]>
Commit | Line | Data |
---|---|---|
7afe21cc | 1 | /* Common subexpression elimination for GNU compiler. |
5e7b4e25 | 2 | Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998 |
26d107db | 3 | 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc. |
7afe21cc | 4 | |
1322177d | 5 | This file is part of GCC. |
7afe21cc | 6 | |
1322177d LB |
7 | GCC is free software; you can redistribute it and/or modify it under |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 2, or (at your option) any later | |
10 | version. | |
7afe21cc | 11 | |
1322177d LB |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
7afe21cc RK |
16 | |
17 | You should have received a copy of the GNU General Public License | |
1322177d LB |
18 | along with GCC; see the file COPYING. If not, write to the Free |
19 | Software Foundation, 59 Temple Place - Suite 330, Boston, MA | |
20 | 02111-1307, USA. */ | |
7afe21cc | 21 | |
7afe21cc | 22 | #include "config.h" |
670ee920 KG |
23 | /* stdio.h must precede rtl.h for FFS. */ |
24 | #include "system.h" | |
4977bab6 ZW |
25 | #include "coretypes.h" |
26 | #include "tm.h" | |
9c3b4c8b | 27 | |
7afe21cc | 28 | #include "rtl.h" |
6baf1cc8 | 29 | #include "tm_p.h" |
7afe21cc RK |
30 | #include "regs.h" |
31 | #include "hard-reg-set.h" | |
630c79be | 32 | #include "basic-block.h" |
7afe21cc RK |
33 | #include "flags.h" |
34 | #include "real.h" | |
35 | #include "insn-config.h" | |
36 | #include "recog.h" | |
49ad7cfa | 37 | #include "function.h" |
956d6950 | 38 | #include "expr.h" |
50b2596f KG |
39 | #include "toplev.h" |
40 | #include "output.h" | |
1497faf6 | 41 | #include "ggc.h" |
3dec4024 | 42 | #include "timevar.h" |
26771da7 | 43 | #include "except.h" |
3c50106f | 44 | #include "target.h" |
9bf8cfbf | 45 | #include "params.h" |
2f93eea8 | 46 | #include "rtlhooks-def.h" |
7afe21cc RK |
47 | |
48 | /* The basic idea of common subexpression elimination is to go | |
49 | through the code, keeping a record of expressions that would | |
50 | have the same value at the current scan point, and replacing | |
51 | expressions encountered with the cheapest equivalent expression. | |
52 | ||
53 | It is too complicated to keep track of the different possibilities | |
e48a7fbe JL |
54 | when control paths merge in this code; so, at each label, we forget all |
55 | that is known and start fresh. This can be described as processing each | |
56 | extended basic block separately. We have a separate pass to perform | |
57 | global CSE. | |
58 | ||
59 | Note CSE can turn a conditional or computed jump into a nop or | |
60 | an unconditional jump. When this occurs we arrange to run the jump | |
61 | optimizer after CSE to delete the unreachable code. | |
7afe21cc RK |
62 | |
63 | We use two data structures to record the equivalent expressions: | |
1bb98cec DM |
64 | a hash table for most expressions, and a vector of "quantity |
65 | numbers" to record equivalent (pseudo) registers. | |
7afe21cc RK |
66 | |
67 | The use of the special data structure for registers is desirable | |
68 | because it is faster. It is possible because registers references | |
69 | contain a fairly small number, the register number, taken from | |
70 | a contiguously allocated series, and two register references are | |
71 | identical if they have the same number. General expressions | |
72 | do not have any such thing, so the only way to retrieve the | |
73 | information recorded on an expression other than a register | |
74 | is to keep it in a hash table. | |
75 | ||
76 | Registers and "quantity numbers": | |
278a83b2 | 77 | |
7afe21cc RK |
78 | At the start of each basic block, all of the (hardware and pseudo) |
79 | registers used in the function are given distinct quantity | |
80 | numbers to indicate their contents. During scan, when the code | |
81 | copies one register into another, we copy the quantity number. | |
82 | When a register is loaded in any other way, we allocate a new | |
83 | quantity number to describe the value generated by this operation. | |
84 | `reg_qty' records what quantity a register is currently thought | |
85 | of as containing. | |
86 | ||
87 | All real quantity numbers are greater than or equal to `max_reg'. | |
88 | If register N has not been assigned a quantity, reg_qty[N] will equal N. | |
89 | ||
1bb98cec DM |
90 | Quantity numbers below `max_reg' do not exist and none of the `qty_table' |
91 | entries should be referenced with an index below `max_reg'. | |
7afe21cc RK |
92 | |
93 | We also maintain a bidirectional chain of registers for each | |
1bb98cec DM |
94 | quantity number. The `qty_table` members `first_reg' and `last_reg', |
95 | and `reg_eqv_table' members `next' and `prev' hold these chains. | |
7afe21cc RK |
96 | |
97 | The first register in a chain is the one whose lifespan is least local. | |
98 | Among equals, it is the one that was seen first. | |
99 | We replace any equivalent register with that one. | |
100 | ||
101 | If two registers have the same quantity number, it must be true that | |
1bb98cec | 102 | REG expressions with qty_table `mode' must be in the hash table for both |
7afe21cc RK |
103 | registers and must be in the same class. |
104 | ||
105 | The converse is not true. Since hard registers may be referenced in | |
106 | any mode, two REG expressions might be equivalent in the hash table | |
107 | but not have the same quantity number if the quantity number of one | |
108 | of the registers is not the same mode as those expressions. | |
278a83b2 | 109 | |
7afe21cc RK |
110 | Constants and quantity numbers |
111 | ||
112 | When a quantity has a known constant value, that value is stored | |
1bb98cec | 113 | in the appropriate qty_table `const_rtx'. This is in addition to |
7afe21cc RK |
114 | putting the constant in the hash table as is usual for non-regs. |
115 | ||
d45cf215 | 116 | Whether a reg or a constant is preferred is determined by the configuration |
7afe21cc RK |
117 | macro CONST_COSTS and will often depend on the constant value. In any |
118 | event, expressions containing constants can be simplified, by fold_rtx. | |
119 | ||
120 | When a quantity has a known nearly constant value (such as an address | |
1bb98cec DM |
121 | of a stack slot), that value is stored in the appropriate qty_table |
122 | `const_rtx'. | |
7afe21cc RK |
123 | |
124 | Integer constants don't have a machine mode. However, cse | |
125 | determines the intended machine mode from the destination | |
