* xcoffout.c (xcoff_tls_data_section_name): Define.
[official-gcc.git] / gcc / cse.c
blobff91b9d93f54d8f8de178240221645428931a37c
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010,
4 2011 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "regs.h"
30 #include "basic-block.h"
31 #include "flags.h"
32 #include "insn-config.h"
33 #include "recog.h"
34 #include "function.h"
35 #include "expr.h"
36 #include "diagnostic-core.h"
37 #include "toplev.h"
38 #include "ggc.h"
39 #include "except.h"
40 #include "target.h"
41 #include "params.h"
42 #include "rtlhooks-def.h"
43 #include "tree-pass.h"
44 #include "df.h"
45 #include "dbgcnt.h"
46 #include "pointer-set.h"
48 /* The basic idea of common subexpression elimination is to go
49 through the code, keeping a record of expressions that would
50 have the same value at the current scan point, and replacing
51 expressions encountered with the cheapest equivalent expression.
53 It is too complicated to keep track of the different possibilities
54 when control paths merge in this code; so, at each label, we forget all
55 that is known and start fresh. This can be described as processing each
56 extended basic block separately. We have a separate pass to perform
57 global CSE.
59 Note CSE can turn a conditional or computed jump into a nop or
60 an unconditional jump. When this occurs we arrange to run the jump
61 optimizer after CSE to delete the unreachable code.
63 We use two data structures to record the equivalent expressions:
64 a hash table for most expressions, and a vector of "quantity
65 numbers" to record equivalent (pseudo) registers.
67 The use of the special data structure for registers is desirable
68 because it is faster. It is possible because registers references
69 contain a fairly small number, the register number, taken from
70 a contiguously allocated series, and two register references are
71 identical if they have the same number. General expressions
72 do not have any such thing, so the only way to retrieve the
73 information recorded on an expression other than a register
74 is to keep it in a hash table.
76 Registers and "quantity numbers":
78 At the start of each basic block, all of the (hardware and pseudo)
79 registers used in the function are given distinct quantity
80 numbers to indicate their contents. During scan, when the code
81 copies one register into another, we copy the quantity number.
82 When a register is loaded in any other way, we allocate a new
83 quantity number to describe the value generated by this operation.
84 `REG_QTY (N)' records what quantity register N is currently thought
85 of as containing.
87 All real quantity numbers are greater than or equal to zero.
88 If register N has not been assigned a quantity, `REG_QTY (N)' will
89 equal -N - 1, which is always negative.
91 Quantity numbers below zero do not exist and none of the `qty_table'
92 entries should be referenced with a negative index.
94 We also maintain a bidirectional chain of registers for each
95 quantity number. The `qty_table` members `first_reg' and `last_reg',
96 and `reg_eqv_table' members `next' and `prev' hold these chains.
98 The first register in a chain is the one whose lifespan is least local.
99 Among equals, it is the one that was seen first.
100 We replace any equivalent register with that one.
102 If two registers have the same quantity number, it must be true that
103 REG expressions with qty_table `mode' must be in the hash table for both
104 registers and must be in the same class.
106 The converse is not true. Since hard registers may be referenced in
107 any mode, two REG expressions might be equivalent in the hash table
108 but not have the same quantity number if the quantity number of one
109 of the registers is not the same mode as those expressions.
111 Constants and quantity numbers
113 When a quantity has a known constant value, that value is stored
114 in the appropriate qty_table `const_rtx'. This is in addition to
115 putting the constant in the hash table as is usual for non-regs.
117 Whether a reg or a constant is preferred is determined by the configuration
118 macro CONST_COSTS and will often depend on the constant value. In any
119 event, expressions containing constants can be simplified, by fold_rtx.
121 When a quantity has a known nearly constant value (such as an address
122 of a stack slot), that value is stored in the appropriate qty_table
123 `const_rtx'.
125 Integer constants don't have a machine mode. However, cse
126 determines the intended machine mode from the destination
127 of the instruction that moves the constant. The machine mode
128 is recorded in the hash table along with the actual RTL
129 constant expression so that different modes are kept separate.
131 Other expressions:
133 To record known equivalences among expressions in general
134 we use a hash table called `table'. It has a fixed number of buckets
135 that contain chains of `struct table_elt' elements for expressions.
136 These chains connect the elements whose expressions have the same
137 hash codes.
139 Other chains through the same elements connect the elements which
140 currently have equivalent values.
142 Register references in an expression are canonicalized before hashing
143 the expression. This is done using `reg_qty' and qty_table `first_reg'.
144 The hash code of a register reference is computed using the quantity
145 number, not the register number.
147 When the value of an expression changes, it is necessary to remove from the
148 hash table not just that expression but all expressions whose values
149 could be different as a result.
151 1. If the value changing is in memory, except in special cases
152 ANYTHING referring to memory could be changed. That is because
153 nobody knows where a pointer does not point.
154 The function `invalidate_memory' removes what is necessary.
156 The special cases are when the address is constant or is
157 a constant plus a fixed register such as the frame pointer
158 or a static chain pointer. When such addresses are stored in,
159 we can tell exactly which other such addresses must be invalidated
160 due to overlap. `invalidate' does this.
161 All expressions that refer to non-constant
162 memory addresses are also invalidated. `invalidate_memory' does this.
164 2. If the value changing is a register, all expressions
165 containing references to that register, and only those,
166 must be removed.
168 Because searching the entire hash table for expressions that contain
169 a register is very slow, we try to figure out when it isn't necessary.
170 Precisely, this is necessary only when expressions have been
171 entered in the hash table using this register, and then the value has
172 changed, and then another expression wants to be added to refer to
173 the register's new value. This sequence of circumstances is rare
174 within any one basic block.
176 `REG_TICK' and `REG_IN_TABLE', accessors for members of
177 cse_reg_info, are used to detect this case. REG_TICK (i) is
178 incremented whenever a value is stored in register i.
179 REG_IN_TABLE (i) holds -1 if no references to register i have been
180 entered in the table; otherwise, it contains the value REG_TICK (i)
181 had when the references were entered. If we want to enter a
182 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
183 remove old references. Until we want to enter a new entry, the
184 mere fact that the two vectors don't match makes the entries be
185 ignored if anyone tries to match them.
187 Registers themselves are entered in the hash table as well as in
188 the equivalent-register chains. However, `REG_TICK' and
189 `REG_IN_TABLE' do not apply to expressions which are simple
190 register references. These expressions are removed from the table
191 immediately when they become invalid, and this can be done even if
192 we do not immediately search for all the expressions that refer to
193 the register.
195 A CLOBBER rtx in an instruction invalidates its operand for further
196 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
197 invalidates everything that resides in memory.
199 Related expressions:
201 Constant expressions that differ only by an additive integer
202 are called related. When a constant expression is put in
203 the table, the related expression with no constant term
204 is also entered. These are made to point at each other
205 so that it is possible to find out if there exists any
206 register equivalent to an expression related to a given expression. */
208 /* Length of qty_table vector. We know in advance we will not need
209 a quantity number this big. */
211 static int max_qty;
213 /* Next quantity number to be allocated.
214 This is 1 + the largest number needed so far. */
216 static int next_qty;
218 /* Per-qty information tracking.
220 `first_reg' and `last_reg' track the head and tail of the
221 chain of registers which currently contain this quantity.
223 `mode' contains the machine mode of this quantity.
225 `const_rtx' holds the rtx of the constant value of this
226 quantity, if known. A summations of the frame/arg pointer
227 and a constant can also be entered here. When this holds
228 a known value, `const_insn' is the insn which stored the
229 constant value.
231 `comparison_{code,const,qty}' are used to track when a
232 comparison between a quantity and some constant or register has
233 been passed. In such a case, we know the results of the comparison
234 in case we see it again. These members record a comparison that
235 is known to be true. `comparison_code' holds the rtx code of such
236 a comparison, else it is set to UNKNOWN and the other two
237 comparison members are undefined. `comparison_const' holds
238 the constant being compared against, or zero if the comparison
239 is not against a constant. `comparison_qty' holds the quantity
240 being compared against when the result is known. If the comparison
241 is not with a register, `comparison_qty' is -1. */
243 struct qty_table_elem
245 rtx const_rtx;
246 rtx const_insn;
247 rtx comparison_const;
248 int comparison_qty;
249 unsigned int first_reg, last_reg;
250 /* The sizes of these fields should match the sizes of the
251 code and mode fields of struct rtx_def (see rtl.h). */
252 ENUM_BITFIELD(rtx_code) comparison_code : 16;
253 ENUM_BITFIELD(machine_mode) mode : 8;
256 /* The table of all qtys, indexed by qty number. */
257 static struct qty_table_elem *qty_table;
259 /* Structure used to pass arguments via for_each_rtx to function
260 cse_change_cc_mode. */
261 struct change_cc_mode_args
263 rtx insn;
264 rtx newreg;
267 #ifdef HAVE_cc0
268 /* For machines that have a CC0, we do not record its value in the hash
269 table since its use is guaranteed to be the insn immediately following
270 its definition and any other insn is presumed to invalidate it.
272 Instead, we store below the current and last value assigned to CC0.
273 If it should happen to be a constant, it is stored in preference
274 to the actual assigned value. In case it is a constant, we store
275 the mode in which the constant should be interpreted. */
277 static rtx this_insn_cc0, prev_insn_cc0;
278 static enum machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
279 #endif
281 /* Insn being scanned. */
283 static rtx this_insn;
284 static bool optimize_this_for_speed_p;
286 /* Index by register number, gives the number of the next (or
287 previous) register in the chain of registers sharing the same
288 value.
290 Or -1 if this register is at the end of the chain.
292 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
294 /* Per-register equivalence chain. */
295 struct reg_eqv_elem
297 int next, prev;
300 /* The table of all register equivalence chains. */
301 static struct reg_eqv_elem *reg_eqv_table;
303 struct cse_reg_info
305 /* The timestamp at which this register is initialized. */
306 unsigned int timestamp;
308 /* The quantity number of the register's current contents. */
309 int reg_qty;
311 /* The number of times the register has been altered in the current
312 basic block. */
313 int reg_tick;
315 /* The REG_TICK value at which rtx's containing this register are
316 valid in the hash table. If this does not equal the current
317 reg_tick value, such expressions existing in the hash table are
318 invalid. */
319 int reg_in_table;
321 /* The SUBREG that was set when REG_TICK was last incremented. Set
322 to -1 if the last store was to the whole register, not a subreg. */
323 unsigned int subreg_ticked;
326 /* A table of cse_reg_info indexed by register numbers. */
327 static struct cse_reg_info *cse_reg_info_table;
329 /* The size of the above table. */
330 static unsigned int cse_reg_info_table_size;
332 /* The index of the first entry that has not been initialized. */
333 static unsigned int cse_reg_info_table_first_uninitialized;
335 /* The timestamp at the beginning of the current run of
336 cse_extended_basic_block. We increment this variable at the beginning of
337 the current run of cse_extended_basic_block. The timestamp field of a
338 cse_reg_info entry matches the value of this variable if and only
339 if the entry has been initialized during the current run of
340 cse_extended_basic_block. */
341 static unsigned int cse_reg_info_timestamp;
343 /* A HARD_REG_SET containing all the hard registers for which there is
344 currently a REG expression in the hash table. Note the difference
345 from the above variables, which indicate if the REG is mentioned in some
346 expression in the table. */
348 static HARD_REG_SET hard_regs_in_table;
350 /* True if CSE has altered the CFG. */
351 static bool cse_cfg_altered;
353 /* True if CSE has altered conditional jump insns in such a way
354 that jump optimization should be redone. */
355 static bool cse_jumps_altered;
357 /* True if we put a LABEL_REF into the hash table for an INSN
358 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
359 to put in the note. */
360 static bool recorded_label_ref;
362 /* canon_hash stores 1 in do_not_record
363 if it notices a reference to CC0, PC, or some other volatile
364 subexpression. */
366 static int do_not_record;
368 /* canon_hash stores 1 in hash_arg_in_memory
369 if it notices a reference to memory within the expression being hashed. */
371 static int hash_arg_in_memory;
373 /* The hash table contains buckets which are chains of `struct table_elt's,
374 each recording one expression's information.
375 That expression is in the `exp' field.
377 The canon_exp field contains a canonical (from the point of view of
378 alias analysis) version of the `exp' field.
380 Those elements with the same hash code are chained in both directions
381 through the `next_same_hash' and `prev_same_hash' fields.
383 Each set of expressions with equivalent values
384 are on a two-way chain through the `next_same_value'
385 and `prev_same_value' fields, and all point with
386 the `first_same_value' field at the first element in
387 that chain. The chain is in order of increasing cost.
388 Each element's cost value is in its `cost' field.
390 The `in_memory' field is nonzero for elements that
391 involve any reference to memory. These elements are removed
392 whenever a write is done to an unidentified location in memory.
393 To be safe, we assume that a memory address is unidentified unless
394 the address is either a symbol constant or a constant plus
395 the frame pointer or argument pointer.
397 The `related_value' field is used to connect related expressions
398 (that differ by adding an integer).
399 The related expressions are chained in a circular fashion.
400 `related_value' is zero for expressions for which this
401 chain is not useful.
403 The `cost' field stores the cost of this element's expression.
404 The `regcost' field stores the value returned by approx_reg_cost for
405 this element's expression.
407 The `is_const' flag is set if the element is a constant (including
408 a fixed address).
410 The `flag' field is used as a temporary during some search routines.
412 The `mode' field is usually the same as GET_MODE (`exp'), but
413 if `exp' is a CONST_INT and has no machine mode then the `mode'
414 field is the mode it was being used as. Each constant is
415 recorded separately for each mode it is used with. */
417 struct table_elt
419 rtx exp;
420 rtx canon_exp;
421 struct table_elt *next_same_hash;
422 struct table_elt *prev_same_hash;
423 struct table_elt *next_same_value;
424 struct table_elt *prev_same_value;
425 struct table_elt *first_same_value;
426 struct table_elt *related_value;
427 int cost;
428 int regcost;
429 /* The size of this field should match the size
430 of the mode field of struct rtx_def (see rtl.h). */
431 ENUM_BITFIELD(machine_mode) mode : 8;
432 char in_memory;
433 char is_const;
434 char flag;
437 /* We don't want a lot of buckets, because we rarely have very many
438 things stored in the hash table, and a lot of buckets slows
439 down a lot of loops that happen frequently. */
440 #define HASH_SHIFT 5
441 #define HASH_SIZE (1 << HASH_SHIFT)
442 #define HASH_MASK (HASH_SIZE - 1)
444 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
445 register (hard registers may require `do_not_record' to be set). */
447 #define HASH(X, M) \
448 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
449 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
450 : canon_hash (X, M)) & HASH_MASK)
452 /* Like HASH, but without side-effects. */
453 #define SAFE_HASH(X, M) \
454 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
455 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
456 : safe_hash (X, M)) & HASH_MASK)
458 /* Determine whether register number N is considered a fixed register for the
459 purpose of approximating register costs.
460 It is desirable to replace other regs with fixed regs, to reduce need for
461 non-fixed hard regs.
462 A reg wins if it is either the frame pointer or designated as fixed. */
463 #define FIXED_REGNO_P(N) \
464 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
465 || fixed_regs[N] || global_regs[N])
467 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
468 hard registers and pointers into the frame are the cheapest with a cost
469 of 0. Next come pseudos with a cost of one and other hard registers with
470 a cost of 2. Aside from these special cases, call `rtx_cost'. */
472 #define CHEAP_REGNO(N) \
473 (REGNO_PTR_FRAME_P(N) \
474 || (HARD_REGISTER_NUM_P (N) \
475 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
477 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET, 1))
478 #define COST_IN(X, OUTER, OPNO) (REG_P (X) ? 0 : notreg_cost (X, OUTER, OPNO))
480 /* Get the number of times this register has been updated in this
481 basic block. */
483 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
485 /* Get the point at which REG was recorded in the table. */
487 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
489 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
490 SUBREG). */
492 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
494 /* Get the quantity number for REG. */
496 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
498 /* Determine if the quantity number for register X represents a valid index
499 into the qty_table. */
501 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
503 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
505 #define CHEAPER(X, Y) \
506 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
508 static struct table_elt *table[HASH_SIZE];
510 /* Chain of `struct table_elt's made so far for this function
511 but currently removed from the table. */
513 static struct table_elt *free_element_chain;
515 /* Set to the cost of a constant pool reference if one was found for a
516 symbolic constant. If this was found, it means we should try to
517 convert constants into constant pool entries if they don't fit in
518 the insn. */
520 static int constant_pool_entries_cost;
521 static int constant_pool_entries_regcost;
523 /* Trace a patch through the CFG. */
525 struct branch_path
527 /* The basic block for this path entry. */
528 basic_block bb;
531 /* This data describes a block that will be processed by
532 cse_extended_basic_block. */
534 struct cse_basic_block_data
536 /* Total number of SETs in block. */
537 int nsets;
538 /* Size of current branch path, if any. */
539 int path_size;
540 /* Current path, indicating which basic_blocks will be processed. */
541 struct branch_path *path;
545 /* Pointers to the live in/live out bitmaps for the boundaries of the
546 current EBB. */
547 static bitmap cse_ebb_live_in, cse_ebb_live_out;
549 /* A simple bitmap to track which basic blocks have been visited
550 already as part of an already processed extended basic block. */
551 static sbitmap cse_visited_basic_blocks;
553 static bool fixed_base_plus_p (rtx x);
554 static int notreg_cost (rtx, enum rtx_code, int);
555 static int approx_reg_cost_1 (rtx *, void *);
556 static int approx_reg_cost (rtx);
557 static int preferable (int, int, int, int);
558 static void new_basic_block (void);
559 static void make_new_qty (unsigned int, enum machine_mode);
560 static void make_regs_eqv (unsigned int, unsigned int);
561 static void delete_reg_equiv (unsigned int);
562 static int mention_regs (rtx);
563 static int insert_regs (rtx, struct table_elt *, int);
564 static void remove_from_table (struct table_elt *, unsigned);
565 static void remove_pseudo_from_table (rtx, unsigned);
566 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
567 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
568 static rtx lookup_as_function (rtx, enum rtx_code);
569 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
570 enum machine_mode, int, int);
571 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
572 enum machine_mode);
573 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
574 static void invalidate (rtx, enum machine_mode);
575 static void remove_invalid_refs (unsigned int);
576 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
577 enum machine_mode);
578 static void rehash_using_reg (rtx);
579 static void invalidate_memory (void);
580 static void invalidate_for_call (void);
581 static rtx use_related_value (rtx, struct table_elt *);
583 static inline unsigned canon_hash (rtx, enum machine_mode);
584 static inline unsigned safe_hash (rtx, enum machine_mode);
585 static inline unsigned hash_rtx_string (const char *);
587 static rtx canon_reg (rtx, rtx);
588 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
589 enum machine_mode *,
590 enum machine_mode *);
591 static rtx fold_rtx (rtx, rtx);
592 static rtx equiv_constant (rtx);
593 static void record_jump_equiv (rtx, bool);
594 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
595 int);
596 static void cse_insn (rtx);
597 static void cse_prescan_path (struct cse_basic_block_data *);
598 static void invalidate_from_clobbers (rtx);
599 static void invalidate_from_sets_and_clobbers (rtx);
600 static rtx cse_process_notes (rtx, rtx, bool *);
601 static void cse_extended_basic_block (struct cse_basic_block_data *);
602 static int check_for_label_ref (rtx *, void *);
603 extern void dump_class (struct table_elt*);
604 static void get_cse_reg_info_1 (unsigned int regno);
605 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
606 static int check_dependence (rtx *, void *);
608 static void flush_hash_table (void);
609 static bool insn_live_p (rtx, int *);
610 static bool set_live_p (rtx, rtx, int *);
611 static int cse_change_cc_mode (rtx *, void *);
612 static void cse_change_cc_mode_insn (rtx, rtx);
613 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
614 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
615 bool);
618 #undef RTL_HOOKS_GEN_LOWPART
619 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
621 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
623 /* Nonzero if X has the form (PLUS frame-pointer integer). */
625 static bool
626 fixed_base_plus_p (rtx x)
628 switch (GET_CODE (x))
630 case REG:
631 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
632 return true;
633 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
634 return true;
635 return false;
637 case PLUS:
638 if (!CONST_INT_P (XEXP (x, 1)))
639 return false;
640 return fixed_base_plus_p (XEXP (x, 0));
642 default:
643 return false;
647 /* Dump the expressions in the equivalence class indicated by CLASSP.
648 This function is used only for debugging. */
649 DEBUG_FUNCTION void
650 dump_class (struct table_elt *classp)
652 struct table_elt *elt;
654 fprintf (stderr, "Equivalence chain for ");
655 print_rtl (stderr, classp->exp);
656 fprintf (stderr, ": \n");
658 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
660 print_rtl (stderr, elt->exp);
661 fprintf (stderr, "\n");
665 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
667 static int
668 approx_reg_cost_1 (rtx *xp, void *data)
670 rtx x = *xp;
671 int *cost_p = (int *) data;
673 if (x && REG_P (x))
675 unsigned int regno = REGNO (x);
677 if (! CHEAP_REGNO (regno))
679 if (regno < FIRST_PSEUDO_REGISTER)
681 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
682 return 1;
683 *cost_p += 2;
685 else
686 *cost_p += 1;
690 return 0;
693 /* Return an estimate of the cost of the registers used in an rtx.
694 This is mostly the number of different REG expressions in the rtx;
695 however for some exceptions like fixed registers we use a cost of
696 0. If any other hard register reference occurs, return MAX_COST. */
698 static int
699 approx_reg_cost (rtx x)
701 int cost = 0;
703 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
704 return MAX_COST;
706 return cost;
709 /* Return a negative value if an rtx A, whose costs are given by COST_A
710 and REGCOST_A, is more desirable than an rtx B.
711 Return a positive value if A is less desirable, or 0 if the two are
712 equally good. */
713 static int
714 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
716 /* First, get rid of cases involving expressions that are entirely
717 unwanted. */
718 if (cost_a != cost_b)
720 if (cost_a == MAX_COST)
721 return 1;
722 if (cost_b == MAX_COST)
723 return -1;
726 /* Avoid extending lifetimes of hardregs. */
727 if (regcost_a != regcost_b)
729 if (regcost_a == MAX_COST)
730 return 1;
731 if (regcost_b == MAX_COST)
732 return -1;
735 /* Normal operation costs take precedence. */
736 if (cost_a != cost_b)
737 return cost_a - cost_b;
738 /* Only if these are identical consider effects on register pressure. */
739 if (regcost_a != regcost_b)
740 return regcost_a - regcost_b;
741 return 0;
744 /* Internal function, to compute cost when X is not a register; called
745 from COST macro to keep it simple. */
747 static int
748 notreg_cost (rtx x, enum rtx_code outer, int opno)
750 return ((GET_CODE (x) == SUBREG
751 && REG_P (SUBREG_REG (x))
752 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
753 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
754 && (GET_MODE_SIZE (GET_MODE (x))
755 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
756 && subreg_lowpart_p (x)
757 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (x),
758 GET_MODE (SUBREG_REG (x))))
760 : rtx_cost (x, outer, opno, optimize_this_for_speed_p) * 2);
764 /* Initialize CSE_REG_INFO_TABLE. */
766 static void
767 init_cse_reg_info (unsigned int nregs)
769 /* Do we need to grow the table? */
770 if (nregs > cse_reg_info_table_size)
772 unsigned int new_size;
774 if (cse_reg_info_table_size < 2048)
776 /* Compute a new size that is a power of 2 and no smaller
777 than the large of NREGS and 64. */
778 new_size = (cse_reg_info_table_size
779 ? cse_reg_info_table_size : 64);
781 while (new_size < nregs)
782 new_size *= 2;
784 else
786 /* If we need a big table, allocate just enough to hold
787 NREGS registers. */
788 new_size = nregs;
791 /* Reallocate the table with NEW_SIZE entries. */
792 free (cse_reg_info_table);
793 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
794 cse_reg_info_table_size = new_size;
795 cse_reg_info_table_first_uninitialized = 0;
798 /* Do we have all of the first NREGS entries initialized? */
799 if (cse_reg_info_table_first_uninitialized < nregs)
801 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
802 unsigned int i;
804 /* Put the old timestamp on newly allocated entries so that they
805 will all be considered out of date. We do not touch those
806 entries beyond the first NREGS entries to be nice to the
807 virtual memory. */
808 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
809 cse_reg_info_table[i].timestamp = old_timestamp;
811 cse_reg_info_table_first_uninitialized = nregs;
815 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
817 static void
818 get_cse_reg_info_1 (unsigned int regno)
820 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
821 entry will be considered to have been initialized. */
822 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
824 /* Initialize the rest of the entry. */
825 cse_reg_info_table[regno].reg_tick = 1;
826 cse_reg_info_table[regno].reg_in_table = -1;
827 cse_reg_info_table[regno].subreg_ticked = -1;
828 cse_reg_info_table[regno].reg_qty = -regno - 1;
831 /* Find a cse_reg_info entry for REGNO. */
833 static inline struct cse_reg_info *
834 get_cse_reg_info (unsigned int regno)
836 struct cse_reg_info *p = &cse_reg_info_table[regno];
838 /* If this entry has not been initialized, go ahead and initialize
839 it. */
840 if (p->timestamp != cse_reg_info_timestamp)
841 get_cse_reg_info_1 (regno);
843 return p;
846 /* Clear the hash table and initialize each register with its own quantity,
847 for a new basic block. */
849 static void
850 new_basic_block (void)
852 int i;
854 next_qty = 0;
856 /* Invalidate cse_reg_info_table. */
857 cse_reg_info_timestamp++;
859 /* Clear out hash table state for this pass. */
860 CLEAR_HARD_REG_SET (hard_regs_in_table);
862 /* The per-quantity values used to be initialized here, but it is
863 much faster to initialize each as it is made in `make_new_qty'. */
865 for (i = 0; i < HASH_SIZE; i++)
867 struct table_elt *first;
869 first = table[i];
870 if (first != NULL)
872 struct table_elt *last = first;
874 table[i] = NULL;
876 while (last->next_same_hash != NULL)
877 last = last->next_same_hash;
879 /* Now relink this hash entire chain into
880 the free element list. */
882 last->next_same_hash = free_element_chain;
883 free_element_chain = first;
887 #ifdef HAVE_cc0
888 prev_insn_cc0 = 0;
889 #endif
892 /* Say that register REG contains a quantity in mode MODE not in any
893 register before and initialize that quantity. */
895 static void
896 make_new_qty (unsigned int reg, enum machine_mode mode)
898 int q;
899 struct qty_table_elem *ent;
900 struct reg_eqv_elem *eqv;
902 gcc_assert (next_qty < max_qty);
904 q = REG_QTY (reg) = next_qty++;
905 ent = &qty_table[q];
906 ent->first_reg = reg;
907 ent->last_reg = reg;
908 ent->mode = mode;
909 ent->const_rtx = ent->const_insn = NULL_RTX;
910 ent->comparison_code = UNKNOWN;
912 eqv = &reg_eqv_table[reg];
913 eqv->next = eqv->prev = -1;
916 /* Make reg NEW equivalent to reg OLD.
