]>
Commit | Line | Data |
---|---|---|
7afe21cc | 1 | /* Common subexpression elimination for GNU compiler. |
747215f1 | 2 | Copyright (C) 1987, 88, 89, 92-7, 1998, 1999 Free Software Foundation, Inc. |
7afe21cc RK |
3 | |
4 | This file is part of GNU CC. | |
5 | ||
6 | GNU CC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8 | the Free Software Foundation; either version 2, or (at your option) | |
9 | any later version. | |
10 | ||
11 | GNU CC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GNU CC; see the file COPYING. If not, write to | |
940d9d63 RK |
18 | the Free Software Foundation, 59 Temple Place - Suite 330, |
19 | Boston, MA 02111-1307, USA. */ | |
7afe21cc RK |
20 | |
21 | ||
22 | #include "config.h" | |
670ee920 KG |
23 | /* stdio.h must precede rtl.h for FFS. */ |
24 | #include "system.h" | |
50b2596f | 25 | #include <setjmp.h> |
9c3b4c8b | 26 | |
7afe21cc RK |
27 | #include "rtl.h" |
28 | #include "regs.h" | |
29 | #include "hard-reg-set.h" | |
30 | #include "flags.h" | |
31 | #include "real.h" | |
32 | #include "insn-config.h" | |
33 | #include "recog.h" | |
956d6950 | 34 | #include "expr.h" |
50b2596f KG |
35 | #include "toplev.h" |
36 | #include "output.h" | |
30f72379 | 37 | #include "splay-tree.h" |
7afe21cc RK |
38 | |
39 | /* The basic idea of common subexpression elimination is to go | |
40 | through the code, keeping a record of expressions that would | |
41 | have the same value at the current scan point, and replacing | |
42 | expressions encountered with the cheapest equivalent expression. | |
43 | ||
44 | It is too complicated to keep track of the different possibilities | |
45 | when control paths merge; so, at each label, we forget all that is | |
46 | known and start fresh. This can be described as processing each | |
47 | basic block separately. Note, however, that these are not quite | |
48 | the same as the basic blocks found by a later pass and used for | |
49 | data flow analysis and register packing. We do not need to start fresh | |
50 | after a conditional jump instruction if there is no label there. | |
51 | ||
52 | We use two data structures to record the equivalent expressions: | |
53 | a hash table for most expressions, and several vectors together | |
54 | with "quantity numbers" to record equivalent (pseudo) registers. | |
55 | ||
56 | The use of the special data structure for registers is desirable | |
57 | because it is faster. It is possible because registers references | |
58 | contain a fairly small number, the register number, taken from | |
59 | a contiguously allocated series, and two register references are | |
60 | identical if they have the same number. General expressions | |
61 | do not have any such thing, so the only way to retrieve the | |
62 | information recorded on an expression other than a register | |
63 | is to keep it in a hash table. | |
64 | ||
65 | Registers and "quantity numbers": | |
66 | ||
67 | At the start of each basic block, all of the (hardware and pseudo) | |
68 | registers used in the function are given distinct quantity | |
69 | numbers to indicate their contents. During scan, when the code | |
70 | copies one register into another, we copy the quantity number. | |
71 | When a register is loaded in any other way, we allocate a new | |
72 | quantity number to describe the value generated by this operation. | |
73 | `reg_qty' records what quantity a register is currently thought | |
74 | of as containing. | |
75 | ||
76 | All real quantity numbers are greater than or equal to `max_reg'. | |
77 | If register N has not been assigned a quantity, reg_qty[N] will equal N. | |
78 | ||
79 | Quantity numbers below `max_reg' do not exist and none of the `qty_...' | |
80 | variables should be referenced with an index below `max_reg'. | |
81 | ||
82 | We also maintain a bidirectional chain of registers for each | |
83 | quantity number. `qty_first_reg', `qty_last_reg', | |
84 | `reg_next_eqv' and `reg_prev_eqv' hold these chains. | |
85 | ||
86 | The first register in a chain is the one whose lifespan is least local. | |
87 | Among equals, it is the one that was seen first. | |
88 | We replace any equivalent register with that one. | |
89 | ||
90 | If two registers have the same quantity number, it must be true that | |
91 | REG expressions with `qty_mode' must be in the hash table for both | |
92 | registers and must be in the same class. | |
93 | ||
94 | The converse is not true. Since hard registers may be referenced in | |
95 | any mode, two REG expressions might be equivalent in the hash table | |
96 | but not have the same quantity number if the quantity number of one | |
97 | of the registers is not the same mode as those expressions. | |
98 | ||
99 | Constants and quantity numbers | |
100 | ||
101 | When a quantity has a known constant value, that value is stored | |
102 | in the appropriate element of qty_const. This is in addition to | |
103 | putting the constant in the hash table as is usual for non-regs. | |
104 | ||
d45cf215 | 105 | Whether a reg or a constant is preferred is determined by the configuration |
7afe21cc RK |
106 | macro CONST_COSTS and will often depend on the constant value. In any |
107 | event, expressions containing constants can be simplified, by fold_rtx. | |
108 | ||
109 | When a quantity has a known nearly constant value (such as an address | |
110 | of a stack slot), that value is stored in the appropriate element | |
111 | of qty_const. | |
112 | ||
113 | Integer constants don't have a machine mode. However, cse | |
114 | determines the intended machine mode from the destination | |
115 | of the instruction that moves the constant. The machine mode | |
116 | is recorded in the hash table along with the actual RTL | |
117 | constant expression so that different modes are kept separate. | |
118 | ||
119 | Other expressions: | |
120 | ||
121 | To record known equivalences among expressions in general | |
122 | we use a hash table called `table'. It has a fixed number of buckets | |
123 | that contain chains of `struct table_elt' elements for expressions. | |
124 | These chains connect the elements whose expressions have the same | |
125 | hash codes. | |
126 | ||
127 | Other chains through the same elements connect the elements which | |
128 | currently have equivalent values. | |
129 | ||
130 | Register references in an expression are canonicalized before hashing | |
131 | the expression. This is done using `reg_qty' and `qty_first_reg'. | |
132 | The hash code of a register reference is computed using the quantity | |
133 | number, not the register number. | |
134 | ||
135 | When the value of an expression changes, it is necessary to remove from the | |
136 | hash table not just that expression but all expressions whose values | |
137 | could be different as a result. | |
138 | ||
139 | 1. If the value changing is in memory, except in special cases | |
140 | ANYTHING referring to memory could be changed. That is because | |
141 | nobody knows where a pointer does not point. | |
142 | The function `invalidate_memory' removes what is necessary. | |
143 | ||
144 | The special cases are when the address is constant or is | |
145 | a constant plus a fixed register such as the frame pointer | |
146 | or a static chain pointer. When such addresses are stored in, | |
147 | we can tell exactly which other such addresses must be invalidated | |
148 | due to overlap. `invalidate' does this. | |
149 | All expressions that refer to non-constant | |
150 | memory addresses are also invalidated. `invalidate_memory' does this. | |
151 | ||
152 | 2. If the value changing is a register, all expressions | |
153 | containing references to that register, and only those, | |
154 | must be removed. | |
155 | ||
156 | Because searching the entire hash table for expressions that contain | |
157 | a register is very slow, we try to figure out when it isn't necessary. | |
158 | Precisely, this is necessary only when expressions have been | |
159 | entered in the hash table using this register, and then the value has | |
160 | changed, and then another expression wants to be added to refer to | |
161 | the register's new value. This sequence of circumstances is rare | |
162 | within any one basic block. | |
163 | ||
164 | The vectors `reg_tick' and `reg_in_table' are used to detect this case. | |
165 | reg_tick[i] is incremented whenever a value is stored in register i. | |
166 | reg_in_table[i] holds -1 if no references to register i have been | |
167 | entered in the table; otherwise, it contains the value reg_tick[i] had | |
168 | when the references were entered. If we want to enter a reference | |
169 | and reg_in_table[i] != reg_tick[i], we must scan and remove old references. | |
170 | Until we want to enter a new entry, the mere fact that the two vectors | |
171 | don't match makes the entries be ignored if anyone tries to match them. | |
172 | ||
173 | Registers themselves are entered in the hash table as well as in | |
174 | the equivalent-register chains. However, the vectors `reg_tick' | |
175 | and `reg_in_table' do not apply to expressions which are simple | |
176 | register references. These expressions are removed from the table | |
177 | immediately when they become invalid, and this can be done even if | |
178 | we do not immediately search for all the expressions that refer to | |
179 | the register. | |
180 | ||
181 | A CLOBBER rtx in an instruction invalidates its operand for further | |
182 | reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK | |
183 | invalidates everything that resides in memory. | |
184 | ||
185 | Related expressions: | |
186 | ||
187 | Constant expressions that differ only by an additive integer | |
188 | are called related. When a constant expression is put in | |
189 | the table, the related expression with no constant term | |
190 | is also entered. These are made to point at each other | |
191 | so that it is possible to find out if there exists any | |
192 | register equivalent to an expression related to a given expression. */ | |
193 | ||
194 | /* One plus largest register number used in this function. */ | |
195 | ||
196 | static int max_reg; | |
197 | ||
556c714b JW |
198 | /* One plus largest instruction UID used in this function at time of |
199 | cse_main call. */ | |
200 | ||
201 | static int max_insn_uid; | |
202 | ||
7afe21cc RK |
203 | /* Length of vectors indexed by quantity number. |
204 | We know in advance we will not need a quantity number this big. */ | |
205 | ||
206 | static int max_qty; | |
207 | ||
208 | /* Next quantity number to be allocated. | |
209 | This is 1 + the largest number needed so far. */ | |
210 | ||
211 | static int next_qty; | |
212 | ||
71d306d1 | 213 | /* Indexed by quantity number, gives the first (or last) register |
7afe21cc RK |
214 | in the chain of registers that currently contain this quantity. */ |
215 | ||
216 | static int *qty_first_reg; | |
217 | static int *qty_last_reg; | |
218 | ||
219 | /* Index by quantity number, gives the mode of the quantity. */ | |
220 | ||
221 | static enum machine_mode *qty_mode; | |
222 | ||
223 | /* Indexed by quantity number, gives the rtx of the constant value of the | |
224 | quantity, or zero if it does not have a known value. | |
225 | A sum of the frame pointer (or arg pointer) plus a constant | |
226 | can also be entered here. */ | |
227 | ||
228 | static rtx *qty_const; | |
229 | ||
230 | /* Indexed by qty number, gives the insn that stored the constant value | |
231 | recorded in `qty_const'. */ | |
232 | ||
233 | static rtx *qty_const_insn; | |
234 | ||
235 | /* The next three variables are used to track when a comparison between a | |
236 | quantity and some constant or register has been passed. In that case, we | |
237 | know the results of the comparison in case we see it again. These variables | |
238 | record a comparison that is known to be true. */ | |
239 | ||
240 | /* Indexed by qty number, gives the rtx code of a comparison with a known | |
241 | result involving this quantity. If none, it is UNKNOWN. */ | |
242 | static enum rtx_code *qty_comparison_code; | |
243 | ||
244 | /* Indexed by qty number, gives the constant being compared against in a | |
245 | comparison of known result. If no such comparison, it is undefined. | |
246 | If the comparison is not with a constant, it is zero. */ | |
247 | ||
248 | static rtx *qty_comparison_const; | |
249 | ||
250 | /* Indexed by qty number, gives the quantity being compared against in a | |
251 | comparison of known result. If no such comparison, if it undefined. | |
252 | If the comparison is not with a register, it is -1. */ | |
253 | ||
254 | static int *qty_comparison_qty; | |
255 | ||
256 | #ifdef HAVE_cc0 | |
257 | /* For machines that have a CC0, we do not record its value in the hash | |
258 | table since its use is guaranteed to be the insn immediately following | |
259 | its definition and any other insn is presumed to invalidate it. | |
260 | ||
261 | Instead, we store below the value last assigned to CC0. If it should | |
262 | happen to be a constant, it is stored in preference to the actual | |
263 | assigned value. In case it is a constant, we store the mode in which | |
264 | the constant should be interpreted. */ | |
265 | ||
266 | static rtx prev_insn_cc0; | |
267 | static enum machine_mode prev_insn_cc0_mode; | |
268 | #endif | |
269 | ||
270 | /* Previous actual insn. 0 if at first insn of basic block. */ | |
271 | ||
272 | static rtx prev_insn; | |
273 | ||
274 | /* Insn being scanned. */ | |
275 | ||
276 | static rtx this_insn; | |
277 | ||
71d306d1 DE |
278 | /* Index by register number, gives the number of the next (or |
279 | previous) register in the chain of registers sharing the same | |
7afe21cc RK |
280 | value. |
281 | ||
282 | Or -1 if this register is at the end of the chain. | |
283 | ||
284 | If reg_qty[N] == N, reg_next_eqv[N] is undefined. */ | |
285 | ||
286 | static int *reg_next_eqv; | |
287 | static int *reg_prev_eqv; | |
288 | ||
30f72379 MM |
289 | struct cse_reg_info { |
290 | union { | |
291 | /* The number of times the register has been altered in the current | |
292 | basic block. */ | |
293 | int reg_tick; | |
294 | ||
295 | /* The next cse_reg_info structure in the free list. */ | |
296 | struct cse_reg_info* next; | |
297 | } variant; | |
298 | ||
299 | /* The REG_TICK value at which rtx's containing this register are | |
300 | valid in the hash table. If this does not equal the current | |
301 | reg_tick value, such expressions existing in the hash table are | |
302 | invalid. */ | |
303 | int reg_in_table; | |
304 | ||
305 | /* The quantity number of the register's current contents. */ | |
306 | int reg_qty; | |
307 | }; | |
7afe21cc | 308 | |
30f72379 MM |
309 | /* A free list of cse_reg_info entries. */ |
310 | static struct cse_reg_info *cse_reg_info_free_list; | |
7afe21cc | 311 | |
30f72379 MM |
312 | /* A mapping from registers to cse_reg_info data structures. */ |
313 | static splay_tree cse_reg_info_tree; | |
7afe21cc | 314 | |
30f72379 MM |
315 | /* The last lookup we did into the cse_reg_info_tree. This allows us |
316 | to cache repeated lookups. */ | |
317 | static int cached_regno; | |
318 | static struct cse_reg_info *cached_cse_reg_info; | |
7afe21cc RK |
319 | |
320 | /* A HARD_REG_SET containing all the hard registers for which there is | |
321 | currently a REG expression in the hash table. Note the difference | |
322 | from the above variables, which indicate if the REG is mentioned in some | |
323 | expression in the table. */ | |
324 | ||
325 | static HARD_REG_SET hard_regs_in_table; | |
326 | ||
327 | /* A HARD_REG_SET containing all the hard registers that are invalidated | |
328 | by a CALL_INSN. */ | |
329 | ||
330 | static HARD_REG_SET regs_invalidated_by_call; | |
331 | ||
7afe21cc RK |
332 | /* CUID of insn that starts the basic block currently being cse-processed. */ |
333 | ||
334 | static int cse_basic_block_start; | |
335 | ||
336 | /* CUID of insn that ends the basic block currently being cse-processed. */ | |
337 | ||
338 | static int cse_basic_block_end; | |
339 | ||
340 | /* Vector mapping INSN_UIDs to cuids. | |
d45cf215 | 341 | The cuids are like uids but increase monotonically always. |
7afe21cc RK |
342 | We use them to see whether a reg is used outside a given basic block. */ |
343 | ||
906c4e36 | 344 | static int *uid_cuid; |
7afe21cc | 345 | |
164c8956 RK |
346 | /* Highest UID in UID_CUID. */ |
347 | static int max_uid; | |
348 | ||
7afe21cc RK |
349 | /* Get the cuid of an insn. */ |
350 | ||
351 | #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)]) | |
352 | ||
353 | /* Nonzero if cse has altered conditional jump insns | |
354 | in such a way that jump optimization should be redone. */ | |
355 | ||
356 | static int cse_jumps_altered; | |
357 | ||
a5dfb4ee RK |
358 | /* Nonzero if we put a LABEL_REF into the hash table. Since we may have put |
359 | it into an INSN without a REG_LABEL, we have to rerun jump after CSE | |
360 | to put in the note. */ | |
361 | static int recorded_label_ref; | |
362 | ||
7afe21cc RK |
363 | /* canon_hash stores 1 in do_not_record |
364 | if it notices a reference to CC0, PC, or some other volatile | |
365 | subexpression. */ | |
366 | ||
367 | static int do_not_record; | |
368 | ||
7bac1be0 RK |
369 | #ifdef LOAD_EXTEND_OP |
370 | ||
371 | /* Scratch rtl used when looking for load-extended copy of a MEM. */ | |
372 | static rtx memory_extend_rtx; | |
373 | #endif | |
374 | ||
7afe21cc RK |
375 | /* canon_hash stores 1 in hash_arg_in_memory |
376 | if it notices a reference to memory within the expression being hashed. */ | |
377 | ||
378 | static int hash_arg_in_memory; | |
379 | ||
380 | /* canon_hash stores 1 in hash_arg_in_struct | |
381 | if it notices a reference to memory that's part of a structure. */ | |
382 | ||
383 | static int hash_arg_in_struct; | |
384 | ||
385 | /* The hash table contains buckets which are chains of `struct table_elt's, | |
386 | each recording one expression's information. | |
387 | That expression is in the `exp' field. | |
388 | ||
389 | Those elements with the same hash code are chained in both directions | |
390 | through the `next_same_hash' and `prev_same_hash' fields. | |
391 | ||
392 | Each set of expressions with equivalent values | |
393 | are on a two-way chain through the `next_same_value' | |
394 | and `prev_same_value' fields, and all point with | |
395 | the `first_same_value' field at the first element in | |
396 | that chain. The chain is in order of increasing cost. | |
397 | Each element's cost value is in its `cost' field. | |
398 | ||
399 | The `in_memory' field is nonzero for elements that | |
400 | involve any reference to memory. These elements are removed | |
401 | whenever a write is done to an unidentified location in memory. | |
402 | To be safe, we assume that a memory address is unidentified unless | |
403 | the address is either a symbol constant or a constant plus | |
404 | the frame pointer or argument pointer. | |
405 | ||
406 | The `in_struct' field is nonzero for elements that | |
407 | involve any reference to memory inside a structure or array. | |
408 | ||
409 | The `related_value' field is used to connect related expressions | |
410 | (that differ by adding an integer). | |
411 | The related expressions are chained in a circular fashion. | |
412 | `related_value' is zero for expressions for which this | |
413 | chain is not useful. | |
414 | ||
415 | The `cost' field stores the cost of this element's expression. | |
416 | ||
417 | The `is_const' flag is set if the element is a constant (including | |
418 | a fixed address). | |
419 | ||
420 | The `flag' field is used as a temporary during some search routines. | |
421 | ||
422 | The `mode' field is usually the same as GET_MODE (`exp'), but | |
423 | if `exp' is a CONST_INT and has no machine mode then the `mode' | |
424 | field is the mode it was being used as. Each constant is | |
425 | recorded separately for each mode it is used with. */ | |
426 | ||
427 | ||
428 | struct table_elt | |
429 | { | |
430 | rtx exp; | |
431 | struct table_elt *next_same_hash; | |
432 | struct table_elt *prev_same_hash; | |
433 | struct table_elt *next_same_value; | |
434 | struct table_elt *prev_same_value; | |
435 | struct table_elt *first_same_value; | |
436 | struct table_elt *related_value; | |
437 | int cost; | |
438 | enum machine_mode mode; | |
439 | char in_memory; | |
440 | char in_struct; | |
441 | char is_const; | |
442 | char flag; | |
443 | }; | |
444 | ||
7afe21cc RK |
445 | /* We don't want a lot of buckets, because we rarely have very many |
446 | things stored in the hash table, and a lot of buckets slows | |
447 | down a lot of loops that happen frequently. */ | |
448 | #define NBUCKETS 31 | |
449 | ||
450 | /* Compute hash code of X in mode M. Special-case case where X is a pseudo | |
451 | register (hard registers may require `do_not_record' to be set). */ | |
452 | ||
453 | #define HASH(X, M) \ | |
454 | (GET_CODE (X) == REG && REGNO (X) >= FIRST_PSEUDO_REGISTER \ | |
30f72379 | 455 | ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) % NBUCKETS \ |
7afe21cc RK |
456 | : canon_hash (X, M) % NBUCKETS) |
457 | ||
458 | /* Determine whether register number N is considered a fixed register for CSE. | |
459 | It is desirable to replace other regs with fixed regs, to reduce need for | |
460 | non-fixed hard regs. | |
461 | A reg wins if it is either the frame pointer or designated as fixed, | |
462 | but not if it is an overlapping register. */ | |
463 | #ifdef OVERLAPPING_REGNO_P | |
464 | #define FIXED_REGNO_P(N) \ | |
8bc169f2 | 465 | (((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \ |
6ab832bc | 466 | || fixed_regs[N] || global_regs[N]) \ |
7afe21cc RK |
467 | && ! OVERLAPPING_REGNO_P ((N))) |
468 | #else | |
469 | #define FIXED_REGNO_P(N) \ | |
8bc169f2 | 470 | ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \ |
6ab832bc | 471 | || fixed_regs[N] || global_regs[N]) |
7afe21cc RK |
472 | #endif |
473 | ||
474 | /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed | |
ac07e066 RK |
475 | hard registers and pointers into the frame are the cheapest with a cost |
476 | of 0. Next come pseudos with a cost of one and other hard registers with | |
477 | a cost of 2. Aside from these special cases, call `rtx_cost'. */ | |
478 | ||
6ab832bc | 479 | #define CHEAP_REGNO(N) \ |
8bc169f2 DE |
480 | ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \ |
481 | || (N) == STACK_POINTER_REGNUM || (N) == ARG_POINTER_REGNUM \ | |
482 | || ((N) >= FIRST_VIRTUAL_REGISTER && (N) <= LAST_VIRTUAL_REGISTER) \ | |
483 | || ((N) < FIRST_PSEUDO_REGISTER \ | |
e7bb59fa | 484 | && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS)) |
7afe21cc | 485 | |
6ab832bc RK |
486 | /* A register is cheap if it is a user variable assigned to the register |
487 | or if its register number always corresponds to a cheap register. */ | |
488 | ||
489 | #define CHEAP_REG(N) \ | |
490 | ((REG_USERVAR_P (N) && REGNO (N) < FIRST_PSEUDO_REGISTER) \ | |
491 | || CHEAP_REGNO (REGNO (N))) | |
492 | ||
38734e55 ILT |
493 | #define COST(X) \ |
494 | (GET_CODE (X) == REG \ | |
495 | ? (CHEAP_REG (X) ? 0 \ | |
496 | : REGNO (X) >= FIRST_PSEUDO_REGISTER ? 1 \ | |
497 | : 2) \ | |
954a5693 | 498 | : notreg_cost(X)) |
7afe21cc | 499 | |
30f72379 MM |
500 | /* Get the info associated with register N. */ |
501 | ||
502 | #define GET_CSE_REG_INFO(N) \ | |
503 | (((N) == cached_regno && cached_cse_reg_info) \ | |
504 | ? cached_cse_reg_info : get_cse_reg_info ((N))) | |
505 | ||
506 | /* Get the number of times this register has been updated in this | |
507 | basic block. */ | |
508 | ||
509 | #define REG_TICK(N) ((GET_CSE_REG_INFO (N))->variant.reg_tick) | |
510 | ||
511 | /* Get the point at which REG was recorded in the table. */ | |
512 | ||
513 | #define REG_IN_TABLE(N) ((GET_CSE_REG_INFO (N))->reg_in_table) | |
514 | ||
515 | /* Get the quantity number for REG. */ | |
516 | ||
517 | #define REG_QTY(N) ((GET_CSE_REG_INFO (N))->reg_qty) | |
518 | ||
7afe21cc RK |
519 | /* Determine if the quantity number for register X represents a valid index |
520 | into the `qty_...' variables. */ | |
521 | ||
30f72379 | 522 | #define REGNO_QTY_VALID_P(N) (REG_QTY (N) != (N)) |
7afe21cc | 523 | |
2f541799 MM |
524 | #ifdef ADDRESS_COST |
525 | /* The ADDRESS_COST macro does not deal with ADDRESSOF nodes. But, | |
526 | during CSE, such nodes are present. Using an ADDRESSOF node which | |
527 | refers to the address of a REG is a good thing because we can then | |
528 | turn (MEM (ADDRESSSOF (REG))) into just plain REG. */ | |
529 | #define CSE_ADDRESS_COST(RTX) \ | |
530 | ((GET_CODE (RTX) == ADDRESSOF && REG_P (XEXP ((RTX), 0))) \ | |
531 | ? -1 : ADDRESS_COST(RTX)) | |
532 | #endif | |
533 | ||
7afe21cc RK |
534 | static struct table_elt *table[NBUCKETS]; |
535 | ||
536 | /* Chain of `struct table_elt's made so far for this function | |
537 | but currently removed from the table. */ | |
538 | ||
539 | static struct table_elt *free_element_chain; | |
540 | ||
541 | /* Number of `struct table_elt' structures made so far for this function. */ | |
542 | ||
543 | static int n_elements_made; | |
544 | ||
545 | /* Maximum value `n_elements_made' has had so far in this compilation | |
546 | for functions previously processed. */ | |
547 | ||
548 | static int max_elements_made; | |
549 | ||
550 | /* Surviving equivalence class when two equivalence classes are merged | |
551 | by recording the effects of a jump in the last insn. Zero if the | |
552 | last insn was not a conditional jump. */ | |
553 | ||
554 | static struct table_elt *last_jump_equiv_class; | |
555 | ||
556 | /* Set to the cost of a constant pool reference if one was found for a | |
557 | symbolic constant. If this was found, it means we should try to | |
558 | convert constants into constant pool entries if they don't fit in | |
559 | the insn. */ | |
560 | ||
561 | static int constant_pool_entries_cost; | |
562 | ||
6cd4575e RK |
563 | /* Define maximum length of a branch path. */ |
564 | ||
565 | #define PATHLENGTH 10 | |
566 | ||
567 | /* This data describes a block that will be processed by cse_basic_block. */ | |
568 | ||
569 | struct cse_basic_block_data { | |
570 | /* Lowest CUID value of insns in block. */ | |
571 | int low_cuid; | |
572 | /* Highest CUID value of insns in block. */ | |
573 | int high_cuid; | |
574 | /* Total number of SETs in block. */ | |
575 | int nsets; | |
576 | /* Last insn in the block. */ | |
577 | rtx last; | |
578 | /* Size of current branch path, if any. */ | |
579 | int path_size; | |
580 | /* Current branch path, indicating which branches will be taken. */ | |
581 | struct branch_path { | |
0f41302f | 582 | /* The branch insn. */ |
6cd4575e RK |
583 | rtx branch; |
584 | /* Whether it should be taken or not. AROUND is the same as taken | |
585 | except that it is used when the destination label is not preceded | |
586 | by a BARRIER. */ | |
587 | enum taken {TAKEN, NOT_TAKEN, AROUND} status; | |
588 | } path[PATHLENGTH]; | |
589 | }; | |
590 | ||
7afe21cc RK |
591 | /* Nonzero if X has the form (PLUS frame-pointer integer). We check for |
592 | virtual regs here because the simplify_*_operation routines are called | |
593 | by integrate.c, which is called before virtual register instantiation. */ | |
594 | ||
595 | #define FIXED_BASE_PLUS_P(X) \ | |
8bc169f2 DE |
596 | ((X) == frame_pointer_rtx || (X) == hard_frame_pointer_rtx \ |
597 | || (X) == arg_pointer_rtx \ | |
7afe21cc RK |
598 | || (X) == virtual_stack_vars_rtx \ |
599 | || (X) == virtual_incoming_args_rtx \ | |
600 | || (GET_CODE (X) == PLUS && GET_CODE (XEXP (X, 1)) == CONST_INT \ | |
601 | && (XEXP (X, 0) == frame_pointer_rtx \ | |
8bc169f2 | 602 | || XEXP (X, 0) == hard_frame_pointer_rtx \ |
7afe21cc RK |
603 | || XEXP (X, 0) == arg_pointer_rtx \ |
604 | || XEXP (X, 0) == virtual_stack_vars_rtx \ | |
e9a25f70 JL |
605 | || XEXP (X, 0) == virtual_incoming_args_rtx)) \ |
606 | || GET_CODE (X) == ADDRESSOF) | |
7afe21cc | 607 | |
6f90e075 JW |
608 | /* Similar, but also allows reference to the stack pointer. |
609 | ||
610 | This used to include FIXED_BASE_PLUS_P, however, we can't assume that | |
611 | arg_pointer_rtx by itself is nonzero, because on at least one machine, | |
612 | the i960, the arg pointer is zero when it is unused. */ | |
7afe21cc RK |
613 | |
614 | #define NONZERO_BASE_PLUS_P(X) \ | |
8bc169f2 | 615 | ((X) == frame_pointer_rtx || (X) == hard_frame_pointer_rtx \ |
6f90e075 JW |
616 | || (X) == virtual_stack_vars_rtx \ |
617 | || (X) == virtual_incoming_args_rtx \ | |
618 | || (GET_CODE (X) == PLUS && GET_CODE (XEXP (X, 1)) == CONST_INT \ | |
619 | && (XEXP (X, 0) == frame_pointer_rtx \ | |
8bc169f2 | 620 | || XEXP (X, 0) == hard_frame_pointer_rtx \ |
6f90e075 JW |
621 | || XEXP (X, 0) == arg_pointer_rtx \ |
622 | || XEXP (X, 0) == virtual_stack_vars_rtx \ | |
623 | || XEXP (X, 0) == virtual_incoming_args_rtx)) \ | |
7afe21cc RK |
624 | || (X) == stack_pointer_rtx \ |
625 | || (X) == virtual_stack_dynamic_rtx \ | |
626 | || (X) == virtual_outgoing_args_rtx \ | |
627 | || (GET_CODE (X) == PLUS && GET_CODE (XEXP (X, 1)) == CONST_INT \ | |
628 | && (XEXP (X, 0) == stack_pointer_rtx \ | |
629 | || XEXP (X, 0) == virtual_stack_dynamic_rtx \ | |
e9a25f70 JL |
630 | || XEXP (X, 0) == virtual_outgoing_args_rtx)) \ |
631 | || GET_CODE (X) == ADDRESSOF) | |
7afe21cc | 632 | |
954a5693 | 633 | static int notreg_cost PROTO((rtx)); |
6cd4575e RK |
634 | static void new_basic_block PROTO((void)); |
635 | static void make_new_qty PROTO((int)); | |
636 | static void make_regs_eqv PROTO((int, int)); | |
637 | static void delete_reg_equiv PROTO((int)); | |
638 | static int mention_regs PROTO((rtx)); | |
639 | static int insert_regs PROTO((rtx, struct table_elt *, int)); | |
640 | static void free_element PROTO((struct table_elt *)); | |
2197a88a | 641 | static void remove_from_table PROTO((struct table_elt *, unsigned)); |
6cd4575e | 642 | static struct table_elt *get_element PROTO((void)); |
2197a88a RK |
643 | static struct table_elt *lookup PROTO((rtx, unsigned, enum machine_mode)), |
644 | *lookup_for_remove PROTO((rtx, unsigned, enum machine_mode)); | |
6cd4575e | 645 | static rtx lookup_as_function PROTO((rtx, enum rtx_code)); |
2197a88a | 646 | static struct table_elt *insert PROTO((rtx, struct table_elt *, unsigned, |
6cd4575e RK |
647 | enum machine_mode)); |
648 | static void merge_equiv_classes PROTO((struct table_elt *, | |
649 | struct table_elt *)); | |
68c1e173 | 650 | static void invalidate PROTO((rtx, enum machine_mode)); |
9ae8ffe7 | 651 | static int cse_rtx_varies_p PROTO((rtx)); |
6cd4575e | 652 | static void remove_invalid_refs PROTO((int)); |
34c73909 | 653 | static void remove_invalid_subreg_refs PROTO((int, int, enum machine_mode)); |
6cd4575e | 654 | static void rehash_using_reg PROTO((rtx)); |
9ae8ffe7 | 655 | static void invalidate_memory PROTO((void)); |
6cd4575e RK |
656 | static void invalidate_for_call PROTO((void)); |
657 | static rtx use_related_value PROTO((rtx, struct table_elt *)); | |
2197a88a RK |
658 | static unsigned canon_hash PROTO((rtx, enum machine_mode)); |
659 | static unsigned safe_hash PROTO((rtx, enum machine_mode)); | |
6cd4575e | 660 | static int exp_equiv_p PROTO((rtx, rtx, int, int)); |
f451db89 | 661 | static void set_nonvarying_address_components PROTO((rtx, int, rtx *, |
6500fb43 RK |
662 | HOST_WIDE_INT *, |
663 | HOST_WIDE_INT *)); | |
6cd4575e | 664 | static int refers_to_p PROTO((rtx, rtx)); |
6cd4575e RK |
665 | static rtx canon_reg PROTO((rtx, rtx)); |
666 | static void find_best_addr PROTO((rtx, rtx *)); | |
667 | static enum rtx_code find_comparison_args PROTO((enum rtx_code, rtx *, rtx *, | |
668 | enum machine_mode *, | |
669 | enum machine_mode *)); | |
96b0e481 RK |
670 | static rtx cse_gen_binary PROTO((enum rtx_code, enum machine_mode, |
671 | rtx, rtx)); | |
672 | static rtx simplify_plus_minus PROTO((enum rtx_code, enum machine_mode, | |
673 | rtx, rtx)); | |
6cd4575e RK |
674 | static rtx fold_rtx PROTO((rtx, rtx)); |
675 | static rtx equiv_constant PROTO((rtx)); | |
676 | static void record_jump_equiv PROTO((rtx, int)); | |
677 | static void record_jump_cond PROTO((enum rtx_code, enum machine_mode, | |
678 | rtx, rtx, int)); | |
7bd8b2a8 | 679 | static void cse_insn PROTO((rtx, rtx)); |
9ae8ffe7 JL |
680 | static int note_mem_written PROTO((rtx)); |
681 | static void invalidate_from_clobbers PROTO((rtx)); | |
6cd4575e RK |
682 | static rtx cse_process_notes PROTO((rtx, rtx)); |
683 | static void cse_around_loop PROTO((rtx)); | |
684 | static void invalidate_skipped_set PROTO((rtx, rtx)); | |
685 | static void invalidate_skipped_block PROTO((rtx)); | |
686 | static void cse_check_loop_start PROTO((rtx, rtx)); | |
687 | static void cse_set_around_loop PROTO((rtx, rtx, rtx)); | |
688 | static rtx cse_basic_block PROTO((rtx, rtx, struct branch_path *, int)); | |
79644f06 | 689 | static void count_reg_usage PROTO((rtx, int *, rtx, int)); |
a0153051 | 690 | extern void dump_class PROTO((struct table_elt*)); |
1a87eea2 | 691 | static void check_fold_consts PROTO((PTR)); |
30f72379 MM |
692 | static struct cse_reg_info* get_cse_reg_info PROTO((int)); |
693 | static void free_cse_reg_info PROTO((splay_tree_value)); | |
01e752d3 | 694 | static void flush_hash_table PROTO((void)); |
c407b802 RK |
695 | |
696 | extern int rtx_equal_function_value_matters; | |
7afe21cc | 697 | \f |
a4c6502a MM |
698 | /* Dump the expressions in the equivalence class indicated by CLASSP. |
699 | This function is used only for debugging. */ | |
a0153051 | 700 | void |
a4c6502a MM |
701 | dump_class (classp) |
702 | struct table_elt *classp; | |
703 | { | |
704 | struct table_elt *elt; | |
705 | ||
706 | fprintf (stderr, "Equivalence chain for "); | |
707 | print_rtl (stderr, classp->exp); | |
708 | fprintf (stderr, ": \n"); | |
709 | ||
710 | for (elt = classp->first_same_value; elt; elt = elt->next_same_value) | |
711 | { | |
712 | print_rtl (stderr, elt->exp); | |
713 | fprintf (stderr, "\n"); | |
714 | } | |
715 | } | |
716 | ||
7afe21cc RK |
717 | /* Return an estimate of the cost of computing rtx X. |
718 | One use is in cse, to decide which expression to keep in the hash table. | |
719 | Another is in rtl generation, to pick the cheapest way to multiply. | |
720 | Other uses like the latter are expected in the future. */ | |
721 | ||
954a5693 RK |
722 | /* Internal function, to compute cost when X is not a register; called |
723 | from COST macro to keep it simple. */ | |
724 | ||
725 | static int | |
726 | notreg_cost (x) | |
727 | rtx x; | |
728 | { | |
729 | return ((GET_CODE (x) == SUBREG | |
730 | && GET_CODE (SUBREG_REG (x)) == REG | |
731 | && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT | |
732 | && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT | |
733 | && (GET_MODE_SIZE (GET_MODE (x)) | |
734 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
735 | && subreg_lowpart_p (x) | |
736 | && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)), | |
737 | GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))))) | |
738 | ? (CHEAP_REG (SUBREG_REG (x)) ? 0 | |
739 | : (REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER ? 1 | |
740 | : 2)) | |
741 | : rtx_cost (x, SET) * 2); | |
742 | } | |
743 | ||
7afe21cc RK |
744 | /* Return the right cost to give to an operation |
745 | to make the cost of the corresponding register-to-register instruction | |
746 | N times that of a fast register-to-register instruction. */ | |
747 | ||
748 | #define COSTS_N_INSNS(N) ((N) * 4 - 2) | |
749 | ||
750 | int | |
e5f6a288 | 751 | rtx_cost (x, outer_code) |
7afe21cc | 752 | rtx x; |
79c9824e | 753 | enum rtx_code outer_code ATTRIBUTE_UNUSED; |
7afe21cc RK |
754 | { |
755 | register int i, j; | |
756 | register enum rtx_code code; | |
757 | register char *fmt; | |
758 | register int total; | |
759 | ||
760 | if (x == 0) | |
761 | return 0; | |
762 | ||
763 | /* Compute the default costs of certain things. | |
764 | Note that RTX_COSTS can override the defaults. */ | |
765 | ||
766 | code = GET_CODE (x); | |
767 | switch (code) | |
768 | { | |
769 | case MULT: | |
770 | /* Count multiplication by 2**n as a shift, | |
771 | because if we are considering it, we would output it as a shift. */ | |
772 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
773 | && exact_log2 (INTVAL (XEXP (x, 1))) >= 0) | |
774 | total = 2; | |
775 | else | |
776 | total = COSTS_N_INSNS (5); | |
777 | break; | |
778 | case DIV: | |
779 | case UDIV: | |
780 | case MOD: | |
781 | case UMOD: | |
782 | total = COSTS_N_INSNS (7); | |
783 | break; | |
784 | case USE: | |
785 | /* Used in loop.c and combine.c as a marker. */ | |
786 | total = 0; | |
787 | break; | |
538b78e7 RS |
788 | case ASM_OPERANDS: |
789 | /* We don't want these to be used in substitutions because | |
790 | we have no way of validating the resulting insn. So assign | |
791 | anything containing an ASM_OPERANDS a very high cost. */ | |
792 | total = 1000; | |
793 | break; | |
7afe21cc RK |
794 | default: |
795 | total = 2; | |
796 | } | |
797 | ||
798 | switch (code) | |
799 | { | |
800 | case REG: | |
6ab832bc | 801 | return ! CHEAP_REG (x); |
ac07e066 | 802 | |
7afe21cc | 803 | case SUBREG: |
fc3ffe83 RK |
804 | /* If we can't tie these modes, make this expensive. The larger |
805 | the mode, the more expensive it is. */ | |
806 | if (! MODES_TIEABLE_P (GET_MODE (x), GET_MODE (SUBREG_REG (x)))) | |
807 | return COSTS_N_INSNS (2 | |
808 | + GET_MODE_SIZE (GET_MODE (x)) / UNITS_PER_WORD); | |
7afe21cc RK |
809 | return 2; |
810 | #ifdef RTX_COSTS | |
e5f6a288 | 811 | RTX_COSTS (x, code, outer_code); |
7afe21cc | 812 | #endif |
47a0b68f | 813 | #ifdef CONST_COSTS |
e5f6a288 | 814 | CONST_COSTS (x, code, outer_code); |
47a0b68f | 815 | #endif |
8625fab5 KG |
816 | |
817 | default: | |
818 | #ifdef DEFAULT_RTX_COSTS | |
819 | DEFAULT_RTX_COSTS(x, code, outer_code); | |
820 | #endif | |
821 | break; | |
7afe21cc RK |
822 | } |
823 | ||
824 | /* Sum the costs of the sub-rtx's, plus cost of this operation, | |
825 | which is already in total. */ | |
826 | ||
827 | fmt = GET_RTX_FORMAT (code); | |
828 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
829 | if (fmt[i] == 'e') | |
e5f6a288 | 830 | total += rtx_cost (XEXP (x, i), code); |
7afe21cc RK |
831 | else if (fmt[i] == 'E') |
832 | for (j = 0; j < XVECLEN (x, i); j++) | |
e5f6a288 | 833 | total += rtx_cost (XVECEXP (x, i, j), code); |
7afe21cc RK |
834 | |
835 | return total; | |
836 | } | |
837 | \f | |
30f72379 MM |
838 | static struct cse_reg_info * |
839 | get_cse_reg_info (regno) | |
840 | int regno; | |
841 | { | |
842 | struct cse_reg_info *cri; | |
843 | splay_tree_node n; | |
844 | ||
845 | /* See if we already have this entry. */ | |
846 | n = splay_tree_lookup (cse_reg_info_tree, | |
847 | (splay_tree_key) regno); | |
848 | if (n) | |
849 | cri = (struct cse_reg_info *) (n->value); | |
850 | else | |
851 | { | |
852 | /* Get a new cse_reg_info structure. */ | |
853 | if (cse_reg_info_free_list) | |
854 | { | |
855 | cri = cse_reg_info_free_list; | |
856 | cse_reg_info_free_list = cri->variant.next; | |
857 | } | |
858 | else | |
859 | cri = (struct cse_reg_info *) xmalloc (sizeof (struct cse_reg_info)); | |
860 | ||
861 | /* Initialize it. */ | |
862 | cri->variant.reg_tick = 0; | |
863 | cri->reg_in_table = -1; | |
864 | cri->reg_qty = regno; | |
865 | ||
866 | splay_tree_insert (cse_reg_info_tree, | |
867 | (splay_tree_key) regno, | |
868 | (splay_tree_value) cri); | |
869 | } | |
870 | ||
871 | /* Cache this lookup; we tend to be looking up information about the | |
872 | same register several times in a row. */ | |
873 | cached_regno = regno; | |
874 | cached_cse_reg_info = cri; | |
875 | ||
876 | return cri; | |
877 | } | |
878 | ||
879 | static void | |
880 | free_cse_reg_info (v) | |
881 | splay_tree_value v; | |
882 | { | |
883 | struct cse_reg_info *cri = (struct cse_reg_info *) v; | |
884 | ||
885 | cri->variant.next = cse_reg_info_free_list; | |
886 | cse_reg_info_free_list = cri; | |
887 | } | |
888 | ||
7afe21cc RK |
889 | /* Clear the hash table and initialize each register with its own quantity, |
890 | for a new basic block. */ | |
891 | ||
892 | static void | |
893 | new_basic_block () | |
894 | { | |
895 | register int i; | |
896 | ||
897 | next_qty = max_reg; | |
898 | ||
30f72379 MM |
899 | if (cse_reg_info_tree) |
900 | { | |
901 | splay_tree_delete (cse_reg_info_tree); | |
902 | cached_cse_reg_info = 0; | |
903 | } | |
904 | ||
905 | cse_reg_info_tree = splay_tree_new (splay_tree_compare_ints, 0, | |
906 | free_cse_reg_info); | |
7afe21cc | 907 | |
7afe21cc RK |
908 | CLEAR_HARD_REG_SET (hard_regs_in_table); |
909 | ||
910 | /* The per-quantity values used to be initialized here, but it is | |
911 | much faster to initialize each as it is made in `make_new_qty'. */ | |
912 | ||
913 | for (i = 0; i < NBUCKETS; i++) | |
914 | { | |
915 | register struct table_elt *this, *next; | |
916 | for (this = table[i]; this; this = next) | |
917 | { | |
918 | next = this->next_same_hash; | |
919 | free_element (this); | |
920 | } | |
921 | } | |
922 | ||
4c9a05bc | 923 | bzero ((char *) table, sizeof table); |
7afe21cc RK |
924 | |
925 | prev_insn = 0; | |
926 | ||
927 | #ifdef HAVE_cc0 | |
928 | prev_insn_cc0 = 0; | |
929 | #endif | |
930 | } | |
931 | ||
932 | /* Say that register REG contains a quantity not in any register before | |
933 | and initialize that quantity. */ | |
934 | ||
935 | static void | |
936 | make_new_qty (reg) | |
937 | register int reg; | |
938 | { | |
939 | register int q; | |
940 | ||
941 | if (next_qty >= max_qty) | |
942 | abort (); | |
943 | ||
30f72379 | 944 | q = REG_QTY (reg) = next_qty++; |
7afe21cc RK |
945 | qty_first_reg[q] = reg; |
946 | qty_last_reg[q] = reg; | |
947 | qty_const[q] = qty_const_insn[q] = 0; | |
948 | qty_comparison_code[q] = UNKNOWN; | |
949 | ||
950 | reg_next_eqv[reg] = reg_prev_eqv[reg] = -1; | |
951 | } | |
952 | ||
953 | /* Make reg NEW equivalent to reg OLD. | |
954 | OLD is not changing; NEW is. */ | |
955 | ||
956 | static void | |
957 | make_regs_eqv (new, old) | |
958 | register int new, old; | |
959 | { | |
960 | register int lastr, firstr; | |
30f72379 | 961 | register int q = REG_QTY (old); |
7afe21cc RK |
962 | |
963 | /* Nothing should become eqv until it has a "non-invalid" qty number. */ | |
964 | if (! REGNO_QTY_VALID_P (old)) | |
965 | abort (); | |
966 | ||
30f72379 | 967 | REG_QTY (new) = q; |
7afe21cc RK |
968 | firstr = qty_first_reg[q]; |
969 | lastr = qty_last_reg[q]; | |
970 | ||
971 | /* Prefer fixed hard registers to anything. Prefer pseudo regs to other | |
972 | hard regs. Among pseudos, if NEW will live longer than any other reg | |
973 | of the same qty, and that is beyond the current basic block, | |
974 | make it the new canonical replacement for this qty. */ | |
975 | if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr)) | |
976 | /* Certain fixed registers might be of the class NO_REGS. This means | |
977 | that not only can they not be allocated by the compiler, but | |
830a38ee | 978 | they cannot be used in substitutions or canonicalizations |
7afe21cc RK |
979 | either. */ |
980 | && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS) | |
981 | && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new)) | |
982 | || (new >= FIRST_PSEUDO_REGISTER | |
983 | && (firstr < FIRST_PSEUDO_REGISTER | |
b1f21e0a MM |
984 | || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end |
985 | || (uid_cuid[REGNO_FIRST_UID (new)] | |
7afe21cc | 986 | < cse_basic_block_start)) |
b1f21e0a MM |
987 | && (uid_cuid[REGNO_LAST_UID (new)] |
988 | > uid_cuid[REGNO_LAST_UID (firstr)])))))) | |
7afe21cc RK |
989 | { |
990 | reg_prev_eqv[firstr] = new; | |
991 | reg_next_eqv[new] = firstr; | |
992 | reg_prev_eqv[new] = -1; | |
993 | qty_first_reg[q] = new; | |
994 | } | |
995 | else | |
996 | { | |
997 | /* If NEW is a hard reg (known to be non-fixed), insert at end. | |
998 | Otherwise, insert before any non-fixed hard regs that are at the | |
999 | end. Registers of class NO_REGS cannot be used as an | |
1000 | equivalent for anything. */ | |
1001 | while (lastr < FIRST_PSEUDO_REGISTER && reg_prev_eqv[lastr] >= 0 | |
1002 | && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr)) | |
1003 | && new >= FIRST_PSEUDO_REGISTER) | |
1004 | lastr = reg_prev_eqv[lastr]; | |
1005 | reg_next_eqv[new] = reg_next_eqv[lastr]; | |
1006 | if (reg_next_eqv[lastr] >= 0) | |
1007 | reg_prev_eqv[reg_next_eqv[lastr]] = new; | |
1008 | else | |
1009 | qty_last_reg[q] = new; | |
1010 | reg_next_eqv[lastr] = new; | |
1011 | reg_prev_eqv[new] = lastr; | |
1012 | } | |
1013 | } | |
1014 | ||
1015 | /* Remove REG from its equivalence class. */ | |
1016 | ||
1017 | static void | |
1018 | delete_reg_equiv (reg) | |
1019 | register int reg; | |
1020 | { | |
30f72379 | 1021 | register int q = REG_QTY (reg); |
a4e262bc | 1022 | register int p, n; |
7afe21cc | 1023 | |
a4e262bc | 1024 | /* If invalid, do nothing. */ |
7afe21cc RK |
1025 | if (q == reg) |
1026 | return; | |
1027 | ||
a4e262bc RK |
1028 | p = reg_prev_eqv[reg]; |
1029 | n = reg_next_eqv[reg]; | |
1030 | ||
7afe21cc RK |
1031 | if (n != -1) |
1032 | reg_prev_eqv[n] = p; | |
1033 | else | |
1034 | qty_last_reg[q] = p; | |
1035 | if (p != -1) | |
1036 | reg_next_eqv[p] = n; | |
1037 | else | |
1038 | qty_first_reg[q] = n; | |
1039 | ||
30f72379 | 1040 | REG_QTY (reg) = reg; |
7afe21cc RK |
1041 | } |
1042 | ||
1043 | /* Remove any invalid expressions from the hash table | |
1044 | that refer to any of the registers contained in expression X. | |
1045 | ||
1046 | Make sure that newly inserted references to those registers | |
1047 | as subexpressions will be considered valid. | |
1048 | ||
1049 | mention_regs is not called when a register itself | |
1050 | is being stored in the table. | |
1051 | ||
1052 | Return 1 if we have done something that may have changed the hash code | |
1053 | of X. */ | |
1054 | ||
1055 | static int | |
1056 | mention_regs (x) | |
1057 | rtx x; | |
1058 | { | |
1059 | register enum rtx_code code; | |
1060 | register int i, j; | |
1061 | register char *fmt; | |
1062 | register int changed = 0; | |
1063 | ||
1064 | if (x == 0) | |
e5f6a288 | 1065 | return 0; |
7afe21cc RK |
1066 | |
1067 | code = GET_CODE (x); | |
1068 | if (code == REG) | |
1069 | { | |
1070 | register int regno = REGNO (x); | |
1071 | register int endregno | |
1072 | = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1 | |
1073 | : HARD_REGNO_NREGS (regno, GET_MODE (x))); | |
1074 | int i; | |
1075 | ||
1076 | for (i = regno; i < endregno; i++) | |
1077 | { | |
30f72379 | 1078 | if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i)) |
7afe21cc RK |
1079 | remove_invalid_refs (i); |
1080 | ||
30f72379 | 1081 | REG_IN_TABLE (i) = REG_TICK (i); |
7afe21cc RK |
1082 | } |
1083 | ||
1084 | return 0; | |
1085 | } | |
1086 | ||
34c73909 R |