126 | of the instruction that moves the constant. The machine mode | |
127 | is recorded in the hash table along with the actual RTL | |
128 | constant expression so that different modes are kept separate. | |
129 | ||
130 | Other expressions: | |
131 | ||
132 | To record known equivalences among expressions in general | |
133 | we use a hash table called `table'. It has a fixed number of buckets | |
134 | that contain chains of `struct table_elt' elements for expressions. | |
135 | These chains connect the elements whose expressions have the same | |
136 | hash codes. | |
137 | ||
138 | Other chains through the same elements connect the elements which | |
139 | currently have equivalent values. | |
140 | ||
141 | Register references in an expression are canonicalized before hashing | |
1bb98cec | 142 | the expression. This is done using `reg_qty' and qty_table `first_reg'. |
7afe21cc RK |
143 | The hash code of a register reference is computed using the quantity |
144 | number, not the register number. | |
145 | ||
146 | When the value of an expression changes, it is necessary to remove from the | |
147 | hash table not just that expression but all expressions whose values | |
148 | could be different as a result. | |
149 | ||
150 | 1. If the value changing is in memory, except in special cases | |
151 | ANYTHING referring to memory could be changed. That is because | |
152 | nobody knows where a pointer does not point. | |
153 | The function `invalidate_memory' removes what is necessary. | |
154 | ||
155 | The special cases are when the address is constant or is | |
156 | a constant plus a fixed register such as the frame pointer | |
157 | or a static chain pointer. When such addresses are stored in, | |
158 | we can tell exactly which other such addresses must be invalidated | |
159 | due to overlap. `invalidate' does this. | |
160 | All expressions that refer to non-constant | |
161 | memory addresses are also invalidated. `invalidate_memory' does this. | |
162 | ||
163 | 2. If the value changing is a register, all expressions | |
164 | containing references to that register, and only those, | |
165 | must be removed. | |
166 | ||
167 | Because searching the entire hash table for expressions that contain | |
168 | a register is very slow, we try to figure out when it isn't necessary. | |
169 | Precisely, this is necessary only when expressions have been | |
170 | entered in the hash table using this register, and then the value has | |
171 | changed, and then another expression wants to be added to refer to | |
172 | the register's new value. This sequence of circumstances is rare | |
173 | within any one basic block. | |
174 | ||
175 | The vectors `reg_tick' and `reg_in_table' are used to detect this case. | |
176 | reg_tick[i] is incremented whenever a value is stored in register i. | |
177 | reg_in_table[i] holds -1 if no references to register i have been | |
178 | entered in the table; otherwise, it contains the value reg_tick[i] had | |
179 | when the references were entered. If we want to enter a reference | |
180 | and reg_in_table[i] != reg_tick[i], we must scan and remove old references. | |
181 | Until we want to enter a new entry, the mere fact that the two vectors | |
182 | don't match makes the entries be ignored if anyone tries to match them. | |
183 | ||
184 | Registers themselves are entered in the hash table as well as in | |
185 | the equivalent-register chains. However, the vectors `reg_tick' | |
186 | and `reg_in_table' do not apply to expressions which are simple | |
187 | register references. These expressions are removed from the table | |
188 | immediately when they become invalid, and this can be done even if | |
189 | we do not immediately search for all the expressions that refer to | |
190 | the register. | |
191 | ||
192 | A CLOBBER rtx in an instruction invalidates its operand for further | |
193 | reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK | |
194 | invalidates everything that resides in memory. | |
195 | ||
196 | Related expressions: | |
197 | ||
198 | Constant expressions that differ only by an additive integer | |
199 | are called related. When a constant expression is put in | |
200 | the table, the related expression with no constant term | |
201 | is also entered. These are made to point at each other | |
202 | so that it is possible to find out if there exists any | |
203 | register equivalent to an expression related to a given expression. */ | |
278a83b2 | 204 | |
7afe21cc RK |
205 | /* One plus largest register number used in this function. */ |
206 | ||
207 | static int max_reg; | |
208 | ||
556c714b JW |
209 | /* One plus largest instruction UID used in this function at time of |
210 | cse_main call. */ | |
211 | ||
212 | static int max_insn_uid; | |
213 | ||
1bb98cec DM |
214 | /* Length of qty_table vector. We know in advance we will not need |
215 | a quantity number this big. */ | |
7afe21cc RK |
216 | |
217 | static int max_qty; | |
218 | ||
219 | /* Next quantity number to be allocated. | |
220 | This is 1 + the largest number needed so far. */ | |
221 | ||
222 | static int next_qty; | |
223 | ||
1bb98cec | 224 | /* Per-qty information tracking. |
7afe21cc | 225 | |
1bb98cec DM |
226 | `first_reg' and `last_reg' track the head and tail of the |
227 | chain of registers which currently contain this quantity. | |
7afe21cc | 228 | |
1bb98cec | 229 | `mode' contains the machine mode of this quantity. |
7afe21cc | 230 | |
1bb98cec DM |
231 | `const_rtx' holds the rtx of the constant value of this |
232 | quantity, if known. A summations of the frame/arg pointer | |
233 | and a constant can also be entered here. When this holds | |
234 | a known value, `const_insn' is the insn which stored the | |
235 | constant value. | |
7afe21cc | 236 | |
1bb98cec DM |
237 | `comparison_{code,const,qty}' are used to track when a |
238 | comparison between a quantity and some constant or register has | |
239 | been passed. In such a case, we know the results of the comparison | |
240 | in case we see it again. These members record a comparison that | |
241 | is known to be true. `comparison_code' holds the rtx code of such | |
242 | a comparison, else it is set to UNKNOWN and the other two | |
243 | comparison members are undefined. `comparison_const' holds | |
244 | the constant being compared against, or zero if the comparison | |
245 | is not against a constant. `comparison_qty' holds the quantity | |