917 OLD is not changing; NEW is. */
919 static void
920 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
922 unsigned int lastr, firstr;
923 int q = REG_QTY (old_reg);
924 struct qty_table_elem *ent;
926 ent = &qty_table[q];
928 /* Nothing should become eqv until it has a "non-invalid" qty number. */
929 gcc_assert (REGNO_QTY_VALID_P (old_reg));
931 REG_QTY (new_reg) = q;
932 firstr = ent->first_reg;
933 lastr = ent->last_reg;
935 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
936 hard regs. Among pseudos, if NEW will live longer than any other reg
937 of the same qty, and that is beyond the current basic block,
938 make it the new canonical replacement for this qty. */
939 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
940 /* Certain fixed registers might be of the class NO_REGS. This means
941 that not only can they not be allocated by the compiler, but
942 they cannot be used in substitutions or canonicalizations
943 either. */
944 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
945 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
946 || (new_reg >= FIRST_PSEUDO_REGISTER
947 && (firstr < FIRST_PSEUDO_REGISTER
948 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
949 && !bitmap_bit_p (cse_ebb_live_out, firstr))
950 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
951 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
953 reg_eqv_table[firstr].prev = new_reg;
954 reg_eqv_table[new_reg].next = firstr;
955 reg_eqv_table[new_reg].prev = -1;
956 ent->first_reg = new_reg;
958 else
960 /* If NEW is a hard reg (known to be non-fixed), insert at end.
961 Otherwise, insert before any non-fixed hard regs that are at the
962 end. Registers of class NO_REGS cannot be used as an
963 equivalent for anything. */
964 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
965 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
966 && new_reg >= FIRST_PSEUDO_REGISTER)
967 lastr = reg_eqv_table[lastr].prev;
968 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
969 if (reg_eqv_table[lastr].next >= 0)
970 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
971 else
972 qty_table[q].last_reg = new_reg;
973 reg_eqv_table[lastr].next = new_reg;
974 reg_eqv_table[new_reg].prev = lastr;
978 /* Remove REG from its equivalence class. */
980 static void
981 delete_reg_equiv (unsigned int reg)
983 struct qty_table_elem *ent;
984 int q = REG_QTY (reg);
985 int p, n;
987 /* If invalid, do nothing. */
988 if (! REGNO_QTY_VALID_P (reg))
989 return;
991 ent = &qty_table[q];
993 p = reg_eqv_table[reg].prev;
994 n = reg_eqv_table[reg].next;
996 if (n != -1)
997 reg_eqv_table[n].prev = p;
998 else
999 ent->last_reg = p;
1000 if (p != -1)
1001 reg_eqv_table[p].next = n;
1002 else
1003 ent->first_reg = n;
1005 REG_QTY (reg) = -reg - 1;
1008 /* Remove any invalid expressions from the hash table
1009 that refer to any of the registers contained in expression X.
1011 Make sure that newly inserted references to those registers
1012 as subexpressions will be considered valid.
1014 mention_regs is not called when a register itself
1015 is being stored in the table.
1017 Return 1 if we have done something that may have changed the hash code
1018 of X. */
1020 static int
1021 mention_regs (rtx x)
1023 enum rtx_code code;
1024 int i, j;
1025 const char *fmt;
1026 int changed = 0;
1028 if (x == 0)
1029 return 0;
1031 code = GET_CODE (x);
1032 if (code == REG)
1034 unsigned int regno = REGNO (x);
1035 unsigned int endregno = END_REGNO (x);
1036 unsigned int i;
1038 for (i = regno; i < endregno; i++)
1040 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1041 remove_invalid_refs (i);
1043 REG_IN_TABLE (i) = REG_TICK (i);
1044 SUBREG_TICKED (i) = -1;
1047 return 0;
1050 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1051 pseudo if they don't use overlapping words. We handle only pseudos
1052 here for simplicity. */
1053 if (code == SUBREG && REG_P (SUBREG_REG (x))
1054 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1056 unsigned int i = REGNO (SUBREG_REG (x));
1058 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1060 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1061 the last store to this register really stored into this
1062 subreg, then remove the memory of this subreg.
1063 Otherwise, remove any memory of the entire register and
1064 all its subregs from the table. */
1065 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1066 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1067 remove_invalid_refs (i);
1068 else
1069 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1072 REG_IN_TABLE (i) = REG_TICK (i);
1073 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1074 return 0;
1077 /* If X is a comparison or a COMPARE and either operand is a register
1078 that does not have a quantity, give it one. This is so that a later
1079 call to record_jump_equiv won't cause X to be assigned a different
1080 hash code and not found in the table after that call.
1082 It is not necessary to do this here, since rehash_using_reg can
1083 fix up the table later, but doing this here eliminates the need to
1084 call that expensive function in the most common case where the only
1085 use of the register is in the comparison. */
1087 if (code == COMPARE || COMPARISON_P (x))
1089 if (REG_P (XEXP (x, 0))
1090 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1091 if (insert_regs (XEXP (x, 0), NULL, 0))
1093 rehash_using_reg (XEXP (x, 0));
1094 changed = 1;
1097 if (REG_P (XEXP (x, 1))
1098 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1099 if (insert_regs (XEXP (x, 1), NULL, 0))
1101 rehash_using_reg (XEXP (x, 1));
1102 changed = 1;
1106 fmt = GET_RTX_FORMAT (code);
1107 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1108 if (fmt[i] == 'e')
1109 changed |= mention_regs (XEXP (x, i));
1110 else if (fmt[i] == 'E')
1111 for (j = 0; j < XVECLEN (x, i); j++)
1112 changed |= mention_regs (XVECEXP (x, i, j));
1114 return changed;
1117 /* Update the register quantities for inserting X into the hash table
1118 with a value equivalent to CLASSP.
1119 (If the class does not contain a REG, it is irrelevant.)
1120 If MODIFIED is nonzero, X is a destination; it is being modified.
1121 Note that delete_reg_equiv should be called on a register
1122 before insert_regs is done on that register with MODIFIED != 0.
1124 Nonzero value means that elements of reg_qty have changed
1125 so X's hash code may be different. */
1127 static int
1128 insert_regs (rtx x, struct table_elt *classp, int modified)
1130 if (REG_P (x))
1132 unsigned int regno = REGNO (x);
1133 int qty_valid;
1135 /* If REGNO is in the equivalence table already but is of the
1136 wrong mode for that equivalence, don't do anything here. */
1138 qty_valid = REGNO_QTY_VALID_P (regno);
1139 if (qty_valid)
1141 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1143 if (ent->mode != GET_MODE (x))
1144 return 0;
1147 if (modified || ! qty_valid)
1149 if (classp)
1150 for (classp = classp->first_same_value;
1151 classp != 0;
1152 classp = classp->next_same_value)
1153 if (REG_P (classp->exp)
1154 && GET_MODE (classp->exp) == GET_MODE (x))
1156 unsigned c_regno = REGNO (classp->exp);
1158 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1160 /* Suppose that 5 is hard reg and 100 and 101 are
1161 pseudos. Consider
1163 (set (reg:si 100) (reg:si 5))
1164 (set (reg:si 5) (reg:si 100))
1165 (set (reg:di 101) (reg:di 5))
1167 We would now set REG_QTY (101) = REG_QTY (5), but the
1168 entry for 5 is in SImode. When we use this later in
1169 copy propagation, we get the register in wrong mode. */
1170 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1171 continue;
1173 make_regs_eqv (regno, c_regno);
1174 return 1;
1177 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1178 than REG_IN_TABLE to find out if there was only a single preceding
1179 invalidation - for the SUBREG - or another one, which would be
1180 for the full register. However, if we find here that REG_TICK
1181 indicates that the register is invalid, it means that it has
1182 been invalidated in a separate operation. The SUBREG might be used
1183 now (then this is a recursive call), or we might use the full REG
1184 now and a SUBREG of it later. So bump up REG_TICK so that
1185 mention_regs will do the right thing. */
1186 if (! modified
1187 && REG_IN_TABLE (regno) >= 0
1188 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1189 REG_TICK (regno)++;
1190 make_new_qty (regno, GET_MODE (x));
1191 return 1;
1194 return 0;
1197 /* If X is a SUBREG, we will likely be inserting the inner register in the
1198 table. If that register doesn't have an assigned quantity number at
1199 this point but does later, the insertion that we will be doing now will
1200 not be accessible because its hash code will have changed. So assign
1201 a quantity number now. */
1203 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1204 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1206 insert_regs (SUBREG_REG (x), NULL, 0);
1207 mention_regs (x);
1208 return 1;
1210 else
1211 return mention_regs (x);
1215 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1216 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1217 CST is equal to an anchor. */
1219 static bool
1220 compute_const_anchors (rtx cst,
1221 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1222 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1224 HOST_WIDE_INT n = INTVAL (cst);
1226 *lower_base = n & ~(targetm.const_anchor - 1);
1227 if (*lower_base == n)
1228 return false;
1230 *upper_base =
1231 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1232 *upper_offs = n - *upper_base;
1233 *lower_offs = n - *lower_base;
1234 return true;
1237 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1239 static void
1240 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1241 enum machine_mode mode)
1243 struct table_elt *elt;
1244 unsigned hash;
1245 rtx anchor_exp;
1246 rtx exp;
1248 anchor_exp = GEN_INT (anchor);
1249 hash = HASH (anchor_exp, mode);
1250 elt = lookup (anchor_exp, hash, mode);
1251 if (!elt)
1252 elt = insert (anchor_exp, NULL, hash, mode);
1254 exp = plus_constant (mode, reg, offs);
1255 /* REG has just been inserted and the hash codes recomputed. */
1256 mention_regs (exp);
1257 hash = HASH (exp, mode);
1259 /* Use the cost of the register rather than the whole expression. When
1260 looking up constant anchors we will further offset the corresponding
1261 expression therefore it does not make sense to prefer REGs over
1262 reg-immediate additions. Prefer instead the oldest expression. Also
1263 don't prefer pseudos over hard regs so that we derive constants in
1264 argument registers from other argument registers rather than from the
1265 original pseudo that was used to synthesize the constant. */
1266 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1269 /* The constant CST is equivalent to the register REG. Create
1270 equivalences between the two anchors of CST and the corresponding
1271 register-offset expressions using REG. */
1273 static void
1274 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1276 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1278 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1279 &upper_base, &upper_offs))
1280 return;
1282 /* Ignore anchors of value 0. Constants accessible from zero are
1283 simple. */
1284 if (lower_base != 0)
1285 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1287 if (upper_base != 0)
1288 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1291 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1292 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1293 valid expression. Return the cheapest and oldest of such expressions. In
1294 *OLD, return how old the resulting expression is compared to the other
1295 equivalent expressions. */
1297 static rtx
1298 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1299 unsigned *old)
1301 struct table_elt *elt;
1302 unsigned idx;
1303 struct table_elt *match_elt;
1304 rtx match;
1306 /* Find the cheapest and *oldest* expression to maximize the chance of
1307 reusing the same pseudo. */
1309 match_elt = NULL;
1310 match = NULL_RTX;
1311 for (elt = anchor_elt->first_same_value, idx = 0;
1312 elt;
1313 elt = elt->next_same_value, idx++)
1315 if (match_elt && CHEAPER (match_elt, elt))
1316 return match;
1318 if (REG_P (elt->exp)
1319 || (GET_CODE (elt->exp) == PLUS
1320 && REG_P (XEXP (elt->exp, 0))
1321 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1323 rtx x;
1325 /* Ignore expressions that are no longer valid. */
1326 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1327 continue;
1329 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1330 if (REG_P (x)
1331 || (GET_CODE (x) == PLUS
1332 && IN_RANGE (INTVAL (XEXP (x, 1)),
1333 -targetm.const_anchor,
1334 targetm.const_anchor - 1)))
1336 match = x;
1337 match_elt = elt;
1338 *old = idx;
1343 return match;
1346 /* Try to express the constant SRC_CONST using a register+offset expression
1347 derived from a constant anchor. Return it if successful or NULL_RTX,
1348 otherwise. */
1350 static rtx
1351 try_const_anchors (rtx src_const, enum machine_mode mode)
1353 struct table_elt *lower_elt, *upper_elt;
1354 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1355 rtx lower_anchor_rtx, upper_anchor_rtx;
1356 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1357 unsigned lower_old, upper_old;
1359 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1360 &upper_base, &upper_offs))
1361 return NULL_RTX;
1363 lower_anchor_rtx = GEN_INT (lower_base);
1364 upper_anchor_rtx = GEN_INT (upper_base);
1365 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1366 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1368 if (lower_elt)
1369 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1370 if (upper_elt)
1371 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1373 if (!lower_exp)
1374 return upper_exp;
1375 if (!upper_exp)
1376 return lower_exp;
1378 /* Return the older expression. */
1379 return (upper_old > lower_old ? upper_exp : lower_exp);
1382 /* Look in or update the hash table. */
1384 /* Remove table element ELT from use in the table.
1385 HASH is its hash code, made using the HASH macro.
1386 It's an argument because often that is known in advance
1387 and we save much time not recomputing it. */
1389 static void
1390 remove_from_table (struct table_elt *elt, unsigned int hash)
1392 if (elt == 0)
1393 return;
1395 /* Mark this element as removed. See cse_insn. */
1396 elt->first_same_value = 0;
1398 /* Remove the table element from its equivalence class. */
1401 struct table_elt *prev = elt->prev_same_value;
1402 struct table_elt *next = elt->next_same_value;
1404 if (next)
1405 next->prev_same_value = prev;
1407 if (prev)
1408 prev->next_same_value = next;
1409 else
1411 struct table_elt *newfirst = next;
1412 while (next)
1414 next->first_same_value = newfirst;
1415 next = next->next_same_value;
1420 /* Remove the table element from its hash bucket. */
1423 struct table_elt *prev = elt->prev_same_hash;
1424 struct table_elt *next = elt->next_same_hash;
1426 if (next)
1427 next->prev_same_hash = prev;
1429 if (prev)
1430 prev->next_same_hash = next;
1431 else if (table[hash] == elt)
1432 table[hash] = next;
1433 else
1435 /* This entry is not in the proper hash bucket. This can happen
1436 when two classes were merged by `merge_equiv_classes'. Search
1437 for the hash bucket that it heads. This happens only very
1438 rarely, so the cost is acceptable. */
1439 for (hash = 0; hash < HASH_SIZE; hash++)
1440 if (table[hash] == elt)
1441 table[hash] = next;
1445 /* Remove the table element from its related-value circular chain. */
1447 if (elt->related_value != 0 && elt->related_value != elt)
1449 struct table_elt *p = elt->related_value;
1451 while (p->related_value != elt)
1452 p = p->related_value;
1453 p->related_value = elt->related_value;
1454 if (p->related_value == p)
1455 p->related_value = 0;
1458 /* Now add it to the free element chain. */
1459 elt->next_same_hash = free_element_chain;
1460 free_element_chain = elt;
1463 /* Same as above, but X is a pseudo-register. */
1465 static void
1466 remove_pseudo_from_table (rtx x, unsigned int hash)
1468 struct table_elt *elt;
1470 /* Because a pseudo-register can be referenced in more than one
1471 mode, we might have to remove more than one table entry. */
1472 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1473 remove_from_table (elt, hash);
1476 /* Look up X in the hash table and return its table element,
1477 or 0 if X is not in the table.
1479 MODE is the machine-mode of X, or if X is an integer constant
1480 with VOIDmode then MODE is the mode with which X will be used.
1482 Here we are satisfied to find an expression whose tree structure
1483 looks like X. */
1485 static struct table_elt *
1486 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1488 struct table_elt *p;
1490 for (p = table[hash]; p; p = p->next_same_hash)
1491 if (mode == p->mode && ((x == p->exp && REG_P (x))
1492 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1493 return p;
1495 return 0;
1498 /* Like `lookup' but don't care whether the table element uses invalid regs.
1499 Also ignore discrepancies in the machine mode of a register. */
1501 static struct table_elt *
1502 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1504 struct table_elt *p;
1506 if (REG_P (x))
1508 unsigned int regno = REGNO (x);
1510 /* Don't check the machine mode when comparing registers;
1511 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1512 for (p = table[hash]; p; p = p->next_same_hash)
1513 if (REG_P (p->exp)
1514 && REGNO (p->exp) == regno)
1515 return p;
1517 else
1519 for (p = table[hash]; p; p = p->next_same_hash)
1520 if (mode == p->mode
1521 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1522 return p;
1525 return 0;
1528 /* Look for an expression equivalent to X and with code CODE.
1529 If one is found, return that expression. */
1531 static rtx
1532 lookup_as_function (rtx x, enum rtx_code code)
1534 struct table_elt *p
1535 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1537 if (p == 0)
1538 return 0;
1540 for (p = p->first_same_value; p; p = p->next_same_value)
1541 if (GET_CODE (p->exp) == code
1542 /* Make sure this is a valid entry in the table. */
1543 && exp_equiv_p (p->exp, p->exp, 1, false))
1544 return p->exp;
1546 return 0;
1549 /* Insert X in the hash table, assuming HASH is its hash code and
1550 CLASSP is an element of the class it should go in (or 0 if a new
1551 class should be made). COST is the code of X and reg_cost is the
1552 cost of registers in X. It is inserted at the proper position to
1553 keep the class in the order cheapest first.
1555 MODE is the machine-mode of X, or if X is an integer constant
1556 with VOIDmode then MODE is the mode with which X will be used.
1558 For elements of equal cheapness, the most recent one
1559 goes in front, except that the first element in the list
1560 remains first unless a cheaper element is added. The order of
1561 pseudo-registers does not matter, as canon_reg will be called to
1562 find the cheapest when a register is retrieved from the table.
1564 The in_memory field in the hash table element is set to 0.
1565 The caller must set it nonzero if appropriate.
1567 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1568 and if insert_regs returns a nonzero value
1569 you must then recompute its hash code before calling here.
1571 If necessary, update table showing constant values of quantities. */
1573 static struct table_elt *
1574 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1575 enum machine_mode mode, int cost, int reg_cost)
1577 struct table_elt *elt;
1579 /* If X is a register and we haven't made a quantity for it,
1580 something is wrong. */
1581 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1583 /* If X is a hard register, show it is being put in the table. */
1584 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1585 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1587 /* Put an element for X into the right hash bucket. */
1589 elt = free_element_chain;
1590 if (elt)
1591 free_element_chain = elt->next_same_hash;
1592 else
1593 elt = XNEW (struct table_elt);
1595 elt->exp = x;
1596 elt->canon_exp = NULL_RTX;
1597 elt->cost = cost;
1598 elt->regcost = reg_cost;
1599 elt->next_same_value = 0;
1600 elt->prev_same_value = 0;
1601 elt->next_same_hash = table[hash];
1602 elt->prev_same_hash = 0;
1603 elt->related_value = 0;
1604 elt->in_memory = 0;
1605 elt->mode = mode;
1606 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1608 if (table[hash])
1609 table[hash]->prev_same_hash = elt;
1610 table[hash] = elt;
1612 /* Put it into the proper value-class. */
1613 if (classp)
1615 classp = classp->first_same_value;
1616 if (CHEAPER (elt, classp))
1617 /* Insert at the head of the class. */
1619 struct table_elt *p;
1620 elt->next_same_value = classp;
1621 classp->prev_same_value = elt;
1622 elt->first_same_value = elt;
1624 for (p = classp; p; p = p->next_same_value)
1625 p->first_same_value = elt;
1627 else
1629 /* Insert not at head of the class. */
1630 /* Put it after the last element cheaper than X. */
1631 struct table_elt *p, *next;
1633 for (p = classp;
1634 (next = p->next_same_value) && CHEAPER (next, elt);
1635 p = next)
1638 /* Put it after P and before NEXT. */
1639 elt->next_same_value = next;
1640 if (next)
1641 next->prev_same_value = elt;
1643 elt->prev_same_value = p;
1644 p->next_same_value = elt;
1645 elt->first_same_value = classp;
1648 else
1649 elt->first_same_value = elt;
1651 /* If this is a constant being set equivalent to a register or a register
1652 being set equivalent to a constant, note the constant equivalence.
1654 If this is a constant, it cannot be equivalent to a different constant,
1655 and a constant is the only thing that can be cheaper than a register. So
1656 we know the register is the head of the class (before the constant was
1657 inserted).
1659 If this is a register that is not already known equivalent to a
1660 constant, we must check the entire class.
1662 If this is a register that is already known equivalent to an insn,
1663 update the qtys `const_insn' to show that `this_insn' is the latest
1664 insn making that quantity equivalent to the constant. */
1666 if (elt->is_const && classp && REG_P (classp->exp)
1667 && !REG_P (x))
1669 int exp_q = REG_QTY (REGNO (classp->exp));
1670 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1672 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1673 exp_ent->const_insn = this_insn;
1676 else if (REG_P (x)
1677 && classp
1678 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1679 && ! elt->is_const)
1681 struct table_elt *p;
1683 for (p = classp; p != 0; p = p->next_same_value)
1685 if (p->is_const && !REG_P (p->exp))
1687 int x_q = REG_QTY (REGNO (x));
1688 struct qty_table_elem *x_ent = &qty_table[x_q];
1690 x_ent->const_rtx
1691 = gen_lowpart (GET_MODE (x), p->exp);
1692 x_ent->const_insn = this_insn;
1693 break;
1698 else if (REG_P (x)
1699 && qty_table[REG_QTY (REGNO (x))].const_rtx
1700 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1701 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1703 /* If this is a constant with symbolic value,
1704 and it has a term with an explicit integer value,
1705 link it up with related expressions. */
1706 if (GET_CODE (x) == CONST)
1708 rtx subexp = get_related_value (x);
1709 unsigned subhash;
1710 struct table_elt *subelt, *subelt_prev;
1712 if (subexp != 0)
1714 /* Get the integer-free subexpression in the hash table. */
1715 subhash = SAFE_HASH (subexp, mode);
1716 subelt = lookup (subexp, subhash, mode);
1717 if (subelt == 0)
1718 subelt = insert (subexp, NULL, subhash, mode);
1719 /* Initialize SUBELT's circular chain if it has none. */
1720 if (subelt->related_value == 0)
1721 subelt->related_value = subelt;
1722 /* Find the element in the circular chain that precedes SUBELT. */
1723 subelt_prev = subelt;
1724 while (subelt_prev->related_value != subelt)
1725 subelt_prev = subelt_prev->related_value;
1726 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1727 This way the element that follows SUBELT is the oldest one. */
1728 elt->related_value = subelt_prev->related_value;
1729 subelt_prev->related_value = elt;
1733 return elt;
1736 /* Wrap insert_with_costs by passing the default costs. */
1738 static struct table_elt *
1739 insert (rtx x, struct table_elt *classp, unsigned int hash,
1740 enum machine_mode mode)
1742 return
1743 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1747 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1748 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1749 the two classes equivalent.
1751 CLASS1 will be the surviving class; CLASS2 should not be used after this
1752 call.
1754 Any invalid entries in CLASS2 will not be copied. */
1756 static void
1757 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1759 struct table_elt *elt, *next, *new_elt;
1761 /* Ensure we start with the head of the classes. */
1762 class1 = class1->first_same_value;
1763 class2 = class2->first_same_value;
1765 /* If they were already equal, forget it. */
1766 if (class1 == class2)
1767 return;
1769 for (elt = class2; elt; elt = next)
1771 unsigned int hash;
1772 rtx exp = elt->exp;
1773 enum machine_mode mode = elt->mode;
1775 next = elt->next_same_value;
1777 /* Remove old entry, make a new one in CLASS1's class.
1778 Don't do this for invalid entries as we cannot find their
1779 hash code (it also isn't necessary). */
1780 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1782 bool need_rehash = false;
1784 hash_arg_in_memory = 0;
1785 hash = HASH (exp, mode);
1787 if (REG_P (exp))
1789 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1790 delete_reg_equiv (REGNO (exp));
1793 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1794 remove_pseudo_from_table (exp, hash);
1795 else
1796 remove_from_table (elt, hash);
1798 if (insert_regs (exp, class1, 0) || need_rehash)
1800 rehash_using_reg (exp);
1801 hash = HASH (exp, mode);
1803 new_elt = insert (exp, class1, hash, mode);
1804 new_elt->in_memory = hash_arg_in_memory;
1809 /* Flush the entire hash table. */
1811 static void
1812 flush_hash_table (void)
1814 int i;
1815 struct table_elt *p;
1817 for (i = 0; i < HASH_SIZE; i++)
1818 for (p = table[i]; p; p = table[i])
1820 /* Note that invalidate can remove elements
1821 after P in the current hash chain. */
1822 if (REG_P (p->exp))
1823 invalidate (p->exp, VOIDmode);
1824 else
1825 remove_from_table (p, i);
1829 /* Function called for each rtx to check whether true dependence exist. */
1830 struct check_dependence_data
1832 enum machine_mode mode;
1833 rtx exp;
1834 rtx addr;
1837 static int
1838 check_dependence (rtx *x, void *data)
1840 struct check_dependence_data *d = (struct check_dependence_data *) data;
1841 if (*x && MEM_P (*x))
1842 return canon_true_dependence (d->exp, d->mode, d->addr, *x, NULL_RTX);
1843 else
1844 return 0;
1847 /* Remove from the hash table, or mark as invalid, all expressions whose
1848 values could be altered by storing in X. X is a register, a subreg, or
1849 a memory reference with nonvarying address (because, when a memory
1850 reference with a varying address is stored in, all memory references are
1851 removed by invalidate_memory so specific invalidation is superfluous).
1852 FULL_MODE, if not VOIDmode, indicates that this much should be
1853 invalidated instead of just the amount indicated by the mode of X. This
1854 is only used for bitfield stores into memory.
1856 A nonvarying address may be just a register or just a symbol reference,
1857 or it may be either of those plus a numeric offset. */
1859 static void
1860 invalidate (rtx x, enum machine_mode full_mode)
1862 int i;
1863 struct table_elt *p;
1864 rtx addr;
1866 switch (GET_CODE (x))
1868 case REG:
1870 /* If X is a register, dependencies on its contents are recorded
1871 through the qty number mechanism. Just change the qty number of
1872 the register, mark it as invalid for expressions that refer to it,
1873 and remove it itself. */
1874 unsigned int regno = REGNO (x);
1875 unsigned int hash = HASH (x, GET_MODE (x));
1877 /* Remove REGNO from any quantity list it might be on and indicate
1878 that its value might have changed. If it is a pseudo, remove its
1879 entry from the hash table.