1087 | /* If this is a SUBREG, we don't want to discard other SUBREGs of the same |
1088 | pseudo if they don't use overlapping words. We handle only pseudos | |
1089 | here for simplicity. */ | |
1090 | if (code == SUBREG && GET_CODE (SUBREG_REG (x)) == REG | |
1091 | && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER) | |
1092 | { | |
1093 | int i = REGNO (SUBREG_REG (x)); | |
1094 | ||
30f72379 | 1095 | if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i)) |
34c73909 R |
1096 | { |
1097 | /* If reg_tick has been incremented more than once since | |
1098 | reg_in_table was last set, that means that the entire | |
1099 | register has been set before, so discard anything memorized | |
1100 | for the entrire register, including all SUBREG expressions. */ | |
30f72379 | 1101 | if (REG_IN_TABLE (i) != REG_TICK (i) - 1) |
34c73909 R |
1102 | remove_invalid_refs (i); |
1103 | else | |
1104 | remove_invalid_subreg_refs (i, SUBREG_WORD (x), GET_MODE (x)); | |
1105 | } | |
1106 | ||
30f72379 | 1107 | REG_IN_TABLE (i) = REG_TICK (i); |
34c73909 R |
1108 | return 0; |
1109 | } | |
1110 | ||
7afe21cc RK |
1111 | /* If X is a comparison or a COMPARE and either operand is a register |
1112 | that does not have a quantity, give it one. This is so that a later | |
1113 | call to record_jump_equiv won't cause X to be assigned a different | |
1114 | hash code and not found in the table after that call. | |
1115 | ||
1116 | It is not necessary to do this here, since rehash_using_reg can | |
1117 | fix up the table later, but doing this here eliminates the need to | |
1118 | call that expensive function in the most common case where the only | |
1119 | use of the register is in the comparison. */ | |
1120 | ||
1121 | if (code == COMPARE || GET_RTX_CLASS (code) == '<') | |
1122 | { | |
1123 | if (GET_CODE (XEXP (x, 0)) == REG | |
1124 | && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))) | |
906c4e36 | 1125 | if (insert_regs (XEXP (x, 0), NULL_PTR, 0)) |
7afe21cc RK |
1126 | { |
1127 | rehash_using_reg (XEXP (x, 0)); | |
1128 | changed = 1; | |
1129 | } | |
1130 | ||
1131 | if (GET_CODE (XEXP (x, 1)) == REG | |
1132 | && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1)))) | |
906c4e36 | 1133 | if (insert_regs (XEXP (x, 1), NULL_PTR, 0)) |
7afe21cc RK |
1134 | { |
1135 | rehash_using_reg (XEXP (x, 1)); | |
1136 | changed = 1; | |
1137 | } | |
1138 | } | |
1139 | ||
1140 | fmt = GET_RTX_FORMAT (code); | |
1141 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1142 | if (fmt[i] == 'e') | |
1143 | changed |= mention_regs (XEXP (x, i)); | |
1144 | else if (fmt[i] == 'E') | |
1145 | for (j = 0; j < XVECLEN (x, i); j++) | |
1146 | changed |= mention_regs (XVECEXP (x, i, j)); | |
1147 | ||
1148 | return changed; | |
1149 | } | |
1150 | ||
1151 | /* Update the register quantities for inserting X into the hash table | |
1152 | with a value equivalent to CLASSP. | |
1153 | (If the class does not contain a REG, it is irrelevant.) | |
1154 | If MODIFIED is nonzero, X is a destination; it is being modified. | |
1155 | Note that delete_reg_equiv should be called on a register | |
1156 | before insert_regs is done on that register with MODIFIED != 0. | |
1157 | ||
1158 | Nonzero value means that elements of reg_qty have changed | |
1159 | so X's hash code may be different. */ | |
1160 | ||
1161 | static int | |
1162 | insert_regs (x, classp, modified) | |
1163 | rtx x; | |
1164 | struct table_elt *classp; | |
1165 | int modified; | |
1166 | { | |
1167 | if (GET_CODE (x) == REG) | |
1168 | { | |
1169 | register int regno = REGNO (x); | |
1170 | ||
1ff0c00d RK |
1171 | /* If REGNO is in the equivalence table already but is of the |
1172 | wrong mode for that equivalence, don't do anything here. */ | |
1173 | ||
1174 | if (REGNO_QTY_VALID_P (regno) | |
30f72379 | 1175 | && qty_mode[REG_QTY (regno)] != GET_MODE (x)) |
1ff0c00d RK |
1176 | return 0; |
1177 | ||
1178 | if (modified || ! REGNO_QTY_VALID_P (regno)) | |
7afe21cc RK |
1179 | { |
1180 | if (classp) | |
1181 | for (classp = classp->first_same_value; | |
1182 | classp != 0; | |
1183 | classp = classp->next_same_value) | |
1184 | if (GET_CODE (classp->exp) == REG | |
1185 | && GET_MODE (classp->exp) == GET_MODE (x)) | |
1186 | { | |
1187 | make_regs_eqv (regno, REGNO (classp->exp)); | |
1188 | return 1; | |
1189 | } | |
1190 | ||
1191 | make_new_qty (regno); | |
30f72379 | 1192 | qty_mode[REG_QTY (regno)] = GET_MODE (x); |
7afe21cc RK |
1193 | return 1; |
1194 | } | |
cdf4112f TG |
1195 | |
1196 | return 0; | |
7afe21cc | 1197 | } |
c610adec RK |
1198 | |
1199 | /* If X is a SUBREG, we will likely be inserting the inner register in the | |
1200 | table. If that register doesn't have an assigned quantity number at | |
1201 | this point but does later, the insertion that we will be doing now will | |
1202 | not be accessible because its hash code will have changed. So assign | |
1203 | a quantity number now. */ | |
1204 | ||
1205 | else if (GET_CODE (x) == SUBREG && GET_CODE (SUBREG_REG (x)) == REG | |
1206 | && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x)))) | |
1207 | { | |
34c73909 R |
1208 | int regno = REGNO (SUBREG_REG (x)); |
1209 | ||
906c4e36 | 1210 | insert_regs (SUBREG_REG (x), NULL_PTR, 0); |
34c73909 R |
1211 | /* Mention_regs checks if REG_TICK is exactly one larger than |
1212 | REG_IN_TABLE to find out if there was only a single preceding | |
1213 | invalidation - for the SUBREG - or another one, which would be | |
1214 | for the full register. Since we don't invalidate the SUBREG | |
1215 | here first, we might have to bump up REG_TICK so that mention_regs | |
1216 | will do the right thing. */ | |
30f72379 MM |
1217 | if (REG_IN_TABLE (regno) >= 0 |
1218 | && REG_TICK (regno) == REG_IN_TABLE (regno) + 1) | |
1219 | REG_TICK (regno)++; | |
34c73909 | 1220 | mention_regs (x); |
c610adec RK |
1221 | return 1; |
1222 | } | |
7afe21cc RK |
1223 | else |
1224 | return mention_regs (x); | |
1225 | } | |
1226 | \f | |
1227 | /* Look in or update the hash table. */ | |
1228 | ||
1229 | /* Put the element ELT on the list of free elements. */ | |
1230 | ||
1231 | static void | |
1232 | free_element (elt) | |
1233 | struct table_elt *elt; | |
1234 | { | |
1235 | elt->next_same_hash = free_element_chain; | |
1236 | free_element_chain = elt; | |
1237 | } | |
1238 | ||
1239 | /* Return an element that is free for use. */ | |
1240 | ||
1241 | static struct table_elt * | |
1242 | get_element () | |
1243 | { | |
1244 | struct table_elt *elt = free_element_chain; | |
1245 | if (elt) | |
1246 | { | |
1247 | free_element_chain = elt->next_same_hash; | |
1248 | return elt; | |
1249 | } | |
1250 | n_elements_made++; | |
1251 | return (struct table_elt *) oballoc (sizeof (struct table_elt)); | |
1252 | } | |
1253 | ||
1254 | /* Remove table element ELT from use in the table. | |
1255 | HASH is its hash code, made using the HASH macro. | |
1256 | It's an argument because often that is known in advance | |
1257 | and we save much time not recomputing it. */ | |
1258 | ||
1259 | static void | |
1260 | remove_from_table (elt, hash) | |
1261 | register struct table_elt *elt; | |
2197a88a | 1262 | unsigned hash; |
7afe21cc RK |
1263 | { |
1264 | if (elt == 0) | |
1265 | return; | |
1266 | ||
1267 | /* Mark this element as removed. See cse_insn. */ | |
1268 | elt->first_same_value = 0; | |
1269 | ||
1270 | /* Remove the table element from its equivalence class. */ | |
1271 | ||
1272 | { | |
1273 | register struct table_elt *prev = elt->prev_same_value; | |
1274 | register struct table_elt *next = elt->next_same_value; | |
1275 | ||
1276 | if (next) next->prev_same_value = prev; | |
1277 | ||
1278 | if (prev) | |
1279 | prev->next_same_value = next; | |
1280 | else | |
1281 | { | |
1282 | register struct table_elt *newfirst = next; | |
1283 | while (next) | |
1284 | { | |
1285 | next->first_same_value = newfirst; | |
1286 | next = next->next_same_value; | |
1287 | } | |
1288 | } | |
1289 | } | |
1290 | ||
1291 | /* Remove the table element from its hash bucket. */ | |
1292 | ||
1293 | { | |
1294 | register struct table_elt *prev = elt->prev_same_hash; | |
1295 | register struct table_elt *next = elt->next_same_hash; | |
1296 | ||
1297 | if (next) next->prev_same_hash = prev; | |
1298 | ||
1299 | if (prev) | |
1300 | prev->next_same_hash = next; | |
1301 | else if (table[hash] == elt) | |
1302 | table[hash] = next; | |
1303 | else | |
1304 | { | |
1305 | /* This entry is not in the proper hash bucket. This can happen | |
1306 | when two classes were merged by `merge_equiv_classes'. Search | |
1307 | for the hash bucket that it heads. This happens only very | |
1308 | rarely, so the cost is acceptable. */ | |
1309 | for (hash = 0; hash < NBUCKETS; hash++) | |
1310 | if (table[hash] == elt) | |
1311 | table[hash] = next; | |
1312 | } | |
1313 | } | |
1314 | ||
1315 | /* Remove the table element from its related-value circular chain. */ | |
1316 | ||
1317 | if (elt->related_value != 0 && elt->related_value != elt) | |
1318 | { | |
1319 | register struct table_elt *p = elt->related_value; | |
1320 | while (p->related_value != elt) | |
1321 | p = p->related_value; | |
1322 | p->related_value = elt->related_value; | |
1323 | if (p->related_value == p) | |
1324 | p->related_value = 0; | |
1325 | } | |
1326 | ||
1327 | free_element (elt); | |
1328 | } | |
1329 | ||
1330 | /* Look up X in the hash table and return its table element, | |
1331 | or 0 if X is not in the table. | |
1332 | ||
1333 | MODE is the machine-mode of X, or if X is an integer constant | |
1334 | with VOIDmode then MODE is the mode with which X will be used. | |
1335 | ||
1336 | Here we are satisfied to find an expression whose tree structure | |
1337 | looks like X. */ | |
1338 | ||
1339 | static struct table_elt * | |
1340 | lookup (x, hash, mode) | |
1341 | rtx x; | |
2197a88a | 1342 | unsigned hash; |
7afe21cc RK |
1343 | enum machine_mode mode; |
1344 | { | |
1345 | register struct table_elt *p; | |
1346 | ||
1347 | for (p = table[hash]; p; p = p->next_same_hash) | |
1348 | if (mode == p->mode && ((x == p->exp && GET_CODE (x) == REG) | |
1349 | || exp_equiv_p (x, p->exp, GET_CODE (x) != REG, 0))) | |
1350 | return p; | |
1351 | ||
1352 | return 0; | |
1353 | } | |
1354 | ||
1355 | /* Like `lookup' but don't care whether the table element uses invalid regs. | |
1356 | Also ignore discrepancies in the machine mode of a register. */ | |
1357 | ||
1358 | static struct table_elt * | |
1359 | lookup_for_remove (x, hash, mode) | |
1360 | rtx x; | |
2197a88a | 1361 | unsigned hash; |
7afe21cc RK |
1362 | enum machine_mode mode; |
1363 | { | |
1364 | register struct table_elt *p; | |
1365 | ||
1366 | if (GET_CODE (x) == REG) | |
1367 | { | |
1368 | int regno = REGNO (x); | |
1369 | /* Don't check the machine mode when comparing registers; | |
1370 | invalidating (REG:SI 0) also invalidates (REG:DF 0). */ | |
1371 | for (p = table[hash]; p; p = p->next_same_hash) | |
1372 | if (GET_CODE (p->exp) == REG | |
1373 | && REGNO (p->exp) == regno) | |
1374 | return p; | |
1375 | } | |
1376 | else | |
1377 | { | |
1378 | for (p = table[hash]; p; p = p->next_same_hash) | |
1379 | if (mode == p->mode && (x == p->exp || exp_equiv_p (x, p->exp, 0, 0))) | |
1380 | return p; | |
1381 | } | |
1382 | ||
1383 | return 0; | |
1384 | } | |
1385 | ||
1386 | /* Look for an expression equivalent to X and with code CODE. | |
1387 | If one is found, return that expression. */ | |
1388 | ||
1389 | static rtx | |
1390 | lookup_as_function (x, code) | |
1391 | rtx x; | |
1392 | enum rtx_code code; | |
1393 | { | |
1394 | register struct table_elt *p = lookup (x, safe_hash (x, VOIDmode) % NBUCKETS, | |
1395 | GET_MODE (x)); | |
34c73909 R |
1396 | /* If we are looking for a CONST_INT, the mode doesn't really matter, as |
1397 | long as we are narrowing. So if we looked in vain for a mode narrower | |
1398 | than word_mode before, look for word_mode now. */ | |
1399 | if (p == 0 && code == CONST_INT | |
1400 | && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode)) | |
1401 | { | |
1402 | x = copy_rtx (x); | |
1403 | PUT_MODE (x, word_mode); | |
1404 | p = lookup (x, safe_hash (x, VOIDmode) % NBUCKETS, word_mode); | |
1405 | } | |
1406 | ||
7afe21cc RK |
1407 | if (p == 0) |
1408 | return 0; | |
1409 | ||
1410 | for (p = p->first_same_value; p; p = p->next_same_value) | |
1411 | { | |
1412 | if (GET_CODE (p->exp) == code | |
1413 | /* Make sure this is a valid entry in the table. */ | |
1414 | && exp_equiv_p (p->exp, p->exp, 1, 0)) | |
1415 | return p->exp; | |
1416 | } | |
1417 | ||
1418 | return 0; | |
1419 | } | |
1420 | ||
1421 | /* Insert X in the hash table, assuming HASH is its hash code | |
1422 | and CLASSP is an element of the class it should go in | |
1423 | (or 0 if a new class should be made). | |
1424 | It is inserted at the proper position to keep the class in | |
1425 | the order cheapest first. | |
1426 | ||
1427 | MODE is the machine-mode of X, or if X is an integer constant | |
1428 | with VOIDmode then MODE is the mode with which X will be used. | |
1429 | ||
1430 | For elements of equal cheapness, the most recent one | |
1431 | goes in front, except that the first element in the list | |
1432 | remains first unless a cheaper element is added. The order of | |
1433 | pseudo-registers does not matter, as canon_reg will be called to | |
830a38ee | 1434 | find the cheapest when a register is retrieved from the table. |
7afe21cc RK |
1435 | |
1436 | The in_memory field in the hash table element is set to 0. | |
1437 | The caller must set it nonzero if appropriate. | |
1438 | ||
1439 | You should call insert_regs (X, CLASSP, MODIFY) before calling here, | |
1440 | and if insert_regs returns a nonzero value | |
1441 | you must then recompute its hash code before calling here. | |
1442 | ||
1443 | If necessary, update table showing constant values of quantities. */ | |
1444 | ||
1445 | #define CHEAPER(X,Y) ((X)->cost < (Y)->cost) | |
1446 | ||
1447 | static struct table_elt * | |
1448 | insert (x, classp, hash, mode) | |
1449 | register rtx x; | |
1450 | register struct table_elt *classp; | |
2197a88a | 1451 | unsigned hash; |
7afe21cc RK |
1452 | enum machine_mode mode; |
1453 | { | |
1454 | register struct table_elt *elt; | |
1455 | ||
1456 | /* If X is a register and we haven't made a quantity for it, | |
1457 | something is wrong. */ | |
1458 | if (GET_CODE (x) == REG && ! REGNO_QTY_VALID_P (REGNO (x))) | |
1459 | abort (); | |
1460 | ||
1461 | /* If X is a hard register, show it is being put in the table. */ | |
1462 | if (GET_CODE (x) == REG && REGNO (x) < FIRST_PSEUDO_REGISTER) | |
1463 | { | |
1464 | int regno = REGNO (x); | |
1465 | int endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
1466 | int i; | |
1467 | ||
1468 | for (i = regno; i < endregno; i++) | |
1469 | SET_HARD_REG_BIT (hard_regs_in_table, i); | |
1470 | } | |
1471 | ||
a5dfb4ee | 1472 | /* If X is a label, show we recorded it. */ |
970c9ace RK |
1473 | if (GET_CODE (x) == LABEL_REF |
1474 | || (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS | |
1475 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF)) | |
a5dfb4ee | 1476 | recorded_label_ref = 1; |
7afe21cc RK |
1477 | |
1478 | /* Put an element for X into the right hash bucket. */ | |
1479 | ||
1480 | elt = get_element (); | |
1481 | elt->exp = x; | |
1482 | elt->cost = COST (x); | |
1483 | elt->next_same_value = 0; | |
1484 | elt->prev_same_value = 0; | |
1485 | elt->next_same_hash = table[hash]; | |
1486 | elt->prev_same_hash = 0; | |
1487 | elt->related_value = 0; | |
1488 | elt->in_memory = 0; | |
1489 | elt->mode = mode; | |
1490 | elt->is_const = (CONSTANT_P (x) | |
1491 | /* GNU C++ takes advantage of this for `this' | |
1492 | (and other const values). */ | |
1493 | || (RTX_UNCHANGING_P (x) | |
1494 | && GET_CODE (x) == REG | |
1495 | && REGNO (x) >= FIRST_PSEUDO_REGISTER) | |
1496 | || FIXED_BASE_PLUS_P (x)); | |
1497 | ||
1498 | if (table[hash]) | |
1499 | table[hash]->prev_same_hash = elt; | |
1500 | table[hash] = elt; | |
1501 | ||
1502 | /* Put it into the proper value-class. */ | |
1503 | if (classp) | |
1504 | { | |
1505 | classp = classp->first_same_value; | |
1506 | if (CHEAPER (elt, classp)) | |
1507 | /* Insert at the head of the class */ | |
1508 | { | |
1509 | register struct table_elt *p; | |
1510 | elt->next_same_value = classp; | |
1511 | classp->prev_same_value = elt; | |
1512 | elt->first_same_value = elt; | |
1513 | ||
1514 | for (p = classp; p; p = p->next_same_value) | |
1515 | p->first_same_value = elt; | |
1516 | } | |
1517 | else | |
1518 | { | |
1519 | /* Insert not at head of the class. */ | |
1520 | /* Put it after the last element cheaper than X. */ | |
1521 | register struct table_elt *p, *next; | |
1522 | for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt); | |
1523 | p = next); | |
1524 | /* Put it after P and before NEXT. */ | |
1525 | elt->next_same_value = next; | |
1526 | if (next) | |
1527 | next->prev_same_value = elt; | |
1528 | elt->prev_same_value = p; | |
1529 | p->next_same_value = elt; | |
1530 | elt->first_same_value = classp; | |
1531 | } | |
1532 | } | |
1533 | else | |
1534 | elt->first_same_value = elt; | |
1535 | ||
1536 | /* If this is a constant being set equivalent to a register or a register | |
1537 | being set equivalent to a constant, note the constant equivalence. | |
1538 | ||
1539 | If this is a constant, it cannot be equivalent to a different constant, | |
1540 | and a constant is the only thing that can be cheaper than a register. So | |
1541 | we know the register is the head of the class (before the constant was | |
1542 | inserted). | |
1543 | ||
1544 | If this is a register that is not already known equivalent to a | |
1545 | constant, we must check the entire class. | |
1546 | ||
1547 | If this is a register that is already known equivalent to an insn, | |
1548 | update `qty_const_insn' to show that `this_insn' is the latest | |
1549 | insn making that quantity equivalent to the constant. */ | |
1550 | ||
f353588a RK |
1551 | if (elt->is_const && classp && GET_CODE (classp->exp) == REG |
1552 | && GET_CODE (x) != REG) | |
7afe21cc | 1553 | { |
30f72379 MM |
1554 | qty_const[REG_QTY (REGNO (classp->exp))] |
1555 | = gen_lowpart_if_possible (qty_mode[REG_QTY (REGNO (classp->exp))], x); | |
1556 | qty_const_insn[REG_QTY (REGNO (classp->exp))] = this_insn; | |
7afe21cc RK |
1557 | } |
1558 | ||
30f72379 | 1559 | else if (GET_CODE (x) == REG && classp && ! qty_const[REG_QTY (REGNO (x))] |
f353588a | 1560 | && ! elt->is_const) |
7afe21cc RK |
1561 | { |
1562 | register struct table_elt *p; | |
1563 | ||
1564 | for (p = classp; p != 0; p = p->next_same_value) | |
1565 | { | |
f353588a | 1566 | if (p->is_const && GET_CODE (p->exp) != REG) |
7afe21cc | 1567 | { |
30f72379 | 1568 | qty_const[REG_QTY (REGNO (x))] |
7afe21cc | 1569 | = gen_lowpart_if_possible (GET_MODE (x), p->exp); |
30f72379 | 1570 | qty_const_insn[REG_QTY (REGNO (x))] = this_insn; |
7afe21cc RK |
1571 | break; |
1572 | } | |
1573 | } | |
1574 | } | |
1575 | ||
30f72379 MM |
1576 | else if (GET_CODE (x) == REG && qty_const[REG_QTY (REGNO (x))] |
1577 | && GET_MODE (x) == qty_mode[REG_QTY (REGNO (x))]) | |
1578 | qty_const_insn[REG_QTY (REGNO (x))] = this_insn; | |
7afe21cc RK |
1579 | |
1580 | /* If this is a constant with symbolic value, | |
1581 | and it has a term with an explicit integer value, | |
1582 | link it up with related expressions. */ | |
1583 | if (GET_CODE (x) == CONST) | |
1584 | { | |
1585 | rtx subexp = get_related_value (x); | |
2197a88a | 1586 | unsigned subhash; |
7afe21cc RK |
1587 | struct table_elt *subelt, *subelt_prev; |
1588 | ||
1589 | if (subexp != 0) | |
1590 | { | |
1591 | /* Get the integer-free subexpression in the hash table. */ | |
1592 | subhash = safe_hash (subexp, mode) % NBUCKETS; | |
1593 | subelt = lookup (subexp, subhash, mode); | |
1594 | if (subelt == 0) | |
906c4e36 | 1595 | subelt = insert (subexp, NULL_PTR, subhash, mode); |
7afe21cc RK |
1596 | /* Initialize SUBELT's circular chain if it has none. */ |
1597 | if (subelt->related_value == 0) | |
1598 | subelt->related_value = subelt; | |
1599 | /* Find the element in the circular chain that precedes SUBELT. */ | |
1600 | subelt_prev = subelt; | |
1601 | while (subelt_prev->related_value != subelt) | |
1602 | subelt_prev = subelt_prev->related_value; | |
1603 | /* Put new ELT into SUBELT's circular chain just before SUBELT. | |
1604 | This way the element that follows SUBELT is the oldest one. */ | |
1605 | elt->related_value = subelt_prev->related_value; | |
1606 | subelt_prev->related_value = elt; | |
1607 | } | |
1608 | } | |
1609 | ||
1610 | return elt; | |
1611 | } | |
1612 | \f | |
1613 | /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from | |
1614 | CLASS2 into CLASS1. This is done when we have reached an insn which makes | |
1615 | the two classes equivalent. | |
1616 | ||
1617 | CLASS1 will be the surviving class; CLASS2 should not be used after this | |
1618 | call. | |
1619 | ||
1620 | Any invalid entries in CLASS2 will not be copied. */ | |
1621 | ||
1622 | static void | |
1623 | merge_equiv_classes (class1, class2) | |
1624 | struct table_elt *class1, *class2; | |
1625 | { | |
1626 | struct table_elt *elt, *next, *new; | |
1627 | ||
1628 | /* Ensure we start with the head of the classes. */ | |
1629 | class1 = class1->first_same_value; | |
1630 | class2 = class2->first_same_value; | |
1631 | ||
1632 | /* If they were already equal, forget it. */ | |
1633 | if (class1 == class2) | |
1634 | return; | |
1635 | ||
1636 | for (elt = class2; elt; elt = next) | |
1637 | { | |
2197a88a | 1638 | unsigned hash; |
7afe21cc RK |
1639 | rtx exp = elt->exp; |
1640 | enum machine_mode mode = elt->mode; | |
1641 | ||
1642 | next = elt->next_same_value; | |
1643 | ||
1644 | /* Remove old entry, make a new one in CLASS1's class. | |
1645 | Don't do this for invalid entries as we cannot find their | |
0f41302f | 1646 | hash code (it also isn't necessary). */ |
7afe21cc RK |
1647 | if (GET_CODE (exp) == REG || exp_equiv_p (exp, exp, 1, 0)) |
1648 | { | |
1649 | hash_arg_in_memory = 0; | |
1650 | hash_arg_in_struct = 0; | |
1651 | hash = HASH (exp, mode); | |
1652 | ||
1653 | if (GET_CODE (exp) == REG) | |
1654 | delete_reg_equiv (REGNO (exp)); | |
1655 | ||
1656 | remove_from_table (elt, hash); | |
1657 | ||
1658 | if (insert_regs (exp, class1, 0)) | |
8ae2b8f6 JW |
1659 | { |
1660 | rehash_using_reg (exp); | |
1661 | hash = HASH (exp, mode); | |
1662 | } | |
7afe21cc RK |
1663 | new = insert (exp, class1, hash, mode); |
1664 | new->in_memory = hash_arg_in_memory; | |
1665 | new->in_struct = hash_arg_in_struct; | |
1666 | } | |
1667 | } | |
1668 | } | |
1669 | \f | |
01e752d3 JL |
1670 | |
1671 | /* Flush the entire hash table. */ | |
1672 | ||
1673 | static void | |
1674 | flush_hash_table () | |
1675 | { | |
1676 | int i; | |
1677 | struct table_elt *p; | |
1678 | ||
1679 | for (i = 0; i < NBUCKETS; i++) | |
1680 | for (p = table[i]; p; p = table[i]) | |
1681 | { | |
1682 | /* Note that invalidate can remove elements | |
1683 | after P in the current hash chain. */ | |
1684 | if (GET_CODE (p->exp) == REG) | |
1685 | invalidate (p->exp, p->mode); | |
1686 | else | |
1687 | remove_from_table (p, i); | |
1688 | } | |
1689 | } | |
1690 | ||
1691 | ||
7afe21cc RK |
1692 | /* Remove from the hash table, or mark as invalid, |
1693 | all expressions whose values could be altered by storing in X. | |
1694 | X is a register, a subreg, or a memory reference with nonvarying address | |
1695 | (because, when a memory reference with a varying address is stored in, | |
1696 | all memory references are removed by invalidate_memory | |
1697 | so specific invalidation is superfluous). | |
bb4034b3 JW |
1698 | FULL_MODE, if not VOIDmode, indicates that this much should be invalidated |
1699 | instead of just the amount indicated by the mode of X. This is only used | |
1700 | for bitfield stores into memory. | |
7afe21cc RK |
1701 | |
1702 | A nonvarying address may be just a register or just | |
1703 | a symbol reference, or it may be either of those plus | |
1704 | a numeric offset. */ | |
1705 | ||
1706 | static void | |
bb4034b3 | 1707 | invalidate (x, full_mode) |
7afe21cc | 1708 | rtx x; |
bb4034b3 | 1709 | enum machine_mode full_mode; |
7afe21cc RK |
1710 | { |
1711 | register int i; | |
1712 | register struct table_elt *p; | |
7afe21cc RK |
1713 | |
1714 | /* If X is a register, dependencies on its contents | |
1715 | are recorded through the qty number mechanism. | |
1716 | Just change the qty number of the register, | |
1717 | mark it as invalid for expressions that refer to it, | |
1718 | and remove it itself. */ | |
1719 | ||
1720 | if (GET_CODE (x) == REG) | |
1721 | { | |
1722 | register int regno = REGNO (x); | |
2197a88a | 1723 | register unsigned hash = HASH (x, GET_MODE (x)); |
7afe21cc RK |
1724 | |
1725 | /* Remove REGNO from any quantity list it might be on and indicate | |
9ec36da5 | 1726 | that its value might have changed. If it is a pseudo, remove its |
7afe21cc RK |
1727 | entry from the hash table. |
1728 | ||
1729 | For a hard register, we do the first two actions above for any | |
1730 | additional hard registers corresponding to X. Then, if any of these | |
1731 | registers are in the table, we must remove any REG entries that | |
1732 | overlap these registers. */ | |
1733 | ||
1734 | delete_reg_equiv (regno); | |
30f72379 | 1735 | REG_TICK (regno)++; |
7afe21cc RK |
1736 | |
1737 | if (regno >= FIRST_PSEUDO_REGISTER) | |
85e4d983 RK |
1738 | { |
1739 | /* Because a register can be referenced in more than one mode, | |
1740 | we might have to remove more than one table entry. */ | |
1741 | ||
1742 | struct table_elt *elt; | |
1743 | ||
2d8b0f3a | 1744 | while ((elt = lookup_for_remove (x, hash, GET_MODE (x)))) |
85e4d983 RK |
1745 | remove_from_table (elt, hash); |
1746 | } | |
7afe21cc RK |
1747 | else |
1748 | { | |
54b1de55 RK |
1749 | HOST_WIDE_INT in_table |
1750 | = TEST_HARD_REG_BIT (hard_regs_in_table, regno); | |
7afe21cc RK |
1751 | int endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); |
1752 | int tregno, tendregno; | |
1753 | register struct table_elt *p, *next; | |
1754 | ||
1755 | CLEAR_HARD_REG_BIT (hard_regs_in_table, regno); | |
1756 | ||
1757 | for (i = regno + 1; i < endregno; i++) | |
1758 | { | |
1759 | in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, i); | |
1760 | CLEAR_HARD_REG_BIT (hard_regs_in_table, i); | |
1761 | delete_reg_equiv (i); | |
30f72379 | 1762 | REG_TICK (i)++; |
7afe21cc RK |
1763 | } |
1764 | ||
1765 | if (in_table) | |
1766 | for (hash = 0; hash < NBUCKETS; hash++) | |
1767 | for (p = table[hash]; p; p = next) | |
1768 | { | |
1769 | next = p->next_same_hash; | |
1770 | ||
1771 | if (GET_CODE (p->exp) != REG | |
1772 | || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER) | |
1773 | continue; | |
1774 | ||
1775 | tregno = REGNO (p->exp); | |
1776 | tendregno | |
1777 | = tregno + HARD_REGNO_NREGS (tregno, GET_MODE (p->exp)); | |
1778 | if (tendregno > regno && tregno < endregno) | |
925be47c | 1779 | remove_from_table (p, hash); |
7afe21cc RK |
1780 | } |
1781 | } | |
1782 | ||
1783 | return; | |
1784 | } | |
1785 | ||
1786 | if (GET_CODE (x) == SUBREG) | |
1787 | { | |
1788 | if (GET_CODE (SUBREG_REG (x)) != REG) | |
1789 | abort (); | |
bb4034b3 | 1790 | invalidate (SUBREG_REG (x), VOIDmode); |
7afe21cc RK |
1791 | return; |
1792 | } | |
1793 | ||
aac5cc16 RH |
1794 | /* If X is a parallel, invalidate all of its elements. */ |
1795 | ||
1796 | if (GET_CODE (x) == PARALLEL) | |
1797 | { | |
1798 | for (i = XVECLEN (x, 0) - 1; i >= 0 ; --i) | |
1799 | invalidate (XVECEXP (x, 0, i), VOIDmode); | |
1800 | return; | |
1801 | } | |
1802 | ||
1803 | /* If X is an expr_list, this is part of a disjoint return value; | |
1804 | extract the location in question ignoring the offset. */ | |
1805 | ||
1806 | if (GET_CODE (x) == EXPR_LIST) | |
1807 | { | |
1808 | invalidate (XEXP (x, 0), VOIDmode); | |
1809 | return; | |
1810 | } | |
1811 | ||
7afe21cc RK |
1812 | /* X is not a register; it must be a memory reference with |
1813 | a nonvarying address. Remove all hash table elements | |
1814 | that refer to overlapping pieces of memory. */ | |
1815 | ||
1816 | if (GET_CODE (x) != MEM) | |
1817 | abort (); | |
7afe21cc | 1818 | |
bb4034b3 JW |
1819 | if (full_mode == VOIDmode) |
1820 | full_mode = GET_MODE (x); | |
1821 | ||
7afe21cc RK |
1822 | for (i = 0; i < NBUCKETS; i++) |
1823 | { | |
1824 | register struct table_elt *next; | |
1825 | for (p = table[i]; p; p = next) | |
1826 | { | |
1827 | next = p->next_same_hash; | |
9ae8ffe7 JL |
1828 | /* Invalidate ASM_OPERANDS which reference memory (this is easier |
1829 | than checking all the aliases). */ | |
1830 | if (p->in_memory | |
1831 | && (GET_CODE (p->exp) != MEM | |
1832 | || true_dependence (x, full_mode, p->exp, cse_rtx_varies_p))) | |
7afe21cc RK |
1833 | remove_from_table (p, i); |
1834 | } | |
1835 | } | |
1836 | } | |
1837 | ||
1838 | /* Remove all expressions that refer to register REGNO, | |
1839 | since they are already invalid, and we are about to | |
1840 | mark that register valid again and don't want the old | |
1841 | expressions to reappear as valid. */ | |
1842 | ||
1843 | static void | |
1844 | remove_invalid_refs (regno) | |
1845 | int regno; | |
1846 | { | |
1847 | register int i; | |
1848 | register struct table_elt *p, *next; | |
1849 | ||
1850 | for (i = 0; i < NBUCKETS; i++) | |
1851 | for (p = table[i]; p; p = next) | |
1852 | { | |
1853 | next = p->next_same_hash; | |
1854 | if (GET_CODE (p->exp) != REG | |
906c4e36 | 1855 | && refers_to_regno_p (regno, regno + 1, p->exp, NULL_PTR)) |
7afe21cc RK |
1856 | remove_from_table (p, i); |
1857 | } | |
1858 | } | |
34c73909 R |
1859 | |
1860 | /* Likewise for a subreg with subreg_reg WORD and mode MODE. */ | |
1861 | static void | |
1862 | remove_invalid_subreg_refs (regno, word, mode) | |
1863 | int regno; | |
1864 | int word; | |
1865 | enum machine_mode mode; | |
1866 | { | |
1867 | register int i; | |
1868 | register struct table_elt *p, *next; | |
1869 | int end = word + (GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD; | |
1870 | ||
1871 | for (i = 0; i < NBUCKETS; i++) | |
1872 | for (p = table[i]; p; p = next) | |
1873 | { | |
1874 | rtx exp; | |
1875 | next = p->next_same_hash; | |
1876 | ||
1877 | exp = p->exp; | |
1878 | if (GET_CODE (p->exp) != REG | |
1879 | && (GET_CODE (exp) != SUBREG | |
1880 | || GET_CODE (SUBREG_REG (exp)) != REG | |
1881 | || REGNO (SUBREG_REG (exp)) != regno | |
1882 | || (((SUBREG_WORD (exp) | |
1883 | + (GET_MODE_SIZE (GET_MODE (exp)) - 1) / UNITS_PER_WORD) | |
1884 | >= word) | |
1885 | && SUBREG_WORD (exp) <= end)) | |
1886 | && refers_to_regno_p (regno, regno + 1, p->exp, NULL_PTR)) | |
1887 | remove_from_table (p, i); | |
1888 | } | |
1889 | } | |
7afe21cc RK |
1890 | \f |
1891 | /* Recompute the hash codes of any valid entries in the hash table that | |
1892 | reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG. | |
1893 | ||
1894 | This is called when we make a jump equivalence. */ | |
1895 | ||
1896 | static void | |
1897 | rehash_using_reg (x) | |
1898 | rtx x; | |
1899 | { | |
973838fd | 1900 | unsigned int i; |
7afe21cc | 1901 | struct table_elt *p, *next; |
2197a88a | 1902 | unsigned hash; |
7afe21cc RK |
1903 | |
1904 | if (GET_CODE (x) == SUBREG) | |
1905 | x = SUBREG_REG (x); | |
1906 | ||
1907 | /* If X is not a register or if the register is known not to be in any | |
1908 | valid entries in the table, we have no work to do. */ | |
1909 | ||
1910 | if (GET_CODE (x) != REG | |
30f72379 MM |
1911 | || REG_IN_TABLE (REGNO (x)) < 0 |
1912 | || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x))) | |
7afe21cc RK |
1913 | return; |
1914 | ||
1915 | /* Scan all hash chains looking for valid entries that mention X. | |
1916 | If we find one and it is in the wrong hash chain, move it. We can skip | |
1917 | objects that are registers, since they are handled specially. */ | |
1918 | ||
1919 | for (i = 0; i < NBUCKETS; i++) | |
1920 | for (p = table[i]; p; p = next) | |
1921 | { | |
1922 | next = p->next_same_hash; | |
1923 | if (GET_CODE (p->exp) != REG && reg_mentioned_p (x, p->exp) | |
538b78e7 | 1924 | && exp_equiv_p (p->exp, p->exp, 1, 0) |
7afe21cc RK |
1925 | && i != (hash = safe_hash (p->exp, p->mode) % NBUCKETS)) |
1926 | { | |
1927 | if (p->next_same_hash) | |
1928 | p->next_same_hash->prev_same_hash = p->prev_same_hash; | |
1929 | ||
1930 | if (p->prev_same_hash) | |
1931 | p->prev_same_hash->next_same_hash = p->next_same_hash; | |
1932 | else | |
1933 | table[i] = p->next_same_hash; | |
1934 | ||
1935 | p->next_same_hash = table[hash]; | |
1936 | p->prev_same_hash = 0; | |
1937 | if (table[hash]) | |
1938 | table[hash]->prev_same_hash = p; | |
1939 | table[hash] = p; | |
1940 | } | |
1941 | } | |
1942 | } | |
1943 | \f | |
7afe21cc RK |
1944 | /* Remove from the hash table any expression that is a call-clobbered |
1945 | register. Also update their TICK values. */ | |
1946 | ||
1947 | static void | |
1948 | invalidate_for_call () | |
1949 | { | |
1950 | int regno, endregno; | |
1951 | int i; | |
2197a88a | 1952 | unsigned hash; |
7afe21cc RK |
1953 | struct table_elt *p, *next; |
1954 | int in_table = 0; | |
1955 | ||
1956 | /* Go through all the hard registers. For each that is clobbered in | |
1957 | a CALL_INSN, remove the register from quantity chains and update | |
1958 | reg_tick if defined. Also see if any of these registers is currently | |
1959 | in the table. */ | |
1960 | ||
1961 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
1962 | if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno)) | |
1963 | { | |
1964 | delete_reg_equiv (regno); | |
30f72379 MM |
1965 | if (REG_TICK (regno) >= 0) |
1966 | REG_TICK (regno)++; | |
7afe21cc | 1967 | |
0e227018 | 1968 | in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0); |
7afe21cc RK |
1969 | } |
1970 | ||
1971 | /* In the case where we have no call-clobbered hard registers in the | |
1972 | table, we are done. Otherwise, scan the table and remove any | |
1973 | entry that overlaps a call-clobbered register. */ | |
1974 | ||
1975 | if (in_table) | |
1976 | for (hash = 0; hash < NBUCKETS; hash++) | |
1977 | for (p = table[hash]; p; p = next) | |
1978 | { | |
1979 | next = p->next_same_hash; | |
1980 | ||
9ae8ffe7 JL |
1981 | if (p->in_memory) |
1982 | { | |
1983 | remove_from_table (p, hash); | |
1984 | continue; | |
1985 | } | |
1986 | ||
7afe21cc RK |
1987 | if (GET_CODE (p->exp) != REG |
1988 | || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER) | |
1989 | continue; | |
1990 | ||
1991 | regno = REGNO (p->exp); | |
1992 | endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (p->exp)); | |
1993 | ||
1994 | for (i = regno; i < endregno; i++) | |
1995 | if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)) | |
1996 | { | |
1997 | remove_from_table (p, hash); | |
1998 | break; | |
1999 | } | |
2000 | } | |
2001 | } | |
2002 | \f | |
2003 | /* Given an expression X of type CONST, | |
2004 | and ELT which is its table entry (or 0 if it | |
2005 | is not in the hash table), | |
2006 | return an alternate expression for X as a register plus integer. | |
2007 | If none can be found, return 0. */ | |
2008 | ||
2009 | static rtx | |
2010 | use_related_value (x, elt) | |
2011 | rtx x; | |
2012 | struct table_elt *elt; | |
2013 | { | |
2014 | register struct table_elt *relt = 0; | |
2015 | register struct table_elt *p, *q; | |
906c4e36 | 2016 | HOST_WIDE_INT offset; |
7afe21cc RK |
2017 | |
2018 | /* First, is there anything related known? | |
2019 | If we have a table element, we can tell from that. | |
2020 | Otherwise, must look it up. */ | |
2021 | ||
2022 | if (elt != 0 && elt->related_value != 0) | |
2023 | relt = elt; | |
2024 | else if (elt == 0 && GET_CODE (x) == CONST) | |
2025 | { | |
2026 | rtx subexp = get_related_value (x); | |
2027 | if (subexp != 0) | |
2028 | relt = lookup (subexp, | |
2029 | safe_hash (subexp, GET_MODE (subexp)) % NBUCKETS, | |
2030 | GET_MODE (subexp)); | |
2031 | } | |
2032 | ||
2033 | if (relt == 0) | |
2034 | return 0; | |
2035 | ||
2036 | /* Search all related table entries for one that has an | |
2037 | equivalent register. */ | |
2038 | ||
2039 | p = relt; | |
2040 | while (1) | |
2041 | { | |
2042 | /* This loop is strange in that it is executed in two different cases. | |
2043 | The first is when X is already in the table. Then it is searching | |
2044 | the RELATED_VALUE list of X's class (RELT). The second case is when | |
2045 | X is not in the table. Then RELT points to a class for the related | |
2046 | value. | |
2047 | ||
2048 | Ensure that, whatever case we are in, that we ignore classes that have | |
2049 | the same value as X. */ | |
2050 | ||
2051 | if (rtx_equal_p (x, p->exp)) | |
2052 | q = 0; | |
2053 | else | |
2054 | for (q = p->first_same_value; q; q = q->next_same_value) | |
2055 | if (GET_CODE (q->exp) == REG) | |
2056 | break; | |
2057 | ||
2058 | if (q) | |
2059 | break; | |
2060 | ||
2061 | p = p->related_value; | |
2062 | ||
2063 | /* We went all the way around, so there is nothing to be found. | |
2064 | Alternatively, perhaps RELT was in the table for some other reason | |
2065 | and it has no related values recorded. */ | |
2066 | if (p == relt || p == 0) | |
2067 | break; | |
2068 | } | |
2069 | ||
2070 | if (q == 0) | |
2071 | return 0; | |
2072 | ||
2073 | offset = (get_integer_term (x) - get_integer_term (p->exp)); | |
2074 | /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */ | |
2075 | return plus_constant (q->exp, offset); | |
2076 | } | |
2077 | \f | |
2078 | /* Hash an rtx. We are careful to make sure the value is never negative. | |
2079 | Equivalent registers hash identically. | |
2080 | MODE is used in hashing for CONST_INTs only; | |
2081 | otherwise the mode of X is used. | |
2082 | ||
2083 | Store 1 in do_not_record if any subexpression is volatile. | |
2084 | ||
2085 | Store 1 in hash_arg_in_memory if X contains a MEM rtx | |
2086 | which does not have the RTX_UNCHANGING_P bit set. | |
2087 | In this case, also store 1 in hash_arg_in_struct | |
2088 | if there is a MEM rtx which has the MEM_IN_STRUCT_P bit set. | |
2089 | ||
2090 | Note that cse_insn knows that the hash code of a MEM expression | |
2091 | is just (int) MEM plus the hash code of the address. */ | |
2092 | ||
2197a88a | 2093 | static unsigned |
7afe21cc RK |
2094 | canon_hash (x, mode) |
2095 | rtx x; | |
2096 | enum machine_mode mode; | |
2097 | { | |
2098 | register int i, j; | |
2197a88a | 2099 | register unsigned hash = 0; |
7afe21cc RK |
2100 | register enum rtx_code code; |
2101 | register char *fmt; | |
2102 | ||
2103 | /* repeat is used to turn tail-recursion into iteration. */ | |
2104 | repeat: | |
2105 | if (x == 0) | |
2106 | return hash; | |
2107 | ||
2108 | code = GET_CODE (x); | |
2109 | switch (code) | |
2110 | { | |
2111 | case REG: | |
2112 | { | |
2113 | register int regno = REGNO (x); | |
2114 | ||
2115 | /* On some machines, we can't record any non-fixed hard register, | |
2116 | because extending its life will cause reload problems. We | |
9a794e50 RH |
2117 | consider ap, fp, and sp to be fixed for this purpose. |
2118 | ||
2119 | We also consider CCmode registers to be fixed for this purpose; | |
2120 | failure to do so leads to failure to simplify 0<100 type of | |
2121 | conditionals. | |
2122 | ||
0f41302f | 2123 | On all machines, we can't record any global registers. */ |
7afe21cc RK |
2124 | |
2125 | if (regno < FIRST_PSEUDO_REGISTER | |
2126 | && (global_regs[regno] | |
f95182a4 ILT |
2127 | || (SMALL_REGISTER_CLASSES |
2128 | && ! fixed_regs[regno] | |
7afe21cc | 2129 | && regno != FRAME_POINTER_REGNUM |
8bc169f2 | 2130 | && regno != HARD_FRAME_POINTER_REGNUM |
7afe21cc | 2131 | && regno != ARG_POINTER_REGNUM |
9a794e50 RH |
2132 | && regno != STACK_POINTER_REGNUM |
2133 | && GET_MODE_CLASS (GET_MODE (x)) != MODE_CC))) | |
7afe21cc RK |
2134 | { |
2135 | do_not_record = 1; | |
2136 | return 0; | |
2137 | } | |
30f72379 | 2138 | hash += ((unsigned) REG << 7) + (unsigned) REG_QTY (regno); |
2197a88a | 2139 | return hash; |
7afe21cc RK |
2140 | } |
2141 | ||
34c73909 R |
2142 | /* We handle SUBREG of a REG specially because the underlying |
2143 | reg changes its hash value with every value change; we don't | |
2144 | want to have to forget unrelated subregs when one subreg changes. */ | |
2145 | case SUBREG: | |
2146 | { | |
2147 | if (GET_CODE (SUBREG_REG (x)) == REG) | |
2148 | { | |
2149 | hash += (((unsigned) SUBREG << 7) | |
2150 | + REGNO (SUBREG_REG (x)) + SUBREG_WORD (x)); | |
2151 | return hash; | |
2152 | } | |
2153 | break; | |
2154 | } | |
2155 | ||
7afe21cc | 2156 | case CONST_INT: |
2197a88a RK |
2157 | { |
2158 | unsigned HOST_WIDE_INT tem = INTVAL (x); | |
2159 | hash += ((unsigned) CONST_INT << 7) + (unsigned) mode + tem; | |
2160 | return hash; | |
2161 | } | |
7afe21cc RK |
2162 | |
2163 | case CONST_DOUBLE: | |
2164 | /* This is like the general case, except that it only counts | |
2165 | the integers representing the constant. */ | |
2197a88a | 2166 | hash += (unsigned) code + (unsigned) GET_MODE (x); |
969c8517 RK |
2167 | if (GET_MODE (x) != VOIDmode) |
2168 | for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++) | |
2169 | { | |
2170 | unsigned tem = XINT (x, i); | |
2171 | hash += tem; | |
2172 | } | |
2173 | else | |
2174 | hash += ((unsigned) CONST_DOUBLE_LOW (x) | |
2175 | + (unsigned) CONST_DOUBLE_HIGH (x)); | |
7afe21cc RK |
2176 | return hash; |
2177 | ||
2178 | /* Assume there is only one rtx object for any given label. */ | |
2179 | case LABEL_REF: | |
3c543775 | 2180 | hash |
7bcac048 | 2181 | += ((unsigned) LABEL_REF << 7) + (unsigned long) XEXP (x, 0); |
2197a88a | 2182 | return hash; |
7afe21cc RK |
2183 | |
2184 | case SYMBOL_REF: | |
3c543775 | 2185 | hash |
7bcac048 | 2186 | += ((unsigned) SYMBOL_REF << 7) + (unsigned long) XSTR (x, 0); |
2197a88a | 2187 | return hash; |
7afe21cc RK |
2188 | |
2189 | case MEM: | |
2190 | if (MEM_VOLATILE_P (x)) | |
2191 | { | |
2192 | do_not_record = 1; | |
2193 | return 0; | |
2194 | } | |
9ad91d71 | 2195 | if (! RTX_UNCHANGING_P (x) || FIXED_BASE_PLUS_P (XEXP (x, 0))) |
7afe21cc RK |
2196 | { |
2197 | hash_arg_in_memory = 1; | |
2198 | if (MEM_IN_STRUCT_P (x)) hash_arg_in_struct = 1; | |
2199 | } | |
2200 | /* Now that we have already found this special case, | |
2201 | might as well speed it up as much as possible. */ | |
2197a88a | 2202 | hash += (unsigned) MEM; |
7afe21cc RK |
2203 | x = XEXP (x, 0); |
2204 | goto repeat; | |
2205 | ||
2206 | case PRE_DEC: | |
2207 | case PRE_INC: | |
2208 | case POST_DEC: | |
2209 | case POST_INC: | |
2210 | case PC: | |
2211 | case CC0: | |
2212 | case CALL: | |
2213 | case UNSPEC_VOLATILE: | |
2214 | do_not_record = 1; | |
2215 | return 0; | |
2216 | ||
2217 | case ASM_OPERANDS: | |
2218 | if (MEM_VOLATILE_P (x)) | |
2219 | { | |
2220 | do_not_record = 1; | |
2221 | return 0; | |
2222 | } | |
e9a25f70 JL |
2223 | break; |
2224 | ||
2225 | default: | |
2226 | break; | |
7afe21cc RK |
2227 | } |
2228 | ||
2229 | i = GET_RTX_LENGTH (code) - 1; | |
2197a88a | 2230 | hash += (unsigned) code + (unsigned) GET_MODE (x); |
7afe21cc RK |
2231 | fmt = GET_RTX_FORMAT (code); |
2232 | for (; i >= 0; i--) | |
2233 | { | |
2234 | if (fmt[i] == 'e') | |
2235 | { | |
2236 | rtx tem = XEXP (x, i); | |
7afe21cc RK |
2237 | |
2238 | /* If we are about to do the last recursive call | |
2239 | needed at this level, change it into iteration. | |
2240 | This function is called enough to be worth it. */ | |
2241 | if (i == 0) | |
2242 | { | |
2243 | x = tem; | |
2244 | goto repeat; | |
2245 | } | |
2246 | hash += canon_hash (tem, 0); | |
2247 | } | |
2248 | else if (fmt[i] == 'E') | |
2249 | for (j = 0; j < XVECLEN (x, i); j++) | |
2250 | hash += canon_hash (XVECEXP (x, i, j), 0); | |
2251 | else if (fmt[i] == 's') | |
2252 | { | |
2197a88a | 2253 | register unsigned char *p = (unsigned char *) XSTR (x, i); |
7afe21cc RK |
2254 | if (p) |
2255 | while (*p) | |
2197a88a | 2256 | hash += *p++; |
7afe21cc RK |
2257 | } |
2258 | else if (fmt[i] == 'i') | |
2259 | { | |
2197a88a RK |
2260 | register unsigned tem = XINT (x, i); |
2261 | hash += tem; | |
7afe21cc | 2262 | } |
e9a25f70 JL |
2263 | else if (fmt[i] == '0') |
2264 | /* unused */; | |
7afe21cc RK |
2265 | else |
2266 | abort (); | |
2267 | } | |
2268 | return hash; | |
2269 | } | |
2270 | ||
2271 | /* Like canon_hash but with no side effects. */ | |
2272 | ||
2197a88a | 2273 | static unsigned |
7afe21cc RK |
2274 | safe_hash (x, mode) |
2275 | rtx x; | |
2276 | enum machine_mode mode; | |
2277 | { | |
2278 | int save_do_not_record = do_not_record; | |
2279 | int save_hash_arg_in_memory = hash_arg_in_memory; | |
2280 | int save_hash_arg_in_struct = hash_arg_in_struct; | |
2197a88a | 2281 | unsigned hash = canon_hash (x, mode); |
7afe21cc RK |
2282 | hash_arg_in_memory = save_hash_arg_in_memory; |
2283 | hash_arg_in_struct = save_hash_arg_in_struct; | |
2284 | do_not_record = save_do_not_record; | |
2285 | return hash; | |
2286 | } | |
2287 | \f | |
2288 | /* Return 1 iff X and Y would canonicalize into the same thing, | |
2289 | without actually constructing the canonicalization of either one. | |
2290 | If VALIDATE is nonzero, | |
2291 | we assume X is an expression being processed from the rtl | |
2292 | and Y was found in the hash table. We check register refs | |
2293 | in Y for being marked as valid. | |
2294 | ||
2295 | If EQUAL_VALUES is nonzero, we allow a register to match a constant value | |
2296 | that is known to be in the register. Ordinarily, we don't allow them | |
2297 | to match, because letting them match would cause unpredictable results | |
2298 | in all the places that search a hash table chain for an equivalent | |
2299 | for a given value. A possible equivalent that has different structure | |
2300 | has its hash code computed from different data. Whether the hash code | |
38e01259 | 2301 | is the same as that of the given value is pure luck. */ |
7afe21cc RK |
2302 | |
2303 | static int | |
2304 | exp_equiv_p (x, y, validate, equal_values) | |
2305 | rtx x, y; | |
2306 | int validate; | |
2307 | int equal_values; | |
2308 | { | |
906c4e36 | 2309 | register int i, j; |
7afe21cc RK |
2310 | register enum rtx_code code; |
2311 | register char *fmt; | |
2312 | ||
2313 | /* Note: it is incorrect to assume an expression is equivalent to itself | |
2314 | if VALIDATE is nonzero. */ | |
2315 | if (x == y && !validate) | |
2316 | return 1; | |
2317 | if (x == 0 || y == 0) | |
2318 | return x == y; | |
2319 | ||
2320 | code = GET_CODE (x); | |
2321 | if (code != GET_CODE (y)) | |
2322 | { | |
2323 | if (!equal_values) | |
2324 | return 0; | |
2325 | ||
2326 | /* If X is a constant and Y is a register or vice versa, they may be | |
2327 | equivalent. We only have to validate if Y is a register. */ | |