246 | being compared against when the result is known. If the comparison | |
247 | is not with a register, `comparison_qty' is -1. */ | |
7afe21cc | 248 | |
1bb98cec DM |
249 | struct qty_table_elem |
250 | { | |
251 | rtx const_rtx; | |
252 | rtx const_insn; | |
253 | rtx comparison_const; | |
254 | int comparison_qty; | |
770ae6cc | 255 | unsigned int first_reg, last_reg; |
496324d0 DN |
256 | /* The sizes of these fields should match the sizes of the |
257 | code and mode fields of struct rtx_def (see rtl.h). */ | |
258 | ENUM_BITFIELD(rtx_code) comparison_code : 16; | |
259 | ENUM_BITFIELD(machine_mode) mode : 8; | |
1bb98cec | 260 | }; |
7afe21cc | 261 | |
1bb98cec DM |
262 | /* The table of all qtys, indexed by qty number. */ |
263 | static struct qty_table_elem *qty_table; | |
7afe21cc RK |
264 | |
265 | #ifdef HAVE_cc0 | |
266 | /* For machines that have a CC0, we do not record its value in the hash | |
267 | table since its use is guaranteed to be the insn immediately following | |
268 | its definition and any other insn is presumed to invalidate it. | |
269 | ||
270 | Instead, we store below the value last assigned to CC0. If it should | |
271 | happen to be a constant, it is stored in preference to the actual | |
272 | assigned value. In case it is a constant, we store the mode in which | |
273 | the constant should be interpreted. */ | |
274 | ||
275 | static rtx prev_insn_cc0; | |
276 | static enum machine_mode prev_insn_cc0_mode; | |
7afe21cc RK |
277 | |
278 | /* Previous actual insn. 0 if at first insn of basic block. */ | |
279 | ||
280 | static rtx prev_insn; | |
4977bab6 | 281 | #endif |
7afe21cc RK |
282 | |
283 | /* Insn being scanned. */ | |
284 | ||
285 | static rtx this_insn; | |
286 | ||
71d306d1 DE |
287 | /* Index by register number, gives the number of the next (or |
288 | previous) register in the chain of registers sharing the same | |
7afe21cc RK |
289 | value. |
290 | ||
291 | Or -1 if this register is at the end of the chain. | |
292 | ||
1bb98cec DM |
293 | If reg_qty[N] == N, reg_eqv_table[N].next is undefined. */ |
294 | ||
295 | /* Per-register equivalence chain. */ | |
296 | struct reg_eqv_elem | |
297 | { | |
298 | int next, prev; | |
299 | }; | |
7afe21cc | 300 | |
1bb98cec DM |
301 | /* The table of all register equivalence chains. */ |
302 | static struct reg_eqv_elem *reg_eqv_table; | |
7afe21cc | 303 | |
14a774a9 RK |
304 | struct cse_reg_info |
305 | { | |
9b1549b8 DM |
306 | /* Next in hash chain. */ |
307 | struct cse_reg_info *hash_next; | |
c1edba58 VM |
308 | |
309 | /* The next cse_reg_info structure in the free or used list. */ | |
14a774a9 | 310 | struct cse_reg_info *next; |
30f72379 | 311 | |
9b1549b8 | 312 | /* Search key */ |
770ae6cc | 313 | unsigned int regno; |
9b1549b8 DM |
314 | |
315 | /* The quantity number of the register's current contents. */ | |
316 | int reg_qty; | |
317 | ||
318 | /* The number of times the register has been altered in the current | |
319 | basic block. */ | |
320 | int reg_tick; | |
321 | ||
30f72379 MM |
322 | /* The REG_TICK value at which rtx's containing this register are |
323 | valid in the hash table. If this does not equal the current | |
324 | reg_tick value, such expressions existing in the hash table are | |
325 | invalid. */ | |
326 | int reg_in_table; | |
46081bb3 SH |
327 | |
328 | /* The SUBREG that was set when REG_TICK was last incremented. Set | |
329 | to -1 if the last store was to the whole register, not a subreg. */ | |
5dd78e9a | 330 | unsigned int subreg_ticked; |
30f72379 | 331 | }; |
7afe21cc | 332 | |
30f72379 MM |
333 | /* A free list of cse_reg_info entries. */ |
334 | static struct cse_reg_info *cse_reg_info_free_list; | |
7afe21cc | 335 | |
c1edba58 VM |
336 | /* A used list of cse_reg_info entries. */ |
337 | static struct cse_reg_info *cse_reg_info_used_list; | |
338 | static struct cse_reg_info *cse_reg_info_used_list_end; | |
339 | ||
30f72379 | 340 | /* A mapping from registers to cse_reg_info data structures. */ |
9b1549b8 DM |
341 | #define REGHASH_SHIFT 7 |
342 | #define REGHASH_SIZE (1 << REGHASH_SHIFT) | |
343 | #define REGHASH_MASK (REGHASH_SIZE - 1) | |
344 | static struct cse_reg_info *reg_hash[REGHASH_SIZE]; | |
345 | ||
346 | #define REGHASH_FN(REGNO) \ | |
347 | (((REGNO) ^ ((REGNO) >> REGHASH_SHIFT)) & REGHASH_MASK) | |
7afe21cc | 348 | |
30f72379 MM |
349 | /* The last lookup we did into the cse_reg_info_tree. This allows us |
350 | to cache repeated lookups. */ | |
770ae6cc | 351 | static unsigned int cached_regno; |
30f72379 | 352 | static struct cse_reg_info *cached_cse_reg_info; |
7afe21cc | 353 | |
278a83b2 | 354 | /* A HARD_REG_SET containing all the hard registers for which there is |
7afe21cc RK |
355 | currently a REG expression in the hash table. Note the difference |
356 | from the above variables, which indicate if the REG is mentioned in some | |
357 | expression in the table. */ | |
358 | ||
359 | static HARD_REG_SET hard_regs_in_table; | |
360 | ||
7afe21cc RK |
361 | /* CUID of insn that starts the basic block currently being cse-processed. */ |
362 | ||
363 | static int cse_basic_block_start; | |
364 | ||
365 | /* CUID of insn that ends the basic block currently being cse-processed. */ | |
366 | ||
367 | static int cse_basic_block_end; | |
368 | ||
369 | /* Vector mapping INSN_UIDs to cuids. | |
d45cf215 | 370 | The cuids are like uids but increase monotonically always. |
7afe21cc RK |
371 | We use them to see whether a reg is used outside a given basic block. */ |
372 | ||
906c4e36 | 373 | static int *uid_cuid; |
7afe21cc | 374 | |
164c8956 RK |
375 | /* Highest UID in UID_CUID. */ |
376 | static int max_uid; | |
377 | ||
7afe21cc RK |
378 | /* Get the cuid of an insn. */ |
379 | ||
380 | #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)]) | |
381 | ||
4eadede7 ZW |
382 | /* Nonzero if this pass has made changes, and therefore it's |
383 | worthwhile to run the garbage collector. */ | |
384 | ||
385 | static int cse_altered; | |
386 | ||
7afe21cc RK |
387 | /* Nonzero if cse has altered conditional jump insns |
388 | in such a way that jump optimization should be redone. */ | |
389 | ||
390 | static int cse_jumps_altered; | |
391 | ||
f85cc4cb RK |
392 | /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a |
393 | REG_LABEL, we have to rerun jump after CSE to put in the note. */ | |
a5dfb4ee RK |
394 | static int recorded_label_ref; |
395 | ||
7afe21cc RK |
396 | /* canon_hash stores 1 in do_not_record |
