equiv_classes (elt, src_const_elt);
4737 else if (src_const_elt && elt == 0)
4738 elt = src_const_elt;
4740 /* See if there is a register linearly related to a constant
4741 equivalent of SRC. */
4742 if (src_const
4743 && (GET_CODE (src_const) == CONST
4744 || (src_const_elt && src_const_elt->related_value != 0)))
4746 src_related = use_related_value (src_const, src_const_elt);
4747 if (src_related)
4749 struct table_elt *src_related_elt
4750 = lookup (src_related, HASH (src_related, mode), mode);
4751 if (src_related_elt && elt)
4753 if (elt->first_same_value
4754 != src_related_elt->first_same_value)
4755 /* This can occur when we previously saw a CONST
4756 involving a SYMBOL_REF and then see the SYMBOL_REF
4757 twice. Merge the involved classes. */
4758 merge_equiv_classes (elt, src_related_elt);
4760 src_related = 0;
4761 src_related_elt = 0;
4763 else if (src_related_elt && elt == 0)
4764 elt = src_related_elt;
4768 /* See if we have a CONST_INT that is already in a register in a
4769 wider mode. */
4771 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4772 && GET_MODE_CLASS (mode) == MODE_INT
4773 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4775 enum machine_mode wider_mode;
4777 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4778 wider_mode != VOIDmode
4779 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4780 && src_related == 0;
4781 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4783 struct table_elt *const_elt
4784 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4786 if (const_elt == 0)
4787 continue;
4789 for (const_elt = const_elt->first_same_value;
4790 const_elt; const_elt = const_elt->next_same_value)
4791 if (REG_P (const_elt->exp))
4793 src_related = gen_lowpart (mode, const_elt->exp);
4794 break;
4799 /* Another possibility is that we have an AND with a constant in
4800 a mode narrower than a word. If so, it might have been generated
4801 as part of an "if" which would narrow the AND. If we already
4802 have done the AND in a wider mode, we can use a SUBREG of that
4803 value. */
4805 if (flag_expensive_optimizations && ! src_related
4806 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4807 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4809 enum machine_mode tmode;
4810 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4812 for (tmode = GET_MODE_WIDER_MODE (mode);
4813 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4814 tmode = GET_MODE_WIDER_MODE (tmode))
4816 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4817 struct table_elt *larger_elt;
4819 if (inner)
4821 PUT_MODE (new_and, tmode);
4822 XEXP (new_and, 0) = inner;
4823 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4824 if (larger_elt == 0)
4825 continue;
4827 for (larger_elt = larger_elt->first_same_value;
4828 larger_elt; larger_elt = larger_elt->next_same_value)
4829 if (REG_P (larger_elt->exp))
4831 src_related
4832 = gen_lowpart (mode, larger_elt->exp);
4833 break;
4836 if (src_related)
4837 break;
4842 #ifdef LOAD_EXTEND_OP
4843 /* See if a MEM has already been loaded with a widening operation;
4844 if it has, we can use a subreg of that. Many CISC machines
4845 also have such operations, but this is only likely to be
4846 beneficial on these machines. */
4848 if (flag_expensive_optimizations && src_related == 0
4849 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4850 && GET_MODE_CLASS (mode) == MODE_INT
4851 && MEM_P (src) && ! do_not_record
4852 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4854 struct rtx_def memory_extend_buf;
4855 rtx memory_extend_rtx = &memory_extend_buf;
4856 enum machine_mode tmode;
4858 /* Set what we are trying to extend and the operation it might
4859 have been extended with. */
4860 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4861 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4862 XEXP (memory_extend_rtx, 0) = src;
4864 for (tmode = GET_MODE_WIDER_MODE (mode);
4865 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4866 tmode = GET_MODE_WIDER_MODE (tmode))
4868 struct table_elt *larger_elt;
4870 PUT_MODE (memory_extend_rtx, tmode);
4871 larger_elt = lookup (memory_extend_rtx,
4872 HASH (memory_extend_rtx, tmode), tmode);
4873 if (larger_elt == 0)
4874 continue;
4876 for (larger_elt = larger_elt->first_same_value;
4877 larger_elt; larger_elt = larger_elt->next_same_value)
4878 if (REG_P (larger_elt->exp))
4880 src_related = gen_lowpart (mode, larger_elt->exp);
4881 break;
4884 if (src_related)
4885 break;
4888 #endif /* LOAD_EXTEND_OP */
4890 /* Try to express the constant using a register+offset expression
4891 derived from a constant anchor. */
4893 if (targetm.const_anchor
4894 && !src_related
4895 && src_const
4896 && GET_CODE (src_const) == CONST_INT)
4898 src_related = try_const_anchors (src_const, mode);
4899 src_related_is_const_anchor = src_related != NULL_RTX;
4903 if (src == src_folded)
4904 src_folded = 0;
4906 /* At this point, ELT, if nonzero, points to a class of expressions
4907 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4908 and SRC_RELATED, if nonzero, each contain additional equivalent
4909 expressions. Prune these latter expressions by deleting expressions
4910 already in the equivalence class.
4912 Check for an equivalent identical to the destination. If found,
4913 this is the preferred equivalent since it will likely lead to
4914 elimination of the insn. Indicate this by placing it in
4915 `src_related'. */
4917 if (elt)
4918 elt = elt->first_same_value;
4919 for (p = elt; p; p = p->next_same_value)
4921 enum rtx_code code = GET_CODE (p->exp);
4923 /* If the expression is not valid, ignore it. Then we do not
4924 have to check for validity below. In most cases, we can use
4925 `rtx_equal_p', since canonicalization has already been done. */
4926 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4927 continue;
4929 /* Also skip paradoxical subregs, unless that's what we're
4930 looking for. */
4931 if (paradoxical_subreg_p (p->exp)
4932 && ! (src != 0
4933 && GET_CODE (src) == SUBREG
4934 && GET_MODE (src) == GET_MODE (p->exp)
4935 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4936 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4937 continue;
4939 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4940 src = 0;
4941 else if (src_folded && GET_CODE (src_folded) == code
4942 && rtx_equal_p (src_folded, p->exp))
4943 src_folded = 0;
4944 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4945 && rtx_equal_p (src_eqv_here, p->exp))
4946 src_eqv_here = 0;
4947 else if (src_related && GET_CODE (src_related) == code
4948 && rtx_equal_p (src_related, p->exp))
4949 src_related = 0;
4951 /* This is the same as the destination of the insns, we want
4952 to prefer it. Copy it to src_related. The code below will
4953 then give it a negative cost. */
4954 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4955 src_related = dest;
4958 /* Find the cheapest valid equivalent, trying all the available
4959 possibilities. Prefer items not in the hash table to ones
4960 that are when they are equal cost. Note that we can never
4961 worsen an insn as the current contents will also succeed.
4962 If we find an equivalent identical to the destination, use it as best,
4963 since this insn will probably be eliminated in that case. */
4964 if (src)
4966 if (rtx_equal_p (src, dest))
4967 src_cost = src_regcost = -1;
4968 else
4970 src_cost = COST (src);
4971 src_regcost = approx_reg_cost (src);
4975 if (src_eqv_here)
4977 if (rtx_equal_p (src_eqv_here, dest))
4978 src_eqv_cost = src_eqv_regcost = -1;
4979 else
4981 src_eqv_cost = COST (src_eqv_here);
4982 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4986 if (src_folded)
4988 if (rtx_equal_p (src_folded, dest))
4989 src_folded_cost = src_folded_regcost = -1;
4990 else
4992 src_folded_cost = COST (src_folded);
4993 src_folded_regcost = approx_reg_cost (src_folded);
4997 if (src_related)
4999 if (rtx_equal_p (src_related, dest))
5000 src_related_cost = src_related_regcost = -1;
5001 else
5003 src_related_cost = COST (src_related);
5004 src_related_regcost = approx_reg_cost (src_related);
5006 /* If a const-anchor is used to synthesize a constant that
5007 normally requires multiple instructions then slightly prefer
5008 it over the original sequence. These instructions are likely
5009 to become redundant now. We can't compare against the cost
5010 of src_eqv_here because, on MIPS for example, multi-insn
5011 constants have zero cost; they are assumed to be hoisted from
5012 loops. */
5013 if (src_related_is_const_anchor
5014 && src_related_cost == src_cost
5015 && src_eqv_here)
5016 src_related_cost--;
5020 /* If this was an indirect jump insn, a known label will really be
5021 cheaper even though it looks more expensive. */
5022 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5023 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5025 /* Terminate loop when replacement made. This must terminate since
5026 the current contents will be tested and will always be valid. */
5027 while (1)
5029 rtx trial;
5031 /* Skip invalid entries. */
5032 while (elt && !REG_P (elt->exp)
5033 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5034 elt = elt->next_same_value;
5036 /* A paradoxical subreg would be bad here: it'll be the right
5037 size, but later may be adjusted so that the upper bits aren't
5038 what we want. So reject it. */
5039 if (elt != 0
5040 && paradoxical_subreg_p (elt->exp)
5041 /* It is okay, though, if the rtx we're trying to match
5042 will ignore any of the bits we can't predict. */
5043 && ! (src != 0
5044 && GET_CODE (src) == SUBREG
5045 && GET_MODE (src) == GET_MODE (elt->exp)
5046 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5047 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5049 elt = elt->next_same_value;
5050 continue;
5053 if (elt)
5055 src_elt_cost = elt->cost;
5056 src_elt_regcost = elt->regcost;
5059 /* Find cheapest and skip it for the next time. For items
5060 of equal cost, use this order:
5061 src_folded, src, src_eqv, src_related and hash table entry. */
5062 if (src_folded
5063 && preferable (src_folded_cost, src_folded_regcost,
5064 src_cost, src_regcost) <= 0
5065 && preferable (src_folded_cost, src_folded_regcost,
5066 src_eqv_cost, src_eqv_regcost) <= 0
5067 && preferable (src_folded_cost, src_folded_regcost,
5068 src_related_cost, src_related_regcost) <= 0
5069 && preferable (src_folded_cost, src_folded_regcost,
5070 src_elt_cost, src_elt_regcost) <= 0)
5072 trial = src_folded, src_folded_cost = MAX_COST;
5073 if (src_folded_force_flag)
5075 rtx forced = force_const_mem (mode, trial);
5076 if (forced)
5077 trial = forced;
5080 else if (src
5081 && preferable (src_cost, src_regcost,
5082 src_eqv_cost, src_eqv_regcost) <= 0
5083 && preferable (src_cost, src_regcost,
5084 src_related_cost, src_related_regcost) <= 0
5085 && preferable (src_cost, src_regcost,