2328 | if (CONSTANT_P (x) && GET_CODE (y) == REG | |
2329 | && REGNO_QTY_VALID_P (REGNO (y)) | |
30f72379 MM |
2330 | && GET_MODE (y) == qty_mode[REG_QTY (REGNO (y))] |
2331 | && rtx_equal_p (x, qty_const[REG_QTY (REGNO (y))]) | |
2332 | && (! validate || REG_IN_TABLE (REGNO (y)) == REG_TICK (REGNO (y)))) | |
7afe21cc RK |
2333 | return 1; |
2334 | ||
2335 | if (CONSTANT_P (y) && code == REG | |
2336 | && REGNO_QTY_VALID_P (REGNO (x)) | |
30f72379 MM |
2337 | && GET_MODE (x) == qty_mode[REG_QTY (REGNO (x))] |
2338 | && rtx_equal_p (y, qty_const[REG_QTY (REGNO (x))])) | |
7afe21cc RK |
2339 | return 1; |
2340 | ||
2341 | return 0; | |
2342 | } | |
2343 | ||
2344 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */ | |
2345 | if (GET_MODE (x) != GET_MODE (y)) | |
2346 | return 0; | |
2347 | ||
2348 | switch (code) | |
2349 | { | |
2350 | case PC: | |
2351 | case CC0: | |
2352 | return x == y; | |
2353 | ||
2354 | case CONST_INT: | |
58c8c593 | 2355 | return INTVAL (x) == INTVAL (y); |
7afe21cc RK |
2356 | |
2357 | case LABEL_REF: | |
7afe21cc RK |
2358 | return XEXP (x, 0) == XEXP (y, 0); |
2359 | ||
f54d4924 RK |
2360 | case SYMBOL_REF: |
2361 | return XSTR (x, 0) == XSTR (y, 0); | |
2362 | ||
7afe21cc RK |
2363 | case REG: |
2364 | { | |
2365 | int regno = REGNO (y); | |
2366 | int endregno | |
2367 | = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1 | |
2368 | : HARD_REGNO_NREGS (regno, GET_MODE (y))); | |
2369 | int i; | |
2370 | ||
2371 | /* If the quantities are not the same, the expressions are not | |
2372 | equivalent. If there are and we are not to validate, they | |
2373 | are equivalent. Otherwise, ensure all regs are up-to-date. */ | |
2374 | ||
30f72379 | 2375 | if (REG_QTY (REGNO (x)) != REG_QTY (regno)) |
7afe21cc RK |
2376 | return 0; |
2377 | ||
2378 | if (! validate) | |
2379 | return 1; | |
2380 | ||
2381 | for (i = regno; i < endregno; i++) | |
30f72379 | 2382 | if (REG_IN_TABLE (i) != REG_TICK (i)) |
7afe21cc RK |
2383 | return 0; |
2384 | ||
2385 | return 1; | |
2386 | } | |
2387 | ||
2388 | /* For commutative operations, check both orders. */ | |
2389 | case PLUS: | |
2390 | case MULT: | |
2391 | case AND: | |
2392 | case IOR: | |
2393 | case XOR: | |
2394 | case NE: | |
2395 | case EQ: | |
2396 | return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0), validate, equal_values) | |
2397 | && exp_equiv_p (XEXP (x, 1), XEXP (y, 1), | |
2398 | validate, equal_values)) | |
2399 | || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1), | |
2400 | validate, equal_values) | |
2401 | && exp_equiv_p (XEXP (x, 1), XEXP (y, 0), | |
2402 | validate, equal_values))); | |
e9a25f70 JL |
2403 | |
2404 | default: | |
2405 | break; | |
7afe21cc RK |
2406 | } |
2407 | ||
2408 | /* Compare the elements. If any pair of corresponding elements | |
2409 | fail to match, return 0 for the whole things. */ | |
2410 | ||
2411 | fmt = GET_RTX_FORMAT (code); | |
2412 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2413 | { | |
906c4e36 | 2414 | switch (fmt[i]) |
7afe21cc | 2415 | { |
906c4e36 | 2416 | case 'e': |
7afe21cc RK |
2417 | if (! exp_equiv_p (XEXP (x, i), XEXP (y, i), validate, equal_values)) |
2418 | return 0; | |
906c4e36 RK |
2419 | break; |
2420 | ||
2421 | case 'E': | |
7afe21cc RK |
2422 | if (XVECLEN (x, i) != XVECLEN (y, i)) |
2423 | return 0; | |
2424 | for (j = 0; j < XVECLEN (x, i); j++) | |
2425 | if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j), | |
2426 | validate, equal_values)) | |
2427 | return 0; | |
906c4e36 RK |
2428 | break; |
2429 | ||
2430 | case 's': | |
7afe21cc RK |
2431 | if (strcmp (XSTR (x, i), XSTR (y, i))) |
2432 | return 0; | |
906c4e36 RK |
2433 | break; |
2434 | ||
2435 | case 'i': | |
7afe21cc RK |
2436 | if (XINT (x, i) != XINT (y, i)) |
2437 | return 0; | |
906c4e36 RK |
2438 | break; |
2439 | ||
2440 | case 'w': | |
2441 | if (XWINT (x, i) != XWINT (y, i)) | |
2442 | return 0; | |
2443 | break; | |
2444 | ||
2445 | case '0': | |
2446 | break; | |
2447 | ||
2448 | default: | |
2449 | abort (); | |
7afe21cc | 2450 | } |
906c4e36 RK |
2451 | } |
2452 | ||
7afe21cc RK |
2453 | return 1; |
2454 | } | |
2455 | \f | |
2456 | /* Return 1 iff any subexpression of X matches Y. | |
2457 | Here we do not require that X or Y be valid (for registers referred to) | |
2458 | for being in the hash table. */ | |
2459 | ||
6cd4575e | 2460 | static int |
7afe21cc RK |
2461 | refers_to_p (x, y) |
2462 | rtx x, y; | |
2463 | { | |
2464 | register int i; | |
2465 | register enum rtx_code code; | |
2466 | register char *fmt; | |
2467 | ||
2468 | repeat: | |
2469 | if (x == y) | |
2470 | return 1; | |
2471 | if (x == 0 || y == 0) | |
2472 | return 0; | |
2473 | ||
2474 | code = GET_CODE (x); | |
2475 | /* If X as a whole has the same code as Y, they may match. | |
2476 | If so, return 1. */ | |
2477 | if (code == GET_CODE (y)) | |
2478 | { | |
2479 | if (exp_equiv_p (x, y, 0, 1)) | |
2480 | return 1; | |
2481 | } | |
2482 | ||
2483 | /* X does not match, so try its subexpressions. */ | |
2484 | ||
2485 | fmt = GET_RTX_FORMAT (code); | |
2486 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2487 | if (fmt[i] == 'e') | |
2488 | { | |
2489 | if (i == 0) | |
2490 | { | |
2491 | x = XEXP (x, 0); | |
2492 | goto repeat; | |
2493 | } | |
2494 | else | |
2495 | if (refers_to_p (XEXP (x, i), y)) | |
2496 | return 1; | |
2497 | } | |
2498 | else if (fmt[i] == 'E') | |
2499 | { | |
2500 | int j; | |
2501 | for (j = 0; j < XVECLEN (x, i); j++) | |
2502 | if (refers_to_p (XVECEXP (x, i, j), y)) | |
2503 | return 1; | |
2504 | } | |
2505 | ||
2506 | return 0; | |
2507 | } | |
2508 | \f | |
f451db89 JL |
2509 | /* Given ADDR and SIZE (a memory address, and the size of the memory reference), |
2510 | set PBASE, PSTART, and PEND which correspond to the base of the address, | |
2511 | the starting offset, and ending offset respectively. | |
2512 | ||
bb4034b3 | 2513 | ADDR is known to be a nonvarying address. */ |
f451db89 | 2514 | |
bb4034b3 JW |
2515 | /* ??? Despite what the comments say, this function is in fact frequently |
2516 | passed varying addresses. This does not appear to cause any problems. */ | |
f451db89 JL |
2517 | |
2518 | static void | |
2519 | set_nonvarying_address_components (addr, size, pbase, pstart, pend) | |
2520 | rtx addr; | |
2521 | int size; | |
2522 | rtx *pbase; | |
6500fb43 | 2523 | HOST_WIDE_INT *pstart, *pend; |
f451db89 JL |
2524 | { |
2525 | rtx base; | |
c85663b1 | 2526 | HOST_WIDE_INT start, end; |
f451db89 JL |
2527 | |
2528 | base = addr; | |
2529 | start = 0; | |
2530 | end = 0; | |
2531 | ||
e5e809f4 JL |
2532 | if (flag_pic && GET_CODE (base) == PLUS |
2533 | && XEXP (base, 0) == pic_offset_table_rtx) | |
2534 | base = XEXP (base, 1); | |
2535 | ||
f451db89 JL |
2536 | /* Registers with nonvarying addresses usually have constant equivalents; |
2537 | but the frame pointer register is also possible. */ | |
2538 | if (GET_CODE (base) == REG | |
2539 | && qty_const != 0 | |
2540 | && REGNO_QTY_VALID_P (REGNO (base)) | |
30f72379 MM |
2541 | && qty_mode[REG_QTY (REGNO (base))] == GET_MODE (base) |
2542 | && qty_const[REG_QTY (REGNO (base))] != 0) | |
2543 | base = qty_const[REG_QTY (REGNO (base))]; | |
f451db89 JL |
2544 | else if (GET_CODE (base) == PLUS |
2545 | && GET_CODE (XEXP (base, 1)) == CONST_INT | |
2546 | && GET_CODE (XEXP (base, 0)) == REG | |
2547 | && qty_const != 0 | |
2548 | && REGNO_QTY_VALID_P (REGNO (XEXP (base, 0))) | |
30f72379 | 2549 | && (qty_mode[REG_QTY (REGNO (XEXP (base, 0)))] |
f451db89 | 2550 | == GET_MODE (XEXP (base, 0))) |
30f72379 | 2551 | && qty_const[REG_QTY (REGNO (XEXP (base, 0)))]) |
f451db89 JL |
2552 | { |
2553 | start = INTVAL (XEXP (base, 1)); | |
30f72379 | 2554 | base = qty_const[REG_QTY (REGNO (XEXP (base, 0)))]; |
f451db89 | 2555 | } |
9c6b0bae | 2556 | /* This can happen as the result of virtual register instantiation, |
abc95ed3 | 2557 | if the initial offset is too large to be a valid address. */ |
9c6b0bae RK |
2558 | else if (GET_CODE (base) == PLUS |
2559 | && GET_CODE (XEXP (base, 0)) == REG | |
2560 | && GET_CODE (XEXP (base, 1)) == REG | |
2561 | && qty_const != 0 | |
2562 | && REGNO_QTY_VALID_P (REGNO (XEXP (base, 0))) | |
30f72379 | 2563 | && (qty_mode[REG_QTY (REGNO (XEXP (base, 0)))] |
9c6b0bae | 2564 | == GET_MODE (XEXP (base, 0))) |
30f72379 | 2565 | && qty_const[REG_QTY (REGNO (XEXP (base, 0)))] |
9c6b0bae | 2566 | && REGNO_QTY_VALID_P (REGNO (XEXP (base, 1))) |
30f72379 | 2567 | && (qty_mode[REG_QTY (REGNO (XEXP (base, 1)))] |
9c6b0bae | 2568 | == GET_MODE (XEXP (base, 1))) |
30f72379 | 2569 | && qty_const[REG_QTY (REGNO (XEXP (base, 1)))]) |
9c6b0bae | 2570 | { |
30f72379 MM |
2571 | rtx tem = qty_const[REG_QTY (REGNO (XEXP (base, 1)))]; |
2572 | base = qty_const[REG_QTY (REGNO (XEXP (base, 0)))]; | |
9c6b0bae RK |
2573 | |
2574 | /* One of the two values must be a constant. */ | |
2575 | if (GET_CODE (base) != CONST_INT) | |
2576 | { | |
2577 | if (GET_CODE (tem) != CONST_INT) | |
2578 | abort (); | |
2579 | start = INTVAL (tem); | |
2580 | } | |
2581 | else | |
2582 | { | |
2583 | start = INTVAL (base); | |
2584 | base = tem; | |
2585 | } | |
2586 | } | |
f451db89 | 2587 | |
c85663b1 RK |
2588 | /* Handle everything that we can find inside an address that has been |
2589 | viewed as constant. */ | |
f451db89 | 2590 | |
c85663b1 | 2591 | while (1) |
f451db89 | 2592 | { |
c85663b1 RK |
2593 | /* If no part of this switch does a "continue", the code outside |
2594 | will exit this loop. */ | |
2595 | ||
2596 | switch (GET_CODE (base)) | |
2597 | { | |
2598 | case LO_SUM: | |
2599 | /* By definition, operand1 of a LO_SUM is the associated constant | |
2600 | address. Use the associated constant address as the base | |
2601 | instead. */ | |
2602 | base = XEXP (base, 1); | |
2603 | continue; | |
2604 | ||
2605 | case CONST: | |
2606 | /* Strip off CONST. */ | |
2607 | base = XEXP (base, 0); | |
2608 | continue; | |
2609 | ||
2610 | case PLUS: | |
2611 | if (GET_CODE (XEXP (base, 1)) == CONST_INT) | |
2612 | { | |
2613 | start += INTVAL (XEXP (base, 1)); | |
2614 | base = XEXP (base, 0); | |
2615 | continue; | |
2616 | } | |
2617 | break; | |
2618 | ||
2619 | case AND: | |
2620 | /* Handle the case of an AND which is the negative of a power of | |
2621 | two. This is used to represent unaligned memory operations. */ | |
2622 | if (GET_CODE (XEXP (base, 1)) == CONST_INT | |
2623 | && exact_log2 (- INTVAL (XEXP (base, 1))) > 0) | |
2624 | { | |
2625 | set_nonvarying_address_components (XEXP (base, 0), size, | |
2626 | pbase, pstart, pend); | |
2627 | ||
2628 | /* Assume the worst misalignment. START is affected, but not | |
2629 | END, so compensate but adjusting SIZE. Don't lose any | |
2630 | constant we already had. */ | |
2631 | ||
2632 | size = *pend - *pstart - INTVAL (XEXP (base, 1)) - 1; | |
89046535 RK |
2633 | start += *pstart + INTVAL (XEXP (base, 1)) + 1; |
2634 | end += *pend; | |
c85663b1 RK |
2635 | base = *pbase; |
2636 | } | |
2637 | break; | |
e9a25f70 JL |
2638 | |
2639 | default: | |
2640 | break; | |
c85663b1 RK |
2641 | } |
2642 | ||
2643 | break; | |
f451db89 JL |
2644 | } |
2645 | ||
336d6f0a RK |
2646 | if (GET_CODE (base) == CONST_INT) |
2647 | { | |
2648 | start += INTVAL (base); | |
2649 | base = const0_rtx; | |
2650 | } | |
2651 | ||
f451db89 JL |
2652 | end = start + size; |
2653 | ||
2654 | /* Set the return values. */ | |
2655 | *pbase = base; | |
2656 | *pstart = start; | |
2657 | *pend = end; | |
2658 | } | |
2659 | ||
9ae8ffe7 JL |
2660 | /* Return 1 if X has a value that can vary even between two |
2661 | executions of the program. 0 means X can be compared reliably | |
2662 | against certain constants or near-constants. */ | |
7afe21cc RK |
2663 | |
2664 | static int | |
9ae8ffe7 JL |
2665 | cse_rtx_varies_p (x) |
2666 | register rtx x; | |
7afe21cc RK |
2667 | { |
2668 | /* We need not check for X and the equivalence class being of the same | |
2669 | mode because if X is equivalent to a constant in some mode, it | |
2670 | doesn't vary in any mode. */ | |
2671 | ||
9ae8ffe7 JL |
2672 | if (GET_CODE (x) == REG |
2673 | && REGNO_QTY_VALID_P (REGNO (x)) | |
30f72379 MM |
2674 | && GET_MODE (x) == qty_mode[REG_QTY (REGNO (x))] |
2675 | && qty_const[REG_QTY (REGNO (x))] != 0) | |
7afe21cc RK |
2676 | return 0; |
2677 | ||
9ae8ffe7 JL |
2678 | if (GET_CODE (x) == PLUS |
2679 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
2680 | && GET_CODE (XEXP (x, 0)) == REG | |
2681 | && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))) | |
2682 | && (GET_MODE (XEXP (x, 0)) | |
30f72379 MM |
2683 | == qty_mode[REG_QTY (REGNO (XEXP (x, 0)))]) |
2684 | && qty_const[REG_QTY (REGNO (XEXP (x, 0)))]) | |
7afe21cc RK |
2685 | return 0; |
2686 | ||
9c6b0bae RK |
2687 | /* This can happen as the result of virtual register instantiation, if |
2688 | the initial constant is too large to be a valid address. This gives | |
2689 | us a three instruction sequence, load large offset into a register, | |
2690 | load fp minus a constant into a register, then a MEM which is the | |
2691 | sum of the two `constant' registers. */ | |
9ae8ffe7 JL |
2692 | if (GET_CODE (x) == PLUS |
2693 | && GET_CODE (XEXP (x, 0)) == REG | |
2694 | && GET_CODE (XEXP (x, 1)) == REG | |
2695 | && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))) | |
2696 | && (GET_MODE (XEXP (x, 0)) | |
30f72379 MM |
2697 | == qty_mode[REG_QTY (REGNO (XEXP (x, 0)))]) |
2698 | && qty_const[REG_QTY (REGNO (XEXP (x, 0)))] | |
9ae8ffe7 JL |
2699 | && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))) |
2700 | && (GET_MODE (XEXP (x, 1)) | |
30f72379 MM |
2701 | == qty_mode[REG_QTY (REGNO (XEXP (x, 1)))]) |
2702 | && qty_const[REG_QTY (REGNO (XEXP (x, 1)))]) | |
9c6b0bae RK |
2703 | return 0; |
2704 | ||
9ae8ffe7 | 2705 | return rtx_varies_p (x); |
7afe21cc RK |
2706 | } |
2707 | \f | |
2708 | /* Canonicalize an expression: | |
2709 | replace each register reference inside it | |
2710 | with the "oldest" equivalent register. | |
2711 | ||
2712 | If INSN is non-zero and we are replacing a pseudo with a hard register | |
7722328e RK |
2713 | or vice versa, validate_change is used to ensure that INSN remains valid |
2714 | after we make our substitution. The calls are made with IN_GROUP non-zero | |
2715 | so apply_change_group must be called upon the outermost return from this | |
2716 | function (unless INSN is zero). The result of apply_change_group can | |
2717 | generally be discarded since the changes we are making are optional. */ | |
7afe21cc RK |
2718 | |
2719 | static rtx | |
2720 | canon_reg (x, insn) | |
2721 | rtx x; | |
2722 | rtx insn; | |
2723 | { | |
2724 | register int i; | |
2725 | register enum rtx_code code; | |
2726 | register char *fmt; | |
2727 | ||
2728 | if (x == 0) | |
2729 | return x; | |
2730 | ||
2731 | code = GET_CODE (x); | |
2732 | switch (code) | |
2733 | { | |
2734 | case PC: | |
2735 | case CC0: | |
2736 | case CONST: | |
2737 | case CONST_INT: | |
2738 | case CONST_DOUBLE: | |
2739 | case SYMBOL_REF: | |
2740 | case LABEL_REF: | |
2741 | case ADDR_VEC: | |
2742 | case ADDR_DIFF_VEC: | |
2743 | return x; | |
2744 | ||
2745 | case REG: | |
2746 | { | |
2747 | register int first; | |
2748 | ||
2749 | /* Never replace a hard reg, because hard regs can appear | |
2750 | in more than one machine mode, and we must preserve the mode | |
2751 | of each occurrence. Also, some hard regs appear in | |
2752 | MEMs that are shared and mustn't be altered. Don't try to | |
2753 | replace any reg that maps to a reg of class NO_REGS. */ | |
2754 | if (REGNO (x) < FIRST_PSEUDO_REGISTER | |
2755 | || ! REGNO_QTY_VALID_P (REGNO (x))) | |
2756 | return x; | |
2757 | ||
30f72379 | 2758 | first = qty_first_reg[REG_QTY (REGNO (x))]; |
7afe21cc RK |
2759 | return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first] |
2760 | : REGNO_REG_CLASS (first) == NO_REGS ? x | |
30f72379 | 2761 | : gen_rtx_REG (qty_mode[REG_QTY (REGNO (x))], first)); |
7afe21cc | 2762 | } |
e9a25f70 JL |
2763 | |
2764 | default: | |
2765 | break; | |
7afe21cc RK |
2766 | } |
2767 | ||
2768 | fmt = GET_RTX_FORMAT (code); | |
2769 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2770 | { | |
2771 | register int j; | |
2772 | ||
2773 | if (fmt[i] == 'e') | |
2774 | { | |
2775 | rtx new = canon_reg (XEXP (x, i), insn); | |
58873255 | 2776 | int insn_code; |
7afe21cc RK |
2777 | |
2778 | /* If replacing pseudo with hard reg or vice versa, ensure the | |
178c39f6 | 2779 | insn remains valid. Likewise if the insn has MATCH_DUPs. */ |
aee9dc31 RS |
2780 | if (insn != 0 && new != 0 |
2781 | && GET_CODE (new) == REG && GET_CODE (XEXP (x, i)) == REG | |
178c39f6 RK |
2782 | && (((REGNO (new) < FIRST_PSEUDO_REGISTER) |
2783 | != (REGNO (XEXP (x, i)) < FIRST_PSEUDO_REGISTER)) | |
58873255 RK |
2784 | || (insn_code = recog_memoized (insn)) < 0 |
2785 | || insn_n_dups[insn_code] > 0)) | |
77fa0940 | 2786 | validate_change (insn, &XEXP (x, i), new, 1); |
7afe21cc RK |
2787 | else |
2788 | XEXP (x, i) = new; | |
2789 | } | |
2790 | else if (fmt[i] == 'E') | |
2791 | for (j = 0; j < XVECLEN (x, i); j++) | |
2792 | XVECEXP (x, i, j) = canon_reg (XVECEXP (x, i, j), insn); | |
2793 | } | |
2794 | ||
2795 | return x; | |
2796 | } | |
2797 | \f | |
a2cabb29 | 2798 | /* LOC is a location within INSN that is an operand address (the contents of |
7afe21cc RK |
2799 | a MEM). Find the best equivalent address to use that is valid for this |
2800 | insn. | |
2801 | ||
2802 | On most CISC machines, complicated address modes are costly, and rtx_cost | |
2803 | is a good approximation for that cost. However, most RISC machines have | |
2804 | only a few (usually only one) memory reference formats. If an address is | |
2805 | valid at all, it is often just as cheap as any other address. Hence, for | |
2806 | RISC machines, we use the configuration macro `ADDRESS_COST' to compare the | |
2807 | costs of various addresses. For two addresses of equal cost, choose the one | |
2808 | with the highest `rtx_cost' value as that has the potential of eliminating | |
2809 | the most insns. For equal costs, we choose the first in the equivalence | |
2810 | class. Note that we ignore the fact that pseudo registers are cheaper | |
2811 | than hard registers here because we would also prefer the pseudo registers. | |
2812 | */ | |
2813 | ||
6cd4575e | 2814 | static void |
7afe21cc RK |
2815 | find_best_addr (insn, loc) |
2816 | rtx insn; | |
2817 | rtx *loc; | |
2818 | { | |
7a87758d | 2819 | struct table_elt *elt; |
7afe21cc | 2820 | rtx addr = *loc; |
7a87758d AS |
2821 | #ifdef ADDRESS_COST |
2822 | struct table_elt *p; | |
7afe21cc | 2823 | int found_better = 1; |
7a87758d | 2824 | #endif |
7afe21cc RK |
2825 | int save_do_not_record = do_not_record; |
2826 | int save_hash_arg_in_memory = hash_arg_in_memory; | |
2827 | int save_hash_arg_in_struct = hash_arg_in_struct; | |
7afe21cc RK |
2828 | int addr_volatile; |
2829 | int regno; | |
2197a88a | 2830 | unsigned hash; |
7afe21cc RK |
2831 | |
2832 | /* Do not try to replace constant addresses or addresses of local and | |
2833 | argument slots. These MEM expressions are made only once and inserted | |
2834 | in many instructions, as well as being used to control symbol table | |
2835 | output. It is not safe to clobber them. | |
2836 | ||
2837 | There are some uncommon cases where the address is already in a register | |
2838 | for some reason, but we cannot take advantage of that because we have | |
2839 | no easy way to unshare the MEM. In addition, looking up all stack | |
2840 | addresses is costly. */ | |
2841 | if ((GET_CODE (addr) == PLUS | |
2842 | && GET_CODE (XEXP (addr, 0)) == REG | |
2843 | && GET_CODE (XEXP (addr, 1)) == CONST_INT | |
2844 | && (regno = REGNO (XEXP (addr, 0)), | |
8bc169f2 DE |
2845 | regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM |
2846 | || regno == ARG_POINTER_REGNUM)) | |
7afe21cc | 2847 | || (GET_CODE (addr) == REG |
8bc169f2 DE |
2848 | && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM |
2849 | || regno == HARD_FRAME_POINTER_REGNUM | |
2850 | || regno == ARG_POINTER_REGNUM)) | |
e9a25f70 | 2851 | || GET_CODE (addr) == ADDRESSOF |
7afe21cc RK |
2852 | || CONSTANT_ADDRESS_P (addr)) |
2853 | return; | |
2854 | ||
2855 | /* If this address is not simply a register, try to fold it. This will | |
2856 | sometimes simplify the expression. Many simplifications | |
2857 | will not be valid, but some, usually applying the associative rule, will | |
2858 | be valid and produce better code. */ | |
8c87f107 RK |
2859 | if (GET_CODE (addr) != REG) |
2860 | { | |
2861 | rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX); | |
2862 | ||
2863 | if (1 | |
2864 | #ifdef ADDRESS_COST | |
2f541799 MM |
2865 | && (CSE_ADDRESS_COST (folded) < CSE_ADDRESS_COST (addr) |
2866 | || (CSE_ADDRESS_COST (folded) == CSE_ADDRESS_COST (addr) | |
9a252d29 | 2867 | && rtx_cost (folded, MEM) > rtx_cost (addr, MEM))) |
8c87f107 | 2868 | #else |
9a252d29 | 2869 | && rtx_cost (folded, MEM) < rtx_cost (addr, MEM) |
8c87f107 RK |
2870 | #endif |
2871 | && validate_change (insn, loc, folded, 0)) | |
2872 | addr = folded; | |
2873 | } | |
7afe21cc | 2874 | |
42495ca0 RK |
2875 | /* If this address is not in the hash table, we can't look for equivalences |
2876 | of the whole address. Also, ignore if volatile. */ | |
2877 | ||
7afe21cc | 2878 | do_not_record = 0; |
2197a88a | 2879 | hash = HASH (addr, Pmode); |
7afe21cc RK |
2880 | addr_volatile = do_not_record; |
2881 | do_not_record = save_do_not_record; | |
2882 | hash_arg_in_memory = save_hash_arg_in_memory; | |
2883 | hash_arg_in_struct = save_hash_arg_in_struct; | |
2884 | ||
2885 | if (addr_volatile) | |
2886 | return; | |
2887 | ||
2197a88a | 2888 | elt = lookup (addr, hash, Pmode); |
7afe21cc | 2889 | |
7afe21cc | 2890 | #ifndef ADDRESS_COST |
42495ca0 RK |
2891 | if (elt) |
2892 | { | |
2d8b0f3a | 2893 | int our_cost = elt->cost; |
42495ca0 RK |
2894 | |
2895 | /* Find the lowest cost below ours that works. */ | |
2896 | for (elt = elt->first_same_value; elt; elt = elt->next_same_value) | |
2897 | if (elt->cost < our_cost | |
2898 | && (GET_CODE (elt->exp) == REG | |
2899 | || exp_equiv_p (elt->exp, elt->exp, 1, 0)) | |
2900 | && validate_change (insn, loc, | |
906c4e36 | 2901 | canon_reg (copy_rtx (elt->exp), NULL_RTX), 0)) |
42495ca0 RK |
2902 | return; |
2903 | } | |
2904 | #else | |
7afe21cc | 2905 | |
42495ca0 RK |
2906 | if (elt) |
2907 | { | |
2908 | /* We need to find the best (under the criteria documented above) entry | |
2909 | in the class that is valid. We use the `flag' field to indicate | |
2910 | choices that were invalid and iterate until we can't find a better | |
2911 | one that hasn't already been tried. */ | |
7afe21cc | 2912 | |
42495ca0 RK |
2913 | for (p = elt->first_same_value; p; p = p->next_same_value) |
2914 | p->flag = 0; | |
7afe21cc | 2915 | |
42495ca0 RK |
2916 | while (found_better) |
2917 | { | |
2f541799 | 2918 | int best_addr_cost = CSE_ADDRESS_COST (*loc); |
42495ca0 RK |
2919 | int best_rtx_cost = (elt->cost + 1) >> 1; |
2920 | struct table_elt *best_elt = elt; | |
2921 | ||
2922 | found_better = 0; | |
2923 | for (p = elt->first_same_value; p; p = p->next_same_value) | |
2f541799 | 2924 | if (! p->flag) |
42495ca0 | 2925 | { |
2f541799 MM |
2926 | if ((GET_CODE (p->exp) == REG |
2927 | || exp_equiv_p (p->exp, p->exp, 1, 0)) | |
2928 | && (CSE_ADDRESS_COST (p->exp) < best_addr_cost | |
2929 | || (CSE_ADDRESS_COST (p->exp) == best_addr_cost | |
2930 | && (p->cost + 1) >> 1 > best_rtx_cost))) | |
2931 | { | |
2932 | found_better = 1; | |
2933 | best_addr_cost = CSE_ADDRESS_COST (p->exp); | |
2934 | best_rtx_cost = (p->cost + 1) >> 1; | |
2935 | best_elt = p; | |
2936 | } | |
42495ca0 | 2937 | } |
7afe21cc | 2938 | |
42495ca0 RK |
2939 | if (found_better) |
2940 | { | |
2941 | if (validate_change (insn, loc, | |
906c4e36 RK |
2942 | canon_reg (copy_rtx (best_elt->exp), |
2943 | NULL_RTX), 0)) | |
42495ca0 RK |
2944 | return; |
2945 | else | |
2946 | best_elt->flag = 1; | |
2947 | } | |
2948 | } | |
2949 | } | |
7afe21cc | 2950 | |
42495ca0 RK |
2951 | /* If the address is a binary operation with the first operand a register |
2952 | and the second a constant, do the same as above, but looking for | |
2953 | equivalences of the register. Then try to simplify before checking for | |
2954 | the best address to use. This catches a few cases: First is when we | |
2955 | have REG+const and the register is another REG+const. We can often merge | |
2956 | the constants and eliminate one insn and one register. It may also be | |
2957 | that a machine has a cheap REG+REG+const. Finally, this improves the | |
2958 | code on the Alpha for unaligned byte stores. */ | |
2959 | ||
2960 | if (flag_expensive_optimizations | |
2961 | && (GET_RTX_CLASS (GET_CODE (*loc)) == '2' | |
2962 | || GET_RTX_CLASS (GET_CODE (*loc)) == 'c') | |
2963 | && GET_CODE (XEXP (*loc, 0)) == REG | |
2964 | && GET_CODE (XEXP (*loc, 1)) == CONST_INT) | |
7afe21cc | 2965 | { |
42495ca0 RK |
2966 | rtx c = XEXP (*loc, 1); |
2967 | ||
2968 | do_not_record = 0; | |
2197a88a | 2969 | hash = HASH (XEXP (*loc, 0), Pmode); |
42495ca0 RK |
2970 | do_not_record = save_do_not_record; |
2971 | hash_arg_in_memory = save_hash_arg_in_memory; | |
2972 | hash_arg_in_struct = save_hash_arg_in_struct; | |
2973 | ||
2197a88a | 2974 | elt = lookup (XEXP (*loc, 0), hash, Pmode); |
42495ca0 RK |
2975 | if (elt == 0) |
2976 | return; | |
2977 | ||
2978 | /* We need to find the best (under the criteria documented above) entry | |
2979 | in the class that is valid. We use the `flag' field to indicate | |
2980 | choices that were invalid and iterate until we can't find a better | |
2981 | one that hasn't already been tried. */ | |
7afe21cc | 2982 | |
7afe21cc | 2983 | for (p = elt->first_same_value; p; p = p->next_same_value) |
42495ca0 | 2984 | p->flag = 0; |
7afe21cc | 2985 | |
42495ca0 | 2986 | while (found_better) |
7afe21cc | 2987 | { |
2f541799 | 2988 | int best_addr_cost = CSE_ADDRESS_COST (*loc); |
42495ca0 RK |
2989 | int best_rtx_cost = (COST (*loc) + 1) >> 1; |
2990 | struct table_elt *best_elt = elt; | |
2991 | rtx best_rtx = *loc; | |
f6516aee JW |
2992 | int count; |
2993 | ||
2994 | /* This is at worst case an O(n^2) algorithm, so limit our search | |
2995 | to the first 32 elements on the list. This avoids trouble | |
2996 | compiling code with very long basic blocks that can easily | |
2997 | call cse_gen_binary so many times that we run out of memory. */ | |
42495ca0 RK |
2998 | |
2999 | found_better = 0; | |
f6516aee JW |
3000 | for (p = elt->first_same_value, count = 0; |
3001 | p && count < 32; | |
3002 | p = p->next_same_value, count++) | |
42495ca0 RK |
3003 | if (! p->flag |
3004 | && (GET_CODE (p->exp) == REG | |
3005 | || exp_equiv_p (p->exp, p->exp, 1, 0))) | |
3006 | { | |
96b0e481 | 3007 | rtx new = cse_gen_binary (GET_CODE (*loc), Pmode, p->exp, c); |
42495ca0 | 3008 | |
2f541799 MM |
3009 | if ((CSE_ADDRESS_COST (new) < best_addr_cost |
3010 | || (CSE_ADDRESS_COST (new) == best_addr_cost | |
42495ca0 RK |
3011 | && (COST (new) + 1) >> 1 > best_rtx_cost))) |
3012 | { | |
3013 | found_better = 1; | |
2f541799 | 3014 | best_addr_cost = CSE_ADDRESS_COST (new); |
42495ca0 RK |
3015 | best_rtx_cost = (COST (new) + 1) >> 1; |
3016 | best_elt = p; | |
3017 | best_rtx = new; | |
3018 | } | |
3019 | } | |
3020 | ||
3021 | if (found_better) | |
3022 | { | |
3023 | if (validate_change (insn, loc, | |
906c4e36 RK |
3024 | canon_reg (copy_rtx (best_rtx), |
3025 | NULL_RTX), 0)) | |
42495ca0 RK |
3026 | return; |
3027 | else | |
3028 | best_elt->flag = 1; | |
3029 | } | |
7afe21cc RK |
3030 | } |
3031 | } | |
3032 | #endif | |
3033 | } | |
3034 | \f | |
3035 | /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison | |
3036 | operation (EQ, NE, GT, etc.), follow it back through the hash table and | |
3037 | what values are being compared. | |
3038 | ||
3039 | *PARG1 and *PARG2 are updated to contain the rtx representing the values | |
3040 | actually being compared. For example, if *PARG1 was (cc0) and *PARG2 | |
3041 | was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were | |
3042 | compared to produce cc0. | |
3043 | ||
3044 | The return value is the comparison operator and is either the code of | |
3045 | A or the code corresponding to the inverse of the comparison. */ | |
3046 | ||
3047 | static enum rtx_code | |
13c9910f | 3048 | find_comparison_args (code, parg1, parg2, pmode1, pmode2) |
7afe21cc RK |
3049 | enum rtx_code code; |
3050 | rtx *parg1, *parg2; | |
13c9910f | 3051 | enum machine_mode *pmode1, *pmode2; |
7afe21cc RK |
3052 | { |
3053 | rtx arg1, arg2; | |
3054 | ||
3055 | arg1 = *parg1, arg2 = *parg2; | |
3056 | ||
3057 | /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */ | |
3058 | ||
b2796a4b | 3059 | while (arg2 == CONST0_RTX (GET_MODE (arg1))) |
7afe21cc RK |
3060 | { |
3061 | /* Set non-zero when we find something of interest. */ | |
3062 | rtx x = 0; | |
3063 | int reverse_code = 0; | |
3064 | struct table_elt *p = 0; | |
3065 | ||
3066 | /* If arg1 is a COMPARE, extract the comparison arguments from it. | |
3067 | On machines with CC0, this is the only case that can occur, since | |
3068 | fold_rtx will return the COMPARE or item being compared with zero | |
3069 | when given CC0. */ | |
3070 | ||
3071 | if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx) | |
3072 | x = arg1; | |
3073 | ||
3074 | /* If ARG1 is a comparison operator and CODE is testing for | |
3075 | STORE_FLAG_VALUE, get the inner arguments. */ | |
3076 | ||
3077 | else if (GET_RTX_CLASS (GET_CODE (arg1)) == '<') | |
3078 | { | |
c610adec RK |
3079 | if (code == NE |
3080 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT | |
3081 | && code == LT && STORE_FLAG_VALUE == -1) | |
3082 | #ifdef FLOAT_STORE_FLAG_VALUE | |
3083 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT | |
3084 | && FLOAT_STORE_FLAG_VALUE < 0) | |
3085 | #endif | |
3086 | ) | |
7afe21cc | 3087 | x = arg1; |
c610adec RK |
3088 | else if (code == EQ |
3089 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT | |
3090 | && code == GE && STORE_FLAG_VALUE == -1) | |
3091 | #ifdef FLOAT_STORE_FLAG_VALUE | |
3092 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT | |
3093 | && FLOAT_STORE_FLAG_VALUE < 0) | |
3094 | #endif | |
3095 | ) | |
7afe21cc RK |
3096 | x = arg1, reverse_code = 1; |
3097 | } | |
3098 | ||
3099 | /* ??? We could also check for | |
3100 | ||
3101 | (ne (and (eq (...) (const_int 1))) (const_int 0)) | |
3102 | ||
3103 | and related forms, but let's wait until we see them occurring. */ | |
3104 | ||
3105 | if (x == 0) | |
3106 | /* Look up ARG1 in the hash table and see if it has an equivalence | |
3107 | that lets us see what is being compared. */ | |
3108 | p = lookup (arg1, safe_hash (arg1, GET_MODE (arg1)) % NBUCKETS, | |
3109 | GET_MODE (arg1)); | |
3110 | if (p) p = p->first_same_value; | |
3111 | ||
3112 | for (; p; p = p->next_same_value) | |
3113 | { | |
3114 | enum machine_mode inner_mode = GET_MODE (p->exp); | |
3115 | ||
3116 | /* If the entry isn't valid, skip it. */ | |
3117 | if (! exp_equiv_p (p->exp, p->exp, 1, 0)) | |
3118 | continue; | |
3119 | ||
3120 | if (GET_CODE (p->exp) == COMPARE | |
3121 | /* Another possibility is that this machine has a compare insn | |
3122 | that includes the comparison code. In that case, ARG1 would | |
3123 | be equivalent to a comparison operation that would set ARG1 to | |
3124 | either STORE_FLAG_VALUE or zero. If this is an NE operation, | |
3125 | ORIG_CODE is the actual comparison being done; if it is an EQ, | |
3126 | we must reverse ORIG_CODE. On machine with a negative value | |
3127 | for STORE_FLAG_VALUE, also look at LT and GE operations. */ | |
3128 | || ((code == NE | |
3129 | || (code == LT | |
c610adec | 3130 | && GET_MODE_CLASS (inner_mode) == MODE_INT |
906c4e36 RK |
3131 | && (GET_MODE_BITSIZE (inner_mode) |
3132 | <= HOST_BITS_PER_WIDE_INT) | |
7afe21cc | 3133 | && (STORE_FLAG_VALUE |
906c4e36 RK |
3134 | & ((HOST_WIDE_INT) 1 |
3135 | << (GET_MODE_BITSIZE (inner_mode) - 1)))) | |
c610adec RK |
3136 | #ifdef FLOAT_STORE_FLAG_VALUE |
3137 | || (code == LT | |
3138 | && GET_MODE_CLASS (inner_mode) == MODE_FLOAT | |
3139 | && FLOAT_STORE_FLAG_VALUE < 0) | |
3140 | #endif | |
3141 | ) | |
7afe21cc RK |
3142 | && GET_RTX_CLASS (GET_CODE (p->exp)) == '<')) |
3143 | { | |
3144 | x = p->exp; | |
3145 | break; | |
3146 | } | |
3147 | else if ((code == EQ | |
3148 | || (code == GE | |
c610adec | 3149 | && GET_MODE_CLASS (inner_mode) == MODE_INT |
906c4e36 RK |
3150 | && (GET_MODE_BITSIZE (inner_mode) |
3151 | <= HOST_BITS_PER_WIDE_INT) | |
7afe21cc | 3152 | && (STORE_FLAG_VALUE |
906c4e36 RK |
3153 | & ((HOST_WIDE_INT) 1 |
3154 | << (GET_MODE_BITSIZE (inner_mode) - 1)))) | |
c610adec RK |
3155 | #ifdef FLOAT_STORE_FLAG_VALUE |
3156 | || (code == GE | |
3157 | && GET_MODE_CLASS (inner_mode) == MODE_FLOAT | |
3158 | && FLOAT_STORE_FLAG_VALUE < 0) | |
3159 | #endif | |
3160 | ) | |
7afe21cc RK |
3161 | && GET_RTX_CLASS (GET_CODE (p->exp)) == '<') |
3162 | { | |
3163 | reverse_code = 1; | |
3164 | x = p->exp; | |
3165 | break; | |
3166 | } | |
3167 | ||
3168 | /* If this is fp + constant, the equivalent is a better operand since | |
3169 | it may let us predict the value of the comparison. */ | |
3170 | else if (NONZERO_BASE_PLUS_P (p->exp)) | |
3171 | { | |
3172 | arg1 = p->exp; | |
3173 | continue; | |
3174 | } | |
3175 | } | |
3176 | ||
3177 | /* If we didn't find a useful equivalence for ARG1, we are done. | |
3178 | Otherwise, set up for the next iteration. */ | |
3179 | if (x == 0) | |
3180 | break; | |
3181 | ||
3182 | arg1 = XEXP (x, 0), arg2 = XEXP (x, 1); | |
3183 | if (GET_RTX_CLASS (GET_CODE (x)) == '<') | |
3184 | code = GET_CODE (x); | |
3185 | ||
3186 | if (reverse_code) | |
3187 | code = reverse_condition (code); | |
3188 | } | |
3189 | ||
13c9910f RS |
3190 | /* Return our results. Return the modes from before fold_rtx |
3191 | because fold_rtx might produce const_int, and then it's too late. */ | |
3192 | *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2); | |
7afe21cc RK |
3193 | *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0); |
3194 | ||
3195 | return code; | |
3196 | } | |
3197 | \f | |
3198 | /* Try to simplify a unary operation CODE whose output mode is to be | |
3199 | MODE with input operand OP whose mode was originally OP_MODE. | |
3200 | Return zero if no simplification can be made. */ | |
3201 | ||
3202 | rtx | |
3203 | simplify_unary_operation (code, mode, op, op_mode) | |
3204 | enum rtx_code code; | |
3205 | enum machine_mode mode; | |
3206 | rtx op; | |
3207 | enum machine_mode op_mode; | |
3208 | { | |
3209 | register int width = GET_MODE_BITSIZE (mode); | |
3210 | ||
3211 | /* The order of these tests is critical so that, for example, we don't | |
3212 | check the wrong mode (input vs. output) for a conversion operation, | |
3213 | such as FIX. At some point, this should be simplified. */ | |
3214 | ||
62c0ea12 | 3215 | #if !defined(REAL_IS_NOT_DOUBLE) || defined(REAL_ARITHMETIC) |
7afe21cc | 3216 | |
62c0ea12 RK |
3217 | if (code == FLOAT && GET_MODE (op) == VOIDmode |
3218 | && (GET_CODE (op) == CONST_DOUBLE || GET_CODE (op) == CONST_INT)) | |
7afe21cc | 3219 | { |
62c0ea12 | 3220 | HOST_WIDE_INT hv, lv; |
7afe21cc RK |
3221 | REAL_VALUE_TYPE d; |
3222 | ||
62c0ea12 RK |
3223 | if (GET_CODE (op) == CONST_INT) |
3224 | lv = INTVAL (op), hv = INTVAL (op) < 0 ? -1 : 0; | |
3225 | else | |
7ac4a266 | 3226 | lv = CONST_DOUBLE_LOW (op), hv = CONST_DOUBLE_HIGH (op); |
7afe21cc RK |
3227 | |
3228 | #ifdef REAL_ARITHMETIC | |
2ebcccf3 | 3229 | REAL_VALUE_FROM_INT (d, lv, hv, mode); |
7afe21cc | 3230 | #else |
62c0ea12 | 3231 | if (hv < 0) |
7afe21cc | 3232 | { |
62c0ea12 | 3233 | d = (double) (~ hv); |
906c4e36 RK |
3234 | d *= ((double) ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) |
3235 | * (double) ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2))); | |
62c0ea12 | 3236 | d += (double) (unsigned HOST_WIDE_INT) (~ lv); |
7afe21cc RK |
3237 | d = (- d - 1.0); |
3238 | } | |
3239 | else | |
3240 | { | |
62c0ea12 | 3241 | d = (double) hv; |
906c4e36 RK |
3242 | d *= ((double) ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) |
3243 | * (double) ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2))); | |
62c0ea12 | 3244 | d += (double) (unsigned HOST_WIDE_INT) lv; |
7afe21cc RK |
3245 | } |
3246 | #endif /* REAL_ARITHMETIC */ | |
940fd0b5 | 3247 | d = real_value_truncate (mode, d); |
7afe21cc RK |
3248 | return CONST_DOUBLE_FROM_REAL_VALUE (d, mode); |
3249 | } | |
62c0ea12 RK |
3250 | else if (code == UNSIGNED_FLOAT && GET_MODE (op) == VOIDmode |
3251 | && (GET_CODE (op) == CONST_DOUBLE || GET_CODE (op) == CONST_INT)) | |
7afe21cc | 3252 | { |
62c0ea12 | 3253 | HOST_WIDE_INT hv, lv; |
7afe21cc RK |
3254 | REAL_VALUE_TYPE d; |
3255 | ||
62c0ea12 RK |
3256 | if (GET_CODE (op) == CONST_INT) |
3257 | lv = INTVAL (op), hv = INTVAL (op) < 0 ? -1 : 0; | |
3258 | else | |
7ac4a266 | 3259 | lv = CONST_DOUBLE_LOW (op), hv = CONST_DOUBLE_HIGH (op); |
62c0ea12 | 3260 | |
a9c6464d RK |
3261 | if (op_mode == VOIDmode) |
3262 | { | |
3263 | /* We don't know how to interpret negative-looking numbers in | |
3264 | this case, so don't try to fold those. */ | |
3265 | if (hv < 0) | |
3266 | return 0; | |
3267 | } | |
3268 | else if (GET_MODE_BITSIZE (op_mode) >= HOST_BITS_PER_WIDE_INT * 2) | |
62c0ea12 RK |
3269 | ; |
3270 | else | |
3271 | hv = 0, lv &= GET_MODE_MASK (op_mode); | |
3272 | ||
7afe21cc | 3273 | #ifdef REAL_ARITHMETIC |
2ebcccf3 | 3274 | REAL_VALUE_FROM_UNSIGNED_INT (d, lv, hv, mode); |
7afe21cc | 3275 | #else |
62c0ea12 | 3276 | |
138cec59 | 3277 | d = (double) (unsigned HOST_WIDE_INT) hv; |
906c4e36 RK |
3278 | d *= ((double) ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) |
3279 | * (double) ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2))); | |
62c0ea12 | 3280 | d += (double) (unsigned HOST_WIDE_INT) lv; |
7afe21cc | 3281 | #endif /* REAL_ARITHMETIC */ |
940fd0b5 | 3282 | d = real_value_truncate (mode, d); |
7afe21cc RK |
3283 | return CONST_DOUBLE_FROM_REAL_VALUE (d, mode); |
3284 | } | |
3285 | #endif | |
3286 | ||
f89e32e9 RK |
3287 | if (GET_CODE (op) == CONST_INT |
3288 | && width <= HOST_BITS_PER_WIDE_INT && width > 0) | |
7afe21cc | 3289 | { |
906c4e36 RK |
3290 | register HOST_WIDE_INT arg0 = INTVAL (op); |
3291 | register HOST_WIDE_INT val; | |
7afe21cc RK |
3292 | |
3293 | switch (code) | |
3294 | { | |
3295 | case NOT: | |
3296 | val = ~ arg0; | |
3297 | break; | |
3298 | ||
3299 | case NEG: | |
3300 | val = - arg0; | |
3301 | break; | |
3302 | ||
3303 | case ABS: | |
3304 | val = (arg0 >= 0 ? arg0 : - arg0); | |
3305 | break; | |
3306 | ||
3307 | case FFS: | |
3308 | /* Don't use ffs here. Instead, get low order bit and then its | |
3309 | number. If arg0 is zero, this will return 0, as desired. */ | |
3310 | arg0 &= GET_MODE_MASK (mode); | |
3311 | val = exact_log2 (arg0 & (- arg0)) + 1; | |
3312 | break; | |
3313 | ||
3314 | case TRUNCATE: | |
3315 | val = arg0; | |
3316 | break; | |
3317 | ||
3318 | case ZERO_EXTEND: | |
3319 | if (op_mode == VOIDmode) | |
3320 | op_mode = mode; | |
82a5e898 | 3321 | if (GET_MODE_BITSIZE (op_mode) == HOST_BITS_PER_WIDE_INT) |
d80e9fd7 RS |
3322 | { |
3323 | /* If we were really extending the mode, | |
3324 | we would have to distinguish between zero-extension | |
3325 | and sign-extension. */ | |
3326 | if (width != GET_MODE_BITSIZE (op_mode)) | |
3327 | abort (); | |
3328 | val = arg0; | |
3329 | } | |
82a5e898 CH |
3330 | else if (GET_MODE_BITSIZE (op_mode) < HOST_BITS_PER_WIDE_INT) |
3331 | val = arg0 & ~((HOST_WIDE_INT) (-1) << GET_MODE_BITSIZE (op_mode)); | |
7afe21cc RK |
3332 | else |
3333 | return 0; | |
3334 | break; | |
3335 | ||
3336 | case SIGN_EXTEND: | |
3337 | if (op_mode == VOIDmode) | |
3338 | op_mode = mode; | |
82a5e898 | 3339 | if (GET_MODE_BITSIZE (op_mode) == HOST_BITS_PER_WIDE_INT) |
d80e9fd7 RS |
3340 | { |
3341 | /* If we were really extending the mode, | |
3342 | we would have to distinguish between zero-extension | |
3343 | and sign-extension. */ | |
3344 | if (width != GET_MODE_BITSIZE (op_mode)) | |
3345 | abort (); | |
3346 | val = arg0; | |
3347 | } | |
f12564b4 | 3348 | else if (GET_MODE_BITSIZE (op_mode) < HOST_BITS_PER_WIDE_INT) |
7afe21cc | 3349 | { |
82a5e898 CH |
3350 | val |
3351 | = arg0 & ~((HOST_WIDE_INT) (-1) << GET_MODE_BITSIZE (op_mode)); | |
3352 | if (val | |
3353 | & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (op_mode) - 1))) | |
3354 | val -= (HOST_WIDE_INT) 1 << GET_MODE_BITSIZE (op_mode); | |
7afe21cc RK |
3355 | } |
3356 | else | |
3357 | return 0; | |
3358 | break; | |
3359 | ||
d45cf215 RS |
3360 | case SQRT: |
3361 | return 0; | |
3362 | ||
7afe21cc RK |
3363 | default: |
3364 | abort (); | |
3365 | } | |
3366 | ||
3367 | /* Clear the bits that don't belong in our mode, | |
3368 | unless they and our sign bit are all one. | |
3369 | So we get either a reasonable negative value or a reasonable | |
3370 | unsigned value for this mode. */ | |
906c4e36 RK |
3371 | if (width < HOST_BITS_PER_WIDE_INT |
3372 | && ((val & ((HOST_WIDE_INT) (-1) << (width - 1))) | |
3373 | != ((HOST_WIDE_INT) (-1) << (width - 1)))) | |
4879acf6 | 3374 | val &= ((HOST_WIDE_INT) 1 << width) - 1; |
7afe21cc | 3375 | |
737e7965 JW |
3376 | /* If this would be an entire word for the target, but is not for |
3377 | the host, then sign-extend on the host so that the number will look | |
3378 | the same way on the host that it would on the target. | |
3379 | ||
3380 | For example, when building a 64 bit alpha hosted 32 bit sparc | |
3381 | targeted compiler, then we want the 32 bit unsigned value -1 to be | |
3382 | represented as a 64 bit value -1, and not as 0x00000000ffffffff. | |
3383 | The later confuses the sparc backend. */ | |
3384 | ||
3385 | if (BITS_PER_WORD < HOST_BITS_PER_WIDE_INT && BITS_PER_WORD == width | |
3386 | && (val & ((HOST_WIDE_INT) 1 << (width - 1)))) | |
3387 | val |= ((HOST_WIDE_INT) (-1) << width); | |
3388 | ||
906c4e36 | 3389 | return GEN_INT (val); |
7afe21cc RK |
3390 | } |
3391 | ||
3392 | /* We can do some operations on integer CONST_DOUBLEs. Also allow | |
0f41302f | 3393 | for a DImode operation on a CONST_INT. */ |
8e0ac43b | 3394 | else if (GET_MODE (op) == VOIDmode && width <= HOST_BITS_PER_INT * 2 |
7afe21cc RK |
3395 | && (GET_CODE (op) == CONST_DOUBLE || GET_CODE (op) == CONST_INT)) |
3396 | { | |
906c4e36 | 3397 | HOST_WIDE_INT l1, h1, lv, hv; |
7afe21cc RK |
3398 | |
3399 | if (GET_CODE (op) == CONST_DOUBLE) | |
3400 | l1 = CONST_DOUBLE_LOW (op), h1 = CONST_DOUBLE_HIGH (op); | |
3401 | else | |
3402 | l1 = INTVAL (op), h1 = l1 < 0 ? -1 : 0; | |
3403 | ||
3404 | switch (code) | |
3405 | { | |
3406 | case NOT: | |
3407 | lv = ~ l1; | |
3408 | hv = ~ h1; | |
3409 | break; | |
3410 | ||
3411 | case NEG: | |
3412 | neg_double (l1, h1, &lv, &hv); | |
3413 | break; | |
3414 | ||
3415 | case ABS: | |
3416 | if (h1 < 0) | |
3417 | neg_double (l1, h1, &lv, &hv); | |
3418 | else | |
3419 | lv = l1, hv = h1; | |
3420 | break; | |
3421 | ||
3422 | case FFS: | |
3423 | hv = 0; | |
3424 | if (l1 == 0) | |
906c4e36 | 3425 | lv = HOST_BITS_PER_WIDE_INT + exact_log2 (h1 & (-h1)) + 1; |
7afe21cc RK |
3426 | else |
3427 | lv = exact_log2 (l1 & (-l1)) + 1; | |
3428 | break; | |
3429 | ||
3430 | case TRUNCATE: | |
8e0ac43b | 3431 | /* This is just a change-of-mode, so do nothing. */ |
d50d63c0 | 3432 | lv = l1, hv = h1; |
7afe21cc RK |
3433 | break; |
3434 | ||
f72aed24 RS |
3435 | case ZERO_EXTEND: |
3436 | if (op_mode == VOIDmode | |
906c4e36 | 3437 | || GET_MODE_BITSIZE (op_mode) > HOST_BITS_PER_WIDE_INT) |
f72aed24 RS |
3438 | return 0; |
3439 | ||
3440 | hv = 0; | |
3441 | lv = l1 & GET_MODE_MASK (op_mode); | |
3442 | break; | |
3443 | ||
3444 | case SIGN_EXTEND: | |
3445 | if (op_mode == VOIDmode | |
906c4e36 | 3446 | || GET_MODE_BITSIZE (op_mode) > HOST_BITS_PER_WIDE_INT) |
f72aed24 RS |
3447 | return 0; |
3448 | else | |
3449 | { | |
3450 | lv = l1 & GET_MODE_MASK (op_mode); | |
906c4e36 RK |
3451 | if (GET_MODE_BITSIZE (op_mode) < HOST_BITS_PER_WIDE_INT |
3452 | && (lv & ((HOST_WIDE_INT) 1 | |
3453 | << (GET_MODE_BITSIZE (op_mode) - 1))) != 0) | |
3454 | lv -= (HOST_WIDE_INT) 1 << GET_MODE_BITSIZE (op_mode); | |
f72aed24 | 3455 | |
906c4e36 | 3456 | hv = (lv < 0) ? ~ (HOST_WIDE_INT) 0 : 0; |
f72aed24 RS |
3457 | } |
3458 | break; | |
3459 | ||
d45cf215 RS |
3460 | case SQRT: |
3461 | return 0; | |
3462 | ||
7afe21cc RK |
3463 | default: |
3464 | return 0; | |
3465 | } | |
3466 | ||
3467 | return immed_double_const (lv, hv, mode); | |
3468 | } | |
3469 | ||
3470 | #if ! defined (REAL_IS_NOT_DOUBLE) || defined (REAL_ARITHMETIC) | |
3471 | else if (GET_CODE (op) == CONST_DOUBLE | |
3472 | && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
3473 | { | |
3474 | REAL_VALUE_TYPE d; | |
3475 | jmp_buf handler; | |
3476 | rtx x; | |
3477 | ||
3478 | if (setjmp (handler)) | |
3479 | /* There used to be a warning here, but that is inadvisable. | |
3480 | People may want to cause traps, and the natural way | |
3481 | to do it should not get a warning. */ | |
3482 | return 0; | |
3483 | ||
3484 | set_float_handler (handler); | |
3485 | ||
3486 | REAL_VALUE_FROM_CONST_DOUBLE (d, op); | |
3487 | ||
3488 | switch (code) | |
3489 | { | |
3490 | case NEG: | |
3491 | d = REAL_VALUE_NEGATE (d); | |
3492 | break; | |
3493 | ||
3494 | case ABS: | |
8b3686ed | 3495 | if (REAL_VALUE_NEGATIVE (d)) |
7afe21cc RK |
3496 | d = REAL_VALUE_NEGATE (d); |
3497 | break; | |
3498 | ||
3499 | case FLOAT_TRUNCATE: | |
d3159aee | 3500 | d = real_value_truncate (mode, d); |
7afe21cc RK |
3501 | break; |
3502 | ||
3503 | case FLOAT_EXTEND: | |
3504 | /* All this does is change the mode. */ | |
3505 | break; | |
3506 | ||
3507 | case FIX: | |
d3159aee | 3508 | d = REAL_VALUE_RNDZINT (d); |
7afe21cc RK |
3509 | break; |
3510 | ||
3511 | case UNSIGNED_FIX: | |
d3159aee | 3512 | d = REAL_VALUE_UNSIGNED_RNDZINT (d); |
7afe21cc RK |
3513 | break; |
3514 | ||
d45cf215 RS |
3515 | case SQRT: |
3516 | return 0; | |
3517 | ||
7afe21cc RK |
3518 | default: |
3519 | abort (); | |
3520 | } | |
3521 | ||
560c94a2 | 3522 | x = CONST_DOUBLE_FROM_REAL_VALUE (d, mode); |
906c4e36 | 3523 | set_float_handler (NULL_PTR); |
7afe21cc RK |
3524 | return x; |
3525 | } | |
8e0ac43b RK |
3526 | |
3527 | else if (GET_CODE (op) == CONST_DOUBLE | |
3528 | && GET_MODE_CLASS (GET_MODE (op)) == MODE_FLOAT | |
3529 | && GET_MODE_CLASS (mode) == MODE_INT | |
906c4e36 | 3530 | && width <= HOST_BITS_PER_WIDE_INT && width > 0) |
7afe21cc RK |
3531 | { |
3532 | REAL_VALUE_TYPE d; | |
3533 | jmp_buf handler; | |
906c4e36 | 3534 | HOST_WIDE_INT val; |
7afe21cc RK |
3535 | |
3536 | if (setjmp (handler)) | |
3537 | return 0; | |
3538 | ||
3539 | set_float_handler (handler); | |
3540 | ||
3541 | REAL_VALUE_FROM_CONST_DOUBLE (d, op); | |
3542 | ||
3543 | switch (code) | |
3544 | { | |
3545 | case FIX: | |
3546 | val = REAL_VALUE_FIX (d); | |