397 | if it notices a reference to CC0, PC, or some other volatile | |
398 | subexpression. */ | |
399 | ||
400 | static int do_not_record; | |
401 | ||
7bac1be0 RK |
402 | #ifdef LOAD_EXTEND_OP |
403 | ||
404 | /* Scratch rtl used when looking for load-extended copy of a MEM. */ | |
405 | static rtx memory_extend_rtx; | |
406 | #endif | |
407 | ||
7afe21cc RK |
408 | /* canon_hash stores 1 in hash_arg_in_memory |
409 | if it notices a reference to memory within the expression being hashed. */ | |
410 | ||
411 | static int hash_arg_in_memory; | |
412 | ||
7afe21cc RK |
413 | /* The hash table contains buckets which are chains of `struct table_elt's, |
414 | each recording one expression's information. | |
415 | That expression is in the `exp' field. | |
416 | ||
db048faf MM |
417 | The canon_exp field contains a canonical (from the point of view of |
418 | alias analysis) version of the `exp' field. | |
419 | ||
7afe21cc RK |
420 | Those elements with the same hash code are chained in both directions |
421 | through the `next_same_hash' and `prev_same_hash' fields. | |
422 | ||
423 | Each set of expressions with equivalent values | |
424 | are on a two-way chain through the `next_same_value' | |
425 | and `prev_same_value' fields, and all point with | |
426 | the `first_same_value' field at the first element in | |
427 | that chain. The chain is in order of increasing cost. | |
428 | Each element's cost value is in its `cost' field. | |
429 | ||
430 | The `in_memory' field is nonzero for elements that | |
431 | involve any reference to memory. These elements are removed | |
432 | whenever a write is done to an unidentified location in memory. | |
433 | To be safe, we assume that a memory address is unidentified unless | |
434 | the address is either a symbol constant or a constant plus | |
435 | the frame pointer or argument pointer. | |
436 | ||
7afe21cc RK |
437 | The `related_value' field is used to connect related expressions |
438 | (that differ by adding an integer). | |
439 | The related expressions are chained in a circular fashion. | |
440 | `related_value' is zero for expressions for which this | |
441 | chain is not useful. | |
442 | ||
443 | The `cost' field stores the cost of this element's expression. | |
630c79be BS |
444 | The `regcost' field stores the value returned by approx_reg_cost for |
445 | this element's expression. | |
7afe21cc RK |
446 | |
447 | The `is_const' flag is set if the element is a constant (including | |
448 | a fixed address). | |
449 | ||
450 | The `flag' field is used as a temporary during some search routines. | |
451 | ||
452 | The `mode' field is usually the same as GET_MODE (`exp'), but | |
453 | if `exp' is a CONST_INT and has no machine mode then the `mode' | |
454 | field is the mode it was being used as. Each constant is | |
455 | recorded separately for each mode it is used with. */ | |
456 | ||
7afe21cc RK |
457 | struct table_elt |
458 | { | |
459 | rtx exp; | |
db048faf | 460 | rtx canon_exp; |
7afe21cc RK |
461 | struct table_elt *next_same_hash; |
462 | struct table_elt *prev_same_hash; | |
463 | struct table_elt *next_same_value; | |
464 | struct table_elt *prev_same_value; | |
465 | struct table_elt *first_same_value; | |
466 | struct table_elt *related_value; | |
467 | int cost; | |
630c79be | 468 | int regcost; |
496324d0 DN |
469 | /* The size of this field should match the size |
470 | of the mode field of struct rtx_def (see rtl.h). */ | |
471 | ENUM_BITFIELD(machine_mode) mode : 8; | |
7afe21cc | 472 | char in_memory; |
7afe21cc RK |
473 | char is_const; |
474 | char flag; | |
475 | }; | |
476 | ||
7afe21cc RK |
477 | /* We don't want a lot of buckets, because we rarely have very many |
478 | things stored in the hash table, and a lot of buckets slows | |
479 | down a lot of loops that happen frequently. */ | |
9b1549b8 DM |
480 | #define HASH_SHIFT 5 |
481 | #define HASH_SIZE (1 << HASH_SHIFT) | |
482 | #define HASH_MASK (HASH_SIZE - 1) | |
7afe21cc RK |
483 | |
484 | /* Compute hash code of X in mode M. Special-case case where X is a pseudo | |
485 | register (hard registers may require `do_not_record' to be set). */ | |
486 | ||
487 | #define HASH(X, M) \ | |
f8cfc6aa | 488 | ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \ |
9b1549b8 DM |
489 | ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \ |
490 | : canon_hash (X, M)) & HASH_MASK) | |
7afe21cc | 491 | |
630c79be BS |
492 | /* Determine whether register number N is considered a fixed register for the |
493 | purpose of approximating register costs. | |
7afe21cc RK |
494 | It is desirable to replace other regs with fixed regs, to reduce need for |
495 | non-fixed hard regs. | |
553687c9 | 496 | A reg wins if it is either the frame pointer or designated as fixed. */ |
7afe21cc | 497 | #define FIXED_REGNO_P(N) \ |
8bc169f2 | 498 | ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \ |
6ab832bc | 499 | || fixed_regs[N] || global_regs[N]) |
7afe21cc RK |
500 | |
501 | /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed | |
ac07e066 RK |
502 | hard registers and pointers into the frame are the cheapest with a cost |
503 | of 0. Next come pseudos with a cost of one and other hard registers with | |
504 | a cost of 2. Aside from these special cases, call `rtx_cost'. */ | |
505 | ||
6ab832bc | 506 | #define CHEAP_REGNO(N) \ |
7080f735 AJ |
507 | ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \ |
508 | || (N) == STACK_POINTER_REGNUM || (N) == ARG_POINTER_REGNUM \ | |
509 | || ((N) >= FIRST_VIRTUAL_REGISTER && (N) <= LAST_VIRTUAL_REGISTER) \ | |
8bc169f2 | 510 | || ((N) < FIRST_PSEUDO_REGISTER \ |
e7bb59fa | 511 | && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS)) |
7afe21cc | 512 | |
f8cfc6aa JQ |
513 | #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET)) |
514 | #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER)) | |
7afe21cc | 515 | |
30f72379 MM |
516 | /* Get the info associated with register N. */ |
517 | ||
7080f735 | 518 | #define GET_CSE_REG_INFO(N) \ |
30f72379 MM |
519 | (((N) == cached_regno && cached_cse_reg_info) \ |
520 | ? cached_cse_reg_info : get_cse_reg_info ((N))) | |
521 | ||
522 | /* Get the number of times this register has been updated in this | |
523 | basic block. */ | |
524 | ||
c1edba58 | 525 | #define REG_TICK(N) ((GET_CSE_REG_INFO (N))->reg_tick) |
30f72379 MM |
526 | |
527 | /* Get the point at which REG was recorded in the table. */ | |
528 | ||