1881 For a hard register, we do the first two actions above for any
1882 additional hard registers corresponding to X. Then, if any of these
1883 registers are in the table, we must remove any REG entries that
1884 overlap these registers. */
1886 delete_reg_equiv (regno);
1887 REG_TICK (regno)++;
1888 SUBREG_TICKED (regno) = -1;
1890 if (regno >= FIRST_PSEUDO_REGISTER)
1891 remove_pseudo_from_table (x, hash);
1892 else
1894 HOST_WIDE_INT in_table
1895 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1896 unsigned int endregno = END_HARD_REGNO (x);
1897 unsigned int tregno, tendregno, rn;
1898 struct table_elt *p, *next;
1900 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1902 for (rn = regno + 1; rn < endregno; rn++)
1904 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1905 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1906 delete_reg_equiv (rn);
1907 REG_TICK (rn)++;
1908 SUBREG_TICKED (rn) = -1;
1911 if (in_table)
1912 for (hash = 0; hash < HASH_SIZE; hash++)
1913 for (p = table[hash]; p; p = next)
1915 next = p->next_same_hash;
1917 if (!REG_P (p->exp)
1918 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1919 continue;
1921 tregno = REGNO (p->exp);
1922 tendregno = END_HARD_REGNO (p->exp);
1923 if (tendregno > regno && tregno < endregno)
1924 remove_from_table (p, hash);
1928 return;
1930 case SUBREG:
1931 invalidate (SUBREG_REG (x), VOIDmode);
1932 return;
1934 case PARALLEL:
1935 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1936 invalidate (XVECEXP (x, 0, i), VOIDmode);
1937 return;
1939 case EXPR_LIST:
1940 /* This is part of a disjoint return value; extract the location in
1941 question ignoring the offset. */
1942 invalidate (XEXP (x, 0), VOIDmode);
1943 return;
1945 case MEM:
1946 addr = canon_rtx (get_addr (XEXP (x, 0)));
1947 /* Calculate the canonical version of X here so that
1948 true_dependence doesn't generate new RTL for X on each call. */
1949 x = canon_rtx (x);
1951 /* Remove all hash table elements that refer to overlapping pieces of
1952 memory. */
1953 if (full_mode == VOIDmode)
1954 full_mode = GET_MODE (x);
1956 for (i = 0; i < HASH_SIZE; i++)
1958 struct table_elt *next;
1960 for (p = table[i]; p; p = next)
1962 next = p->next_same_hash;
1963 if (p->in_memory)
1965 struct check_dependence_data d;
1967 /* Just canonicalize the expression once;
1968 otherwise each time we call invalidate
1969 true_dependence will canonicalize the
1970 expression again. */
1971 if (!p->canon_exp)
1972 p->canon_exp = canon_rtx (p->exp);
1973 d.exp = x;
1974 d.addr = addr;
1975 d.mode = full_mode;
1976 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1977 remove_from_table (p, i);
1981 return;
1983 default:
1984 gcc_unreachable ();
1988 /* Remove all expressions that refer to register REGNO,
1989 since they are already invalid, and we are about to
1990 mark that register valid again and don't want the old
1991 expressions to reappear as valid. */
1993 static void
1994 remove_invalid_refs (unsigned int regno)
1996 unsigned int i;
1997 struct table_elt *p, *next;
1999 for (i = 0; i < HASH_SIZE; i++)
2000 for (p = table[i]; p; p = next)
2002 next = p->next_same_hash;
2003 if (!REG_P (p->exp)
2004 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2005 remove_from_table (p, i);
2009 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2010 and mode MODE. */
2011 static void
2012 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2013 enum machine_mode mode)
2015 unsigned int i;
2016 struct table_elt *p, *next;
2017 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2019 for (i = 0; i < HASH_SIZE; i++)
2020 for (p = table[i]; p; p = next)
2022 rtx exp = p->exp;
2023 next = p->next_same_hash;
2025 if (!REG_P (exp)
2026 && (GET_CODE (exp) != SUBREG
2027 || !REG_P (SUBREG_REG (exp))
2028 || REGNO (SUBREG_REG (exp)) != regno
2029 || (((SUBREG_BYTE (exp)
2030 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2031 && SUBREG_BYTE (exp) <= end))
2032 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2033 remove_from_table (p, i);
2037 /* Recompute the hash codes of any valid entries in the hash table that
2038 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2040 This is called when we make a jump equivalence. */
2042 static void
2043 rehash_using_reg (rtx x)
2045 unsigned int i;
2046 struct table_elt *p, *next;
2047 unsigned hash;
2049 if (GET_CODE (x) == SUBREG)
2050 x = SUBREG_REG (x);
2052 /* If X is not a register or if the register is known not to be in any
2053 valid entries in the table, we have no work to do. */
2055 if (!REG_P (x)
2056 || REG_IN_TABLE (REGNO (x)) < 0
2057 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2058 return;
2060 /* Scan all hash chains looking for valid entries that mention X.
2061 If we find one and it is in the wrong hash chain, move it. */
2063 for (i = 0; i < HASH_SIZE; i++)
2064 for (p = table[i]; p; p = next)
2066 next = p->next_same_hash;
2067 if (reg_mentioned_p (x, p->exp)
2068 && exp_equiv_p (p->exp, p->exp, 1, false)
2069 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2071 if (p->next_same_hash)
2072 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2074 if (p->prev_same_hash)
2075 p->prev_same_hash->next_same_hash = p->next_same_hash;
2076 else
2077 table[i] = p->next_same_hash;
2079 p->next_same_hash = table[hash];
2080 p->prev_same_hash = 0;
2081 if (table[hash])
2082 table[hash]->prev_same_hash = p;
2083 table[hash] = p;
2088 /* Remove from the hash table any expression that is a call-clobbered
2089 register. Also update their TICK values. */
2091 static void
2092 invalidate_for_call (void)
2094 unsigned int regno, endregno;
2095 unsigned int i;
2096 unsigned hash;
2097 struct table_elt *p, *next;
2098 int in_table = 0;
2099 hard_reg_set_iterator hrsi;
2101 /* Go through all the hard registers. For each that is clobbered in
2102 a CALL_INSN, remove the register from quantity chains and update
2103 reg_tick if defined. Also see if any of these registers is currently
2104 in the table. */
2105 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2107 delete_reg_equiv (regno);
2108 if (REG_TICK (regno) >= 0)
2110 REG_TICK (regno)++;
2111 SUBREG_TICKED (regno) = -1;
2113 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2116 /* In the case where we have no call-clobbered hard registers in the
2117 table, we are done. Otherwise, scan the table and remove any
2118 entry that overlaps a call-clobbered register. */
2120 if (in_table)
2121 for (hash = 0; hash < HASH_SIZE; hash++)
2122 for (p = table[hash]; p; p = next)
2124 next = p->next_same_hash;
2126 if (!REG_P (p->exp)
2127 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2128 continue;
2130 regno = REGNO (p->exp);
2131 endregno = END_HARD_REGNO (p->exp);
2133 for (i = regno; i < endregno; i++)
2134 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2136 remove_from_table (p, hash);
2137 break;
2142 /* Given an expression X of type CONST,
2143 and ELT which is its table entry (or 0 if it
2144 is not in the hash table),
2145 return an alternate expression for X as a register plus integer.
2146 If none can be found, return 0. */
2148 static rtx
2149 use_related_value (rtx x, struct table_elt *elt)
2151 struct table_elt *relt = 0;
2152 struct table_elt *p, *q;
2153 HOST_WIDE_INT offset;
2155 /* First, is there anything related known?
2156 If we have a table element, we can tell from that.
2157 Otherwise, must look it up. */
2159 if (elt != 0 && elt->related_value != 0)
2160 relt = elt;
2161 else if (elt == 0 && GET_CODE (x) == CONST)
2163 rtx subexp = get_related_value (x);
2164 if (subexp != 0)
2165 relt = lookup (subexp,
2166 SAFE_HASH (subexp, GET_MODE (subexp)),
2167 GET_MODE (subexp));
2170 if (relt == 0)
2171 return 0;
2173 /* Search all related table entries for one that has an
2174 equivalent register. */
2176 p = relt;
2177 while (1)
2179 /* This loop is strange in that it is executed in two different cases.
2180 The first is when X is already in the table. Then it is searching
2181 the RELATED_VALUE list of X's class (RELT). The second case is when
2182 X is not in the table. Then RELT points to a class for the related
2183 value.
2185 Ensure that, whatever case we are in, that we ignore classes that have
2186 the same value as X. */
2188 if (rtx_equal_p (x, p->exp))
2189 q = 0;
2190 else
2191 for (q = p->first_same_value; q; q = q->next_same_value)
2192 if (REG_P (q->exp))
2193 break;
2195 if (q)
2196 break;
2198 p = p->related_value;
2200 /* We went all the way around, so there is nothing to be found.
2201 Alternatively, perhaps RELT was in the table for some other reason
2202 and it has no related values recorded. */
2203 if (p == relt || p == 0)
2204 break;
2207 if (q == 0)
2208 return 0;
2210 offset = (get_integer_term (x) - get_integer_term (p->exp));
2211 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2212 return plus_constant (q->mode, q->exp, offset);
2216 /* Hash a string. Just add its bytes up. */
2217 static inline unsigned
2218 hash_rtx_string (const char *ps)
2220 unsigned hash = 0;
2221 const unsigned char *p = (const unsigned char *) ps;
2223 if (p)
2224 while (*p)
2225 hash += *p++;
2227 return hash;
2230 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2231 When the callback returns true, we continue with the new rtx. */
2233 unsigned
2234 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2235 int *do_not_record_p, int *hash_arg_in_memory_p,
2236 bool have_reg_qty, hash_rtx_callback_function cb)
2238 int i, j;
2239 unsigned hash = 0;
2240 enum rtx_code code;
2241 const char *fmt;
2242 enum machine_mode newmode;
2243 rtx newx;
2245 /* Used to turn recursion into iteration. We can't rely on GCC's
2246 tail-recursion elimination since we need to keep accumulating values
2247 in HASH. */
2248 repeat:
2249 if (x == 0)
2250 return hash;
2252 /* Invoke the callback first. */
2253 if (cb != NULL
2254 && ((*cb) (x, mode, &newx, &newmode)))
2256 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2257 hash_arg_in_memory_p, have_reg_qty, cb);
2258 return hash;
2261 code = GET_CODE (x);
2262 switch (code)
2264 case REG:
2266 unsigned int regno = REGNO (x);
2268 if (do_not_record_p && !reload_completed)
2270 /* On some machines, we can't record any non-fixed hard register,
2271 because extending its life will cause reload problems. We
2272 consider ap, fp, sp, gp to be fixed for this purpose.
2274 We also consider CCmode registers to be fixed for this purpose;
2275 failure to do so leads to failure to simplify 0<100 type of
2276 conditionals.
2278 On all machines, we can't record any global registers.
2279 Nor should we record any register that is in a small
2280 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2281 bool record;
2283 if (regno >= FIRST_PSEUDO_REGISTER)
2284 record = true;
2285 else if (x == frame_pointer_rtx
2286 || x == hard_frame_pointer_rtx
2287 || x == arg_pointer_rtx
2288 || x == stack_pointer_rtx
2289 || x == pic_offset_table_rtx)
2290 record = true;
2291 else if (global_regs[regno])
2292 record = false;
2293 else if (fixed_regs[regno])
2294 record = true;
2295 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2296 record = true;
2297 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2298 record = false;
2299 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2300 record = false;
2301 else
2302 record = true;
2304 if (!record)
2306 *do_not_record_p = 1;
2307 return 0;
2311 hash += ((unsigned int) REG << 7);
2312 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2313 return hash;
2316 /* We handle SUBREG of a REG specially because the underlying
2317 reg changes its hash value with every value change; we don't
2318 want to have to forget unrelated subregs when one subreg changes. */
2319 case SUBREG:
2321 if (REG_P (SUBREG_REG (x)))
2323 hash += (((unsigned int) SUBREG << 7)
2324 + REGNO (SUBREG_REG (x))
2325 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2326 return hash;
2328 break;
2331 case CONST_INT:
2332 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2333 + (unsigned int) INTVAL (x));
2334 return hash;
2336 case CONST_DOUBLE:
2337 /* This is like the general case, except that it only counts
2338 the integers representing the constant. */
2339 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2340 if (GET_MODE (x) != VOIDmode)
2341 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2342 else
2343 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2344 + (unsigned int) CONST_DOUBLE_HIGH (x));
2345 return hash;
2347 case CONST_FIXED:
2348 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2349 hash += fixed_hash (CONST_FIXED_VALUE (x));
2350 return hash;
2352 case CONST_VECTOR:
2354 int units;
2355 rtx elt;
2357 units = CONST_VECTOR_NUNITS (x);
2359 for (i = 0; i < units; ++i)
2361 elt = CONST_VECTOR_ELT (x, i);
2362 hash += hash_rtx_cb (elt, GET_MODE (elt),
2363 do_not_record_p, hash_arg_in_memory_p,
2364 have_reg_qty, cb);
2367 return hash;
2370 /* Assume there is only one rtx object for any given label. */
2371 case LABEL_REF:
2372 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2373 differences and differences between each stage's debugging dumps. */
2374 hash += (((unsigned int) LABEL_REF << 7)
2375 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2376 return hash;
2378 case SYMBOL_REF:
2380 /* Don't hash on the symbol's address to avoid bootstrap differences.
2381 Different hash values may cause expressions to be recorded in
2382 different orders and thus different registers to be used in the
2383 final assembler. This also avoids differences in the dump files
2384 between various stages. */
2385 unsigned int h = 0;
2386 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2388 while (*p)
2389 h += (h << 7) + *p++; /* ??? revisit */
2391 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2392 return hash;
2395 case MEM:
2396 /* We don't record if marked volatile or if BLKmode since we don't
2397 know the size of the move. */
2398 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2400 *do_not_record_p = 1;
2401 return 0;
2403 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2404 *hash_arg_in_memory_p = 1;
2406 /* Now that we have already found this special case,
2407 might as well speed it up as much as possible. */
2408 hash += (unsigned) MEM;
2409 x = XEXP (x, 0);
2410 goto repeat;
2412 case USE:
2413 /* A USE that mentions non-volatile memory needs special
2414 handling since the MEM may be BLKmode which normally
2415 prevents an entry from being made. Pure calls are
2416 marked by a USE which mentions BLKmode memory.
2417 See calls.c:emit_call_1. */
2418 if (MEM_P (XEXP (x, 0))
2419 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2421 hash += (unsigned) USE;
2422 x = XEXP (x, 0);
2424 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2425 *hash_arg_in_memory_p = 1;
2427 /* Now that we have already found this special case,
2428 might as well speed it up as much as possible. */
2429 hash += (unsigned) MEM;
2430 x = XEXP (x, 0);
2431 goto repeat;
2433 break;
2435 case PRE_DEC:
2436 case PRE_INC:
2437 case POST_DEC:
2438 case POST_INC:
2439 case PRE_MODIFY:
2440 case POST_MODIFY:
2441 case PC:
2442 case CC0:
2443 case CALL:
2444 case UNSPEC_VOLATILE:
2445 if (do_not_record_p) {
2446 *do_not_record_p = 1;
2447 return 0;
2449 else
2450 return hash;
2451 break;
2453 case ASM_OPERANDS:
2454 if (do_not_record_p && MEM_VOLATILE_P (x))
2456 *do_not_record_p = 1;
2457 return 0;
2459 else
2461 /* We don't want to take the filename and line into account. */
2462 hash += (unsigned) code + (unsigned) GET_MODE (x)
2463 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2464 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2465 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2467 if (ASM_OPERANDS_INPUT_LENGTH (x))
2469 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2471 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2472 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2473 do_not_record_p, hash_arg_in_memory_p,
2474 have_reg_qty, cb)
2475 + hash_rtx_string
2476 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2479 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2480 x = ASM_OPERANDS_INPUT (x, 0);
2481 mode = GET_MODE (x);
2482 goto repeat;
2485 return hash;
2487 break;
2489 default:
2490 break;
2493 i = GET_RTX_LENGTH (code) - 1;
2494 hash += (unsigned) code + (unsigned) GET_MODE (x);
2495 fmt = GET_RTX_FORMAT (code);
2496 for (; i >= 0; i--)
2498 switch (fmt[i])
2500 case 'e':
2501 /* If we are about to do the last recursive call
2502 needed at this level, change it into iteration.
2503 This function is called enough to be worth it. */
2504 if (i == 0)
2506 x = XEXP (x, i);
2507 goto repeat;
2510 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2511 hash_arg_in_memory_p,
2512 have_reg_qty, cb);
2513 break;
2515 case 'E':
2516 for (j = 0; j < XVECLEN (x, i); j++)
2517 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2518 hash_arg_in_memory_p,
2519 have_reg_qty, cb);
2520 break;
2522 case 's':
2523 hash += hash_rtx_string (XSTR (x, i));
2524 break;
2526 case 'i':
2527 hash += (unsigned int) XINT (x, i);
2528 break;
2530 case '0': case 't':
2531 /* Unused. */
2532 break;
2534 default:
2535 gcc_unreachable ();
2539 return hash;
2542 /* Hash an rtx. We are careful to make sure the value is never negative.
2543 Equivalent registers hash identically.
2544 MODE is used in hashing for CONST_INTs only;
2545 otherwise the mode of X is used.
2547 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2549 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2550 a MEM rtx which does not have the MEM_READONLY_P flag set.
2552 Note that cse_insn knows that the hash code of a MEM expression
2553 is just (int) MEM plus the hash code of the address. */
2555 unsigned
2556 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2557 int *hash_arg_in_memory_p, bool have_reg_qty)
2559 return hash_rtx_cb (x, mode, do_not_record_p,
2560 hash_arg_in_memory_p, have_reg_qty, NULL);
2563 /* Hash an rtx X for cse via hash_rtx.
2564 Stores 1 in do_not_record if any subexpression is volatile.
2565 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2566 does not have the MEM_READONLY_P flag set. */
2568 static inline unsigned
2569 canon_hash (rtx x, enum machine_mode mode)
2571 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2574 /* Like canon_hash but with no side effects, i.e. do_not_record
2575 and hash_arg_in_memory are not changed. */
2577 static inline unsigned
2578 safe_hash (rtx x, enum machine_mode mode)
2580 int dummy_do_not_record;
2581 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2584 /* Return 1 iff X and Y would canonicalize into the same thing,
2585 without actually constructing the canonicalization of either one.
2586 If VALIDATE is nonzero,
2587 we assume X is an expression being processed from the rtl
2588 and Y was found in the hash table. We check register refs
2589 in Y for being marked as valid.
2591 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2594 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2596 int i, j;
2597 enum rtx_code code;
2598 const char *fmt;
2600 /* Note: it is incorrect to assume an expression is equivalent to itself
2601 if VALIDATE is nonzero. */
2602 if (x == y && !validate)
2603 return 1;
2605 if (x == 0 || y == 0)
2606 return x == y;
2608 code = GET_CODE (x);
2609 if (code != GET_CODE (y))
2610 return 0;
2612 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2613 if (GET_MODE (x) != GET_MODE (y))
2614 return 0;
2616 /* MEMs referring to different address space are not equivalent. */
2617 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2618 return 0;
2620 switch (code)
2622 case PC:
2623 case CC0:
2624 CASE_CONST_UNIQUE:
2625 return x == y;
2627 case LABEL_REF:
2628 return XEXP (x, 0) == XEXP (y, 0);
2630 case SYMBOL_REF:
2631 return XSTR (x, 0) == XSTR (y, 0);
2633 case REG:
2634 if (for_gcse)
2635 return REGNO (x) == REGNO (y);
2636 else
2638 unsigned int regno = REGNO (y);
2639 unsigned int i;
2640 unsigned int endregno = END_REGNO (y);
2642 /* If the quantities are not the same, the expressions are not
2643 equivalent. If there are and we are not to validate, they
2644 are equivalent. Otherwise, ensure all regs are up-to-date. */
2646 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2647 return 0;
2649 if (! validate)
2650 return 1;
2652 for (i = regno; i < endregno; i++)
2653 if (REG_IN_TABLE (i) != REG_TICK (i))
2654 return 0;
2656 return 1;
2659 case MEM:
2660 if (for_gcse)
2662 /* A volatile mem should not be considered equivalent to any
2663 other. */
2664 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2665 return 0;
2667 /* Can't merge two expressions in different alias sets, since we
2668 can decide that the expression is transparent in a block when
2669 it isn't, due to it being set with the different alias set.
2671 Also, can't merge two expressions with different MEM_ATTRS.
2672 They could e.g. be two different entities allocated into the
2673 same space on the stack (see e.g. PR25130). In that case, the
2674 MEM addresses can be the same, even though the two MEMs are
2675 absolutely not equivalent.
2677 But because really all MEM attributes should be the same for
2678 equivalent MEMs, we just use the invariant that MEMs that have
2679 the same attributes share the same mem_attrs data structure. */
2680 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2681 return 0;
2683 break;
2685 /* For commutative operations, check both orders. */
2686 case PLUS:
2687 case MULT:
2688 case AND:
2689 case IOR:
2690 case XOR:
2691 case NE:
2692 case EQ:
2693 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2694 validate, for_gcse)
2695 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2696 validate, for_gcse))
2697 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2698 validate, for_gcse)
2699 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2700 validate, for_gcse)));
2702 case ASM_OPERANDS:
2703 /* We don't use the generic code below because we want to
2704 disregard filename and line numbers. */
2706 /* A volatile asm isn't equivalent to any other. */
2707 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2708 return 0;
2710 if (GET_MODE (x) != GET_MODE (y)
2711 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2712 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2713 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2714 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2715 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2716 return 0;
2718 if (ASM_OPERANDS_INPUT_LENGTH (x))
2720 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2721 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2722 ASM_OPERANDS_INPUT (y, i),
2723 validate, for_gcse)
2724 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2725 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2726 return 0;
2729 return 1;
2731 default:
2732 break;
2735 /* Compare the elements. If any pair of corresponding elements
2736 fail to match, return 0 for the whole thing. */
2738 fmt = GET_RTX_FORMAT (code);
2739 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2741 switch (fmt[i])
2743 case 'e':
2744 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2745 validate, for_gcse))
2746 return 0;
2747 break;
2749 case 'E':
2750 if (XVECLEN (x, i) != XVECLEN (y, i))
2751 return 0;
2752 for (j = 0; j < XVECLEN (x, i); j++)
2753 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2754 validate, for_gcse))
2755 return 0;
2756 break;
2758 case 's':
2759 if (strcmp (XSTR (x, i), XSTR (y, i)))
2760 return 0;
2761 break;
2763 case 'i':
2764 if (XINT (x, i) != XINT (y, i))
2765 return 0;
2766 break;
2768 case 'w':
2769 if (XWINT (x, i) != XWINT (y, i))
2770 return 0;
2771 break;
2773 case '0':
2774 case 't':
2775 break;
2777 default:
2778 gcc_unreachable ();
2782 return 1;
2785 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2786 the result if necessary. INSN is as for canon_reg. */
2788 static void
2789 validate_canon_reg (rtx *xloc, rtx insn)
2791 if (*xloc)
2793 rtx new_rtx = canon_reg (*xloc, insn);
2795 /* If replacing pseudo with hard reg or vice versa, ensure the
2796 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2797 gcc_assert (insn && new_rtx);
2798 validate_change (insn, xloc, new_rtx, 1);
2802 /* Canonicalize an expression:
2803 replace each register reference inside it
2804 with the "oldest" equivalent register.
2806 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2807 after we make our substitution. The calls are made with IN_GROUP nonzero
2808 so apply_change_group must be called upon the outermost return from this
2809 function (unless INSN is zero). The result of apply_change_group can
2810 generally be discarded since the changes we are making are optional. */
2812 static rtx
2813 canon_reg (rtx x, rtx insn)
2815 int i;
2816 enum rtx_code code;
2817 const char *fmt;
2819 if (x == 0)
2820 return x;
2822 code = GET_CODE (x);
2823 switch (code)
2825 case PC:
2826 case CC0:
2827 case CONST:
2828 CASE_CONST_ANY:
2829 case SYMBOL_REF:
2830 case LABEL_REF:
2831 case ADDR_VEC:
2832 case ADDR_DIFF_VEC:
2833 return x;
2835 case REG:
2837 int first;
2838 int q;
2839 struct qty_table_elem *ent;
2841 /* Never replace a hard reg, because hard regs can appear
2842 in more than one machine mode, and we must preserve the mode
2843 of each occurrence. Also, some hard regs appear in
2844 MEMs that are shared and mustn't be altered. Don't try to
2845 replace any reg that maps to a reg of class NO_REGS. */
2846 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2847 || ! REGNO_QTY_VALID_P (REGNO (x)))
2848 return x;
2850 q = REG_QTY (REGNO (x));
2851 ent = &qty_table[q];
2852 first = ent->first_reg;
2853 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2854 : REGNO_REG_CLASS (first) == NO_REGS ? x
2855 : gen_rtx_REG (ent->mode, first));
2858 default:
2859 break;
2862 fmt = GET_RTX_FORMAT (code);
2863 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2865 int j;
2867 if (fmt[i] == 'e')
2868 validate_canon_reg (&XEXP (x, i), insn);
2869 else if (fmt[i] == 'E')
2870 for (j = 0; j < XVECLEN (x, i); j++)
2871 validate_canon_reg (&XVECEXP (x, i, j), insn);
2874 return x;
2877 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2878 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2879 what values are being compared.
2881 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2882 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2883 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2884 compared to produce cc0.
2886 The return value is the comparison operator and is either the code of
2887 A or the code corresponding to the inverse of the comparison. */
2889 static enum rtx_code
2890 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2891 enum machine_mode *pmode1, enum machine_mode *pmode2)
2893 rtx arg1, arg2;
2894 struct pointer_set_t *visited = NULL;
2895 /* Set nonzero when we find something of interest. */
2896 rtx x = NULL;
2898 arg1 = *parg1, arg2 = *parg2;
2900 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2902 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2904 int reverse_code = 0;
2905 struct table_elt *p = 0;
2907 /* Remember state from previous iteration. */
2908 if (x)
2910 if (!visited)
2911 visited = pointer_set_create ();
2912 pointer_set_insert (visited, x);
2913 x = 0;
2916 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2917 On machines with CC0, this is the only case that can occur, since
2918 fold_rtx will return the COMPARE or item being compared with zero
2919 when given CC0. */
2921 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2922 x = arg1;
2924 /* If ARG1 is a comparison operator and CODE is testing for
2925 STORE_FLAG_VALUE, get the inner arguments. */
2927 else if (COMPARISON_P (arg1))
2929 #ifdef FLOAT_STORE_FLAG_VALUE
2930 REAL_VALUE_TYPE fsfv;
2931 #endif
2933 if (code == NE
2934 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2935 && code == LT && STORE_FLAG_VALUE == -1)
2936 #ifdef FLOAT_STORE_FLAG_VALUE
2937 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2938 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2939 REAL_VALUE_NEGATIVE (fsfv)))
2940 #endif
2942 x = arg1;
2943 else if (code == EQ
2944 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2945 && code == GE && STORE_FLAG_VALUE == -1)
2946 #ifdef FLOAT_STORE_FLAG_VALUE
2947 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2948 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2949 REAL_VALUE_NEGATIVE (fsfv)))
2950 #endif
2952 x = arg1, reverse_code = 1;
2955 /* ??? We could also check for
2957 (ne (and (eq (...) (const_int 1))) (const_int 0))
2959 and related forms, but let's wait until we see them occurring. */
2961 if (x == 0)
2962 /* Look up ARG1 in the hash table and see if it has an equivalence
2963 that lets us see what is being compared. */
2964 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2965 if (p)
2967 p = p->first_same_value;
2969 /* If what we compare is already known to be constant, that is as
2970 good as it gets.