3547 | break; | |
3548 | ||
3549 | case UNSIGNED_FIX: | |
3550 | val = REAL_VALUE_UNSIGNED_FIX (d); | |
3551 | break; | |
3552 | ||
3553 | default: | |
3554 | abort (); | |
3555 | } | |
3556 | ||
906c4e36 | 3557 | set_float_handler (NULL_PTR); |
7afe21cc RK |
3558 | |
3559 | /* Clear the bits that don't belong in our mode, | |
3560 | unless they and our sign bit are all one. | |
3561 | So we get either a reasonable negative value or a reasonable | |
3562 | unsigned value for this mode. */ | |
906c4e36 RK |
3563 | if (width < HOST_BITS_PER_WIDE_INT |
3564 | && ((val & ((HOST_WIDE_INT) (-1) << (width - 1))) | |
3565 | != ((HOST_WIDE_INT) (-1) << (width - 1)))) | |
3566 | val &= ((HOST_WIDE_INT) 1 << width) - 1; | |
7afe21cc | 3567 | |
ad89d6f6 TG |
3568 | /* If this would be an entire word for the target, but is not for |
3569 | the host, then sign-extend on the host so that the number will look | |
3570 | the same way on the host that it would on the target. | |
3571 | ||
3572 | For example, when building a 64 bit alpha hosted 32 bit sparc | |
3573 | targeted compiler, then we want the 32 bit unsigned value -1 to be | |
3574 | represented as a 64 bit value -1, and not as 0x00000000ffffffff. | |
3575 | The later confuses the sparc backend. */ | |
3576 | ||
3577 | if (BITS_PER_WORD < HOST_BITS_PER_WIDE_INT && BITS_PER_WORD == width | |
3578 | && (val & ((HOST_WIDE_INT) 1 << (width - 1)))) | |
3579 | val |= ((HOST_WIDE_INT) (-1) << width); | |
3580 | ||
906c4e36 | 3581 | return GEN_INT (val); |
7afe21cc RK |
3582 | } |
3583 | #endif | |
a6acbe15 RS |
3584 | /* This was formerly used only for non-IEEE float. |
3585 | eggert@twinsun.com says it is safe for IEEE also. */ | |
3586 | else | |
7afe21cc RK |
3587 | { |
3588 | /* There are some simplifications we can do even if the operands | |
a6acbe15 | 3589 | aren't constant. */ |
7afe21cc RK |
3590 | switch (code) |
3591 | { | |
3592 | case NEG: | |
3593 | case NOT: | |
3594 | /* (not (not X)) == X, similarly for NEG. */ | |
3595 | if (GET_CODE (op) == code) | |
3596 | return XEXP (op, 0); | |
3597 | break; | |
3598 | ||
3599 | case SIGN_EXTEND: | |
3600 | /* (sign_extend (truncate (minus (label_ref L1) (label_ref L2)))) | |
3601 | becomes just the MINUS if its mode is MODE. This allows | |
3602 | folding switch statements on machines using casesi (such as | |
3603 | the Vax). */ | |
3604 | if (GET_CODE (op) == TRUNCATE | |
3605 | && GET_MODE (XEXP (op, 0)) == mode | |
3606 | && GET_CODE (XEXP (op, 0)) == MINUS | |
3607 | && GET_CODE (XEXP (XEXP (op, 0), 0)) == LABEL_REF | |
3608 | && GET_CODE (XEXP (XEXP (op, 0), 1)) == LABEL_REF) | |
3609 | return XEXP (op, 0); | |
cceb347c RK |
3610 | |
3611 | #ifdef POINTERS_EXTEND_UNSIGNED | |
3612 | if (! POINTERS_EXTEND_UNSIGNED | |
3613 | && mode == Pmode && GET_MODE (op) == ptr_mode | |
3614 | && CONSTANT_P (op)) | |
3615 | return convert_memory_address (Pmode, op); | |
3616 | #endif | |
3617 | break; | |
3618 | ||
3619 | #ifdef POINTERS_EXTEND_UNSIGNED | |
3620 | case ZERO_EXTEND: | |
3621 | if (POINTERS_EXTEND_UNSIGNED | |
3622 | && mode == Pmode && GET_MODE (op) == ptr_mode | |
3623 | && CONSTANT_P (op)) | |
3624 | return convert_memory_address (Pmode, op); | |
7afe21cc | 3625 | break; |
cceb347c | 3626 | #endif |
e9a25f70 JL |
3627 | |
3628 | default: | |
3629 | break; | |
7afe21cc RK |
3630 | } |
3631 | ||
3632 | return 0; | |
3633 | } | |
7afe21cc RK |
3634 | } |
3635 | \f | |
3636 | /* Simplify a binary operation CODE with result mode MODE, operating on OP0 | |
3637 | and OP1. Return 0 if no simplification is possible. | |
3638 | ||
3639 | Don't use this for relational operations such as EQ or LT. | |
3640 | Use simplify_relational_operation instead. */ | |
3641 | ||
3642 | rtx | |
3643 | simplify_binary_operation (code, mode, op0, op1) | |
3644 | enum rtx_code code; | |
3645 | enum machine_mode mode; | |
3646 | rtx op0, op1; | |
3647 | { | |
906c4e36 RK |
3648 | register HOST_WIDE_INT arg0, arg1, arg0s, arg1s; |
3649 | HOST_WIDE_INT val; | |
7afe21cc | 3650 | int width = GET_MODE_BITSIZE (mode); |
96b0e481 | 3651 | rtx tem; |
7afe21cc RK |
3652 | |
3653 | /* Relational operations don't work here. We must know the mode | |
3654 | of the operands in order to do the comparison correctly. | |
3655 | Assuming a full word can give incorrect results. | |
3656 | Consider comparing 128 with -128 in QImode. */ | |
3657 | ||
3658 | if (GET_RTX_CLASS (code) == '<') | |
3659 | abort (); | |
3660 | ||
3661 | #if ! defined (REAL_IS_NOT_DOUBLE) || defined (REAL_ARITHMETIC) | |
3662 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
3663 | && GET_CODE (op0) == CONST_DOUBLE && GET_CODE (op1) == CONST_DOUBLE | |
3664 | && mode == GET_MODE (op0) && mode == GET_MODE (op1)) | |
3665 | { | |
3666 | REAL_VALUE_TYPE f0, f1, value; | |
3667 | jmp_buf handler; | |
3668 | ||
3669 | if (setjmp (handler)) | |
3670 | return 0; | |
3671 | ||
3672 | set_float_handler (handler); | |
3673 | ||
3674 | REAL_VALUE_FROM_CONST_DOUBLE (f0, op0); | |
3675 | REAL_VALUE_FROM_CONST_DOUBLE (f1, op1); | |
5352b11a RS |
3676 | f0 = real_value_truncate (mode, f0); |
3677 | f1 = real_value_truncate (mode, f1); | |
7afe21cc RK |
3678 | |
3679 | #ifdef REAL_ARITHMETIC | |
956d6950 JL |
3680 | #ifndef REAL_INFINITY |
3681 | if (code == DIV && REAL_VALUES_EQUAL (f1, dconst0)) | |
3682 | return 0; | |
3683 | #endif | |
d3159aee | 3684 | REAL_ARITHMETIC (value, rtx_to_tree_code (code), f0, f1); |
7afe21cc RK |
3685 | #else |
3686 | switch (code) | |
3687 | { | |
3688 | case PLUS: | |
3689 | value = f0 + f1; | |
3690 | break; | |
3691 | case MINUS: | |
3692 | value = f0 - f1; | |
3693 | break; | |
3694 | case MULT: | |
3695 | value = f0 * f1; | |
3696 | break; | |
3697 | case DIV: | |
3698 | #ifndef REAL_INFINITY | |
3699 | if (f1 == 0) | |
21d12b80 | 3700 | return 0; |
7afe21cc RK |
3701 | #endif |
3702 | value = f0 / f1; | |
3703 | break; | |
3704 | case SMIN: | |
3705 | value = MIN (f0, f1); | |
3706 | break; | |
3707 | case SMAX: | |
3708 | value = MAX (f0, f1); | |
3709 | break; | |
3710 | default: | |
3711 | abort (); | |
3712 | } | |
3713 | #endif | |
3714 | ||
5352b11a | 3715 | value = real_value_truncate (mode, value); |
831522a4 | 3716 | set_float_handler (NULL_PTR); |
560c94a2 | 3717 | return CONST_DOUBLE_FROM_REAL_VALUE (value, mode); |
7afe21cc | 3718 | } |
6076248a | 3719 | #endif /* not REAL_IS_NOT_DOUBLE, or REAL_ARITHMETIC */ |
7afe21cc RK |
3720 | |
3721 | /* We can fold some multi-word operations. */ | |
6076248a | 3722 | if (GET_MODE_CLASS (mode) == MODE_INT |
33085906 | 3723 | && width == HOST_BITS_PER_WIDE_INT * 2 |
fe873240 | 3724 | && (GET_CODE (op0) == CONST_DOUBLE || GET_CODE (op0) == CONST_INT) |
6076248a | 3725 | && (GET_CODE (op1) == CONST_DOUBLE || GET_CODE (op1) == CONST_INT)) |
7afe21cc | 3726 | { |
906c4e36 | 3727 | HOST_WIDE_INT l1, l2, h1, h2, lv, hv; |
7afe21cc | 3728 | |
fe873240 RK |
3729 | if (GET_CODE (op0) == CONST_DOUBLE) |
3730 | l1 = CONST_DOUBLE_LOW (op0), h1 = CONST_DOUBLE_HIGH (op0); | |
3731 | else | |
3732 | l1 = INTVAL (op0), h1 = l1 < 0 ? -1 : 0; | |
7afe21cc RK |
3733 | |
3734 | if (GET_CODE (op1) == CONST_DOUBLE) | |
3735 | l2 = CONST_DOUBLE_LOW (op1), h2 = CONST_DOUBLE_HIGH (op1); | |
3736 | else | |
3737 | l2 = INTVAL (op1), h2 = l2 < 0 ? -1 : 0; | |
3738 | ||
3739 | switch (code) | |
3740 | { | |
3741 | case MINUS: | |
3742 | /* A - B == A + (-B). */ | |
3743 | neg_double (l2, h2, &lv, &hv); | |
3744 | l2 = lv, h2 = hv; | |
3745 | ||
0f41302f | 3746 | /* .. fall through ... */ |
7afe21cc RK |
3747 | |
3748 | case PLUS: | |
3749 | add_double (l1, h1, l2, h2, &lv, &hv); | |
3750 | break; | |
3751 | ||
3752 | case MULT: | |
3753 | mul_double (l1, h1, l2, h2, &lv, &hv); | |
3754 | break; | |
3755 | ||
3756 | case DIV: case MOD: case UDIV: case UMOD: | |
3757 | /* We'd need to include tree.h to do this and it doesn't seem worth | |
3758 | it. */ | |
3759 | return 0; | |
3760 | ||
3761 | case AND: | |
3762 | lv = l1 & l2, hv = h1 & h2; | |
3763 | break; | |
3764 | ||
3765 | case IOR: | |
3766 | lv = l1 | l2, hv = h1 | h2; | |
3767 | break; | |
3768 | ||
3769 | case XOR: | |
3770 | lv = l1 ^ l2, hv = h1 ^ h2; | |
3771 | break; | |
3772 | ||
3773 | case SMIN: | |
906c4e36 RK |
3774 | if (h1 < h2 |
3775 | || (h1 == h2 | |
3776 | && ((unsigned HOST_WIDE_INT) l1 | |
3777 | < (unsigned HOST_WIDE_INT) l2))) | |
7afe21cc RK |
3778 | lv = l1, hv = h1; |
3779 | else | |
3780 | lv = l2, hv = h2; | |
3781 | break; | |
3782 | ||
3783 | case SMAX: | |
906c4e36 RK |
3784 | if (h1 > h2 |
3785 | || (h1 == h2 | |
3786 | && ((unsigned HOST_WIDE_INT) l1 | |
3787 | > (unsigned HOST_WIDE_INT) l2))) | |
7afe21cc RK |
3788 | lv = l1, hv = h1; |
3789 | else | |
3790 | lv = l2, hv = h2; | |
3791 | break; | |
3792 | ||
3793 | case UMIN: | |
906c4e36 RK |
3794 | if ((unsigned HOST_WIDE_INT) h1 < (unsigned HOST_WIDE_INT) h2 |
3795 | || (h1 == h2 | |
3796 | && ((unsigned HOST_WIDE_INT) l1 | |
3797 | < (unsigned HOST_WIDE_INT) l2))) | |
7afe21cc RK |
3798 | lv = l1, hv = h1; |
3799 | else | |
3800 | lv = l2, hv = h2; | |
3801 | break; | |
3802 | ||
3803 | case UMAX: | |
906c4e36 RK |
3804 | if ((unsigned HOST_WIDE_INT) h1 > (unsigned HOST_WIDE_INT) h2 |
3805 | || (h1 == h2 | |
3806 | && ((unsigned HOST_WIDE_INT) l1 | |
3807 | > (unsigned HOST_WIDE_INT) l2))) | |
7afe21cc RK |
3808 | lv = l1, hv = h1; |
3809 | else | |
3810 | lv = l2, hv = h2; | |
3811 | break; | |
3812 | ||
3813 | case LSHIFTRT: case ASHIFTRT: | |
45620ed4 | 3814 | case ASHIFT: |
7afe21cc RK |
3815 | case ROTATE: case ROTATERT: |
3816 | #ifdef SHIFT_COUNT_TRUNCATED | |
85c0a556 RK |
3817 | if (SHIFT_COUNT_TRUNCATED) |
3818 | l2 &= (GET_MODE_BITSIZE (mode) - 1), h2 = 0; | |
7afe21cc RK |
3819 | #endif |
3820 | ||
3821 | if (h2 != 0 || l2 < 0 || l2 >= GET_MODE_BITSIZE (mode)) | |
3822 | return 0; | |
3823 | ||
3824 | if (code == LSHIFTRT || code == ASHIFTRT) | |
3825 | rshift_double (l1, h1, l2, GET_MODE_BITSIZE (mode), &lv, &hv, | |
3826 | code == ASHIFTRT); | |
45620ed4 RK |
3827 | else if (code == ASHIFT) |
3828 | lshift_double (l1, h1, l2, GET_MODE_BITSIZE (mode), &lv, &hv, 1); | |
7afe21cc RK |
3829 | else if (code == ROTATE) |
3830 | lrotate_double (l1, h1, l2, GET_MODE_BITSIZE (mode), &lv, &hv); | |
3831 | else /* code == ROTATERT */ | |
3832 | rrotate_double (l1, h1, l2, GET_MODE_BITSIZE (mode), &lv, &hv); | |
3833 | break; | |
3834 | ||
3835 | default: | |
3836 | return 0; | |
3837 | } | |
3838 | ||
3839 | return immed_double_const (lv, hv, mode); | |
3840 | } | |
7afe21cc RK |
3841 | |
3842 | if (GET_CODE (op0) != CONST_INT || GET_CODE (op1) != CONST_INT | |
906c4e36 | 3843 | || width > HOST_BITS_PER_WIDE_INT || width == 0) |
7afe21cc RK |
3844 | { |
3845 | /* Even if we can't compute a constant result, | |
3846 | there are some cases worth simplifying. */ | |
3847 | ||
3848 | switch (code) | |
3849 | { | |
3850 | case PLUS: | |
3851 | /* In IEEE floating point, x+0 is not the same as x. Similarly | |
3852 | for the other optimizations below. */ | |
3853 | if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT | |
a83afb65 | 3854 | && FLOAT_MODE_P (mode) && ! flag_fast_math) |
7afe21cc RK |
3855 | break; |
3856 | ||
3857 | if (op1 == CONST0_RTX (mode)) | |
3858 | return op0; | |
3859 | ||
7afe21cc RK |
3860 | /* ((-a) + b) -> (b - a) and similarly for (a + (-b)) */ |
3861 | if (GET_CODE (op0) == NEG) | |
96b0e481 | 3862 | return cse_gen_binary (MINUS, mode, op1, XEXP (op0, 0)); |
7afe21cc | 3863 | else if (GET_CODE (op1) == NEG) |
96b0e481 | 3864 | return cse_gen_binary (MINUS, mode, op0, XEXP (op1, 0)); |
7afe21cc | 3865 | |
96b0e481 RK |
3866 | /* Handle both-operands-constant cases. We can only add |
3867 | CONST_INTs to constants since the sum of relocatable symbols | |
fe873240 RK |
3868 | can't be handled by most assemblers. Don't add CONST_INT |
3869 | to CONST_INT since overflow won't be computed properly if wider | |
3870 | than HOST_BITS_PER_WIDE_INT. */ | |
7afe21cc | 3871 | |
fe873240 RK |
3872 | if (CONSTANT_P (op0) && GET_MODE (op0) != VOIDmode |
3873 | && GET_CODE (op1) == CONST_INT) | |
96b0e481 | 3874 | return plus_constant (op0, INTVAL (op1)); |
fe873240 RK |
3875 | else if (CONSTANT_P (op1) && GET_MODE (op1) != VOIDmode |
3876 | && GET_CODE (op0) == CONST_INT) | |
96b0e481 | 3877 | return plus_constant (op1, INTVAL (op0)); |
7afe21cc | 3878 | |
30d69925 RK |
3879 | /* See if this is something like X * C - X or vice versa or |
3880 | if the multiplication is written as a shift. If so, we can | |
3881 | distribute and make a new multiply, shift, or maybe just | |
3882 | have X (if C is 2 in the example above). But don't make | |
3883 | real multiply if we didn't have one before. */ | |
3884 | ||
3885 | if (! FLOAT_MODE_P (mode)) | |
3886 | { | |
3887 | HOST_WIDE_INT coeff0 = 1, coeff1 = 1; | |
3888 | rtx lhs = op0, rhs = op1; | |
3889 | int had_mult = 0; | |
3890 | ||
3891 | if (GET_CODE (lhs) == NEG) | |
3892 | coeff0 = -1, lhs = XEXP (lhs, 0); | |
3893 | else if (GET_CODE (lhs) == MULT | |
3894 | && GET_CODE (XEXP (lhs, 1)) == CONST_INT) | |
3895 | { | |
3896 | coeff0 = INTVAL (XEXP (lhs, 1)), lhs = XEXP (lhs, 0); | |
3897 | had_mult = 1; | |
3898 | } | |
3899 | else if (GET_CODE (lhs) == ASHIFT | |
3900 | && GET_CODE (XEXP (lhs, 1)) == CONST_INT | |
3901 | && INTVAL (XEXP (lhs, 1)) >= 0 | |
3902 | && INTVAL (XEXP (lhs, 1)) < HOST_BITS_PER_WIDE_INT) | |
3903 | { | |
3904 | coeff0 = ((HOST_WIDE_INT) 1) << INTVAL (XEXP (lhs, 1)); | |
3905 | lhs = XEXP (lhs, 0); | |
3906 | } | |
3907 | ||
3908 | if (GET_CODE (rhs) == NEG) | |
3909 | coeff1 = -1, rhs = XEXP (rhs, 0); | |
3910 | else if (GET_CODE (rhs) == MULT | |
3911 | && GET_CODE (XEXP (rhs, 1)) == CONST_INT) | |
3912 | { | |
3913 | coeff1 = INTVAL (XEXP (rhs, 1)), rhs = XEXP (rhs, 0); | |
3914 | had_mult = 1; | |
3915 | } | |
3916 | else if (GET_CODE (rhs) == ASHIFT | |
3917 | && GET_CODE (XEXP (rhs, 1)) == CONST_INT | |
3918 | && INTVAL (XEXP (rhs, 1)) >= 0 | |
3919 | && INTVAL (XEXP (rhs, 1)) < HOST_BITS_PER_WIDE_INT) | |
3920 | { | |
3921 | coeff1 = ((HOST_WIDE_INT) 1) << INTVAL (XEXP (rhs, 1)); | |
3922 | rhs = XEXP (rhs, 0); | |
3923 | } | |
3924 | ||
3925 | if (rtx_equal_p (lhs, rhs)) | |
3926 | { | |
3927 | tem = cse_gen_binary (MULT, mode, lhs, | |
3928 | GEN_INT (coeff0 + coeff1)); | |
3929 | return (GET_CODE (tem) == MULT && ! had_mult) ? 0 : tem; | |
3930 | } | |
3931 | } | |
3932 | ||
96b0e481 RK |
3933 | /* If one of the operands is a PLUS or a MINUS, see if we can |
3934 | simplify this by the associative law. | |
3935 | Don't use the associative law for floating point. | |
3936 | The inaccuracy makes it nonassociative, | |
3937 | and subtle programs can break if operations are associated. */ | |
7afe21cc | 3938 | |
cbf6a543 | 3939 | if (INTEGRAL_MODE_P (mode) |
96b0e481 RK |
3940 | && (GET_CODE (op0) == PLUS || GET_CODE (op0) == MINUS |
3941 | || GET_CODE (op1) == PLUS || GET_CODE (op1) == MINUS) | |
3942 | && (tem = simplify_plus_minus (code, mode, op0, op1)) != 0) | |
3943 | return tem; | |
7afe21cc RK |
3944 | break; |
3945 | ||
3946 | case COMPARE: | |
3947 | #ifdef HAVE_cc0 | |
3948 | /* Convert (compare FOO (const_int 0)) to FOO unless we aren't | |
3949 | using cc0, in which case we want to leave it as a COMPARE | |
3950 | so we can distinguish it from a register-register-copy. | |
3951 | ||
3952 | In IEEE floating point, x-0 is not the same as x. */ | |
3953 | ||
3954 | if ((TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
a83afb65 | 3955 | || ! FLOAT_MODE_P (mode) || flag_fast_math) |
7afe21cc RK |
3956 | && op1 == CONST0_RTX (mode)) |
3957 | return op0; | |
3958 | #else | |
3959 | /* Do nothing here. */ | |
3960 | #endif | |
3961 | break; | |
3962 | ||
3963 | case MINUS: | |
21648b45 RK |
3964 | /* None of these optimizations can be done for IEEE |
3965 | floating point. */ | |
3966 | if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT | |
a83afb65 | 3967 | && FLOAT_MODE_P (mode) && ! flag_fast_math) |
21648b45 RK |
3968 | break; |
3969 | ||
a83afb65 RK |
3970 | /* We can't assume x-x is 0 even with non-IEEE floating point, |
3971 | but since it is zero except in very strange circumstances, we | |
3972 | will treat it as zero with -ffast-math. */ | |
7afe21cc RK |
3973 | if (rtx_equal_p (op0, op1) |
3974 | && ! side_effects_p (op0) | |
a83afb65 RK |
3975 | && (! FLOAT_MODE_P (mode) || flag_fast_math)) |
3976 | return CONST0_RTX (mode); | |
7afe21cc RK |
3977 | |
3978 | /* Change subtraction from zero into negation. */ | |
3979 | if (op0 == CONST0_RTX (mode)) | |
38a448ca | 3980 | return gen_rtx_NEG (mode, op1); |
7afe21cc | 3981 | |
96b0e481 RK |
3982 | /* (-1 - a) is ~a. */ |
3983 | if (op0 == constm1_rtx) | |
38a448ca | 3984 | return gen_rtx_NOT (mode, op1); |
96b0e481 | 3985 | |
7afe21cc RK |
3986 | /* Subtracting 0 has no effect. */ |
3987 | if (op1 == CONST0_RTX (mode)) | |
3988 | return op0; | |
3989 | ||
30d69925 RK |
3990 | /* See if this is something like X * C - X or vice versa or |
3991 | if the multiplication is written as a shift. If so, we can | |
3992 | distribute and make a new multiply, shift, or maybe just | |
3993 | have X (if C is 2 in the example above). But don't make | |
3994 | real multiply if we didn't have one before. */ | |
3995 | ||
3996 | if (! FLOAT_MODE_P (mode)) | |
3997 | { | |
3998 | HOST_WIDE_INT coeff0 = 1, coeff1 = 1; | |
3999 | rtx lhs = op0, rhs = op1; | |
4000 | int had_mult = 0; | |
4001 | ||
4002 | if (GET_CODE (lhs) == NEG) | |
4003 | coeff0 = -1, lhs = XEXP (lhs, 0); | |
4004 | else if (GET_CODE (lhs) == MULT | |
4005 | && GET_CODE (XEXP (lhs, 1)) == CONST_INT) | |
4006 | { | |
4007 | coeff0 = INTVAL (XEXP (lhs, 1)), lhs = XEXP (lhs, 0); | |
4008 | had_mult = 1; | |
4009 | } | |
4010 | else if (GET_CODE (lhs) == ASHIFT | |
4011 | && GET_CODE (XEXP (lhs, 1)) == CONST_INT | |
4012 | && INTVAL (XEXP (lhs, 1)) >= 0 | |
4013 | && INTVAL (XEXP (lhs, 1)) < HOST_BITS_PER_WIDE_INT) | |
4014 | { | |
4015 | coeff0 = ((HOST_WIDE_INT) 1) << INTVAL (XEXP (lhs, 1)); | |
4016 | lhs = XEXP (lhs, 0); | |
4017 | } | |
4018 | ||
4019 | if (GET_CODE (rhs) == NEG) | |
4020 | coeff1 = - 1, rhs = XEXP (rhs, 0); | |
4021 | else if (GET_CODE (rhs) == MULT | |
4022 | && GET_CODE (XEXP (rhs, 1)) == CONST_INT) | |
4023 | { | |
4024 | coeff1 = INTVAL (XEXP (rhs, 1)), rhs = XEXP (rhs, 0); | |
4025 | had_mult = 1; | |
4026 | } | |
4027 | else if (GET_CODE (rhs) == ASHIFT | |
4028 | && GET_CODE (XEXP (rhs, 1)) == CONST_INT | |
4029 | && INTVAL (XEXP (rhs, 1)) >= 0 | |
4030 | && INTVAL (XEXP (rhs, 1)) < HOST_BITS_PER_WIDE_INT) | |
4031 | { | |
4032 | coeff1 = ((HOST_WIDE_INT) 1) << INTVAL (XEXP (rhs, 1)); | |
4033 | rhs = XEXP (rhs, 0); | |
4034 | } | |
4035 | ||
4036 | if (rtx_equal_p (lhs, rhs)) | |
4037 | { | |
4038 | tem = cse_gen_binary (MULT, mode, lhs, | |
4039 | GEN_INT (coeff0 - coeff1)); | |
4040 | return (GET_CODE (tem) == MULT && ! had_mult) ? 0 : tem; | |
4041 | } | |
4042 | } | |
4043 | ||
7afe21cc RK |
4044 | /* (a - (-b)) -> (a + b). */ |
4045 | if (GET_CODE (op1) == NEG) | |
96b0e481 | 4046 | return cse_gen_binary (PLUS, mode, op0, XEXP (op1, 0)); |
7afe21cc | 4047 | |
96b0e481 RK |
4048 | /* If one of the operands is a PLUS or a MINUS, see if we can |
4049 | simplify this by the associative law. | |
4050 | Don't use the associative law for floating point. | |
7afe21cc RK |
4051 | The inaccuracy makes it nonassociative, |
4052 | and subtle programs can break if operations are associated. */ | |
7afe21cc | 4053 | |
cbf6a543 | 4054 | if (INTEGRAL_MODE_P (mode) |
96b0e481 RK |
4055 | && (GET_CODE (op0) == PLUS || GET_CODE (op0) == MINUS |
4056 | || GET_CODE (op1) == PLUS || GET_CODE (op1) == MINUS) | |
4057 | && (tem = simplify_plus_minus (code, mode, op0, op1)) != 0) | |
4058 | return tem; | |
7afe21cc RK |
4059 | |
4060 | /* Don't let a relocatable value get a negative coeff. */ | |
b5a09c41 | 4061 | if (GET_CODE (op1) == CONST_INT && GET_MODE (op0) != VOIDmode) |
7afe21cc | 4062 | return plus_constant (op0, - INTVAL (op1)); |
29d72c4b TG |
4063 | |
4064 | /* (x - (x & y)) -> (x & ~y) */ | |
4065 | if (GET_CODE (op1) == AND) | |
4066 | { | |
4067 | if (rtx_equal_p (op0, XEXP (op1, 0))) | |
38a448ca | 4068 | return cse_gen_binary (AND, mode, op0, gen_rtx_NOT (mode, XEXP (op1, 1))); |
29d72c4b | 4069 | if (rtx_equal_p (op0, XEXP (op1, 1))) |
38a448ca | 4070 | return cse_gen_binary (AND, mode, op0, gen_rtx_NOT (mode, XEXP (op1, 0))); |
29d72c4b | 4071 | } |
7afe21cc RK |
4072 | break; |
4073 | ||
4074 | case MULT: | |
4075 | if (op1 == constm1_rtx) | |
4076 | { | |
96b0e481 | 4077 | tem = simplify_unary_operation (NEG, mode, op0, mode); |
7afe21cc | 4078 | |
38a448ca | 4079 | return tem ? tem : gen_rtx_NEG (mode, op0); |
7afe21cc RK |
4080 | } |
4081 | ||
4082 | /* In IEEE floating point, x*0 is not always 0. */ | |
4083 | if ((TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
a83afb65 | 4084 | || ! FLOAT_MODE_P (mode) || flag_fast_math) |
7afe21cc RK |
4085 | && op1 == CONST0_RTX (mode) |
4086 | && ! side_effects_p (op0)) | |
4087 | return op1; | |
4088 | ||
4089 | /* In IEEE floating point, x*1 is not equivalent to x for nans. | |
4090 | However, ANSI says we can drop signals, | |
4091 | so we can do this anyway. */ | |
4092 | if (op1 == CONST1_RTX (mode)) | |
4093 | return op0; | |
4094 | ||
c407b802 RK |
4095 | /* Convert multiply by constant power of two into shift unless |
4096 | we are still generating RTL. This test is a kludge. */ | |
7afe21cc | 4097 | if (GET_CODE (op1) == CONST_INT |
c407b802 | 4098 | && (val = exact_log2 (INTVAL (op1))) >= 0 |
2d917903 JW |
4099 | /* If the mode is larger than the host word size, and the |
4100 | uppermost bit is set, then this isn't a power of two due | |
4101 | to implicit sign extension. */ | |
4102 | && (width <= HOST_BITS_PER_WIDE_INT | |
4103 | || val != HOST_BITS_PER_WIDE_INT - 1) | |
c407b802 | 4104 | && ! rtx_equal_function_value_matters) |
38a448ca | 4105 | return gen_rtx_ASHIFT (mode, op0, GEN_INT (val)); |
7afe21cc RK |
4106 | |
4107 | if (GET_CODE (op1) == CONST_DOUBLE | |
4108 | && GET_MODE_CLASS (GET_MODE (op1)) == MODE_FLOAT) | |
4109 | { | |
4110 | REAL_VALUE_TYPE d; | |
5a3d4bef RK |
4111 | jmp_buf handler; |
4112 | int op1is2, op1ism1; | |
4113 | ||
4114 | if (setjmp (handler)) | |
4115 | return 0; | |
4116 | ||
4117 | set_float_handler (handler); | |
7afe21cc | 4118 | REAL_VALUE_FROM_CONST_DOUBLE (d, op1); |
5a3d4bef RK |
4119 | op1is2 = REAL_VALUES_EQUAL (d, dconst2); |
4120 | op1ism1 = REAL_VALUES_EQUAL (d, dconstm1); | |
4121 | set_float_handler (NULL_PTR); | |
7afe21cc RK |
4122 | |
4123 | /* x*2 is x+x and x*(-1) is -x */ | |
5a3d4bef | 4124 | if (op1is2 && GET_MODE (op0) == mode) |
38a448ca | 4125 | return gen_rtx_PLUS (mode, op0, copy_rtx (op0)); |
7afe21cc | 4126 | |
5a3d4bef | 4127 | else if (op1ism1 && GET_MODE (op0) == mode) |
38a448ca | 4128 | return gen_rtx_NEG (mode, op0); |
7afe21cc RK |
4129 | } |
4130 | break; | |
4131 | ||
4132 | case IOR: | |
4133 | if (op1 == const0_rtx) | |
4134 | return op0; | |
4135 | if (GET_CODE (op1) == CONST_INT | |
4136 | && (INTVAL (op1) & GET_MODE_MASK (mode)) == GET_MODE_MASK (mode)) | |
4137 | return op1; | |
4138 | if (rtx_equal_p (op0, op1) && ! side_effects_p (op0)) | |
4139 | return op0; | |
4140 | /* A | (~A) -> -1 */ | |
4141 | if (((GET_CODE (op0) == NOT && rtx_equal_p (XEXP (op0, 0), op1)) | |
4142 | || (GET_CODE (op1) == NOT && rtx_equal_p (XEXP (op1, 0), op0))) | |
31dcf83f | 4143 | && ! side_effects_p (op0) |
8e7e5365 | 4144 | && GET_MODE_CLASS (mode) != MODE_CC) |
7afe21cc RK |
4145 | return constm1_rtx; |
4146 | break; | |
4147 | ||
4148 | case XOR: | |
4149 | if (op1 == const0_rtx) | |
4150 | return op0; | |
4151 | if (GET_CODE (op1) == CONST_INT | |
4152 | && (INTVAL (op1) & GET_MODE_MASK (mode)) == GET_MODE_MASK (mode)) | |
38a448ca | 4153 | return gen_rtx_NOT (mode, op0); |
31dcf83f | 4154 | if (op0 == op1 && ! side_effects_p (op0) |
8e7e5365 | 4155 | && GET_MODE_CLASS (mode) != MODE_CC) |
7afe21cc RK |
4156 | return const0_rtx; |
4157 | break; | |
4158 | ||
4159 | case AND: | |
4160 | if (op1 == const0_rtx && ! side_effects_p (op0)) | |
4161 | return const0_rtx; | |
4162 | if (GET_CODE (op1) == CONST_INT | |
4163 | && (INTVAL (op1) & GET_MODE_MASK (mode)) == GET_MODE_MASK (mode)) | |
4164 | return op0; | |
31dcf83f | 4165 | if (op0 == op1 && ! side_effects_p (op0) |
8e7e5365 | 4166 | && GET_MODE_CLASS (mode) != MODE_CC) |
7afe21cc RK |
4167 | return op0; |
4168 | /* A & (~A) -> 0 */ | |
4169 | if (((GET_CODE (op0) == NOT && rtx_equal_p (XEXP (op0, 0), op1)) | |
4170 | || (GET_CODE (op1) == NOT && rtx_equal_p (XEXP (op1, 0), op0))) | |
709ab4fc | 4171 | && ! side_effects_p (op0) |
8e7e5365 | 4172 | && GET_MODE_CLASS (mode) != MODE_CC) |
7afe21cc RK |
4173 | return const0_rtx; |
4174 | break; | |
4175 | ||
4176 | case UDIV: | |
4177 | /* Convert divide by power of two into shift (divide by 1 handled | |
4178 | below). */ | |
4179 | if (GET_CODE (op1) == CONST_INT | |
4180 | && (arg1 = exact_log2 (INTVAL (op1))) > 0) | |
38a448ca | 4181 | return gen_rtx_LSHIFTRT (mode, op0, GEN_INT (arg1)); |
7afe21cc | 4182 | |
0f41302f | 4183 | /* ... fall through ... */ |
7afe21cc RK |
4184 | |
4185 | case DIV: | |
4186 | if (op1 == CONST1_RTX (mode)) | |
4187 | return op0; | |
e7a522ba RS |
4188 | |
4189 | /* In IEEE floating point, 0/x is not always 0. */ | |
4190 | if ((TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
a83afb65 | 4191 | || ! FLOAT_MODE_P (mode) || flag_fast_math) |
e7a522ba RS |
4192 | && op0 == CONST0_RTX (mode) |
4193 | && ! side_effects_p (op1)) | |
7afe21cc | 4194 | return op0; |
e7a522ba | 4195 | |
7afe21cc | 4196 | #if ! defined (REAL_IS_NOT_DOUBLE) || defined (REAL_ARITHMETIC) |
a83afb65 RK |
4197 | /* Change division by a constant into multiplication. Only do |
4198 | this with -ffast-math until an expert says it is safe in | |
4199 | general. */ | |
7afe21cc RK |
4200 | else if (GET_CODE (op1) == CONST_DOUBLE |
4201 | && GET_MODE_CLASS (GET_MODE (op1)) == MODE_FLOAT | |
a83afb65 RK |
4202 | && op1 != CONST0_RTX (mode) |
4203 | && flag_fast_math) | |
7afe21cc RK |
4204 | { |
4205 | REAL_VALUE_TYPE d; | |
4206 | REAL_VALUE_FROM_CONST_DOUBLE (d, op1); | |
a83afb65 RK |
4207 | |
4208 | if (! REAL_VALUES_EQUAL (d, dconst0)) | |
4209 | { | |
7afe21cc | 4210 | #if defined (REAL_ARITHMETIC) |
a83afb65 | 4211 | REAL_ARITHMETIC (d, rtx_to_tree_code (DIV), dconst1, d); |
38a448ca RH |
4212 | return gen_rtx_MULT (mode, op0, |
4213 | CONST_DOUBLE_FROM_REAL_VALUE (d, mode)); | |
7afe21cc | 4214 | #else |
38a448ca RH |
4215 | return gen_rtx_MULT (mode, op0, |
4216 | CONST_DOUBLE_FROM_REAL_VALUE (1./d, mode)); | |
7afe21cc | 4217 | #endif |
a83afb65 RK |
4218 | } |
4219 | } | |
7afe21cc RK |
4220 | #endif |
4221 | break; | |
4222 | ||
4223 | case UMOD: | |
4224 | /* Handle modulus by power of two (mod with 1 handled below). */ | |
4225 | if (GET_CODE (op1) == CONST_INT | |
4226 | && exact_log2 (INTVAL (op1)) > 0) | |
38a448ca | 4227 | return gen_rtx_AND (mode, op0, GEN_INT (INTVAL (op1) - 1)); |
7afe21cc | 4228 | |
0f41302f | 4229 | /* ... fall through ... */ |
7afe21cc RK |
4230 | |
4231 | case MOD: | |
4232 | if ((op0 == const0_rtx || op1 == const1_rtx) | |
4233 | && ! side_effects_p (op0) && ! side_effects_p (op1)) | |
4234 | return const0_rtx; | |
4235 | break; | |
4236 | ||
4237 | case ROTATERT: | |
4238 | case ROTATE: | |
4239 | /* Rotating ~0 always results in ~0. */ | |
906c4e36 | 4240 | if (GET_CODE (op0) == CONST_INT && width <= HOST_BITS_PER_WIDE_INT |
7afe21cc RK |
4241 | && INTVAL (op0) == GET_MODE_MASK (mode) |
4242 | && ! side_effects_p (op1)) | |
4243 | return op0; | |
4244 | ||
0f41302f | 4245 | /* ... fall through ... */ |
7afe21cc | 4246 | |
7afe21cc RK |
4247 | case ASHIFT: |
4248 | case ASHIFTRT: | |
4249 | case LSHIFTRT: | |
4250 | if (op1 == const0_rtx) | |
4251 | return op0; | |
4252 | if (op0 == const0_rtx && ! side_effects_p (op1)) | |
4253 | return op0; | |
4254 | break; | |
4255 | ||
4256 | case SMIN: | |
906c4e36 RK |
4257 | if (width <= HOST_BITS_PER_WIDE_INT && GET_CODE (op1) == CONST_INT |
4258 | && INTVAL (op1) == (HOST_WIDE_INT) 1 << (width -1) | |
7afe21cc RK |
4259 | && ! side_effects_p (op0)) |
4260 | return op1; | |
4261 | else if (rtx_equal_p (op0, op1) && ! side_effects_p (op0)) | |
4262 | return op0; | |
4263 | break; | |
4264 | ||
4265 | case SMAX: | |
906c4e36 | 4266 | if (width <= HOST_BITS_PER_WIDE_INT && GET_CODE (op1) == CONST_INT |
dbbe6445 RK |
4267 | && (INTVAL (op1) |
4268 | == (unsigned HOST_WIDE_INT) GET_MODE_MASK (mode) >> 1) | |
7afe21cc RK |
4269 | && ! side_effects_p (op0)) |
4270 | return op1; | |
4271 | else if (rtx_equal_p (op0, op1) && ! side_effects_p (op0)) | |
4272 | return op0; | |
4273 | break; | |
4274 | ||
4275 | case UMIN: | |
4276 | if (op1 == const0_rtx && ! side_effects_p (op0)) | |
4277 | return op1; | |
4278 | else if (rtx_equal_p (op0, op1) && ! side_effects_p (op0)) | |
4279 | return op0; | |
4280 | break; | |
4281 | ||
4282 | case UMAX: | |
4283 | if (op1 == constm1_rtx && ! side_effects_p (op0)) | |
4284 | return op1; | |
4285 | else if (rtx_equal_p (op0, op1) && ! side_effects_p (op0)) | |
4286 | return op0; | |
4287 | break; | |
4288 | ||
4289 | default: | |
4290 | abort (); | |
4291 | } | |
4292 | ||
4293 | return 0; | |
4294 | } | |
4295 | ||
4296 | /* Get the integer argument values in two forms: | |
4297 | zero-extended in ARG0, ARG1 and sign-extended in ARG0S, ARG1S. */ | |
4298 | ||
4299 | arg0 = INTVAL (op0); | |
4300 | arg1 = INTVAL (op1); | |
4301 | ||
906c4e36 | 4302 | if (width < HOST_BITS_PER_WIDE_INT) |
7afe21cc | 4303 | { |
906c4e36 RK |
4304 | arg0 &= ((HOST_WIDE_INT) 1 << width) - 1; |
4305 | arg1 &= ((HOST_WIDE_INT) 1 << width) - 1; | |
7afe21cc RK |
4306 | |
4307 | arg0s = arg0; | |
906c4e36 RK |
4308 | if (arg0s & ((HOST_WIDE_INT) 1 << (width - 1))) |
4309 | arg0s |= ((HOST_WIDE_INT) (-1) << width); | |
7afe21cc RK |
4310 | |
4311 | arg1s = arg1; | |
906c4e36 RK |
4312 | if (arg1s & ((HOST_WIDE_INT) 1 << (width - 1))) |
4313 | arg1s |= ((HOST_WIDE_INT) (-1) << width); | |
7afe21cc RK |
4314 | } |
4315 | else | |
4316 | { | |
4317 | arg0s = arg0; | |
4318 | arg1s = arg1; | |
4319 | } | |
4320 | ||
4321 | /* Compute the value of the arithmetic. */ | |
4322 | ||
4323 | switch (code) | |
4324 | { | |
4325 | case PLUS: | |
538b78e7 | 4326 | val = arg0s + arg1s; |
7afe21cc RK |
4327 | break; |
4328 | ||
4329 | case MINUS: | |
538b78e7 | 4330 | val = arg0s - arg1s; |
7afe21cc RK |
4331 | break; |
4332 | ||
4333 | case MULT: | |
4334 | val = arg0s * arg1s; | |
4335 | break; | |
4336 | ||
4337 | case DIV: | |
4338 | if (arg1s == 0) | |
4339 | return 0; | |
4340 | val = arg0s / arg1s; | |
4341 | break; | |
4342 | ||
4343 | case MOD: | |
4344 | if (arg1s == 0) | |
4345 | return 0; | |
4346 | val = arg0s % arg1s; | |
4347 | break; | |
4348 | ||
4349 | case UDIV: | |
4350 | if (arg1 == 0) | |
4351 | return 0; | |
906c4e36 | 4352 | val = (unsigned HOST_WIDE_INT) arg0 / arg1; |
7afe21cc RK |
4353 | break; |
4354 | ||
4355 | case UMOD: | |
4356 | if (arg1 == 0) | |
4357 | return 0; | |
906c4e36 | 4358 | val = (unsigned HOST_WIDE_INT) arg0 % arg1; |
7afe21cc RK |
4359 | break; |
4360 | ||
4361 | case AND: | |
4362 | val = arg0 & arg1; | |
4363 | break; | |
4364 | ||
4365 | case IOR: | |
4366 | val = arg0 | arg1; | |
4367 | break; | |
4368 | ||
4369 | case XOR: | |
4370 | val = arg0 ^ arg1; | |
4371 | break; | |
4372 | ||
4373 | case LSHIFTRT: | |
4374 | /* If shift count is undefined, don't fold it; let the machine do | |
4375 | what it wants. But truncate it if the machine will do that. */ | |
4376 | if (arg1 < 0) | |
4377 | return 0; | |
4378 | ||
4379 | #ifdef SHIFT_COUNT_TRUNCATED | |
85c0a556 | 4380 | if (SHIFT_COUNT_TRUNCATED) |
4d61f8c5 | 4381 | arg1 %= width; |
7afe21cc RK |
4382 | #endif |
4383 | ||
906c4e36 | 4384 | val = ((unsigned HOST_WIDE_INT) arg0) >> arg1; |
7afe21cc RK |
4385 | break; |
4386 | ||
4387 | case ASHIFT: | |
7afe21cc RK |
4388 | if (arg1 < 0) |
4389 | return 0; | |
4390 | ||
4391 | #ifdef SHIFT_COUNT_TRUNCATED | |
85c0a556 | 4392 | if (SHIFT_COUNT_TRUNCATED) |
4d61f8c5 | 4393 | arg1 %= width; |
7afe21cc RK |
4394 | #endif |
4395 | ||
906c4e36 | 4396 | val = ((unsigned HOST_WIDE_INT) arg0) << arg1; |
7afe21cc RK |
4397 | break; |
4398 | ||
4399 | case ASHIFTRT: | |
4400 | if (arg1 < 0) | |
4401 | return 0; | |
4402 | ||
4403 | #ifdef SHIFT_COUNT_TRUNCATED | |
85c0a556 | 4404 | if (SHIFT_COUNT_TRUNCATED) |
4d61f8c5 | 4405 | arg1 %= width; |
7afe21cc RK |
4406 | #endif |
4407 | ||
7afe21cc | 4408 | val = arg0s >> arg1; |
2166571b RS |
4409 | |
4410 | /* Bootstrap compiler may not have sign extended the right shift. | |
4411 | Manually extend the sign to insure bootstrap cc matches gcc. */ | |
4412 | if (arg0s < 0 && arg1 > 0) | |
4413 | val |= ((HOST_WIDE_INT) -1) << (HOST_BITS_PER_WIDE_INT - arg1); | |
4414 | ||
7afe21cc RK |
4415 | break; |
4416 | ||
4417 | case ROTATERT: | |
4418 | if (arg1 < 0) | |
4419 | return 0; | |
4420 | ||
4421 | arg1 %= width; | |
906c4e36 RK |
4422 | val = ((((unsigned HOST_WIDE_INT) arg0) << (width - arg1)) |
4423 | | (((unsigned HOST_WIDE_INT) arg0) >> arg1)); | |
7afe21cc RK |
4424 | break; |
4425 | ||
4426 | case ROTATE: | |
4427 | if (arg1 < 0) | |
4428 | return 0; | |
4429 | ||
4430 | arg1 %= width; | |
906c4e36 RK |
4431 | val = ((((unsigned HOST_WIDE_INT) arg0) << arg1) |
4432 | | (((unsigned HOST_WIDE_INT) arg0) >> (width - arg1))); | |
7afe21cc RK |
4433 | break; |
4434 | ||
4435 | case COMPARE: | |
4436 | /* Do nothing here. */ | |
4437 | return 0; | |
4438 | ||
830a38ee RS |
4439 | case SMIN: |
4440 | val = arg0s <= arg1s ? arg0s : arg1s; | |
4441 | break; | |
4442 | ||
4443 | case UMIN: | |
906c4e36 RK |
4444 | val = ((unsigned HOST_WIDE_INT) arg0 |
4445 | <= (unsigned HOST_WIDE_INT) arg1 ? arg0 : arg1); | |
830a38ee RS |
4446 | break; |
4447 | ||
4448 | case SMAX: | |
4449 | val = arg0s > arg1s ? arg0s : arg1s; | |
4450 | break; | |
4451 | ||
4452 | case UMAX: | |
906c4e36 RK |
4453 | val = ((unsigned HOST_WIDE_INT) arg0 |
4454 | > (unsigned HOST_WIDE_INT) arg1 ? arg0 : arg1); | |
830a38ee RS |
4455 | break; |
4456 | ||
7afe21cc RK |
4457 | default: |
4458 | abort (); | |
4459 | } | |
4460 | ||
4461 | /* Clear the bits that don't belong in our mode, unless they and our sign | |
4462 | bit are all one. So we get either a reasonable negative value or a | |
4463 | reasonable unsigned value for this mode. */ | |
906c4e36 RK |
4464 | if (width < HOST_BITS_PER_WIDE_INT |
4465 | && ((val & ((HOST_WIDE_INT) (-1) << (width - 1))) | |
4466 | != ((HOST_WIDE_INT) (-1) << (width - 1)))) | |
4467 | val &= ((HOST_WIDE_INT) 1 << width) - 1; | |
4468 | ||
ad89d6f6 TG |
4469 | /* If this would be an entire word for the target, but is not for |
4470 | the host, then sign-extend on the host so that the number will look | |
4471 | the same way on the host that it would on the target. | |
4472 | ||
4473 | For example, when building a 64 bit alpha hosted 32 bit sparc | |
4474 | targeted compiler, then we want the 32 bit unsigned value -1 to be | |
4475 | represented as a 64 bit value -1, and not as 0x00000000ffffffff. | |
4476 | The later confuses the sparc backend. */ | |
4477 | ||
4478 | if (BITS_PER_WORD < HOST_BITS_PER_WIDE_INT && BITS_PER_WORD == width | |
4479 | && (val & ((HOST_WIDE_INT) 1 << (width - 1)))) | |
4480 | val |= ((HOST_WIDE_INT) (-1) << width); | |
4481 | ||
906c4e36 | 4482 | return GEN_INT (val); |
7afe21cc RK |
4483 | } |
4484 | \f | |
96b0e481 RK |
4485 | /* Simplify a PLUS or MINUS, at least one of whose operands may be another |
4486 | PLUS or MINUS. | |
4487 | ||
4488 | Rather than test for specific case, we do this by a brute-force method | |
4489 | and do all possible simplifications until no more changes occur. Then | |
4490 | we rebuild the operation. */ | |
4491 | ||
4492 | static rtx | |
4493 | simplify_plus_minus (code, mode, op0, op1) | |
4494 | enum rtx_code code; | |
4495 | enum machine_mode mode; | |
4496 | rtx op0, op1; | |
4497 | { | |
4498 | rtx ops[8]; | |
4499 | int negs[8]; | |
4500 | rtx result, tem; | |
fb5c8ce6 | 4501 | int n_ops = 2, input_ops = 2, input_consts = 0, n_consts = 0; |
96b0e481 | 4502 | int first = 1, negate = 0, changed; |
fb5c8ce6 | 4503 | int i, j; |
96b0e481 | 4504 | |
4c9a05bc | 4505 | bzero ((char *) ops, sizeof ops); |
96b0e481 RK |
4506 | |
4507 | /* Set up the two operands and then expand them until nothing has been | |
4508 | changed. If we run out of room in our array, give up; this should | |
4509 | almost never happen. */ | |
4510 | ||
4511 | ops[0] = op0, ops[1] = op1, negs[0] = 0, negs[1] = (code == MINUS); | |
4512 | ||
4513 | changed = 1; | |
4514 | while (changed) | |
4515 | { | |
4516 | changed = 0; | |
4517 | ||
4518 | for (i = 0; i < n_ops; i++) | |
4519 | switch (GET_CODE (ops[i])) | |
4520 | { | |
4521 | case PLUS: | |
4522 | case MINUS: | |
4523 | if (n_ops == 7) | |
4524 | return 0; | |
4525 | ||
4526 | ops[n_ops] = XEXP (ops[i], 1); | |
4527 | negs[n_ops++] = GET_CODE (ops[i]) == MINUS ? !negs[i] : negs[i]; | |
4528 | ops[i] = XEXP (ops[i], 0); | |
b7d9299b | 4529 | input_ops++; |
96b0e481 RK |
4530 | changed = 1; |
4531 | break; | |
4532 | ||
4533 | case NEG: | |
4534 | ops[i] = XEXP (ops[i], 0); | |
4535 | negs[i] = ! negs[i]; | |
4536 | changed = 1; | |
4537 | break; | |
4538 | ||
4539 | case CONST: | |
4540 | ops[i] = XEXP (ops[i], 0); | |
fb5c8ce6 | 4541 | input_consts++; |
96b0e481 RK |
4542 | changed = 1; |
4543 | break; | |
4544 | ||
4545 | case NOT: | |
4546 | /* ~a -> (-a - 1) */ | |
4547 | if (n_ops != 7) | |
4548 | { | |
4549 | ops[n_ops] = constm1_rtx; | |
5931019b | 4550 | negs[n_ops++] = negs[i]; |
96b0e481 RK |
4551 | ops[i] = XEXP (ops[i], 0); |
4552 | negs[i] = ! negs[i]; | |
4553 | changed = 1; | |
4554 | } | |
4555 | break; | |
4556 | ||
4557 | case CONST_INT: | |
4558 | if (negs[i]) | |
4559 | ops[i] = GEN_INT (- INTVAL (ops[i])), negs[i] = 0, changed = 1; | |
4560 | break; | |
e9a25f70 JL |
4561 | |
4562 | default: | |
4563 | break; | |
96b0e481 RK |
4564 | } |
4565 | } | |
4566 | ||
4567 | /* If we only have two operands, we can't do anything. */ | |
4568 | if (n_ops <= 2) | |
4569 | return 0; | |
4570 | ||
4571 | /* Now simplify each pair of operands until nothing changes. The first | |
4572 | time through just simplify constants against each other. */ | |
4573 | ||
4574 | changed = 1; | |
4575 | while (changed) | |
4576 | { | |
4577 | changed = first; | |
4578 | ||
4579 | for (i = 0; i < n_ops - 1; i++) | |
4580 | for (j = i + 1; j < n_ops; j++) | |
4581 | if (ops[i] != 0 && ops[j] != 0 | |
4582 | && (! first || (CONSTANT_P (ops[i]) && CONSTANT_P (ops[j])))) | |
4583 | { | |
4584 | rtx lhs = ops[i], rhs = ops[j]; | |
4585 | enum rtx_code ncode = PLUS; | |
4586 | ||
4587 | if (negs[i] && ! negs[j]) | |
4588 | lhs = ops[j], rhs = ops[i], ncode = MINUS; | |
4589 | else if (! negs[i] && negs[j]) | |
4590 | ncode = MINUS; | |
4591 | ||
4592 | tem = simplify_binary_operation (ncode, mode, lhs, rhs); | |
b7d9299b | 4593 | if (tem) |
96b0e481 RK |
4594 | { |
4595 | ops[i] = tem, ops[j] = 0; | |
4596 | negs[i] = negs[i] && negs[j]; | |
4597 | if (GET_CODE (tem) == NEG) | |
4598 | ops[i] = XEXP (tem, 0), negs[i] = ! negs[i]; | |
4599 | ||
4600 | if (GET_CODE (ops[i]) == CONST_INT && negs[i]) | |
4601 | ops[i] = GEN_INT (- INTVAL (ops[i])), negs[i] = 0; | |
4602 | changed = 1; | |
4603 | } | |
4604 | } | |
4605 | ||
4606 | first = 0; | |
4607 | } | |
4608 | ||
4609 | /* Pack all the operands to the lower-numbered entries and give up if | |
91a60f37 | 4610 | we didn't reduce the number of operands we had. Make sure we |
fb5c8ce6 RK |
4611 | count a CONST as two operands. If we have the same number of |
4612 | operands, but have made more CONSTs than we had, this is also | |
4613 | an improvement, so accept it. */ | |
91a60f37 | 4614 | |
fb5c8ce6 | 4615 | for (i = 0, j = 0; j < n_ops; j++) |
96b0e481 | 4616 | if (ops[j] != 0) |
91a60f37 RK |
4617 | { |
4618 | ops[i] = ops[j], negs[i++] = negs[j]; | |
4619 | if (GET_CODE (ops[j]) == CONST) | |
fb5c8ce6 | 4620 | n_consts++; |
91a60f37 | 4621 | } |
96b0e481 | 4622 | |
fb5c8ce6 RK |
4623 | if (i + n_consts > input_ops |
4624 | || (i + n_consts == input_ops && n_consts <= input_consts)) | |
96b0e481 RK |
4625 | return 0; |
4626 | ||
4627 | n_ops = i; | |
4628 | ||
4629 | /* If we have a CONST_INT, put it last. */ | |
4630 | for (i = 0; i < n_ops - 1; i++) | |
4631 | if (GET_CODE (ops[i]) == CONST_INT) | |
4632 | { | |
4633 | tem = ops[n_ops - 1], ops[n_ops - 1] = ops[i] , ops[i] = tem; | |
4634 | j = negs[n_ops - 1], negs[n_ops - 1] = negs[i], negs[i] = j; | |
4635 | } | |
4636 | ||
4637 | /* Put a non-negated operand first. If there aren't any, make all | |
4638 | operands positive and negate the whole thing later. */ | |
4639 | for (i = 0; i < n_ops && negs[i]; i++) | |
4640 | ; | |
4641 | ||
4642 | if (i == n_ops) | |
4643 | { | |
4644 | for (i = 0; i < n_ops; i++) | |
4645 | negs[i] = 0; | |
4646 | negate = 1; | |
4647 | } | |
4648 | else if (i != 0) | |
4649 | { | |
4650 | tem = ops[0], ops[0] = ops[i], ops[i] = tem; | |
4651 | j = negs[0], negs[0] = negs[i], negs[i] = j; | |
4652 | } | |
4653 | ||
4654 | /* Now make the result by performing the requested operations. */ | |
4655 | result = ops[0]; | |
4656 | for (i = 1; i < n_ops; i++) | |
4657 | result = cse_gen_binary (negs[i] ? MINUS : PLUS, mode, result, ops[i]); | |
4658 | ||
38a448ca | 4659 | return negate ? gen_rtx_NEG (mode, result) : result; |
96b0e481 RK |
4660 | } |
4661 | \f | |
4662 | /* Make a binary operation by properly ordering the operands and | |
4663 | seeing if the expression folds. */ | |
4664 | ||
4665 | static rtx | |
4666 | cse_gen_binary (code, mode, op0, op1) | |
4667 | enum rtx_code code; | |
4668 | enum machine_mode mode; | |
4669 | rtx op0, op1; | |
4670 | { | |
4671 | rtx tem; | |
4672 | ||
4673 | /* Put complex operands first and constants second if commutative. */ | |
4674 | if (GET_RTX_CLASS (code) == 'c' | |
4675 | && ((CONSTANT_P (op0) && GET_CODE (op1) != CONST_INT) | |
4676 | || (GET_RTX_CLASS (GET_CODE (op0)) == 'o' | |
4677 | && GET_RTX_CLASS (GET_CODE (op1)) != 'o') | |
4678 | || (GET_CODE (op0) == SUBREG | |
4679 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (op0))) == 'o' | |
4680 | && GET_RTX_CLASS (GET_CODE (op1)) != 'o'))) | |
4681 | tem = op0, op0 = op1, op1 = tem; | |
4682 | ||
4683 | /* If this simplifies, do it. */ | |
4684 | tem = simplify_binary_operation (code, mode, op0, op1); | |
4685 | ||
4686 | if (tem) | |
4687 | return tem; | |
4688 | ||
4689 | /* Handle addition and subtraction of CONST_INT specially. Otherwise, | |
4690 | just form the operation. */ | |
4691 | ||
4692 | if (code == PLUS && GET_CODE (op1) == CONST_INT | |
4693 | && GET_MODE (op0) != VOIDmode) | |
4694 | return plus_constant (op0, INTVAL (op1)); | |
4695 | else if (code == MINUS && GET_CODE (op1) == CONST_INT | |
4696 | && GET_MODE (op0) != VOIDmode) | |
4697 | return plus_constant (op0, - INTVAL (op1)); | |
4698 | else | |
38a448ca | 4699 | return gen_rtx_fmt_ee (code, mode, op0, op1); |
96b0e481 RK |
4700 | } |
4701 | \f | |
1a87eea2 KG |
4702 | struct cfc_args |
4703 | { | |
4704 | /* Input */ | |
4705 | rtx op0, op1; | |
4706 | /* Output */ | |
4707 | int equal, op0lt, op1lt; | |
4708 | }; | |
4709 | ||
4710 | static void | |
4711 | check_fold_consts (data) | |
4712 | PTR data; | |
4713 | { | |
4714 | struct cfc_args * args = (struct cfc_args *) data; | |
4715 | REAL_VALUE_TYPE d0, d1; | |
4716 | ||
4717 | REAL_VALUE_FROM_CONST_DOUBLE (d0, args->op0); | |
4718 | REAL_VALUE_FROM_CONST_DOUBLE (d1, args->op1); | |
4719 | args->equal = REAL_VALUES_EQUAL (d0, d1); | |
4720 | args->op0lt = REAL_VALUES_LESS (d0, d1); | |
4721 | args->op1lt = REAL_VALUES_LESS (d1, d0); | |
4722 | } | |
4723 | ||
7afe21cc | 4724 | /* Like simplify_binary_operation except used for relational operators. |
a432f20d RK |
4725 | MODE is the mode of the operands, not that of the result. If MODE |
4726 | is VOIDmode, both operands must also be VOIDmode and we compare the | |
4727 | operands in "infinite precision". | |
4728 | ||
4729 | If no simplification is possible, this function returns zero. Otherwise, | |
4730 | it returns either const_true_rtx or const0_rtx. */ | |
7afe21cc RK |
4731 | |
4732 | rtx | |
4733 | simplify_relational_operation (code, mode, op0, op1) | |
4734 | enum rtx_code code; | |
4735 | enum machine_mode mode; | |
4736 | rtx op0, op1; | |
4737 | { | |
a432f20d RK |
4738 | int equal, op0lt, op0ltu, op1lt, op1ltu; |
4739 | rtx tem; | |
7afe21cc RK |
4740 | |
4741 | /* If op0 is a compare, extract the comparison arguments from it. */ | |
4742 | if (GET_CODE (op0) == COMPARE && op1 == const0_rtx) | |
4743 | op1 = XEXP (op0, 1), op0 = XEXP (op0, 0); | |
4744 | ||
28bad1cb RK |
4745 | /* We can't simplify MODE_CC values since we don't know what the |
4746 | actual comparison is. */ | |
4747 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC | |
4748 | #ifdef HAVE_cc0 | |
4749 | || op0 == cc0_rtx | |
4750 | #endif | |
4751 | ) | |
31dcf83f RS |
4752 | return 0; |
4753 | ||
a432f20d RK |
4754 | /* For integer comparisons of A and B maybe we can simplify A - B and can |
4755 | then simplify a comparison of that with zero. If A and B are both either | |
4756 | a register or a CONST_INT, this can't help; testing for these cases will | |
4757 | prevent infinite recursion here and speed things up. | |
4758 | ||
c27b5c62 JW |
4759 | If CODE is an unsigned comparison, then we can never do this optimization, |
4760 | because it gives an incorrect result if the subtraction wraps around zero. | |
4761 | ANSI C defines unsigned operations such that they never overflow, and | |
4762 | thus such cases can not be ignored. */ | |
a432f20d RK |
4763 | |
4764 | if (INTEGRAL_MODE_P (mode) && op1 != const0_rtx | |
4765 | && ! ((GET_CODE (op0) == REG || GET_CODE (op0) == CONST_INT) | |
4766 | && (GET_CODE (op1) == REG || GET_CODE (op1) == CONST_INT)) | |
4767 | && 0 != (tem = simplify_binary_operation (MINUS, mode, op0, op1)) | |
c27b5c62 | 4768 | && code != GTU && code != GEU && code != LTU && code != LEU) |
a432f20d RK |
4769 | return simplify_relational_operation (signed_condition (code), |
4770 | mode, tem, const0_rtx); | |
4771 | ||
4772 | /* For non-IEEE floating-point, if the two operands are equal, we know the | |
4773 | result. */ | |
4774 | if (rtx_equal_p (op0, op1) | |
4775 | && (TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
4776 | || ! FLOAT_MODE_P (GET_MODE (op0)) || flag_fast_math)) | |
4777 | equal = 1, op0lt = 0, op0ltu = 0, op1lt = 0, op1ltu = 0; | |
4778 | ||
4779 | /* If the operands are floating-point constants, see if we can fold | |
4780 | the result. */ | |
6076248a | 4781 | #if ! defined (REAL_IS_NOT_DOUBLE) || defined (REAL_ARITHMETIC) |
a432f20d RK |
4782 | else if (GET_CODE (op0) == CONST_DOUBLE && GET_CODE (op1) == CONST_DOUBLE |
4783 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
4784 | { | |
1a87eea2 KG |
4785 | struct cfc_args args; |
4786 | ||
4787 | /* Setup input for check_fold_consts() */ | |