529 | #define REG_IN_TABLE(N) ((GET_CSE_REG_INFO (N))->reg_in_table) | |
530 | ||
46081bb3 SH |
531 | /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a |
532 | SUBREG). */ | |
533 | ||
534 | #define SUBREG_TICKED(N) ((GET_CSE_REG_INFO (N))->subreg_ticked) | |
535 | ||
30f72379 MM |
536 | /* Get the quantity number for REG. */ |
537 | ||
538 | #define REG_QTY(N) ((GET_CSE_REG_INFO (N))->reg_qty) | |
539 | ||
7afe21cc | 540 | /* Determine if the quantity number for register X represents a valid index |
1bb98cec | 541 | into the qty_table. */ |
7afe21cc | 542 | |
770ae6cc | 543 | #define REGNO_QTY_VALID_P(N) (REG_QTY (N) != (int) (N)) |
7afe21cc | 544 | |
9b1549b8 | 545 | static struct table_elt *table[HASH_SIZE]; |
7afe21cc RK |
546 | |
547 | /* Chain of `struct table_elt's made so far for this function | |
548 | but currently removed from the table. */ | |
549 | ||
550 | static struct table_elt *free_element_chain; | |
551 | ||
552 | /* Number of `struct table_elt' structures made so far for this function. */ | |
553 | ||
554 | static int n_elements_made; | |
555 | ||
556 | /* Maximum value `n_elements_made' has had so far in this compilation | |
557 | for functions previously processed. */ | |
558 | ||
559 | static int max_elements_made; | |
560 | ||
278a83b2 | 561 | /* Surviving equivalence class when two equivalence classes are merged |
7afe21cc RK |
562 | by recording the effects of a jump in the last insn. Zero if the |
563 | last insn was not a conditional jump. */ | |
564 | ||
565 | static struct table_elt *last_jump_equiv_class; | |
566 | ||
567 | /* Set to the cost of a constant pool reference if one was found for a | |
568 | symbolic constant. If this was found, it means we should try to | |
569 | convert constants into constant pool entries if they don't fit in | |
570 | the insn. */ | |
571 | ||
572 | static int constant_pool_entries_cost; | |
dd0ba281 | 573 | static int constant_pool_entries_regcost; |
7afe21cc | 574 | |
6cd4575e RK |
575 | /* This data describes a block that will be processed by cse_basic_block. */ |
576 | ||
14a774a9 RK |
577 | struct cse_basic_block_data |
578 | { | |
6cd4575e RK |
579 | /* Lowest CUID value of insns in block. */ |
580 | int low_cuid; | |
581 | /* Highest CUID value of insns in block. */ | |
582 | int high_cuid; | |
583 | /* Total number of SETs in block. */ | |
584 | int nsets; | |
585 | /* Last insn in the block. */ | |
586 | rtx last; | |
587 | /* Size of current branch path, if any. */ | |
588 | int path_size; | |
589 | /* Current branch path, indicating which branches will be taken. */ | |
14a774a9 RK |
590 | struct branch_path |
591 | { | |
592 | /* The branch insn. */ | |
593 | rtx branch; | |
594 | /* Whether it should be taken or not. AROUND is the same as taken | |
595 | except that it is used when the destination label is not preceded | |
6cd4575e | 596 | by a BARRIER. */ |
6de9cd9a | 597 | enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status; |
9bf8cfbf | 598 | } *path; |
6cd4575e RK |
599 | }; |
600 | ||
7080f735 AJ |
601 | static bool fixed_base_plus_p (rtx x); |
602 | static int notreg_cost (rtx, enum rtx_code); | |
603 | static int approx_reg_cost_1 (rtx *, void *); | |
604 | static int approx_reg_cost (rtx); | |
56ae04af | 605 | static int preferable (int, int, int, int); |
7080f735 AJ |
606 | static void new_basic_block (void); |
607 | static void make_new_qty (unsigned int, enum machine_mode); | |
608 | static void make_regs_eqv (unsigned int, unsigned int); | |
609 | static void delete_reg_equiv (unsigned int); | |
610 | static int mention_regs (rtx); | |
611 | static int insert_regs (rtx, struct table_elt *, int); | |
612 | static void remove_from_table (struct table_elt *, unsigned); | |
613 | static struct table_elt *lookup (rtx, unsigned, enum machine_mode); | |
614 | static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode); | |
615 | static rtx lookup_as_function (rtx, enum rtx_code); | |
616 | static struct table_elt *insert (rtx, struct table_elt *, unsigned, | |
617 | enum machine_mode); | |
618 | static void merge_equiv_classes (struct table_elt *, struct table_elt *); | |
619 | static void invalidate (rtx, enum machine_mode); | |
620 | static int cse_rtx_varies_p (rtx, int); | |
621 | static void remove_invalid_refs (unsigned int); | |
622 | static void remove_invalid_subreg_refs (unsigned int, unsigned int, | |
623 | enum machine_mode); | |
624 | static void rehash_using_reg (rtx); | |
625 | static void invalidate_memory (void); | |
626 | static void invalidate_for_call (void); | |
627 | static rtx use_related_value (rtx, struct table_elt *); | |
628 | static unsigned canon_hash (rtx, enum machine_mode); | |
629 | static unsigned canon_hash_string (const char *); | |
630 | static unsigned safe_hash (rtx, enum machine_mode); | |
631 | static int exp_equiv_p (rtx, rtx, int, int); | |
632 | static rtx canon_reg (rtx, rtx); | |
633 | static void find_best_addr (rtx, rtx *, enum machine_mode); | |
634 | static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *, | |
635 | enum machine_mode *, | |
636 | enum machine_mode *); | |
637 | static rtx fold_rtx (rtx, rtx); | |
638 | static rtx equiv_constant (rtx); | |
639 | static void record_jump_equiv (rtx, int); | |
640 | static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx, | |
641 | int); | |
642 | static void cse_insn (rtx, rtx); | |
86caf04d PB |
643 | static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *, |
644 | int, int, int); | |
7080f735 AJ |
645 | static int addr_affects_sp_p (rtx); |
646 | static void invalidate_from_clobbers (rtx); | |
647 | static rtx cse_process_notes (rtx, rtx); | |
648 | static void cse_around_loop (rtx); | |
649 | static void invalidate_skipped_set (rtx, rtx, void *); | |
650 | static void invalidate_skipped_block (rtx); | |
651 | static void cse_check_loop_start (rtx, rtx, void *); | |
652 | static void cse_set_around_loop (rtx, rtx, rtx); | |
653 | static rtx cse_basic_block (rtx, rtx, struct branch_path *, int); | |
9ab81df2 | 654 | static void count_reg_usage (rtx, int *, int); |
7080f735 AJ |
655 | static int check_for_label_ref (rtx *, void *); |
656 | extern void dump_class (struct table_elt*); | |
657 | static struct cse_reg_info * get_cse_reg_info (unsigned int); | |
658 | static int check_dependence (rtx *, void *); | |
659 | ||
660 | static void flush_hash_table (void); | |
661 | static bool insn_live_p (rtx, int *); | |
662 | static bool set_live_p (rtx, rtx, int *); | |