2971 We need to break the loop in this case, because otherwise we
2972 can have an infinite loop when looking at a reg that is known
2973 to be a constant which is the same as a comparison of a reg
2974 against zero which appears later in the insn stream, which in
2975 turn is constant and the same as the comparison of the first reg
2976 against zero... */
2977 if (p->is_const)
2978 break;
2981 for (; p; p = p->next_same_value)
2983 enum machine_mode inner_mode = GET_MODE (p->exp);
2984 #ifdef FLOAT_STORE_FLAG_VALUE
2985 REAL_VALUE_TYPE fsfv;
2986 #endif
2988 /* If the entry isn't valid, skip it. */
2989 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2990 continue;
2992 /* If it's a comparison we've used before, skip it. */
2993 if (visited && pointer_set_contains (visited, p->exp))
2994 continue;
2996 if (GET_CODE (p->exp) == COMPARE
2997 /* Another possibility is that this machine has a compare insn
2998 that includes the comparison code. In that case, ARG1 would
2999 be equivalent to a comparison operation that would set ARG1 to
3000 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3001 ORIG_CODE is the actual comparison being done; if it is an EQ,
3002 we must reverse ORIG_CODE. On machine with a negative value
3003 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3004 || ((code == NE
3005 || (code == LT
3006 && val_signbit_known_set_p (inner_mode,
3007 STORE_FLAG_VALUE))
3008 #ifdef FLOAT_STORE_FLAG_VALUE
3009 || (code == LT
3010 && SCALAR_FLOAT_MODE_P (inner_mode)
3011 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3012 REAL_VALUE_NEGATIVE (fsfv)))
3013 #endif
3015 && COMPARISON_P (p->exp)))
3017 x = p->exp;
3018 break;
3020 else if ((code == EQ
3021 || (code == GE
3022 && val_signbit_known_set_p (inner_mode,
3023 STORE_FLAG_VALUE))
3024 #ifdef FLOAT_STORE_FLAG_VALUE
3025 || (code == GE
3026 && SCALAR_FLOAT_MODE_P (inner_mode)
3027 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3028 REAL_VALUE_NEGATIVE (fsfv)))
3029 #endif
3031 && COMPARISON_P (p->exp))
3033 reverse_code = 1;
3034 x = p->exp;
3035 break;
3038 /* If this non-trapping address, e.g. fp + constant, the
3039 equivalent is a better operand since it may let us predict
3040 the value of the comparison. */
3041 else if (!rtx_addr_can_trap_p (p->exp))
3043 arg1 = p->exp;
3044 continue;
3048 /* If we didn't find a useful equivalence for ARG1, we are done.
3049 Otherwise, set up for the next iteration. */
3050 if (x == 0)
3051 break;
3053 /* If we need to reverse the comparison, make sure that that is
3054 possible -- we can't necessarily infer the value of GE from LT
3055 with floating-point operands. */
3056 if (reverse_code)
3058 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3059 if (reversed == UNKNOWN)
3060 break;
3061 else
3062 code = reversed;
3064 else if (COMPARISON_P (x))
3065 code = GET_CODE (x);
3066 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3069 /* Return our results. Return the modes from before fold_rtx
3070 because fold_rtx might produce const_int, and then it's too late. */
3071 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3072 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3074 if (visited)
3075 pointer_set_destroy (visited);
3076 return code;
3079 /* If X is a nontrivial arithmetic operation on an argument for which
3080 a constant value can be determined, return the result of operating
3081 on that value, as a constant. Otherwise, return X, possibly with
3082 one or more operands changed to a forward-propagated constant.
3084 If X is a register whose contents are known, we do NOT return
3085 those contents here; equiv_constant is called to perform that task.
3086 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3088 INSN is the insn that we may be modifying. If it is 0, make a copy
3089 of X before modifying it. */
3091 static rtx
3092 fold_rtx (rtx x, rtx insn)
3094 enum rtx_code code;
3095 enum machine_mode mode;
3096 const char *fmt;
3097 int i;
3098 rtx new_rtx = 0;
3099 int changed = 0;
3101 /* Operands of X. */
3102 rtx folded_arg0;
3103 rtx folded_arg1;
3105 /* Constant equivalents of first three operands of X;
3106 0 when no such equivalent is known. */
3107 rtx const_arg0;
3108 rtx const_arg1;
3109 rtx const_arg2;
3111 /* The mode of the first operand of X. We need this for sign and zero
3112 extends. */
3113 enum machine_mode mode_arg0;
3115 if (x == 0)
3116 return x;
3118 /* Try to perform some initial simplifications on X. */
3119 code = GET_CODE (x);
3120 switch (code)
3122 case MEM:
3123 case SUBREG:
3124 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3125 return new_rtx;
3126 return x;
3128 case CONST:
3129 CASE_CONST_ANY:
3130 case SYMBOL_REF:
3131 case LABEL_REF:
3132 case REG:
3133 case PC:
3134 /* No use simplifying an EXPR_LIST
3135 since they are used only for lists of args
3136 in a function call's REG_EQUAL note. */
3137 case EXPR_LIST:
3138 return x;
3140 #ifdef HAVE_cc0
3141 case CC0:
3142 return prev_insn_cc0;
3143 #endif
3145 case ASM_OPERANDS:
3146 if (insn)
3148 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3149 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3150 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3152 return x;
3154 #ifdef NO_FUNCTION_CSE
3155 case CALL:
3156 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3157 return x;
3158 break;
3159 #endif
3161 /* Anything else goes through the loop below. */
3162 default:
3163 break;
3166 mode = GET_MODE (x);
3167 const_arg0 = 0;
3168 const_arg1 = 0;
3169 const_arg2 = 0;
3170 mode_arg0 = VOIDmode;
3172 /* Try folding our operands.
3173 Then see which ones have constant values known. */
3175 fmt = GET_RTX_FORMAT (code);
3176 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3177 if (fmt[i] == 'e')
3179 rtx folded_arg = XEXP (x, i), const_arg;
3180 enum machine_mode mode_arg = GET_MODE (folded_arg);
3182 switch (GET_CODE (folded_arg))
3184 case MEM:
3185 case REG:
3186 case SUBREG:
3187 const_arg = equiv_constant (folded_arg);
3188 break;
3190 case CONST:
3191 CASE_CONST_ANY:
3192 case SYMBOL_REF:
3193 case LABEL_REF:
3194 const_arg = folded_arg;
3195 break;
3197 #ifdef HAVE_cc0
3198 case CC0:
3199 folded_arg = prev_insn_cc0;
3200 mode_arg = prev_insn_cc0_mode;
3201 const_arg = equiv_constant (folded_arg);
3202 break;
3203 #endif
3205 default:
3206 folded_arg = fold_rtx (folded_arg, insn);
3207 const_arg = equiv_constant (folded_arg);
3208 break;
3211 /* For the first three operands, see if the operand
3212 is constant or equivalent to a constant. */
3213 switch (i)
3215 case 0:
3216 folded_arg0 = folded_arg;
3217 const_arg0 = const_arg;
3218 mode_arg0 = mode_arg;
3219 break;
3220 case 1:
3221 folded_arg1 = folded_arg;
3222 const_arg1 = const_arg;
3223 break;
3224 case 2:
3225 const_arg2 = const_arg;
3226 break;
3229 /* Pick the least expensive of the argument and an equivalent constant
3230 argument. */
3231 if (const_arg != 0
3232 && const_arg != folded_arg
3233 && COST_IN (const_arg, code, i) <= COST_IN (folded_arg, code, i)
3235 /* It's not safe to substitute the operand of a conversion
3236 operator with a constant, as the conversion's identity
3237 depends upon the mode of its operand. This optimization
3238 is handled by the call to simplify_unary_operation. */
3239 && (GET_RTX_CLASS (code) != RTX_UNARY
3240 || GET_MODE (const_arg) == mode_arg0
3241 || (code != ZERO_EXTEND
3242 && code != SIGN_EXTEND
3243 && code != TRUNCATE
3244 && code != FLOAT_TRUNCATE
3245 && code != FLOAT_EXTEND
3246 && code != FLOAT
3247 && code != FIX
3248 && code != UNSIGNED_FLOAT
3249 && code != UNSIGNED_FIX)))
3250 folded_arg = const_arg;
3252 if (folded_arg == XEXP (x, i))
3253 continue;
3255 if (insn == NULL_RTX && !changed)
3256 x = copy_rtx (x);
3257 changed = 1;
3258 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3261 if (changed)
3263 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3264 consistent with the order in X. */
3265 if (canonicalize_change_group (insn, x))
3267 rtx tem;
3268 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3269 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3272 apply_change_group ();
3275 /* If X is an arithmetic operation, see if we can simplify it. */
3277 switch (GET_RTX_CLASS (code))
3279 case RTX_UNARY:
3281 /* We can't simplify extension ops unless we know the
3282 original mode. */
3283 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3284 && mode_arg0 == VOIDmode)
3285 break;
3287 new_rtx = simplify_unary_operation (code, mode,
3288 const_arg0 ? const_arg0 : folded_arg0,
3289 mode_arg0);
3291 break;
3293 case RTX_COMPARE:
3294 case RTX_COMM_COMPARE:
3295 /* See what items are actually being compared and set FOLDED_ARG[01]
3296 to those values and CODE to the actual comparison code. If any are
3297 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3298 do anything if both operands are already known to be constant. */
3300 /* ??? Vector mode comparisons are not supported yet. */
3301 if (VECTOR_MODE_P (mode))
3302 break;
3304 if (const_arg0 == 0 || const_arg1 == 0)
3306 struct table_elt *p0, *p1;
3307 rtx true_rtx, false_rtx;
3308 enum machine_mode mode_arg1;
3310 if (SCALAR_FLOAT_MODE_P (mode))
3312 #ifdef FLOAT_STORE_FLAG_VALUE
3313 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3314 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3315 #else
3316 true_rtx = NULL_RTX;
3317 #endif
3318 false_rtx = CONST0_RTX (mode);
3320 else
3322 true_rtx = const_true_rtx;
3323 false_rtx = const0_rtx;
3326 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3327 &mode_arg0, &mode_arg1);
3329 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3330 what kinds of things are being compared, so we can't do
3331 anything with this comparison. */
3333 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3334 break;
3336 const_arg0 = equiv_constant (folded_arg0);
3337 const_arg1 = equiv_constant (folded_arg1);
3339 /* If we do not now have two constants being compared, see
3340 if we can nevertheless deduce some things about the
3341 comparison. */
3342 if (const_arg0 == 0 || const_arg1 == 0)
3344 if (const_arg1 != NULL)
3346 rtx cheapest_simplification;
3347 int cheapest_cost;
3348 rtx simp_result;
3349 struct table_elt *p;
3351 /* See if we can find an equivalent of folded_arg0
3352 that gets us a cheaper expression, possibly a
3353 constant through simplifications. */
3354 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3355 mode_arg0);
3357 if (p != NULL)
3359 cheapest_simplification = x;
3360 cheapest_cost = COST (x);
3362 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3364 int cost;
3366 /* If the entry isn't valid, skip it. */
3367 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3368 continue;
3370 /* Try to simplify using this equivalence. */
3371 simp_result
3372 = simplify_relational_operation (code, mode,
3373 mode_arg0,
3374 p->exp,
3375 const_arg1);
3377 if (simp_result == NULL)
3378 continue;
3380 cost = COST (simp_result);
3381 if (cost < cheapest_cost)
3383 cheapest_cost = cost;
3384 cheapest_simplification = simp_result;
3388 /* If we have a cheaper expression now, use that
3389 and try folding it further, from the top. */
3390 if (cheapest_simplification != x)
3391 return fold_rtx (copy_rtx (cheapest_simplification),
3392 insn);
3396 /* See if the two operands are the same. */
3398 if ((REG_P (folded_arg0)
3399 && REG_P (folded_arg1)
3400 && (REG_QTY (REGNO (folded_arg0))
3401 == REG_QTY (REGNO (folded_arg1))))
3402 || ((p0 = lookup (folded_arg0,
3403 SAFE_HASH (folded_arg0, mode_arg0),
3404 mode_arg0))
3405 && (p1 = lookup (folded_arg1,
3406 SAFE_HASH (folded_arg1, mode_arg0),
3407 mode_arg0))
3408 && p0->first_same_value == p1->first_same_value))
3409 folded_arg1 = folded_arg0;
3411 /* If FOLDED_ARG0 is a register, see if the comparison we are
3412 doing now is either the same as we did before or the reverse
3413 (we only check the reverse if not floating-point). */
3414 else if (REG_P (folded_arg0))
3416 int qty = REG_QTY (REGNO (folded_arg0));
3418 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3420 struct qty_table_elem *ent = &qty_table[qty];
3422 if ((comparison_dominates_p (ent->comparison_code, code)
3423 || (! FLOAT_MODE_P (mode_arg0)
3424 && comparison_dominates_p (ent->comparison_code,
3425 reverse_condition (code))))
3426 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3427 || (const_arg1
3428 && rtx_equal_p (ent->comparison_const,
3429 const_arg1))
3430 || (REG_P (folded_arg1)
3431 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3433 if (comparison_dominates_p (ent->comparison_code, code))
3435 if (true_rtx)
3436 return true_rtx;
3437 else
3438 break;
3440 else
3441 return false_rtx;
3448 /* If we are comparing against zero, see if the first operand is
3449 equivalent to an IOR with a constant. If so, we may be able to
3450 determine the result of this comparison. */
3451 if (const_arg1 == const0_rtx && !const_arg0)
3453 rtx y = lookup_as_function (folded_arg0, IOR);
3454 rtx inner_const;
3456 if (y != 0
3457 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3458 && CONST_INT_P (inner_const)
3459 && INTVAL (inner_const) != 0)
3460 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3464 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3465 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3466 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3467 op0, op1);
3469 break;
3471 case RTX_BIN_ARITH:
3472 case RTX_COMM_ARITH:
3473 switch (code)
3475 case PLUS:
3476 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3477 with that LABEL_REF as its second operand. If so, the result is
3478 the first operand of that MINUS. This handles switches with an
3479 ADDR_DIFF_VEC table. */
3480 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3482 rtx y
3483 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3484 : lookup_as_function (folded_arg0, MINUS);
3486 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3487 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3488 return XEXP (y, 0);
3490 /* Now try for a CONST of a MINUS like the above. */
3491 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3492 : lookup_as_function (folded_arg0, CONST))) != 0
3493 && GET_CODE (XEXP (y, 0)) == MINUS
3494 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3495 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3496 return XEXP (XEXP (y, 0), 0);
3499 /* Likewise if the operands are in the other order. */
3500 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3502 rtx y
3503 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3504 : lookup_as_function (folded_arg1, MINUS);
3506 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3507 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3508 return XEXP (y, 0);
3510 /* Now try for a CONST of a MINUS like the above. */
3511 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3512 : lookup_as_function (folded_arg1, CONST))) != 0
3513 && GET_CODE (XEXP (y, 0)) == MINUS
3514 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3515 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3516 return XEXP (XEXP (y, 0), 0);
3519 /* If second operand is a register equivalent to a negative
3520 CONST_INT, see if we can find a register equivalent to the
3521 positive constant. Make a MINUS if so. Don't do this for
3522 a non-negative constant since we might then alternate between
3523 choosing positive and negative constants. Having the positive
3524 constant previously-used is the more common case. Be sure
3525 the resulting constant is non-negative; if const_arg1 were
3526 the smallest negative number this would overflow: depending
3527 on the mode, this would either just be the same value (and
3528 hence not save anything) or be incorrect. */
3529 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3530 && INTVAL (const_arg1) < 0
3531 /* This used to test
3533 -INTVAL (const_arg1) >= 0
3535 But The Sun V5.0 compilers mis-compiled that test. So
3536 instead we test for the problematic value in a more direct
3537 manner and hope the Sun compilers get it correct. */
3538 && INTVAL (const_arg1) !=
3539 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3540 && REG_P (folded_arg1))
3542 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3543 struct table_elt *p
3544 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3546 if (p)
3547 for (p = p->first_same_value; p; p = p->next_same_value)
3548 if (REG_P (p->exp))
3549 return simplify_gen_binary (MINUS, mode, folded_arg0,
3550 canon_reg (p->exp, NULL_RTX));
3552 goto from_plus;
3554 case MINUS:
3555 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3556 If so, produce (PLUS Z C2-C). */
3557 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3559 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3560 if (y && CONST_INT_P (XEXP (y, 1)))
3561 return fold_rtx (plus_constant (mode, copy_rtx (y),
3562 -INTVAL (const_arg1)),
3563 NULL_RTX);
3566 /* Fall through. */
3568 from_plus:
3569 case SMIN: case SMAX: case UMIN: case UMAX:
3570 case IOR: case AND: case XOR:
3571 case MULT:
3572 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3573 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3574 is known to be of similar form, we may be able to replace the
3575 operation with a combined operation. This may eliminate the
3576 intermediate operation if every use is simplified in this way.
3577 Note that the similar optimization done by combine.c only works
3578 if the intermediate operation's result has only one reference. */
3580 if (REG_P (folded_arg0)
3581 && const_arg1 && CONST_INT_P (const_arg1))
3583 int is_shift
3584 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3585 rtx y, inner_const, new_const;
3586 rtx canon_const_arg1 = const_arg1;
3587 enum rtx_code associate_code;
3589 if (is_shift
3590 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3591 || INTVAL (const_arg1) < 0))
3593 if (SHIFT_COUNT_TRUNCATED)
3594 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3595 & (GET_MODE_BITSIZE (mode)
3596 - 1));
3597 else
3598 break;
3601 y = lookup_as_function (folded_arg0, code);
3602 if (y == 0)
3603 break;
3605 /* If we have compiled a statement like
3606 "if (x == (x & mask1))", and now are looking at
3607 "x & mask2", we will have a case where the first operand
3608 of Y is the same as our first operand. Unless we detect
3609 this case, an infinite loop will result. */
3610 if (XEXP (y, 0) == folded_arg0)
3611 break;
3613 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3614 if (!inner_const || !CONST_INT_P (inner_const))
3615 break;
3617 /* Don't associate these operations if they are a PLUS with the
3618 same constant and it is a power of two. These might be doable
3619 with a pre- or post-increment. Similarly for two subtracts of
3620 identical powers of two with post decrement. */
3622 if (code == PLUS && const_arg1 == inner_const
3623 && ((HAVE_PRE_INCREMENT
3624 && exact_log2 (INTVAL (const_arg1)) >= 0)
3625 || (HAVE_POST_INCREMENT
3626 && exact_log2 (INTVAL (const_arg1)) >= 0)
3627 || (HAVE_PRE_DECREMENT
3628 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3629 || (HAVE_POST_DECREMENT
3630 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3631 break;
3633 /* ??? Vector mode shifts by scalar
3634 shift operand are not supported yet. */
3635 if (is_shift && VECTOR_MODE_P (mode))
3636 break;
3638 if (is_shift
3639 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3640 || INTVAL (inner_const) < 0))
3642 if (SHIFT_COUNT_TRUNCATED)
3643 inner_const = GEN_INT (INTVAL (inner_const)
3644 & (GET_MODE_BITSIZE (mode) - 1));
3645 else
3646 break;
3649 /* Compute the code used to compose the constants. For example,
3650 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3652 associate_code = (is_shift || code == MINUS ? PLUS : code);
3654 new_const = simplify_binary_operation (associate_code, mode,
3655 canon_const_arg1,
3656 inner_const);
3658 if (new_const == 0)
3659 break;
3661 /* If we are associating shift operations, don't let this
3662 produce a shift of the size of the object or larger.
3663 This could occur when we follow a sign-extend by a right
3664 shift on a machine that does a sign-extend as a pair
3665 of shifts. */
3667 if (is_shift
3668 && CONST_INT_P (new_const)
3669 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3671 /* As an exception, we can turn an ASHIFTRT of this
3672 form into a shift of the number of bits - 1. */
3673 if (code == ASHIFTRT)
3674 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3675 else if (!side_effects_p (XEXP (y, 0)))
3676 return CONST0_RTX (mode);
3677 else
3678 break;
3681 y = copy_rtx (XEXP (y, 0));
3683 /* If Y contains our first operand (the most common way this
3684 can happen is if Y is a MEM), we would do into an infinite
3685 loop if we tried to fold it. So don't in that case. */
3687 if (! reg_mentioned_p (folded_arg0, y))
3688 y = fold_rtx (y, insn);
3690 return simplify_gen_binary (code, mode, y, new_const);
3692 break;
3694 case DIV: case UDIV:
3695 /* ??? The associative optimization performed immediately above is
3696 also possible for DIV and UDIV using associate_code of MULT.
3697 However, we would need extra code to verify that the
3698 multiplication does not overflow, that is, there is no overflow
3699 in the calculation of new_const. */
3700 break;
3702 default:
3703 break;
3706 new_rtx = simplify_binary_operation (code, mode,
3707 const_arg0 ? const_arg0 : folded_arg0,
3708 const_arg1 ? const_arg1 : folded_arg1);
3709 break;
3711 case RTX_OBJ:
3712 /* (lo_sum (high X) X) is simply X. */
3713 if (code == LO_SUM && const_arg0 != 0
3714 && GET_CODE (const_arg0) == HIGH
3715 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3716 return const_arg1;
3717 break;
3719 case RTX_TERNARY:
3720 case RTX_BITFIELD_OPS:
3721 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3722 const_arg0 ? const_arg0 : folded_arg0,
3723 const_arg1 ? const_arg1 : folded_arg1,
3724 const_arg2 ? const_arg2 : XEXP (x, 2));
3725 break;
3727 default:
3728 break;
3731 return new_rtx ? new_rtx : x;
3734 /* Return a constant value currently equivalent to X.
3735 Return 0 if we don't know one. */
3737 static rtx
3738 equiv_constant (rtx x)
3740 if (REG_P (x)
3741 && REGNO_QTY_VALID_P (REGNO (x)))
3743 int x_q = REG_QTY (REGNO (x));
3744 struct qty_table_elem *x_ent = &qty_table[x_q];
3746 if (x_ent->const_rtx)
3747 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3750 if (x == 0 || CONSTANT_P (x))
3751 return x;
3753 if (GET_CODE (x) == SUBREG)
3755 enum machine_mode mode = GET_MODE (x);
3756 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3757 rtx new_rtx;
3759 /* See if we previously assigned a constant value to this SUBREG. */
3760 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3761 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3762 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3763 return new_rtx;
3765 /* If we didn't and if doing so makes sense, see if we previously
3766 assigned a constant value to the enclosing word mode SUBREG. */
3767 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3768 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3770 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3771 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3773 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3774 new_rtx = lookup_as_function (y, CONST_INT);
3775 if (new_rtx)
3776 return gen_lowpart (mode, new_rtx);
3780 /* Otherwise see if we already have a constant for the inner REG,
3781 and if that is enough to calculate an equivalent constant for
3782 the subreg. Note that the upper bits of paradoxical subregs
3783 are undefined, so they cannot be said to equal anything. */
3784 if (REG_P (SUBREG_REG (x))
3785 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3786 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3787 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3789 return 0;
3792 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3793 the hash table in case its value was seen before. */
3795 if (MEM_P (x))
3797 struct table_elt *elt;
3799 x = avoid_constant_pool_reference (x);
3800 if (CONSTANT_P (x))
3801 return x;
3803 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3804 if (elt == 0)
3805 return 0;
3807 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3808 if (elt->is_const && CONSTANT_P (elt->exp))
3809 return elt->exp;
3812 return 0;
3815 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3816 "taken" branch.
3818 In certain cases, this can cause us to add an equivalence. For example,
3819 if we are following the taken case of
3820 if (i == 2)
3821 we can add the fact that `i' and '2' are now equivalent.
3823 In any case, we can record that this comparison was passed. If the same
3824 comparison is seen later, we will know its value. */
3826 static void
3827 record_jump_equiv (rtx insn, bool taken)
3829 int cond_known_true;
3830 rtx op0, op1;
3831 rtx set;
3832 enum machine_mode mode, mode0, mode1;
3833 int reversed_nonequality = 0;
3834 enum rtx_code code;
3836 /* Ensure this is the right kind of insn. */
3837 gcc_assert (any_condjump_p (insn));
3839 set = pc_set (insn);
3841 /* See if this jump condition is known true or false. */
3842 if (taken)
3843 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3844 else
3845 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3847 /* Get the type of comparison being done and the operands being compared.