5086 src_elt_cost, src_elt_regcost) <= 0)
5087 trial = src, src_cost = MAX_COST;
5088 else if (src_eqv_here
5089 && preferable (src_eqv_cost, src_eqv_regcost,
5090 src_related_cost, src_related_regcost) <= 0
5091 && preferable (src_eqv_cost, src_eqv_regcost,
5092 src_elt_cost, src_elt_regcost) <= 0)
5093 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5094 else if (src_related
5095 && preferable (src_related_cost, src_related_regcost,
5096 src_elt_cost, src_elt_regcost) <= 0)
5097 trial = src_related, src_related_cost = MAX_COST;
5098 else
5100 trial = elt->exp;
5101 elt = elt->next_same_value;
5102 src_elt_cost = MAX_COST;
5105 /* Avoid creation of overlapping memory moves. */
5106 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5108 rtx src, dest;
5110 /* BLKmode moves are not handled by cse anyway. */
5111 if (GET_MODE (trial) == BLKmode)
5112 break;
5114 src = canon_rtx (trial);
5115 dest = canon_rtx (SET_DEST (sets[i].rtl));
5117 if (!MEM_P (src) || !MEM_P (dest)
5118 || !nonoverlapping_memrefs_p (src, dest, false))
5119 break;
5122 /* Try to optimize
5123 (set (reg:M N) (const_int A))
5124 (set (reg:M2 O) (const_int B))
5125 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5126 (reg:M2 O)). */
5127 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5128 && CONST_INT_P (trial)
5129 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5130 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5131 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5132 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5133 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5134 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5135 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5136 <= HOST_BITS_PER_WIDE_INT))
5138 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5139 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5140 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5141 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5142 struct table_elt *dest_elt
5143 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5144 rtx dest_cst = NULL;
5146 if (dest_elt)
5147 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5148 if (p->is_const && CONST_INT_P (p->exp))
5150 dest_cst = p->exp;
5151 break;
5153 if (dest_cst)
5155 HOST_WIDE_INT val = INTVAL (dest_cst);
5156 HOST_WIDE_INT mask;
5157 unsigned int shift;
5158 if (BITS_BIG_ENDIAN)
5159 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5160 - INTVAL (pos) - INTVAL (width);
5161 else
5162 shift = INTVAL (pos);
5163 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5164 mask = ~(HOST_WIDE_INT) 0;
5165 else
5166 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5167 val &= ~(mask << shift);
5168 val |= (INTVAL (trial) & mask) << shift;
5169 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5170 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5171 dest_reg, 1);
5172 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5173 GEN_INT (val), 1);
5174 if (apply_change_group ())
5176 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5177 if (note)
5179 remove_note (insn, note);
5180 df_notes_rescan (insn);
5182 src_eqv = NULL_RTX;
5183 src_eqv_elt = NULL;
5184 src_eqv_volatile = 0;
5185 src_eqv_in_memory = 0;
5186 src_eqv_hash = 0;
5187 repeat = true;
5188 break;
5193 /* We don't normally have an insn matching (set (pc) (pc)), so
5194 check for this separately here. We will delete such an
5195 insn below.
5197 For other cases such as a table jump or conditional jump
5198 where we know the ultimate target, go ahead and replace the
5199 operand. While that may not make a valid insn, we will
5200 reemit the jump below (and also insert any necessary
5201 barriers). */
5202 if (n_sets == 1 && dest == pc_rtx
5203 && (trial == pc_rtx
5204 || (GET_CODE (trial) == LABEL_REF
5205 && ! condjump_p (insn))))
5207 /* Don't substitute non-local labels, this confuses CFG. */
5208 if (GET_CODE (trial) == LABEL_REF
5209 && LABEL_REF_NONLOCAL_P (trial))
5210 continue;
5212 SET_SRC (sets[i].rtl) = trial;
5213 cse_jumps_altered = true;
5214 break;
5217 /* Reject certain invalid forms of CONST that we create. */
5218 else if (CONSTANT_P (trial)
5219 && GET_CODE (trial) == CONST
5220 /* Reject cases that will cause decode_rtx_const to
5221 die. On the alpha when simplifying a switch, we
5222 get (const (truncate (minus (label_ref)
5223 (label_ref)))). */
5224 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5225 /* Likewise on IA-64, except without the
5226 truncate. */
5227 || (GET_CODE (XEXP (trial, 0)) == MINUS
5228 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5229 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5230 /* Do nothing for this case. */
5233 /* Look for a substitution that makes a valid insn. */
5234 else if (validate_unshare_change
5235 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5237 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5239 /* The result of apply_change_group can be ignored; see
5240 canon_reg. */
5242 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5243 apply_change_group ();
5245 break;
5248 /* If we previously found constant pool entries for
5249 constants and this is a constant, try making a
5250 pool entry. Put it in src_folded unless we already have done
5251 this since that is where it likely came from. */
5253 else if (constant_pool_entries_cost
5254 && CONSTANT_P (trial)
5255 && (src_folded == 0
5256 || (!MEM_P (src_folded)
5257 && ! src_folded_force_flag))
5258 && GET_MODE_CLASS (mode) != MODE_CC
5259 && mode != VOIDmode)
5261 src_folded_force_flag = 1;
5262 src_folded = trial;
5263 src_folded_cost = constant_pool_entries_cost;
5264 src_folded_regcost = constant_pool_entries_regcost;
5268 /* If we changed the insn too much, handle this set from scratch. */
5269 if (repeat)
5271 i--;
5272 continue;
5275 src = SET_SRC (sets[i].rtl);
5277 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5278 However, there is an important exception: If both are registers
5279 that are not the head of their equivalence class, replace SET_SRC
5280 with the head of the class. If we do not do this, we will have
5281 both registers live over a portion of the basic block. This way,
5282 their lifetimes will likely abut instead of overlapping. */
5283 if (REG_P (dest)
5284 && REGNO_QTY_VALID_P (REGNO (dest)))
5286 int dest_q = REG_QTY (REGNO (dest));
5287 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5289 if (dest_ent->mode == GET_MODE (dest)
5290 && dest_ent->first_reg != REGNO (dest)
5291 && REG_P (src) && REGNO (src) == REGNO (dest)
5292 /* Don't do this if the original insn had a hard reg as
5293 SET_SRC or SET_DEST. */
5294 && (!REG_P (sets[i].src)
5295 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5296 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5297 /* We can't call canon_reg here because it won't do anything if
5298 SRC is a hard register. */
5300 int src_q = REG_QTY (REGNO (src));
5301 struct qty_table_elem *src_ent = &qty_table[src_q];
5302 int first = src_ent->first_reg;
5303 rtx new_src
5304 = (first >= FIRST_PSEUDO_REGISTER
5305 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5307 /* We must use validate-change even for this, because this
5308 might be a special no-op instruction, suitable only to
5309 tag notes onto. */
5310 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5312 src = new_src;
5313 /* If we had a constant that is cheaper than what we are now
5314 setting SRC to, use that constant. We ignored it when we
5315 thought we could make this into a no-op. */
5316 if (src_const && COST (src_const) < COST (src)
5317 && validate_change (insn, &SET_SRC (sets[i].rtl),
5318 src_const, 0))
5319 src = src_const;
5324 /* If we made a change, recompute SRC values. */
5325 if (src != sets[i].src)
5327 do_not_record = 0;
5328 hash_arg_in_memory = 0;
5329 sets[i].src = src;
5330 sets[i].src_hash = HASH (src, mode);
5331 sets[i].src_volatile = do_not_record;
5332 sets[i].src_in_memory = hash_arg_in_memory;
5333 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5336 /* If this is a single SET, we are setting a register, and we have an
5337 equivalent constant, we want to add a REG_EQUAL note if the constant
5338 is different from the source. We don't want to do it for a constant
5339 pseudo since verifying that this pseudo hasn't been eliminated is a
5340 pain; moreover such a note won't help anything.
5342 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5343 which can be created for a reference to a compile time computable
5344 entry in a jump table. */
5345 if (n_sets == 1
5346 && REG_P (dest)
5347 && src_const
5348 && !REG_P (src_const)
5349 && !(GET_CODE (src_const) == SUBREG
5350 && REG_P (SUBREG_REG (src_const)))
5351 && !(GET_CODE (src_const) == CONST
5352 && GET_CODE (XEXP (src_const, 0)) == MINUS
5353 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5354 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5355 && !rtx_equal_p (src, src_const))
5357 /* Make sure that the rtx is not shared. */
5358 src_const = copy_rtx (src_const);
5360 /* Record the actual constant value in a REG_EQUAL note,
5361 making a new one if one does not already exist. */
5362 set_unique_reg_note (insn, REG_EQUAL, src_const);
5363 df_notes_rescan (insn);
5366 /* Now deal with the destination. */
5367 do_not_record = 0;
5369 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5370 while (GET_CODE (dest) == SUBREG
5371 || GET_CODE (dest) == ZERO_EXTRACT
5372 || GET_CODE (dest) == STRICT_LOW_PART)
5373 dest = XEXP (dest, 0);
5375 sets[i].inner_dest = dest;
5377 if (MEM_P (dest))
5379 #ifdef PUSH_ROUNDING
5380 /* Stack pushes invalidate the stack pointer. */
5381 rtx addr = XEXP (dest, 0);
5382 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5383 && XEXP (addr, 0) == stack_pointer_rtx)
5384 invalidate (stack_pointer_rtx, VOIDmode);
5385 #endif
5386 dest = fold_rtx (dest, insn);
5389 /* Compute the hash code of the destination now,
5390 before the effects of this instruction are recorded,
5391 since the register values used in the address computation
5392 are those before this instruction. */
5393 sets[i].dest_hash = HASH (dest, mode);
5395 /* Don't enter a bit-field in the hash table
5396 because the value in it after the store
5397 may not equal what was stored, due to truncation. */
5399 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5401 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5403 if (src_const != 0 && CONST_INT_P (src_const)
5404 && CONST_INT_P (width)
5405 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5406 && ! (INTVAL (src_const)
5407 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5408 /* Exception: if the value is constant,
5409 and it won't be truncated, record it. */
5411 else
5413 /* This is chosen so that the destination will be invalidated
5414 but no new value will be recorded.