4788 | args.op0 = op0; | |
4789 | args.op1 = op1; | |
a432f20d | 4790 | |
1a87eea2 KG |
4791 | if (do_float_handler(check_fold_consts, (PTR) &args) == 0) |
4792 | /* We got an exception from check_fold_consts() */ | |
a432f20d | 4793 | return 0; |
7afe21cc | 4794 | |
1a87eea2 KG |
4795 | /* Receive output from check_fold_consts() */ |
4796 | equal = args.equal; | |
4797 | op0lt = op0ltu = args.op0lt; | |
4798 | op1lt = op1ltu = args.op1lt; | |
a432f20d RK |
4799 | } |
4800 | #endif /* not REAL_IS_NOT_DOUBLE, or REAL_ARITHMETIC */ | |
7afe21cc | 4801 | |
a432f20d RK |
4802 | /* Otherwise, see if the operands are both integers. */ |
4803 | else if ((GET_MODE_CLASS (mode) == MODE_INT || mode == VOIDmode) | |
4804 | && (GET_CODE (op0) == CONST_DOUBLE || GET_CODE (op0) == CONST_INT) | |
4805 | && (GET_CODE (op1) == CONST_DOUBLE || GET_CODE (op1) == CONST_INT)) | |
4806 | { | |
4807 | int width = GET_MODE_BITSIZE (mode); | |
64812ded RK |
4808 | HOST_WIDE_INT l0s, h0s, l1s, h1s; |
4809 | unsigned HOST_WIDE_INT l0u, h0u, l1u, h1u; | |
7afe21cc | 4810 | |
a432f20d RK |
4811 | /* Get the two words comprising each integer constant. */ |
4812 | if (GET_CODE (op0) == CONST_DOUBLE) | |
4813 | { | |
4814 | l0u = l0s = CONST_DOUBLE_LOW (op0); | |
4815 | h0u = h0s = CONST_DOUBLE_HIGH (op0); | |
7afe21cc | 4816 | } |
a432f20d | 4817 | else |
6076248a | 4818 | { |
a432f20d | 4819 | l0u = l0s = INTVAL (op0); |
cb3bb2a7 | 4820 | h0u = h0s = l0s < 0 ? -1 : 0; |
a432f20d | 4821 | } |
6076248a | 4822 | |
a432f20d RK |
4823 | if (GET_CODE (op1) == CONST_DOUBLE) |
4824 | { | |
4825 | l1u = l1s = CONST_DOUBLE_LOW (op1); | |
4826 | h1u = h1s = CONST_DOUBLE_HIGH (op1); | |
4827 | } | |
4828 | else | |
4829 | { | |
4830 | l1u = l1s = INTVAL (op1); | |
cb3bb2a7 | 4831 | h1u = h1s = l1s < 0 ? -1 : 0; |
a432f20d RK |
4832 | } |
4833 | ||
4834 | /* If WIDTH is nonzero and smaller than HOST_BITS_PER_WIDE_INT, | |
4835 | we have to sign or zero-extend the values. */ | |
4836 | if (width != 0 && width <= HOST_BITS_PER_WIDE_INT) | |
4837 | h0u = h1u = 0, h0s = l0s < 0 ? -1 : 0, h1s = l1s < 0 ? -1 : 0; | |
6076248a | 4838 | |
a432f20d RK |
4839 | if (width != 0 && width < HOST_BITS_PER_WIDE_INT) |
4840 | { | |
4841 | l0u &= ((HOST_WIDE_INT) 1 << width) - 1; | |
4842 | l1u &= ((HOST_WIDE_INT) 1 << width) - 1; | |
6076248a | 4843 | |
a432f20d RK |
4844 | if (l0s & ((HOST_WIDE_INT) 1 << (width - 1))) |
4845 | l0s |= ((HOST_WIDE_INT) (-1) << width); | |
6076248a | 4846 | |
a432f20d RK |
4847 | if (l1s & ((HOST_WIDE_INT) 1 << (width - 1))) |
4848 | l1s |= ((HOST_WIDE_INT) (-1) << width); | |
6076248a RK |
4849 | } |
4850 | ||
a432f20d RK |
4851 | equal = (h0u == h1u && l0u == l1u); |
4852 | op0lt = (h0s < h1s || (h0s == h1s && l0s < l1s)); | |
4853 | op1lt = (h1s < h0s || (h1s == h0s && l1s < l0s)); | |
4854 | op0ltu = (h0u < h1u || (h0u == h1u && l0u < l1u)); | |
4855 | op1ltu = (h1u < h0u || (h1u == h0u && l1u < l0u)); | |
4856 | } | |
4857 | ||
4858 | /* Otherwise, there are some code-specific tests we can make. */ | |
4859 | else | |
4860 | { | |
7afe21cc RK |
4861 | switch (code) |
4862 | { | |
4863 | case EQ: | |
a432f20d RK |
4864 | /* References to the frame plus a constant or labels cannot |
4865 | be zero, but a SYMBOL_REF can due to #pragma weak. */ | |
4866 | if (((NONZERO_BASE_PLUS_P (op0) && op1 == const0_rtx) | |
4867 | || GET_CODE (op0) == LABEL_REF) | |
1a7c818b | 4868 | #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM |
a432f20d RK |
4869 | /* On some machines, the ap reg can be 0 sometimes. */ |
4870 | && op0 != arg_pointer_rtx | |
7afe21cc | 4871 | #endif |
a432f20d RK |
4872 | ) |
4873 | return const0_rtx; | |
4874 | break; | |
7afe21cc RK |
4875 | |
4876 | case NE: | |
a432f20d RK |
4877 | if (((NONZERO_BASE_PLUS_P (op0) && op1 == const0_rtx) |
4878 | || GET_CODE (op0) == LABEL_REF) | |
1a7c818b | 4879 | #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM |
a432f20d | 4880 | && op0 != arg_pointer_rtx |
7afe21cc | 4881 | #endif |
a432f20d | 4882 | ) |
7afe21cc RK |
4883 | return const_true_rtx; |
4884 | break; | |
4885 | ||
4886 | case GEU: | |
a432f20d RK |
4887 | /* Unsigned values are never negative. */ |
4888 | if (op1 == const0_rtx) | |
7afe21cc RK |
4889 | return const_true_rtx; |
4890 | break; | |
4891 | ||
4892 | case LTU: | |
a432f20d | 4893 | if (op1 == const0_rtx) |
7afe21cc RK |
4894 | return const0_rtx; |
4895 | break; | |
4896 | ||
4897 | case LEU: | |
4898 | /* Unsigned values are never greater than the largest | |
4899 | unsigned value. */ | |
4900 | if (GET_CODE (op1) == CONST_INT | |
4901 | && INTVAL (op1) == GET_MODE_MASK (mode) | |
a432f20d RK |
4902 | && INTEGRAL_MODE_P (mode)) |
4903 | return const_true_rtx; | |
7afe21cc RK |
4904 | break; |
4905 | ||
4906 | case GTU: | |
4907 | if (GET_CODE (op1) == CONST_INT | |
4908 | && INTVAL (op1) == GET_MODE_MASK (mode) | |
cbf6a543 | 4909 | && INTEGRAL_MODE_P (mode)) |
7afe21cc RK |
4910 | return const0_rtx; |
4911 | break; | |
e9a25f70 JL |
4912 | |
4913 | default: | |
4914 | break; | |
7afe21cc RK |
4915 | } |
4916 | ||
4917 | return 0; | |
4918 | } | |
4919 | ||
a432f20d RK |
4920 | /* If we reach here, EQUAL, OP0LT, OP0LTU, OP1LT, and OP1LTU are set |
4921 | as appropriate. */ | |
7afe21cc RK |
4922 | switch (code) |
4923 | { | |
7afe21cc | 4924 | case EQ: |
a432f20d RK |
4925 | return equal ? const_true_rtx : const0_rtx; |
4926 | case NE: | |
4927 | return ! equal ? const_true_rtx : const0_rtx; | |
7afe21cc | 4928 | case LT: |
a432f20d | 4929 | return op0lt ? const_true_rtx : const0_rtx; |
7afe21cc | 4930 | case GT: |
a432f20d | 4931 | return op1lt ? const_true_rtx : const0_rtx; |
7afe21cc | 4932 | case LTU: |
a432f20d | 4933 | return op0ltu ? const_true_rtx : const0_rtx; |
7afe21cc | 4934 | case GTU: |
a432f20d RK |
4935 | return op1ltu ? const_true_rtx : const0_rtx; |
4936 | case LE: | |
4937 | return equal || op0lt ? const_true_rtx : const0_rtx; | |
4938 | case GE: | |
4939 | return equal || op1lt ? const_true_rtx : const0_rtx; | |
4940 | case LEU: | |
4941 | return equal || op0ltu ? const_true_rtx : const0_rtx; | |
4942 | case GEU: | |
4943 | return equal || op1ltu ? const_true_rtx : const0_rtx; | |
e9a25f70 JL |
4944 | default: |
4945 | abort (); | |
7afe21cc | 4946 | } |
7afe21cc RK |
4947 | } |
4948 | \f | |
4949 | /* Simplify CODE, an operation with result mode MODE and three operands, | |
4950 | OP0, OP1, and OP2. OP0_MODE was the mode of OP0 before it became | |
4951 | a constant. Return 0 if no simplifications is possible. */ | |
4952 | ||
4953 | rtx | |
4954 | simplify_ternary_operation (code, mode, op0_mode, op0, op1, op2) | |
4955 | enum rtx_code code; | |
4956 | enum machine_mode mode, op0_mode; | |
4957 | rtx op0, op1, op2; | |
4958 | { | |
4959 | int width = GET_MODE_BITSIZE (mode); | |
4960 | ||
4961 | /* VOIDmode means "infinite" precision. */ | |
4962 | if (width == 0) | |
906c4e36 | 4963 | width = HOST_BITS_PER_WIDE_INT; |
7afe21cc RK |
4964 | |
4965 | switch (code) | |
4966 | { | |
4967 | case SIGN_EXTRACT: | |
4968 | case ZERO_EXTRACT: | |
4969 | if (GET_CODE (op0) == CONST_INT | |
4970 | && GET_CODE (op1) == CONST_INT | |
4971 | && GET_CODE (op2) == CONST_INT | |
4972 | && INTVAL (op1) + INTVAL (op2) <= GET_MODE_BITSIZE (op0_mode) | |
906c4e36 | 4973 | && width <= HOST_BITS_PER_WIDE_INT) |
7afe21cc RK |
4974 | { |
4975 | /* Extracting a bit-field from a constant */ | |
906c4e36 | 4976 | HOST_WIDE_INT val = INTVAL (op0); |
7afe21cc | 4977 | |
f76b9db2 ILT |
4978 | if (BITS_BIG_ENDIAN) |
4979 | val >>= (GET_MODE_BITSIZE (op0_mode) | |
4980 | - INTVAL (op2) - INTVAL (op1)); | |
4981 | else | |
4982 | val >>= INTVAL (op2); | |
4983 | ||
906c4e36 | 4984 | if (HOST_BITS_PER_WIDE_INT != INTVAL (op1)) |
7afe21cc RK |
4985 | { |
4986 | /* First zero-extend. */ | |
906c4e36 | 4987 | val &= ((HOST_WIDE_INT) 1 << INTVAL (op1)) - 1; |
7afe21cc | 4988 | /* If desired, propagate sign bit. */ |
906c4e36 RK |
4989 | if (code == SIGN_EXTRACT |
4990 | && (val & ((HOST_WIDE_INT) 1 << (INTVAL (op1) - 1)))) | |
4991 | val |= ~ (((HOST_WIDE_INT) 1 << INTVAL (op1)) - 1); | |
7afe21cc RK |
4992 | } |
4993 | ||
4994 | /* Clear the bits that don't belong in our mode, | |
4995 | unless they and our sign bit are all one. | |
4996 | So we get either a reasonable negative value or a reasonable | |
4997 | unsigned value for this mode. */ | |
906c4e36 RK |
4998 | if (width < HOST_BITS_PER_WIDE_INT |
4999 | && ((val & ((HOST_WIDE_INT) (-1) << (width - 1))) | |
5000 | != ((HOST_WIDE_INT) (-1) << (width - 1)))) | |
5001 | val &= ((HOST_WIDE_INT) 1 << width) - 1; | |
7afe21cc | 5002 | |
906c4e36 | 5003 | return GEN_INT (val); |
7afe21cc RK |
5004 | } |
5005 | break; | |
5006 | ||
5007 | case IF_THEN_ELSE: | |
5008 | if (GET_CODE (op0) == CONST_INT) | |
5009 | return op0 != const0_rtx ? op1 : op2; | |
3bf1b082 JW |
5010 | |
5011 | /* Convert a == b ? b : a to "a". */ | |
5012 | if (GET_CODE (op0) == NE && ! side_effects_p (op0) | |
5013 | && rtx_equal_p (XEXP (op0, 0), op1) | |
5014 | && rtx_equal_p (XEXP (op0, 1), op2)) | |
5015 | return op1; | |
5016 | else if (GET_CODE (op0) == EQ && ! side_effects_p (op0) | |
5017 | && rtx_equal_p (XEXP (op0, 1), op1) | |
5018 | && rtx_equal_p (XEXP (op0, 0), op2)) | |
5019 | return op2; | |
e82ad93d | 5020 | else if (GET_RTX_CLASS (GET_CODE (op0)) == '<' && ! side_effects_p (op0)) |
ed1ecb19 JL |
5021 | { |
5022 | rtx temp; | |
5023 | temp = simplify_relational_operation (GET_CODE (op0), op0_mode, | |
5024 | XEXP (op0, 0), XEXP (op0, 1)); | |
5025 | /* See if any simplifications were possible. */ | |
5026 | if (temp == const0_rtx) | |
5027 | return op2; | |
5028 | else if (temp == const1_rtx) | |
5029 | return op1; | |
5030 | } | |
7afe21cc RK |
5031 | break; |
5032 | ||
5033 | default: | |
5034 | abort (); | |
5035 | } | |
5036 | ||
5037 | return 0; | |
5038 | } | |
5039 | \f | |
5040 | /* If X is a nontrivial arithmetic operation on an argument | |
5041 | for which a constant value can be determined, return | |
5042 | the result of operating on that value, as a constant. | |
5043 | Otherwise, return X, possibly with one or more operands | |
5044 | modified by recursive calls to this function. | |
5045 | ||
e7bb59fa RK |
5046 | If X is a register whose contents are known, we do NOT |
5047 | return those contents here. equiv_constant is called to | |
5048 | perform that task. | |
7afe21cc RK |
5049 | |
5050 | INSN is the insn that we may be modifying. If it is 0, make a copy | |
5051 | of X before modifying it. */ | |
5052 | ||
5053 | static rtx | |
5054 | fold_rtx (x, insn) | |
5055 | rtx x; | |
5056 | rtx insn; | |
5057 | { | |
5058 | register enum rtx_code code; | |
5059 | register enum machine_mode mode; | |
5060 | register char *fmt; | |
906c4e36 | 5061 | register int i; |
7afe21cc RK |
5062 | rtx new = 0; |
5063 | int copied = 0; | |
5064 | int must_swap = 0; | |
5065 | ||
5066 | /* Folded equivalents of first two operands of X. */ | |
5067 | rtx folded_arg0; | |
5068 | rtx folded_arg1; | |
5069 | ||
5070 | /* Constant equivalents of first three operands of X; | |
5071 | 0 when no such equivalent is known. */ | |
5072 | rtx const_arg0; | |
5073 | rtx const_arg1; | |
5074 | rtx const_arg2; | |
5075 | ||
5076 | /* The mode of the first operand of X. We need this for sign and zero | |
5077 | extends. */ | |
5078 | enum machine_mode mode_arg0; | |
5079 | ||
5080 | if (x == 0) | |
5081 | return x; | |
5082 | ||
5083 | mode = GET_MODE (x); | |
5084 | code = GET_CODE (x); | |
5085 | switch (code) | |
5086 | { | |
5087 | case CONST: | |
5088 | case CONST_INT: | |
5089 | case CONST_DOUBLE: | |
5090 | case SYMBOL_REF: | |
5091 | case LABEL_REF: | |
5092 | case REG: | |
5093 | /* No use simplifying an EXPR_LIST | |
5094 | since they are used only for lists of args | |
5095 | in a function call's REG_EQUAL note. */ | |
5096 | case EXPR_LIST: | |
956d6950 JL |
5097 | /* Changing anything inside an ADDRESSOF is incorrect; we don't |
5098 | want to (e.g.,) make (addressof (const_int 0)) just because | |
5099 | the location is known to be zero. */ | |
5100 | case ADDRESSOF: | |
7afe21cc RK |
5101 | return x; |
5102 | ||
5103 | #ifdef HAVE_cc0 | |
5104 | case CC0: | |
5105 | return prev_insn_cc0; | |
5106 | #endif | |
5107 | ||
5108 | case PC: | |
5109 | /* If the next insn is a CODE_LABEL followed by a jump table, | |
5110 | PC's value is a LABEL_REF pointing to that label. That | |
5111 | lets us fold switch statements on the Vax. */ | |
5112 | if (insn && GET_CODE (insn) == JUMP_INSN) | |
5113 | { | |
5114 | rtx next = next_nonnote_insn (insn); | |
5115 | ||
5116 | if (next && GET_CODE (next) == CODE_LABEL | |
5117 | && NEXT_INSN (next) != 0 | |
5118 | && GET_CODE (NEXT_INSN (next)) == JUMP_INSN | |
5119 | && (GET_CODE (PATTERN (NEXT_INSN (next))) == ADDR_VEC | |
5120 | || GET_CODE (PATTERN (NEXT_INSN (next))) == ADDR_DIFF_VEC)) | |
38a448ca | 5121 | return gen_rtx_LABEL_REF (Pmode, next); |
7afe21cc RK |
5122 | } |
5123 | break; | |
5124 | ||
5125 | case SUBREG: | |
c610adec RK |
5126 | /* See if we previously assigned a constant value to this SUBREG. */ |
5127 | if ((new = lookup_as_function (x, CONST_INT)) != 0 | |
5128 | || (new = lookup_as_function (x, CONST_DOUBLE)) != 0) | |
7afe21cc RK |
5129 | return new; |
5130 | ||
4b980e20 RK |
5131 | /* If this is a paradoxical SUBREG, we have no idea what value the |
5132 | extra bits would have. However, if the operand is equivalent | |
5133 | to a SUBREG whose operand is the same as our mode, and all the | |
5134 | modes are within a word, we can just use the inner operand | |
31c85c78 RK |
5135 | because these SUBREGs just say how to treat the register. |
5136 | ||
5137 | Similarly if we find an integer constant. */ | |
4b980e20 | 5138 | |
e5f6a288 | 5139 | if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) |
4b980e20 RK |
5140 | { |
5141 | enum machine_mode imode = GET_MODE (SUBREG_REG (x)); | |
5142 | struct table_elt *elt; | |
5143 | ||
5144 | if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD | |
5145 | && GET_MODE_SIZE (imode) <= UNITS_PER_WORD | |
5146 | && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode), | |
5147 | imode)) != 0) | |
31c85c78 RK |
5148 | for (elt = elt->first_same_value; |
5149 | elt; elt = elt->next_same_value) | |
5150 | { | |
5151 | if (CONSTANT_P (elt->exp) | |
5152 | && GET_MODE (elt->exp) == VOIDmode) | |
5153 | return elt->exp; | |
5154 | ||
4b980e20 RK |
5155 | if (GET_CODE (elt->exp) == SUBREG |
5156 | && GET_MODE (SUBREG_REG (elt->exp)) == mode | |
906c4e36 | 5157 | && exp_equiv_p (elt->exp, elt->exp, 1, 0)) |
4b980e20 RK |
5158 | return copy_rtx (SUBREG_REG (elt->exp)); |
5159 | } | |
5160 | ||
5161 | return x; | |
5162 | } | |
e5f6a288 | 5163 | |
7afe21cc RK |
5164 | /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG. |
5165 | We might be able to if the SUBREG is extracting a single word in an | |
5166 | integral mode or extracting the low part. */ | |
5167 | ||
5168 | folded_arg0 = fold_rtx (SUBREG_REG (x), insn); | |
5169 | const_arg0 = equiv_constant (folded_arg0); | |
5170 | if (const_arg0) | |
5171 | folded_arg0 = const_arg0; | |
5172 | ||
5173 | if (folded_arg0 != SUBREG_REG (x)) | |
5174 | { | |
5175 | new = 0; | |
5176 | ||
5177 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5178 | && GET_MODE_SIZE (mode) == UNITS_PER_WORD | |
5179 | && GET_MODE (SUBREG_REG (x)) != VOIDmode) | |
5180 | new = operand_subword (folded_arg0, SUBREG_WORD (x), 0, | |
5181 | GET_MODE (SUBREG_REG (x))); | |
5182 | if (new == 0 && subreg_lowpart_p (x)) | |
5183 | new = gen_lowpart_if_possible (mode, folded_arg0); | |
5184 | if (new) | |
5185 | return new; | |
5186 | } | |
e5f6a288 RK |
5187 | |
5188 | /* If this is a narrowing SUBREG and our operand is a REG, see if | |
858a47b1 | 5189 | we can find an equivalence for REG that is an arithmetic operation |
e5f6a288 RK |
5190 | in a wider mode where both operands are paradoxical SUBREGs |
5191 | from objects of our result mode. In that case, we couldn't report | |
5192 | an equivalent value for that operation, since we don't know what the | |
5193 | extra bits will be. But we can find an equivalence for this SUBREG | |
5194 | by folding that operation is the narrow mode. This allows us to | |
5195 | fold arithmetic in narrow modes when the machine only supports | |
4b980e20 RK |
5196 | word-sized arithmetic. |
5197 | ||
5198 | Also look for a case where we have a SUBREG whose operand is the | |
5199 | same as our result. If both modes are smaller than a word, we | |
5200 | are simply interpreting a register in different modes and we | |
5201 | can use the inner value. */ | |
e5f6a288 RK |
5202 | |
5203 | if (GET_CODE (folded_arg0) == REG | |
e8d76a39 RS |
5204 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)) |
5205 | && subreg_lowpart_p (x)) | |
e5f6a288 RK |
5206 | { |
5207 | struct table_elt *elt; | |
5208 | ||
5209 | /* We can use HASH here since we know that canon_hash won't be | |
5210 | called. */ | |
5211 | elt = lookup (folded_arg0, | |
5212 | HASH (folded_arg0, GET_MODE (folded_arg0)), | |
5213 | GET_MODE (folded_arg0)); | |
5214 | ||
5215 | if (elt) | |
5216 | elt = elt->first_same_value; | |
5217 | ||
5218 | for (; elt; elt = elt->next_same_value) | |
5219 | { | |
e8d76a39 RS |
5220 | enum rtx_code eltcode = GET_CODE (elt->exp); |
5221 | ||
e5f6a288 RK |
5222 | /* Just check for unary and binary operations. */ |
5223 | if (GET_RTX_CLASS (GET_CODE (elt->exp)) == '1' | |
5224 | && GET_CODE (elt->exp) != SIGN_EXTEND | |
5225 | && GET_CODE (elt->exp) != ZERO_EXTEND | |
5226 | && GET_CODE (XEXP (elt->exp, 0)) == SUBREG | |
5227 | && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode) | |
5228 | { | |
5229 | rtx op0 = SUBREG_REG (XEXP (elt->exp, 0)); | |
5230 | ||
5231 | if (GET_CODE (op0) != REG && ! CONSTANT_P (op0)) | |
906c4e36 | 5232 | op0 = fold_rtx (op0, NULL_RTX); |
e5f6a288 RK |
5233 | |
5234 | op0 = equiv_constant (op0); | |
5235 | if (op0) | |
5236 | new = simplify_unary_operation (GET_CODE (elt->exp), mode, | |
5237 | op0, mode); | |
5238 | } | |
5239 | else if ((GET_RTX_CLASS (GET_CODE (elt->exp)) == '2' | |
5240 | || GET_RTX_CLASS (GET_CODE (elt->exp)) == 'c') | |
e8d76a39 RS |
5241 | && eltcode != DIV && eltcode != MOD |
5242 | && eltcode != UDIV && eltcode != UMOD | |
5243 | && eltcode != ASHIFTRT && eltcode != LSHIFTRT | |
5244 | && eltcode != ROTATE && eltcode != ROTATERT | |
e5f6a288 RK |
5245 | && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG |
5246 | && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) | |
5247 | == mode)) | |
5248 | || CONSTANT_P (XEXP (elt->exp, 0))) | |
5249 | && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG | |
5250 | && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1))) | |
5251 | == mode)) | |
5252 | || CONSTANT_P (XEXP (elt->exp, 1)))) | |
5253 | { | |
5254 | rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0)); | |
5255 | rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1)); | |
5256 | ||
5257 | if (op0 && GET_CODE (op0) != REG && ! CONSTANT_P (op0)) | |
906c4e36 | 5258 | op0 = fold_rtx (op0, NULL_RTX); |
e5f6a288 RK |
5259 | |
5260 | if (op0) | |
5261 | op0 = equiv_constant (op0); | |
5262 | ||
5263 | if (op1 && GET_CODE (op1) != REG && ! CONSTANT_P (op1)) | |
906c4e36 | 5264 | op1 = fold_rtx (op1, NULL_RTX); |
e5f6a288 RK |
5265 | |
5266 | if (op1) | |
5267 | op1 = equiv_constant (op1); | |
5268 | ||
76fb0b60 RS |
5269 | /* If we are looking for the low SImode part of |
5270 | (ashift:DI c (const_int 32)), it doesn't work | |
5271 | to compute that in SImode, because a 32-bit shift | |
5272 | in SImode is unpredictable. We know the value is 0. */ | |
5273 | if (op0 && op1 | |
45620ed4 | 5274 | && GET_CODE (elt->exp) == ASHIFT |
76fb0b60 RS |
5275 | && GET_CODE (op1) == CONST_INT |
5276 | && INTVAL (op1) >= GET_MODE_BITSIZE (mode)) | |
5277 | { | |
5278 | if (INTVAL (op1) < GET_MODE_BITSIZE (GET_MODE (elt->exp))) | |
5279 | ||
5280 | /* If the count fits in the inner mode's width, | |
5281 | but exceeds the outer mode's width, | |
5282 | the value will get truncated to 0 | |
5283 | by the subreg. */ | |
5284 | new = const0_rtx; | |
5285 | else | |
5286 | /* If the count exceeds even the inner mode's width, | |
5287 | don't fold this expression. */ | |
5288 | new = 0; | |
5289 | } | |
5290 | else if (op0 && op1) | |
e5f6a288 RK |
5291 | new = simplify_binary_operation (GET_CODE (elt->exp), mode, |
5292 | op0, op1); | |
5293 | } | |
5294 | ||
4b980e20 RK |
5295 | else if (GET_CODE (elt->exp) == SUBREG |
5296 | && GET_MODE (SUBREG_REG (elt->exp)) == mode | |
5297 | && (GET_MODE_SIZE (GET_MODE (folded_arg0)) | |
5298 | <= UNITS_PER_WORD) | |
906c4e36 | 5299 | && exp_equiv_p (elt->exp, elt->exp, 1, 0)) |
4b980e20 RK |
5300 | new = copy_rtx (SUBREG_REG (elt->exp)); |
5301 | ||
e5f6a288 RK |
5302 | if (new) |
5303 | return new; | |
5304 | } | |
5305 | } | |
5306 | ||
7afe21cc RK |
5307 | return x; |
5308 | ||
5309 | case NOT: | |
5310 | case NEG: | |
5311 | /* If we have (NOT Y), see if Y is known to be (NOT Z). | |
5312 | If so, (NOT Y) simplifies to Z. Similarly for NEG. */ | |
5313 | new = lookup_as_function (XEXP (x, 0), code); | |
5314 | if (new) | |
5315 | return fold_rtx (copy_rtx (XEXP (new, 0)), insn); | |
5316 | break; | |
13c9910f | 5317 | |
7afe21cc RK |
5318 | case MEM: |
5319 | /* If we are not actually processing an insn, don't try to find the | |
5320 | best address. Not only don't we care, but we could modify the | |
5321 | MEM in an invalid way since we have no insn to validate against. */ | |
5322 | if (insn != 0) | |
5323 | find_best_addr (insn, &XEXP (x, 0)); | |
5324 | ||
5325 | { | |
5326 | /* Even if we don't fold in the insn itself, | |
5327 | we can safely do so here, in hopes of getting a constant. */ | |
906c4e36 | 5328 | rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX); |
7afe21cc | 5329 | rtx base = 0; |
906c4e36 | 5330 | HOST_WIDE_INT offset = 0; |
7afe21cc RK |
5331 | |
5332 | if (GET_CODE (addr) == REG | |
5333 | && REGNO_QTY_VALID_P (REGNO (addr)) | |
30f72379 MM |
5334 | && GET_MODE (addr) == qty_mode[REG_QTY (REGNO (addr))] |
5335 | && qty_const[REG_QTY (REGNO (addr))] != 0) | |
5336 | addr = qty_const[REG_QTY (REGNO (addr))]; | |
7afe21cc RK |
5337 | |
5338 | /* If address is constant, split it into a base and integer offset. */ | |
5339 | if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF) | |
5340 | base = addr; | |
5341 | else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS | |
5342 | && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT) | |
5343 | { | |
5344 | base = XEXP (XEXP (addr, 0), 0); | |
5345 | offset = INTVAL (XEXP (XEXP (addr, 0), 1)); | |
5346 | } | |
5347 | else if (GET_CODE (addr) == LO_SUM | |
5348 | && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF) | |
5349 | base = XEXP (addr, 1); | |
e9a25f70 | 5350 | else if (GET_CODE (addr) == ADDRESSOF) |
956d6950 | 5351 | return change_address (x, VOIDmode, addr); |
7afe21cc RK |
5352 | |
5353 | /* If this is a constant pool reference, we can fold it into its | |
5354 | constant to allow better value tracking. */ | |
5355 | if (base && GET_CODE (base) == SYMBOL_REF | |
5356 | && CONSTANT_POOL_ADDRESS_P (base)) | |
5357 | { | |
5358 | rtx constant = get_pool_constant (base); | |
5359 | enum machine_mode const_mode = get_pool_mode (base); | |
5360 | rtx new; | |
5361 | ||
5362 | if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT) | |
5363 | constant_pool_entries_cost = COST (constant); | |
5364 | ||
5365 | /* If we are loading the full constant, we have an equivalence. */ | |
5366 | if (offset == 0 && mode == const_mode) | |
5367 | return constant; | |
5368 | ||
9faa82d8 | 5369 | /* If this actually isn't a constant (weird!), we can't do |
7afe21cc RK |
5370 | anything. Otherwise, handle the two most common cases: |
5371 | extracting a word from a multi-word constant, and extracting | |
5372 | the low-order bits. Other cases don't seem common enough to | |
5373 | worry about. */ | |
5374 | if (! CONSTANT_P (constant)) | |
5375 | return x; | |
5376 | ||
5377 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5378 | && GET_MODE_SIZE (mode) == UNITS_PER_WORD | |
5379 | && offset % UNITS_PER_WORD == 0 | |
5380 | && (new = operand_subword (constant, | |
5381 | offset / UNITS_PER_WORD, | |
5382 | 0, const_mode)) != 0) | |
5383 | return new; | |
5384 | ||
5385 | if (((BYTES_BIG_ENDIAN | |
5386 | && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1) | |
5387 | || (! BYTES_BIG_ENDIAN && offset == 0)) | |
5388 | && (new = gen_lowpart_if_possible (mode, constant)) != 0) | |
5389 | return new; | |
5390 | } | |
5391 | ||
5392 | /* If this is a reference to a label at a known position in a jump | |
5393 | table, we also know its value. */ | |
5394 | if (base && GET_CODE (base) == LABEL_REF) | |
5395 | { | |
5396 | rtx label = XEXP (base, 0); | |
5397 | rtx table_insn = NEXT_INSN (label); | |
5398 | ||
5399 | if (table_insn && GET_CODE (table_insn) == JUMP_INSN | |
5400 | && GET_CODE (PATTERN (table_insn)) == ADDR_VEC) | |
5401 | { | |
5402 | rtx table = PATTERN (table_insn); | |
5403 | ||
5404 | if (offset >= 0 | |
5405 | && (offset / GET_MODE_SIZE (GET_MODE (table)) | |
5406 | < XVECLEN (table, 0))) | |
5407 | return XVECEXP (table, 0, | |
5408 | offset / GET_MODE_SIZE (GET_MODE (table))); | |
5409 | } | |
5410 | if (table_insn && GET_CODE (table_insn) == JUMP_INSN | |
5411 | && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC) | |
5412 | { | |
5413 | rtx table = PATTERN (table_insn); | |
5414 | ||
5415 | if (offset >= 0 | |
5416 | && (offset / GET_MODE_SIZE (GET_MODE (table)) | |
5417 | < XVECLEN (table, 1))) | |
5418 | { | |
5419 | offset /= GET_MODE_SIZE (GET_MODE (table)); | |
38a448ca RH |
5420 | new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset), |
5421 | XEXP (table, 0)); | |
7afe21cc RK |
5422 | |
5423 | if (GET_MODE (table) != Pmode) | |
38a448ca | 5424 | new = gen_rtx_TRUNCATE (GET_MODE (table), new); |
7afe21cc | 5425 | |
67a37737 RK |
5426 | /* Indicate this is a constant. This isn't a |
5427 | valid form of CONST, but it will only be used | |
5428 | to fold the next insns and then discarded, so | |
ac7ef8d5 FS |
5429 | it should be safe. |
5430 | ||
5431 | Note this expression must be explicitly discarded, | |
5432 | by cse_insn, else it may end up in a REG_EQUAL note | |
5433 | and "escape" to cause problems elsewhere. */ | |
38a448ca | 5434 | return gen_rtx_CONST (GET_MODE (new), new); |
7afe21cc RK |
5435 | } |
5436 | } | |
5437 | } | |
5438 | ||
5439 | return x; | |
5440 | } | |
9255709c RK |
5441 | |
5442 | case ASM_OPERANDS: | |
5443 | for (i = XVECLEN (x, 3) - 1; i >= 0; i--) | |
5444 | validate_change (insn, &XVECEXP (x, 3, i), | |
5445 | fold_rtx (XVECEXP (x, 3, i), insn), 0); | |
5446 | break; | |
e9a25f70 JL |
5447 | |
5448 | default: | |
5449 | break; | |
7afe21cc RK |
5450 | } |
5451 | ||
5452 | const_arg0 = 0; | |
5453 | const_arg1 = 0; | |
5454 | const_arg2 = 0; | |
5455 | mode_arg0 = VOIDmode; | |
5456 | ||
5457 | /* Try folding our operands. | |
5458 | Then see which ones have constant values known. */ | |
5459 | ||
5460 | fmt = GET_RTX_FORMAT (code); | |
5461 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
5462 | if (fmt[i] == 'e') | |
5463 | { | |
5464 | rtx arg = XEXP (x, i); | |
5465 | rtx folded_arg = arg, const_arg = 0; | |
5466 | enum machine_mode mode_arg = GET_MODE (arg); | |
5467 | rtx cheap_arg, expensive_arg; | |
5468 | rtx replacements[2]; | |
5469 | int j; | |
5470 | ||
5471 | /* Most arguments are cheap, so handle them specially. */ | |
5472 | switch (GET_CODE (arg)) | |
5473 | { | |
5474 | case REG: | |
5475 | /* This is the same as calling equiv_constant; it is duplicated | |
5476 | here for speed. */ | |
5477 | if (REGNO_QTY_VALID_P (REGNO (arg)) | |
30f72379 MM |
5478 | && qty_const[REG_QTY (REGNO (arg))] != 0 |
5479 | && GET_CODE (qty_const[REG_QTY (REGNO (arg))]) != REG | |
5480 | && GET_CODE (qty_const[REG_QTY (REGNO (arg))]) != PLUS) | |
7afe21cc RK |
5481 | const_arg |
5482 | = gen_lowpart_if_possible (GET_MODE (arg), | |
30f72379 | 5483 | qty_const[REG_QTY (REGNO (arg))]); |
7afe21cc RK |
5484 | break; |
5485 | ||
5486 | case CONST: | |
5487 | case CONST_INT: | |
5488 | case SYMBOL_REF: | |
5489 | case LABEL_REF: | |
5490 | case CONST_DOUBLE: | |
5491 | const_arg = arg; | |
5492 | break; | |
5493 | ||
5494 | #ifdef HAVE_cc0 | |
5495 | case CC0: | |
5496 | folded_arg = prev_insn_cc0; | |
5497 | mode_arg = prev_insn_cc0_mode; | |
5498 | const_arg = equiv_constant (folded_arg); | |
5499 | break; | |
5500 | #endif | |
5501 | ||
5502 | default: | |
5503 | folded_arg = fold_rtx (arg, insn); | |
5504 | const_arg = equiv_constant (folded_arg); | |
5505 | } | |
5506 | ||
5507 | /* For the first three operands, see if the operand | |
5508 | is constant or equivalent to a constant. */ | |
5509 | switch (i) | |
5510 | { | |
5511 | case 0: | |
5512 | folded_arg0 = folded_arg; | |
5513 | const_arg0 = const_arg; | |
5514 | mode_arg0 = mode_arg; | |
5515 | break; | |
5516 | case 1: | |
5517 | folded_arg1 = folded_arg; | |
5518 | const_arg1 = const_arg; | |
5519 | break; | |
5520 | case 2: | |
5521 | const_arg2 = const_arg; | |
5522 | break; | |
5523 | } | |
5524 | ||
5525 | /* Pick the least expensive of the folded argument and an | |
5526 | equivalent constant argument. */ | |
5527 | if (const_arg == 0 || const_arg == folded_arg | |
5528 | || COST (const_arg) > COST (folded_arg)) | |
5529 | cheap_arg = folded_arg, expensive_arg = const_arg; | |
5530 | else | |
5531 | cheap_arg = const_arg, expensive_arg = folded_arg; | |
5532 | ||
5533 | /* Try to replace the operand with the cheapest of the two | |
5534 | possibilities. If it doesn't work and this is either of the first | |
5535 | two operands of a commutative operation, try swapping them. | |
5536 | If THAT fails, try the more expensive, provided it is cheaper | |
5537 | than what is already there. */ | |
5538 | ||
5539 | if (cheap_arg == XEXP (x, i)) | |
5540 | continue; | |
5541 | ||
5542 | if (insn == 0 && ! copied) | |
5543 | { | |
5544 | x = copy_rtx (x); | |
5545 | copied = 1; | |
5546 | } | |
5547 | ||
5548 | replacements[0] = cheap_arg, replacements[1] = expensive_arg; | |
5549 | for (j = 0; | |
5550 | j < 2 && replacements[j] | |
5551 | && COST (replacements[j]) < COST (XEXP (x, i)); | |
5552 | j++) | |
5553 | { | |
5554 | if (validate_change (insn, &XEXP (x, i), replacements[j], 0)) | |
5555 | break; | |
5556 | ||
5557 | if (code == NE || code == EQ || GET_RTX_CLASS (code) == 'c') | |
5558 | { | |
5559 | validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1); | |
5560 | validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1); | |
5561 | ||
5562 | if (apply_change_group ()) | |
5563 | { | |
5564 | /* Swap them back to be invalid so that this loop can | |
5565 | continue and flag them to be swapped back later. */ | |
5566 | rtx tem; | |
5567 | ||
5568 | tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1); | |
5569 | XEXP (x, 1) = tem; | |
5570 | must_swap = 1; | |
5571 | break; | |
5572 | } | |
5573 | } | |
5574 | } | |
5575 | } | |
5576 | ||
2d8b0f3a JL |
5577 | else |
5578 | { | |
5579 | if (fmt[i] == 'E') | |
5580 | /* Don't try to fold inside of a vector of expressions. | |
5581 | Doing nothing is harmless. */ | |
5582 | {;} | |
5583 | } | |
7afe21cc RK |
5584 | |
5585 | /* If a commutative operation, place a constant integer as the second | |
5586 | operand unless the first operand is also a constant integer. Otherwise, | |
5587 | place any constant second unless the first operand is also a constant. */ | |
5588 | ||
5589 | if (code == EQ || code == NE || GET_RTX_CLASS (code) == 'c') | |
5590 | { | |
5591 | if (must_swap || (const_arg0 | |
5592 | && (const_arg1 == 0 | |
5593 | || (GET_CODE (const_arg0) == CONST_INT | |
5594 | && GET_CODE (const_arg1) != CONST_INT)))) | |
5595 | { | |
5596 | register rtx tem = XEXP (x, 0); | |
5597 | ||
5598 | if (insn == 0 && ! copied) | |
5599 | { | |
5600 | x = copy_rtx (x); | |
5601 | copied = 1; | |
5602 | } | |
5603 | ||
5604 | validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1); | |
5605 | validate_change (insn, &XEXP (x, 1), tem, 1); | |
5606 | if (apply_change_group ()) | |
5607 | { | |
5608 | tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem; | |
5609 | tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem; | |
5610 | } | |
5611 | } | |
5612 | } | |
5613 | ||
5614 | /* If X is an arithmetic operation, see if we can simplify it. */ | |
5615 | ||
5616 | switch (GET_RTX_CLASS (code)) | |
5617 | { | |
5618 | case '1': | |
67a37737 RK |
5619 | { |
5620 | int is_const = 0; | |
5621 | ||
5622 | /* We can't simplify extension ops unless we know the | |
5623 | original mode. */ | |
5624 | if ((code == ZERO_EXTEND || code == SIGN_EXTEND) | |
5625 | && mode_arg0 == VOIDmode) | |
5626 | break; | |
5627 | ||
5628 | /* If we had a CONST, strip it off and put it back later if we | |
5629 | fold. */ | |
5630 | if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST) | |
5631 | is_const = 1, const_arg0 = XEXP (const_arg0, 0); | |
5632 | ||
5633 | new = simplify_unary_operation (code, mode, | |
5634 | const_arg0 ? const_arg0 : folded_arg0, | |
5635 | mode_arg0); | |
5636 | if (new != 0 && is_const) | |
38a448ca | 5637 | new = gen_rtx_CONST (mode, new); |
67a37737 | 5638 | } |
7afe21cc RK |
5639 | break; |
5640 | ||
5641 | case '<': | |
5642 | /* See what items are actually being compared and set FOLDED_ARG[01] | |
5643 | to those values and CODE to the actual comparison code. If any are | |
5644 | constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't | |
5645 | do anything if both operands are already known to be constant. */ | |
5646 | ||
5647 | if (const_arg0 == 0 || const_arg1 == 0) | |
5648 | { | |
5649 | struct table_elt *p0, *p1; | |
c610adec | 5650 | rtx true = const_true_rtx, false = const0_rtx; |
13c9910f | 5651 | enum machine_mode mode_arg1; |
c610adec RK |
5652 | |
5653 | #ifdef FLOAT_STORE_FLAG_VALUE | |
c7c955ee | 5654 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
c610adec | 5655 | { |
560c94a2 RK |
5656 | true = CONST_DOUBLE_FROM_REAL_VALUE (FLOAT_STORE_FLAG_VALUE, |
5657 | mode); | |
c610adec RK |
5658 | false = CONST0_RTX (mode); |
5659 | } | |
5660 | #endif | |
7afe21cc | 5661 | |
13c9910f RS |
5662 | code = find_comparison_args (code, &folded_arg0, &folded_arg1, |
5663 | &mode_arg0, &mode_arg1); | |
7afe21cc RK |
5664 | const_arg0 = equiv_constant (folded_arg0); |
5665 | const_arg1 = equiv_constant (folded_arg1); | |
5666 | ||
13c9910f RS |
5667 | /* If the mode is VOIDmode or a MODE_CC mode, we don't know |
5668 | what kinds of things are being compared, so we can't do | |
5669 | anything with this comparison. */ | |
7afe21cc RK |
5670 | |
5671 | if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC) | |
5672 | break; | |
5673 | ||
0f41302f MS |
5674 | /* If we do not now have two constants being compared, see |
5675 | if we can nevertheless deduce some things about the | |
5676 | comparison. */ | |
7afe21cc RK |
5677 | if (const_arg0 == 0 || const_arg1 == 0) |
5678 | { | |
0f41302f MS |
5679 | /* Is FOLDED_ARG0 frame-pointer plus a constant? Or |
5680 | non-explicit constant? These aren't zero, but we | |
5681 | don't know their sign. */ | |
7afe21cc RK |
5682 | if (const_arg1 == const0_rtx |
5683 | && (NONZERO_BASE_PLUS_P (folded_arg0) | |
5684 | #if 0 /* Sad to say, on sysvr4, #pragma weak can make a symbol address | |
5685 | come out as 0. */ | |
5686 | || GET_CODE (folded_arg0) == SYMBOL_REF | |
5687 | #endif | |
5688 | || GET_CODE (folded_arg0) == LABEL_REF | |
5689 | || GET_CODE (folded_arg0) == CONST)) | |
5690 | { | |
5691 | if (code == EQ) | |
c610adec | 5692 | return false; |
7afe21cc | 5693 | else if (code == NE) |
c610adec | 5694 | return true; |
7afe21cc RK |
5695 | } |
5696 | ||
5697 | /* See if the two operands are the same. We don't do this | |
5698 | for IEEE floating-point since we can't assume x == x | |
5699 | since x might be a NaN. */ | |
5700 | ||
5701 | if ((TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
a83afb65 | 5702 | || ! FLOAT_MODE_P (mode_arg0) || flag_fast_math) |
7afe21cc RK |
5703 | && (folded_arg0 == folded_arg1 |
5704 | || (GET_CODE (folded_arg0) == REG | |
5705 | && GET_CODE (folded_arg1) == REG | |
30f72379 MM |
5706 | && (REG_QTY (REGNO (folded_arg0)) |
5707 | == REG_QTY (REGNO (folded_arg1)))) | |
7afe21cc RK |
5708 | || ((p0 = lookup (folded_arg0, |
5709 | (safe_hash (folded_arg0, mode_arg0) | |
5710 | % NBUCKETS), mode_arg0)) | |
5711 | && (p1 = lookup (folded_arg1, | |
5712 | (safe_hash (folded_arg1, mode_arg0) | |
5713 | % NBUCKETS), mode_arg0)) | |
5714 | && p0->first_same_value == p1->first_same_value))) | |
5715 | return ((code == EQ || code == LE || code == GE | |
5716 | || code == LEU || code == GEU) | |
c610adec | 5717 | ? true : false); |
7afe21cc RK |
5718 | |
5719 | /* If FOLDED_ARG0 is a register, see if the comparison we are | |
5720 | doing now is either the same as we did before or the reverse | |
5721 | (we only check the reverse if not floating-point). */ | |
5722 | else if (GET_CODE (folded_arg0) == REG) | |
5723 | { | |
30f72379 | 5724 | int qty = REG_QTY (REGNO (folded_arg0)); |
7afe21cc RK |
5725 | |
5726 | if (REGNO_QTY_VALID_P (REGNO (folded_arg0)) | |
5727 | && (comparison_dominates_p (qty_comparison_code[qty], code) | |
5728 | || (comparison_dominates_p (qty_comparison_code[qty], | |
5729 | reverse_condition (code)) | |
cbf6a543 | 5730 | && ! FLOAT_MODE_P (mode_arg0))) |
7afe21cc RK |
5731 | && (rtx_equal_p (qty_comparison_const[qty], folded_arg1) |
5732 | || (const_arg1 | |
5733 | && rtx_equal_p (qty_comparison_const[qty], | |
5734 | const_arg1)) | |
5735 | || (GET_CODE (folded_arg1) == REG | |
30f72379 | 5736 | && (REG_QTY (REGNO (folded_arg1)) |
7afe21cc RK |
5737 | == qty_comparison_qty[qty])))) |
5738 | return (comparison_dominates_p (qty_comparison_code[qty], | |
5739 | code) | |
c610adec | 5740 | ? true : false); |
7afe21cc RK |
5741 | } |
5742 | } | |
5743 | } | |
5744 | ||
5745 | /* If we are comparing against zero, see if the first operand is | |
5746 | equivalent to an IOR with a constant. If so, we may be able to | |
5747 | determine the result of this comparison. */ | |
5748 | ||
5749 | if (const_arg1 == const0_rtx) | |
5750 | { | |
5751 | rtx y = lookup_as_function (folded_arg0, IOR); | |
5752 | rtx inner_const; | |
5753 | ||
5754 | if (y != 0 | |
5755 | && (inner_const = equiv_constant (XEXP (y, 1))) != 0 | |
5756 | && GET_CODE (inner_const) == CONST_INT | |
5757 | && INTVAL (inner_const) != 0) | |
5758 | { | |
5759 | int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1; | |
906c4e36 RK |
5760 | int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum |
5761 | && (INTVAL (inner_const) | |
5762 | & ((HOST_WIDE_INT) 1 << sign_bitnum))); | |
c610adec RK |
5763 | rtx true = const_true_rtx, false = const0_rtx; |
5764 | ||
5765 | #ifdef FLOAT_STORE_FLAG_VALUE | |
c7c955ee | 5766 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
c610adec | 5767 | { |
560c94a2 RK |
5768 | true = CONST_DOUBLE_FROM_REAL_VALUE (FLOAT_STORE_FLAG_VALUE, |
5769 | mode); | |
c610adec RK |
5770 | false = CONST0_RTX (mode); |
5771 | } | |
5772 | #endif | |
7afe21cc RK |
5773 | |
5774 | switch (code) | |
5775 | { | |
5776 | case EQ: | |
c610adec | 5777 | return false; |
7afe21cc | 5778 | case NE: |
c610adec | 5779 | return true; |
7afe21cc RK |
5780 | case LT: case LE: |
5781 | if (has_sign) | |
c610adec | 5782 | return true; |
7afe21cc RK |
5783 | break; |
5784 | case GT: case GE: | |
5785 | if (has_sign) | |
c610adec | 5786 | return false; |
7afe21cc | 5787 | break; |
e9a25f70 JL |
5788 | default: |
5789 | break; | |
7afe21cc RK |
5790 | } |
5791 | } | |
5792 | } | |
5793 | ||
5794 | new = simplify_relational_operation (code, mode_arg0, | |
5795 | const_arg0 ? const_arg0 : folded_arg0, | |
5796 | const_arg1 ? const_arg1 : folded_arg1); | |
c610adec RK |
5797 | #ifdef FLOAT_STORE_FLAG_VALUE |
5798 | if (new != 0 && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5799 | new = ((new == const0_rtx) ? CONST0_RTX (mode) | |
560c94a2 | 5800 | : CONST_DOUBLE_FROM_REAL_VALUE (FLOAT_STORE_FLAG_VALUE, mode)); |
c610adec | 5801 | #endif |
7afe21cc RK |
5802 | break; |
5803 | ||
5804 | case '2': | |
5805 | case 'c': | |
5806 | switch (code) | |
5807 | { | |
5808 | case PLUS: | |
5809 | /* If the second operand is a LABEL_REF, see if the first is a MINUS | |
5810 | with that LABEL_REF as its second operand. If so, the result is | |
5811 | the first operand of that MINUS. This handles switches with an | |
5812 | ADDR_DIFF_VEC table. */ | |
5813 | if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF) | |
5814 | { | |
e650cbda RK |
5815 | rtx y |
5816 | = GET_CODE (folded_arg0) == MINUS ? folded_arg0 | |
5817 | : lookup_as_function (folded_arg0, MINUS); | |
7afe21cc RK |
5818 | |
5819 | if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF | |
5820 | && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0)) | |
5821 | return XEXP (y, 0); | |
67a37737 RK |
5822 | |
5823 | /* Now try for a CONST of a MINUS like the above. */ | |
e650cbda RK |
5824 | if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0 |
5825 | : lookup_as_function (folded_arg0, CONST))) != 0 | |
67a37737 RK |
5826 | && GET_CODE (XEXP (y, 0)) == MINUS |
5827 | && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF | |
5828 | && XEXP (XEXP (XEXP (y, 0),1), 0) == XEXP (const_arg1, 0)) | |
5829 | return XEXP (XEXP (y, 0), 0); | |
7afe21cc | 5830 | } |
c2cc0778 | 5831 | |
e650cbda RK |
5832 | /* Likewise if the operands are in the other order. */ |
5833 | if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF) | |
5834 | { | |
5835 | rtx y | |
5836 | = GET_CODE (folded_arg1) == MINUS ? folded_arg1 | |
5837 | : lookup_as_function (folded_arg1, MINUS); | |
5838 | ||
5839 | if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF | |
5840 | && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0)) | |
5841 | return XEXP (y, 0); | |
5842 | ||
5843 | /* Now try for a CONST of a MINUS like the above. */ | |
5844 | if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1 | |
5845 | : lookup_as_function (folded_arg1, CONST))) != 0 | |
5846 | && GET_CODE (XEXP (y, 0)) == MINUS | |
5847 | && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF | |
5848 | && XEXP (XEXP (XEXP (y, 0),1), 0) == XEXP (const_arg0, 0)) | |
5849 | return XEXP (XEXP (y, 0), 0); | |
5850 | } | |
5851 | ||
c2cc0778 RK |
5852 | /* If second operand is a register equivalent to a negative |
5853 | CONST_INT, see if we can find a register equivalent to the | |
5854 | positive constant. Make a MINUS if so. Don't do this for | |
5d595063 | 5855 | a non-negative constant since we might then alternate between |
c2cc0778 | 5856 | chosing positive and negative constants. Having the positive |
5d595063 RK |
5857 | constant previously-used is the more common case. Be sure |
5858 | the resulting constant is non-negative; if const_arg1 were | |
5859 | the smallest negative number this would overflow: depending | |
5860 | on the mode, this would either just be the same value (and | |
5861 | hence not save anything) or be incorrect. */ | |
5862 | if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT | |
5863 | && INTVAL (const_arg1) < 0 | |
5864 | && - INTVAL (const_arg1) >= 0 | |
5865 | && GET_CODE (folded_arg1) == REG) | |
c2cc0778 RK |
5866 | { |
5867 | rtx new_const = GEN_INT (- INTVAL (const_arg1)); | |
5868 | struct table_elt *p | |
5869 | = lookup (new_const, safe_hash (new_const, mode) % NBUCKETS, | |
5870 | mode); | |
5871 | ||
5872 | if (p) | |
5873 | for (p = p->first_same_value; p; p = p->next_same_value) | |
5874 | if (GET_CODE (p->exp) == REG) | |
5875 | return cse_gen_binary (MINUS, mode, folded_arg0, | |
5876 | canon_reg (p->exp, NULL_RTX)); | |
5877 | } | |
13c9910f RS |
5878 | goto from_plus; |
5879 | ||
5880 | case MINUS: | |
5881 | /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2). | |
5882 | If so, produce (PLUS Z C2-C). */ | |
5883 | if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT) | |
5884 | { | |
5885 | rtx y = lookup_as_function (XEXP (x, 0), PLUS); | |
5886 | if (y && GET_CODE (XEXP (y, 1)) == CONST_INT) | |
f3becefd RK |
5887 | return fold_rtx (plus_constant (copy_rtx (y), |
5888 | -INTVAL (const_arg1)), | |
a3b5c94a | 5889 | NULL_RTX); |
13c9910f | 5890 | } |
7afe21cc | 5891 | |
0f41302f | 5892 | /* ... fall through ... */ |
7afe21cc | 5893 | |
13c9910f | 5894 | from_plus: |
7afe21cc RK |
5895 | case SMIN: case SMAX: case UMIN: case UMAX: |
5896 | case IOR: case AND: case XOR: | |
5897 | case MULT: case DIV: case UDIV: | |
5898 | case ASHIFT: case LSHIFTRT: case ASHIFTRT: | |
5899 | /* If we have (<op> <reg> <const_int>) for an associative OP and REG | |
5900 | is known to be of similar form, we may be able to replace the | |
5901 | operation with a combined operation. This may eliminate the | |
5902 | intermediate operation if every use is simplified in this way. | |
5903 | Note that the similar optimization done by combine.c only works | |
5904 | if the intermediate operation's result has only one reference. */ | |
5905 | ||
5906 | if (GET_CODE (folded_arg0) == REG | |
5907 | && const_arg1 && GET_CODE (const_arg1) == CONST_INT) | |
5908 | { | |
5909 | int is_shift | |
5910 | = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT); | |
5911 | rtx y = lookup_as_function (folded_arg0, code); | |
5912 | rtx inner_const; | |
5913 | enum rtx_code associate_code; | |
5914 | rtx new_const; | |
5915 | ||
5916 | if (y == 0 | |
5917 | || 0 == (inner_const | |
5918 | = equiv_constant (fold_rtx (XEXP (y, 1), 0))) | |
5919 | || GET_CODE (inner_const) != CONST_INT | |
5920 | /* If we have compiled a statement like | |
5921 | "if (x == (x & mask1))", and now are looking at | |
5922 | "x & mask2", we will have a case where the first operand | |
5923 | of Y is the same as our first operand. Unless we detect | |
5924 | this case, an infinite loop will result. */ | |
5925 | || XEXP (y, 0) == folded_arg0) | |
5926 | break; | |
5927 | ||
5928 | /* Don't associate these operations if they are a PLUS with the | |
5929 | same constant and it is a power of two. These might be doable | |
5930 | with a pre- or post-increment. Similarly for two subtracts of | |
5931 | identical powers of two with post decrement. */ | |
5932 | ||
5933 | if (code == PLUS && INTVAL (const_arg1) == INTVAL (inner_const) | |
940da324 JL |
5934 | && ((HAVE_PRE_INCREMENT |
5935 | && exact_log2 (INTVAL (const_arg1)) >= 0) | |
5936 | || (HAVE_POST_INCREMENT | |
5937 | && exact_log2 (INTVAL (const_arg1)) >= 0) | |
5938 | || (HAVE_PRE_DECREMENT | |
5939 | && exact_log2 (- INTVAL (const_arg1)) >= 0) | |
5940 | || (HAVE_POST_DECREMENT | |
5941 | && exact_log2 (- INTVAL (const_arg1)) >= 0))) | |
7afe21cc RK |
5942 | break; |
5943 | ||
5944 | /* Compute the code used to compose the constants. For example, | |
5945 | A/C1/C2 is A/(C1 * C2), so if CODE == DIV, we want MULT. */ | |
5946 | ||
5947 | associate_code | |