3848 If we had to reverse a non-equality condition, record that fact so we
3849 know that it isn't valid for floating-point. */
3850 code = GET_CODE (XEXP (SET_SRC (set), 0));
3851 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3852 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3854 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3855 if (! cond_known_true)
3857 code = reversed_comparison_code_parts (code, op0, op1, insn);
3859 /* Don't remember if we can't find the inverse. */
3860 if (code == UNKNOWN)
3861 return;
3864 /* The mode is the mode of the non-constant. */
3865 mode = mode0;
3866 if (mode1 != VOIDmode)
3867 mode = mode1;
3869 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3872 /* Yet another form of subreg creation. In this case, we want something in
3873 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3875 static rtx
3876 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3878 enum machine_mode op_mode = GET_MODE (op);
3879 if (op_mode == mode || op_mode == VOIDmode)
3880 return op;
3881 return lowpart_subreg (mode, op, op_mode);
3884 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3885 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3886 Make any useful entries we can with that information. Called from
3887 above function and called recursively. */
3889 static void
3890 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3891 rtx op1, int reversed_nonequality)
3893 unsigned op0_hash, op1_hash;
3894 int op0_in_memory, op1_in_memory;
3895 struct table_elt *op0_elt, *op1_elt;
3897 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3898 we know that they are also equal in the smaller mode (this is also
3899 true for all smaller modes whether or not there is a SUBREG, but
3900 is not worth testing for with no SUBREG). */
3902 /* Note that GET_MODE (op0) may not equal MODE. */
3903 if (code == EQ && paradoxical_subreg_p (op0))
3905 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3906 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3907 if (tem)
3908 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3909 reversed_nonequality);
3912 if (code == EQ && paradoxical_subreg_p (op1))
3914 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3915 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3916 if (tem)
3917 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3918 reversed_nonequality);
3921 /* Similarly, if this is an NE comparison, and either is a SUBREG
3922 making a smaller mode, we know the whole thing is also NE. */
3924 /* Note that GET_MODE (op0) may not equal MODE;
3925 if we test MODE instead, we can get an infinite recursion
3926 alternating between two modes each wider than MODE. */
3928 if (code == NE && GET_CODE (op0) == SUBREG
3929 && subreg_lowpart_p (op0)
3930 && (GET_MODE_SIZE (GET_MODE (op0))
3931 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3933 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3934 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3935 if (tem)
3936 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3937 reversed_nonequality);
3940 if (code == NE && GET_CODE (op1) == SUBREG
3941 && subreg_lowpart_p (op1)
3942 && (GET_MODE_SIZE (GET_MODE (op1))
3943 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3945 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3946 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3947 if (tem)
3948 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3949 reversed_nonequality);
3952 /* Hash both operands. */
3954 do_not_record = 0;
3955 hash_arg_in_memory = 0;
3956 op0_hash = HASH (op0, mode);
3957 op0_in_memory = hash_arg_in_memory;
3959 if (do_not_record)
3960 return;
3962 do_not_record = 0;
3963 hash_arg_in_memory = 0;
3964 op1_hash = HASH (op1, mode);
3965 op1_in_memory = hash_arg_in_memory;
3967 if (do_not_record)
3968 return;
3970 /* Look up both operands. */
3971 op0_elt = lookup (op0, op0_hash, mode);
3972 op1_elt = lookup (op1, op1_hash, mode);
3974 /* If both operands are already equivalent or if they are not in the
3975 table but are identical, do nothing. */
3976 if ((op0_elt != 0 && op1_elt != 0
3977 && op0_elt->first_same_value == op1_elt->first_same_value)
3978 || op0 == op1 || rtx_equal_p (op0, op1))
3979 return;
3981 /* If we aren't setting two things equal all we can do is save this
3982 comparison. Similarly if this is floating-point. In the latter
3983 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
3984 If we record the equality, we might inadvertently delete code
3985 whose intent was to change -0 to +0. */
3987 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
3989 struct qty_table_elem *ent;
3990 int qty;
3992 /* If we reversed a floating-point comparison, if OP0 is not a
3993 register, or if OP1 is neither a register or constant, we can't
3994 do anything. */
3996 if (!REG_P (op1))
3997 op1 = equiv_constant (op1);
3999 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4000 || !REG_P (op0) || op1 == 0)
4001 return;
4003 /* Put OP0 in the hash table if it isn't already. This gives it a
4004 new quantity number. */
4005 if (op0_elt == 0)
4007 if (insert_regs (op0, NULL, 0))
4009 rehash_using_reg (op0);
4010 op0_hash = HASH (op0, mode);
4012 /* If OP0 is contained in OP1, this changes its hash code
4013 as well. Faster to rehash than to check, except
4014 for the simple case of a constant. */
4015 if (! CONSTANT_P (op1))
4016 op1_hash = HASH (op1,mode);
4019 op0_elt = insert (op0, NULL, op0_hash, mode);
4020 op0_elt->in_memory = op0_in_memory;
4023 qty = REG_QTY (REGNO (op0));
4024 ent = &qty_table[qty];
4026 ent->comparison_code = code;
4027 if (REG_P (op1))
4029 /* Look it up again--in case op0 and op1 are the same. */
4030 op1_elt = lookup (op1, op1_hash, mode);
4032 /* Put OP1 in the hash table so it gets a new quantity number. */
4033 if (op1_elt == 0)
4035 if (insert_regs (op1, NULL, 0))
4037 rehash_using_reg (op1);
4038 op1_hash = HASH (op1, mode);
4041 op1_elt = insert (op1, NULL, op1_hash, mode);
4042 op1_elt->in_memory = op1_in_memory;
4045 ent->comparison_const = NULL_RTX;
4046 ent->comparison_qty = REG_QTY (REGNO (op1));
4048 else
4050 ent->comparison_const = op1;
4051 ent->comparison_qty = -1;
4054 return;
4057 /* If either side is still missing an equivalence, make it now,
4058 then merge the equivalences. */
4060 if (op0_elt == 0)
4062 if (insert_regs (op0, NULL, 0))
4064 rehash_using_reg (op0);
4065 op0_hash = HASH (op0, mode);
4068 op0_elt = insert (op0, NULL, op0_hash, mode);
4069 op0_elt->in_memory = op0_in_memory;
4072 if (op1_elt == 0)
4074 if (insert_regs (op1, NULL, 0))
4076 rehash_using_reg (op1);
4077 op1_hash = HASH (op1, mode);
4080 op1_elt = insert (op1, NULL, op1_hash, mode);
4081 op1_elt->in_memory = op1_in_memory;
4084 merge_equiv_classes (op0_elt, op1_elt);
4087 /* CSE processing for one instruction.
4089 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4090 but the few that "leak through" are cleaned up by cse_insn, and complex
4091 addressing modes are often formed here.
4093 The main function is cse_insn, and between here and that function
4094 a couple of helper functions is defined to keep the size of cse_insn
4095 within reasonable proportions.
4097 Data is shared between the main and helper functions via STRUCT SET,
4098 that contains all data related for every set in the instruction that
4099 is being processed.
4101 Note that cse_main processes all sets in the instruction. Most
4102 passes in GCC only process simple SET insns or single_set insns, but
4103 CSE processes insns with multiple sets as well. */
4105 /* Data on one SET contained in the instruction. */
4107 struct set
4109 /* The SET rtx itself. */
4110 rtx rtl;
4111 /* The SET_SRC of the rtx (the original value, if it is changing). */
4112 rtx src;
4113 /* The hash-table element for the SET_SRC of the SET. */
4114 struct table_elt *src_elt;
4115 /* Hash value for the SET_SRC. */
4116 unsigned src_hash;
4117 /* Hash value for the SET_DEST. */
4118 unsigned dest_hash;
4119 /* The SET_DEST, with SUBREG, etc., stripped. */
4120 rtx inner_dest;
4121 /* Nonzero if the SET_SRC is in memory. */
4122 char src_in_memory;
4123 /* Nonzero if the SET_SRC contains something
4124 whose value cannot be predicted and understood. */
4125 char src_volatile;
4126 /* Original machine mode, in case it becomes a CONST_INT.
4127 The size of this field should match the size of the mode
4128 field of struct rtx_def (see rtl.h). */
4129 ENUM_BITFIELD(machine_mode) mode : 8;
4130 /* A constant equivalent for SET_SRC, if any. */
4131 rtx src_const;
4132 /* Hash value of constant equivalent for SET_SRC. */
4133 unsigned src_const_hash;
4134 /* Table entry for constant equivalent for SET_SRC, if any. */
4135 struct table_elt *src_const_elt;
4136 /* Table entry for the destination address. */
4137 struct table_elt *dest_addr_elt;
4140 /* Special handling for (set REG0 REG1) where REG0 is the
4141 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4142 be used in the sequel, so (if easily done) change this insn to
4143 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4144 that computed their value. Then REG1 will become a dead store
4145 and won't cloud the situation for later optimizations.
4147 Do not make this change if REG1 is a hard register, because it will
4148 then be used in the sequel and we may be changing a two-operand insn
4149 into a three-operand insn.
4151 This is the last transformation that cse_insn will try to do. */
4153 static void
4154 try_back_substitute_reg (rtx set, rtx insn)
4156 rtx dest = SET_DEST (set);
4157 rtx src = SET_SRC (set);
4159 if (REG_P (dest)
4160 && REG_P (src) && ! HARD_REGISTER_P (src)
4161 && REGNO_QTY_VALID_P (REGNO (src)))
4163 int src_q = REG_QTY (REGNO (src));
4164 struct qty_table_elem *src_ent = &qty_table[src_q];
4166 if (src_ent->first_reg == REGNO (dest))
4168 /* Scan for the previous nonnote insn, but stop at a basic
4169 block boundary. */
4170 rtx prev = insn;
4171 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4174 prev = PREV_INSN (prev);
4176 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4178 /* Do not swap the registers around if the previous instruction
4179 attaches a REG_EQUIV note to REG1.
4181 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4182 from the pseudo that originally shadowed an incoming argument
4183 to another register. Some uses of REG_EQUIV might rely on it
4184 being attached to REG1 rather than REG2.
4186 This section previously turned the REG_EQUIV into a REG_EQUAL
4187 note. We cannot do that because REG_EQUIV may provide an
4188 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4189 if (NONJUMP_INSN_P (prev)
4190 && GET_CODE (PATTERN (prev)) == SET
4191 && SET_DEST (PATTERN (prev)) == src
4192 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4194 rtx note;
4196 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4197 validate_change (insn, &SET_DEST (set), src, 1);
4198 validate_change (insn, &SET_SRC (set), dest, 1);
4199 apply_change_group ();
4201 /* If INSN has a REG_EQUAL note, and this note mentions
4202 REG0, then we must delete it, because the value in
4203 REG0 has changed. If the note's value is REG1, we must
4204 also delete it because that is now this insn's dest. */
4205 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4206 if (note != 0
4207 && (reg_mentioned_p (dest, XEXP (note, 0))
4208 || rtx_equal_p (src, XEXP (note, 0))))
4209 remove_note (insn, note);
4215 /* Record all the SETs in this instruction into SETS_PTR,
4216 and return the number of recorded sets. */
4217 static int
4218 find_sets_in_insn (rtx insn, struct set **psets)
4220 struct set *sets = *psets;
4221 int n_sets = 0;
4222 rtx x = PATTERN (insn);
4224 if (GET_CODE (x) == SET)
4226 /* Ignore SETs that are unconditional jumps.
4227 They never need cse processing, so this does not hurt.
4228 The reason is not efficiency but rather
4229 so that we can test at the end for instructions
4230 that have been simplified to unconditional jumps
4231 and not be misled by unchanged instructions
4232 that were unconditional jumps to begin with. */
4233 if (SET_DEST (x) == pc_rtx
4234 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4236 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4237 The hard function value register is used only once, to copy to
4238 someplace else, so it isn't worth cse'ing. */
4239 else if (GET_CODE (SET_SRC (x)) == CALL)
4241 else
4242 sets[n_sets++].rtl = x;
4244 else if (GET_CODE (x) == PARALLEL)
4246 int i, lim = XVECLEN (x, 0);
4248 /* Go over the epressions of the PARALLEL in forward order, to
4249 put them in the same order in the SETS array. */
4250 for (i = 0; i < lim; i++)
4252 rtx y = XVECEXP (x, 0, i);
4253 if (GET_CODE (y) == SET)
4255 /* As above, we ignore unconditional jumps and call-insns and
4256 ignore the result of apply_change_group. */
4257 if (SET_DEST (y) == pc_rtx
4258 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4260 else if (GET_CODE (SET_SRC (y)) == CALL)
4262 else
4263 sets[n_sets++].rtl = y;
4268 return n_sets;
4271 /* Where possible, substitute every register reference in the N_SETS
4272 number of SETS in INSN with the the canonical register.
4274 Register canonicalization propagatest the earliest register (i.e.
4275 one that is set before INSN) with the same value. This is a very
4276 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4277 to RTL. For instance, a CONST for an address is usually expanded
4278 multiple times to loads into different registers, thus creating many
4279 subexpressions of the form:
4281 (set (reg1) (some_const))
4282 (set (mem (... reg1 ...) (thing)))
4283 (set (reg2) (some_const))
4284 (set (mem (... reg2 ...) (thing)))
4286 After canonicalizing, the code takes the following form:
4288 (set (reg1) (some_const))
4289 (set (mem (... reg1 ...) (thing)))
4290 (set (reg2) (some_const))
4291 (set (mem (... reg1 ...) (thing)))
4293 The set to reg2 is now trivially dead, and the memory reference (or
4294 address, or whatever) may be a candidate for further CSEing.
4296 In this function, the result of apply_change_group can be ignored;
4297 see canon_reg. */
4299 static void
4300 canonicalize_insn (rtx insn, struct set **psets, int n_sets)
4302 struct set *sets = *psets;
4303 rtx tem;
4304 rtx x = PATTERN (insn);
4305 int i;
4307 if (CALL_P (insn))
4309 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4310 if (GET_CODE (XEXP (tem, 0)) != SET)
4311 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4314 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4316 canon_reg (SET_SRC (x), insn);
4317 apply_change_group ();
4318 fold_rtx (SET_SRC (x), insn);
4320 else if (GET_CODE (x) == CLOBBER)
4322 /* If we clobber memory, canon the address.
4323 This does nothing when a register is clobbered
4324 because we have already invalidated the reg. */
4325 if (MEM_P (XEXP (x, 0)))
4326 canon_reg (XEXP (x, 0), insn);
4328 else if (GET_CODE (x) == USE
4329 && ! (REG_P (XEXP (x, 0))
4330 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4331 /* Canonicalize a USE of a pseudo register or memory location. */
4332 canon_reg (x, insn);
4333 else if (GET_CODE (x) == ASM_OPERANDS)
4335 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4337 rtx input = ASM_OPERANDS_INPUT (x, i);
4338 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4340 input = canon_reg (input, insn);
4341 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4345 else if (GET_CODE (x) == CALL)
4347 canon_reg (x, insn);
4348 apply_change_group ();
4349 fold_rtx (x, insn);
4351 else if (DEBUG_INSN_P (insn))
4352 canon_reg (PATTERN (insn), insn);
4353 else if (GET_CODE (x) == PARALLEL)
4355 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4357 rtx y = XVECEXP (x, 0, i);
4358 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4360 canon_reg (SET_SRC (y), insn);
4361 apply_change_group ();
4362 fold_rtx (SET_SRC (y), insn);
4364 else if (GET_CODE (y) == CLOBBER)
4366 if (MEM_P (XEXP (y, 0)))
4367 canon_reg (XEXP (y, 0), insn);
4369 else if (GET_CODE (y) == USE
4370 && ! (REG_P (XEXP (y, 0))
4371 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4372 canon_reg (y, insn);
4373 else if (GET_CODE (y) == CALL)
4375 canon_reg (y, insn);
4376 apply_change_group ();
4377 fold_rtx (y, insn);
4382 if (n_sets == 1 && REG_NOTES (insn) != 0
4383 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4385 /* We potentially will process this insn many times. Therefore,
4386 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4387 unique set in INSN.
4389 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4390 because cse_insn handles those specially. */
4391 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4392 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4393 remove_note (insn, tem);
4394 else
4396 canon_reg (XEXP (tem, 0), insn);
4397 apply_change_group ();
4398 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4399 df_notes_rescan (insn);
4403 /* Canonicalize sources and addresses of destinations.
4404 We do this in a separate pass to avoid problems when a MATCH_DUP is
4405 present in the insn pattern. In that case, we want to ensure that
4406 we don't break the duplicate nature of the pattern. So we will replace
4407 both operands at the same time. Otherwise, we would fail to find an
4408 equivalent substitution in the loop calling validate_change below.
4410 We used to suppress canonicalization of DEST if it appears in SRC,
4411 but we don't do this any more. */
4413 for (i = 0; i < n_sets; i++)
4415 rtx dest = SET_DEST (sets[i].rtl);
4416 rtx src = SET_SRC (sets[i].rtl);
4417 rtx new_rtx = canon_reg (src, insn);
4419 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4421 if (GET_CODE (dest) == ZERO_EXTRACT)
4423 validate_change (insn, &XEXP (dest, 1),
4424 canon_reg (XEXP (dest, 1), insn), 1);
4425 validate_change (insn, &XEXP (dest, 2),
4426 canon_reg (XEXP (dest, 2), insn), 1);
4429 while (GET_CODE (dest) == SUBREG
4430 || GET_CODE (dest) == ZERO_EXTRACT
4431 || GET_CODE (dest) == STRICT_LOW_PART)
4432 dest = XEXP (dest, 0);
4434 if (MEM_P (dest))
4435 canon_reg (dest, insn);
4438 /* Now that we have done all the replacements, we can apply the change
4439 group and see if they all work. Note that this will cause some
4440 canonicalizations that would have worked individually not to be applied
4441 because some other canonicalization didn't work, but this should not
4442 occur often.
4444 The result of apply_change_group can be ignored; see canon_reg. */
4446 apply_change_group ();
4449 /* Main function of CSE.
4450 First simplify sources and addresses of all assignments
4451 in the instruction, using previously-computed equivalents values.
4452 Then install the new sources and destinations in the table
4453 of available values. */
4455 static void
4456 cse_insn (rtx insn)
4458 rtx x = PATTERN (insn);
4459 int i;
4460 rtx tem;
4461 int n_sets = 0;
4463 rtx src_eqv = 0;
4464 struct table_elt *src_eqv_elt = 0;
4465 int src_eqv_volatile = 0;
4466 int src_eqv_in_memory = 0;
4467 unsigned src_eqv_hash = 0;
4469 struct set *sets = (struct set *) 0;
4471 if (GET_CODE (x) == SET)
4472 sets = XALLOCA (struct set);
4473 else if (GET_CODE (x) == PARALLEL)
4474 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4476 this_insn = insn;
4477 #ifdef HAVE_cc0
4478 /* Records what this insn does to set CC0. */
4479 this_insn_cc0 = 0;
4480 this_insn_cc0_mode = VOIDmode;
4481 #endif
4483 /* Find all regs explicitly clobbered in this insn,
4484 to ensure they are not replaced with any other regs
4485 elsewhere in this insn. */
4486 invalidate_from_sets_and_clobbers (insn);
4488 /* Record all the SETs in this instruction. */
4489 n_sets = find_sets_in_insn (insn, &sets);
4491 /* Substitute the canonical register where possible. */
4492 canonicalize_insn (insn, &sets, n_sets);
4494 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4495 if different, or if the DEST is a STRICT_LOW_PART. The latter condition
4496 is necessary because SRC_EQV is handled specially for this case, and if
4497 it isn't set, then there will be no equivalence for the destination. */
4498 if (n_sets == 1 && REG_NOTES (insn) != 0
4499 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4500 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4501 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4502 src_eqv = copy_rtx (XEXP (tem, 0));
4504 /* Set sets[i].src_elt to the class each source belongs to.
4505 Detect assignments from or to volatile things
4506 and set set[i] to zero so they will be ignored
4507 in the rest of this function.
4509 Nothing in this loop changes the hash table or the register chains. */
4511 for (i = 0; i < n_sets; i++)
4513 bool repeat = false;
4514 rtx src, dest;
4515 rtx src_folded;
4516 struct table_elt *elt = 0, *p;
4517 enum machine_mode mode;
4518 rtx src_eqv_here;
4519 rtx src_const = 0;
4520 rtx src_related = 0;
4521 bool src_related_is_const_anchor = false;
4522 struct table_elt *src_const_elt = 0;
4523 int src_cost = MAX_COST;
4524 int src_eqv_cost = MAX_COST;
4525 int src_folded_cost = MAX_COST;
4526 int src_related_cost = MAX_COST;
4527 int src_elt_cost = MAX_COST;
4528 int src_regcost = MAX_COST;
4529 int src_eqv_regcost = MAX_COST;
4530 int src_folded_regcost = MAX_COST;
4531 int src_related_regcost = MAX_COST;
4532 int src_elt_regcost = MAX_COST;
4533 /* Set nonzero if we need to call force_const_mem on with the
4534 contents of src_folded before using it. */
4535 int src_folded_force_flag = 0;
4537 dest = SET_DEST (sets[i].rtl);
4538 src = SET_SRC (sets[i].rtl);
4540 /* If SRC is a constant that has no machine mode,
4541 hash it with the destination's machine mode.
4542 This way we can keep different modes separate. */
4544 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4545 sets[i].mode = mode;
4547 if (src_eqv)
4549 enum machine_mode eqvmode = mode;
4550 if (GET_CODE (dest) == STRICT_LOW_PART)
4551 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4552 do_not_record = 0;
4553 hash_arg_in_memory = 0;
4554 src_eqv_hash = HASH (src_eqv, eqvmode);
4556 /* Find the equivalence class for the equivalent expression. */
4558 if (!do_not_record)
4559 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4561 src_eqv_volatile = do_not_record;
4562 src_eqv_in_memory = hash_arg_in_memory;
4565 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4566 value of the INNER register, not the destination. So it is not
4567 a valid substitution for the source. But save it for later. */
4568 if (GET_CODE (dest) == STRICT_LOW_PART)
4569 src_eqv_here = 0;
4570 else
4571 src_eqv_here = src_eqv;
4573 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4574 simplified result, which may not necessarily be valid. */
4575 src_folded = fold_rtx (src, insn);
4577 #if 0
4578 /* ??? This caused bad code to be generated for the m68k port with -O2.
4579 Suppose src is (CONST_INT -1), and that after truncation src_folded
4580 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4581 At the end we will add src and src_const to the same equivalence
4582 class. We now have 3 and -1 on the same equivalence class. This
4583 causes later instructions to be mis-optimized. */
4584 /* If storing a constant in a bitfield, pre-truncate the constant
4585 so we will be able to record it later. */
4586 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4588 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4590 if (CONST_INT_P (src)
4591 && CONST_INT_P (width)
4592 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4593 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4594 src_folded
4595 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4596 << INTVAL (width)) - 1));
4598 #endif
4600 /* Compute SRC's hash code, and also notice if it
4601 should not be recorded at all. In that case,
4602 prevent any further processing of this assignment. */
4603 do_not_record = 0;
4604 hash_arg_in_memory = 0;
4606 sets[i].src = src;
4607 sets[i].src_hash = HASH (src, mode);
4608 sets[i].src_volatile = do_not_record;
4609 sets[i].src_in_memory = hash_arg_in_memory;
4611 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4612 a pseudo, do not record SRC. Using SRC as a replacement for
4613 anything else will be incorrect in that situation. Note that
4614 this usually occurs only for stack slots, in which case all the
4615 RTL would be referring to SRC, so we don't lose any optimization
4616 opportunities by not having SRC in the hash table. */
4618 if (MEM_P (src)
4619 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4620 && REG_P (dest)
4621 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4622 sets[i].src_volatile = 1;
4624 #if 0
4625 /* It is no longer clear why we used to do this, but it doesn't
4626 appear to still be needed. So let's try without it since this
4627 code hurts cse'ing widened ops. */
4628 /* If source is a paradoxical subreg (such as QI treated as an SI),
4629 treat it as volatile. It may do the work of an SI in one context
4630 where the extra bits are not being used, but cannot replace an SI
4631 in general. */
4632 if (paradoxical_subreg_p (src))
4633 sets[i].src_volatile = 1;
4634 #endif
4636 /* Locate all possible equivalent forms for SRC. Try to replace
4637 SRC in the insn with each cheaper equivalent.
4639 We have the following types of equivalents: SRC itself, a folded
4640 version, a value given in a REG_EQUAL note, or a value related
4641 to a constant.
4643 Each of these equivalents may be part of an additional class
4644 of equivalents (if more than one is in the table, they must be in
4645 the same class; we check for this).
4647 If the source is volatile, we don't do any table lookups.
4649 We note any constant equivalent for possible later use in a
4650 REG_NOTE. */
4652 if (!sets[i].src_volatile)
4653 elt = lookup (src, sets[i].src_hash, mode);
4655 sets[i].src_elt = elt;
4657 if (elt && src_eqv_here && src_eqv_elt)
4659 if (elt->first_same_value != src_eqv_elt->first_same_value)
4661 /* The REG_EQUAL is indicating that two formerly distinct
4662 classes are now equivalent. So merge them. */
4663 merge_equiv_classes (elt, src_eqv_elt);
4664 src_eqv_hash = HASH (src_eqv, elt->mode);
4665 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4668 src_eqv_here = 0;
4671 else if (src_eqv_elt)
4672 elt = src_eqv_elt;
4674 /* Try to find a constant somewhere and record it in `src_const'.