5415 We must invalidate because sometimes constant
5416 values can be recorded for bitfields. */
5417 sets[i].src_elt = 0;
5418 sets[i].src_volatile = 1;
5419 src_eqv = 0;
5420 src_eqv_elt = 0;
5424 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5425 the insn. */
5426 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5428 /* One less use of the label this insn used to jump to. */
5429 delete_insn_and_edges (insn);
5430 cse_jumps_altered = true;
5431 /* No more processing for this set. */
5432 sets[i].rtl = 0;
5435 /* If this SET is now setting PC to a label, we know it used to
5436 be a conditional or computed branch. */
5437 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5438 && !LABEL_REF_NONLOCAL_P (src))
5440 /* We reemit the jump in as many cases as possible just in
5441 case the form of an unconditional jump is significantly
5442 different than a computed jump or conditional jump.
5444 If this insn has multiple sets, then reemitting the
5445 jump is nontrivial. So instead we just force rerecognition
5446 and hope for the best. */
5447 if (n_sets == 1)
5449 rtx new_rtx, note;
5451 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5452 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5453 LABEL_NUSES (XEXP (src, 0))++;
5455 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5456 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5457 if (note)
5459 XEXP (note, 1) = NULL_RTX;
5460 REG_NOTES (new_rtx) = note;
5463 delete_insn_and_edges (insn);
5464 insn = new_rtx;
5466 else
5467 INSN_CODE (insn) = -1;
5469 /* Do not bother deleting any unreachable code, let jump do it. */
5470 cse_jumps_altered = true;
5471 sets[i].rtl = 0;
5474 /* If destination is volatile, invalidate it and then do no further
5475 processing for this assignment. */
5477 else if (do_not_record)
5479 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5480 invalidate (dest, VOIDmode);
5481 else if (MEM_P (dest))
5482 invalidate (dest, VOIDmode);
5483 else if (GET_CODE (dest) == STRICT_LOW_PART
5484 || GET_CODE (dest) == ZERO_EXTRACT)
5485 invalidate (XEXP (dest, 0), GET_MODE (dest));
5486 sets[i].rtl = 0;
5489 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5490 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5492 #ifdef HAVE_cc0
5493 /* If setting CC0, record what it was set to, or a constant, if it
5494 is equivalent to a constant. If it is being set to a floating-point
5495 value, make a COMPARE with the appropriate constant of 0. If we
5496 don't do this, later code can interpret this as a test against
5497 const0_rtx, which can cause problems if we try to put it into an
5498 insn as a floating-point operand. */
5499 if (dest == cc0_rtx)
5501 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5502 this_insn_cc0_mode = mode;
5503 if (FLOAT_MODE_P (mode))
5504 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5505 CONST0_RTX (mode));
5507 #endif
5510 /* Now enter all non-volatile source expressions in the hash table
5511 if they are not already present.
5512 Record their equivalence classes in src_elt.
5513 This way we can insert the corresponding destinations into
5514 the same classes even if the actual sources are no longer in them
5515 (having been invalidated). */
5517 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5518 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5520 struct table_elt *elt;
5521 struct table_elt *classp = sets[0].src_elt;
5522 rtx dest = SET_DEST (sets[0].rtl);
5523 enum machine_mode eqvmode = GET_MODE (dest);
5525 if (GET_CODE (dest) == STRICT_LOW_PART)
5527 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5528 classp = 0;
5530 if (insert_regs (src_eqv, classp, 0))
5532 rehash_using_reg (src_eqv);
5533 src_eqv_hash = HASH (src_eqv, eqvmode);
5535 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5536 elt->in_memory = src_eqv_in_memory;
5537 src_eqv_elt = elt;
5539 /* Check to see if src_eqv_elt is the same as a set source which
5540 does not yet have an elt, and if so set the elt of the set source
5541 to src_eqv_elt. */
5542 for (i = 0; i < n_sets; i++)
5543 if (sets[i].rtl && sets[i].src_elt == 0
5544 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5545 sets[i].src_elt = src_eqv_elt;
5548 for (i = 0; i < n_sets; i++)
5549 if (sets[i].rtl && ! sets[i].src_volatile
5550 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5552 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5554 /* REG_EQUAL in setting a STRICT_LOW_PART
5555 gives an equivalent for the entire destination register,
5556 not just for the subreg being stored in now.
5557 This is a more interesting equivalence, so we arrange later
5558 to treat the entire reg as the destination. */
5559 sets[i].src_elt = src_eqv_elt;
5560 sets[i].src_hash = src_eqv_hash;
5562 else
5564 /* Insert source and constant equivalent into hash table, if not
5565 already present. */
5566 struct table_elt *classp = src_eqv_elt;
5567 rtx src = sets[i].src;
5568 rtx dest = SET_DEST (sets[i].rtl);
5569 enum machine_mode mode
5570 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5572 /* It's possible that we have a source value known to be
5573 constant but don't have a REG_EQUAL note on the insn.
5574 Lack of a note will mean src_eqv_elt will be NULL. This
5575 can happen where we've generated a SUBREG to access a
5576 CONST_INT that is already in a register in a wider mode.
5577 Ensure that the source expression is put in the proper
5578 constant class. */
5579 if (!classp)
5580 classp = sets[i].src_const_elt;
5582 if (sets[i].src_elt == 0)
5584 struct table_elt *elt;
5586 /* Note that these insert_regs calls cannot remove
5587 any of the src_elt's, because they would have failed to
5588 match if not still valid. */
5589 if (insert_regs (src, classp, 0))
5591 rehash_using_reg (src);
5592 sets[i].src_hash = HASH (src, mode);
5594 elt = insert (src, classp, sets[i].src_hash, mode);
5595 elt->in_memory = sets[i].src_in_memory;
5596 sets[i].src_elt = classp = elt;
5598 if (sets[i].src_const && sets[i].src_const_elt == 0
5599 && src != sets[i].src_const
5600 && ! rtx_equal_p (sets[i].src_const, src))
5601 sets[i].src_elt = insert (sets[i].src_const, classp,
5602 sets[i].src_const_hash, mode);
5605 else if (sets[i].src_elt == 0)
5606 /* If we did not insert the source into the hash table (e.g., it was
5607 volatile), note the equivalence class for the REG_EQUAL value, if any,
5608 so that the destination goes into that class. */
5609 sets[i].src_elt = src_eqv_elt;
5611 /* Record destination addresses in the hash table. This allows us to
5612 check if they are invalidated by other sets. */
5613 for (i = 0; i < n_sets; i++)
5615 if (sets[i].rtl)
5617 rtx x = sets[i].inner_dest;
5618 struct table_elt *elt;
5619 enum machine_mode mode;
5620 unsigned hash;
5622 if (MEM_P (x))
5624 x = XEXP (x, 0);
5625 mode = GET_MODE (x);
5626 hash = HASH (x, mode);
5627 elt = lookup (x, hash, mode);
5628 if (!elt)
5630 if (insert_regs (x, NULL, 0))
5632 rtx dest = SET_DEST (sets[i].rtl);
5634 rehash_using_reg (x);
5635 hash = HASH (x, mode);
5636 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5638 elt = insert (x, NULL, hash, mode);
5641 sets[i].dest_addr_elt = elt;
5643 else
5644 sets[i].dest_addr_elt = NULL;
5648 invalidate_from_clobbers (insn);
5650 /* Some registers are invalidated by subroutine calls. Memory is
5651 invalidated by non-constant calls. */
5653 if (CALL_P (insn))
5655 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5656 invalidate_memory ();
5657 invalidate_for_call ();
5660 /* Now invalidate everything set by this instruction.
5661 If a SUBREG or other funny destination is being set,
5662 sets[i].rtl is still nonzero, so here we invalidate the reg
5663 a part of which is being set. */
5665 for (i = 0; i < n_sets; i++)
5666 if (sets[i].rtl)
5668 /* We can't use the inner dest, because the mode associated with
5669 a ZERO_EXTRACT is significant. */
5670 rtx dest = SET_DEST (sets[i].rtl);
5672 /* Needed for registers to remove the register from its
5673 previous quantity's chain.
5674 Needed for memory if this is a nonvarying address, unless
5675 we have just done an invalidate_memory that covers even those. */
5676 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5677 invalidate (dest, VOIDmode);
5678 else if (MEM_P (dest))
5679 invalidate (dest, VOIDmode);
5680 else if (GET_CODE (dest) == STRICT_LOW_PART
5681 || GET_CODE (dest) == ZERO_EXTRACT)
5682 invalidate (XEXP (dest, 0), GET_MODE (dest));
5685 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5686 the regs restored by the longjmp come from a later time
5687 than the setjmp. */
5688 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5690 flush_hash_table ();
5691 goto done;
5694 /* Make sure registers mentioned in destinations
5695 are safe for use in an expression to be inserted.
5696 This removes from the hash table
5697 any invalid entry that refers to one of these registers.
5699 We don't care about the return value from mention_regs because
5700 we are going to hash the SET_DEST values unconditionally. */
5702 for (i = 0; i < n_sets; i++)
5704 if (sets[i].rtl)
5706 rtx x = SET_DEST (sets[i].rtl);
5708 if (!REG_P (x))
5709 mention_regs (x);
5710 else
5712 /* We used to rely on all references to a register becoming
5713 inaccessible when a register changes to a new quantity,
5714 since that changes the hash code. However, that is not
5715 safe, since after HASH_SIZE new quantities we get a
5716 hash 'collision' of a register with its own invalid
5717 entries. And since SUBREGs have been changed not to
5718 change their hash code with the hash code of the register,
5719 it wouldn't work any longer at all. So we have to check
5720 for any invalid references lying around now.
5721 This code is similar to the REG case in mention_regs,
5722 but it knows that reg_tick has been incremented, and
5723 it leaves reg_in_table as -1 . */
5724 unsigned int regno = REGNO (x);
5725 unsigned int endregno = END_REGNO (x);
5726 unsigned int i;
5728 for (i = regno; i < endregno; i++)
5730 if (REG_IN_TABLE (i) >= 0)
5732 remove_invalid_refs (i);
5733 REG_IN_TABLE (i) = -1;
5740 /* We may have just removed some of the src_elt's from the hash table.