663 | static bool dead_libcall_p (rtx, int *); | |
e129d93a ILT |
664 | static int cse_change_cc_mode (rtx *, void *); |
665 | static void cse_change_cc_mode_insns (rtx, rtx, rtx); | |
666 | static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool); | |
7afe21cc | 667 | \f |
2f93eea8 PB |
668 | |
669 | #undef RTL_HOOKS_GEN_LOWPART | |
670 | #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible | |
671 | ||
672 | static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER; | |
673 | \f | |
4977bab6 ZW |
674 | /* Nonzero if X has the form (PLUS frame-pointer integer). We check for |
675 | virtual regs here because the simplify_*_operation routines are called | |
676 | by integrate.c, which is called before virtual register instantiation. */ | |
677 | ||
678 | static bool | |
7080f735 | 679 | fixed_base_plus_p (rtx x) |
4977bab6 ZW |
680 | { |
681 | switch (GET_CODE (x)) | |
682 | { | |
683 | case REG: | |
684 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx) | |
685 | return true; | |
686 | if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
687 | return true; | |
688 | if (REGNO (x) >= FIRST_VIRTUAL_REGISTER | |
689 | && REGNO (x) <= LAST_VIRTUAL_REGISTER) | |
690 | return true; | |
691 | return false; | |
692 | ||
693 | case PLUS: | |
694 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
695 | return false; | |
696 | return fixed_base_plus_p (XEXP (x, 0)); | |
697 | ||
4977bab6 ZW |
698 | default: |
699 | return false; | |
700 | } | |
701 | } | |
702 | ||
a4c6502a MM |
703 | /* Dump the expressions in the equivalence class indicated by CLASSP. |
704 | This function is used only for debugging. */ | |
a0153051 | 705 | void |
7080f735 | 706 | dump_class (struct table_elt *classp) |
a4c6502a MM |
707 | { |
708 | struct table_elt *elt; | |
709 | ||
710 | fprintf (stderr, "Equivalence chain for "); | |
711 | print_rtl (stderr, classp->exp); | |
712 | fprintf (stderr, ": \n"); | |
278a83b2 | 713 | |
a4c6502a MM |
714 | for (elt = classp->first_same_value; elt; elt = elt->next_same_value) |
715 | { | |
716 | print_rtl (stderr, elt->exp); | |
717 | fprintf (stderr, "\n"); | |
718 | } | |
719 | } | |
720 | ||
630c79be | 721 | /* Subroutine of approx_reg_cost; called through for_each_rtx. */ |
be8ac49a | 722 | |
630c79be | 723 | static int |
7080f735 | 724 | approx_reg_cost_1 (rtx *xp, void *data) |
630c79be BS |
725 | { |
726 | rtx x = *xp; | |
c863f8c2 | 727 | int *cost_p = data; |
630c79be | 728 | |
f8cfc6aa | 729 | if (x && REG_P (x)) |
c863f8c2 DM |
730 | { |
731 | unsigned int regno = REGNO (x); | |
732 | ||
733 | if (! CHEAP_REGNO (regno)) | |
734 | { | |
735 | if (regno < FIRST_PSEUDO_REGISTER) | |
736 | { | |
737 | if (SMALL_REGISTER_CLASSES) | |
738 | return 1; | |
739 | *cost_p += 2; | |
740 | } | |
741 | else | |
742 | *cost_p += 1; | |
743 | } | |
744 | } | |
745 | ||
630c79be BS |
746 | return 0; |
747 | } | |
748 | ||
749 | /* Return an estimate of the cost of the registers used in an rtx. | |
750 | This is mostly the number of different REG expressions in the rtx; | |
a1f300c0 | 751 | however for some exceptions like fixed registers we use a cost of |
f1c1dfc3 | 752 | 0. If any other hard register reference occurs, return MAX_COST. */ |
630c79be BS |
753 | |
754 | static int | |
7080f735 | 755 | approx_reg_cost (rtx x) |
630c79be | 756 | { |
630c79be | 757 | int cost = 0; |
f1c1dfc3 | 758 | |
c863f8c2 DM |
759 | if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost)) |
760 | return MAX_COST; | |
630c79be | 761 | |
c863f8c2 | 762 | return cost; |
630c79be BS |
763 | } |
764 | ||
765 | /* Return a negative value if an rtx A, whose costs are given by COST_A | |
766 | and REGCOST_A, is more desirable than an rtx B. | |
767 | Return a positive value if A is less desirable, or 0 if the two are | |
768 | equally good. */ | |
769 | static int | |
56ae04af | 770 | preferable (int cost_a, int regcost_a, int cost_b, int regcost_b) |
630c79be | 771 | { |
423adbb9 | 772 | /* First, get rid of cases involving expressions that are entirely |
f1c1dfc3 BS |
773 | unwanted. */ |
774 | if (cost_a != cost_b) | |
775 | { | |
776 | if (cost_a == MAX_COST) | |
777 | return 1; | |
778 | if (cost_b == MAX_COST) | |
779 | return -1; | |
780 | } | |
781 | ||
782 | /* Avoid extending lifetimes of hardregs. */ | |
783 | if (regcost_a != regcost_b) | |
784 | { | |
785 | if (regcost_a == MAX_COST) | |
786 | return 1; | |
787 | if (regcost_b == MAX_COST) | |
788 | return -1; | |
789 | } | |
790 | ||
791 | /* Normal operation costs take precedence. */ | |
630c79be BS |
792 | if (cost_a != cost_b) |
793 | return cost_a - cost_b; | |
f1c1dfc3 | 794 | /* Only if these are identical consider effects on register pressure. */ |
630c79be BS |
795 | if (regcost_a != regcost_b) |
796 | return regcost_a - regcost_b; | |
797 | return 0; | |
798 | } | |
799 | ||
954a5693 RK |
800 | /* Internal function, to compute cost when X is not a register; called |
801 | from COST macro to keep it simple. */ | |
802 | ||
803 | static int | |
7080f735 | 804 | notreg_cost (rtx x, enum rtx_code outer) |
954a5693 RK |
805 | { |
806 | return ((GET_CODE (x) == SUBREG | |
f8cfc6aa | 807 | && REG_P (SUBREG_REG (x)) |
954a5693 RK |
808 | && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT |
809 | && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT | |
810 | && (GET_MODE_SIZE (GET_MODE (x)) | |
811 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
812 | && subreg_lowpart_p (x) | |
813 | && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)), | |
814 | GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))))) | |
630c79be | 815 | ? 0 |
f2fa288f | 816 | : rtx_cost (x, outer) * 2); |
954a5693 RK |
817 | } |
818 | ||
01329426 | 819 | \f |
30f72379 | 820 | static struct cse_reg_info * |
7080f735 | 821 | get_cse_reg_info (unsigned int regno) |
30f72379 | 822 | { |
9b1549b8 DM |
823 | struct cse_reg_info **hash_head = ®_hash[REGHASH_FN (regno)]; |
824 | struct cse_reg_info *p; | |
825 | ||
278a83b2 | 826 | for (p = *hash_head; p != NULL; p = p->hash_next) |
9b1549b8 DM |
827 | if (p->regno == regno) |
828 | break; | |
829 | ||
830 | if (p == NULL) | |
30f72379 MM |
831 | { |
832 | /* Get a new cse_reg_info structure. */ | |
9b1549b8 | 833 | if (cse_reg_info_free_list) |
30f72379 | 834 | { |
9b1549b8 DM |
835 | p = cse_reg_info_free_list; |
836 | cse_reg_info_free_list = p->next; | |
30f72379 MM |
837 | } |
838 | else | |
703ad42b | 839 | p = xmalloc (sizeof (struct cse_reg_info)); |