4675 Record its table element, if any, in `src_const_elt'. Look in
4676 any known equivalences first. (If the constant is not in the
4677 table, also set `sets[i].src_const_hash'). */
4678 if (elt)
4679 for (p = elt->first_same_value; p; p = p->next_same_value)
4680 if (p->is_const)
4682 src_const = p->exp;
4683 src_const_elt = elt;
4684 break;
4687 if (src_const == 0
4688 && (CONSTANT_P (src_folded)
4689 /* Consider (minus (label_ref L1) (label_ref L2)) as
4690 "constant" here so we will record it. This allows us
4691 to fold switch statements when an ADDR_DIFF_VEC is used. */
4692 || (GET_CODE (src_folded) == MINUS
4693 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4694 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4695 src_const = src_folded, src_const_elt = elt;
4696 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4697 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4699 /* If we don't know if the constant is in the table, get its
4700 hash code and look it up. */
4701 if (src_const && src_const_elt == 0)
4703 sets[i].src_const_hash = HASH (src_const, mode);
4704 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4707 sets[i].src_const = src_const;
4708 sets[i].src_const_elt = src_const_elt;
4710 /* If the constant and our source are both in the table, mark them as
4711 equivalent. Otherwise, if a constant is in the table but the source
4712 isn't, set ELT to it. */
4713 if (src_const_elt && elt
4714 && src_const_elt->first_same_value != elt->first_same_value)
4715 merge_equiv_classes (elt, src_const_elt);
4716 else if (src_const_elt && elt == 0)
4717 elt = src_const_elt;
4719 /* See if there is a register linearly related to a constant
4720 equivalent of SRC. */
4721 if (src_const
4722 && (GET_CODE (src_const) == CONST
4723 || (src_const_elt && src_const_elt->related_value != 0)))
4725 src_related = use_related_value (src_const, src_const_elt);
4726 if (src_related)
4728 struct table_elt *src_related_elt
4729 = lookup (src_related, HASH (src_related, mode), mode);
4730 if (src_related_elt && elt)
4732 if (elt->first_same_value
4733 != src_related_elt->first_same_value)
4734 /* This can occur when we previously saw a CONST
4735 involving a SYMBOL_REF and then see the SYMBOL_REF
4736 twice. Merge the involved classes. */
4737 merge_equiv_classes (elt, src_related_elt);
4739 src_related = 0;
4740 src_related_elt = 0;
4742 else if (src_related_elt && elt == 0)
4743 elt = src_related_elt;
4747 /* See if we have a CONST_INT that is already in a register in a
4748 wider mode. */
4750 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4751 && GET_MODE_CLASS (mode) == MODE_INT
4752 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4754 enum machine_mode wider_mode;
4756 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4757 wider_mode != VOIDmode
4758 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4759 && src_related == 0;
4760 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4762 struct table_elt *const_elt
4763 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4765 if (const_elt == 0)
4766 continue;
4768 for (const_elt = const_elt->first_same_value;
4769 const_elt; const_elt = const_elt->next_same_value)
4770 if (REG_P (const_elt->exp))
4772 src_related = gen_lowpart (mode, const_elt->exp);
4773 break;
4778 /* Another possibility is that we have an AND with a constant in
4779 a mode narrower than a word. If so, it might have been generated
4780 as part of an "if" which would narrow the AND. If we already
4781 have done the AND in a wider mode, we can use a SUBREG of that
4782 value. */
4784 if (flag_expensive_optimizations && ! src_related
4785 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4786 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4788 enum machine_mode tmode;
4789 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4791 for (tmode = GET_MODE_WIDER_MODE (mode);
4792 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4793 tmode = GET_MODE_WIDER_MODE (tmode))
4795 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4796 struct table_elt *larger_elt;
4798 if (inner)
4800 PUT_MODE (new_and, tmode);
4801 XEXP (new_and, 0) = inner;
4802 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4803 if (larger_elt == 0)
4804 continue;
4806 for (larger_elt = larger_elt->first_same_value;
4807 larger_elt; larger_elt = larger_elt->next_same_value)
4808 if (REG_P (larger_elt->exp))
4810 src_related
4811 = gen_lowpart (mode, larger_elt->exp);
4812 break;
4815 if (src_related)
4816 break;
4821 #ifdef LOAD_EXTEND_OP
4822 /* See if a MEM has already been loaded with a widening operation;
4823 if it has, we can use a subreg of that. Many CISC machines
4824 also have such operations, but this is only likely to be
4825 beneficial on these machines. */
4827 if (flag_expensive_optimizations && src_related == 0
4828 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4829 && GET_MODE_CLASS (mode) == MODE_INT
4830 && MEM_P (src) && ! do_not_record
4831 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4833 struct rtx_def memory_extend_buf;
4834 rtx memory_extend_rtx = &memory_extend_buf;
4835 enum machine_mode tmode;
4837 /* Set what we are trying to extend and the operation it might
4838 have been extended with. */
4839 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4840 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4841 XEXP (memory_extend_rtx, 0) = src;
4843 for (tmode = GET_MODE_WIDER_MODE (mode);
4844 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4845 tmode = GET_MODE_WIDER_MODE (tmode))
4847 struct table_elt *larger_elt;
4849 PUT_MODE (memory_extend_rtx, tmode);
4850 larger_elt = lookup (memory_extend_rtx,
4851 HASH (memory_extend_rtx, tmode), tmode);
4852 if (larger_elt == 0)
4853 continue;
4855 for (larger_elt = larger_elt->first_same_value;
4856 larger_elt; larger_elt = larger_elt->next_same_value)
4857 if (REG_P (larger_elt->exp))
4859 src_related = gen_lowpart (mode, larger_elt->exp);
4860 break;
4863 if (src_related)
4864 break;
4867 #endif /* LOAD_EXTEND_OP */
4869 /* Try to express the constant using a register+offset expression
4870 derived from a constant anchor. */
4872 if (targetm.const_anchor
4873 && !src_related
4874 && src_const
4875 && GET_CODE (src_const) == CONST_INT)
4877 src_related = try_const_anchors (src_const, mode);
4878 src_related_is_const_anchor = src_related != NULL_RTX;
4882 if (src == src_folded)
4883 src_folded = 0;
4885 /* At this point, ELT, if nonzero, points to a class of expressions
4886 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4887 and SRC_RELATED, if nonzero, each contain additional equivalent
4888 expressions. Prune these latter expressions by deleting expressions
4889 already in the equivalence class.
4891 Check for an equivalent identical to the destination. If found,
4892 this is the preferred equivalent since it will likely lead to
4893 elimination of the insn. Indicate this by placing it in
4894 `src_related'. */
4896 if (elt)
4897 elt = elt->first_same_value;
4898 for (p = elt; p; p = p->next_same_value)
4900 enum rtx_code code = GET_CODE (p->exp);
4902 /* If the expression is not valid, ignore it. Then we do not
4903 have to check for validity below. In most cases, we can use
4904 `rtx_equal_p', since canonicalization has already been done. */
4905 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4906 continue;
4908 /* Also skip paradoxical subregs, unless that's what we're
4909 looking for. */
4910 if (paradoxical_subreg_p (p->exp)
4911 && ! (src != 0
4912 && GET_CODE (src) == SUBREG
4913 && GET_MODE (src) == GET_MODE (p->exp)
4914 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4915 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4916 continue;
4918 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4919 src = 0;
4920 else if (src_folded && GET_CODE (src_folded) == code
4921 && rtx_equal_p (src_folded, p->exp))
4922 src_folded = 0;
4923 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4924 && rtx_equal_p (src_eqv_here, p->exp))
4925 src_eqv_here = 0;
4926 else if (src_related && GET_CODE (src_related) == code
4927 && rtx_equal_p (src_related, p->exp))
4928 src_related = 0;
4930 /* This is the same as the destination of the insns, we want
4931 to prefer it. Copy it to src_related. The code below will
4932 then give it a negative cost. */
4933 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4934 src_related = dest;
4937 /* Find the cheapest valid equivalent, trying all the available
4938 possibilities. Prefer items not in the hash table to ones
4939 that are when they are equal cost. Note that we can never
4940 worsen an insn as the current contents will also succeed.
4941 If we find an equivalent identical to the destination, use it as best,
4942 since this insn will probably be eliminated in that case. */
4943 if (src)
4945 if (rtx_equal_p (src, dest))
4946 src_cost = src_regcost = -1;
4947 else
4949 src_cost = COST (src);
4950 src_regcost = approx_reg_cost (src);
4954 if (src_eqv_here)
4956 if (rtx_equal_p (src_eqv_here, dest))
4957 src_eqv_cost = src_eqv_regcost = -1;
4958 else
4960 src_eqv_cost = COST (src_eqv_here);
4961 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4965 if (src_folded)
4967 if (rtx_equal_p (src_folded, dest))
4968 src_folded_cost = src_folded_regcost = -1;
4969 else
4971 src_folded_cost = COST (src_folded);
4972 src_folded_regcost = approx_reg_cost (src_folded);
4976 if (src_related)
4978 if (rtx_equal_p (src_related, dest))
4979 src_related_cost = src_related_regcost = -1;
4980 else
4982 src_related_cost = COST (src_related);
4983 src_related_regcost = approx_reg_cost (src_related);
4985 /* If a const-anchor is used to synthesize a constant that
4986 normally requires multiple instructions then slightly prefer
4987 it over the original sequence. These instructions are likely
4988 to become redundant now. We can't compare against the cost
4989 of src_eqv_here because, on MIPS for example, multi-insn
4990 constants have zero cost; they are assumed to be hoisted from
4991 loops. */
4992 if (src_related_is_const_anchor
4993 && src_related_cost == src_cost
4994 && src_eqv_here)
4995 src_related_cost--;
4999 /* If this was an indirect jump insn, a known label will really be
5000 cheaper even though it looks more expensive. */
5001 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5002 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5004 /* Terminate loop when replacement made. This must terminate since
5005 the current contents will be tested and will always be valid. */
5006 while (1)
5008 rtx trial;
5010 /* Skip invalid entries. */
5011 while (elt && !REG_P (elt->exp)
5012 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5013 elt = elt->next_same_value;
5015 /* A paradoxical subreg would be bad here: it'll be the right
5016 size, but later may be adjusted so that the upper bits aren't
5017 what we want. So reject it. */
5018 if (elt != 0
5019 && paradoxical_subreg_p (elt->exp)
5020 /* It is okay, though, if the rtx we're trying to match
5021 will ignore any of the bits we can't predict. */
5022 && ! (src != 0
5023 && GET_CODE (src) == SUBREG
5024 && GET_MODE (src) == GET_MODE (elt->exp)
5025 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5026 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5028 elt = elt->next_same_value;
5029 continue;
5032 if (elt)
5034 src_elt_cost = elt->cost;
5035 src_elt_regcost = elt->regcost;
5038 /* Find cheapest and skip it for the next time. For items
5039 of equal cost, use this order:
5040 src_folded, src, src_eqv, src_related and hash table entry. */
5041 if (src_folded
5042 && preferable (src_folded_cost, src_folded_regcost,
5043 src_cost, src_regcost) <= 0
5044 && preferable (src_folded_cost, src_folded_regcost,
5045 src_eqv_cost, src_eqv_regcost) <= 0
5046 && preferable (src_folded_cost, src_folded_regcost,
5047 src_related_cost, src_related_regcost) <= 0
5048 && preferable (src_folded_cost, src_folded_regcost,
5049 src_elt_cost, src_elt_regcost) <= 0)
5051 trial = src_folded, src_folded_cost = MAX_COST;
5052 if (src_folded_force_flag)
5054 rtx forced = force_const_mem (mode, trial);
5055 if (forced)
5056 trial = forced;
5059 else if (src
5060 && preferable (src_cost, src_regcost,
5061 src_eqv_cost, src_eqv_regcost) <= 0
5062 && preferable (src_cost, src_regcost,
5063 src_related_cost, src_related_regcost) <= 0
5064 && preferable (src_cost, src_regcost,
5065 src_elt_cost, src_elt_regcost) <= 0)
5066 trial = src, src_cost = MAX_COST;
5067 else if (src_eqv_here
5068 && preferable (src_eqv_cost, src_eqv_regcost,
5069 src_related_cost, src_related_regcost) <= 0
5070 && preferable (src_eqv_cost, src_eqv_regcost,
5071 src_elt_cost, src_elt_regcost) <= 0)
5072 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5073 else if (src_related
5074 && preferable (src_related_cost, src_related_regcost,
5075 src_elt_cost, src_elt_regcost) <= 0)
5076 trial = src_related, src_related_cost = MAX_COST;
5077 else
5079 trial = elt->exp;
5080 elt = elt->next_same_value;
5081 src_elt_cost = MAX_COST;
5084 /* Avoid creation of overlapping memory moves. */
5085 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5087 rtx src, dest;
5089 /* BLKmode moves are not handled by cse anyway. */
5090 if (GET_MODE (trial) == BLKmode)
5091 break;
5093 src = canon_rtx (trial);
5094 dest = canon_rtx (SET_DEST (sets[i].rtl));
5096 if (!MEM_P (src) || !MEM_P (dest)
5097 || !nonoverlapping_memrefs_p (src, dest, false))
5098 break;
5101 /* Try to optimize
5102 (set (reg:M N) (const_int A))
5103 (set (reg:M2 O) (const_int B))
5104 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5105 (reg:M2 O)). */
5106 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5107 && CONST_INT_P (trial)
5108 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5109 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5110 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5111 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5112 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5113 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5114 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5115 <= HOST_BITS_PER_WIDE_INT))
5117 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5118 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5119 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5120 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5121 struct table_elt *dest_elt
5122 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5123 rtx dest_cst = NULL;
5125 if (dest_elt)
5126 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5127 if (p->is_const && CONST_INT_P (p->exp))
5129 dest_cst = p->exp;
5130 break;
5132 if (dest_cst)
5134 HOST_WIDE_INT val = INTVAL (dest_cst);
5135 HOST_WIDE_INT mask;
5136 unsigned int shift;
5137 if (BITS_BIG_ENDIAN)
5138 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5139 - INTVAL (pos) - INTVAL (width);
5140 else
5141 shift = INTVAL (pos);
5142 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5143 mask = ~(HOST_WIDE_INT) 0;
5144 else
5145 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5146 val &= ~(mask << shift);
5147 val |= (INTVAL (trial) & mask) << shift;
5148 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5149 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5150 dest_reg, 1);
5151 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5152 GEN_INT (val), 1);
5153 if (apply_change_group ())
5155 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5156 if (note)
5158 remove_note (insn, note);
5159 df_notes_rescan (insn);
5161 src_eqv = NULL_RTX;
5162 src_eqv_elt = NULL;
5163 src_eqv_volatile = 0;
5164 src_eqv_in_memory = 0;
5165 src_eqv_hash = 0;
5166 repeat = true;
5167 break;
5172 /* We don't normally have an insn matching (set (pc) (pc)), so
5173 check for this separately here. We will delete such an
5174 insn below.
5176 For other cases such as a table jump or conditional jump
5177 where we know the ultimate target, go ahead and replace the
5178 operand. While that may not make a valid insn, we will
5179 reemit the jump below (and also insert any necessary
5180 barriers). */
5181 if (n_sets == 1 && dest == pc_rtx
5182 && (trial == pc_rtx
5183 || (GET_CODE (trial) == LABEL_REF
5184 && ! condjump_p (insn))))
5186 /* Don't substitute non-local labels, this confuses CFG. */
5187 if (GET_CODE (trial) == LABEL_REF
5188 && LABEL_REF_NONLOCAL_P (trial))
5189 continue;
5191 SET_SRC (sets[i].rtl) = trial;
5192 cse_jumps_altered = true;
5193 break;
5196 /* Reject certain invalid forms of CONST that we create. */
5197 else if (CONSTANT_P (trial)
5198 && GET_CODE (trial) == CONST
5199 /* Reject cases that will cause decode_rtx_const to
5200 die. On the alpha when simplifying a switch, we
5201 get (const (truncate (minus (label_ref)
5202 (label_ref)))). */
5203 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5204 /* Likewise on IA-64, except without the
5205 truncate. */
5206 || (GET_CODE (XEXP (trial, 0)) == MINUS
5207 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5208 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5209 /* Do nothing for this case. */
5212 /* Look for a substitution that makes a valid insn. */
5213 else if (validate_unshare_change
5214 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5216 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5218 /* The result of apply_change_group can be ignored; see
5219 canon_reg. */
5221 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5222 apply_change_group ();
5224 break;
5227 /* If we previously found constant pool entries for
5228 constants and this is a constant, try making a
5229 pool entry. Put it in src_folded unless we already have done
5230 this since that is where it likely came from. */
5232 else if (constant_pool_entries_cost
5233 && CONSTANT_P (trial)
5234 && (src_folded == 0
5235 || (!MEM_P (src_folded)
5236 && ! src_folded_force_flag))
5237 && GET_MODE_CLASS (mode) != MODE_CC
5238 && mode != VOIDmode)
5240 src_folded_force_flag = 1;
5241 src_folded = trial;
5242 src_folded_cost = constant_pool_entries_cost;
5243 src_folded_regcost = constant_pool_entries_regcost;
5247 /* If we changed the insn too much, handle this set from scratch. */
5248 if (repeat)
5250 i--;
5251 continue;
5254 src = SET_SRC (sets[i].rtl);
5256 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5257 However, there is an important exception: If both are registers
5258 that are not the head of their equivalence class, replace SET_SRC
5259 with the head of the class. If we do not do this, we will have
5260 both registers live over a portion of the basic block. This way,
5261 their lifetimes will likely abut instead of overlapping. */
5262 if (REG_P (dest)
5263 && REGNO_QTY_VALID_P (REGNO (dest)))
5265 int dest_q = REG_QTY (REGNO (dest));
5266 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5268 if (dest_ent->mode == GET_MODE (dest)
5269 && dest_ent->first_reg != REGNO (dest)
5270 && REG_P (src) && REGNO (src) == REGNO (dest)
5271 /* Don't do this if the original insn had a hard reg as
5272 SET_SRC or SET_DEST. */
5273 && (!REG_P (sets[i].src)
5274 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5275 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5276 /* We can't call canon_reg here because it won't do anything if
5277 SRC is a hard register. */
5279 int src_q = REG_QTY (REGNO (src));
5280 struct qty_table_elem *src_ent = &qty_table[src_q];
5281 int first = src_ent->first_reg;
5282 rtx new_src
5283 = (first >= FIRST_PSEUDO_REGISTER
5284 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5286 /* We must use validate-change even for this, because this
5287 might be a special no-op instruction, suitable only to
5288 tag notes onto. */
5289 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5291 src = new_src;
5292 /* If we had a constant that is cheaper than what we are now
5293 setting SRC to, use that constant. We ignored it when we
5294 thought we could make this into a no-op. */
5295 if (src_const && COST (src_const) < COST (src)
5296 && validate_change (insn, &SET_SRC (sets[i].rtl),
5297 src_const, 0))
5298 src = src_const;
5303 /* If we made a change, recompute SRC values. */
5304 if (src != sets[i].src)
5306 do_not_record = 0;
5307 hash_arg_in_memory = 0;
5308 sets[i].src = src;
5309 sets[i].src_hash = HASH (src, mode);
5310 sets[i].src_volatile = do_not_record;
5311 sets[i].src_in_memory = hash_arg_in_memory;
5312 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5315 /* If this is a single SET, we are setting a register, and we have an
5316 equivalent constant, we want to add a REG_NOTE. We don't want
5317 to write a REG_EQUAL note for a constant pseudo since verifying that
5318 that pseudo hasn't been eliminated is a pain. Such a note also
5319 won't help anything.
5321 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5322 which can be created for a reference to a compile time computable
5323 entry in a jump table. */
5325 if (n_sets == 1 && src_const && REG_P (dest)
5326 && !REG_P (src_const)
5327 && ! (GET_CODE (src_const) == CONST
5328 && GET_CODE (XEXP (src_const, 0)) == MINUS
5329 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5330 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5332 /* We only want a REG_EQUAL note if src_const != src. */
5333 if (! rtx_equal_p (src, src_const))
5335 /* Make sure that the rtx is not shared. */
5336 src_const = copy_rtx (src_const);
5338 /* Record the actual constant value in a REG_EQUAL note,
5339 making a new one if one does not already exist. */
5340 set_unique_reg_note (insn, REG_EQUAL, src_const);
5341 df_notes_rescan (insn);
5345 /* Now deal with the destination. */
5346 do_not_record = 0;
5348 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5349 while (GET_CODE (dest) == SUBREG
5350 || GET_CODE (dest) == ZERO_EXTRACT
5351 || GET_CODE (dest) == STRICT_LOW_PART)
5352 dest = XEXP (dest, 0);
5354 sets[i].inner_dest = dest;
5356 if (MEM_P (dest))
5358 #ifdef PUSH_ROUNDING
5359 /* Stack pushes invalidate the stack pointer. */
5360 rtx addr = XEXP (dest, 0);
5361 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5362 && XEXP (addr, 0) == stack_pointer_rtx)
5363 invalidate (stack_pointer_rtx, VOIDmode);
5364 #endif
5365 dest = fold_rtx (dest, insn);
5368 /* Compute the hash code of the destination now,
5369 before the effects of this instruction are recorded,
5370 since the register values used in the address computation
5371 are those before this instruction. */
5372 sets[i].dest_hash = HASH (dest, mode);
5374 /* Don't enter a bit-field in the hash table
5375 because the value in it after the store
5376 may not equal what was stored, due to truncation. */
5378 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5380 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5382 if (src_const != 0 && CONST_INT_P (src_const)
5383 && CONST_INT_P (width)
5384 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5385 && ! (INTVAL (src_const)
5386 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5387 /* Exception: if the value is constant,
5388 and it won't be truncated, record it. */
5390 else
5392 /* This is chosen so that the destination will be invalidated
5393 but no new value will be recorded.
5394 We must invalidate because sometimes constant
5395 values can be recorded for bitfields. */
5396 sets[i].src_elt = 0;
5397 sets[i].src_volatile = 1;
5398 src_eqv = 0;
5399 src_eqv_elt = 0;
5403 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5404 the insn. */
5405 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5407 /* One less use of the label this insn used to jump to. */
5408 delete_insn_and_edges (insn);
5409 cse_jumps_altered = true;
5410 /* No more processing for this set. */
5411 sets[i].rtl = 0;
5414 /* If this SET is now setting PC to a label, we know it used to
5415 be a conditional or computed branch. */
5416 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5417 && !LABEL_REF_NONLOCAL_P (src))
5419 /* We reemit the jump in as many cases as possible just in
5420 case the form of an unconditional jump is significantly
5421 different than a computed jump or conditional jump.
5423 If this insn has multiple sets, then reemitting the
5424 jump is nontrivial. So instead we just force rerecognition
5425 and hope for the best. */
5426 if (n_sets == 1)
5428 rtx new_rtx, note;
5430 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5431 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5432 LABEL_NUSES (XEXP (src, 0))++;
5434 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5435 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5436 if (note)
5438 XEXP (note, 1) = NULL_RTX;
5439 REG_NOTES (new_rtx) = note;
5442 delete_insn_and_edges (insn);
5443 insn = new_rtx;
5445 else
5446 INSN_CODE (insn) = -1;
5448 /* Do not bother deleting any unreachable code, let jump do it. */
5449 cse_jumps_altered = true;
5450 sets[i].rtl = 0;
5453 /* If destination is volatile, invalidate it and then do no further
5454 processing for this assignment. */
5456 else if (do_not_record)
5458 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5459 invalidate (dest, VOIDmode);
5460 else if (MEM_P (dest))
5461 invalidate (dest, VOIDmode);
5462 else if (GET_CODE (dest) == STRICT_LOW_PART
5463 || GET_CODE (dest) == ZERO_EXTRACT)
5464 invalidate (XEXP (dest, 0), GET_MODE (dest));
5465 sets[i].rtl = 0;
5468 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5469 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5471 #ifdef HAVE_cc0
5472 /* If setting CC0, record what it was set to, or a constant, if it
5473 is equivalent to a constant. If it is being set to a floating-point
5474 value, make a COMPARE with the appropriate constant of 0. If we
5475 don't do this, later code can interpret this as a test against
5476 const0_rtx, which can cause problems if we try to put it into an
5477 insn as a floating-point operand. */
5478 if (dest == cc0_rtx)
5480 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5481 this_insn_cc0_mode = mode;
5482 if (FLOAT_MODE_P (mode))
5483 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5484 CONST0_RTX (mode));
5486 #endif
5489 /* Now enter all non-volatile source expressions in the hash table
5490 if they are not already present.
5491 Record their equivalence classes in src_elt.
5492 This way we can insert the corresponding destinations into
5493 the same classes even if the actual sources are no longer in them
5494 (having been invalidated). */
5496 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5497 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5499 struct table_elt *elt;
5500 struct table_elt *classp = sets[0].src_elt;
5501 rtx dest = SET_DEST (sets[0].rtl);
5502 enum machine_mode eqvmode = GET_MODE (dest);
5504 if (GET_CODE (dest) == STRICT_LOW_PART)
5506 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5507 classp = 0;
5509 if (insert_regs (src_eqv, classp, 0))
5511 rehash_using_reg (src_eqv);
5512 src_eqv_hash = HASH (src_eqv, eqvmode);
5514 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5515 elt->in_memory = src_eqv_in_memory;
5516 src_eqv_elt = elt;
5518 /* Check to see if src_eqv_elt is the same as a set source which
5519 does not yet have an elt, and if so set the elt of the set source
5520 to src_eqv_elt. */
5521 for (i = 0; i < n_sets; i++)
5522 if (sets[i].rtl && sets[i].src_elt == 0
5523 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5524 sets[i].src_elt = src_eqv_elt;
5527 for (i = 0; i < n_sets; i++)
5528 if (sets[i].rtl && ! sets[i].src_volatile
5529 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5531 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5533 /* REG_EQUAL in setting a STRICT_LOW_PART
5534 gives an equivalent for the entire destination register,
5535 not just for the subreg being stored in now.
5536 This is a more interesting equivalence, so we arrange later
5537 to treat the entire reg as the destination. */
5538 sets[i].src_elt = src_eqv_elt;
5539 sets[i].src_hash = src_eqv_hash;
5541 else
5543 /* Insert source and constant equivalent into hash table, if not
5544 already present. */
5545 struct table_elt *classp = src_eqv_elt;
5546 rtx src = sets[i].src;
5547 rtx dest = SET_DEST (sets[i].rtl);
5548 enum machine_mode mode
5549 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5551 /* It's possible that we have a source value known to be
5552 constant but don't have a REG_EQUAL note on the insn.
5553 Lack of a note will mean src_eqv_elt will be NULL. This
5554 can happen where we've generated a SUBREG to access a
5555 CONST_INT that is already in a register in a wider mode.
5556 Ensure that the source expression is put in the proper
5557 constant class. */
5558 if (!classp)
5559 classp = sets[i].src_const_elt;
5561 if (sets[i].src_elt == 0)
5563 struct table_elt *elt;
5565 /* Note that these insert_regs calls cannot remove
5566 any of the src_elt's, because they would have failed to
5567 match if not still valid. */
5568 if (insert_regs (src, classp, 0))
5570 rehash_using_reg (src);
5571 sets[i].src_hash = HASH (src, mode);
5573 elt = insert (src, classp, sets[i].src_hash, mode);
5574 elt->in_memory = sets[i].src_in_memory;
5575 sets[i].src_elt = classp = elt;
5577 if (sets[i].src_const && sets[i].src_const_elt == 0
5578 && src != sets[i].src_const
5579 && ! rtx_equal_p (sets[i].src_const, src))
5580 sets[i].src_elt = insert (sets[i].src_const, classp,
5581 sets[i].src_const_hash, mode);
5584 else if (sets[i].src_elt == 0)
5585 /* If we did not insert the source into the hash table (e.g., it was
5586 volatile), note the equivalence class for the REG_EQUAL value, if any,
5587 so that the destination goes into that class. */
5588 sets[i].src_elt = src_eqv_elt;
5590 /* Record destination addresses in the hash table. This allows us to
5591 check if they are invalidated by other sets. */
5592 for (i = 0; i < n_sets; i++)
5594 if (sets[i].rtl)
5596 rtx x = sets[i].inner_dest;
5597 struct table_elt *elt;
5598 enum machine_mode mode;
5599 unsigned hash;
5601 if (MEM_P (x))
5603 x = XEXP (x, 0);
5604 mode = GET_MODE (x);
5605 hash = HASH (x, mode);
5606 elt = lookup (x, hash, mode);
5607 if (!elt)
5609 if (insert_regs (x, NULL, 0))
5611 rtx dest = SET_DEST (sets[i].rtl);
5613 rehash_using_reg (x);
5614 hash = HASH (x, mode);
5615 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5617 elt = insert (x, NULL, hash, mode);
5620 sets[i].dest_addr_elt = elt;
5622 else
5623 sets[i].dest_addr_elt = NULL;
5627 invalidate_from_clobbers (insn);
5629 /* Some registers are invalidated by subroutine calls. Memory is
5630 invalidated by non-constant calls. */
5632 if (CALL_P (insn))
5634 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5635 invalidate_memory ();
5636 invalidate_for_call ();
5639 /* Now invalidate everything set by this instruction.