5741 So replace each one with the current head of the same class.
5742 Also check if destination addresses have been removed. */
5744 for (i = 0; i < n_sets; i++)
5745 if (sets[i].rtl)
5747 if (sets[i].dest_addr_elt
5748 && sets[i].dest_addr_elt->first_same_value == 0)
5750 /* The elt was removed, which means this destination is not
5751 valid after this instruction. */
5752 sets[i].rtl = NULL_RTX;
5754 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5755 /* If elt was removed, find current head of same class,
5756 or 0 if nothing remains of that class. */
5758 struct table_elt *elt = sets[i].src_elt;
5760 while (elt && elt->prev_same_value)
5761 elt = elt->prev_same_value;
5763 while (elt && elt->first_same_value == 0)
5764 elt = elt->next_same_value;
5765 sets[i].src_elt = elt ? elt->first_same_value : 0;
5769 /* Now insert the destinations into their equivalence classes. */
5771 for (i = 0; i < n_sets; i++)
5772 if (sets[i].rtl)
5774 rtx dest = SET_DEST (sets[i].rtl);
5775 struct table_elt *elt;
5777 /* Don't record value if we are not supposed to risk allocating
5778 floating-point values in registers that might be wider than
5779 memory. */
5780 if ((flag_float_store
5781 && MEM_P (dest)
5782 && FLOAT_MODE_P (GET_MODE (dest)))
5783 /* Don't record BLKmode values, because we don't know the
5784 size of it, and can't be sure that other BLKmode values
5785 have the same or smaller size. */
5786 || GET_MODE (dest) == BLKmode
5787 /* If we didn't put a REG_EQUAL value or a source into the hash
5788 table, there is no point is recording DEST. */
5789 || sets[i].src_elt == 0
5790 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5791 or SIGN_EXTEND, don't record DEST since it can cause
5792 some tracking to be wrong.
5794 ??? Think about this more later. */
5795 || (paradoxical_subreg_p (dest)
5796 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5797 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5798 continue;
5800 /* STRICT_LOW_PART isn't part of the value BEING set,
5801 and neither is the SUBREG inside it.
5802 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5803 if (GET_CODE (dest) == STRICT_LOW_PART)
5804 dest = SUBREG_REG (XEXP (dest, 0));
5806 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5807 /* Registers must also be inserted into chains for quantities. */
5808 if (insert_regs (dest, sets[i].src_elt, 1))
5810 /* If `insert_regs' changes something, the hash code must be
5811 recalculated. */
5812 rehash_using_reg (dest);
5813 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5816 elt = insert (dest, sets[i].src_elt,
5817 sets[i].dest_hash, GET_MODE (dest));
5819 /* If this is a constant, insert the constant anchors with the
5820 equivalent register-offset expressions using register DEST. */
5821 if (targetm.const_anchor
5822 && REG_P (dest)
5823 && SCALAR_INT_MODE_P (GET_MODE (dest))
5824 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5825 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5827 elt->in_memory = (MEM_P (sets[i].inner_dest)
5828 && !MEM_READONLY_P (sets[i].inner_dest));
5830 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5831 narrower than M2, and both M1 and M2 are the same number of words,
5832 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5833 make that equivalence as well.
5835 However, BAR may have equivalences for which gen_lowpart
5836 will produce a simpler value than gen_lowpart applied to
5837 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5838 BAR's equivalences. If we don't get a simplified form, make
5839 the SUBREG. It will not be used in an equivalence, but will
5840 cause two similar assignments to be detected.
5842 Note the loop below will find SUBREG_REG (DEST) since we have
5843 already entered SRC and DEST of the SET in the table. */
5845 if (GET_CODE (dest) == SUBREG
5846 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5847 / UNITS_PER_WORD)
5848 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5849 && (GET_MODE_SIZE (GET_MODE (dest))
5850 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5851 && sets[i].src_elt != 0)
5853 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5854 struct table_elt *elt, *classp = 0;
5856 for (elt = sets[i].src_elt->first_same_value; elt;
5857 elt = elt->next_same_value)
5859 rtx new_src = 0;
5860 unsigned src_hash;
5861 struct table_elt *src_elt;
5862 int byte = 0;
5864 /* Ignore invalid entries. */
5865 if (!REG_P (elt->exp)
5866 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5867 continue;
5869 /* We may have already been playing subreg games. If the
5870 mode is already correct for the destination, use it. */
5871 if (GET_MODE (elt->exp) == new_mode)
5872 new_src = elt->exp;
5873 else
5875 /* Calculate big endian correction for the SUBREG_BYTE.
5876 We have already checked that M1 (GET_MODE (dest))
5877 is not narrower than M2 (new_mode). */
5878 if (BYTES_BIG_ENDIAN)
5879 byte = (GET_MODE_SIZE (GET_MODE (dest))
5880 - GET_MODE_SIZE (new_mode));
5882 new_src = simplify_gen_subreg (new_mode, elt->exp,
5883 GET_MODE (dest), byte);
5886 /* The call to simplify_gen_subreg fails if the value
5887 is VOIDmode, yet we can't do any simplification, e.g.
5888 for EXPR_LISTs denoting function call results.
5889 It is invalid to construct a SUBREG with a VOIDmode
5890 SUBREG_REG, hence a zero new_src means we can't do
5891 this substitution. */
5892 if (! new_src)
5893 continue;
5895 src_hash = HASH (new_src, new_mode);
5896 src_elt = lookup (new_src, src_hash, new_mode);
5898 /* Put the new source in the hash table is if isn't
5899 already. */
5900 if (src_elt == 0)
5902 if (insert_regs (new_src, classp, 0))
5904 rehash_using_reg (new_src);
5905 src_hash = HASH (new_src, new_mode);
5907 src_elt = insert (new_src, classp, src_hash, new_mode);
5908 src_elt->in_memory = elt->in_memory;
5910 else if (classp && classp != src_elt->first_same_value)
5911 /* Show that two things that we've seen before are
5912 actually the same. */
5913 merge_equiv_classes (src_elt, classp);
5915 classp = src_elt->first_same_value;
5916 /* Ignore invalid entries. */
5917 while (classp
5918 && !REG_P (classp->exp)
5919 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5920 classp = classp->next_same_value;
5925 /* Special handling for (set REG0 REG1) where REG0 is the
5926 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5927 be used in the sequel, so (if easily done) change this insn to
5928 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5929 that computed their value. Then REG1 will become a dead store
5930 and won't cloud the situation for later optimizations.
5932 Do not make this change if REG1 is a hard register, because it will
5933 then be used in the sequel and we may be changing a two-operand insn
5934 into a three-operand insn.
5936 Also do not do this if we are operating on a copy of INSN. */
5938 if (n_sets == 1 && sets[0].rtl)
5939 try_back_substitute_reg (sets[0].rtl, insn);
5941 done:;
5944 /* Remove from the hash table all expressions that reference memory. */
5946 static void
5947 invalidate_memory (void)
5949 int i;
5950 struct table_elt *p, *next;
5952 for (i = 0; i < HASH_SIZE; i++)
5953 for (p = table[i]; p; p = next)
5955 next = p->next_same_hash;
5956 if (p->in_memory)
5957 remove_from_table (p, i);
5961 /* Perform invalidation on the basis of everything about INSN,
5962 except for invalidating the actual places that are SET in it.
5963 This includes the places CLOBBERed, and anything that might
5964 alias with something that is SET or CLOBBERed. */
5966 static void
5967 invalidate_from_clobbers (rtx insn)
5969 rtx x = PATTERN (insn);
5971 if (GET_CODE (x) == CLOBBER)
5973 rtx ref = XEXP (x, 0);
5974 if (ref)
5976 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5977 || MEM_P (ref))
5978 invalidate (ref, VOIDmode);
5979 else if (GET_CODE (ref) == STRICT_LOW_PART
5980 || GET_CODE (ref) == ZERO_EXTRACT)
5981 invalidate (XEXP (ref, 0), GET_MODE (ref));
5984 else if (GET_CODE (x) == PARALLEL)
5986 int i;
5987 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5989 rtx y = XVECEXP (x, 0, i);
5990 if (GET_CODE (y) == CLOBBER)
5992 rtx ref = XEXP (y, 0);
5993 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5994 || MEM_P (ref))
5995 invalidate (ref, VOIDmode);
5996 else if (GET_CODE (ref) == STRICT_LOW_PART
5997 || GET_CODE (ref) == ZERO_EXTRACT)
5998 invalidate (XEXP (ref, 0), GET_MODE (ref));
6004 /* Perform invalidation on the basis of everything about INSN.
6005 This includes the places CLOBBERed, and anything that might
6006 alias with something that is SET or CLOBBERed. */
6008 static void
6009 invalidate_from_sets_and_clobbers (rtx insn)
6011 rtx tem;
6012 rtx x = PATTERN (insn);
6014 if (CALL_P (insn))
6016 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6017 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6018 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6021 /* Ensure we invalidate the destination register of a CALL insn.
6022 This is necessary for machines where this register is a fixed_reg,
6023 because no other code would invalidate it. */
6024 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6025 invalidate (SET_DEST (x), VOIDmode);
6027 else if (GET_CODE (x) == PARALLEL)
6029 int i;
6031 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6033 rtx y = XVECEXP (x, 0, i);
6034 if (GET_CODE (y) == CLOBBER)
6036 rtx clobbered = XEXP (y, 0);
6038 if (REG_P (clobbered)
6039 || GET_CODE (clobbered) == SUBREG)
6040 invalidate (clobbered, VOIDmode);
6041 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6042 || GET_CODE (clobbered) == ZERO_EXTRACT)
6043 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6045 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6046 invalidate (SET_DEST (y), VOIDmode);
6051 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6052 and replace any registers in them with either an equivalent constant
6053 or the canonical form of the register. If we are inside an address,
6054 only do this if the address remains valid.