5948 | = (code == MULT || code == DIV || code == UDIV ? MULT | |
5949 | : is_shift || code == PLUS || code == MINUS ? PLUS : code); | |
5950 | ||
5951 | new_const = simplify_binary_operation (associate_code, mode, | |
5952 | const_arg1, inner_const); | |
5953 | ||
5954 | if (new_const == 0) | |
5955 | break; | |
5956 | ||
5957 | /* If we are associating shift operations, don't let this | |
4908e508 RS |
5958 | produce a shift of the size of the object or larger. |
5959 | This could occur when we follow a sign-extend by a right | |
5960 | shift on a machine that does a sign-extend as a pair | |
5961 | of shifts. */ | |
7afe21cc RK |
5962 | |
5963 | if (is_shift && GET_CODE (new_const) == CONST_INT | |
4908e508 RS |
5964 | && INTVAL (new_const) >= GET_MODE_BITSIZE (mode)) |
5965 | { | |
5966 | /* As an exception, we can turn an ASHIFTRT of this | |
5967 | form into a shift of the number of bits - 1. */ | |
5968 | if (code == ASHIFTRT) | |
5969 | new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1); | |
5970 | else | |
5971 | break; | |
5972 | } | |
7afe21cc RK |
5973 | |
5974 | y = copy_rtx (XEXP (y, 0)); | |
5975 | ||
5976 | /* If Y contains our first operand (the most common way this | |
5977 | can happen is if Y is a MEM), we would do into an infinite | |
5978 | loop if we tried to fold it. So don't in that case. */ | |
5979 | ||
5980 | if (! reg_mentioned_p (folded_arg0, y)) | |
5981 | y = fold_rtx (y, insn); | |
5982 | ||
96b0e481 | 5983 | return cse_gen_binary (code, mode, y, new_const); |
7afe21cc | 5984 | } |
e9a25f70 JL |
5985 | break; |
5986 | ||
5987 | default: | |
5988 | break; | |
7afe21cc RK |
5989 | } |
5990 | ||
5991 | new = simplify_binary_operation (code, mode, | |
5992 | const_arg0 ? const_arg0 : folded_arg0, | |
5993 | const_arg1 ? const_arg1 : folded_arg1); | |
5994 | break; | |
5995 | ||
5996 | case 'o': | |
5997 | /* (lo_sum (high X) X) is simply X. */ | |
5998 | if (code == LO_SUM && const_arg0 != 0 | |
5999 | && GET_CODE (const_arg0) == HIGH | |
6000 | && rtx_equal_p (XEXP (const_arg0, 0), const_arg1)) | |
6001 | return const_arg1; | |
6002 | break; | |
6003 | ||
6004 | case '3': | |
6005 | case 'b': | |
6006 | new = simplify_ternary_operation (code, mode, mode_arg0, | |
6007 | const_arg0 ? const_arg0 : folded_arg0, | |
6008 | const_arg1 ? const_arg1 : folded_arg1, | |
6009 | const_arg2 ? const_arg2 : XEXP (x, 2)); | |
6010 | break; | |
ee5332b8 RH |
6011 | |
6012 | case 'x': | |
6013 | /* Always eliminate CONSTANT_P_RTX at this stage. */ | |
6014 | if (code == CONSTANT_P_RTX) | |
6015 | return (const_arg0 ? const1_rtx : const0_rtx); | |
6016 | break; | |
7afe21cc RK |
6017 | } |
6018 | ||
6019 | return new ? new : x; | |
6020 | } | |
6021 | \f | |
6022 | /* Return a constant value currently equivalent to X. | |
6023 | Return 0 if we don't know one. */ | |
6024 | ||
6025 | static rtx | |
6026 | equiv_constant (x) | |
6027 | rtx x; | |
6028 | { | |
6029 | if (GET_CODE (x) == REG | |
6030 | && REGNO_QTY_VALID_P (REGNO (x)) | |
30f72379 MM |
6031 | && qty_const[REG_QTY (REGNO (x))]) |
6032 | x = gen_lowpart_if_possible (GET_MODE (x), qty_const[REG_QTY (REGNO (x))]); | |
7afe21cc | 6033 | |
2ce5e1b4 | 6034 | if (x == 0 || CONSTANT_P (x)) |
7afe21cc RK |
6035 | return x; |
6036 | ||
fc3ffe83 RK |
6037 | /* If X is a MEM, try to fold it outside the context of any insn to see if |
6038 | it might be equivalent to a constant. That handles the case where it | |
6039 | is a constant-pool reference. Then try to look it up in the hash table | |
6040 | in case it is something whose value we have seen before. */ | |
6041 | ||
6042 | if (GET_CODE (x) == MEM) | |
6043 | { | |
6044 | struct table_elt *elt; | |
6045 | ||
906c4e36 | 6046 | x = fold_rtx (x, NULL_RTX); |
fc3ffe83 RK |
6047 | if (CONSTANT_P (x)) |
6048 | return x; | |
6049 | ||
6050 | elt = lookup (x, safe_hash (x, GET_MODE (x)) % NBUCKETS, GET_MODE (x)); | |
6051 | if (elt == 0) | |
6052 | return 0; | |
6053 | ||
6054 | for (elt = elt->first_same_value; elt; elt = elt->next_same_value) | |
6055 | if (elt->is_const && CONSTANT_P (elt->exp)) | |
6056 | return elt->exp; | |
6057 | } | |
6058 | ||
7afe21cc RK |
6059 | return 0; |
6060 | } | |
6061 | \f | |
6062 | /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point | |
6063 | number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the | |
6064 | least-significant part of X. | |
6065 | MODE specifies how big a part of X to return. | |
6066 | ||
6067 | If the requested operation cannot be done, 0 is returned. | |
6068 | ||
6069 | This is similar to gen_lowpart in emit-rtl.c. */ | |
6070 | ||
6071 | rtx | |
6072 | gen_lowpart_if_possible (mode, x) | |
6073 | enum machine_mode mode; | |
6074 | register rtx x; | |
6075 | { | |
6076 | rtx result = gen_lowpart_common (mode, x); | |
6077 | ||
6078 | if (result) | |
6079 | return result; | |
6080 | else if (GET_CODE (x) == MEM) | |
6081 | { | |
6082 | /* This is the only other case we handle. */ | |
6083 | register int offset = 0; | |
6084 | rtx new; | |
6085 | ||
f76b9db2 ILT |
6086 | if (WORDS_BIG_ENDIAN) |
6087 | offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD) | |
6088 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)); | |
6089 | if (BYTES_BIG_ENDIAN) | |
6090 | /* Adjust the address so that the address-after-the-data is | |
6091 | unchanged. */ | |
6092 | offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode)) | |
6093 | - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x)))); | |
38a448ca | 6094 | new = gen_rtx_MEM (mode, plus_constant (XEXP (x, 0), offset)); |
7afe21cc RK |
6095 | if (! memory_address_p (mode, XEXP (new, 0))) |
6096 | return 0; | |
7afe21cc | 6097 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (x); |
c6df88cb | 6098 | MEM_COPY_ATTRIBUTES (new, x); |
7afe21cc RK |
6099 | return new; |
6100 | } | |
6101 | else | |
6102 | return 0; | |
6103 | } | |
6104 | \f | |
6105 | /* Given INSN, a jump insn, TAKEN indicates if we are following the "taken" | |
6106 | branch. It will be zero if not. | |
6107 | ||
6108 | In certain cases, this can cause us to add an equivalence. For example, | |
6109 | if we are following the taken case of | |
6110 | if (i == 2) | |
6111 | we can add the fact that `i' and '2' are now equivalent. | |
6112 | ||
6113 | In any case, we can record that this comparison was passed. If the same | |
6114 | comparison is seen later, we will know its value. */ | |
6115 | ||
6116 | static void | |
6117 | record_jump_equiv (insn, taken) | |
6118 | rtx insn; | |
6119 | int taken; | |
6120 | { | |
6121 | int cond_known_true; | |
6122 | rtx op0, op1; | |
13c9910f | 6123 | enum machine_mode mode, mode0, mode1; |
7afe21cc RK |
6124 | int reversed_nonequality = 0; |
6125 | enum rtx_code code; | |
6126 | ||
6127 | /* Ensure this is the right kind of insn. */ | |
6128 | if (! condjump_p (insn) || simplejump_p (insn)) | |
6129 | return; | |
6130 | ||
6131 | /* See if this jump condition is known true or false. */ | |
6132 | if (taken) | |
6133 | cond_known_true = (XEXP (SET_SRC (PATTERN (insn)), 2) == pc_rtx); | |
6134 | else | |
6135 | cond_known_true = (XEXP (SET_SRC (PATTERN (insn)), 1) == pc_rtx); | |
6136 | ||
6137 | /* Get the type of comparison being done and the operands being compared. | |
6138 | If we had to reverse a non-equality condition, record that fact so we | |
6139 | know that it isn't valid for floating-point. */ | |
6140 | code = GET_CODE (XEXP (SET_SRC (PATTERN (insn)), 0)); | |
6141 | op0 = fold_rtx (XEXP (XEXP (SET_SRC (PATTERN (insn)), 0), 0), insn); | |
6142 | op1 = fold_rtx (XEXP (XEXP (SET_SRC (PATTERN (insn)), 0), 1), insn); | |
6143 | ||
13c9910f | 6144 | code = find_comparison_args (code, &op0, &op1, &mode0, &mode1); |
7afe21cc RK |
6145 | if (! cond_known_true) |
6146 | { | |
6147 | reversed_nonequality = (code != EQ && code != NE); | |
6148 | code = reverse_condition (code); | |
6149 | } | |
6150 | ||
6151 | /* The mode is the mode of the non-constant. */ | |
13c9910f RS |
6152 | mode = mode0; |
6153 | if (mode1 != VOIDmode) | |
6154 | mode = mode1; | |
7afe21cc RK |
6155 | |
6156 | record_jump_cond (code, mode, op0, op1, reversed_nonequality); | |
6157 | } | |
6158 | ||
6159 | /* We know that comparison CODE applied to OP0 and OP1 in MODE is true. | |
6160 | REVERSED_NONEQUALITY is nonzero if CODE had to be swapped. | |
6161 | Make any useful entries we can with that information. Called from | |
6162 | above function and called recursively. */ | |
6163 | ||
6164 | static void | |
6165 | record_jump_cond (code, mode, op0, op1, reversed_nonequality) | |
6166 | enum rtx_code code; | |
6167 | enum machine_mode mode; | |
6168 | rtx op0, op1; | |
6169 | int reversed_nonequality; | |
6170 | { | |
2197a88a | 6171 | unsigned op0_hash, op1_hash; |
7afe21cc RK |
6172 | int op0_in_memory, op0_in_struct, op1_in_memory, op1_in_struct; |
6173 | struct table_elt *op0_elt, *op1_elt; | |
6174 | ||
6175 | /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG, | |
6176 | we know that they are also equal in the smaller mode (this is also | |
6177 | true for all smaller modes whether or not there is a SUBREG, but | |
ac7ef8d5 | 6178 | is not worth testing for with no SUBREG). */ |
7afe21cc | 6179 | |
2e794ee8 | 6180 | /* Note that GET_MODE (op0) may not equal MODE. */ |
7afe21cc | 6181 | if (code == EQ && GET_CODE (op0) == SUBREG |
2e794ee8 RS |
6182 | && (GET_MODE_SIZE (GET_MODE (op0)) |
6183 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))) | |
7afe21cc RK |
6184 | { |
6185 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0)); | |
6186 | rtx tem = gen_lowpart_if_possible (inner_mode, op1); | |
6187 | ||
6188 | record_jump_cond (code, mode, SUBREG_REG (op0), | |
38a448ca | 6189 | tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0), |
7afe21cc RK |
6190 | reversed_nonequality); |
6191 | } | |
6192 | ||
6193 | if (code == EQ && GET_CODE (op1) == SUBREG | |
2e794ee8 RS |
6194 | && (GET_MODE_SIZE (GET_MODE (op1)) |
6195 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1))))) | |
7afe21cc RK |
6196 | { |
6197 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1)); | |
6198 | rtx tem = gen_lowpart_if_possible (inner_mode, op0); | |
6199 | ||
6200 | record_jump_cond (code, mode, SUBREG_REG (op1), | |
38a448ca | 6201 | tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0), |
7afe21cc RK |
6202 | reversed_nonequality); |
6203 | } | |
6204 | ||
6205 | /* Similarly, if this is an NE comparison, and either is a SUBREG | |
6206 | making a smaller mode, we know the whole thing is also NE. */ | |
6207 | ||
2e794ee8 RS |
6208 | /* Note that GET_MODE (op0) may not equal MODE; |
6209 | if we test MODE instead, we can get an infinite recursion | |
6210 | alternating between two modes each wider than MODE. */ | |
6211 | ||
7afe21cc RK |
6212 | if (code == NE && GET_CODE (op0) == SUBREG |
6213 | && subreg_lowpart_p (op0) | |
2e794ee8 RS |
6214 | && (GET_MODE_SIZE (GET_MODE (op0)) |
6215 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))) | |
7afe21cc RK |
6216 | { |
6217 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0)); | |
6218 | rtx tem = gen_lowpart_if_possible (inner_mode, op1); | |
6219 | ||
6220 | record_jump_cond (code, mode, SUBREG_REG (op0), | |
38a448ca | 6221 | tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0), |
7afe21cc RK |
6222 | reversed_nonequality); |
6223 | } | |
6224 | ||
6225 | if (code == NE && GET_CODE (op1) == SUBREG | |
6226 | && subreg_lowpart_p (op1) | |
2e794ee8 RS |
6227 | && (GET_MODE_SIZE (GET_MODE (op1)) |
6228 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1))))) | |
7afe21cc RK |
6229 | { |
6230 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1)); | |
6231 | rtx tem = gen_lowpart_if_possible (inner_mode, op0); | |
6232 | ||
6233 | record_jump_cond (code, mode, SUBREG_REG (op1), | |
38a448ca | 6234 | tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0), |
7afe21cc RK |
6235 | reversed_nonequality); |
6236 | } | |
6237 | ||
6238 | /* Hash both operands. */ | |
6239 | ||
6240 | do_not_record = 0; | |
6241 | hash_arg_in_memory = 0; | |
6242 | hash_arg_in_struct = 0; | |
2197a88a | 6243 | op0_hash = HASH (op0, mode); |
7afe21cc RK |
6244 | op0_in_memory = hash_arg_in_memory; |
6245 | op0_in_struct = hash_arg_in_struct; | |
6246 | ||
6247 | if (do_not_record) | |
6248 | return; | |
6249 | ||
6250 | do_not_record = 0; | |
6251 | hash_arg_in_memory = 0; | |
6252 | hash_arg_in_struct = 0; | |
2197a88a | 6253 | op1_hash = HASH (op1, mode); |
7afe21cc RK |
6254 | op1_in_memory = hash_arg_in_memory; |
6255 | op1_in_struct = hash_arg_in_struct; | |
6256 | ||
6257 | if (do_not_record) | |
6258 | return; | |
6259 | ||
6260 | /* Look up both operands. */ | |
2197a88a RK |
6261 | op0_elt = lookup (op0, op0_hash, mode); |
6262 | op1_elt = lookup (op1, op1_hash, mode); | |
7afe21cc | 6263 | |
af3869c1 RK |
6264 | /* If both operands are already equivalent or if they are not in the |
6265 | table but are identical, do nothing. */ | |
6266 | if ((op0_elt != 0 && op1_elt != 0 | |
6267 | && op0_elt->first_same_value == op1_elt->first_same_value) | |
6268 | || op0 == op1 || rtx_equal_p (op0, op1)) | |
6269 | return; | |
6270 | ||
7afe21cc | 6271 | /* If we aren't setting two things equal all we can do is save this |
b2796a4b RK |
6272 | comparison. Similarly if this is floating-point. In the latter |
6273 | case, OP1 might be zero and both -0.0 and 0.0 are equal to it. | |
6274 | If we record the equality, we might inadvertently delete code | |
6275 | whose intent was to change -0 to +0. */ | |
6276 | ||
cbf6a543 | 6277 | if (code != EQ || FLOAT_MODE_P (GET_MODE (op0))) |
7afe21cc RK |
6278 | { |
6279 | /* If we reversed a floating-point comparison, if OP0 is not a | |
6280 | register, or if OP1 is neither a register or constant, we can't | |
6281 | do anything. */ | |
6282 | ||
6283 | if (GET_CODE (op1) != REG) | |
6284 | op1 = equiv_constant (op1); | |
6285 | ||
cbf6a543 | 6286 | if ((reversed_nonequality && FLOAT_MODE_P (mode)) |
7afe21cc RK |
6287 | || GET_CODE (op0) != REG || op1 == 0) |
6288 | return; | |
6289 | ||
6290 | /* Put OP0 in the hash table if it isn't already. This gives it a | |
6291 | new quantity number. */ | |
6292 | if (op0_elt == 0) | |
6293 | { | |
906c4e36 | 6294 | if (insert_regs (op0, NULL_PTR, 0)) |
7afe21cc RK |
6295 | { |
6296 | rehash_using_reg (op0); | |
2197a88a | 6297 | op0_hash = HASH (op0, mode); |
2bb81c86 RK |
6298 | |
6299 | /* If OP0 is contained in OP1, this changes its hash code | |
6300 | as well. Faster to rehash than to check, except | |
6301 | for the simple case of a constant. */ | |
6302 | if (! CONSTANT_P (op1)) | |
2197a88a | 6303 | op1_hash = HASH (op1,mode); |
7afe21cc RK |
6304 | } |
6305 | ||
2197a88a | 6306 | op0_elt = insert (op0, NULL_PTR, op0_hash, mode); |
7afe21cc RK |
6307 | op0_elt->in_memory = op0_in_memory; |
6308 | op0_elt->in_struct = op0_in_struct; | |
6309 | } | |
6310 | ||
30f72379 | 6311 | qty_comparison_code[REG_QTY (REGNO (op0))] = code; |
7afe21cc RK |
6312 | if (GET_CODE (op1) == REG) |
6313 | { | |
5d5ea909 | 6314 | /* Look it up again--in case op0 and op1 are the same. */ |
2197a88a | 6315 | op1_elt = lookup (op1, op1_hash, mode); |
5d5ea909 | 6316 | |
7afe21cc RK |
6317 | /* Put OP1 in the hash table so it gets a new quantity number. */ |
6318 | if (op1_elt == 0) | |
6319 | { | |
906c4e36 | 6320 | if (insert_regs (op1, NULL_PTR, 0)) |
7afe21cc RK |
6321 | { |
6322 | rehash_using_reg (op1); | |
2197a88a | 6323 | op1_hash = HASH (op1, mode); |
7afe21cc RK |
6324 | } |
6325 | ||
2197a88a | 6326 | op1_elt = insert (op1, NULL_PTR, op1_hash, mode); |
7afe21cc RK |
6327 | op1_elt->in_memory = op1_in_memory; |
6328 | op1_elt->in_struct = op1_in_struct; | |
6329 | } | |
6330 | ||
30f72379 MM |
6331 | qty_comparison_qty[REG_QTY (REGNO (op0))] = REG_QTY (REGNO (op1)); |
6332 | qty_comparison_const[REG_QTY (REGNO (op0))] = 0; | |
7afe21cc RK |
6333 | } |
6334 | else | |
6335 | { | |
30f72379 MM |
6336 | qty_comparison_qty[REG_QTY (REGNO (op0))] = -1; |
6337 | qty_comparison_const[REG_QTY (REGNO (op0))] = op1; | |
7afe21cc RK |
6338 | } |
6339 | ||
6340 | return; | |
6341 | } | |
6342 | ||
eb5ad42a RS |
6343 | /* If either side is still missing an equivalence, make it now, |
6344 | then merge the equivalences. */ | |
7afe21cc | 6345 | |
7afe21cc RK |
6346 | if (op0_elt == 0) |
6347 | { | |
eb5ad42a | 6348 | if (insert_regs (op0, NULL_PTR, 0)) |
7afe21cc RK |
6349 | { |
6350 | rehash_using_reg (op0); | |
2197a88a | 6351 | op0_hash = HASH (op0, mode); |
7afe21cc RK |
6352 | } |
6353 | ||
2197a88a | 6354 | op0_elt = insert (op0, NULL_PTR, op0_hash, mode); |
7afe21cc RK |
6355 | op0_elt->in_memory = op0_in_memory; |
6356 | op0_elt->in_struct = op0_in_struct; | |
7afe21cc RK |
6357 | } |
6358 | ||
6359 | if (op1_elt == 0) | |
6360 | { | |
eb5ad42a | 6361 | if (insert_regs (op1, NULL_PTR, 0)) |
7afe21cc RK |
6362 | { |
6363 | rehash_using_reg (op1); | |
2197a88a | 6364 | op1_hash = HASH (op1, mode); |
7afe21cc RK |
6365 | } |
6366 | ||
2197a88a | 6367 | op1_elt = insert (op1, NULL_PTR, op1_hash, mode); |
7afe21cc RK |
6368 | op1_elt->in_memory = op1_in_memory; |
6369 | op1_elt->in_struct = op1_in_struct; | |
7afe21cc | 6370 | } |
eb5ad42a RS |
6371 | |
6372 | merge_equiv_classes (op0_elt, op1_elt); | |
6373 | last_jump_equiv_class = op0_elt; | |
7afe21cc RK |
6374 | } |
6375 | \f | |
6376 | /* CSE processing for one instruction. | |
6377 | First simplify sources and addresses of all assignments | |
6378 | in the instruction, using previously-computed equivalents values. | |
6379 | Then install the new sources and destinations in the table | |
6380 | of available values. | |
6381 | ||
1ed0205e VM |
6382 | If LIBCALL_INSN is nonzero, don't record any equivalence made in |
6383 | the insn. It means that INSN is inside libcall block. In this | |
6384 | case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */ | |
7afe21cc RK |
6385 | |
6386 | /* Data on one SET contained in the instruction. */ | |
6387 | ||
6388 | struct set | |
6389 | { | |
6390 | /* The SET rtx itself. */ | |
6391 | rtx rtl; | |
6392 | /* The SET_SRC of the rtx (the original value, if it is changing). */ | |
6393 | rtx src; | |
6394 | /* The hash-table element for the SET_SRC of the SET. */ | |
6395 | struct table_elt *src_elt; | |
2197a88a RK |
6396 | /* Hash value for the SET_SRC. */ |
6397 | unsigned src_hash; | |
6398 | /* Hash value for the SET_DEST. */ | |
6399 | unsigned dest_hash; | |
7afe21cc RK |
6400 | /* The SET_DEST, with SUBREG, etc., stripped. */ |
6401 | rtx inner_dest; | |
6402 | /* Place where the pointer to the INNER_DEST was found. */ | |
6403 | rtx *inner_dest_loc; | |
6404 | /* Nonzero if the SET_SRC is in memory. */ | |
6405 | char src_in_memory; | |
6406 | /* Nonzero if the SET_SRC is in a structure. */ | |
6407 | char src_in_struct; | |
6408 | /* Nonzero if the SET_SRC contains something | |
6409 | whose value cannot be predicted and understood. */ | |
6410 | char src_volatile; | |
6411 | /* Original machine mode, in case it becomes a CONST_INT. */ | |
6412 | enum machine_mode mode; | |
6413 | /* A constant equivalent for SET_SRC, if any. */ | |
6414 | rtx src_const; | |
2197a88a RK |
6415 | /* Hash value of constant equivalent for SET_SRC. */ |
6416 | unsigned src_const_hash; | |
7afe21cc RK |
6417 | /* Table entry for constant equivalent for SET_SRC, if any. */ |
6418 | struct table_elt *src_const_elt; | |
6419 | }; | |
6420 | ||
6421 | static void | |
7bd8b2a8 | 6422 | cse_insn (insn, libcall_insn) |
7afe21cc | 6423 | rtx insn; |
7bd8b2a8 | 6424 | rtx libcall_insn; |
7afe21cc RK |
6425 | { |
6426 | register rtx x = PATTERN (insn); | |
7afe21cc | 6427 | register int i; |
92f9aa51 | 6428 | rtx tem; |
7afe21cc RK |
6429 | register int n_sets = 0; |
6430 | ||
2d8b0f3a | 6431 | #ifdef HAVE_cc0 |
7afe21cc RK |
6432 | /* Records what this insn does to set CC0. */ |
6433 | rtx this_insn_cc0 = 0; | |
135d84b8 | 6434 | enum machine_mode this_insn_cc0_mode = VOIDmode; |
2d8b0f3a | 6435 | #endif |
7afe21cc RK |
6436 | |
6437 | rtx src_eqv = 0; | |
6438 | struct table_elt *src_eqv_elt = 0; | |
6439 | int src_eqv_volatile; | |
6440 | int src_eqv_in_memory; | |
6441 | int src_eqv_in_struct; | |
2197a88a | 6442 | unsigned src_eqv_hash; |
7afe21cc RK |
6443 | |
6444 | struct set *sets; | |
6445 | ||
6446 | this_insn = insn; | |
7afe21cc RK |
6447 | |
6448 | /* Find all the SETs and CLOBBERs in this instruction. | |
6449 | Record all the SETs in the array `set' and count them. | |
6450 | Also determine whether there is a CLOBBER that invalidates | |
6451 | all memory references, or all references at varying addresses. */ | |
6452 | ||
f1e7c95f RK |
6453 | if (GET_CODE (insn) == CALL_INSN) |
6454 | { | |
6455 | for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1)) | |
6456 | if (GET_CODE (XEXP (tem, 0)) == CLOBBER) | |
bb4034b3 | 6457 | invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode); |
f1e7c95f RK |
6458 | } |
6459 | ||
7afe21cc RK |
6460 | if (GET_CODE (x) == SET) |
6461 | { | |
6462 | sets = (struct set *) alloca (sizeof (struct set)); | |
6463 | sets[0].rtl = x; | |
6464 | ||
6465 | /* Ignore SETs that are unconditional jumps. | |
6466 | They never need cse processing, so this does not hurt. | |
6467 | The reason is not efficiency but rather | |
6468 | so that we can test at the end for instructions | |
6469 | that have been simplified to unconditional jumps | |
6470 | and not be misled by unchanged instructions | |
6471 | that were unconditional jumps to begin with. */ | |
6472 | if (SET_DEST (x) == pc_rtx | |
6473 | && GET_CODE (SET_SRC (x)) == LABEL_REF) | |
6474 | ; | |
6475 | ||
6476 | /* Don't count call-insns, (set (reg 0) (call ...)), as a set. | |
6477 | The hard function value register is used only once, to copy to | |
6478 | someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)! | |
6479 | Ensure we invalidate the destination register. On the 80386 no | |
7722328e | 6480 | other code would invalidate it since it is a fixed_reg. |
0f41302f | 6481 | We need not check the return of apply_change_group; see canon_reg. */ |
7afe21cc RK |
6482 | |
6483 | else if (GET_CODE (SET_SRC (x)) == CALL) | |
6484 | { | |
6485 | canon_reg (SET_SRC (x), insn); | |
77fa0940 | 6486 | apply_change_group (); |
7afe21cc | 6487 | fold_rtx (SET_SRC (x), insn); |
bb4034b3 | 6488 | invalidate (SET_DEST (x), VOIDmode); |
7afe21cc RK |
6489 | } |
6490 | else | |
6491 | n_sets = 1; | |
6492 | } | |
6493 | else if (GET_CODE (x) == PARALLEL) | |
6494 | { | |
6495 | register int lim = XVECLEN (x, 0); | |
6496 | ||
6497 | sets = (struct set *) alloca (lim * sizeof (struct set)); | |
6498 | ||
6499 | /* Find all regs explicitly clobbered in this insn, | |
6500 | and ensure they are not replaced with any other regs | |
6501 | elsewhere in this insn. | |
6502 | When a reg that is clobbered is also used for input, | |
6503 | we should presume that that is for a reason, | |
6504 | and we should not substitute some other register | |
6505 | which is not supposed to be clobbered. | |
6506 | Therefore, this loop cannot be merged into the one below | |
830a38ee | 6507 | because a CALL may precede a CLOBBER and refer to the |
7afe21cc RK |
6508 | value clobbered. We must not let a canonicalization do |
6509 | anything in that case. */ | |
6510 | for (i = 0; i < lim; i++) | |
6511 | { | |
6512 | register rtx y = XVECEXP (x, 0, i); | |
2708da92 RS |
6513 | if (GET_CODE (y) == CLOBBER) |
6514 | { | |
6515 | rtx clobbered = XEXP (y, 0); | |
6516 | ||
6517 | if (GET_CODE (clobbered) == REG | |
6518 | || GET_CODE (clobbered) == SUBREG) | |
bb4034b3 | 6519 | invalidate (clobbered, VOIDmode); |
2708da92 RS |
6520 | else if (GET_CODE (clobbered) == STRICT_LOW_PART |
6521 | || GET_CODE (clobbered) == ZERO_EXTRACT) | |
bb4034b3 | 6522 | invalidate (XEXP (clobbered, 0), GET_MODE (clobbered)); |
2708da92 | 6523 | } |
7afe21cc RK |
6524 | } |
6525 | ||
6526 | for (i = 0; i < lim; i++) | |
6527 | { | |
6528 | register rtx y = XVECEXP (x, 0, i); | |
6529 | if (GET_CODE (y) == SET) | |
6530 | { | |
7722328e RK |
6531 | /* As above, we ignore unconditional jumps and call-insns and |
6532 | ignore the result of apply_change_group. */ | |
7afe21cc RK |
6533 | if (GET_CODE (SET_SRC (y)) == CALL) |
6534 | { | |
6535 | canon_reg (SET_SRC (y), insn); | |
77fa0940 | 6536 | apply_change_group (); |
7afe21cc | 6537 | fold_rtx (SET_SRC (y), insn); |
bb4034b3 | 6538 | invalidate (SET_DEST (y), VOIDmode); |
7afe21cc RK |
6539 | } |
6540 | else if (SET_DEST (y) == pc_rtx | |
6541 | && GET_CODE (SET_SRC (y)) == LABEL_REF) | |
6542 | ; | |
6543 | else | |
6544 | sets[n_sets++].rtl = y; | |
6545 | } | |
6546 | else if (GET_CODE (y) == CLOBBER) | |
6547 | { | |
9ae8ffe7 | 6548 | /* If we clobber memory, canon the address. |
7afe21cc RK |
6549 | This does nothing when a register is clobbered |
6550 | because we have already invalidated the reg. */ | |
6551 | if (GET_CODE (XEXP (y, 0)) == MEM) | |
9ae8ffe7 | 6552 | canon_reg (XEXP (y, 0), NULL_RTX); |
7afe21cc RK |
6553 | } |
6554 | else if (GET_CODE (y) == USE | |
6555 | && ! (GET_CODE (XEXP (y, 0)) == REG | |
6556 | && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER)) | |
906c4e36 | 6557 | canon_reg (y, NULL_RTX); |
7afe21cc RK |
6558 | else if (GET_CODE (y) == CALL) |
6559 | { | |
7722328e RK |
6560 | /* The result of apply_change_group can be ignored; see |
6561 | canon_reg. */ | |
7afe21cc | 6562 | canon_reg (y, insn); |
77fa0940 | 6563 | apply_change_group (); |
7afe21cc RK |
6564 | fold_rtx (y, insn); |
6565 | } | |
6566 | } | |
6567 | } | |
6568 | else if (GET_CODE (x) == CLOBBER) | |
6569 | { | |
6570 | if (GET_CODE (XEXP (x, 0)) == MEM) | |
9ae8ffe7 | 6571 | canon_reg (XEXP (x, 0), NULL_RTX); |
7afe21cc RK |
6572 | } |
6573 | ||
6574 | /* Canonicalize a USE of a pseudo register or memory location. */ | |
6575 | else if (GET_CODE (x) == USE | |
6576 | && ! (GET_CODE (XEXP (x, 0)) == REG | |
6577 | && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER)) | |
906c4e36 | 6578 | canon_reg (XEXP (x, 0), NULL_RTX); |
7afe21cc RK |
6579 | else if (GET_CODE (x) == CALL) |
6580 | { | |
7722328e | 6581 | /* The result of apply_change_group can be ignored; see canon_reg. */ |
7afe21cc | 6582 | canon_reg (x, insn); |
77fa0940 | 6583 | apply_change_group (); |
7afe21cc RK |
6584 | fold_rtx (x, insn); |
6585 | } | |
6586 | ||
7b3ab05e JW |
6587 | /* Store the equivalent value in SRC_EQV, if different, or if the DEST |
6588 | is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV | |
6589 | is handled specially for this case, and if it isn't set, then there will | |
9faa82d8 | 6590 | be no equivalence for the destination. */ |
92f9aa51 RK |
6591 | if (n_sets == 1 && REG_NOTES (insn) != 0 |
6592 | && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0 | |
7b3ab05e JW |
6593 | && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)) |
6594 | || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART)) | |
92f9aa51 | 6595 | src_eqv = canon_reg (XEXP (tem, 0), NULL_RTX); |
7afe21cc RK |
6596 | |
6597 | /* Canonicalize sources and addresses of destinations. | |
6598 | We do this in a separate pass to avoid problems when a MATCH_DUP is | |
6599 | present in the insn pattern. In that case, we want to ensure that | |
6600 | we don't break the duplicate nature of the pattern. So we will replace | |
6601 | both operands at the same time. Otherwise, we would fail to find an | |
6602 | equivalent substitution in the loop calling validate_change below. | |
7afe21cc RK |
6603 | |
6604 | We used to suppress canonicalization of DEST if it appears in SRC, | |
77fa0940 | 6605 | but we don't do this any more. */ |
7afe21cc RK |
6606 | |
6607 | for (i = 0; i < n_sets; i++) | |
6608 | { | |
6609 | rtx dest = SET_DEST (sets[i].rtl); | |
6610 | rtx src = SET_SRC (sets[i].rtl); | |
6611 | rtx new = canon_reg (src, insn); | |
58873255 | 6612 | int insn_code; |
7afe21cc | 6613 | |
77fa0940 RK |
6614 | if ((GET_CODE (new) == REG && GET_CODE (src) == REG |
6615 | && ((REGNO (new) < FIRST_PSEUDO_REGISTER) | |
6616 | != (REGNO (src) < FIRST_PSEUDO_REGISTER))) | |
58873255 RK |
6617 | || (insn_code = recog_memoized (insn)) < 0 |
6618 | || insn_n_dups[insn_code] > 0) | |
77fa0940 | 6619 | validate_change (insn, &SET_SRC (sets[i].rtl), new, 1); |
7afe21cc RK |
6620 | else |
6621 | SET_SRC (sets[i].rtl) = new; | |
6622 | ||
6623 | if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT) | |
6624 | { | |
6625 | validate_change (insn, &XEXP (dest, 1), | |
77fa0940 | 6626 | canon_reg (XEXP (dest, 1), insn), 1); |
7afe21cc | 6627 | validate_change (insn, &XEXP (dest, 2), |
77fa0940 | 6628 | canon_reg (XEXP (dest, 2), insn), 1); |
7afe21cc RK |
6629 | } |
6630 | ||
6631 | while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART | |
6632 | || GET_CODE (dest) == ZERO_EXTRACT | |
6633 | || GET_CODE (dest) == SIGN_EXTRACT) | |
6634 | dest = XEXP (dest, 0); | |
6635 | ||
6636 | if (GET_CODE (dest) == MEM) | |
6637 | canon_reg (dest, insn); | |
6638 | } | |
6639 | ||
77fa0940 RK |
6640 | /* Now that we have done all the replacements, we can apply the change |
6641 | group and see if they all work. Note that this will cause some | |
6642 | canonicalizations that would have worked individually not to be applied | |
6643 | because some other canonicalization didn't work, but this should not | |
7722328e RK |
6644 | occur often. |
6645 | ||
6646 | The result of apply_change_group can be ignored; see canon_reg. */ | |
77fa0940 RK |
6647 | |
6648 | apply_change_group (); | |
6649 | ||
7afe21cc RK |
6650 | /* Set sets[i].src_elt to the class each source belongs to. |
6651 | Detect assignments from or to volatile things | |
6652 | and set set[i] to zero so they will be ignored | |
6653 | in the rest of this function. | |
6654 | ||
6655 | Nothing in this loop changes the hash table or the register chains. */ | |
6656 | ||
6657 | for (i = 0; i < n_sets; i++) | |
6658 | { | |
6659 | register rtx src, dest; | |
6660 | register rtx src_folded; | |
6661 | register struct table_elt *elt = 0, *p; | |
6662 | enum machine_mode mode; | |
6663 | rtx src_eqv_here; | |
6664 | rtx src_const = 0; | |
6665 | rtx src_related = 0; | |
6666 | struct table_elt *src_const_elt = 0; | |
6667 | int src_cost = 10000, src_eqv_cost = 10000, src_folded_cost = 10000; | |
6668 | int src_related_cost = 10000, src_elt_cost = 10000; | |
6669 | /* Set non-zero if we need to call force_const_mem on with the | |
6670 | contents of src_folded before using it. */ | |
6671 | int src_folded_force_flag = 0; | |
6672 | ||
6673 | dest = SET_DEST (sets[i].rtl); | |
6674 | src = SET_SRC (sets[i].rtl); | |
6675 | ||
6676 | /* If SRC is a constant that has no machine mode, | |
6677 | hash it with the destination's machine mode. | |
6678 | This way we can keep different modes separate. */ | |
6679 | ||
6680 | mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src); | |
6681 | sets[i].mode = mode; | |
6682 | ||
6683 | if (src_eqv) | |
6684 | { | |
6685 | enum machine_mode eqvmode = mode; | |
6686 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
6687 | eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0))); | |
6688 | do_not_record = 0; | |
6689 | hash_arg_in_memory = 0; | |
6690 | hash_arg_in_struct = 0; | |
6691 | src_eqv = fold_rtx (src_eqv, insn); | |
2197a88a | 6692 | src_eqv_hash = HASH (src_eqv, eqvmode); |
7afe21cc RK |
6693 | |
6694 | /* Find the equivalence class for the equivalent expression. */ | |
6695 | ||
6696 | if (!do_not_record) | |
2197a88a | 6697 | src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode); |
7afe21cc RK |
6698 | |
6699 | src_eqv_volatile = do_not_record; | |
6700 | src_eqv_in_memory = hash_arg_in_memory; | |
6701 | src_eqv_in_struct = hash_arg_in_struct; | |
6702 | } | |
6703 | ||
6704 | /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the | |
6705 | value of the INNER register, not the destination. So it is not | |
3826a3da | 6706 | a valid substitution for the source. But save it for later. */ |
7afe21cc RK |
6707 | if (GET_CODE (dest) == STRICT_LOW_PART) |
6708 | src_eqv_here = 0; | |
6709 | else | |
6710 | src_eqv_here = src_eqv; | |
6711 | ||
6712 | /* Simplify and foldable subexpressions in SRC. Then get the fully- | |
6713 | simplified result, which may not necessarily be valid. */ | |
6714 | src_folded = fold_rtx (src, insn); | |
6715 | ||
e6a125a0 RK |
6716 | #if 0 |
6717 | /* ??? This caused bad code to be generated for the m68k port with -O2. | |
6718 | Suppose src is (CONST_INT -1), and that after truncation src_folded | |
6719 | is (CONST_INT 3). Suppose src_folded is then used for src_const. | |
6720 | At the end we will add src and src_const to the same equivalence | |
6721 | class. We now have 3 and -1 on the same equivalence class. This | |
6722 | causes later instructions to be mis-optimized. */ | |
7afe21cc RK |
6723 | /* If storing a constant in a bitfield, pre-truncate the constant |
6724 | so we will be able to record it later. */ | |
6725 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT | |
6726 | || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT) | |
6727 | { | |
6728 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
6729 | ||
6730 | if (GET_CODE (src) == CONST_INT | |
6731 | && GET_CODE (width) == CONST_INT | |
906c4e36 RK |
6732 | && INTVAL (width) < HOST_BITS_PER_WIDE_INT |
6733 | && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width)))) | |
6734 | src_folded | |
6735 | = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1 | |
6736 | << INTVAL (width)) - 1)); | |
7afe21cc | 6737 | } |
e6a125a0 | 6738 | #endif |
7afe21cc RK |
6739 | |
6740 | /* Compute SRC's hash code, and also notice if it | |
6741 | should not be recorded at all. In that case, | |
6742 | prevent any further processing of this assignment. */ | |
6743 | do_not_record = 0; | |
6744 | hash_arg_in_memory = 0; | |
6745 | hash_arg_in_struct = 0; | |
6746 | ||
6747 | sets[i].src = src; | |
2197a88a | 6748 | sets[i].src_hash = HASH (src, mode); |
7afe21cc RK |
6749 | sets[i].src_volatile = do_not_record; |
6750 | sets[i].src_in_memory = hash_arg_in_memory; | |
6751 | sets[i].src_in_struct = hash_arg_in_struct; | |
6752 | ||
50196afa RK |
6753 | /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is |
6754 | a pseudo that is set more than once, do not record SRC. Using | |
6755 | SRC as a replacement for anything else will be incorrect in that | |
6756 | situation. Note that this usually occurs only for stack slots, | |
956d6950 | 6757 | in which case all the RTL would be referring to SRC, so we don't |
50196afa RK |
6758 | lose any optimization opportunities by not having SRC in the |
6759 | hash table. */ | |
6760 | ||
6761 | if (GET_CODE (src) == MEM | |
6762 | && find_reg_note (insn, REG_EQUIV, src) != 0 | |
6763 | && GET_CODE (dest) == REG | |
6764 | && REGNO (dest) >= FIRST_PSEUDO_REGISTER | |
b1f21e0a | 6765 | && REG_N_SETS (REGNO (dest)) != 1) |
50196afa RK |
6766 | sets[i].src_volatile = 1; |
6767 | ||
0dadecf6 RK |
6768 | #if 0 |
6769 | /* It is no longer clear why we used to do this, but it doesn't | |
6770 | appear to still be needed. So let's try without it since this | |
6771 | code hurts cse'ing widened ops. */ | |
7afe21cc RK |
6772 | /* If source is a perverse subreg (such as QI treated as an SI), |
6773 | treat it as volatile. It may do the work of an SI in one context | |
6774 | where the extra bits are not being used, but cannot replace an SI | |
6775 | in general. */ | |
6776 | if (GET_CODE (src) == SUBREG | |
6777 | && (GET_MODE_SIZE (GET_MODE (src)) | |
6778 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))) | |
6779 | sets[i].src_volatile = 1; | |
0dadecf6 | 6780 | #endif |
7afe21cc RK |
6781 | |
6782 | /* Locate all possible equivalent forms for SRC. Try to replace | |
6783 | SRC in the insn with each cheaper equivalent. | |
6784 | ||
6785 | We have the following types of equivalents: SRC itself, a folded | |
6786 | version, a value given in a REG_EQUAL note, or a value related | |
6787 | to a constant. | |
6788 | ||
6789 | Each of these equivalents may be part of an additional class | |
6790 | of equivalents (if more than one is in the table, they must be in | |
6791 | the same class; we check for this). | |
6792 | ||
6793 | If the source is volatile, we don't do any table lookups. | |
6794 | ||
6795 | We note any constant equivalent for possible later use in a | |
6796 | REG_NOTE. */ | |
6797 | ||
6798 | if (!sets[i].src_volatile) | |
2197a88a | 6799 | elt = lookup (src, sets[i].src_hash, mode); |
7afe21cc RK |
6800 | |
6801 | sets[i].src_elt = elt; | |
6802 | ||
6803 | if (elt && src_eqv_here && src_eqv_elt) | |
6804 | { | |
6805 | if (elt->first_same_value != src_eqv_elt->first_same_value) | |
6806 | { | |
6807 | /* The REG_EQUAL is indicating that two formerly distinct | |
6808 | classes are now equivalent. So merge them. */ | |
6809 | merge_equiv_classes (elt, src_eqv_elt); | |
2197a88a RK |
6810 | src_eqv_hash = HASH (src_eqv, elt->mode); |
6811 | src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode); | |
7afe21cc RK |
6812 | } |
6813 | ||
6814 | src_eqv_here = 0; | |
6815 | } | |
6816 | ||
6817 | else if (src_eqv_elt) | |
6818 | elt = src_eqv_elt; | |
6819 | ||
6820 | /* Try to find a constant somewhere and record it in `src_const'. | |
6821 | Record its table element, if any, in `src_const_elt'. Look in | |
6822 | any known equivalences first. (If the constant is not in the | |
2197a88a | 6823 | table, also set `sets[i].src_const_hash'). */ |
7afe21cc RK |
6824 | if (elt) |
6825 | for (p = elt->first_same_value; p; p = p->next_same_value) | |
6826 | if (p->is_const) | |
6827 | { | |
6828 | src_const = p->exp; | |
6829 | src_const_elt = elt; | |
6830 | break; | |
6831 | } | |
6832 | ||
6833 | if (src_const == 0 | |
6834 | && (CONSTANT_P (src_folded) | |
6835 | /* Consider (minus (label_ref L1) (label_ref L2)) as | |
6836 | "constant" here so we will record it. This allows us | |
6837 | to fold switch statements when an ADDR_DIFF_VEC is used. */ | |
6838 | || (GET_CODE (src_folded) == MINUS | |
6839 | && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF | |
6840 | && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF))) | |
6841 | src_const = src_folded, src_const_elt = elt; | |
6842 | else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here)) | |
6843 | src_const = src_eqv_here, src_const_elt = src_eqv_elt; | |
6844 | ||
6845 | /* If we don't know if the constant is in the table, get its | |
6846 | hash code and look it up. */ | |
6847 | if (src_const && src_const_elt == 0) | |
6848 | { | |
2197a88a RK |
6849 | sets[i].src_const_hash = HASH (src_const, mode); |
6850 | src_const_elt = lookup (src_const, sets[i].src_const_hash, mode); | |
7afe21cc RK |
6851 | } |
6852 | ||
6853 | sets[i].src_const = src_const; | |
6854 | sets[i].src_const_elt = src_const_elt; | |
6855 | ||
6856 | /* If the constant and our source are both in the table, mark them as | |
6857 | equivalent. Otherwise, if a constant is in the table but the source | |
6858 | isn't, set ELT to it. */ | |
6859 | if (src_const_elt && elt | |
6860 | && src_const_elt->first_same_value != elt->first_same_value) | |
6861 | merge_equiv_classes (elt, src_const_elt); | |
6862 | else if (src_const_elt && elt == 0) | |
6863 | elt = src_const_elt; | |
6864 | ||
6865 | /* See if there is a register linearly related to a constant | |
6866 | equivalent of SRC. */ | |
6867 | if (src_const | |
6868 | && (GET_CODE (src_const) == CONST | |
6869 | || (src_const_elt && src_const_elt->related_value != 0))) | |
6870 | { | |
6871 | src_related = use_related_value (src_const, src_const_elt); | |
6872 | if (src_related) | |
6873 | { | |
6874 | struct table_elt *src_related_elt | |
6875 | = lookup (src_related, HASH (src_related, mode), mode); | |
6876 | if (src_related_elt && elt) | |
6877 | { | |
6878 | if (elt->first_same_value | |
6879 | != src_related_elt->first_same_value) | |
6880 | /* This can occur when we previously saw a CONST | |
6881 | involving a SYMBOL_REF and then see the SYMBOL_REF | |
6882 | twice. Merge the involved classes. */ | |
6883 | merge_equiv_classes (elt, src_related_elt); | |
6884 | ||
6885 | src_related = 0; | |
6886 | src_related_elt = 0; | |
6887 | } | |
6888 | else if (src_related_elt && elt == 0) | |
6889 | elt = src_related_elt; | |
6890 | } | |
6891 | } | |
6892 | ||
e4600702 RK |
6893 | /* See if we have a CONST_INT that is already in a register in a |
6894 | wider mode. */ | |
6895 | ||
6896 | if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT | |
6897 | && GET_MODE_CLASS (mode) == MODE_INT | |
6898 | && GET_MODE_BITSIZE (mode) < BITS_PER_WORD) | |
6899 | { | |
6900 | enum machine_mode wider_mode; | |
6901 | ||
6902 | for (wider_mode = GET_MODE_WIDER_MODE (mode); | |
6903 | GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD | |
6904 | && src_related == 0; | |
6905 | wider_mode = GET_MODE_WIDER_MODE (wider_mode)) | |
6906 | { | |
6907 | struct table_elt *const_elt | |
6908 | = lookup (src_const, HASH (src_const, wider_mode), wider_mode); | |
6909 | ||
6910 | if (const_elt == 0) | |
6911 | continue; | |
6912 | ||
6913 | for (const_elt = const_elt->first_same_value; | |
6914 | const_elt; const_elt = const_elt->next_same_value) | |
6915 | if (GET_CODE (const_elt->exp) == REG) | |
6916 | { | |
6917 | src_related = gen_lowpart_if_possible (mode, | |
6918 | const_elt->exp); | |
6919 | break; | |
6920 | } | |
6921 | } | |
6922 | } | |
6923 | ||
d45cf215 RS |
6924 | /* Another possibility is that we have an AND with a constant in |
6925 | a mode narrower than a word. If so, it might have been generated | |
6926 | as part of an "if" which would narrow the AND. If we already | |
6927 | have done the AND in a wider mode, we can use a SUBREG of that | |
6928 | value. */ | |
6929 | ||
6930 | if (flag_expensive_optimizations && ! src_related | |
6931 | && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT | |
6932 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD) | |
6933 | { | |
6934 | enum machine_mode tmode; | |
38a448ca | 6935 | rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1)); |
d45cf215 RS |
6936 | |
6937 | for (tmode = GET_MODE_WIDER_MODE (mode); | |
6938 | GET_MODE_SIZE (tmode) <= UNITS_PER_WORD; | |
6939 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
6940 | { | |
6941 | rtx inner = gen_lowpart_if_possible (tmode, XEXP (src, 0)); | |
6942 | struct table_elt *larger_elt; | |
6943 | ||
6944 | if (inner) | |
6945 | { | |
6946 | PUT_MODE (new_and, tmode); | |
6947 | XEXP (new_and, 0) = inner; | |
6948 | larger_elt = lookup (new_and, HASH (new_and, tmode), tmode); | |
6949 | if (larger_elt == 0) | |
6950 | continue; | |
6951 | ||
6952 | for (larger_elt = larger_elt->first_same_value; | |
6953 | larger_elt; larger_elt = larger_elt->next_same_value) | |
6954 | if (GET_CODE (larger_elt->exp) == REG) | |
6955 | { | |
6956 | src_related | |
6957 | = gen_lowpart_if_possible (mode, larger_elt->exp); | |
6958 | break; | |
6959 | } | |
6960 | ||
6961 | if (src_related) | |
6962 | break; | |
6963 | } | |
6964 | } | |
6965 | } | |
7bac1be0 RK |
6966 | |
6967 | #ifdef LOAD_EXTEND_OP | |
6968 | /* See if a MEM has already been loaded with a widening operation; | |
6969 | if it has, we can use a subreg of that. Many CISC machines | |
6970 | also have such operations, but this is only likely to be | |
6971 | beneficial these machines. */ | |
6972 | ||
6973 | if (flag_expensive_optimizations && src_related == 0 | |
6974 | && (GET_MODE_SIZE (mode) < UNITS_PER_WORD) | |
6975 | && GET_MODE_CLASS (mode) == MODE_INT | |
6976 | && GET_CODE (src) == MEM && ! do_not_record | |
6977 | && LOAD_EXTEND_OP (mode) != NIL) | |
6978 | { | |
6979 | enum machine_mode tmode; | |
6980 | ||
6981 | /* Set what we are trying to extend and the operation it might | |
6982 | have been extended with. */ | |
6983 | PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode)); | |
6984 | XEXP (memory_extend_rtx, 0) = src; | |
6985 | ||
6986 | for (tmode = GET_MODE_WIDER_MODE (mode); | |
6987 | GET_MODE_SIZE (tmode) <= UNITS_PER_WORD; | |
6988 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
6989 | { | |
6990 | struct table_elt *larger_elt; | |
6991 | ||
6992 | PUT_MODE (memory_extend_rtx, tmode); | |
6993 | larger_elt = lookup (memory_extend_rtx, | |
6994 | HASH (memory_extend_rtx, tmode), tmode); | |
6995 | if (larger_elt == 0) | |
6996 | continue; | |
6997 | ||
6998 | for (larger_elt = larger_elt->first_same_value; | |
6999 | larger_elt; larger_elt = larger_elt->next_same_value) | |
7000 | if (GET_CODE (larger_elt->exp) == REG) | |
7001 | { | |
7002 | src_related = gen_lowpart_if_possible (mode, | |
7003 | larger_elt->exp); | |
7004 | break; | |
7005 | } | |
7006 | ||
7007 | if (src_related) | |
7008 | break; | |
7009 | } | |
7010 | } | |
7011 | #endif /* LOAD_EXTEND_OP */ | |
7012 | ||
7afe21cc RK |
7013 | if (src == src_folded) |
7014 | src_folded = 0; | |
7015 | ||
7016 | /* At this point, ELT, if non-zero, points to a class of expressions | |
7017 | equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED, | |
7018 | and SRC_RELATED, if non-zero, each contain additional equivalent | |
7019 | expressions. Prune these latter expressions by deleting expressions | |
7020 | already in the equivalence class. | |
7021 | ||
7022 | Check for an equivalent identical to the destination. If found, | |
7023 | this is the preferred equivalent since it will likely lead to | |
7024 | elimination of the insn. Indicate this by placing it in | |
7025 | `src_related'. */ | |
7026 | ||
7027 | if (elt) elt = elt->first_same_value; | |
7028 | for (p = elt; p; p = p->next_same_value) | |
7029 | { | |
7030 | enum rtx_code code = GET_CODE (p->exp); | |
7031 | ||
7032 | /* If the expression is not valid, ignore it. Then we do not | |
7033 | have to check for validity below. In most cases, we can use | |
7034 | `rtx_equal_p', since canonicalization has already been done. */ | |
7035 | if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, 0)) | |
7036 | continue; | |
7037 | ||
5a03c8c4 RK |
7038 | /* Also skip paradoxical subregs, unless that's what we're |
7039 | looking for. */ | |
7040 | if (code == SUBREG | |
7041 | && (GET_MODE_SIZE (GET_MODE (p->exp)) | |
7042 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp)))) | |
7043 | && ! (src != 0 | |
7044 | && GET_CODE (src) == SUBREG | |
7045 | && GET_MODE (src) == GET_MODE (p->exp) | |
7046 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
7047 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp)))))) | |
7048 | continue; | |
7049 | ||
7afe21cc RK |
7050 | if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp)) |
7051 | src = 0; | |
7052 | else if (src_folded && GET_CODE (src_folded) == code | |
7053 | && rtx_equal_p (src_folded, p->exp)) | |
7054 | src_folded = 0; | |
7055 | else if (src_eqv_here && GET_CODE (src_eqv_here) == code | |
7056 | && rtx_equal_p (src_eqv_here, p->exp)) | |
7057 | src_eqv_here = 0; | |
7058 | else if (src_related && GET_CODE (src_related) == code | |
7059 | && rtx_equal_p (src_related, p->exp)) | |