5640 If a SUBREG or other funny destination is being set,
5641 sets[i].rtl is still nonzero, so here we invalidate the reg
5642 a part of which is being set. */
5644 for (i = 0; i < n_sets; i++)
5645 if (sets[i].rtl)
5647 /* We can't use the inner dest, because the mode associated with
5648 a ZERO_EXTRACT is significant. */
5649 rtx dest = SET_DEST (sets[i].rtl);
5651 /* Needed for registers to remove the register from its
5652 previous quantity's chain.
5653 Needed for memory if this is a nonvarying address, unless
5654 we have just done an invalidate_memory that covers even those. */
5655 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5656 invalidate (dest, VOIDmode);
5657 else if (MEM_P (dest))
5658 invalidate (dest, VOIDmode);
5659 else if (GET_CODE (dest) == STRICT_LOW_PART
5660 || GET_CODE (dest) == ZERO_EXTRACT)
5661 invalidate (XEXP (dest, 0), GET_MODE (dest));
5664 /* A volatile ASM or an UNSPEC_VOLATILE invalidates everything. */
5665 if (NONJUMP_INSN_P (insn)
5666 && volatile_insn_p (PATTERN (insn)))
5667 flush_hash_table ();
5669 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5670 the regs restored by the longjmp come from a later time
5671 than the setjmp. */
5672 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5674 flush_hash_table ();
5675 goto done;
5678 /* Make sure registers mentioned in destinations
5679 are safe for use in an expression to be inserted.
5680 This removes from the hash table
5681 any invalid entry that refers to one of these registers.
5683 We don't care about the return value from mention_regs because
5684 we are going to hash the SET_DEST values unconditionally. */
5686 for (i = 0; i < n_sets; i++)
5688 if (sets[i].rtl)
5690 rtx x = SET_DEST (sets[i].rtl);
5692 if (!REG_P (x))
5693 mention_regs (x);
5694 else
5696 /* We used to rely on all references to a register becoming
5697 inaccessible when a register changes to a new quantity,
5698 since that changes the hash code. However, that is not
5699 safe, since after HASH_SIZE new quantities we get a
5700 hash 'collision' of a register with its own invalid
5701 entries. And since SUBREGs have been changed not to
5702 change their hash code with the hash code of the register,
5703 it wouldn't work any longer at all. So we have to check
5704 for any invalid references lying around now.
5705 This code is similar to the REG case in mention_regs,
5706 but it knows that reg_tick has been incremented, and
5707 it leaves reg_in_table as -1 . */
5708 unsigned int regno = REGNO (x);
5709 unsigned int endregno = END_REGNO (x);
5710 unsigned int i;
5712 for (i = regno; i < endregno; i++)
5714 if (REG_IN_TABLE (i) >= 0)
5716 remove_invalid_refs (i);
5717 REG_IN_TABLE (i) = -1;
5724 /* We may have just removed some of the src_elt's from the hash table.
5725 So replace each one with the current head of the same class.
5726 Also check if destination addresses have been removed. */
5728 for (i = 0; i < n_sets; i++)
5729 if (sets[i].rtl)
5731 if (sets[i].dest_addr_elt
5732 && sets[i].dest_addr_elt->first_same_value == 0)
5734 /* The elt was removed, which means this destination is not
5735 valid after this instruction. */
5736 sets[i].rtl = NULL_RTX;
5738 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5739 /* If elt was removed, find current head of same class,
5740 or 0 if nothing remains of that class. */
5742 struct table_elt *elt = sets[i].src_elt;
5744 while (elt && elt->prev_same_value)
5745 elt = elt->prev_same_value;
5747 while (elt && elt->first_same_value == 0)
5748 elt = elt->next_same_value;
5749 sets[i].src_elt = elt ? elt->first_same_value : 0;
5753 /* Now insert the destinations into their equivalence classes. */
5755 for (i = 0; i < n_sets; i++)
5756 if (sets[i].rtl)
5758 rtx dest = SET_DEST (sets[i].rtl);
5759 struct table_elt *elt;
5761 /* Don't record value if we are not supposed to risk allocating
5762 floating-point values in registers that might be wider than
5763 memory. */
5764 if ((flag_float_store
5765 && MEM_P (dest)
5766 && FLOAT_MODE_P (GET_MODE (dest)))
5767 /* Don't record BLKmode values, because we don't know the
5768 size of it, and can't be sure that other BLKmode values
5769 have the same or smaller size. */
5770 || GET_MODE (dest) == BLKmode
5771 /* If we didn't put a REG_EQUAL value or a source into the hash
5772 table, there is no point is recording DEST. */
5773 || sets[i].src_elt == 0
5774 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5775 or SIGN_EXTEND, don't record DEST since it can cause
5776 some tracking to be wrong.
5778 ??? Think about this more later. */
5779 || (paradoxical_subreg_p (dest)
5780 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5781 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5782 continue;
5784 /* STRICT_LOW_PART isn't part of the value BEING set,
5785 and neither is the SUBREG inside it.
5786 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5787 if (GET_CODE (dest) == STRICT_LOW_PART)
5788 dest = SUBREG_REG (XEXP (dest, 0));
5790 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5791 /* Registers must also be inserted into chains for quantities. */
5792 if (insert_regs (dest, sets[i].src_elt, 1))
5794 /* If `insert_regs' changes something, the hash code must be
5795 recalculated. */
5796 rehash_using_reg (dest);
5797 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5800 elt = insert (dest, sets[i].src_elt,
5801 sets[i].dest_hash, GET_MODE (dest));
5803 /* If this is a constant, insert the constant anchors with the
5804 equivalent register-offset expressions using register DEST. */
5805 if (targetm.const_anchor
5806 && REG_P (dest)
5807 && SCALAR_INT_MODE_P (GET_MODE (dest))
5808 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5809 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5811 elt->in_memory = (MEM_P (sets[i].inner_dest)
5812 && !MEM_READONLY_P (sets[i].inner_dest));
5814 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5815 narrower than M2, and both M1 and M2 are the same number of words,
5816 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5817 make that equivalence as well.
5819 However, BAR may have equivalences for which gen_lowpart
5820 will produce a simpler value than gen_lowpart applied to
5821 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5822 BAR's equivalences. If we don't get a simplified form, make
5823 the SUBREG. It will not be used in an equivalence, but will
5824 cause two similar assignments to be detected.
5826 Note the loop below will find SUBREG_REG (DEST) since we have
5827 already entered SRC and DEST of the SET in the table. */
5829 if (GET_CODE (dest) == SUBREG
5830 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5831 / UNITS_PER_WORD)
5832 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5833 && (GET_MODE_SIZE (GET_MODE (dest))
5834 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5835 && sets[i].src_elt != 0)
5837 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5838 struct table_elt *elt, *classp = 0;
5840 for (elt = sets[i].src_elt->first_same_value; elt;
5841 elt = elt->next_same_value)
5843 rtx new_src = 0;
5844 unsigned src_hash;
5845 struct table_elt *src_elt;
5846 int byte = 0;
5848 /* Ignore invalid entries. */
5849 if (!REG_P (elt->exp)
5850 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5851 continue;
5853 /* We may have already been playing subreg games. If the
5854 mode is already correct for the destination, use it. */
5855 if (GET_MODE (elt->exp) == new_mode)
5856 new_src = elt->exp;
5857 else
5859 /* Calculate big endian correction for the SUBREG_BYTE.
5860 We have already checked that M1 (GET_MODE (dest))
5861 is not narrower than M2 (new_mode). */
5862 if (BYTES_BIG_ENDIAN)
5863 byte = (GET_MODE_SIZE (GET_MODE (dest))
5864 - GET_MODE_SIZE (new_mode));
5866 new_src = simplify_gen_subreg (new_mode, elt->exp,
5867 GET_MODE (dest), byte);
5870 /* The call to simplify_gen_subreg fails if the value
5871 is VOIDmode, yet we can't do any simplification, e.g.
5872 for EXPR_LISTs denoting function call results.
5873 It is invalid to construct a SUBREG with a VOIDmode
5874 SUBREG_REG, hence a zero new_src means we can't do
5875 this substitution. */
5876 if (! new_src)
5877 continue;
5879 src_hash = HASH (new_src, new_mode);
5880 src_elt = lookup (new_src, src_hash, new_mode);
5882 /* Put the new source in the hash table is if isn't
5883 already. */
5884 if (src_elt == 0)
5886 if (insert_regs (new_src, classp, 0))
5888 rehash_using_reg (new_src);
5889 src_hash = HASH (new_src, new_mode);
5891 src_elt = insert (new_src, classp, src_hash, new_mode);
5892 src_elt->in_memory = elt->in_memory;
5894 else if (classp && classp != src_elt->first_same_value)
5895 /* Show that two things that we've seen before are
5896 actually the same. */
5897 merge_equiv_classes (src_elt, classp);
5899 classp = src_elt->first_same_value;
5900 /* Ignore invalid entries. */
5901 while (classp
5902 && !REG_P (classp->exp)
5903 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5904 classp = classp->next_same_value;
5909 /* Special handling for (set REG0 REG1) where REG0 is the
5910 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5911 be used in the sequel, so (if easily done) change this insn to
5912 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5913 that computed their value. Then REG1 will become a dead store
5914 and won't cloud the situation for later optimizations.
5916 Do not make this change if REG1 is a hard register, because it will
5917 then be used in the sequel and we may be changing a two-operand insn
5918 into a three-operand insn.
5920 Also do not do this if we are operating on a copy of INSN. */
5922 if (n_sets == 1 && sets[0].rtl)
5923 try_back_substitute_reg (sets[0].rtl, insn);
5925 done:;
5928 /* Remove from the hash table all expressions that reference memory. */
5930 static void
5931 invalidate_memory (void)
5933 int i;
5934 struct table_elt *p, *next;
5936 for (i = 0; i < HASH_SIZE; i++)
5937 for (p = table[i]; p; p = next)
5939 next = p->next_same_hash;
5940 if (p->in_memory)
5941 remove_from_table (p, i);
5945 /* Perform invalidation on the basis of everything about INSN,
5946 except for invalidating the actual places that are SET in it.
5947 This includes the places CLOBBERed, and anything that might
5948 alias with something that is SET or CLOBBERed. */
5950 static void
5951 invalidate_from_clobbers (rtx insn)
5953 rtx x = PATTERN (insn);
5955 if (GET_CODE (x) == CLOBBER)
5957 rtx ref = XEXP (x, 0);
5958 if (ref)
5960 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5961 || MEM_P (ref))
5962 invalidate (ref, VOIDmode);
5963 else if (GET_CODE (ref) == STRICT_LOW_PART
5964 || GET_CODE (ref) == ZERO_EXTRACT)
5965 invalidate (XEXP (ref, 0), GET_MODE (ref));
5968 else if (GET_CODE (x) == PARALLEL)
5970 int i;
5971 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5973 rtx y = XVECEXP (x, 0, i);
5974 if (GET_CODE (y) == CLOBBER)
5976 rtx ref = XEXP (y, 0);
5977 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5978 || MEM_P (ref))
5979 invalidate (ref, VOIDmode);
5980 else if (GET_CODE (ref) == STRICT_LOW_PART
5981 || GET_CODE (ref) == ZERO_EXTRACT)
5982 invalidate (XEXP (ref, 0), GET_MODE (ref));
5988 /* Perform invalidation on the basis of everything about INSN.
5989 This includes the places CLOBBERed, and anything that might
5990 alias with something that is SET or CLOBBERed. */
5992 static void
5993 invalidate_from_sets_and_clobbers (rtx insn)
5995 rtx tem;
5996 rtx x = PATTERN (insn);
5998 if (CALL_P (insn))
6000 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6001 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6002 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6005 /* Ensure we invalidate the destination register of a CALL insn.
6006 This is necessary for machines where this register is a fixed_reg,
6007 because no other code would invalidate it. */
6008 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6009 invalidate (SET_DEST (x), VOIDmode);
6011 else if (GET_CODE (x) == PARALLEL)
6013 int i;
6015 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6017 rtx y = XVECEXP (x, 0, i);
6018 if (GET_CODE (y) == CLOBBER)
6020 rtx clobbered = XEXP (y, 0);
6022 if (REG_P (clobbered)
6023 || GET_CODE (clobbered) == SUBREG)
6024 invalidate (clobbered, VOIDmode);
6025 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6026 || GET_CODE (clobbered) == ZERO_EXTRACT)
6027 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6029 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6030 invalidate (SET_DEST (y), VOIDmode);
6035 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6036 and replace any registers in them with either an equivalent constant
6037 or the canonical form of the register. If we are inside an address,
6038 only do this if the address remains valid.
6040 OBJECT is 0 except when within a MEM in which case it is the MEM.
6042 Return the replacement for X. */
6044 static rtx
6045 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6047 enum rtx_code code = GET_CODE (x);
6048 const char *fmt = GET_RTX_FORMAT (code);
6049 int i;
6051 switch (code)
6053 case CONST:
6054 case SYMBOL_REF:
6055 case LABEL_REF:
6056 CASE_CONST_ANY:
6057 case PC:
6058 case CC0:
6059 case LO_SUM:
6060 return x;
6062 case MEM:
6063 validate_change (x, &XEXP (x, 0),
6064 cse_process_notes (XEXP (x, 0), x, changed), 0);
6065 return x;
6067 case EXPR_LIST:
6068 case INSN_LIST:
6069 if (REG_NOTE_KIND (x) == REG_EQUAL)
6070 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6071 if (XEXP (x, 1))
6072 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6073 return x;
6075 case SIGN_EXTEND:
6076 case ZERO_EXTEND:
6077 case SUBREG:
6079 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6080 /* We don't substitute VOIDmode constants into these rtx,
6081 since they would impede folding. */
6082 if (GET_MODE (new_rtx) != VOIDmode)
6083 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6084 return x;
6087 case REG:
6088 i = REG_QTY (REGNO (x));
6090 /* Return a constant or a constant register. */
6091 if (REGNO_QTY_VALID_P (REGNO (x)))
6093 struct qty_table_elem *ent = &qty_table[i];
6095 if (ent->const_rtx != NULL_RTX
6096 && (CONSTANT_P (ent->const_rtx)
6097 || REG_P (ent->const_rtx)))
6099 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6100 if (new_rtx)
6101 return copy_rtx (new_rtx);
6105 /* Otherwise, canonicalize this register. */
6106 return canon_reg (x, NULL_RTX);
6108 default:
6109 break;
6112 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6113 if (fmt[i] == 'e')
6114 validate_change (object, &XEXP (x, i),
6115 cse_process_notes (XEXP (x, i), object, changed), 0);
6117 return x;
6120 static rtx
6121 cse_process_notes (rtx x, rtx object, bool *changed)
6123 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6124 if (new_rtx != x)
6125 *changed = true;
6126 return new_rtx;
6130 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6132 DATA is a pointer to a struct cse_basic_block_data, that is used to
6133 describe the path.
6134 It is filled with a queue of basic blocks, starting with FIRST_BB
6135 and following a trace through the CFG.
6137 If all paths starting at FIRST_BB have been followed, or no new path
6138 starting at FIRST_BB can be constructed, this function returns FALSE.
6139 Otherwise, DATA->path is filled and the function returns TRUE indicating
6140 that a path to follow was found.
6142 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6143 block in the path will be FIRST_BB. */
6145 static bool
6146 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6147 int follow_jumps)
6149 basic_block bb;
6150 edge e;
6151 int path_size;
6153 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6155 /* See if there is a previous path. */
6156 path_size = data->path_size;
6158 /* There is a previous path. Make sure it started with FIRST_BB. */
6159 if (path_size)
6160 gcc_assert (data->path[0].bb == first_bb);
6162 /* There was only one basic block in the last path. Clear the path and
6163 return, so that paths starting at another basic block can be tried. */
6164 if (path_size == 1)
6166 path_size = 0;
6167 goto done;
6170 /* If the path was empty from the beginning, construct a new path. */
6171 if (path_size == 0)
6172 data->path[path_size++].bb = first_bb;
6173 else
6175 /* Otherwise, path_size must be equal to or greater than 2, because
6176 a previous path exists that is at least two basic blocks long.
6178 Update the previous branch path, if any. If the last branch was
6179 previously along the branch edge, take the fallthrough edge now. */
6180 while (path_size >= 2)
6182 basic_block last_bb_in_path, previous_bb_in_path;
6183 edge e;
6185 --path_size;
6186 last_bb_in_path = data->path[path_size].bb;
6187 previous_bb_in_path = data->path[path_size - 1].bb;
6189 /* If we previously followed a path along the branch edge, try
6190 the fallthru edge now. */
6191 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6192 && any_condjump_p (BB_END (previous_bb_in_path))
6193 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6194 && e == BRANCH_EDGE (previous_bb_in_path))
6196 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6197 if (bb != EXIT_BLOCK_PTR
6198 && single_pred_p (bb)
6199 /* We used to assert here that we would only see blocks
6200 that we have not visited yet. But we may end up
6201 visiting basic blocks twice if the CFG has changed
6202 in this run of cse_main, because when the CFG changes
6203 the topological sort of the CFG also changes. A basic
6204 blocks that previously had more than two predecessors
6205 may now have a single predecessor, and become part of
6206 a path that starts at another basic block.
6208 We still want to visit each basic block only once, so
6209 halt the path here if we have already visited BB. */
6210 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6212 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6213 data->path[path_size++].bb = bb;
6214 break;
6218 data->path[path_size].bb = NULL;
6221 /* If only one block remains in the path, bail. */
6222 if (path_size == 1)
6224 path_size = 0;
6225 goto done;
6229 /* Extend the path if possible. */
6230 if (follow_jumps)
6232 bb = data->path[path_size - 1].bb;
6233 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6235 if (single_succ_p (bb))
6236 e = single_succ_edge (bb);
6237 else if (EDGE_COUNT (bb->succs) == 2
6238 && any_condjump_p (BB_END (bb)))
6240 /* First try to follow the branch. If that doesn't lead
6241 to a useful path, follow the fallthru edge. */
6242 e = BRANCH_EDGE (bb);
6243 if (!single_pred_p (e->dest))
6244 e = FALLTHRU_EDGE (bb);
6246 else
6247 e = NULL;
6249 if (e
6250 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6251 && e->dest != EXIT_BLOCK_PTR
6252 && single_pred_p (e->dest)
6253 /* Avoid visiting basic blocks twice. The large comment
6254 above explains why this can happen. */
6255 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6257 basic_block bb2 = e->dest;
6258 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6259 data->path[path_size++].bb = bb2;
6260 bb = bb2;
6262 else
6263 bb = NULL;
6267 done:
6268 data->path_size = path_size;
6269 return path_size != 0;
6272 /* Dump the path in DATA to file F. NSETS is the number of sets
6273 in the path. */
6275 static void
6276 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6278 int path_entry;
6280 fprintf (f, ";; Following path with %d sets: ", nsets);
6281 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6282 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6283 fputc ('\n', dump_file);
6284 fflush (f);
6288 /* Return true if BB has exception handling successor edges. */
6290 static bool
6291 have_eh_succ_edges (basic_block bb)
6293 edge e;
6294 edge_iterator ei;
6296 FOR_EACH_EDGE (e, ei, bb->succs)
6297 if (e->flags & EDGE_EH)
6298 return true;
6300 return false;
6304 /* Scan to the end of the path described by DATA. Return an estimate of
6305 the total number of SETs of all insns in the path. */
6307 static void
6308 cse_prescan_path (struct cse_basic_block_data *data)
6310 int nsets = 0;
6311 int path_size = data->path_size;
6312 int path_entry;
6314 /* Scan to end of each basic block in the path. */
6315 for (path_entry = 0; path_entry < path_size; path_entry++)
6317 basic_block bb;
6318 rtx insn;
6320 bb = data->path[path_entry].bb;
6322 FOR_BB_INSNS (bb, insn)
6324 if (!INSN_P (insn))
6325 continue;
6327 /* A PARALLEL can have lots of SETs in it,
6328 especially if it is really an ASM_OPERANDS. */
6329 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6330 nsets += XVECLEN (PATTERN (insn), 0);
6331 else
6332 nsets += 1;
6336 data->nsets = nsets;
6339 /* Process a single extended basic block described by EBB_DATA. */
6341 static void
6342 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6344 int path_size = ebb_data->path_size;
6345 int path_entry;
6346 int num_insns = 0;
6348 /* Allocate the space needed by qty_table. */
6349 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6351 new_basic_block ();
6352 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6353 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6354 for (path_entry = 0; path_entry < path_size; path_entry++)
6356 basic_block bb;
6357 rtx insn;
6359 bb = ebb_data->path[path_entry].bb;
6361 /* Invalidate recorded information for eh regs if there is an EH
6362 edge pointing to that bb. */
6363 if (bb_has_eh_pred (bb))
6365 df_ref *def_rec;
6367 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6369 df_ref def = *def_rec;
6370 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6371 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6375 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6376 FOR_BB_INSNS (bb, insn)
6378 /* If we have processed 1,000 insns, flush the hash table to
6379 avoid extreme quadratic behavior. We must not include NOTEs
6380 in the count since there may be more of them when generating
6381 debugging information. If we clear the table at different
6382 times, code generated with -g -O might be different than code
6383 generated with -O but not -g.
6385 FIXME: This is a real kludge and needs to be done some other
6386 way. */
6387 if (NONDEBUG_INSN_P (insn)
6388 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6390 flush_hash_table ();
6391 num_insns = 0;
6394 if (INSN_P (insn))
6396 /* Process notes first so we have all notes in canonical forms
6397 when looking for duplicate operations. */
6398 if (REG_NOTES (insn))
6400 bool changed = false;
6401 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6402 NULL_RTX, &changed);
6403 if (changed)
6404 df_notes_rescan (insn);
6407 cse_insn (insn);
6409 /* If we haven't already found an insn where we added a LABEL_REF,
6410 check this one. */
6411 if (INSN_P (insn) && !recorded_label_ref
6412 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6413 (void *) insn))
6414 recorded_label_ref = true;
6416 #ifdef HAVE_cc0
6417 if (NONDEBUG_INSN_P (insn))
6419 /* If the previous insn sets CC0 and this insn no
6420 longer references CC0, delete the previous insn.
6421 Here we use fact that nothing expects CC0 to be
6422 valid over an insn, which is true until the final
6423 pass. */
6424 rtx prev_insn, tem;
6426 prev_insn = prev_nonnote_nondebug_insn (insn);
6427 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6428 && (tem = single_set (prev_insn)) != NULL_RTX
6429 && SET_DEST (tem) == cc0_rtx
6430 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6431 delete_insn (prev_insn);
6433 /* If this insn is not the last insn in the basic
6434 block, it will be PREV_INSN(insn) in the next
6435 iteration. If we recorded any CC0-related
6436 information for this insn, remember it. */
6437 if (insn != BB_END (bb))
6439 prev_insn_cc0 = this_insn_cc0;
6440 prev_insn_cc0_mode = this_insn_cc0_mode;
6443 #endif
6447 /* With non-call exceptions, we are not always able to update
6448 the CFG properly inside cse_insn. So clean up possibly
6449 redundant EH edges here. */
6450 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6451 cse_cfg_altered |= purge_dead_edges (bb);
6453 /* If we changed a conditional jump, we may have terminated
6454 the path we are following. Check that by verifying that
6455 the edge we would take still exists. If the edge does
6456 not exist anymore, purge the remainder of the path.
6457 Note that this will cause us to return to the caller. */
6458 if (path_entry < path_size - 1)
6460 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6461 if (!find_edge (bb, next_bb))
6465 path_size--;
6467 /* If we truncate the path, we must also reset the
6468 visited bit on the remaining blocks in the path,
6469 or we will never visit them at all. */
6470 bitmap_clear_bit (cse_visited_basic_blocks,
6471 ebb_data->path[path_size].bb->index);
6472 ebb_data->path[path_size].bb = NULL;
6474 while (path_size - 1 != path_entry);
6475 ebb_data->path_size = path_size;
6479 /* If this is a conditional jump insn, record any known
6480 equivalences due to the condition being tested. */
6481 insn = BB_END (bb);
6482 if (path_entry < path_size - 1
6483 && JUMP_P (insn)
6484 && single_set (insn)
6485 && any_condjump_p (insn))
6487 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6488 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6489 record_jump_equiv (insn, taken);
6492 #ifdef HAVE_cc0
6493 /* Clear the CC0-tracking related insns, they can't provide
6494 useful information across basic block boundaries. */
6495 prev_insn_cc0 = 0;
6496 #endif
6499 gcc_assert (next_qty <= max_qty);
6501 free (qty_table);
6505 /* Perform cse on the instructions of a function.
6506 F is the first instruction.
6507 NREGS is one plus the highest pseudo-reg number used in the instruction.
6509 Return 2 if jump optimizations should be redone due to simplifications
6510 in conditional jump instructions.
6511 Return 1 if the CFG should be cleaned up because it has been modified.
6512 Return 0 otherwise. */
6514 static int
6515 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6517 struct cse_basic_block_data ebb_data;
6518 basic_block bb;
6519 int *rc_order = XNEWVEC (int, last_basic_block);
6520 int i, n_blocks;
6522 df_set_flags (DF_LR_RUN_DCE);
6523 df_analyze ();
6524 df_set_flags (DF_DEFER_INSN_RESCAN);
6526 reg_scan (get_insns (), max_reg_num ());
6527 init_cse_reg_info (nregs);
6529 ebb_data.path = XNEWVEC (struct branch_path,
6530 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6532 cse_cfg_altered = false;
6533 cse_jumps_altered = false;
6534 recorded_label_ref = false;
6535 constant_pool_entries_cost = 0;
6536 constant_pool_entries_regcost = 0;
6537 ebb_data.path_size = 0;
6538 ebb_data.nsets = 0;
6539 rtl_hooks = cse_rtl_hooks;
6541 init_recog ();
6542 init_alias_analysis ();
6544 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6546 /* Set up the table of already visited basic blocks. */
6547 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6548 bitmap_clear (cse_visited_basic_blocks);
6550 /* Loop over basic blocks in reverse completion order (RPO),
6551 excluding the ENTRY and EXIT blocks. */
6552 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6553 i = 0;
6554 while (i < n_blocks)
6556 /* Find the first block in the RPO queue that we have not yet
6557 processed before. */
6560 bb = BASIC_BLOCK (rc_order[i++]);
6562 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6563 && i < n_blocks);
6565 /* Find all paths starting with BB, and process them. */
6566 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6568 /* Pre-scan the path. */
6569 cse_prescan_path (&ebb_data);
6571 /* If this basic block has no sets, skip it. */
6572 if (ebb_data.nsets == 0)
6573 continue;
6575 /* Get a reasonable estimate for the maximum number of qty's
6576 needed for this path. For this, we take the number of sets
6577 and multiply that by MAX_RECOG_OPERANDS. */
6578 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6580 /* Dump the path we're about to process. */
6581 if (dump_file)
6582 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6584 cse_extended_basic_block (&ebb_data);
6588 /* Clean up. */
6589 end_alias_analysis ();
6590 free (reg_eqv_table);
6591 free (ebb_data.path);
6592 sbitmap_free (cse_visited_basic_blocks);
6593 free (rc_order);
6594 rtl_hooks = general_rtl_hooks;
6596 if (cse_jumps_altered || recorded_label_ref)
6597 return 2;
6598 else if (cse_cfg_altered)
6599 return 1;
6600 else
6601 return 0;
6604 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6605 which there isn't a REG_LABEL_OPERAND note.