6056 OBJECT is 0 except when within a MEM in which case it is the MEM.
6058 Return the replacement for X. */
6060 static rtx
6061 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6063 enum rtx_code code = GET_CODE (x);
6064 const char *fmt = GET_RTX_FORMAT (code);
6065 int i;
6067 switch (code)
6069 case CONST:
6070 case SYMBOL_REF:
6071 case LABEL_REF:
6072 CASE_CONST_ANY:
6073 case PC:
6074 case CC0:
6075 case LO_SUM:
6076 return x;
6078 case MEM:
6079 validate_change (x, &XEXP (x, 0),
6080 cse_process_notes (XEXP (x, 0), x, changed), 0);
6081 return x;
6083 case EXPR_LIST:
6084 if (REG_NOTE_KIND (x) == REG_EQUAL)
6085 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6086 /* Fall through. */
6088 case INSN_LIST:
6089 case INT_LIST:
6090 if (XEXP (x, 1))
6091 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6092 return x;
6094 case SIGN_EXTEND:
6095 case ZERO_EXTEND:
6096 case SUBREG:
6098 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6099 /* We don't substitute VOIDmode constants into these rtx,
6100 since they would impede folding. */
6101 if (GET_MODE (new_rtx) != VOIDmode)
6102 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6103 return x;
6106 case UNSIGNED_FLOAT:
6108 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6109 /* We don't substitute negative VOIDmode constants into these rtx,
6110 since they would impede folding. */
6111 if (GET_MODE (new_rtx) != VOIDmode
6112 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6113 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6114 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6115 return x;
6118 case REG:
6119 i = REG_QTY (REGNO (x));
6121 /* Return a constant or a constant register. */
6122 if (REGNO_QTY_VALID_P (REGNO (x)))
6124 struct qty_table_elem *ent = &qty_table[i];
6126 if (ent->const_rtx != NULL_RTX
6127 && (CONSTANT_P (ent->const_rtx)
6128 || REG_P (ent->const_rtx)))
6130 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6131 if (new_rtx)
6132 return copy_rtx (new_rtx);
6136 /* Otherwise, canonicalize this register. */
6137 return canon_reg (x, NULL_RTX);
6139 default:
6140 break;
6143 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6144 if (fmt[i] == 'e')
6145 validate_change (object, &XEXP (x, i),
6146 cse_process_notes (XEXP (x, i), object, changed), 0);
6148 return x;
6151 static rtx
6152 cse_process_notes (rtx x, rtx object, bool *changed)
6154 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6155 if (new_rtx != x)
6156 *changed = true;
6157 return new_rtx;
6161 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6163 DATA is a pointer to a struct cse_basic_block_data, that is used to
6164 describe the path.
6165 It is filled with a queue of basic blocks, starting with FIRST_BB
6166 and following a trace through the CFG.
6168 If all paths starting at FIRST_BB have been followed, or no new path
6169 starting at FIRST_BB can be constructed, this function returns FALSE.
6170 Otherwise, DATA->path is filled and the function returns TRUE indicating
6171 that a path to follow was found.
6173 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6174 block in the path will be FIRST_BB. */
6176 static bool
6177 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6178 int follow_jumps)
6180 basic_block bb;
6181 edge e;
6182 int path_size;
6184 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6186 /* See if there is a previous path. */
6187 path_size = data->path_size;
6189 /* There is a previous path. Make sure it started with FIRST_BB. */
6190 if (path_size)
6191 gcc_assert (data->path[0].bb == first_bb);
6193 /* There was only one basic block in the last path. Clear the path and
6194 return, so that paths starting at another basic block can be tried. */
6195 if (path_size == 1)
6197 path_size = 0;
6198 goto done;
6201 /* If the path was empty from the beginning, construct a new path. */
6202 if (path_size == 0)
6203 data->path[path_size++].bb = first_bb;
6204 else
6206 /* Otherwise, path_size must be equal to or greater than 2, because
6207 a previous path exists that is at least two basic blocks long.
6209 Update the previous branch path, if any. If the last branch was
6210 previously along the branch edge, take the fallthrough edge now. */
6211 while (path_size >= 2)
6213 basic_block last_bb_in_path, previous_bb_in_path;
6214 edge e;
6216 --path_size;
6217 last_bb_in_path = data->path[path_size].bb;
6218 previous_bb_in_path = data->path[path_size - 1].bb;
6220 /* If we previously followed a path along the branch edge, try
6221 the fallthru edge now. */
6222 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6223 && any_condjump_p (BB_END (previous_bb_in_path))
6224 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6225 && e == BRANCH_EDGE (previous_bb_in_path))
6227 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6228 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6229 && single_pred_p (bb)
6230 /* We used to assert here that we would only see blocks
6231 that we have not visited yet. But we may end up
6232 visiting basic blocks twice if the CFG has changed
6233 in this run of cse_main, because when the CFG changes
6234 the topological sort of the CFG also changes. A basic
6235 blocks that previously had more than two predecessors
6236 may now have a single predecessor, and become part of
6237 a path that starts at another basic block.
6239 We still want to visit each basic block only once, so
6240 halt the path here if we have already visited BB. */
6241 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6243 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6244 data->path[path_size++].bb = bb;
6245 break;
6249 data->path[path_size].bb = NULL;
6252 /* If only one block remains in the path, bail. */
6253 if (path_size == 1)
6255 path_size = 0;
6256 goto done;
6260 /* Extend the path if possible. */
6261 if (follow_jumps)
6263 bb = data->path[path_size - 1].bb;
6264 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6266 if (single_succ_p (bb))
6267 e = single_succ_edge (bb);
6268 else if (EDGE_COUNT (bb->succs) == 2
6269 && any_condjump_p (BB_END (bb)))
6271 /* First try to follow the branch. If that doesn't lead
6272 to a useful path, follow the fallthru edge. */
6273 e = BRANCH_EDGE (bb);
6274 if (!single_pred_p (e->dest))
6275 e = FALLTHRU_EDGE (bb);
6277 else
6278 e = NULL;
6280 if (e
6281 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6282 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6283 && single_pred_p (e->dest)
6284 /* Avoid visiting basic blocks twice. The large comment
6285 above explains why this can happen. */
6286 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6288 basic_block bb2 = e->dest;
6289 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6290 data->path[path_size++].bb = bb2;
6291 bb = bb2;
6293 else
6294 bb = NULL;
6298 done:
6299 data->path_size = path_size;
6300 return path_size != 0;
6303 /* Dump the path in DATA to file F. NSETS is the number of sets
6304 in the path. */
6306 static void
6307 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6309 int path_entry;
6311 fprintf (f, ";; Following path with %d sets: ", nsets);
6312 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6313 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6314 fputc ('\n', dump_file);
6315 fflush (f);
6319 /* Return true if BB has exception handling successor edges. */
6321 static bool
6322 have_eh_succ_edges (basic_block bb)
6324 edge e;
6325 edge_iterator ei;
6327 FOR_EACH_EDGE (e, ei, bb->succs)
6328 if (e->flags & EDGE_EH)
6329 return true;
6331 return false;
6335 /* Scan to the end of the path described by DATA. Return an estimate of
6336 the total number of SETs of all insns in the path. */
6338 static void
6339 cse_prescan_path (struct cse_basic_block_data *data)
6341 int nsets = 0;
6342 int path_size = data->path_size;
6343 int path_entry;
6345 /* Scan to end of each basic block in the path. */
6346 for (path_entry = 0; path_entry < path_size; path_entry++)
6348 basic_block bb;
6349 rtx insn;
6351 bb = data->path[path_entry].bb;
6353 FOR_BB_INSNS (bb, insn)
6355 if (!INSN_P (insn))
6356 continue;
6358 /* A PARALLEL can have lots of SETs in it,
6359 especially if it is really an ASM_OPERANDS. */
6360 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6361 nsets += XVECLEN (PATTERN (insn), 0);
6362 else
6363 nsets += 1;
6367 data->nsets = nsets;
6370 /* Process a single extended basic block described by EBB_DATA. */
6372 static void
6373 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6375 int path_size = ebb_data->path_size;
6376 int path_entry;
6377 int num_insns = 0;
6379 /* Allocate the space needed by qty_table. */
6380 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6382 new_basic_block ();
6383 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6384 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6385 for (path_entry = 0; path_entry < path_size; path_entry++)
6387 basic_block bb;
6388 rtx insn;
6390 bb = ebb_data->path[path_entry].bb;
6392 /* Invalidate recorded information for eh regs if there is an EH
6393 edge pointing to that bb. */
6394 if (bb_has_eh_pred (bb))
6396 df_ref *def_rec;
6398 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6400 df_ref def = *def_rec;
6401 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6402 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6406 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6407 FOR_BB_INSNS (bb, insn)
6409 /* If we have processed 1,000 insns, flush the hash table to
6410 avoid extreme quadratic behavior. We must not include NOTEs
6411 in the count since there may be more of them when generating
6412 debugging information. If we clear the table at different
6413 times, code generated with -g -O might be different than code
6414 generated with -O but not -g.
6416 FIXME: This is a real kludge and needs to be done some other
6417 way. */
6418 if (NONDEBUG_INSN_P (insn)
6419 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6421 flush_hash_table ();
6422 num_insns = 0;
6425 if (INSN_P (insn))
6427 /* Process notes first so we have all notes in canonical forms
6428 when looking for duplicate operations. */
6429 if (REG_NOTES (insn))
6431 bool changed = false;
6432 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6433 NULL_RTX, &changed);
6434 if (changed)
6435 df_notes_rescan (insn);
6438 cse_insn (insn);
6440 /* If we haven't already found an insn where we added a LABEL_REF,
6441 check this one. */
6442 if (INSN_P (insn) && !recorded_label_ref
6443 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6444 (void *) insn))
6445 recorded_label_ref = true;
6447 #ifdef HAVE_cc0
6448 if (NONDEBUG_INSN_P (insn))
6450 /* If the previous insn sets CC0 and this insn no
6451 longer references CC0, delete the previous insn.
6452 Here we use fact that nothing expects CC0 to be
6453 valid over an insn, which is true until the final
6454 pass. */
6455 rtx prev_insn, tem;
6457 prev_insn = prev_nonnote_nondebug_insn (insn);
6458 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6459 && (tem = single_set (prev_insn)) != NULL_RTX
6460 && SET_DEST (tem) == cc0_rtx
6461 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6462 delete_insn (prev_insn);
6464 /* If this insn is not the last insn in the basic
6465 block, it will be PREV_INSN(insn) in the next
6466 iteration. If we recorded any CC0-related
6467 information for this insn, remember it. */
6468 if (insn != BB_END (bb))
6470 prev_insn_cc0 = this_insn_cc0;
6471 prev_insn_cc0_mode = this_insn_cc0_mode;
6474 #endif
6478 /* With non-call exceptions, we are not always able to update
6479 the CFG properly inside cse_insn. So clean up possibly
6480 redundant EH edges here. */
6481 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6482 cse_cfg_altered |= purge_dead_edges (bb);
6484 /* If we changed a conditional jump, we may have terminated
6485 the path we are following. Check that by verifying that
6486 the edge we would take still exists. If the edge does
6487 not exist anymore, purge the remainder of the path.