7060 | src_related = 0; | |
7061 | ||
7062 | /* This is the same as the destination of the insns, we want | |
7063 | to prefer it. Copy it to src_related. The code below will | |
7064 | then give it a negative cost. */ | |
7065 | if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest)) | |
7066 | src_related = dest; | |
7067 | ||
7068 | } | |
7069 | ||
7070 | /* Find the cheapest valid equivalent, trying all the available | |
7071 | possibilities. Prefer items not in the hash table to ones | |
7072 | that are when they are equal cost. Note that we can never | |
7073 | worsen an insn as the current contents will also succeed. | |
05c33dd8 | 7074 | If we find an equivalent identical to the destination, use it as best, |
0f41302f | 7075 | since this insn will probably be eliminated in that case. */ |
7afe21cc RK |
7076 | if (src) |
7077 | { | |
7078 | if (rtx_equal_p (src, dest)) | |
7079 | src_cost = -1; | |
7080 | else | |
7081 | src_cost = COST (src); | |
7082 | } | |
7083 | ||
7084 | if (src_eqv_here) | |
7085 | { | |
7086 | if (rtx_equal_p (src_eqv_here, dest)) | |
7087 | src_eqv_cost = -1; | |
7088 | else | |
7089 | src_eqv_cost = COST (src_eqv_here); | |
7090 | } | |
7091 | ||
7092 | if (src_folded) | |
7093 | { | |
7094 | if (rtx_equal_p (src_folded, dest)) | |
7095 | src_folded_cost = -1; | |
7096 | else | |
7097 | src_folded_cost = COST (src_folded); | |
7098 | } | |
7099 | ||
7100 | if (src_related) | |
7101 | { | |
7102 | if (rtx_equal_p (src_related, dest)) | |
7103 | src_related_cost = -1; | |
7104 | else | |
7105 | src_related_cost = COST (src_related); | |
7106 | } | |
7107 | ||
7108 | /* If this was an indirect jump insn, a known label will really be | |
7109 | cheaper even though it looks more expensive. */ | |
7110 | if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF) | |
7111 | src_folded = src_const, src_folded_cost = -1; | |
7112 | ||
7113 | /* Terminate loop when replacement made. This must terminate since | |
7114 | the current contents will be tested and will always be valid. */ | |
7115 | while (1) | |
7116 | { | |
7bd8b2a8 | 7117 | rtx trial, old_src; |
7afe21cc RK |
7118 | |
7119 | /* Skip invalid entries. */ | |
7120 | while (elt && GET_CODE (elt->exp) != REG | |
7121 | && ! exp_equiv_p (elt->exp, elt->exp, 1, 0)) | |
7122 | elt = elt->next_same_value; | |
5a03c8c4 RK |
7123 | |
7124 | /* A paradoxical subreg would be bad here: it'll be the right | |
7125 | size, but later may be adjusted so that the upper bits aren't | |
7126 | what we want. So reject it. */ | |
7127 | if (elt != 0 | |
7128 | && GET_CODE (elt->exp) == SUBREG | |
7129 | && (GET_MODE_SIZE (GET_MODE (elt->exp)) | |
7130 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp)))) | |
7131 | /* It is okay, though, if the rtx we're trying to match | |
7132 | will ignore any of the bits we can't predict. */ | |
7133 | && ! (src != 0 | |
7134 | && GET_CODE (src) == SUBREG | |
7135 | && GET_MODE (src) == GET_MODE (elt->exp) | |
7136 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
7137 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp)))))) | |
7138 | { | |
7139 | elt = elt->next_same_value; | |
7140 | continue; | |
7141 | } | |
7afe21cc RK |
7142 | |
7143 | if (elt) src_elt_cost = elt->cost; | |
7144 | ||
7145 | /* Find cheapest and skip it for the next time. For items | |
7146 | of equal cost, use this order: | |
7147 | src_folded, src, src_eqv, src_related and hash table entry. */ | |
7148 | if (src_folded_cost <= src_cost | |
7149 | && src_folded_cost <= src_eqv_cost | |
7150 | && src_folded_cost <= src_related_cost | |
7151 | && src_folded_cost <= src_elt_cost) | |
7152 | { | |
7153 | trial = src_folded, src_folded_cost = 10000; | |
7154 | if (src_folded_force_flag) | |
7155 | trial = force_const_mem (mode, trial); | |
7156 | } | |
7157 | else if (src_cost <= src_eqv_cost | |
7158 | && src_cost <= src_related_cost | |
7159 | && src_cost <= src_elt_cost) | |
7160 | trial = src, src_cost = 10000; | |
7161 | else if (src_eqv_cost <= src_related_cost | |
7162 | && src_eqv_cost <= src_elt_cost) | |
0af62b41 | 7163 | trial = copy_rtx (src_eqv_here), src_eqv_cost = 10000; |
7afe21cc | 7164 | else if (src_related_cost <= src_elt_cost) |
0af62b41 | 7165 | trial = copy_rtx (src_related), src_related_cost = 10000; |
7afe21cc RK |
7166 | else |
7167 | { | |
05c33dd8 | 7168 | trial = copy_rtx (elt->exp); |
7afe21cc RK |
7169 | elt = elt->next_same_value; |
7170 | src_elt_cost = 10000; | |
7171 | } | |
7172 | ||
7173 | /* We don't normally have an insn matching (set (pc) (pc)), so | |
7174 | check for this separately here. We will delete such an | |
7175 | insn below. | |
7176 | ||
7177 | Tablejump insns contain a USE of the table, so simply replacing | |
7178 | the operand with the constant won't match. This is simply an | |
7179 | unconditional branch, however, and is therefore valid. Just | |
7180 | insert the substitution here and we will delete and re-emit | |
7181 | the insn later. */ | |
7182 | ||
7bd8b2a8 JL |
7183 | /* Keep track of the original SET_SRC so that we can fix notes |
7184 | on libcall instructions. */ | |
7185 | old_src = SET_SRC (sets[i].rtl); | |
7186 | ||
7afe21cc RK |
7187 | if (n_sets == 1 && dest == pc_rtx |
7188 | && (trial == pc_rtx | |
7189 | || (GET_CODE (trial) == LABEL_REF | |
7190 | && ! condjump_p (insn)))) | |
7191 | { | |
7192 | /* If TRIAL is a label in front of a jump table, we are | |
7193 | really falling through the switch (this is how casesi | |
7194 | insns work), so we must branch around the table. */ | |
7195 | if (GET_CODE (trial) == CODE_LABEL | |
7196 | && NEXT_INSN (trial) != 0 | |
7197 | && GET_CODE (NEXT_INSN (trial)) == JUMP_INSN | |
7198 | && (GET_CODE (PATTERN (NEXT_INSN (trial))) == ADDR_DIFF_VEC | |
7199 | || GET_CODE (PATTERN (NEXT_INSN (trial))) == ADDR_VEC)) | |
7200 | ||
38a448ca | 7201 | trial = gen_rtx_LABEL_REF (Pmode, get_label_after (trial)); |
7afe21cc RK |
7202 | |
7203 | SET_SRC (sets[i].rtl) = trial; | |
44333223 | 7204 | cse_jumps_altered = 1; |
7afe21cc RK |
7205 | break; |
7206 | } | |
7207 | ||
7208 | /* Look for a substitution that makes a valid insn. */ | |
7209 | else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0)) | |
05c33dd8 | 7210 | { |
7bd8b2a8 JL |
7211 | /* If we just made a substitution inside a libcall, then we |
7212 | need to make the same substitution in any notes attached | |
7213 | to the RETVAL insn. */ | |
1ed0205e VM |
7214 | if (libcall_insn |
7215 | && (GET_CODE (old_src) == REG | |
7216 | || GET_CODE (old_src) == SUBREG | |
7217 | || GET_CODE (old_src) == MEM)) | |
7bd8b2a8 JL |
7218 | replace_rtx (REG_NOTES (libcall_insn), old_src, |
7219 | canon_reg (SET_SRC (sets[i].rtl), insn)); | |
7220 | ||
7722328e RK |
7221 | /* The result of apply_change_group can be ignored; see |
7222 | canon_reg. */ | |
7223 | ||
7224 | validate_change (insn, &SET_SRC (sets[i].rtl), | |
7225 | canon_reg (SET_SRC (sets[i].rtl), insn), | |
7226 | 1); | |
6702af89 | 7227 | apply_change_group (); |
05c33dd8 RK |
7228 | break; |
7229 | } | |
7afe21cc RK |
7230 | |
7231 | /* If we previously found constant pool entries for | |
7232 | constants and this is a constant, try making a | |
7233 | pool entry. Put it in src_folded unless we already have done | |
7234 | this since that is where it likely came from. */ | |
7235 | ||
7236 | else if (constant_pool_entries_cost | |
7237 | && CONSTANT_P (trial) | |
1bbd065b RK |
7238 | && ! (GET_CODE (trial) == CONST |
7239 | && GET_CODE (XEXP (trial, 0)) == TRUNCATE) | |
7240 | && (src_folded == 0 | |
7241 | || (GET_CODE (src_folded) != MEM | |
7242 | && ! src_folded_force_flag)) | |
9ae8ffe7 JL |
7243 | && GET_MODE_CLASS (mode) != MODE_CC |
7244 | && mode != VOIDmode) | |
7afe21cc RK |
7245 | { |
7246 | src_folded_force_flag = 1; | |
7247 | src_folded = trial; | |
7248 | src_folded_cost = constant_pool_entries_cost; | |
7249 | } | |
7250 | } | |
7251 | ||
7252 | src = SET_SRC (sets[i].rtl); | |
7253 | ||
7254 | /* In general, it is good to have a SET with SET_SRC == SET_DEST. | |
7255 | However, there is an important exception: If both are registers | |
7256 | that are not the head of their equivalence class, replace SET_SRC | |
7257 | with the head of the class. If we do not do this, we will have | |
7258 | both registers live over a portion of the basic block. This way, | |
7259 | their lifetimes will likely abut instead of overlapping. */ | |
7260 | if (GET_CODE (dest) == REG | |
7261 | && REGNO_QTY_VALID_P (REGNO (dest)) | |
30f72379 MM |
7262 | && qty_mode[REG_QTY (REGNO (dest))] == GET_MODE (dest) |
7263 | && qty_first_reg[REG_QTY (REGNO (dest))] != REGNO (dest) | |
7afe21cc RK |
7264 | && GET_CODE (src) == REG && REGNO (src) == REGNO (dest) |
7265 | /* Don't do this if the original insn had a hard reg as | |
7266 | SET_SRC. */ | |
7267 | && (GET_CODE (sets[i].src) != REG | |
7268 | || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)) | |
7269 | /* We can't call canon_reg here because it won't do anything if | |
7270 | SRC is a hard register. */ | |
7271 | { | |
30f72379 | 7272 | int first = qty_first_reg[REG_QTY (REGNO (src))]; |
759bd8b7 R |
7273 | rtx new_src |
7274 | = (first >= FIRST_PSEUDO_REGISTER | |
7275 | ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first)); | |
7276 | ||
7277 | /* We must use validate-change even for this, because this | |
7278 | might be a special no-op instruction, suitable only to | |
7279 | tag notes onto. */ | |
7280 | if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0)) | |
7281 | { | |
7282 | src = new_src; | |
7283 | /* If we had a constant that is cheaper than what we are now | |
7284 | setting SRC to, use that constant. We ignored it when we | |
7285 | thought we could make this into a no-op. */ | |
7286 | if (src_const && COST (src_const) < COST (src) | |
7287 | && validate_change (insn, &SET_SRC (sets[i].rtl), src_const, | |
7288 | 0)) | |
7289 | src = src_const; | |
7290 | } | |
7afe21cc RK |
7291 | } |
7292 | ||
7293 | /* If we made a change, recompute SRC values. */ | |
7294 | if (src != sets[i].src) | |
7295 | { | |
7296 | do_not_record = 0; | |
7297 | hash_arg_in_memory = 0; | |
7298 | hash_arg_in_struct = 0; | |
7299 | sets[i].src = src; | |
2197a88a | 7300 | sets[i].src_hash = HASH (src, mode); |
7afe21cc RK |
7301 | sets[i].src_volatile = do_not_record; |
7302 | sets[i].src_in_memory = hash_arg_in_memory; | |
7303 | sets[i].src_in_struct = hash_arg_in_struct; | |
2197a88a | 7304 | sets[i].src_elt = lookup (src, sets[i].src_hash, mode); |
7afe21cc RK |
7305 | } |
7306 | ||
7307 | /* If this is a single SET, we are setting a register, and we have an | |
7308 | equivalent constant, we want to add a REG_NOTE. We don't want | |
7309 | to write a REG_EQUAL note for a constant pseudo since verifying that | |
d45cf215 | 7310 | that pseudo hasn't been eliminated is a pain. Such a note also |
ac7ef8d5 FS |
7311 | won't help anything. |
7312 | ||
7313 | Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF))) | |
7314 | which can be created for a reference to a compile time computable | |
7315 | entry in a jump table. */ | |
7316 | ||
7afe21cc | 7317 | if (n_sets == 1 && src_const && GET_CODE (dest) == REG |
ac7ef8d5 FS |
7318 | && GET_CODE (src_const) != REG |
7319 | && ! (GET_CODE (src_const) == CONST | |
7320 | && GET_CODE (XEXP (src_const, 0)) == MINUS | |
7321 | && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF | |
7322 | && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)) | |
7afe21cc | 7323 | { |
92f9aa51 | 7324 | tem = find_reg_note (insn, REG_EQUAL, NULL_RTX); |
7afe21cc RK |
7325 | |
7326 | /* Record the actual constant value in a REG_EQUAL note, making | |
7327 | a new one if one does not already exist. */ | |
7328 | if (tem) | |
7329 | XEXP (tem, 0) = src_const; | |
7330 | else | |
38a448ca RH |
7331 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, |
7332 | src_const, REG_NOTES (insn)); | |
7afe21cc RK |
7333 | |
7334 | /* If storing a constant value in a register that | |
7335 | previously held the constant value 0, | |
7336 | record this fact with a REG_WAS_0 note on this insn. | |
7337 | ||
7338 | Note that the *register* is required to have previously held 0, | |
7339 | not just any register in the quantity and we must point to the | |
7340 | insn that set that register to zero. | |
7341 | ||
7342 | Rather than track each register individually, we just see if | |
7343 | the last set for this quantity was for this register. */ | |
7344 | ||
7345 | if (REGNO_QTY_VALID_P (REGNO (dest)) | |
30f72379 | 7346 | && qty_const[REG_QTY (REGNO (dest))] == const0_rtx) |
7afe21cc RK |
7347 | { |
7348 | /* See if we previously had a REG_WAS_0 note. */ | |
906c4e36 | 7349 | rtx note = find_reg_note (insn, REG_WAS_0, NULL_RTX); |
30f72379 | 7350 | rtx const_insn = qty_const_insn[REG_QTY (REGNO (dest))]; |
7afe21cc RK |
7351 | |
7352 | if ((tem = single_set (const_insn)) != 0 | |
7353 | && rtx_equal_p (SET_DEST (tem), dest)) | |
7354 | { | |
7355 | if (note) | |
7356 | XEXP (note, 0) = const_insn; | |
7357 | else | |
38a448ca RH |
7358 | REG_NOTES (insn) = gen_rtx_INSN_LIST (REG_WAS_0, |
7359 | const_insn, | |
7360 | REG_NOTES (insn)); | |
7afe21cc RK |
7361 | } |
7362 | } | |
7363 | } | |
7364 | ||
7365 | /* Now deal with the destination. */ | |
7366 | do_not_record = 0; | |
7367 | sets[i].inner_dest_loc = &SET_DEST (sets[0].rtl); | |
7368 | ||
7369 | /* Look within any SIGN_EXTRACT or ZERO_EXTRACT | |
7370 | to the MEM or REG within it. */ | |
7371 | while (GET_CODE (dest) == SIGN_EXTRACT | |
7372 | || GET_CODE (dest) == ZERO_EXTRACT | |
7373 | || GET_CODE (dest) == SUBREG | |
7374 | || GET_CODE (dest) == STRICT_LOW_PART) | |
7375 | { | |
7376 | sets[i].inner_dest_loc = &XEXP (dest, 0); | |
7377 | dest = XEXP (dest, 0); | |
7378 | } | |
7379 | ||
7380 | sets[i].inner_dest = dest; | |
7381 | ||
7382 | if (GET_CODE (dest) == MEM) | |
7383 | { | |
9ae8ffe7 JL |
7384 | #ifdef PUSH_ROUNDING |
7385 | /* Stack pushes invalidate the stack pointer. */ | |
7386 | rtx addr = XEXP (dest, 0); | |
7387 | if ((GET_CODE (addr) == PRE_DEC || GET_CODE (addr) == PRE_INC | |
7388 | || GET_CODE (addr) == POST_DEC || GET_CODE (addr) == POST_INC) | |
7389 | && XEXP (addr, 0) == stack_pointer_rtx) | |
7390 | invalidate (stack_pointer_rtx, Pmode); | |
7391 | #endif | |
7afe21cc | 7392 | dest = fold_rtx (dest, insn); |
7afe21cc RK |
7393 | } |
7394 | ||
7395 | /* Compute the hash code of the destination now, | |
7396 | before the effects of this instruction are recorded, | |
7397 | since the register values used in the address computation | |
7398 | are those before this instruction. */ | |
2197a88a | 7399 | sets[i].dest_hash = HASH (dest, mode); |
7afe21cc RK |
7400 | |
7401 | /* Don't enter a bit-field in the hash table | |
7402 | because the value in it after the store | |
7403 | may not equal what was stored, due to truncation. */ | |
7404 | ||
7405 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT | |
7406 | || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT) | |
7407 | { | |
7408 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
7409 | ||
7410 | if (src_const != 0 && GET_CODE (src_const) == CONST_INT | |
7411 | && GET_CODE (width) == CONST_INT | |
906c4e36 RK |
7412 | && INTVAL (width) < HOST_BITS_PER_WIDE_INT |
7413 | && ! (INTVAL (src_const) | |
7414 | & ((HOST_WIDE_INT) (-1) << INTVAL (width)))) | |
7afe21cc RK |
7415 | /* Exception: if the value is constant, |
7416 | and it won't be truncated, record it. */ | |
7417 | ; | |
7418 | else | |
7419 | { | |
7420 | /* This is chosen so that the destination will be invalidated | |
7421 | but no new value will be recorded. | |
7422 | We must invalidate because sometimes constant | |
7423 | values can be recorded for bitfields. */ | |
7424 | sets[i].src_elt = 0; | |
7425 | sets[i].src_volatile = 1; | |
7426 | src_eqv = 0; | |
7427 | src_eqv_elt = 0; | |
7428 | } | |
7429 | } | |
7430 | ||
7431 | /* If only one set in a JUMP_INSN and it is now a no-op, we can delete | |
7432 | the insn. */ | |
7433 | else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx) | |
7434 | { | |
7435 | PUT_CODE (insn, NOTE); | |
7436 | NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED; | |
7437 | NOTE_SOURCE_FILE (insn) = 0; | |
7438 | cse_jumps_altered = 1; | |
7439 | /* One less use of the label this insn used to jump to. */ | |
85c3ba60 JL |
7440 | if (JUMP_LABEL (insn) != 0) |
7441 | --LABEL_NUSES (JUMP_LABEL (insn)); | |
7afe21cc RK |
7442 | /* No more processing for this set. */ |
7443 | sets[i].rtl = 0; | |
7444 | } | |
7445 | ||
7446 | /* If this SET is now setting PC to a label, we know it used to | |
7447 | be a conditional or computed branch. So we see if we can follow | |
7448 | it. If it was a computed branch, delete it and re-emit. */ | |
7449 | else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF) | |
7450 | { | |
7451 | rtx p; | |
7452 | ||
7453 | /* If this is not in the format for a simple branch and | |
7454 | we are the only SET in it, re-emit it. */ | |
7455 | if (! simplejump_p (insn) && n_sets == 1) | |
7456 | { | |
7457 | rtx new = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn); | |
7458 | JUMP_LABEL (new) = XEXP (src, 0); | |
7459 | LABEL_NUSES (XEXP (src, 0))++; | |
7460 | delete_insn (insn); | |
7461 | insn = new; | |
7462 | } | |
31dcf83f RS |
7463 | else |
7464 | /* Otherwise, force rerecognition, since it probably had | |
7465 | a different pattern before. | |
7466 | This shouldn't really be necessary, since whatever | |
7467 | changed the source value above should have done this. | |
7468 | Until the right place is found, might as well do this here. */ | |
7469 | INSN_CODE (insn) = -1; | |
7afe21cc RK |
7470 | |
7471 | /* Now that we've converted this jump to an unconditional jump, | |
7472 | there is dead code after it. Delete the dead code until we | |
7473 | reach a BARRIER, the end of the function, or a label. Do | |
7474 | not delete NOTEs except for NOTE_INSN_DELETED since later | |
7475 | phases assume these notes are retained. */ | |
7476 | ||
7477 | p = insn; | |
7478 | ||
7479 | while (NEXT_INSN (p) != 0 | |
7480 | && GET_CODE (NEXT_INSN (p)) != BARRIER | |
7481 | && GET_CODE (NEXT_INSN (p)) != CODE_LABEL) | |
7482 | { | |
7483 | if (GET_CODE (NEXT_INSN (p)) != NOTE | |
7484 | || NOTE_LINE_NUMBER (NEXT_INSN (p)) == NOTE_INSN_DELETED) | |
7485 | delete_insn (NEXT_INSN (p)); | |
7486 | else | |
7487 | p = NEXT_INSN (p); | |
7488 | } | |
7489 | ||
7490 | /* If we don't have a BARRIER immediately after INSN, put one there. | |
7491 | Much code assumes that there are no NOTEs between a JUMP_INSN and | |
7492 | BARRIER. */ | |
7493 | ||
7494 | if (NEXT_INSN (insn) == 0 | |
7495 | || GET_CODE (NEXT_INSN (insn)) != BARRIER) | |
783e5bca | 7496 | emit_barrier_before (NEXT_INSN (insn)); |
7afe21cc RK |
7497 | |
7498 | /* We might have two BARRIERs separated by notes. Delete the second | |
7499 | one if so. */ | |
7500 | ||
538b78e7 RS |
7501 | if (p != insn && NEXT_INSN (p) != 0 |
7502 | && GET_CODE (NEXT_INSN (p)) == BARRIER) | |
7afe21cc RK |
7503 | delete_insn (NEXT_INSN (p)); |
7504 | ||
7505 | cse_jumps_altered = 1; | |
7506 | sets[i].rtl = 0; | |
7507 | } | |
7508 | ||
c2a47e48 RK |
7509 | /* If destination is volatile, invalidate it and then do no further |
7510 | processing for this assignment. */ | |
7afe21cc RK |
7511 | |
7512 | else if (do_not_record) | |
c2a47e48 RK |
7513 | { |
7514 | if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG | |
7515 | || GET_CODE (dest) == MEM) | |
bb4034b3 | 7516 | invalidate (dest, VOIDmode); |
2708da92 RS |
7517 | else if (GET_CODE (dest) == STRICT_LOW_PART |
7518 | || GET_CODE (dest) == ZERO_EXTRACT) | |
bb4034b3 | 7519 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
c2a47e48 RK |
7520 | sets[i].rtl = 0; |
7521 | } | |
7afe21cc RK |
7522 | |
7523 | if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl)) | |
2197a88a | 7524 | sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode); |
7afe21cc RK |
7525 | |
7526 | #ifdef HAVE_cc0 | |
7527 | /* If setting CC0, record what it was set to, or a constant, if it | |
7528 | is equivalent to a constant. If it is being set to a floating-point | |
7529 | value, make a COMPARE with the appropriate constant of 0. If we | |
7530 | don't do this, later code can interpret this as a test against | |
7531 | const0_rtx, which can cause problems if we try to put it into an | |
7532 | insn as a floating-point operand. */ | |
7533 | if (dest == cc0_rtx) | |
7534 | { | |
7535 | this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src; | |
7536 | this_insn_cc0_mode = mode; | |
cbf6a543 | 7537 | if (FLOAT_MODE_P (mode)) |
38a448ca RH |
7538 | this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0, |
7539 | CONST0_RTX (mode)); | |
7afe21cc RK |
7540 | } |
7541 | #endif | |
7542 | } | |
7543 | ||
7544 | /* Now enter all non-volatile source expressions in the hash table | |
7545 | if they are not already present. | |
7546 | Record their equivalence classes in src_elt. | |
7547 | This way we can insert the corresponding destinations into | |
7548 | the same classes even if the actual sources are no longer in them | |
7549 | (having been invalidated). */ | |
7550 | ||
7551 | if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile | |
7552 | && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl))) | |
7553 | { | |
7554 | register struct table_elt *elt; | |
7555 | register struct table_elt *classp = sets[0].src_elt; | |
7556 | rtx dest = SET_DEST (sets[0].rtl); | |
7557 | enum machine_mode eqvmode = GET_MODE (dest); | |
7558 | ||
7559 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
7560 | { | |
7561 | eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0))); | |
7562 | classp = 0; | |
7563 | } | |
7564 | if (insert_regs (src_eqv, classp, 0)) | |
8ae2b8f6 JW |
7565 | { |
7566 | rehash_using_reg (src_eqv); | |
7567 | src_eqv_hash = HASH (src_eqv, eqvmode); | |
7568 | } | |
2197a88a | 7569 | elt = insert (src_eqv, classp, src_eqv_hash, eqvmode); |
7afe21cc RK |
7570 | elt->in_memory = src_eqv_in_memory; |
7571 | elt->in_struct = src_eqv_in_struct; | |
7572 | src_eqv_elt = elt; | |
f7911249 JW |
7573 | |
7574 | /* Check to see if src_eqv_elt is the same as a set source which | |
7575 | does not yet have an elt, and if so set the elt of the set source | |
7576 | to src_eqv_elt. */ | |
7577 | for (i = 0; i < n_sets; i++) | |
7578 | if (sets[i].rtl && sets[i].src_elt == 0 | |
7579 | && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv)) | |
7580 | sets[i].src_elt = src_eqv_elt; | |
7afe21cc RK |
7581 | } |
7582 | ||
7583 | for (i = 0; i < n_sets; i++) | |
7584 | if (sets[i].rtl && ! sets[i].src_volatile | |
7585 | && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl))) | |
7586 | { | |
7587 | if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART) | |
7588 | { | |
7589 | /* REG_EQUAL in setting a STRICT_LOW_PART | |
7590 | gives an equivalent for the entire destination register, | |
7591 | not just for the subreg being stored in now. | |
7592 | This is a more interesting equivalence, so we arrange later | |
7593 | to treat the entire reg as the destination. */ | |
7594 | sets[i].src_elt = src_eqv_elt; | |
2197a88a | 7595 | sets[i].src_hash = src_eqv_hash; |
7afe21cc RK |
7596 | } |
7597 | else | |
7598 | { | |
7599 | /* Insert source and constant equivalent into hash table, if not | |
7600 | already present. */ | |
7601 | register struct table_elt *classp = src_eqv_elt; | |
7602 | register rtx src = sets[i].src; | |
7603 | register rtx dest = SET_DEST (sets[i].rtl); | |
7604 | enum machine_mode mode | |
7605 | = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src); | |
7606 | ||
7607 | if (sets[i].src_elt == 0) | |
7608 | { | |
7609 | register struct table_elt *elt; | |
7610 | ||
7611 | /* Note that these insert_regs calls cannot remove | |
7612 | any of the src_elt's, because they would have failed to | |
7613 | match if not still valid. */ | |
7614 | if (insert_regs (src, classp, 0)) | |
8ae2b8f6 JW |
7615 | { |
7616 | rehash_using_reg (src); | |
7617 | sets[i].src_hash = HASH (src, mode); | |
7618 | } | |
2197a88a | 7619 | elt = insert (src, classp, sets[i].src_hash, mode); |
7afe21cc RK |
7620 | elt->in_memory = sets[i].src_in_memory; |
7621 | elt->in_struct = sets[i].src_in_struct; | |
7622 | sets[i].src_elt = classp = elt; | |
7623 | } | |
7624 | ||
7625 | if (sets[i].src_const && sets[i].src_const_elt == 0 | |
7626 | && src != sets[i].src_const | |
7627 | && ! rtx_equal_p (sets[i].src_const, src)) | |
7628 | sets[i].src_elt = insert (sets[i].src_const, classp, | |
2197a88a | 7629 | sets[i].src_const_hash, mode); |
7afe21cc RK |
7630 | } |
7631 | } | |
7632 | else if (sets[i].src_elt == 0) | |
7633 | /* If we did not insert the source into the hash table (e.g., it was | |
7634 | volatile), note the equivalence class for the REG_EQUAL value, if any, | |
7635 | so that the destination goes into that class. */ | |
7636 | sets[i].src_elt = src_eqv_elt; | |
7637 | ||
9ae8ffe7 | 7638 | invalidate_from_clobbers (x); |
77fa0940 RK |
7639 | |
7640 | /* Some registers are invalidated by subroutine calls. Memory is | |
7641 | invalidated by non-constant calls. */ | |
7642 | ||
7afe21cc RK |
7643 | if (GET_CODE (insn) == CALL_INSN) |
7644 | { | |
77fa0940 | 7645 | if (! CONST_CALL_P (insn)) |
9ae8ffe7 | 7646 | invalidate_memory (); |
7afe21cc RK |
7647 | invalidate_for_call (); |
7648 | } | |
7649 | ||
7650 | /* Now invalidate everything set by this instruction. | |
7651 | If a SUBREG or other funny destination is being set, | |
7652 | sets[i].rtl is still nonzero, so here we invalidate the reg | |
7653 | a part of which is being set. */ | |
7654 | ||
7655 | for (i = 0; i < n_sets; i++) | |
7656 | if (sets[i].rtl) | |
7657 | { | |
bb4034b3 JW |
7658 | /* We can't use the inner dest, because the mode associated with |
7659 | a ZERO_EXTRACT is significant. */ | |
7660 | register rtx dest = SET_DEST (sets[i].rtl); | |
7afe21cc RK |
7661 | |
7662 | /* Needed for registers to remove the register from its | |
7663 | previous quantity's chain. | |
7664 | Needed for memory if this is a nonvarying address, unless | |
7665 | we have just done an invalidate_memory that covers even those. */ | |
7666 | if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG | |
9ae8ffe7 | 7667 | || GET_CODE (dest) == MEM) |
bb4034b3 | 7668 | invalidate (dest, VOIDmode); |
2708da92 RS |
7669 | else if (GET_CODE (dest) == STRICT_LOW_PART |
7670 | || GET_CODE (dest) == ZERO_EXTRACT) | |
bb4034b3 | 7671 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
7afe21cc RK |
7672 | } |
7673 | ||
01e752d3 JL |
7674 | /* A volatile ASM invalidates everything. */ |
7675 | if (GET_CODE (insn) == INSN | |
7676 | && GET_CODE (PATTERN (insn)) == ASM_OPERANDS | |
7677 | && MEM_VOLATILE_P (PATTERN (insn))) | |
7678 | flush_hash_table (); | |
7679 | ||
7afe21cc RK |
7680 | /* Make sure registers mentioned in destinations |
7681 | are safe for use in an expression to be inserted. | |
7682 | This removes from the hash table | |
7683 | any invalid entry that refers to one of these registers. | |
7684 | ||
7685 | We don't care about the return value from mention_regs because | |
7686 | we are going to hash the SET_DEST values unconditionally. */ | |
7687 | ||
7688 | for (i = 0; i < n_sets; i++) | |
34c73909 R |
7689 | { |
7690 | if (sets[i].rtl) | |
7691 | { | |
7692 | rtx x = SET_DEST (sets[i].rtl); | |
7693 | ||
7694 | if (GET_CODE (x) != REG) | |
7695 | mention_regs (x); | |
7696 | else | |
7697 | { | |
7698 | /* We used to rely on all references to a register becoming | |
7699 | inaccessible when a register changes to a new quantity, | |
7700 | since that changes the hash code. However, that is not | |
7701 | safe, since after NBUCKETS new quantities we get a | |
7702 | hash 'collision' of a register with its own invalid | |
7703 | entries. And since SUBREGs have been changed not to | |
7704 | change their hash code with the hash code of the register, | |
7705 | it wouldn't work any longer at all. So we have to check | |
7706 | for any invalid references lying around now. | |
7707 | This code is similar to the REG case in mention_regs, | |
7708 | but it knows that reg_tick has been incremented, and | |
7709 | it leaves reg_in_table as -1 . */ | |
7710 | register int regno = REGNO (x); | |
7711 | register int endregno | |
7712 | = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1 | |
7713 | : HARD_REGNO_NREGS (regno, GET_MODE (x))); | |
7714 | int i; | |
7715 | ||
7716 | for (i = regno; i < endregno; i++) | |
7717 | { | |
30f72379 | 7718 | if (REG_IN_TABLE (i) >= 0) |
34c73909 R |
7719 | { |
7720 | remove_invalid_refs (i); | |
30f72379 | 7721 | REG_IN_TABLE (i) = -1; |
34c73909 R |
7722 | } |
7723 | } | |
7724 | } | |
7725 | } | |
7726 | } | |
7afe21cc RK |
7727 | |
7728 | /* We may have just removed some of the src_elt's from the hash table. | |
7729 | So replace each one with the current head of the same class. */ | |
7730 | ||
7731 | for (i = 0; i < n_sets; i++) | |
7732 | if (sets[i].rtl) | |
7733 | { | |
7734 | if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0) | |
7735 | /* If elt was removed, find current head of same class, | |
7736 | or 0 if nothing remains of that class. */ | |
7737 | { | |
7738 | register struct table_elt *elt = sets[i].src_elt; | |
7739 | ||
7740 | while (elt && elt->prev_same_value) | |
7741 | elt = elt->prev_same_value; | |
7742 | ||
7743 | while (elt && elt->first_same_value == 0) | |
7744 | elt = elt->next_same_value; | |
7745 | sets[i].src_elt = elt ? elt->first_same_value : 0; | |
7746 | } | |
7747 | } | |
7748 | ||
7749 | /* Now insert the destinations into their equivalence classes. */ | |
7750 | ||
7751 | for (i = 0; i < n_sets; i++) | |
7752 | if (sets[i].rtl) | |
7753 | { | |
7754 | register rtx dest = SET_DEST (sets[i].rtl); | |
9de2c71a | 7755 | rtx inner_dest = sets[i].inner_dest; |
7afe21cc RK |
7756 | register struct table_elt *elt; |
7757 | ||
7758 | /* Don't record value if we are not supposed to risk allocating | |
7759 | floating-point values in registers that might be wider than | |
7760 | memory. */ | |
7761 | if ((flag_float_store | |
7762 | && GET_CODE (dest) == MEM | |
cbf6a543 | 7763 | && FLOAT_MODE_P (GET_MODE (dest))) |
bc4ddc77 JW |
7764 | /* Don't record BLKmode values, because we don't know the |
7765 | size of it, and can't be sure that other BLKmode values | |
7766 | have the same or smaller size. */ | |
7767 | || GET_MODE (dest) == BLKmode | |
7afe21cc RK |
7768 | /* Don't record values of destinations set inside a libcall block |
7769 | since we might delete the libcall. Things should have been set | |
7770 | up so we won't want to reuse such a value, but we play it safe | |
7771 | here. */ | |
7bd8b2a8 | 7772 | || libcall_insn |
7afe21cc RK |
7773 | /* If we didn't put a REG_EQUAL value or a source into the hash |
7774 | table, there is no point is recording DEST. */ | |
1a8e9a8e RK |
7775 | || sets[i].src_elt == 0 |
7776 | /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND | |
7777 | or SIGN_EXTEND, don't record DEST since it can cause | |
7778 | some tracking to be wrong. | |
7779 | ||
7780 | ??? Think about this more later. */ | |
7781 | || (GET_CODE (dest) == SUBREG | |
7782 | && (GET_MODE_SIZE (GET_MODE (dest)) | |
7783 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))) | |
7784 | && (GET_CODE (sets[i].src) == SIGN_EXTEND | |
7785 | || GET_CODE (sets[i].src) == ZERO_EXTEND))) | |
7afe21cc RK |
7786 | continue; |
7787 | ||
7788 | /* STRICT_LOW_PART isn't part of the value BEING set, | |
7789 | and neither is the SUBREG inside it. | |
7790 | Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */ | |
7791 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
7792 | dest = SUBREG_REG (XEXP (dest, 0)); | |
7793 | ||
c610adec | 7794 | if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG) |
7afe21cc RK |
7795 | /* Registers must also be inserted into chains for quantities. */ |
7796 | if (insert_regs (dest, sets[i].src_elt, 1)) | |
8ae2b8f6 JW |
7797 | { |
7798 | /* If `insert_regs' changes something, the hash code must be | |
7799 | recalculated. */ | |
7800 | rehash_using_reg (dest); | |
7801 | sets[i].dest_hash = HASH (dest, GET_MODE (dest)); | |
7802 | } | |
7afe21cc | 7803 | |
9de2c71a MM |
7804 | if (GET_CODE (inner_dest) == MEM |
7805 | && GET_CODE (XEXP (inner_dest, 0)) == ADDRESSOF) | |
7806 | /* Given (SET (MEM (ADDRESSOF (X))) Y) we don't want to say | |
7807 | that (MEM (ADDRESSOF (X))) is equivalent to Y. | |
7808 | Consider the case in which the address of the MEM is | |
7809 | passed to a function, which alters the MEM. Then, if we | |
7810 | later use Y instead of the MEM we'll miss the update. */ | |
7811 | elt = insert (dest, 0, sets[i].dest_hash, GET_MODE (dest)); | |
7812 | else | |
7813 | elt = insert (dest, sets[i].src_elt, | |
7814 | sets[i].dest_hash, GET_MODE (dest)); | |
7815 | ||
c256df0b | 7816 | elt->in_memory = (GET_CODE (sets[i].inner_dest) == MEM |
9ad91d71 RK |
7817 | && (! RTX_UNCHANGING_P (sets[i].inner_dest) |
7818 | || FIXED_BASE_PLUS_P (XEXP (sets[i].inner_dest, | |
7819 | 0)))); | |
c256df0b | 7820 | |
7afe21cc RK |
7821 | if (elt->in_memory) |
7822 | { | |
7823 | /* This implicitly assumes a whole struct | |
7824 | need not have MEM_IN_STRUCT_P. | |
7825 | But a whole struct is *supposed* to have MEM_IN_STRUCT_P. */ | |
7826 | elt->in_struct = (MEM_IN_STRUCT_P (sets[i].inner_dest) | |
7827 | || sets[i].inner_dest != SET_DEST (sets[i].rtl)); | |
7828 | } | |
7829 | ||
fc3ffe83 RK |
7830 | /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no |
7831 | narrower than M2, and both M1 and M2 are the same number of words, | |
7832 | we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so | |
7833 | make that equivalence as well. | |
7afe21cc RK |
7834 | |
7835 | However, BAR may have equivalences for which gen_lowpart_if_possible | |
7836 | will produce a simpler value than gen_lowpart_if_possible applied to | |
7837 | BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all | |
7838 | BAR's equivalences. If we don't get a simplified form, make | |
7839 | the SUBREG. It will not be used in an equivalence, but will | |
7840 | cause two similar assignments to be detected. | |
7841 | ||
7842 | Note the loop below will find SUBREG_REG (DEST) since we have | |
7843 | already entered SRC and DEST of the SET in the table. */ | |
7844 | ||
7845 | if (GET_CODE (dest) == SUBREG | |
6cdbaec4 RK |
7846 | && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1) |
7847 | / UNITS_PER_WORD) | |
7848 | == (GET_MODE_SIZE (GET_MODE (dest)) - 1)/ UNITS_PER_WORD) | |
7afe21cc RK |
7849 | && (GET_MODE_SIZE (GET_MODE (dest)) |
7850 | >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))) | |
7851 | && sets[i].src_elt != 0) | |
7852 | { | |
7853 | enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest)); | |
7854 | struct table_elt *elt, *classp = 0; | |
7855 | ||
7856 | for (elt = sets[i].src_elt->first_same_value; elt; | |
7857 | elt = elt->next_same_value) | |
7858 | { | |
7859 | rtx new_src = 0; | |
2197a88a | 7860 | unsigned src_hash; |
7afe21cc RK |
7861 | struct table_elt *src_elt; |
7862 | ||
7863 | /* Ignore invalid entries. */ | |
7864 | if (GET_CODE (elt->exp) != REG | |
7865 | && ! exp_equiv_p (elt->exp, elt->exp, 1, 0)) | |
7866 | continue; | |
7867 | ||
7868 | new_src = gen_lowpart_if_possible (new_mode, elt->exp); | |
7869 | if (new_src == 0) | |
38a448ca | 7870 | new_src = gen_rtx_SUBREG (new_mode, elt->exp, 0); |
7afe21cc RK |
7871 | |
7872 | src_hash = HASH (new_src, new_mode); | |
7873 | src_elt = lookup (new_src, src_hash, new_mode); | |
7874 | ||
7875 | /* Put the new source in the hash table is if isn't | |
7876 | already. */ | |
7877 | if (src_elt == 0) | |
7878 | { | |
7879 | if (insert_regs (new_src, classp, 0)) | |
8ae2b8f6 JW |
7880 | { |
7881 | rehash_using_reg (new_src); | |
7882 | src_hash = HASH (new_src, new_mode); | |
7883 | } | |
7afe21cc RK |
7884 | src_elt = insert (new_src, classp, src_hash, new_mode); |
7885 | src_elt->in_memory = elt->in_memory; | |
7886 | src_elt->in_struct = elt->in_struct; | |
7887 | } | |
7888 | else if (classp && classp != src_elt->first_same_value) | |
7889 | /* Show that two things that we've seen before are | |
7890 | actually the same. */ | |
7891 | merge_equiv_classes (src_elt, classp); | |
7892 | ||
7893 | classp = src_elt->first_same_value; | |
da932f04 JL |
7894 | /* Ignore invalid entries. */ |
7895 | while (classp | |
7896 | && GET_CODE (classp->exp) != REG | |
7897 | && ! exp_equiv_p (classp->exp, classp->exp, 1, 0)) | |
7898 | classp = classp->next_same_value; | |
7afe21cc RK |
7899 | } |
7900 | } | |
7901 | } | |
7902 | ||
7903 | /* Special handling for (set REG0 REG1) | |
7904 | where REG0 is the "cheapest", cheaper than REG1. | |
7905 | After cse, REG1 will probably not be used in the sequel, | |
7906 | so (if easily done) change this insn to (set REG1 REG0) and | |
7907 | replace REG1 with REG0 in the previous insn that computed their value. | |
7908 | Then REG1 will become a dead store and won't cloud the situation | |
7909 | for later optimizations. | |
7910 | ||
7911 | Do not make this change if REG1 is a hard register, because it will | |
7912 | then be used in the sequel and we may be changing a two-operand insn | |
7913 | into a three-operand insn. | |
7914 | ||
50270076 R |
7915 | Also do not do this if we are operating on a copy of INSN. |
7916 | ||
7917 | Also don't do this if INSN ends a libcall; this would cause an unrelated | |
7918 | register to be set in the middle of a libcall, and we then get bad code | |
7919 | if the libcall is deleted. */ | |
7afe21cc RK |
7920 | |
7921 | if (n_sets == 1 && sets[0].rtl && GET_CODE (SET_DEST (sets[0].rtl)) == REG | |
7922 | && NEXT_INSN (PREV_INSN (insn)) == insn | |
7923 | && GET_CODE (SET_SRC (sets[0].rtl)) == REG | |
7924 | && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER | |
7925 | && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))) | |
30f72379 | 7926 | && (qty_first_reg[REG_QTY (REGNO (SET_SRC (sets[0].rtl)))] |
50270076 R |
7927 | == REGNO (SET_DEST (sets[0].rtl))) |
7928 | && ! find_reg_note (insn, REG_RETVAL, NULL_RTX)) | |
7afe21cc RK |
7929 | { |
7930 | rtx prev = PREV_INSN (insn); | |
7931 | while (prev && GET_CODE (prev) == NOTE) | |
7932 | prev = PREV_INSN (prev); | |
7933 | ||
7934 | if (prev && GET_CODE (prev) == INSN && GET_CODE (PATTERN (prev)) == SET | |
7935 | && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)) | |
7936 | { | |
7937 | rtx dest = SET_DEST (sets[0].rtl); | |
906c4e36 | 7938 | rtx note = find_reg_note (prev, REG_EQUIV, NULL_RTX); |
7afe21cc RK |
7939 | |
7940 | validate_change (prev, & SET_DEST (PATTERN (prev)), dest, 1); | |
7941 | validate_change (insn, & SET_DEST (sets[0].rtl), | |
7942 | SET_SRC (sets[0].rtl), 1); | |
7943 | validate_change (insn, & SET_SRC (sets[0].rtl), dest, 1); | |
7944 | apply_change_group (); | |
7945 | ||
7946 | /* If REG1 was equivalent to a constant, REG0 is not. */ | |
7947 | if (note) | |
7948 | PUT_REG_NOTE_KIND (note, REG_EQUAL); | |
7949 | ||
7950 | /* If there was a REG_WAS_0 note on PREV, remove it. Move | |
7951 | any REG_WAS_0 note on INSN to PREV. */ | |
906c4e36 | 7952 | note = find_reg_note (prev, REG_WAS_0, NULL_RTX); |
7afe21cc RK |
7953 | if (note) |
7954 | remove_note (prev, note); | |
7955 | ||
906c4e36 | 7956 | note = find_reg_note (insn, REG_WAS_0, NULL_RTX); |
7afe21cc RK |
7957 | if (note) |
7958 | { | |
7959 | remove_note (insn, note); | |
7960 | XEXP (note, 1) = REG_NOTES (prev); | |
7961 | REG_NOTES (prev) = note; | |
7962 | } | |
98369a0f RK |
7963 | |
7964 | /* If INSN has a REG_EQUAL note, and this note mentions REG0, | |
7965 | then we must delete it, because the value in REG0 has changed. */ | |
7966 | note = find_reg_note (insn, REG_EQUAL, NULL_RTX); | |
7967 | if (note && reg_mentioned_p (dest, XEXP (note, 0))) | |
7968 | remove_note (insn, note); | |
7afe21cc RK |
7969 | } |
7970 | } | |
7971 | ||
7972 | /* If this is a conditional jump insn, record any known equivalences due to | |
7973 | the condition being tested. */ | |
7974 | ||
7975 | last_jump_equiv_class = 0; | |
7976 | if (GET_CODE (insn) == JUMP_INSN | |
7977 | && n_sets == 1 && GET_CODE (x) == SET | |
7978 | && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE) | |
7979 | record_jump_equiv (insn, 0); | |
7980 | ||
7981 | #ifdef HAVE_cc0 | |
7982 | /* If the previous insn set CC0 and this insn no longer references CC0, | |
7983 | delete the previous insn. Here we use the fact that nothing expects CC0 | |
7984 | to be valid over an insn, which is true until the final pass. */ | |
7985 | if (prev_insn && GET_CODE (prev_insn) == INSN | |
7986 | && (tem = single_set (prev_insn)) != 0 | |
7987 | && SET_DEST (tem) == cc0_rtx | |
7988 | && ! reg_mentioned_p (cc0_rtx, x)) | |
7989 | { | |
7990 | PUT_CODE (prev_insn, NOTE); | |
7991 | NOTE_LINE_NUMBER (prev_insn) = NOTE_INSN_DELETED; | |
7992 | NOTE_SOURCE_FILE (prev_insn) = 0; | |
7993 | } | |
7994 | ||
7995 | prev_insn_cc0 = this_insn_cc0; | |
7996 | prev_insn_cc0_mode = this_insn_cc0_mode; | |
7997 | #endif | |
7998 | ||
7999 | prev_insn = insn; | |
8000 | } | |
8001 | \f | |
a4c6502a | 8002 | /* Remove from the hash table all expressions that reference memory. */ |
7afe21cc | 8003 | static void |
9ae8ffe7 | 8004 | invalidate_memory () |
7afe21cc | 8005 | { |
9ae8ffe7 JL |
8006 | register int i; |
8007 | register struct table_elt *p, *next; | |
7afe21cc | 8008 | |
9ae8ffe7 JL |
8009 | for (i = 0; i < NBUCKETS; i++) |
8010 | for (p = table[i]; p; p = next) | |
8011 | { | |
8012 | next = p->next_same_hash; | |
8013 | if (p->in_memory) | |
8014 | remove_from_table (p, i); | |
8015 | } | |
8016 | } | |
8017 | ||
8018 | /* XXX ??? The name of this function bears little resemblance to | |
8019 | what this function actually does. FIXME. */ | |
8020 | static int | |
8021 | note_mem_written (addr) | |
8022 | register rtx addr; | |
8023 | { | |
8024 | /* Pushing or popping the stack invalidates just the stack pointer. */ | |
8025 | if ((GET_CODE (addr) == PRE_DEC || GET_CODE (addr) == PRE_INC | |
8026 | || GET_CODE (addr) == POST_DEC || GET_CODE (addr) == POST_INC) | |
8027 | && GET_CODE (XEXP (addr, 0)) == REG | |
8028 | && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM) | |
7afe21cc | 8029 | { |
30f72379 MM |
8030 | if (REG_TICK (STACK_POINTER_REGNUM) >= 0) |
8031 | REG_TICK (STACK_POINTER_REGNUM)++; | |
9ae8ffe7 JL |
8032 | |
8033 | /* This should be *very* rare. */ | |
8034 | if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM)) | |
8035 | invalidate (stack_pointer_rtx, VOIDmode); | |
8036 | return 1; | |
7afe21cc | 8037 | } |
9ae8ffe7 | 8038 | return 0; |
7afe21cc RK |
8039 | } |
8040 | ||
8041 | /* Perform invalidation on the basis of everything about an insn | |
8042 | except for invalidating the actual places that are SET in it. | |
8043 | This includes the places CLOBBERed, and anything that might | |
8044 | alias with something that is SET or CLOBBERed. | |
8045 | ||
7afe21cc RK |
8046 | X is the pattern of the insn. */ |
8047 | ||
8048 | static void | |
9ae8ffe7 | 8049 | invalidate_from_clobbers (x) |
7afe21cc RK |
8050 | rtx x; |
8051 | { | |
7afe21cc RK |
8052 | if (GET_CODE (x) == CLOBBER) |
8053 | { | |
8054 | rtx ref = XEXP (x, 0); | |
9ae8ffe7 JL |
8055 | if (ref) |
8056 | { | |
8057 | if (GET_CODE (ref) == REG || GET_CODE (ref) == SUBREG | |
8058 | || GET_CODE (ref) == MEM) | |
8059 | invalidate (ref, VOIDmode); | |
8060 | else if (GET_CODE (ref) == STRICT_LOW_PART | |
8061 | || GET_CODE (ref) == ZERO_EXTRACT) | |
8062 | invalidate (XEXP (ref, 0), GET_MODE (ref)); | |
8063 | } | |
7afe21cc RK |
8064 | } |
8065 | else if (GET_CODE (x) == PARALLEL) | |
8066 | { | |
8067 | register int i; | |
8068 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
8069 | { | |
8070 | register rtx y = XVECEXP (x, 0, i); | |
8071 | if (GET_CODE (y) == CLOBBER) | |
8072 | { | |
8073 | rtx ref = XEXP (y, 0); | |
9ae8ffe7 JL |
8074 | if (GET_CODE (ref) == REG || GET_CODE (ref) == SUBREG |
8075 | || GET_CODE (ref) == MEM) | |
8076 | invalidate (ref, VOIDmode); | |
8077 | else if (GET_CODE (ref) == STRICT_LOW_PART | |
8078 | || GET_CODE (ref) == ZERO_EXTRACT) | |
8079 | invalidate (XEXP (ref, 0), GET_MODE (ref)); | |
7afe21cc RK |
8080 | } |
8081 | } | |
8082 | } | |
8083 | } | |
8084 | \f | |
8085 | /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes | |
8086 | and replace any registers in them with either an equivalent constant | |
8087 | or the canonical form of the register. If we are inside an address, | |
8088 | only do this if the address remains valid. | |
8089 | ||
8090 | OBJECT is 0 except when within a MEM in which case it is the MEM. | |
8091 | ||
8092 | Return the replacement for X. */ | |
8093 | ||
8094 | static rtx | |
8095 | cse_process_notes (x, object) | |
8096 | rtx x; | |
8097 | rtx object; | |
8098 | { | |
8099 | enum rtx_code code = GET_CODE (x); | |
8100 | char *fmt = GET_RTX_FORMAT (code); | |
7afe21cc RK |
8101 | int i; |
8102 | ||
8103 | switch (code) | |
8104 | { | |
8105 | case CONST_INT: | |
8106 | case CONST: | |
8107 | case SYMBOL_REF: | |
8108 | case LABEL_REF: | |
8109 | case CONST_DOUBLE: | |
8110 | case PC: | |
8111 | case CC0: | |
8112 | case LO_SUM: | |
8113 | return x; | |
8114 | ||
8115 | case MEM: | |
8116 | XEXP (x, 0) = cse_process_notes (XEXP (x, 0), x); | |
8117 | return x; | |
8118 | ||
8119 | case EXPR_LIST: | |
8120 | case INSN_LIST: | |
8121 | if (REG_NOTE_KIND (x) == REG_EQUAL) | |
906c4e36 | 8122 | XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX); |
7afe21cc | 8123 | if (XEXP (x, 1)) |
906c4e36 | 8124 | XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX); |
7afe21cc RK |
8125 | return x; |
8126 | ||
e4890d45 RS |
8127 | case SIGN_EXTEND: |
8128 | case ZERO_EXTEND: | |
0b0ee36c | 8129 | case SUBREG: |
e4890d45 RS |
8130 | { |
8131 | rtx new = cse_process_notes (XEXP (x, 0), object); | |
8132 | /* We don't substitute VOIDmode constants into these rtx, | |
8133 | since they would impede folding. */ | |
8134 | if (GET_MODE (new) != VOIDmode) | |
8135 | validate_change (object, &XEXP (x, 0), new, 0); | |
8136 | return x; | |
8137 | } | |
8138 | ||
7afe21cc | 8139 | case REG: |
30f72379 | 8140 | i = REG_QTY (REGNO (x)); |
7afe21cc RK |
8141 | |
8142 | /* Return a constant or a constant register. */ | |
8143 | if (REGNO_QTY_VALID_P (REGNO (x)) | |
8144 | && qty_const[i] != 0 | |
8145 | && (CONSTANT_P (qty_const[i]) | |
8146 | || GET_CODE (qty_const[i]) == REG)) | |
8147 | { | |
8148 | rtx new = gen_lowpart_if_possible (GET_MODE (x), qty_const[i]); | |
8149 | if (new) | |
8150 | return new; | |
8151 | } | |
8152 | ||
8153 | /* Otherwise, canonicalize this register. */ | |
906c4e36 | 8154 | return canon_reg (x, NULL_RTX); |
e9a25f70 JL |
8155 | |
8156 | default: | |
8157 | break; | |
7afe21cc RK |
8158 | } |
8159 | ||
8160 | for (i = 0; i < GET_RTX_LENGTH (code); i++) | |
8161 | if (fmt[i] == 'e') | |