9b1549b8 DM |
840 | |
841 | /* Insert into hash table. */ | |
842 | p->hash_next = *hash_head; | |
843 | *hash_head = p; | |
30f72379 MM |
844 | |
845 | /* Initialize it. */ | |
9b1549b8 DM |
846 | p->reg_tick = 1; |
847 | p->reg_in_table = -1; | |
46081bb3 | 848 | p->subreg_ticked = -1; |
9b1549b8 DM |
849 | p->reg_qty = regno; |
850 | p->regno = regno; | |
851 | p->next = cse_reg_info_used_list; | |
852 | cse_reg_info_used_list = p; | |
c1edba58 | 853 | if (!cse_reg_info_used_list_end) |
9b1549b8 | 854 | cse_reg_info_used_list_end = p; |
30f72379 MM |
855 | } |
856 | ||
857 | /* Cache this lookup; we tend to be looking up information about the | |
858 | same register several times in a row. */ | |
859 | cached_regno = regno; | |
9b1549b8 | 860 | cached_cse_reg_info = p; |
30f72379 | 861 | |
9b1549b8 | 862 | return p; |
30f72379 MM |
863 | } |
864 | ||
7afe21cc RK |
865 | /* Clear the hash table and initialize each register with its own quantity, |
866 | for a new basic block. */ | |
867 | ||
868 | static void | |
7080f735 | 869 | new_basic_block (void) |
7afe21cc | 870 | { |
b3694847 | 871 | int i; |
7afe21cc RK |
872 | |
873 | next_qty = max_reg; | |
874 | ||
9b1549b8 DM |
875 | /* Clear out hash table state for this pass. */ |
876 | ||
703ad42b | 877 | memset (reg_hash, 0, sizeof reg_hash); |
9b1549b8 DM |
878 | |
879 | if (cse_reg_info_used_list) | |
30f72379 | 880 | { |
9b1549b8 DM |
881 | cse_reg_info_used_list_end->next = cse_reg_info_free_list; |
882 | cse_reg_info_free_list = cse_reg_info_used_list; | |
883 | cse_reg_info_used_list = cse_reg_info_used_list_end = 0; | |
30f72379 | 884 | } |
9b1549b8 | 885 | cached_cse_reg_info = 0; |
7afe21cc | 886 | |
7afe21cc RK |
887 | CLEAR_HARD_REG_SET (hard_regs_in_table); |
888 | ||
889 | /* The per-quantity values used to be initialized here, but it is | |
890 | much faster to initialize each as it is made in `make_new_qty'. */ | |
891 | ||
9b1549b8 | 892 | for (i = 0; i < HASH_SIZE; i++) |
7afe21cc | 893 | { |
9b1549b8 DM |
894 | struct table_elt *first; |
895 | ||
896 | first = table[i]; | |
897 | if (first != NULL) | |
7afe21cc | 898 | { |
9b1549b8 DM |
899 | struct table_elt *last = first; |
900 | ||
901 | table[i] = NULL; | |
902 | ||
903 | while (last->next_same_hash != NULL) | |
904 | last = last->next_same_hash; | |
905 | ||
906 | /* Now relink this hash entire chain into | |
907 | the free element list. */ | |
908 | ||
909 | last->next_same_hash = free_element_chain; | |
910 | free_element_chain = first; | |
7afe21cc RK |
911 | } |
912 | } | |
913 | ||
7afe21cc | 914 | #ifdef HAVE_cc0 |
4977bab6 | 915 | prev_insn = 0; |
7afe21cc RK |
916 | prev_insn_cc0 = 0; |
917 | #endif | |
918 | } | |
919 | ||
1bb98cec DM |
920 | /* Say that register REG contains a quantity in mode MODE not in any |
921 | register before and initialize that quantity. */ | |
7afe21cc RK |
922 | |
923 | static void | |
7080f735 | 924 | make_new_qty (unsigned int reg, enum machine_mode mode) |
7afe21cc | 925 | { |
b3694847 SS |
926 | int q; |
927 | struct qty_table_elem *ent; | |
928 | struct reg_eqv_elem *eqv; | |
7afe21cc RK |
929 | |
930 | if (next_qty >= max_qty) | |
931 | abort (); | |
932 | ||
30f72379 | 933 | q = REG_QTY (reg) = next_qty++; |
1bb98cec DM |
934 | ent = &qty_table[q]; |
935 | ent->first_reg = reg; | |
936 | ent->last_reg = reg; | |
937 | ent->mode = mode; | |
938 | ent->const_rtx = ent->const_insn = NULL_RTX; | |
939 | ent->comparison_code = UNKNOWN; | |
940 | ||
941 | eqv = ®_eqv_table[reg]; | |
942 | eqv->next = eqv->prev = -1; | |
7afe21cc RK |
943 | } |
944 | ||
945 | /* Make reg NEW equivalent to reg OLD. | |
946 | OLD is not changing; NEW is. */ | |
947 | ||
948 | static void | |
7080f735 | 949 | make_regs_eqv (unsigned int new, unsigned int old) |
7afe21cc | 950 | { |
770ae6cc RK |
951 | unsigned int lastr, firstr; |
952 | int q = REG_QTY (old); | |
953 | struct qty_table_elem *ent; | |
1bb98cec DM |
954 | |
955 | ent = &qty_table[q]; | |
7afe21cc RK |
956 | |
957 | /* Nothing should become eqv until it has a "non-invalid" qty number. */ | |
958 | if (! REGNO_QTY_VALID_P (old)) | |
959 | abort (); | |
960 | ||
30f72379 | 961 | REG_QTY (new) = q; |
1bb98cec DM |
962 | firstr = ent->first_reg; |
963 | lastr = ent->last_reg; | |
7afe21cc RK |
964 | |
965 | /* Prefer fixed hard registers to anything. Prefer pseudo regs to other | |
966 | hard regs. Among pseudos, if NEW will live longer than any other reg | |
967 | of the same qty, and that is beyond the current basic block, | |
968 | make it the new canonical replacement for this qty. */ | |
969 | if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr)) | |
970 | /* Certain fixed registers might be of the class NO_REGS. This means | |
971 | that not only can they not be allocated by the compiler, but | |
830a38ee | 972 | they cannot be used in substitutions or canonicalizations |
7afe21cc RK |
973 | either. */ |
974 | && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS) | |
975 | && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new)) | |
976 | || (new >= FIRST_PSEUDO_REGISTER | |
977 | && (firstr < FIRST_PSEUDO_REGISTER | |
b1f21e0a MM |
978 | || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end |
979 | || (uid_cuid[REGNO_FIRST_UID (new)] | |
7afe21cc | 980 | < cse_basic_block_start)) |
b1f21e0a MM |
981 | && (uid_cuid[REGNO_LAST_UID (new)] |
982 | > uid_cuid[REGNO_LAST_UID (firstr)])))))) | |
7afe21cc | 983 | { |
1bb98cec DM |
984 | reg_eqv_table[firstr].prev = new; |
985 | reg_eqv_table[new].next = firstr; | |
986 | reg_eqv_table[new].prev = -1; | |
987 | ent->first_reg = new; | |
7afe21cc RK |
988 | } |
989 | else | |
990 | { | |
991 | /* If NEW is a hard reg (known to be non-fixed), insert at end. | |
992 | Otherwise, insert before any non-fixed hard regs that are at the | |
993 | end. Registers of class NO_REGS cannot be used as an | |
994 | equivalent for anything. */ | |
1bb98cec | 995 | while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0 |
7afe21cc RK |
996 | && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr)) |
997 | && new >= FIRST_PSEUDO_REGISTER) | |
1bb98cec DM |
998 | lastr = reg_eqv_table[lastr].prev; |
999 | reg_eqv_table[new].next = reg_eqv_table[lastr].next; | |
1000 | if (reg_eqv_table[lastr].next >= 0) | |
1001 | reg_eqv_table[reg_eqv_table[lastr].next].prev = new; | |
7afe21cc | 1002 | else |
1bb98cec DM |
1003 | qty_table[q].last_reg = new; |
1004 | reg_eqv_table[lastr].next = new; | |