6488 Note that this will cause us to return to the caller. */
6489 if (path_entry < path_size - 1)
6491 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6492 if (!find_edge (bb, next_bb))
6496 path_size--;
6498 /* If we truncate the path, we must also reset the
6499 visited bit on the remaining blocks in the path,
6500 or we will never visit them at all. */
6501 bitmap_clear_bit (cse_visited_basic_blocks,
6502 ebb_data->path[path_size].bb->index);
6503 ebb_data->path[path_size].bb = NULL;
6505 while (path_size - 1 != path_entry);
6506 ebb_data->path_size = path_size;
6510 /* If this is a conditional jump insn, record any known
6511 equivalences due to the condition being tested. */
6512 insn = BB_END (bb);
6513 if (path_entry < path_size - 1
6514 && JUMP_P (insn)
6515 && single_set (insn)
6516 && any_condjump_p (insn))
6518 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6519 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6520 record_jump_equiv (insn, taken);
6523 #ifdef HAVE_cc0
6524 /* Clear the CC0-tracking related insns, they can't provide
6525 useful information across basic block boundaries. */
6526 prev_insn_cc0 = 0;
6527 #endif
6530 gcc_assert (next_qty <= max_qty);
6532 free (qty_table);
6536 /* Perform cse on the instructions of a function.
6537 F is the first instruction.
6538 NREGS is one plus the highest pseudo-reg number used in the instruction.
6540 Return 2 if jump optimizations should be redone due to simplifications
6541 in conditional jump instructions.
6542 Return 1 if the CFG should be cleaned up because it has been modified.
6543 Return 0 otherwise. */
6545 static int
6546 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6548 struct cse_basic_block_data ebb_data;
6549 basic_block bb;
6550 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6551 int i, n_blocks;
6553 df_set_flags (DF_LR_RUN_DCE);
6554 df_note_add_problem ();
6555 df_analyze ();
6556 df_set_flags (DF_DEFER_INSN_RESCAN);
6558 reg_scan (get_insns (), max_reg_num ());
6559 init_cse_reg_info (nregs);
6561 ebb_data.path = XNEWVEC (struct branch_path,
6562 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6564 cse_cfg_altered = false;
6565 cse_jumps_altered = false;
6566 recorded_label_ref = false;
6567 constant_pool_entries_cost = 0;
6568 constant_pool_entries_regcost = 0;
6569 ebb_data.path_size = 0;
6570 ebb_data.nsets = 0;
6571 rtl_hooks = cse_rtl_hooks;
6573 init_recog ();
6574 init_alias_analysis ();
6576 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6578 /* Set up the table of already visited basic blocks. */
6579 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6580 bitmap_clear (cse_visited_basic_blocks);
6582 /* Loop over basic blocks in reverse completion order (RPO),
6583 excluding the ENTRY and EXIT blocks. */
6584 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6585 i = 0;
6586 while (i < n_blocks)
6588 /* Find the first block in the RPO queue that we have not yet
6589 processed before. */
6592 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6594 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6595 && i < n_blocks);
6597 /* Find all paths starting with BB, and process them. */
6598 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6600 /* Pre-scan the path. */
6601 cse_prescan_path (&ebb_data);
6603 /* If this basic block has no sets, skip it. */
6604 if (ebb_data.nsets == 0)
6605 continue;
6607 /* Get a reasonable estimate for the maximum number of qty's
6608 needed for this path. For this, we take the number of sets
6609 and multiply that by MAX_RECOG_OPERANDS. */
6610 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6612 /* Dump the path we're about to process. */
6613 if (dump_file)
6614 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6616 cse_extended_basic_block (&ebb_data);
6620 /* Clean up. */
6621 end_alias_analysis ();
6622 free (reg_eqv_table);
6623 free (ebb_data.path);
6624 sbitmap_free (cse_visited_basic_blocks);
6625 free (rc_order);
6626 rtl_hooks = general_rtl_hooks;
6628 if (cse_jumps_altered || recorded_label_ref)
6629 return 2;
6630 else if (cse_cfg_altered)
6631 return 1;
6632 else
6633 return 0;
6636 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6637 which there isn't a REG_LABEL_OPERAND note.
6638 Return one if so. DATA is the insn. */
6640 static int
6641 check_for_label_ref (rtx *rtl, void *data)
6643 rtx insn = (rtx) data;
6645 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6646 note for it, we must rerun jump since it needs to place the note. If
6647 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6648 don't do this since no REG_LABEL_OPERAND will be added. */
6649 return (GET_CODE (*rtl) == LABEL_REF
6650 && ! LABEL_REF_NONLOCAL_P (*rtl)
6651 && (!JUMP_P (insn)
6652 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6653 && LABEL_P (XEXP (*rtl, 0))
6654 && INSN_UID (XEXP (*rtl, 0)) != 0
6655 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6658 /* Count the number of times registers are used (not set) in X.
6659 COUNTS is an array in which we accumulate the count, INCR is how much
6660 we count each register usage.
6662 Don't count a usage of DEST, which is the SET_DEST of a SET which
6663 contains X in its SET_SRC. This is because such a SET does not
6664 modify the liveness of DEST.
6665 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6666 We must then count uses of a SET_DEST regardless, because the insn can't be
6667 deleted here. */
6669 static void
6670 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6672 enum rtx_code code;
6673 rtx note;
6674 const char *fmt;
6675 int i, j;
6677 if (x == 0)
6678 return;
6680 switch (code = GET_CODE (x))
6682 case REG:
6683 if (x != dest)
6684 counts[REGNO (x)] += incr;
6685 return;
6687 case PC:
6688 case CC0:
6689 case CONST:
6690 CASE_CONST_ANY:
6691 case SYMBOL_REF:
6692 case LABEL_REF:
6693 return;
6695 case CLOBBER:
6696 /* If we are clobbering a MEM, mark any registers inside the address
6697 as being used. */
6698 if (MEM_P (XEXP (x, 0)))
6699 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6700 return;
6702 case SET:
6703 /* Unless we are setting a REG, count everything in SET_DEST. */
6704 if (!REG_P (SET_DEST (x)))
6705 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6706 count_reg_usage (SET_SRC (x), counts,
6707 dest ? dest : SET_DEST (x),
6708 incr);
6709 return;
6711 case DEBUG_INSN:
6712 return;
6714 case CALL_INSN:
6715 case INSN:
6716 case JUMP_INSN:
6717 /* We expect dest to be NULL_RTX here. If the insn may throw,
6718 or if it cannot be deleted due to side-effects, mark this fact
6719 by setting DEST to pc_rtx. */
6720 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6721 || side_effects_p (PATTERN (x)))
6722 dest = pc_rtx;
6723 if (code == CALL_INSN)
6724 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6725 count_reg_usage (PATTERN (x), counts, dest, incr);
6727 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6728 use them. */
6730 note = find_reg_equal_equiv_note (x);
6731 if (note)
6733 rtx eqv = XEXP (note, 0);
6735 if (GET_CODE (eqv) == EXPR_LIST)
6736 /* This REG_EQUAL note describes the result of a function call.
6737 Process all the arguments. */
6740 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6741 eqv = XEXP (eqv, 1);
6743 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6744 else
6745 count_reg_usage (eqv, counts, dest, incr);
6747 return;
6749 case EXPR_LIST:
6750 if (REG_NOTE_KIND (x) == REG_EQUAL
6751 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6752 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6753 involving registers in the address. */
6754 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6755 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6757 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6758 return;
6760 case ASM_OPERANDS:
6761 /* Iterate over just the inputs, not the constraints as well. */
6762 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6763 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6764 return;
6766 case INSN_LIST:
6767 case INT_LIST:
6768 gcc_unreachable ();
6770 default:
6771 break;
6774 fmt = GET_RTX_FORMAT (code);
6775 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6777 if (fmt[i] == 'e')
6778 count_reg_usage (XEXP (x, i), counts, dest, incr);
6779 else if (fmt[i] == 'E')
6780 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6781 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6785 /* Return true if X is a dead register. */
6787 static inline int
6788 is_dead_reg (rtx x, int *counts)
6790 return (REG_P (x)
6791 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6792 && counts[REGNO (x)] == 0);
6795 /* Return true if set is live. */
6796 static bool
6797 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6798 int *counts)
6800 #ifdef HAVE_cc0
6801 rtx tem;
6802 #endif
6804 if (set_noop_p (set))
6807 #ifdef HAVE_cc0
6808 else if (GET_CODE (SET_DEST (set)) == CC0
6809 && !side_effects_p (SET_SRC (set))
6810 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6811 || !INSN_P (tem)
6812 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6813 return false;
6814 #endif
6815 else if (!is_dead_reg (SET_DEST (set), counts)
6816 || side_effects_p (SET_SRC (set)))
6817 return true;
6818 return false;
6821 /* Return true if insn is live. */
6823 static bool
6824 insn_live_p (rtx insn, int *counts)
6826 int i;
6827 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6828 return true;
6829 else if (GET_CODE (PATTERN (insn)) == SET)
6830 return set_live_p (PATTERN (insn), insn, counts);
6831 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6833 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6835 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6837 if (GET_CODE (elt) == SET)
6839 if (set_live_p (elt, insn, counts))
6840 return true;
6842 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6843 return true;
6845 return false;
6847 else if (DEBUG_INSN_P (insn))
6849 rtx next;
6851 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6852 if (NOTE_P (next))
6853 continue;
6854 else if (!DEBUG_INSN_P (next))
6855 return true;
6856 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6857 return false;
6859 return true;
6861 else
6862 return true;
6865 /* Count the number of stores into pseudo. Callback for note_stores. */
6867 static void
6868 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6870 int *counts = (int *) data;
6871 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6872 counts[REGNO (x)]++;
6875 struct dead_debug_insn_data
6877 int *counts;
6878 rtx *replacements;
6879 bool seen_repl;
6882 /* Return if a DEBUG_INSN needs to be reset because some dead
6883 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6885 static int
6886 is_dead_debug_insn (rtx *loc, void *data)
6888 rtx x = *loc;
6889 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6891 if (is_dead_reg (x, ddid->counts))
6893 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6894 ddid->seen_repl = true;
6895 else
6896 return 1;
6898 return 0;
6901 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6902 Callback for simplify_replace_fn_rtx. */
6904 static rtx
6905 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6907 rtx *replacements = (rtx *) data;
6909 if (REG_P (x)
6910 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6911 && replacements[REGNO (x)] != NULL_RTX)
6913 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6914 return replacements[REGNO (x)];
6915 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6916 GET_MODE (replacements[REGNO (x)]));
6918 return NULL_RTX;
6921 /* Scan all the insns and delete any that are dead; i.e., they store a register
6922 that is never used or they copy a register to itself.