6606 Return one if so. DATA is the insn. */
6608 static int
6609 check_for_label_ref (rtx *rtl, void *data)
6611 rtx insn = (rtx) data;
6613 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6614 note for it, we must rerun jump since it needs to place the note. If
6615 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6616 don't do this since no REG_LABEL_OPERAND will be added. */
6617 return (GET_CODE (*rtl) == LABEL_REF
6618 && ! LABEL_REF_NONLOCAL_P (*rtl)
6619 && (!JUMP_P (insn)
6620 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6621 && LABEL_P (XEXP (*rtl, 0))
6622 && INSN_UID (XEXP (*rtl, 0)) != 0
6623 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6626 /* Count the number of times registers are used (not set) in X.
6627 COUNTS is an array in which we accumulate the count, INCR is how much
6628 we count each register usage.
6630 Don't count a usage of DEST, which is the SET_DEST of a SET which
6631 contains X in its SET_SRC. This is because such a SET does not
6632 modify the liveness of DEST.
6633 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6634 We must then count uses of a SET_DEST regardless, because the insn can't be
6635 deleted here. */
6637 static void
6638 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6640 enum rtx_code code;
6641 rtx note;
6642 const char *fmt;
6643 int i, j;
6645 if (x == 0)
6646 return;
6648 switch (code = GET_CODE (x))
6650 case REG:
6651 if (x != dest)
6652 counts[REGNO (x)] += incr;
6653 return;
6655 case PC:
6656 case CC0:
6657 case CONST:
6658 CASE_CONST_ANY:
6659 case SYMBOL_REF:
6660 case LABEL_REF:
6661 return;
6663 case CLOBBER:
6664 /* If we are clobbering a MEM, mark any registers inside the address
6665 as being used. */
6666 if (MEM_P (XEXP (x, 0)))
6667 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6668 return;
6670 case SET:
6671 /* Unless we are setting a REG, count everything in SET_DEST. */
6672 if (!REG_P (SET_DEST (x)))
6673 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6674 count_reg_usage (SET_SRC (x), counts,
6675 dest ? dest : SET_DEST (x),
6676 incr);
6677 return;
6679 case DEBUG_INSN:
6680 return;
6682 case CALL_INSN:
6683 case INSN:
6684 case JUMP_INSN:
6685 /* We expect dest to be NULL_RTX here. If the insn may throw,
6686 or if it cannot be deleted due to side-effects, mark this fact
6687 by setting DEST to pc_rtx. */
6688 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6689 || side_effects_p (PATTERN (x)))
6690 dest = pc_rtx;
6691 if (code == CALL_INSN)
6692 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6693 count_reg_usage (PATTERN (x), counts, dest, incr);
6695 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6696 use them. */
6698 note = find_reg_equal_equiv_note (x);
6699 if (note)
6701 rtx eqv = XEXP (note, 0);
6703 if (GET_CODE (eqv) == EXPR_LIST)
6704 /* This REG_EQUAL note describes the result of a function call.
6705 Process all the arguments. */
6708 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6709 eqv = XEXP (eqv, 1);
6711 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6712 else
6713 count_reg_usage (eqv, counts, dest, incr);
6715 return;
6717 case EXPR_LIST:
6718 if (REG_NOTE_KIND (x) == REG_EQUAL
6719 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6720 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6721 involving registers in the address. */
6722 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6723 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6725 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6726 return;
6728 case ASM_OPERANDS:
6729 /* Iterate over just the inputs, not the constraints as well. */
6730 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6731 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6732 return;
6734 case INSN_LIST:
6735 gcc_unreachable ();
6737 default:
6738 break;
6741 fmt = GET_RTX_FORMAT (code);
6742 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6744 if (fmt[i] == 'e')
6745 count_reg_usage (XEXP (x, i), counts, dest, incr);
6746 else if (fmt[i] == 'E')
6747 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6748 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6752 /* Return true if X is a dead register. */
6754 static inline int
6755 is_dead_reg (rtx x, int *counts)
6757 return (REG_P (x)
6758 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6759 && counts[REGNO (x)] == 0);
6762 /* Return true if set is live. */
6763 static bool
6764 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6765 int *counts)
6767 #ifdef HAVE_cc0
6768 rtx tem;
6769 #endif
6771 if (set_noop_p (set))
6774 #ifdef HAVE_cc0
6775 else if (GET_CODE (SET_DEST (set)) == CC0
6776 && !side_effects_p (SET_SRC (set))
6777 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6778 || !INSN_P (tem)
6779 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6780 return false;
6781 #endif
6782 else if (!is_dead_reg (SET_DEST (set), counts)
6783 || side_effects_p (SET_SRC (set)))
6784 return true;
6785 return false;
6788 /* Return true if insn is live. */
6790 static bool
6791 insn_live_p (rtx insn, int *counts)
6793 int i;
6794 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6795 return true;
6796 else if (GET_CODE (PATTERN (insn)) == SET)
6797 return set_live_p (PATTERN (insn), insn, counts);
6798 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6800 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6802 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6804 if (GET_CODE (elt) == SET)
6806 if (set_live_p (elt, insn, counts))
6807 return true;
6809 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6810 return true;
6812 return false;
6814 else if (DEBUG_INSN_P (insn))
6816 rtx next;
6818 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6819 if (NOTE_P (next))
6820 continue;
6821 else if (!DEBUG_INSN_P (next))
6822 return true;
6823 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6824 return false;
6826 return true;
6828 else
6829 return true;
6832 /* Count the number of stores into pseudo. Callback for note_stores. */
6834 static void
6835 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6837 int *counts = (int *) data;
6838 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6839 counts[REGNO (x)]++;
6842 struct dead_debug_insn_data
6844 int *counts;
6845 rtx *replacements;
6846 bool seen_repl;
6849 /* Return if a DEBUG_INSN needs to be reset because some dead
6850 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6852 static int
6853 is_dead_debug_insn (rtx *loc, void *data)
6855 rtx x = *loc;
6856 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6858 if (is_dead_reg (x, ddid->counts))
6860 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6861 ddid->seen_repl = true;
6862 else
6863 return 1;
6865 return 0;
6868 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6869 Callback for simplify_replace_fn_rtx. */
6871 static rtx
6872 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6874 rtx *replacements = (rtx *) data;
6876 if (REG_P (x)
6877 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6878 && replacements[REGNO (x)] != NULL_RTX)
6880 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6881 return replacements[REGNO (x)];
6882 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6883 GET_MODE (replacements[REGNO (x)]));
6885 return NULL_RTX;
6888 /* Scan all the insns and delete any that are dead; i.e., they store a register
6889 that is never used or they copy a register to itself.
6891 This is used to remove insns made obviously dead by cse, loop or other
6892 optimizations. It improves the heuristics in loop since it won't try to
6893 move dead invariants out of loops or make givs for dead quantities. The
6894 remaining passes of the compilation are also sped up. */
6897 delete_trivially_dead_insns (rtx insns, int nreg)
6899 int *counts;
6900 rtx insn, prev;
6901 rtx *replacements = NULL;
6902 int ndead = 0;
6904 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6905 /* First count the number of times each register is used. */
6906 if (MAY_HAVE_DEBUG_INSNS)
6908 counts = XCNEWVEC (int, nreg * 3);
6909 for (insn = insns; insn; insn = NEXT_INSN (insn))
6910 if (DEBUG_INSN_P (insn))
6911 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6912 NULL_RTX, 1);
6913 else if (INSN_P (insn))
6915 count_reg_usage (insn, counts, NULL_RTX, 1);
6916 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6918 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6919 First one counts how many times each pseudo is used outside
6920 of debug insns, second counts how many times each pseudo is
6921 used in debug insns and third counts how many times a pseudo
6922 is stored. */
6924 else
6926 counts = XCNEWVEC (int, nreg);
6927 for (insn = insns; insn; insn = NEXT_INSN (insn))
6928 if (INSN_P (insn))
6929 count_reg_usage (insn, counts, NULL_RTX, 1);
6930 /* If no debug insns can be present, COUNTS is just an array
6931 which counts how many times each pseudo is used. */
6933 /* Go from the last insn to the first and delete insns that only set unused
6934 registers or copy a register to itself. As we delete an insn, remove
6935 usage counts for registers it uses.
6937 The first jump optimization pass may leave a real insn as the last
6938 insn in the function. We must not skip that insn or we may end
6939 up deleting code that is not really dead.
6941 If some otherwise unused register is only used in DEBUG_INSNs,
6942 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6943 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6944 has been created for the unused register, replace it with
6945 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6946 for (insn = get_last_insn (); insn; insn = prev)
6948 int live_insn = 0;
6950 prev = PREV_INSN (insn);
6951 if (!INSN_P (insn))
6952 continue;
6954 live_insn = insn_live_p (insn, counts);
6956 /* If this is a dead insn, delete it and show registers in it aren't
6957 being used. */
6959 if (! live_insn && dbg_cnt (delete_trivial_dead))
6961 if (DEBUG_INSN_P (insn))
6962 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6963 NULL_RTX, -1);
6964 else
6966 rtx set;
6967 if (MAY_HAVE_DEBUG_INSNS
6968 && (set = single_set (insn)) != NULL_RTX
6969 && is_dead_reg (SET_DEST (set), counts)
6970 /* Used at least once in some DEBUG_INSN. */
6971 && counts[REGNO (SET_DEST (set)) + nreg] > 0
6972 /* And set exactly once. */
6973 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
6974 && !side_effects_p (SET_SRC (set))
6975 && asm_noperands (PATTERN (insn)) < 0)
6977 rtx dval, bind;
6979 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
6980 dval = make_debug_expr_from_rtl (SET_DEST (set));
6982 /* Emit a debug bind insn before the insn in which
6983 reg dies. */
6984 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
6985 DEBUG_EXPR_TREE_DECL (dval),
6986 SET_SRC (set),
6987 VAR_INIT_STATUS_INITIALIZED);
6988 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
6990 bind = emit_debug_insn_before (bind, insn);
6991 df_insn_rescan (bind);
6993 if (replacements == NULL)
6994 replacements = XCNEWVEC (rtx, nreg);
6995 replacements[REGNO (SET_DEST (set))] = dval;
6998 count_reg_usage (insn, counts, NULL_RTX, -1);
6999 ndead++;
7001 delete_insn_and_edges (insn);
7005 if (MAY_HAVE_DEBUG_INSNS)
7007 struct dead_debug_insn_data ddid;
7008 ddid.counts = counts;
7009 ddid.replacements = replacements;
7010 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7011 if (DEBUG_INSN_P (insn))
7013 /* If this debug insn references a dead register that wasn't replaced
7014 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7015 ddid.seen_repl = false;
7016 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
7017 is_dead_debug_insn, &ddid))
7019 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7020 df_insn_rescan (insn);
7022 else if (ddid.seen_repl)
7024 INSN_VAR_LOCATION_LOC (insn)
7025 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7026 NULL_RTX, replace_dead_reg,
7027 replacements);
7028 df_insn_rescan (insn);
7031 free (replacements);
7034 if (dump_file && ndead)
7035 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7036 ndead);
7037 /* Clean up. */
7038 free (counts);
7039 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7040 return ndead;
7043 /* This function is called via for_each_rtx. The argument, NEWREG, is
7044 a condition code register with the desired mode. If we are looking
7045 at the same register in a different mode, replace it with
7046 NEWREG. */
7048 static int
7049 cse_change_cc_mode (rtx *loc, void *data)
7051 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7053 if (*loc
7054 && REG_P (*loc)
7055 && REGNO (*loc) == REGNO (args->newreg)
7056 && GET_MODE (*loc) != GET_MODE (args->newreg))
7058 validate_change (args->insn, loc, args->newreg, 1);
7060 return -1;
7062 return 0;
7065 /* Change the mode of any reference to the register REGNO (NEWREG) to
7066 GET_MODE (NEWREG) in INSN. */
7068 static void
7069 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7071 struct change_cc_mode_args args;
7072 int success;
7074 if (!INSN_P (insn))
7075 return;
7077 args.insn = insn;
7078 args.newreg = newreg;
7080 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7081 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7083 /* If the following assertion was triggered, there is most probably
7084 something wrong with the cc_modes_compatible back end function.
7085 CC modes only can be considered compatible if the insn - with the mode
7086 replaced by any of the compatible modes - can still be recognized. */
7087 success = apply_change_group ();
7088 gcc_assert (success);
7091 /* Change the mode of any reference to the register REGNO (NEWREG) to
7092 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7093 any instruction which modifies NEWREG. */
7095 static void
7096 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7098 rtx insn;
7100 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7102 if (! INSN_P (insn))
7103 continue;
7105 if (reg_set_p (newreg, insn))
7106 return;
7108 cse_change_cc_mode_insn (insn, newreg);
7112 /* BB is a basic block which finishes with CC_REG as a condition code
7113 register which is set to CC_SRC. Look through the successors of BB
7114 to find blocks which have a single predecessor (i.e., this one),
7115 and look through those blocks for an assignment to CC_REG which is
7116 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7117 permitted to change the mode of CC_SRC to a compatible mode. This
7118 returns VOIDmode if no equivalent assignments were found.
7119 Otherwise it returns the mode which CC_SRC should wind up with.
7120 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7121 but is passed unmodified down to recursive calls in order to prevent
7122 endless recursion.
7124 The main complexity in this function is handling the mode issues.
7125 We may have more than one duplicate which we can eliminate, and we
7126 try to find a mode which will work for multiple duplicates. */
7128 static enum machine_mode
7129 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7130 bool can_change_mode)
7132 bool found_equiv;
7133 enum machine_mode mode;
7134 unsigned int insn_count;
7135 edge e;
7136 rtx insns[2];
7137 enum machine_mode modes[2];
7138 rtx last_insns[2];
7139 unsigned int i;
7140 rtx newreg;
7141 edge_iterator ei;
7143 /* We expect to have two successors. Look at both before picking
7144 the final mode for the comparison. If we have more successors
7145 (i.e., some sort of table jump, although that seems unlikely),
7146 then we require all beyond the first two to use the same
7147 mode. */
7149 found_equiv = false;
7150 mode = GET_MODE (cc_src);
7151 insn_count = 0;
7152 FOR_EACH_EDGE (e, ei, bb->succs)
7154 rtx insn;
7155 rtx end;
7157 if (e->flags & EDGE_COMPLEX)
7158 continue;
7160 if (EDGE_COUNT (e->dest->preds) != 1
7161 || e->dest == EXIT_BLOCK_PTR
7162 /* Avoid endless recursion on unreachable blocks. */
7163 || e->dest == orig_bb)
7164 continue;
7166 end = NEXT_INSN (BB_END (e->dest));
7167 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7169 rtx set;
7171 if (! INSN_P (insn))
7172 continue;
7174 /* If CC_SRC is modified, we have to stop looking for
7175 something which uses it. */
7176 if (modified_in_p (cc_src, insn))
7177 break;
7179 /* Check whether INSN sets CC_REG to CC_SRC. */
7180 set = single_set (insn);
7181 if (set
7182 && REG_P (SET_DEST (set))
7183 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7185 bool found;
7186 enum machine_mode set_mode;
7187 enum machine_mode comp_mode;
7189 found = false;
7190 set_mode = GET_MODE (SET_SRC (set));
7191 comp_mode = set_mode;
7192 if (rtx_equal_p (cc_src, SET_SRC (set)))
7193 found = true;
7194 else if (GET_CODE (cc_src) == COMPARE
7195 && GET_CODE (SET_SRC (set)) == COMPARE
7196 && mode != set_mode
7197 && rtx_equal_p (XEXP (cc_src, 0),
7198 XEXP (SET_SRC (set), 0))
7199 && rtx_equal_p (XEXP (cc_src, 1),
7200 XEXP (SET_SRC (set), 1)))
7203 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7204 if (comp_mode != VOIDmode
7205 && (can_change_mode || comp_mode == mode))
7206 found = true;
7209 if (found)
7211 found_equiv = true;
7212 if (insn_count < ARRAY_SIZE (insns))
7214 insns[insn_count] = insn;
7215 modes[insn_count] = set_mode;
7216 last_insns[insn_count] = end;
7217 ++insn_count;
7219 if (mode != comp_mode)
7221 gcc_assert (can_change_mode);
7222 mode = comp_mode;
7224 /* The modified insn will be re-recognized later. */
7225 PUT_MODE (cc_src, mode);
7228 else
7230 if (set_mode != mode)
7232 /* We found a matching expression in the
7233 wrong mode, but we don't have room to
7234 store it in the array. Punt. This case
7235 should be rare. */
7236 break;
7238 /* INSN sets CC_REG to a value equal to CC_SRC
7239 with the right mode. We can simply delete
7240 it. */
7241 delete_insn (insn);
7244 /* We found an instruction to delete. Keep looking,
7245 in the hopes of finding a three-way jump. */
7246 continue;
7249 /* We found an instruction which sets the condition
7250 code, so don't look any farther. */
7251 break;
7254 /* If INSN sets CC_REG in some other way, don't look any
7255 farther. */
7256 if (reg_set_p (cc_reg, insn))
7257 break;
7260 /* If we fell off the bottom of the block, we can keep looking
7261 through successors. We pass CAN_CHANGE_MODE as false because
7262 we aren't prepared to handle compatibility between the
7263 further blocks and this block. */
7264 if (insn == end)
7266 enum machine_mode submode;
7268 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7269 if (submode != VOIDmode)
7271 gcc_assert (submode == mode);
7272 found_equiv = true;
7273 can_change_mode = false;
7278 if (! found_equiv)
7279 return VOIDmode;
7281 /* Now INSN_COUNT is the number of instructions we found which set
7282 CC_REG to a value equivalent to CC_SRC. The instructions are in
7283 INSNS. The modes used by those instructions are in MODES. */
7285 newreg = NULL_RTX;
7286 for (i = 0; i < insn_count; ++i)
7288 if (modes[i] != mode)
7290 /* We need to change the mode of CC_REG in INSNS[i] and
7291 subsequent instructions. */
7292 if (! newreg)
7294 if (GET_MODE (cc_reg) == mode)
7295 newreg = cc_reg;
7296 else
7297 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7299 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7300 newreg);
7303 delete_insn_and_edges (insns[i]);
7306 return mode;
7309 /* If we have a fixed condition code register (or two), walk through
7310 the instructions and try to eliminate duplicate assignments. */
7312 static void
7313 cse_condition_code_reg (void)
7315 unsigned int cc_regno_1;
7316 unsigned int cc_regno_2;
7317 rtx cc_reg_1;
7318 rtx cc_reg_2;
7319 basic_block bb;
7321 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7322 return;
7324 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7325 if (cc_regno_2 != INVALID_REGNUM)
7326 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7327 else
7328 cc_reg_2 = NULL_RTX;
7330 FOR_EACH_BB (bb)
7332 rtx last_insn;
7333 rtx cc_reg;
7334 rtx insn;
7335 rtx cc_src_insn;
7336 rtx cc_src;
7337 enum machine_mode mode;
7338 enum machine_mode orig_mode;
7340 /* Look for blocks which end with a conditional jump based on a
7341 condition code register. Then look for the instruction which
7342 sets the condition code register. Then look through the
7343 successor blocks for instructions which set the condition
7344 code register to the same value. There are other possible
7345 uses of the condition code register, but these are by far the
7346 most common and the ones which we are most likely to be able
7347 to optimize. */
7349 last_insn = BB_END (bb);
7350 if (!JUMP_P (last_insn))
7351 continue;
7353 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7354 cc_reg = cc_reg_1;
7355 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7356 cc_reg = cc_reg_2;
7357 else
7358 continue;
7360 cc_src_insn = NULL_RTX;
7361 cc_src = NULL_RTX;
7362 for (insn = PREV_INSN (last_insn);
7363 insn && insn != PREV_INSN (BB_HEAD (bb));
7364 insn = PREV_INSN (insn))
7366 rtx set;
7368 if (! INSN_P (insn))
7369 continue;
7370 set = single_set (insn);
7371 if (set
7372 && REG_P (SET_DEST (set))
7373 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7375 cc_src_insn = insn;
7376 cc_src = SET_SRC (set);
7377 break;
7379 else if (reg_set_p (cc_reg, insn))
7380 break;
7383 if (! cc_src_insn)
7384 continue;
7386 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7387 continue;
7389 /* Now CC_REG is a condition code register used for a
7390 conditional jump at the end of the block, and CC_SRC, in
7391 CC_SRC_INSN, is the value to which that condition code
7392 register is set, and CC_SRC is still meaningful at the end of
7393 the basic block. */
7395 orig_mode = GET_MODE (cc_src);
7396 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7397 if (mode != VOIDmode)
7399 gcc_assert (mode == GET_MODE (cc_src));
7400 if (mode != orig_mode)
7402 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7404 cse_change_cc_mode_insn (cc_src_insn, newreg);
7406 /* Do the same in the following insns that use the
7407 current value of CC_REG within BB. */
7408 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7409 NEXT_INSN (last_insn),
7410 newreg);
7417 /* Perform common subexpression elimination. Nonzero value from
7418 `cse_main' means that jumps were simplified and some code may now
7419 be unreachable, so do jump optimization again. */
7420 static bool
7421 gate_handle_cse (void)
7423 return optimize > 0;
7426 static unsigned int
7427 rest_of_handle_cse (void)
7429 int tem;
7431 if (dump_file)
7432 dump_flow_info (dump_file, dump_flags);
7434 tem = cse_main (get_insns (), max_reg_num ());
7436 /* If we are not running more CSE passes, then we are no longer
7437 expecting CSE to be run. But always rerun it in a cheap mode. */
7438 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7440 if (tem == 2)
7442 timevar_push (TV_JUMP);
7443 rebuild_jump_labels (get_insns ());
7444 cleanup_cfg (CLEANUP_CFG_CHANGED);
7445 timevar_pop (TV_JUMP);
7447 else if (tem == 1 || optimize > 1)
7448 cleanup_cfg (0);
7450 return 0;
7453 struct rtl_opt_pass pass_cse =
7456 RTL_PASS,
7457 "cse1", /* name */
7458 OPTGROUP_NONE, /* optinfo_flags */
7459 gate_handle_cse, /* gate */
7460 rest_of_handle_cse, /* execute */
7461 NULL, /* sub */
7462 NULL, /* next */
7463 0, /* static_pass_number */
7464 TV_CSE, /* tv_id */
7465 0, /* properties_required */
7466 0, /* properties_provided */
7467 0, /* properties_destroyed */
7468 0, /* todo_flags_start */
7469 TODO_df_finish | TODO_verify_rtl_sharing |
7470 TODO_ggc_collect |
7471 TODO_verify_flow, /* todo_flags_finish */
7476 static bool
7477 gate_handle_cse2 (void)
7479 return optimize > 0 && flag_rerun_cse_after_loop;
7482 /* Run second CSE pass after loop optimizations. */
7483 static unsigned int
7484 rest_of_handle_cse2 (void)
7486 int tem;
7488 if (dump_file)
7489 dump_flow_info (dump_file, dump_flags);
7491 tem = cse_main (get_insns (), max_reg_num ());
7493 /* Run a pass to eliminate duplicated assignments to condition code
7494 registers. We have to run this after bypass_jumps, because it
7495 makes it harder for that pass to determine whether a jump can be
7496 bypassed safely. */
7497 cse_condition_code_reg ();
7499 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7501 if (tem == 2)
7503 timevar_push (TV_JUMP);
7504 rebuild_jump_labels (get_insns ());
7505 cleanup_cfg (CLEANUP_CFG_CHANGED);
7506 timevar_pop (TV_JUMP);
7508 else if (tem == 1)
7509 cleanup_cfg (0);
7511 cse_not_expected = 1;
7512 return 0;
7516 struct rtl_opt_pass pass_cse2 =
7519 RTL_PASS,
7520 "cse2", /* name */
7521 OPTGROUP_NONE, /* optinfo_flags */
7522 gate_handle_cse2, /* gate */
7523 rest_of_handle_cse2, /* execute */
7524 NULL, /* sub */
7525 NULL, /* next */
7526 0, /* static_pass_number */
7527 TV_CSE2, /* tv_id */
7528 0, /* properties_required */
7529 0, /* properties_provided */
7530 0, /* properties_destroyed */
7531 0, /* todo_flags_start */
7532 TODO_df_finish | TODO_verify_rtl_sharing |
7533 TODO_ggc_collect |
7534 TODO_verify_flow /* todo_flags_finish */
7538 static bool
7539 gate_handle_cse_after_global_opts (void)
7541 return optimize > 0 && flag_rerun_cse_after_global_opts;
7544 /* Run second CSE pass after loop optimizations. */
7545 static unsigned int
7546 rest_of_handle_cse_after_global_opts (void)
7548 int save_cfj;
7549 int tem;
7551 /* We only want to do local CSE, so don't follow jumps. */
7552 save_cfj = flag_cse_follow_jumps;
7553 flag_cse_follow_jumps = 0;
7555 rebuild_jump_labels (get_insns ());
7556 tem = cse_main (get_insns (), max_reg_num ());
7557 purge_all_dead_edges ();
7558 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7560 cse_not_expected = !flag_rerun_cse_after_loop;
7562 /* If cse altered any jumps, rerun jump opts to clean things up. */
7563 if (tem == 2)
7565 timevar_push (TV_JUMP);
7566 rebuild_jump_labels (get_insns ());
7567 cleanup_cfg (CLEANUP_CFG_CHANGED);
7568 timevar_pop (TV_JUMP);
7570 else if (tem == 1)
7571 cleanup_cfg (0);
7573 flag_cse_follow_jumps = save_cfj;
7574 return 0;
7577 struct rtl_opt_pass pass_cse_after_global_opts =
7580 RTL_PASS,
7581 "cse_local", /* name */
7582 OPTGROUP_NONE, /* optinfo_flags */
7583 gate_handle_cse_after_global_opts, /* gate */
7584 rest_of_handle_cse_after_global_opts, /* execute */
7585 NULL, /* sub */
7586 NULL, /* next */
7587 0, /* static_pass_number */
7588 TV_CSE, /* tv_id */
7589 0, /* properties_required */
7590 0, /* properties_provided */
7591 0, /* properties_destroyed */
7592 0, /* todo_flags_start */
7593 TODO_df_finish | TODO_verify_rtl_sharing |
7594 TODO_ggc_collect |
7595 TODO_verify_flow /* todo_flags_finish */