8162 | validate_change (object, &XEXP (x, i), | |
7fe34fdf | 8163 | cse_process_notes (XEXP (x, i), object), 0); |
7afe21cc RK |
8164 | |
8165 | return x; | |
8166 | } | |
8167 | \f | |
8168 | /* Find common subexpressions between the end test of a loop and the beginning | |
8169 | of the loop. LOOP_START is the CODE_LABEL at the start of a loop. | |
8170 | ||
8171 | Often we have a loop where an expression in the exit test is used | |
8172 | in the body of the loop. For example "while (*p) *q++ = *p++;". | |
8173 | Because of the way we duplicate the loop exit test in front of the loop, | |
8174 | however, we don't detect that common subexpression. This will be caught | |
8175 | when global cse is implemented, but this is a quite common case. | |
8176 | ||
8177 | This function handles the most common cases of these common expressions. | |
8178 | It is called after we have processed the basic block ending with the | |
8179 | NOTE_INSN_LOOP_END note that ends a loop and the previous JUMP_INSN | |
8180 | jumps to a label used only once. */ | |
8181 | ||
8182 | static void | |
8183 | cse_around_loop (loop_start) | |
8184 | rtx loop_start; | |
8185 | { | |
8186 | rtx insn; | |
8187 | int i; | |
8188 | struct table_elt *p; | |
8189 | ||
8190 | /* If the jump at the end of the loop doesn't go to the start, we don't | |
8191 | do anything. */ | |
8192 | for (insn = PREV_INSN (loop_start); | |
8193 | insn && (GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) >= 0); | |
8194 | insn = PREV_INSN (insn)) | |
8195 | ; | |
8196 | ||
8197 | if (insn == 0 | |
8198 | || GET_CODE (insn) != NOTE | |
8199 | || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG) | |
8200 | return; | |
8201 | ||
8202 | /* If the last insn of the loop (the end test) was an NE comparison, | |
8203 | we will interpret it as an EQ comparison, since we fell through | |
f72aed24 | 8204 | the loop. Any equivalences resulting from that comparison are |
7afe21cc RK |
8205 | therefore not valid and must be invalidated. */ |
8206 | if (last_jump_equiv_class) | |
8207 | for (p = last_jump_equiv_class->first_same_value; p; | |
8208 | p = p->next_same_value) | |
51723711 KG |
8209 | { |
8210 | if (GET_CODE (p->exp) == MEM || GET_CODE (p->exp) == REG | |
8211 | || (GET_CODE (p->exp) == SUBREG | |
8212 | && GET_CODE (SUBREG_REG (p->exp)) == REG)) | |
8213 | invalidate (p->exp, VOIDmode); | |
8214 | else if (GET_CODE (p->exp) == STRICT_LOW_PART | |
8215 | || GET_CODE (p->exp) == ZERO_EXTRACT) | |
8216 | invalidate (XEXP (p->exp, 0), GET_MODE (p->exp)); | |
8217 | } | |
7afe21cc RK |
8218 | |
8219 | /* Process insns starting after LOOP_START until we hit a CALL_INSN or | |
8220 | a CODE_LABEL (we could handle a CALL_INSN, but it isn't worth it). | |
8221 | ||
8222 | The only thing we do with SET_DEST is invalidate entries, so we | |
8223 | can safely process each SET in order. It is slightly less efficient | |
556c714b JW |
8224 | to do so, but we only want to handle the most common cases. |
8225 | ||
8226 | The gen_move_insn call in cse_set_around_loop may create new pseudos. | |
8227 | These pseudos won't have valid entries in any of the tables indexed | |
8228 | by register number, such as reg_qty. We avoid out-of-range array | |
8229 | accesses by not processing any instructions created after cse started. */ | |
7afe21cc RK |
8230 | |
8231 | for (insn = NEXT_INSN (loop_start); | |
8232 | GET_CODE (insn) != CALL_INSN && GET_CODE (insn) != CODE_LABEL | |
556c714b | 8233 | && INSN_UID (insn) < max_insn_uid |
7afe21cc RK |
8234 | && ! (GET_CODE (insn) == NOTE |
8235 | && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END); | |
8236 | insn = NEXT_INSN (insn)) | |
8237 | { | |
8238 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
8239 | && (GET_CODE (PATTERN (insn)) == SET | |
8240 | || GET_CODE (PATTERN (insn)) == CLOBBER)) | |
8241 | cse_set_around_loop (PATTERN (insn), insn, loop_start); | |
8242 | else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
8243 | && GET_CODE (PATTERN (insn)) == PARALLEL) | |
8244 | for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) | |
8245 | if (GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == SET | |
8246 | || GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == CLOBBER) | |
8247 | cse_set_around_loop (XVECEXP (PATTERN (insn), 0, i), insn, | |
8248 | loop_start); | |
8249 | } | |
8250 | } | |
8251 | \f | |
8b3686ed RK |
8252 | /* Process one SET of an insn that was skipped. We ignore CLOBBERs |
8253 | since they are done elsewhere. This function is called via note_stores. */ | |
8254 | ||
8255 | static void | |
8256 | invalidate_skipped_set (dest, set) | |
8257 | rtx set; | |
8258 | rtx dest; | |
8259 | { | |
9ae8ffe7 JL |
8260 | enum rtx_code code = GET_CODE (dest); |
8261 | ||
8262 | if (code == MEM | |
8263 | && ! note_mem_written (dest) /* If this is not a stack push ... */ | |
8264 | /* There are times when an address can appear varying and be a PLUS | |
8265 | during this scan when it would be a fixed address were we to know | |
8266 | the proper equivalences. So invalidate all memory if there is | |
8267 | a BLKmode or nonscalar memory reference or a reference to a | |
8268 | variable address. */ | |
8269 | && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode | |
8270 | || cse_rtx_varies_p (XEXP (dest, 0)))) | |
8271 | { | |
8272 | invalidate_memory (); | |
8273 | return; | |
8274 | } | |
ffcf6393 | 8275 | |
f47c02fa RK |
8276 | if (GET_CODE (set) == CLOBBER |
8277 | #ifdef HAVE_cc0 | |
8278 | || dest == cc0_rtx | |
8279 | #endif | |
8280 | || dest == pc_rtx) | |
8281 | return; | |
8282 | ||
9ae8ffe7 | 8283 | if (code == STRICT_LOW_PART || code == ZERO_EXTRACT) |
bb4034b3 | 8284 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
9ae8ffe7 JL |
8285 | else if (code == REG || code == SUBREG || code == MEM) |
8286 | invalidate (dest, VOIDmode); | |
8b3686ed RK |
8287 | } |
8288 | ||
8289 | /* Invalidate all insns from START up to the end of the function or the | |
8290 | next label. This called when we wish to CSE around a block that is | |
8291 | conditionally executed. */ | |
8292 | ||
8293 | static void | |
8294 | invalidate_skipped_block (start) | |
8295 | rtx start; | |
8296 | { | |
8297 | rtx insn; | |
8b3686ed RK |
8298 | |
8299 | for (insn = start; insn && GET_CODE (insn) != CODE_LABEL; | |
8300 | insn = NEXT_INSN (insn)) | |
8301 | { | |
8302 | if (GET_RTX_CLASS (GET_CODE (insn)) != 'i') | |
8303 | continue; | |
8304 | ||
8b3686ed RK |
8305 | if (GET_CODE (insn) == CALL_INSN) |
8306 | { | |
9ae8ffe7 JL |
8307 | if (! CONST_CALL_P (insn)) |
8308 | invalidate_memory (); | |
8b3686ed | 8309 | invalidate_for_call (); |
8b3686ed RK |
8310 | } |
8311 | ||
97577254 | 8312 | invalidate_from_clobbers (PATTERN (insn)); |
8b3686ed | 8313 | note_stores (PATTERN (insn), invalidate_skipped_set); |
8b3686ed RK |
8314 | } |
8315 | } | |
8316 | \f | |
7afe21cc RK |
8317 | /* Used for communication between the following two routines; contains a |
8318 | value to be checked for modification. */ | |
8319 | ||
8320 | static rtx cse_check_loop_start_value; | |
8321 | ||
8322 | /* If modifying X will modify the value in CSE_CHECK_LOOP_START_VALUE, | |
8323 | indicate that fact by setting CSE_CHECK_LOOP_START_VALUE to 0. */ | |
8324 | ||
8325 | static void | |
8326 | cse_check_loop_start (x, set) | |
8327 | rtx x; | |
d6f4ec51 | 8328 | rtx set ATTRIBUTE_UNUSED; |
7afe21cc RK |
8329 | { |
8330 | if (cse_check_loop_start_value == 0 | |
8331 | || GET_CODE (x) == CC0 || GET_CODE (x) == PC) | |
8332 | return; | |
8333 | ||
8334 | if ((GET_CODE (x) == MEM && GET_CODE (cse_check_loop_start_value) == MEM) | |
8335 | || reg_overlap_mentioned_p (x, cse_check_loop_start_value)) | |
8336 | cse_check_loop_start_value = 0; | |
8337 | } | |
8338 | ||
8339 | /* X is a SET or CLOBBER contained in INSN that was found near the start of | |
8340 | a loop that starts with the label at LOOP_START. | |
8341 | ||
8342 | If X is a SET, we see if its SET_SRC is currently in our hash table. | |
8343 | If so, we see if it has a value equal to some register used only in the | |
8344 | loop exit code (as marked by jump.c). | |
8345 | ||
8346 | If those two conditions are true, we search backwards from the start of | |
8347 | the loop to see if that same value was loaded into a register that still | |
8348 | retains its value at the start of the loop. | |
8349 | ||
8350 | If so, we insert an insn after the load to copy the destination of that | |
8351 | load into the equivalent register and (try to) replace our SET_SRC with that | |
8352 | register. | |
8353 | ||
8354 | In any event, we invalidate whatever this SET or CLOBBER modifies. */ | |
8355 | ||
8356 | static void | |
8357 | cse_set_around_loop (x, insn, loop_start) | |
8358 | rtx x; | |
8359 | rtx insn; | |
8360 | rtx loop_start; | |
8361 | { | |
7afe21cc | 8362 | struct table_elt *src_elt; |
7afe21cc RK |
8363 | |
8364 | /* If this is a SET, see if we can replace SET_SRC, but ignore SETs that | |
8365 | are setting PC or CC0 or whose SET_SRC is already a register. */ | |
8366 | if (GET_CODE (x) == SET | |
8367 | && GET_CODE (SET_DEST (x)) != PC && GET_CODE (SET_DEST (x)) != CC0 | |
8368 | && GET_CODE (SET_SRC (x)) != REG) | |
8369 | { | |
8370 | src_elt = lookup (SET_SRC (x), | |
8371 | HASH (SET_SRC (x), GET_MODE (SET_DEST (x))), | |
8372 | GET_MODE (SET_DEST (x))); | |
8373 | ||
8374 | if (src_elt) | |
8375 | for (src_elt = src_elt->first_same_value; src_elt; | |
8376 | src_elt = src_elt->next_same_value) | |
8377 | if (GET_CODE (src_elt->exp) == REG && REG_LOOP_TEST_P (src_elt->exp) | |
8378 | && COST (src_elt->exp) < COST (SET_SRC (x))) | |
8379 | { | |
8380 | rtx p, set; | |
8381 | ||
8382 | /* Look for an insn in front of LOOP_START that sets | |
8383 | something in the desired mode to SET_SRC (x) before we hit | |
8384 | a label or CALL_INSN. */ | |
8385 | ||
8386 | for (p = prev_nonnote_insn (loop_start); | |
8387 | p && GET_CODE (p) != CALL_INSN | |
8388 | && GET_CODE (p) != CODE_LABEL; | |
8389 | p = prev_nonnote_insn (p)) | |
8390 | if ((set = single_set (p)) != 0 | |
8391 | && GET_CODE (SET_DEST (set)) == REG | |
8392 | && GET_MODE (SET_DEST (set)) == src_elt->mode | |
8393 | && rtx_equal_p (SET_SRC (set), SET_SRC (x))) | |
8394 | { | |
8395 | /* We now have to ensure that nothing between P | |
8396 | and LOOP_START modified anything referenced in | |
8397 | SET_SRC (x). We know that nothing within the loop | |
8398 | can modify it, or we would have invalidated it in | |
8399 | the hash table. */ | |
8400 | rtx q; | |
8401 | ||
8402 | cse_check_loop_start_value = SET_SRC (x); | |
8403 | for (q = p; q != loop_start; q = NEXT_INSN (q)) | |
8404 | if (GET_RTX_CLASS (GET_CODE (q)) == 'i') | |
8405 | note_stores (PATTERN (q), cse_check_loop_start); | |
8406 | ||
8407 | /* If nothing was changed and we can replace our | |
8408 | SET_SRC, add an insn after P to copy its destination | |
8409 | to what we will be replacing SET_SRC with. */ | |
8410 | if (cse_check_loop_start_value | |
8411 | && validate_change (insn, &SET_SRC (x), | |
8412 | src_elt->exp, 0)) | |
e89d3e6f R |
8413 | { |
8414 | /* If this creates new pseudos, this is unsafe, | |
8415 | because the regno of new pseudo is unsuitable | |
8416 | to index into reg_qty when cse_insn processes | |
8417 | the new insn. Therefore, if a new pseudo was | |
8418 | created, discard this optimization. */ | |
8419 | int nregs = max_reg_num (); | |
8420 | rtx move | |
8421 | = gen_move_insn (src_elt->exp, SET_DEST (set)); | |
8422 | if (nregs != max_reg_num ()) | |
8423 | { | |
8424 | if (! validate_change (insn, &SET_SRC (x), | |
8425 | SET_SRC (set), 0)) | |
8426 | abort (); | |
8427 | } | |
8428 | else | |
8429 | emit_insn_after (move, p); | |
8430 | } | |
7afe21cc RK |
8431 | break; |
8432 | } | |
8433 | } | |
8434 | } | |
8435 | ||
8436 | /* Now invalidate anything modified by X. */ | |
9ae8ffe7 | 8437 | note_mem_written (SET_DEST (x)); |
7afe21cc | 8438 | |
9ae8ffe7 | 8439 | /* See comment on similar code in cse_insn for explanation of these tests. */ |
7afe21cc | 8440 | if (GET_CODE (SET_DEST (x)) == REG || GET_CODE (SET_DEST (x)) == SUBREG |
9ae8ffe7 | 8441 | || GET_CODE (SET_DEST (x)) == MEM) |
bb4034b3 | 8442 | invalidate (SET_DEST (x), VOIDmode); |
2708da92 RS |
8443 | else if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART |
8444 | || GET_CODE (SET_DEST (x)) == ZERO_EXTRACT) | |
bb4034b3 | 8445 | invalidate (XEXP (SET_DEST (x), 0), GET_MODE (SET_DEST (x))); |
7afe21cc RK |
8446 | } |
8447 | \f | |
8448 | /* Find the end of INSN's basic block and return its range, | |
8449 | the total number of SETs in all the insns of the block, the last insn of the | |
8450 | block, and the branch path. | |
8451 | ||
8452 | The branch path indicates which branches should be followed. If a non-zero | |
8453 | path size is specified, the block should be rescanned and a different set | |
8454 | of branches will be taken. The branch path is only used if | |
8b3686ed | 8455 | FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is non-zero. |
7afe21cc RK |
8456 | |
8457 | DATA is a pointer to a struct cse_basic_block_data, defined below, that is | |
8458 | used to describe the block. It is filled in with the information about | |
8459 | the current block. The incoming structure's branch path, if any, is used | |
8460 | to construct the output branch path. */ | |
8461 | ||
7afe21cc | 8462 | void |
8b3686ed | 8463 | cse_end_of_basic_block (insn, data, follow_jumps, after_loop, skip_blocks) |
7afe21cc RK |
8464 | rtx insn; |
8465 | struct cse_basic_block_data *data; | |
8466 | int follow_jumps; | |
8467 | int after_loop; | |
8b3686ed | 8468 | int skip_blocks; |
7afe21cc RK |
8469 | { |
8470 | rtx p = insn, q; | |
8471 | int nsets = 0; | |
8472 | int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn); | |
fc3ffe83 | 8473 | rtx next = GET_RTX_CLASS (GET_CODE (insn)) == 'i' ? insn : next_real_insn (insn); |
7afe21cc RK |
8474 | int path_size = data->path_size; |
8475 | int path_entry = 0; | |
8476 | int i; | |
8477 | ||
8478 | /* Update the previous branch path, if any. If the last branch was | |
8479 | previously TAKEN, mark it NOT_TAKEN. If it was previously NOT_TAKEN, | |
8480 | shorten the path by one and look at the previous branch. We know that | |
8481 | at least one branch must have been taken if PATH_SIZE is non-zero. */ | |
8482 | while (path_size > 0) | |
8483 | { | |
8b3686ed | 8484 | if (data->path[path_size - 1].status != NOT_TAKEN) |
7afe21cc RK |
8485 | { |
8486 | data->path[path_size - 1].status = NOT_TAKEN; | |
8487 | break; | |
8488 | } | |
8489 | else | |
8490 | path_size--; | |
8491 | } | |
8492 | ||
8493 | /* Scan to end of this basic block. */ | |
8494 | while (p && GET_CODE (p) != CODE_LABEL) | |
8495 | { | |
8496 | /* Don't cse out the end of a loop. This makes a difference | |
8497 | only for the unusual loops that always execute at least once; | |
8498 | all other loops have labels there so we will stop in any case. | |
8499 | Cse'ing out the end of the loop is dangerous because it | |
8500 | might cause an invariant expression inside the loop | |
8501 | to be reused after the end of the loop. This would make it | |
8502 | hard to move the expression out of the loop in loop.c, | |
8503 | especially if it is one of several equivalent expressions | |
8504 | and loop.c would like to eliminate it. | |
8505 | ||
8506 | If we are running after loop.c has finished, we can ignore | |
8507 | the NOTE_INSN_LOOP_END. */ | |
8508 | ||
8509 | if (! after_loop && GET_CODE (p) == NOTE | |
8510 | && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END) | |
8511 | break; | |
8512 | ||
8513 | /* Don't cse over a call to setjmp; on some machines (eg vax) | |
8514 | the regs restored by the longjmp come from | |
8515 | a later time than the setjmp. */ | |
8516 | if (GET_CODE (p) == NOTE | |
8517 | && NOTE_LINE_NUMBER (p) == NOTE_INSN_SETJMP) | |
8518 | break; | |
8519 | ||
8520 | /* A PARALLEL can have lots of SETs in it, | |
8521 | especially if it is really an ASM_OPERANDS. */ | |
8522 | if (GET_RTX_CLASS (GET_CODE (p)) == 'i' | |
8523 | && GET_CODE (PATTERN (p)) == PARALLEL) | |
8524 | nsets += XVECLEN (PATTERN (p), 0); | |
8525 | else if (GET_CODE (p) != NOTE) | |
8526 | nsets += 1; | |
8527 | ||
164c8956 RK |
8528 | /* Ignore insns made by CSE; they cannot affect the boundaries of |
8529 | the basic block. */ | |
8530 | ||
8531 | if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid) | |
8b3686ed | 8532 | high_cuid = INSN_CUID (p); |
164c8956 RK |
8533 | if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid) |
8534 | low_cuid = INSN_CUID (p); | |
7afe21cc RK |
8535 | |
8536 | /* See if this insn is in our branch path. If it is and we are to | |
8537 | take it, do so. */ | |
8538 | if (path_entry < path_size && data->path[path_entry].branch == p) | |
8539 | { | |
8b3686ed | 8540 | if (data->path[path_entry].status != NOT_TAKEN) |
7afe21cc RK |
8541 | p = JUMP_LABEL (p); |
8542 | ||
8543 | /* Point to next entry in path, if any. */ | |
8544 | path_entry++; | |
8545 | } | |
8546 | ||
8547 | /* If this is a conditional jump, we can follow it if -fcse-follow-jumps | |
8548 | was specified, we haven't reached our maximum path length, there are | |
8549 | insns following the target of the jump, this is the only use of the | |
8b3686ed RK |
8550 | jump label, and the target label is preceded by a BARRIER. |
8551 | ||
8552 | Alternatively, we can follow the jump if it branches around a | |
8553 | block of code and there are no other branches into the block. | |
8554 | In this case invalidate_skipped_block will be called to invalidate any | |
8555 | registers set in the block when following the jump. */ | |
8556 | ||
8557 | else if ((follow_jumps || skip_blocks) && path_size < PATHLENGTH - 1 | |
7afe21cc RK |
8558 | && GET_CODE (p) == JUMP_INSN |
8559 | && GET_CODE (PATTERN (p)) == SET | |
8560 | && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE | |
85c3ba60 | 8561 | && JUMP_LABEL (p) != 0 |
7afe21cc RK |
8562 | && LABEL_NUSES (JUMP_LABEL (p)) == 1 |
8563 | && NEXT_INSN (JUMP_LABEL (p)) != 0) | |
8564 | { | |
8565 | for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q)) | |
8566 | if ((GET_CODE (q) != NOTE | |
8567 | || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END | |
8568 | || NOTE_LINE_NUMBER (q) == NOTE_INSN_SETJMP) | |
8569 | && (GET_CODE (q) != CODE_LABEL || LABEL_NUSES (q) != 0)) | |
8570 | break; | |
8571 | ||
8572 | /* If we ran into a BARRIER, this code is an extension of the | |
8573 | basic block when the branch is taken. */ | |
8b3686ed | 8574 | if (follow_jumps && q != 0 && GET_CODE (q) == BARRIER) |
7afe21cc RK |
8575 | { |
8576 | /* Don't allow ourself to keep walking around an | |
8577 | always-executed loop. */ | |
fc3ffe83 RK |
8578 | if (next_real_insn (q) == next) |
8579 | { | |
8580 | p = NEXT_INSN (p); | |
8581 | continue; | |
8582 | } | |
7afe21cc RK |
8583 | |
8584 | /* Similarly, don't put a branch in our path more than once. */ | |
8585 | for (i = 0; i < path_entry; i++) | |
8586 | if (data->path[i].branch == p) | |
8587 | break; | |
8588 | ||
8589 | if (i != path_entry) | |
8590 | break; | |
8591 | ||
8592 | data->path[path_entry].branch = p; | |
8593 | data->path[path_entry++].status = TAKEN; | |
8594 | ||
8595 | /* This branch now ends our path. It was possible that we | |
8596 | didn't see this branch the last time around (when the | |
8597 | insn in front of the target was a JUMP_INSN that was | |
8598 | turned into a no-op). */ | |
8599 | path_size = path_entry; | |
8600 | ||
8601 | p = JUMP_LABEL (p); | |
8602 | /* Mark block so we won't scan it again later. */ | |
8603 | PUT_MODE (NEXT_INSN (p), QImode); | |
8604 | } | |
8b3686ed RK |
8605 | /* Detect a branch around a block of code. */ |
8606 | else if (skip_blocks && q != 0 && GET_CODE (q) != CODE_LABEL) | |
8607 | { | |
8608 | register rtx tmp; | |
8609 | ||
fc3ffe83 RK |
8610 | if (next_real_insn (q) == next) |
8611 | { | |
8612 | p = NEXT_INSN (p); | |
8613 | continue; | |
8614 | } | |
8b3686ed RK |
8615 | |
8616 | for (i = 0; i < path_entry; i++) | |
8617 | if (data->path[i].branch == p) | |
8618 | break; | |
8619 | ||
8620 | if (i != path_entry) | |
8621 | break; | |
8622 | ||
8623 | /* This is no_labels_between_p (p, q) with an added check for | |
8624 | reaching the end of a function (in case Q precedes P). */ | |
8625 | for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp)) | |
8626 | if (GET_CODE (tmp) == CODE_LABEL) | |
8627 | break; | |
8628 | ||
8629 | if (tmp == q) | |
8630 | { | |
8631 | data->path[path_entry].branch = p; | |
8632 | data->path[path_entry++].status = AROUND; | |
8633 | ||
8634 | path_size = path_entry; | |
8635 | ||
8636 | p = JUMP_LABEL (p); | |
8637 | /* Mark block so we won't scan it again later. */ | |
8638 | PUT_MODE (NEXT_INSN (p), QImode); | |
8639 | } | |
8640 | } | |
7afe21cc | 8641 | } |
7afe21cc RK |
8642 | p = NEXT_INSN (p); |
8643 | } | |
8644 | ||
8645 | data->low_cuid = low_cuid; | |
8646 | data->high_cuid = high_cuid; | |
8647 | data->nsets = nsets; | |
8648 | data->last = p; | |
8649 | ||
8650 | /* If all jumps in the path are not taken, set our path length to zero | |
8651 | so a rescan won't be done. */ | |
8652 | for (i = path_size - 1; i >= 0; i--) | |
8b3686ed | 8653 | if (data->path[i].status != NOT_TAKEN) |
7afe21cc RK |
8654 | break; |
8655 | ||
8656 | if (i == -1) | |
8657 | data->path_size = 0; | |
8658 | else | |
8659 | data->path_size = path_size; | |
8660 | ||
8661 | /* End the current branch path. */ | |
8662 | data->path[path_size].branch = 0; | |
8663 | } | |
8664 | \f | |
7afe21cc RK |
8665 | /* Perform cse on the instructions of a function. |
8666 | F is the first instruction. | |
8667 | NREGS is one plus the highest pseudo-reg number used in the instruction. | |
8668 | ||
8669 | AFTER_LOOP is 1 if this is the cse call done after loop optimization | |
8670 | (only if -frerun-cse-after-loop). | |
8671 | ||
8672 | Returns 1 if jump_optimize should be redone due to simplifications | |
8673 | in conditional jump instructions. */ | |
8674 | ||
8675 | int | |
8676 | cse_main (f, nregs, after_loop, file) | |
8677 | rtx f; | |
8678 | int nregs; | |
8679 | int after_loop; | |
8680 | FILE *file; | |
8681 | { | |
8682 | struct cse_basic_block_data val; | |
8683 | register rtx insn = f; | |
8684 | register int i; | |
8685 | ||
8686 | cse_jumps_altered = 0; | |
a5dfb4ee | 8687 | recorded_label_ref = 0; |
7afe21cc RK |
8688 | constant_pool_entries_cost = 0; |
8689 | val.path_size = 0; | |
8690 | ||
8691 | init_recog (); | |
9ae8ffe7 | 8692 | init_alias_analysis (); |
7afe21cc RK |
8693 | |
8694 | max_reg = nregs; | |
8695 | ||
556c714b JW |
8696 | max_insn_uid = get_max_uid (); |
8697 | ||
7afe21cc RK |
8698 | reg_next_eqv = (int *) alloca (nregs * sizeof (int)); |
8699 | reg_prev_eqv = (int *) alloca (nregs * sizeof (int)); | |
7afe21cc | 8700 | |
7bac1be0 RK |
8701 | #ifdef LOAD_EXTEND_OP |
8702 | ||
8703 | /* Allocate scratch rtl here. cse_insn will fill in the memory reference | |
8704 | and change the code and mode as appropriate. */ | |
38a448ca | 8705 | memory_extend_rtx = gen_rtx_ZERO_EXTEND (VOIDmode, NULL_RTX); |
7bac1be0 RK |
8706 | #endif |
8707 | ||
7afe21cc RK |
8708 | /* Discard all the free elements of the previous function |
8709 | since they are allocated in the temporarily obstack. */ | |
4c9a05bc | 8710 | bzero ((char *) table, sizeof table); |
7afe21cc RK |
8711 | free_element_chain = 0; |
8712 | n_elements_made = 0; | |
8713 | ||
8714 | /* Find the largest uid. */ | |
8715 | ||
164c8956 RK |
8716 | max_uid = get_max_uid (); |
8717 | uid_cuid = (int *) alloca ((max_uid + 1) * sizeof (int)); | |
4c9a05bc | 8718 | bzero ((char *) uid_cuid, (max_uid + 1) * sizeof (int)); |
7afe21cc RK |
8719 | |
8720 | /* Compute the mapping from uids to cuids. | |
8721 | CUIDs are numbers assigned to insns, like uids, | |
8722 | except that cuids increase monotonically through the code. | |
8723 | Don't assign cuids to line-number NOTEs, so that the distance in cuids | |
8724 | between two insns is not affected by -g. */ | |
8725 | ||
8726 | for (insn = f, i = 0; insn; insn = NEXT_INSN (insn)) | |
8727 | { | |
8728 | if (GET_CODE (insn) != NOTE | |
8729 | || NOTE_LINE_NUMBER (insn) < 0) | |
8730 | INSN_CUID (insn) = ++i; | |
8731 | else | |
8732 | /* Give a line number note the same cuid as preceding insn. */ | |
8733 | INSN_CUID (insn) = i; | |
8734 | } | |
8735 | ||
8736 | /* Initialize which registers are clobbered by calls. */ | |
8737 | ||
8738 | CLEAR_HARD_REG_SET (regs_invalidated_by_call); | |
8739 | ||
8740 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
8741 | if ((call_used_regs[i] | |
8742 | /* Used to check !fixed_regs[i] here, but that isn't safe; | |
8743 | fixed regs are still call-clobbered, and sched can get | |
8744 | confused if they can "live across calls". | |
8745 | ||
8746 | The frame pointer is always preserved across calls. The arg | |
8747 | pointer is if it is fixed. The stack pointer usually is, unless | |
8748 | RETURN_POPS_ARGS, in which case an explicit CLOBBER | |
8749 | will be present. If we are generating PIC code, the PIC offset | |
8750 | table register is preserved across calls. */ | |
8751 | ||
8752 | && i != STACK_POINTER_REGNUM | |
8753 | && i != FRAME_POINTER_REGNUM | |
8bc169f2 DE |
8754 | #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM |
8755 | && i != HARD_FRAME_POINTER_REGNUM | |
8756 | #endif | |
7afe21cc RK |
8757 | #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM |
8758 | && ! (i == ARG_POINTER_REGNUM && fixed_regs[i]) | |
8759 | #endif | |
be8fe470 | 8760 | #if defined (PIC_OFFSET_TABLE_REGNUM) && !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED) |
7afe21cc RK |
8761 | && ! (i == PIC_OFFSET_TABLE_REGNUM && flag_pic) |
8762 | #endif | |
8763 | ) | |
8764 | || global_regs[i]) | |
8765 | SET_HARD_REG_BIT (regs_invalidated_by_call, i); | |
8766 | ||
8767 | /* Loop over basic blocks. | |
8768 | Compute the maximum number of qty's needed for each basic block | |
8769 | (which is 2 for each SET). */ | |
8770 | insn = f; | |
8771 | while (insn) | |
8772 | { | |
8b3686ed RK |
8773 | cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps, after_loop, |
8774 | flag_cse_skip_blocks); | |
7afe21cc RK |
8775 | |
8776 | /* If this basic block was already processed or has no sets, skip it. */ | |
8777 | if (val.nsets == 0 || GET_MODE (insn) == QImode) | |
8778 | { | |
8779 | PUT_MODE (insn, VOIDmode); | |
8780 | insn = (val.last ? NEXT_INSN (val.last) : 0); | |
8781 | val.path_size = 0; | |
8782 | continue; | |
8783 | } | |
8784 | ||
8785 | cse_basic_block_start = val.low_cuid; | |
8786 | cse_basic_block_end = val.high_cuid; | |
8787 | max_qty = val.nsets * 2; | |
8788 | ||
8789 | if (file) | |
ab87f8c8 | 8790 | fnotice (file, ";; Processing block from %d to %d, %d sets.\n", |
7afe21cc RK |
8791 | INSN_UID (insn), val.last ? INSN_UID (val.last) : 0, |
8792 | val.nsets); | |
8793 | ||
8794 | /* Make MAX_QTY bigger to give us room to optimize | |
8795 | past the end of this basic block, if that should prove useful. */ | |
8796 | if (max_qty < 500) | |
8797 | max_qty = 500; | |
8798 | ||
8799 | max_qty += max_reg; | |
8800 | ||
8801 | /* If this basic block is being extended by following certain jumps, | |
8802 | (see `cse_end_of_basic_block'), we reprocess the code from the start. | |
8803 | Otherwise, we start after this basic block. */ | |
8804 | if (val.path_size > 0) | |
8805 | cse_basic_block (insn, val.last, val.path, 0); | |
8806 | else | |
8807 | { | |
8808 | int old_cse_jumps_altered = cse_jumps_altered; | |
8809 | rtx temp; | |
8810 | ||
8811 | /* When cse changes a conditional jump to an unconditional | |
8812 | jump, we want to reprocess the block, since it will give | |
8813 | us a new branch path to investigate. */ | |
8814 | cse_jumps_altered = 0; | |
8815 | temp = cse_basic_block (insn, val.last, val.path, ! after_loop); | |
8b3686ed RK |
8816 | if (cse_jumps_altered == 0 |
8817 | || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0)) | |
7afe21cc RK |
8818 | insn = temp; |
8819 | ||
8820 | cse_jumps_altered |= old_cse_jumps_altered; | |
8821 | } | |
8822 | ||
8823 | #ifdef USE_C_ALLOCA | |
8824 | alloca (0); | |
8825 | #endif | |
8826 | } | |
8827 | ||
8828 | /* Tell refers_to_mem_p that qty_const info is not available. */ | |
8829 | qty_const = 0; | |
8830 | ||
8831 | if (max_elements_made < n_elements_made) | |
8832 | max_elements_made = n_elements_made; | |
8833 | ||
a5dfb4ee | 8834 | return cse_jumps_altered || recorded_label_ref; |
7afe21cc RK |
8835 | } |
8836 | ||
8837 | /* Process a single basic block. FROM and TO and the limits of the basic | |
8838 | block. NEXT_BRANCH points to the branch path when following jumps or | |
8839 | a null path when not following jumps. | |
8840 | ||
8841 | AROUND_LOOP is non-zero if we are to try to cse around to the start of a | |
8842 | loop. This is true when we are being called for the last time on a | |
8843 | block and this CSE pass is before loop.c. */ | |
8844 | ||
8845 | static rtx | |
8846 | cse_basic_block (from, to, next_branch, around_loop) | |
8847 | register rtx from, to; | |
8848 | struct branch_path *next_branch; | |
8849 | int around_loop; | |
8850 | { | |
8851 | register rtx insn; | |
8852 | int to_usage = 0; | |
7bd8b2a8 | 8853 | rtx libcall_insn = NULL_RTX; |
e9a25f70 | 8854 | int num_insns = 0; |
7afe21cc RK |
8855 | |
8856 | /* Each of these arrays is undefined before max_reg, so only allocate | |
8857 | the space actually needed and adjust the start below. */ | |
8858 | ||
8859 | qty_first_reg = (int *) alloca ((max_qty - max_reg) * sizeof (int)); | |
8860 | qty_last_reg = (int *) alloca ((max_qty - max_reg) * sizeof (int)); | |
8861 | qty_mode= (enum machine_mode *) alloca ((max_qty - max_reg) * sizeof (enum machine_mode)); | |
8862 | qty_const = (rtx *) alloca ((max_qty - max_reg) * sizeof (rtx)); | |
8863 | qty_const_insn = (rtx *) alloca ((max_qty - max_reg) * sizeof (rtx)); | |
8864 | qty_comparison_code | |
8865 | = (enum rtx_code *) alloca ((max_qty - max_reg) * sizeof (enum rtx_code)); | |
8866 | qty_comparison_qty = (int *) alloca ((max_qty - max_reg) * sizeof (int)); | |
8867 | qty_comparison_const = (rtx *) alloca ((max_qty - max_reg) * sizeof (rtx)); | |
8868 | ||
8869 | qty_first_reg -= max_reg; | |
8870 | qty_last_reg -= max_reg; | |
8871 | qty_mode -= max_reg; | |
8872 | qty_const -= max_reg; | |
8873 | qty_const_insn -= max_reg; | |
8874 | qty_comparison_code -= max_reg; | |
8875 | qty_comparison_qty -= max_reg; | |
8876 | qty_comparison_const -= max_reg; | |
8877 | ||
8878 | new_basic_block (); | |
8879 | ||
8880 | /* TO might be a label. If so, protect it from being deleted. */ | |
8881 | if (to != 0 && GET_CODE (to) == CODE_LABEL) | |
8882 | ++LABEL_NUSES (to); | |
8883 | ||
8884 | for (insn = from; insn != to; insn = NEXT_INSN (insn)) | |
8885 | { | |
1d22a2c1 | 8886 | register enum rtx_code code = GET_CODE (insn); |
e9a25f70 | 8887 | |
1d22a2c1 MM |
8888 | /* If we have processed 1,000 insns, flush the hash table to |
8889 | avoid extreme quadratic behavior. We must not include NOTEs | |
8890 | in the count since there may be more or them when generating | |
8891 | debugging information. If we clear the table at different | |
8892 | times, code generated with -g -O might be different than code | |
8893 | generated with -O but not -g. | |
e9a25f70 JL |
8894 | |
8895 | ??? This is a real kludge and needs to be done some other way. | |
8896 | Perhaps for 2.9. */ | |
1d22a2c1 | 8897 | if (code != NOTE && num_insns++ > 1000) |
e9a25f70 | 8898 | { |
01e752d3 | 8899 | flush_hash_table (); |
e9a25f70 JL |
8900 | num_insns = 0; |
8901 | } | |
7afe21cc RK |
8902 | |
8903 | /* See if this is a branch that is part of the path. If so, and it is | |
8904 | to be taken, do so. */ | |
8905 | if (next_branch->branch == insn) | |
8906 | { | |
8b3686ed RK |
8907 | enum taken status = next_branch++->status; |
8908 | if (status != NOT_TAKEN) | |
7afe21cc | 8909 | { |
8b3686ed RK |
8910 | if (status == TAKEN) |
8911 | record_jump_equiv (insn, 1); | |
8912 | else | |
8913 | invalidate_skipped_block (NEXT_INSN (insn)); | |
8914 | ||
7afe21cc RK |
8915 | /* Set the last insn as the jump insn; it doesn't affect cc0. |
8916 | Then follow this branch. */ | |
8917 | #ifdef HAVE_cc0 | |
8918 | prev_insn_cc0 = 0; | |
8919 | #endif | |
8920 | prev_insn = insn; | |
8921 | insn = JUMP_LABEL (insn); | |
8922 | continue; | |
8923 | } | |
8924 | } | |
8925 | ||
7afe21cc RK |
8926 | if (GET_MODE (insn) == QImode) |
8927 | PUT_MODE (insn, VOIDmode); | |
8928 | ||
8929 | if (GET_RTX_CLASS (code) == 'i') | |
8930 | { | |
7bd8b2a8 JL |
8931 | rtx p; |
8932 | ||
7afe21cc RK |
8933 | /* Process notes first so we have all notes in canonical forms when |
8934 | looking for duplicate operations. */ | |
8935 | ||
8936 | if (REG_NOTES (insn)) | |
906c4e36 | 8937 | REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX); |
7afe21cc RK |
8938 | |
8939 | /* Track when we are inside in LIBCALL block. Inside such a block, | |
8940 | we do not want to record destinations. The last insn of a | |
8941 | LIBCALL block is not considered to be part of the block, since | |
830a38ee | 8942 | its destination is the result of the block and hence should be |
7afe21cc RK |
8943 | recorded. */ |
8944 | ||
63be02db | 8945 | if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX))) |
7bd8b2a8 | 8946 | libcall_insn = XEXP (p, 0); |
906c4e36 | 8947 | else if (find_reg_note (insn, REG_RETVAL, NULL_RTX)) |
7bd8b2a8 | 8948 | libcall_insn = NULL_RTX; |
7afe21cc | 8949 | |
7bd8b2a8 | 8950 | cse_insn (insn, libcall_insn); |
7afe21cc RK |
8951 | } |
8952 | ||
8953 | /* If INSN is now an unconditional jump, skip to the end of our | |
8954 | basic block by pretending that we just did the last insn in the | |
8955 | basic block. If we are jumping to the end of our block, show | |
8956 | that we can have one usage of TO. */ | |
8957 | ||
8958 | if (simplejump_p (insn)) | |
8959 | { | |
8960 | if (to == 0) | |
8961 | return 0; | |
8962 | ||
8963 | if (JUMP_LABEL (insn) == to) | |
8964 | to_usage = 1; | |
8965 | ||
6a5293dc RS |
8966 | /* Maybe TO was deleted because the jump is unconditional. |
8967 | If so, there is nothing left in this basic block. */ | |
8968 | /* ??? Perhaps it would be smarter to set TO | |
8969 | to whatever follows this insn, | |
8970 | and pretend the basic block had always ended here. */ | |
8971 | if (INSN_DELETED_P (to)) | |
8972 | break; | |
8973 | ||
7afe21cc RK |
8974 | insn = PREV_INSN (to); |
8975 | } | |
8976 | ||
8977 | /* See if it is ok to keep on going past the label | |
8978 | which used to end our basic block. Remember that we incremented | |
d45cf215 | 8979 | the count of that label, so we decrement it here. If we made |
7afe21cc RK |
8980 | a jump unconditional, TO_USAGE will be one; in that case, we don't |
8981 | want to count the use in that jump. */ | |
8982 | ||
8983 | if (to != 0 && NEXT_INSN (insn) == to | |
8984 | && GET_CODE (to) == CODE_LABEL && --LABEL_NUSES (to) == to_usage) | |
8985 | { | |
8986 | struct cse_basic_block_data val; | |
146135d6 | 8987 | rtx prev; |
7afe21cc RK |
8988 | |
8989 | insn = NEXT_INSN (to); | |
8990 | ||
8991 | if (LABEL_NUSES (to) == 0) | |
146135d6 | 8992 | insn = delete_insn (to); |
7afe21cc | 8993 | |
146135d6 RK |
8994 | /* If TO was the last insn in the function, we are done. */ |
8995 | if (insn == 0) | |
7afe21cc RK |
8996 | return 0; |
8997 | ||
146135d6 RK |
8998 | /* If TO was preceded by a BARRIER we are done with this block |
8999 | because it has no continuation. */ | |
9000 | prev = prev_nonnote_insn (to); | |
9001 | if (prev && GET_CODE (prev) == BARRIER) | |
9002 | return insn; | |
9003 | ||
9004 | /* Find the end of the following block. Note that we won't be | |
9005 | following branches in this case. */ | |
7afe21cc RK |
9006 | to_usage = 0; |
9007 | val.path_size = 0; | |
8b3686ed | 9008 | cse_end_of_basic_block (insn, &val, 0, 0, 0); |
7afe21cc RK |
9009 | |
9010 | /* If the tables we allocated have enough space left | |
9011 | to handle all the SETs in the next basic block, | |
9012 | continue through it. Otherwise, return, | |
9013 | and that block will be scanned individually. */ | |
9014 | if (val.nsets * 2 + next_qty > max_qty) | |
9015 | break; | |
9016 | ||
9017 | cse_basic_block_start = val.low_cuid; | |
9018 | cse_basic_block_end = val.high_cuid; | |
9019 | to = val.last; | |
9020 | ||
9021 | /* Prevent TO from being deleted if it is a label. */ | |
9022 | if (to != 0 && GET_CODE (to) == CODE_LABEL) | |
9023 | ++LABEL_NUSES (to); | |
9024 | ||
9025 | /* Back up so we process the first insn in the extension. */ | |
9026 | insn = PREV_INSN (insn); | |
9027 | } | |
9028 | } | |
9029 | ||
9030 | if (next_qty > max_qty) | |
9031 | abort (); | |
9032 | ||
9033 | /* If we are running before loop.c, we stopped on a NOTE_INSN_LOOP_END, and | |
9034 | the previous insn is the only insn that branches to the head of a loop, | |
9035 | we can cse into the loop. Don't do this if we changed the jump | |
9036 | structure of a loop unless we aren't going to be following jumps. */ | |
9037 | ||
8b3686ed RK |
9038 | if ((cse_jumps_altered == 0 |
9039 | || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0)) | |
7afe21cc RK |
9040 | && around_loop && to != 0 |
9041 | && GET_CODE (to) == NOTE && NOTE_LINE_NUMBER (to) == NOTE_INSN_LOOP_END | |
9042 | && GET_CODE (PREV_INSN (to)) == JUMP_INSN | |
9043 | && JUMP_LABEL (PREV_INSN (to)) != 0 | |
9044 | && LABEL_NUSES (JUMP_LABEL (PREV_INSN (to))) == 1) | |
9045 | cse_around_loop (JUMP_LABEL (PREV_INSN (to))); | |
9046 | ||
9047 | return to ? NEXT_INSN (to) : 0; | |
9048 | } | |
9049 | \f | |
9050 | /* Count the number of times registers are used (not set) in X. | |
9051 | COUNTS is an array in which we accumulate the count, INCR is how much | |
79644f06 RK |
9052 | we count each register usage. |
9053 | ||
9054 | Don't count a usage of DEST, which is the SET_DEST of a SET which | |
9055 | contains X in its SET_SRC. This is because such a SET does not | |
9056 | modify the liveness of DEST. */ | |
7afe21cc RK |
9057 | |
9058 | static void | |
79644f06 | 9059 | count_reg_usage (x, counts, dest, incr) |
7afe21cc RK |
9060 | rtx x; |
9061 | int *counts; | |
79644f06 | 9062 | rtx dest; |
7afe21cc RK |
9063 | int incr; |
9064 | { | |
f1e7c95f | 9065 | enum rtx_code code; |
7afe21cc RK |
9066 | char *fmt; |
9067 | int i, j; | |
9068 | ||
f1e7c95f RK |
9069 | if (x == 0) |
9070 | return; | |
9071 | ||
9072 | switch (code = GET_CODE (x)) | |
7afe21cc RK |
9073 | { |
9074 | case REG: | |
79644f06 RK |
9075 | if (x != dest) |
9076 | counts[REGNO (x)] += incr; | |
7afe21cc RK |
9077 | return; |
9078 | ||
9079 | case PC: | |
9080 | case CC0: | |
9081 | case CONST: | |
9082 | case CONST_INT: | |
9083 | case CONST_DOUBLE: | |
9084 | case SYMBOL_REF: | |
9085 | case LABEL_REF: | |
02e39abc JL |
9086 | return; |
9087 | ||
9088 | case CLOBBER: | |
9089 | /* If we are clobbering a MEM, mark any registers inside the address | |
9090 | as being used. */ | |
9091 | if (GET_CODE (XEXP (x, 0)) == MEM) | |
9092 | count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr); | |
7afe21cc RK |
9093 | return; |
9094 | ||
9095 | case SET: | |
9096 | /* Unless we are setting a REG, count everything in SET_DEST. */ | |
9097 | if (GET_CODE (SET_DEST (x)) != REG) | |
79644f06 | 9098 | count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr); |
9ff08f70 RK |
9099 | |
9100 | /* If SRC has side-effects, then we can't delete this insn, so the | |
9101 | usage of SET_DEST inside SRC counts. | |
9102 | ||
9103 | ??? Strictly-speaking, we might be preserving this insn | |
9104 | because some other SET has side-effects, but that's hard | |
9105 | to do and can't happen now. */ | |
9106 | count_reg_usage (SET_SRC (x), counts, | |
9107 | side_effects_p (SET_SRC (x)) ? NULL_RTX : SET_DEST (x), | |
9108 | incr); | |
7afe21cc RK |
9109 | return; |
9110 | ||
f1e7c95f RK |
9111 | case CALL_INSN: |
9112 | count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, NULL_RTX, incr); | |
9113 | ||
9114 | /* ... falls through ... */ | |
7afe21cc RK |
9115 | case INSN: |
9116 | case JUMP_INSN: | |
79644f06 | 9117 | count_reg_usage (PATTERN (x), counts, NULL_RTX, incr); |
7afe21cc RK |
9118 | |
9119 | /* Things used in a REG_EQUAL note aren't dead since loop may try to | |
9120 | use them. */ | |
9121 | ||
f1e7c95f | 9122 | count_reg_usage (REG_NOTES (x), counts, NULL_RTX, incr); |
7afe21cc RK |
9123 | return; |
9124 | ||
9125 | case EXPR_LIST: | |
9126 | case INSN_LIST: | |
f1e7c95f | 9127 | if (REG_NOTE_KIND (x) == REG_EQUAL |
c6a26dc4 | 9128 | || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)) |
79644f06 | 9129 | count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr); |
f1e7c95f | 9130 | count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr); |
7afe21cc | 9131 | return; |
e9a25f70 JL |
9132 | |
9133 | default: | |
9134 | break; | |
7afe21cc RK |
9135 | } |
9136 | ||
9137 | fmt = GET_RTX_FORMAT (code); | |
9138 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
9139 | { | |
9140 | if (fmt[i] == 'e') | |
79644f06 | 9141 | count_reg_usage (XEXP (x, i), counts, dest, incr); |
7afe21cc RK |
9142 | else if (fmt[i] == 'E') |
9143 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
79644f06 | 9144 | count_reg_usage (XVECEXP (x, i, j), counts, dest, incr); |
7afe21cc RK |
9145 | } |
9146 | } | |
9147 | \f | |
9148 | /* Scan all the insns and delete any that are dead; i.e., they store a register | |
9149 | that is never used or they copy a register to itself. | |
9150 | ||
c6a26dc4 JL |
9151 | This is used to remove insns made obviously dead by cse, loop or other |
9152 | optimizations. It improves the heuristics in loop since it won't try to | |
9153 | move dead invariants out of loops or make givs for dead quantities. The | |
9154 | remaining passes of the compilation are also sped up. */ | |
7afe21cc RK |
9155 | |
9156 | void | |
c6a26dc4 | 9157 | delete_trivially_dead_insns (insns, nreg) |
7afe21cc RK |
9158 | rtx insns; |
9159 | int nreg; | |
9160 | { | |
9161 | int *counts = (int *) alloca (nreg * sizeof (int)); | |
77fa0940 | 9162 | rtx insn, prev; |
51723711 | 9163 | #ifdef HAVE_cc0 |
d45cf215 | 9164 | rtx tem; |
51723711 | 9165 | #endif |
7afe21cc | 9166 | int i; |
614bb5d4 | 9167 | int in_libcall = 0, dead_libcall = 0; |
7afe21cc RK |
9168 | |
9169 | /* First count the number of times each register is used. */ | |
4c9a05bc | 9170 | bzero ((char *) counts, sizeof (int) * nreg); |
7afe21cc | 9171 | for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn)) |
79644f06 | 9172 | count_reg_usage (insn, counts, NULL_RTX, 1); |
7afe21cc RK |
9173 | |
9174 | /* Go from the last insn to the first and delete insns that only set unused | |
9175 | registers or copy a register to itself. As we delete an insn, remove | |
9176 | usage counts for registers it uses. */ | |
77fa0940 | 9177 | for (insn = prev_real_insn (get_last_insn ()); insn; insn = prev) |
7afe21cc RK |
9178 | { |
9179 | int live_insn = 0; | |
614bb5d4 | 9180 | rtx note; |
7afe21cc | 9181 | |
77fa0940 RK |
9182 | prev = prev_real_insn (insn); |
9183 | ||
614bb5d4 JL |
9184 | /* Don't delete any insns that are part of a libcall block unless |
9185 | we can delete the whole libcall block. | |
9186 | ||
77fa0940 RK |
9187 | Flow or loop might get confused if we did that. Remember |
9188 | that we are scanning backwards. */ | |
906c4e36 | 9189 | if (find_reg_note (insn, REG_RETVAL, NULL_RTX)) |
614bb5d4 JL |
9190 | { |
9191 | in_libcall = 1; | |
9192 | live_insn = 1; | |
9193 | dead_libcall = 0; | |
e4890d45 | 9194 | |
614bb5d4 JL |
9195 | /* See if there's a REG_EQUAL note on this insn and try to |
9196 | replace the source with the REG_EQUAL expression. | |
9197 | ||
9198 | We assume that insns with REG_RETVALs can only be reg->reg | |
9199 | copies at this point. */ | |
9200 | note = find_reg_note (insn, REG_EQUAL, NULL_RTX); | |
9201 | if (note) | |
9202 | { | |
9203 | rtx set = single_set (insn); | |
9204 | if (set | |
9205 | && validate_change (insn, &SET_SRC (set), XEXP (note, 0), 0)) | |
9206 | { | |
9207 | remove_note (insn, | |
9208 | find_reg_note (insn, REG_RETVAL, NULL_RTX)); | |
9209 | dead_libcall = 1; | |
9210 | } | |
9211 | } | |
9212 | } | |
9213 | else if (in_libcall) | |
9214 | live_insn = ! dead_libcall; | |
e4890d45 | 9215 | else if (GET_CODE (PATTERN (insn)) == SET) |
7afe21cc RK |
9216 | { |
9217 | if (GET_CODE (SET_DEST (PATTERN (insn))) == REG | |
9218 | && SET_DEST (PATTERN (insn)) == SET_SRC (PATTERN (insn))) | |
9219 | ; | |
9220 | ||
d45cf215 RS |
9221 | #ifdef HAVE_cc0 |
9222 | else if (GET_CODE (SET_DEST (PATTERN (insn))) == CC0 | |
9223 | && ! side_effects_p (SET_SRC (PATTERN (insn))) | |
9224 | && ((tem = next_nonnote_insn (insn)) == 0 | |
9225 | || GET_RTX_CLASS (GET_CODE (tem)) != 'i' | |
9226 | || ! reg_referenced_p (cc0_rtx, PATTERN (tem)))) | |
9227 | ; | |
9228 | #endif | |
7afe21cc RK |
9229 | else if (GET_CODE (SET_DEST (PATTERN (insn))) != REG |
9230 | || REGNO (SET_DEST (PATTERN (insn))) < FIRST_PSEUDO_REGISTER | |
9231 | || counts[REGNO (SET_DEST (PATTERN (insn)))] != 0 | |
9232 | || side_effects_p (SET_SRC (PATTERN (insn)))) | |
9233 | live_insn = 1; | |
9234 | } | |
9235 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
9236 | for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) | |
9237 | { | |
9238 | rtx elt = XVECEXP (PATTERN (insn), 0, i); | |
9239 | ||
9240 | if (GET_CODE (elt) == SET) | |
9241 | { | |
9242 | if (GET_CODE (SET_DEST (elt)) == REG | |
9243 | && SET_DEST (elt) == SET_SRC (elt)) | |
9244 | ; | |
9245 | ||
d45cf215 RS |
9246 | #ifdef HAVE_cc0 |
9247 | else if (GET_CODE (SET_DEST (elt)) == CC0 | |
9248 | && ! side_effects_p (SET_SRC (elt)) | |
9249 | && ((tem = next_nonnote_insn (insn)) == 0 | |
9250 | || GET_RTX_CLASS (GET_CODE (tem)) != 'i' | |
9251 | || ! reg_referenced_p (cc0_rtx, PATTERN (tem)))) | |
9252 | ; | |
9253 | #endif | |
7afe21cc RK |
9254 | else if (GET_CODE (SET_DEST (elt)) != REG |
9255 | || REGNO (SET_DEST (elt)) < FIRST_PSEUDO_REGISTER | |
9256 | || counts[REGNO (SET_DEST (elt))] != 0 | |
9257 | || side_effects_p (SET_SRC (elt))) | |
9258 | live_insn = 1; | |
9259 | } | |
9260 | else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE) | |
9261 | live_insn = 1; | |
9262 | } | |
9263 | else | |
9264 | live_insn = 1; | |
9265 | ||
9266 | /* If this is a dead insn, delete it and show registers in it aren't | |
e4890d45 | 9267 | being used. */ |
7afe21cc | 9268 | |
e4890d45 | 9269 | if (! live_insn) |
7afe21cc | 9270 | { |
79644f06 | 9271 | count_reg_usage (insn, counts, NULL_RTX, -1); |
77fa0940 | 9272 | delete_insn (insn); |
7afe21cc | 9273 | } |
e4890d45 | 9274 | |
906c4e36 | 9275 | if (find_reg_note (insn, REG_LIBCALL, NULL_RTX)) |
614bb5d4 JL |
9276 | { |
9277 | in_libcall = 0; | |
9278 | dead_libcall = 0; | |
9279 | } | |
7afe21cc RK |
9280 | } |
9281 | } |