6924 This is used to remove insns made obviously dead by cse, loop or other
6925 optimizations. It improves the heuristics in loop since it won't try to
6926 move dead invariants out of loops or make givs for dead quantities. The
6927 remaining passes of the compilation are also sped up. */
6930 delete_trivially_dead_insns (rtx insns, int nreg)
6932 int *counts;
6933 rtx insn, prev;
6934 rtx *replacements = NULL;
6935 int ndead = 0;
6937 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6938 /* First count the number of times each register is used. */
6939 if (MAY_HAVE_DEBUG_INSNS)
6941 counts = XCNEWVEC (int, nreg * 3);
6942 for (insn = insns; insn; insn = NEXT_INSN (insn))
6943 if (DEBUG_INSN_P (insn))
6944 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6945 NULL_RTX, 1);
6946 else if (INSN_P (insn))
6948 count_reg_usage (insn, counts, NULL_RTX, 1);
6949 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6951 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6952 First one counts how many times each pseudo is used outside
6953 of debug insns, second counts how many times each pseudo is
6954 used in debug insns and third counts how many times a pseudo
6955 is stored. */
6957 else
6959 counts = XCNEWVEC (int, nreg);
6960 for (insn = insns; insn; insn = NEXT_INSN (insn))
6961 if (INSN_P (insn))
6962 count_reg_usage (insn, counts, NULL_RTX, 1);
6963 /* If no debug insns can be present, COUNTS is just an array
6964 which counts how many times each pseudo is used. */
6966 /* Go from the last insn to the first and delete insns that only set unused
6967 registers or copy a register to itself. As we delete an insn, remove
6968 usage counts for registers it uses.
6970 The first jump optimization pass may leave a real insn as the last
6971 insn in the function. We must not skip that insn or we may end
6972 up deleting code that is not really dead.
6974 If some otherwise unused register is only used in DEBUG_INSNs,
6975 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6976 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6977 has been created for the unused register, replace it with
6978 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6979 for (insn = get_last_insn (); insn; insn = prev)
6981 int live_insn = 0;
6983 prev = PREV_INSN (insn);
6984 if (!INSN_P (insn))
6985 continue;
6987 live_insn = insn_live_p (insn, counts);
6989 /* If this is a dead insn, delete it and show registers in it aren't
6990 being used. */
6992 if (! live_insn && dbg_cnt (delete_trivial_dead))
6994 if (DEBUG_INSN_P (insn))
6995 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6996 NULL_RTX, -1);
6997 else
6999 rtx set;
7000 if (MAY_HAVE_DEBUG_INSNS
7001 && (set = single_set (insn)) != NULL_RTX
7002 && is_dead_reg (SET_DEST (set), counts)
7003 /* Used at least once in some DEBUG_INSN. */
7004 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7005 /* And set exactly once. */
7006 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7007 && !side_effects_p (SET_SRC (set))
7008 && asm_noperands (PATTERN (insn)) < 0)
7010 rtx dval, bind;
7012 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7013 dval = make_debug_expr_from_rtl (SET_DEST (set));
7015 /* Emit a debug bind insn before the insn in which
7016 reg dies. */
7017 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7018 DEBUG_EXPR_TREE_DECL (dval),
7019 SET_SRC (set),
7020 VAR_INIT_STATUS_INITIALIZED);
7021 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
7023 bind = emit_debug_insn_before (bind, insn);
7024 df_insn_rescan (bind);
7026 if (replacements == NULL)
7027 replacements = XCNEWVEC (rtx, nreg);
7028 replacements[REGNO (SET_DEST (set))] = dval;
7031 count_reg_usage (insn, counts, NULL_RTX, -1);
7032 ndead++;
7034 delete_insn_and_edges (insn);
7038 if (MAY_HAVE_DEBUG_INSNS)
7040 struct dead_debug_insn_data ddid;
7041 ddid.counts = counts;
7042 ddid.replacements = replacements;
7043 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7044 if (DEBUG_INSN_P (insn))
7046 /* If this debug insn references a dead register that wasn't replaced
7047 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7048 ddid.seen_repl = false;
7049 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
7050 is_dead_debug_insn, &ddid))
7052 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7053 df_insn_rescan (insn);
7055 else if (ddid.seen_repl)
7057 INSN_VAR_LOCATION_LOC (insn)
7058 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7059 NULL_RTX, replace_dead_reg,
7060 replacements);
7061 df_insn_rescan (insn);
7064 free (replacements);
7067 if (dump_file && ndead)
7068 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7069 ndead);
7070 /* Clean up. */
7071 free (counts);
7072 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7073 return ndead;
7076 /* This function is called via for_each_rtx. The argument, NEWREG, is
7077 a condition code register with the desired mode. If we are looking
7078 at the same register in a different mode, replace it with
7079 NEWREG. */
7081 static int
7082 cse_change_cc_mode (rtx *loc, void *data)
7084 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7086 if (*loc
7087 && REG_P (*loc)
7088 && REGNO (*loc) == REGNO (args->newreg)
7089 && GET_MODE (*loc) != GET_MODE (args->newreg))
7091 validate_change (args->insn, loc, args->newreg, 1);
7093 return -1;
7095 return 0;
7098 /* Change the mode of any reference to the register REGNO (NEWREG) to
7099 GET_MODE (NEWREG) in INSN. */
7101 static void
7102 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7104 struct change_cc_mode_args args;
7105 int success;
7107 if (!INSN_P (insn))
7108 return;
7110 args.insn = insn;
7111 args.newreg = newreg;
7113 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7114 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7116 /* If the following assertion was triggered, there is most probably
7117 something wrong with the cc_modes_compatible back end function.
7118 CC modes only can be considered compatible if the insn - with the mode
7119 replaced by any of the compatible modes - can still be recognized. */
7120 success = apply_change_group ();
7121 gcc_assert (success);
7124 /* Change the mode of any reference to the register REGNO (NEWREG) to
7125 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7126 any instruction which modifies NEWREG. */
7128 static void
7129 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7131 rtx insn;
7133 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7135 if (! INSN_P (insn))
7136 continue;
7138 if (reg_set_p (newreg, insn))
7139 return;
7141 cse_change_cc_mode_insn (insn, newreg);
7145 /* BB is a basic block which finishes with CC_REG as a condition code
7146 register which is set to CC_SRC. Look through the successors of BB
7147 to find blocks which have a single predecessor (i.e., this one),
7148 and look through those blocks for an assignment to CC_REG which is
7149 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7150 permitted to change the mode of CC_SRC to a compatible mode. This
7151 returns VOIDmode if no equivalent assignments were found.
7152 Otherwise it returns the mode which CC_SRC should wind up with.
7153 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7154 but is passed unmodified down to recursive calls in order to prevent
7155 endless recursion.
7157 The main complexity in this function is handling the mode issues.
7158 We may have more than one duplicate which we can eliminate, and we
7159 try to find a mode which will work for multiple duplicates. */
7161 static enum machine_mode
7162 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7163 bool can_change_mode)
7165 bool found_equiv;
7166 enum machine_mode mode;
7167 unsigned int insn_count;
7168 edge e;
7169 rtx insns[2];
7170 enum machine_mode modes[2];
7171 rtx last_insns[2];
7172 unsigned int i;
7173 rtx newreg;
7174 edge_iterator ei;
7176 /* We expect to have two successors. Look at both before picking
7177 the final mode for the comparison. If we have more successors
7178 (i.e., some sort of table jump, although that seems unlikely),
7179 then we require all beyond the first two to use the same
7180 mode. */
7182 found_equiv = false;
7183 mode = GET_MODE (cc_src);
7184 insn_count = 0;
7185 FOR_EACH_EDGE (e, ei, bb->succs)
7187 rtx insn;
7188 rtx end;
7190 if (e->flags & EDGE_COMPLEX)
7191 continue;
7193 if (EDGE_COUNT (e->dest->preds) != 1
7194 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7195 /* Avoid endless recursion on unreachable blocks. */
7196 || e->dest == orig_bb)
7197 continue;
7199 end = NEXT_INSN (BB_END (e->dest));
7200 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7202 rtx set;
7204 if (! INSN_P (insn))
7205 continue;
7207 /* If CC_SRC is modified, we have to stop looking for
7208 something which uses it. */
7209 if (modified_in_p (cc_src, insn))
7210 break;
7212 /* Check whether INSN sets CC_REG to CC_SRC. */
7213 set = single_set (insn);
7214 if (set
7215 && REG_P (SET_DEST (set))
7216 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7218 bool found;
7219 enum machine_mode set_mode;
7220 enum machine_mode comp_mode;
7222 found = false;
7223 set_mode = GET_MODE (SET_SRC (set));
7224 comp_mode = set_mode;
7225 if (rtx_equal_p (cc_src, SET_SRC (set)))
7226 found = true;
7227 else if (GET_CODE (cc_src) == COMPARE
7228 && GET_CODE (SET_SRC (set)) == COMPARE
7229 && mode != set_mode
7230 && rtx_equal_p (XEXP (cc_src, 0),
7231 XEXP (SET_SRC (set), 0))
7232 && rtx_equal_p (XEXP (cc_src, 1),
7233 XEXP (SET_SRC (set), 1)))
7236 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7237 if (comp_mode != VOIDmode
7238 && (can_change_mode || comp_mode == mode))
7239 found = true;
7242 if (found)
7244 found_equiv = true;
7245 if (insn_count < ARRAY_SIZE (insns))
7247 insns[insn_count] = insn;
7248 modes[insn_count] = set_mode;
7249 last_insns[insn_count] = end;
7250 ++insn_count;
7252 if (mode != comp_mode)
7254 gcc_assert (can_change_mode);
7255 mode = comp_mode;
7257 /* The modified insn will be re-recognized later. */
7258 PUT_MODE (cc_src, mode);
7261 else
7263 if (set_mode != mode)
7265 /* We found a matching expression in the
7266 wrong mode, but we don't have room to
7267 store it in the array. Punt. This case
7268 should be rare. */
7269 break;
7271 /* INSN sets CC_REG to a value equal to CC_SRC
7272 with the right mode. We can simply delete
7273 it. */
7274 delete_insn (insn);
7277 /* We found an instruction to delete. Keep looking,
7278 in the hopes of finding a three-way jump. */
7279 continue;
7282 /* We found an instruction which sets the condition
7283 code, so don't look any farther. */
7284 break;
7287 /* If INSN sets CC_REG in some other way, don't look any
7288 farther. */
7289 if (reg_set_p (cc_reg, insn))
7290 break;
7293 /* If we fell off the bottom of the block, we can keep looking
7294 through successors. We pass CAN_CHANGE_MODE as false because
7295 we aren't prepared to handle compatibility between the
7296 further blocks and this block. */
7297 if (insn == end)
7299 enum machine_mode submode;
7301 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7302 if (submode != VOIDmode)
7304 gcc_assert (submode == mode);
7305 found_equiv = true;
7306 can_change_mode = false;
7311 if (! found_equiv)
7312 return VOIDmode;
7314 /* Now INSN_COUNT is the number of instructions we found which set
7315 CC_REG to a value equivalent to CC_SRC. The instructions are in
7316 INSNS. The modes used by those instructions are in MODES. */
7318 newreg = NULL_RTX;
7319 for (i = 0; i < insn_count; ++i)
7321 if (modes[i] != mode)
7323 /* We need to change the mode of CC_REG in INSNS[i] and
7324 subsequent instructions. */
7325 if (! newreg)
7327 if (GET_MODE (cc_reg) == mode)
7328 newreg = cc_reg;
7329 else
7330 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7332 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7333 newreg);
7336 delete_insn_and_edges (insns[i]);
7339 return mode;
7342 /* If we have a fixed condition code register (or two), walk through
7343 the instructions and try to eliminate duplicate assignments. */
7345 static void
7346 cse_condition_code_reg (void)
7348 unsigned int cc_regno_1;
7349 unsigned int cc_regno_2;
7350 rtx cc_reg_1;
7351 rtx cc_reg_2;
7352 basic_block bb;
7354 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7355 return;
7357 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7358 if (cc_regno_2 != INVALID_REGNUM)
7359 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7360 else
7361 cc_reg_2 = NULL_RTX;
7363 FOR_EACH_BB_FN (bb, cfun)
7365 rtx last_insn;
7366 rtx cc_reg;
7367 rtx insn;
7368 rtx cc_src_insn;
7369 rtx cc_src;
7370 enum machine_mode mode;
7371 enum machine_mode orig_mode;
7373 /* Look for blocks which end with a conditional jump based on a
7374 condition code register. Then look for the instruction which
7375 sets the condition code register. Then look through the
7376 successor blocks for instructions which set the condition
7377 code register to the same value. There are other possible
7378 uses of the condition code register, but these are by far the
7379 most common and the ones which we are most likely to be able
7380 to optimize. */
7382 last_insn = BB_END (bb);
7383 if (!JUMP_P (last_insn))
7384 continue;
7386 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7387 cc_reg = cc_reg_1;
7388 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7389 cc_reg = cc_reg_2;
7390 else
7391 continue;
7393 cc_src_insn = NULL_RTX;
7394 cc_src = NULL_RTX;
7395 for (insn = PREV_INSN (last_insn);
7396 insn && insn != PREV_INSN (BB_HEAD (bb));
7397 insn = PREV_INSN (insn))
7399 rtx set;
7401 if (! INSN_P (insn))
7402 continue;
7403 set = single_set (insn);
7404 if (set
7405 && REG_P (SET_DEST (set))
7406 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7408 cc_src_insn = insn;
7409 cc_src = SET_SRC (set);
7410 break;
7412 else if (reg_set_p (cc_reg, insn))
7413 break;
7416 if (! cc_src_insn)
7417 continue;
7419 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7420 continue;
7422 /* Now CC_REG is a condition code register used for a
7423 conditional jump at the end of the block, and CC_SRC, in
7424 CC_SRC_INSN, is the value to which that condition code
7425 register is set, and CC_SRC is still meaningful at the end of
7426 the basic block. */
7428 orig_mode = GET_MODE (cc_src);
7429 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7430 if (mode != VOIDmode)
7432 gcc_assert (mode == GET_MODE (cc_src));
7433 if (mode != orig_mode)
7435 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7437 cse_change_cc_mode_insn (cc_src_insn, newreg);
7439 /* Do the same in the following insns that use the
7440 current value of CC_REG within BB. */
7441 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7442 NEXT_INSN (last_insn),
7443 newreg);
7450 /* Perform common subexpression elimination. Nonzero value from
7451 `cse_main' means that jumps were simplified and some code may now
7452 be unreachable, so do jump optimization again. */
7453 static bool
7454 gate_handle_cse (void)
7456 return optimize > 0;
7459 static unsigned int
7460 rest_of_handle_cse (void)
7462 int tem;
7464 if (dump_file)
7465 dump_flow_info (dump_file, dump_flags);
7467 tem = cse_main (get_insns (), max_reg_num ());
7469 /* If we are not running more CSE passes, then we are no longer
7470 expecting CSE to be run. But always rerun it in a cheap mode. */
7471 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7473 if (tem == 2)
7475 timevar_push (TV_JUMP);
7476 rebuild_jump_labels (get_insns ());
7477 cleanup_cfg (CLEANUP_CFG_CHANGED);
7478 timevar_pop (TV_JUMP);
7480 else if (tem == 1 || optimize > 1)
7481 cleanup_cfg (0);
7483 return 0;
7486 namespace {
7488 const pass_data pass_data_cse =
7490 RTL_PASS, /* type */
7491 "cse1", /* name */
7492 OPTGROUP_NONE, /* optinfo_flags */
7493 true, /* has_gate */
7494 true, /* has_execute */
7495 TV_CSE, /* tv_id */
7496 0, /* properties_required */
7497 0, /* properties_provided */
7498 0, /* properties_destroyed */
7499 0, /* todo_flags_start */
7500 ( TODO_df_finish | TODO_verify_rtl_sharing
7501 | TODO_verify_flow ), /* todo_flags_finish */
7504 class pass_cse : public rtl_opt_pass
7506 public:
7507 pass_cse (gcc::context *ctxt)
7508 : rtl_opt_pass (pass_data_cse, ctxt)
7511 /* opt_pass methods: */
7512 bool gate () { return gate_handle_cse (); }
7513 unsigned int execute () { return rest_of_handle_cse (); }
7515 }; // class pass_cse
7517 } // anon namespace
7519 rtl_opt_pass *
7520 make_pass_cse (gcc::context *ctxt)
7522 return new pass_cse (ctxt);
7526 static bool
7527 gate_handle_cse2 (void)
7529 return optimize > 0 && flag_rerun_cse_after_loop;
7532 /* Run second CSE pass after loop optimizations. */
7533 static unsigned int
7534 rest_of_handle_cse2 (void)
7536 int tem;
7538 if (dump_file)
7539 dump_flow_info (dump_file, dump_flags);
7541 tem = cse_main (get_insns (), max_reg_num ());
7543 /* Run a pass to eliminate duplicated assignments to condition code
7544 registers. We have to run this after bypass_jumps, because it
7545 makes it harder for that pass to determine whether a jump can be
7546 bypassed safely. */
7547 cse_condition_code_reg ();
7549 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7551 if (tem == 2)
7553 timevar_push (TV_JUMP);
7554 rebuild_jump_labels (get_insns ());
7555 cleanup_cfg (CLEANUP_CFG_CHANGED);
7556 timevar_pop (TV_JUMP);
7558 else if (tem == 1)
7559 cleanup_cfg (0);
7561 cse_not_expected = 1;
7562 return 0;
7566 namespace {
7568 const pass_data pass_data_cse2 =
7570 RTL_PASS, /* type */
7571 "cse2", /* name */
7572 OPTGROUP_NONE, /* optinfo_flags */
7573 true, /* has_gate */
7574 true, /* has_execute */
7575 TV_CSE2, /* tv_id */
7576 0, /* properties_required */
7577 0, /* properties_provided */
7578 0, /* properties_destroyed */
7579 0, /* todo_flags_start */
7580 ( TODO_df_finish | TODO_verify_rtl_sharing
7581 | TODO_verify_flow ), /* todo_flags_finish */
7584 class pass_cse2 : public rtl_opt_pass
7586 public:
7587 pass_cse2 (gcc::context *ctxt)
7588 : rtl_opt_pass (pass_data_cse2, ctxt)
7591 /* opt_pass methods: */
7592 bool gate () { return gate_handle_cse2 (); }
7593 unsigned int execute () { return rest_of_handle_cse2 (); }
7595 }; // class pass_cse2
7597 } // anon namespace
7599 rtl_opt_pass *
7600 make_pass_cse2 (gcc::context *ctxt)
7602 return new pass_cse2 (ctxt);
7605 static bool
7606 gate_handle_cse_after_global_opts (void)
7608 return optimize > 0 && flag_rerun_cse_after_global_opts;
7611 /* Run second CSE pass after loop optimizations. */
7612 static unsigned int
7613 rest_of_handle_cse_after_global_opts (void)
7615 int save_cfj;
7616 int tem;
7618 /* We only want to do local CSE, so don't follow jumps. */
7619 save_cfj = flag_cse_follow_jumps;
7620 flag_cse_follow_jumps = 0;
7622 rebuild_jump_labels (get_insns ());
7623 tem = cse_main (get_insns (), max_reg_num ());
7624 purge_all_dead_edges ();
7625 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7627 cse_not_expected = !flag_rerun_cse_after_loop;
7629 /* If cse altered any jumps, rerun jump opts to clean things up. */
7630 if (tem == 2)
7632 timevar_push (TV_JUMP);
7633 rebuild_jump_labels (get_insns ());
7634 cleanup_cfg (CLEANUP_CFG_CHANGED);
7635 timevar_pop (TV_JUMP);
7637 else if (tem == 1)
7638 cleanup_cfg (0);
7640 flag_cse_follow_jumps = save_cfj;
7641 return 0;
7644 namespace {
7646 const pass_data pass_data_cse_after_global_opts =
7648 RTL_PASS, /* type */
7649 "cse_local", /* name */
7650 OPTGROUP_NONE, /* optinfo_flags */
7651 true, /* has_gate */
7652 true, /* has_execute */
7653 TV_CSE, /* tv_id */
7654 0, /* properties_required */
7655 0, /* properties_provided */
7656 0, /* properties_destroyed */
7657 0, /* todo_flags_start */
7658 ( TODO_df_finish | TODO_verify_rtl_sharing
7659 | TODO_verify_flow ), /* todo_flags_finish */
7662 class pass_cse_after_global_opts : public rtl_opt_pass
7664 public:
7665 pass_cse_after_global_opts (gcc::context *ctxt)
7666 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7669 /* opt_pass methods: */
7670 bool gate () { return gate_handle_cse_after_global_opts (); }
7671 unsigned int execute () {
7672 return rest_of_handle_cse_after_global_opts ();
7675 }; // class pass_cse_after_global_opts
7677 } // anon namespace
7679 rtl_opt_pass *
7680 make_pass_cse_after_global_opts (gcc::context *ctxt)
7682 return new pass_cse_after_global_opts (ctxt);