1005 | reg_eqv_table[new].prev = lastr; | |
7afe21cc RK |
1006 | } |
1007 | } | |
1008 | ||
1009 | /* Remove REG from its equivalence class. */ | |
1010 | ||
1011 | static void | |
7080f735 | 1012 | delete_reg_equiv (unsigned int reg) |
7afe21cc | 1013 | { |
b3694847 SS |
1014 | struct qty_table_elem *ent; |
1015 | int q = REG_QTY (reg); | |
1016 | int p, n; | |
7afe21cc | 1017 | |
a4e262bc | 1018 | /* If invalid, do nothing. */ |
770ae6cc | 1019 | if (q == (int) reg) |
7afe21cc RK |
1020 | return; |
1021 | ||
1bb98cec DM |
1022 | ent = &qty_table[q]; |
1023 | ||
1024 | p = reg_eqv_table[reg].prev; | |
1025 | n = reg_eqv_table[reg].next; | |
a4e262bc | 1026 | |
7afe21cc | 1027 | if (n != -1) |
1bb98cec | 1028 | reg_eqv_table[n].prev = p; |
7afe21cc | 1029 | else |
1bb98cec | 1030 | ent->last_reg = p; |
7afe21cc | 1031 | if (p != -1) |
1bb98cec | 1032 | reg_eqv_table[p].next = n; |
7afe21cc | 1033 | else |
1bb98cec | 1034 | ent->first_reg = n; |
7afe21cc | 1035 | |
30f72379 | 1036 | REG_QTY (reg) = reg; |
7afe21cc RK |
1037 | } |
1038 | ||
1039 | /* Remove any invalid expressions from the hash table | |
1040 | that refer to any of the registers contained in expression X. | |
1041 | ||
1042 | Make sure that newly inserted references to those registers | |
1043 | as subexpressions will be considered valid. | |
1044 | ||
1045 | mention_regs is not called when a register itself | |
1046 | is being stored in the table. | |
1047 | ||
1048 | Return 1 if we have done something that may have changed the hash code | |
1049 | of X. */ | |
1050 | ||
1051 | static int | |
7080f735 | 1052 | mention_regs (rtx x) |
7afe21cc | 1053 | { |
b3694847 SS |
1054 | enum rtx_code code; |
1055 | int i, j; | |
1056 | const char *fmt; | |
1057 | int changed = 0; | |
7afe21cc RK |
1058 | |
1059 | if (x == 0) | |
e5f6a288 | 1060 | return 0; |
7afe21cc RK |
1061 | |
1062 | code = GET_CODE (x); | |
1063 | if (code == REG) | |
1064 | { | |
770ae6cc RK |
1065 | unsigned int regno = REGNO (x); |
1066 | unsigned int endregno | |
7afe21cc | 1067 | = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1 |
66fd46b6 | 1068 | : hard_regno_nregs[regno][GET_MODE (x)]); |
770ae6cc | 1069 | unsigned int i; |
7afe21cc RK |
1070 | |
1071 | for (i = regno; i < endregno; i++) | |
1072 | { | |
30f72379 | 1073 | if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i)) |
7afe21cc RK |
1074 | remove_invalid_refs (i); |
1075 | ||
30f72379 | 1076 | REG_IN_TABLE (i) = REG_TICK (i); |
46081bb3 | 1077 | SUBREG_TICKED (i) = -1; |
7afe21cc RK |
1078 | } |
1079 | ||
1080 | return 0; | |
1081 | } | |
1082 | ||
34c73909 R |
1083 | /* If this is a SUBREG, we don't want to discard other SUBREGs of the same |
1084 | pseudo if they don't use overlapping words. We handle only pseudos | |
1085 | here for simplicity. */ | |
f8cfc6aa | 1086 | if (code == SUBREG && REG_P (SUBREG_REG (x)) |
34c73909 R |
1087 | && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER) |
1088 | { | |
770ae6cc | 1089 | unsigned int i = REGNO (SUBREG_REG (x)); |
34c73909 | 1090 | |
30f72379 | 1091 | if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i)) |
34c73909 | 1092 | { |
46081bb3 SH |
1093 | /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and |
1094 | the last store to this register really stored into this | |
1095 | subreg, then remove the memory of this subreg. | |
1096 | Otherwise, remove any memory of the entire register and | |
1097 | all its subregs from the table. */ | |
1098 | if (REG_TICK (i) - REG_IN_TABLE (i) > 1 | |
5dd78e9a | 1099 | || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x))) |
34c73909 R |
1100 | remove_invalid_refs (i); |
1101 | else | |
ddef6bc7 | 1102 | remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x)); |
34c73909 R |
1103 | } |
1104 | ||
30f72379 | 1105 | REG_IN_TABLE (i) = REG_TICK (i); |
5dd78e9a | 1106 | SUBREG_TICKED (i) = REGNO (SUBREG_REG (x)); |
34c73909 R |
1107 | return 0; |
1108 | } | |
1109 | ||
7afe21cc RK |
1110 | /* If X is a comparison or a COMPARE and either operand is a register |
1111 | that does not have a quantity, give it one. This is so that a later | |
1112 | call to record_jump_equiv won't cause X to be assigned a different | |
1113 | hash code and not found in the table after that call. | |
1114 | ||
1115 | It is not necessary to do this here, since rehash_using_reg can | |
1116 | fix up the table later, but doing this here eliminates the need to | |
1117 | call that expensive function in the most common case where the only | |
1118 | use of the register is in the comparison. */ | |
1119 | ||
ec8e098d | 1120 | if (code == COMPARE || COMPARISON_P (x)) |
7afe21cc | 1121 | { |
f8cfc6aa | 1122 | if (REG_P (XEXP (x, 0)) |
7afe21cc | 1123 | && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))) |
9714cf43 | 1124 | if (insert_regs (XEXP (x, 0), NULL, 0)) |
7afe21cc RK |
1125 | { |
1126 | rehash_using_reg (XEXP (x, 0)); | |
1127 | changed = 1; | |
1128 | } | |
1129 | ||
f8cfc6aa | 1130 | if (REG_P (XEXP (x, 1)) |
7afe21cc | 1131 | && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1)))) |
9714cf43 | 1132 | if (insert_regs (XEXP (x, 1), NULL, 0)) |
7afe21cc RK |
1133 | { |
1134 | rehash_using_reg (XEXP (x, 1)); | |
1135 | changed = 1; | |
1136 | } | |
1137 | } | |
1138 | ||
1139 | fmt = GET_RTX_FORMAT (code); | |
1140 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1141 | if (fmt[i] == 'e') | |
1142 | changed |= mention_regs (XEXP (x, i)); | |
1143 | else if (fmt[i] == 'E') | |
1144 | for (j = 0; j < XVECLEN (x, i); j++) | |
1145 | changed |= mention_regs (XVECEXP (x, i, j)); | |
1146 | ||
1147 | return changed; | |
1148 | } | |
1149 | ||
1150 | /* Update the register quantities for inserting X into the hash table | |
1151 | with a value equivalent to CLASSP. | |
1152 | (If the class does not contain a REG, it is irrelevant.) | |
1153 | If MODIFIED is nonzero, X is a destination; it is being modified. | |
1154 | Note that delete_reg_equiv should be called on a register | |
1155 | before insert_regs is done on that register with MODIFIED != 0. | |
1156 | ||
1157 | Nonzero value means that elements of reg_qty have changed | |
1158 | so X's hash code may be different. */ | |
1159 | ||
1160 | static int | |
7080f735 | 1161 | insert_regs (rtx x, struct table_elt *classp, int modified) |
7afe21cc | 1162 | { |
f8cfc6aa | 1163 | if (REG_P (x)) |
7afe21cc | 1164 | { |
770ae6cc RK |
1165 | unsigned int regno = REGNO (x); |
1166 | int qty_valid; | |
7afe21cc | 1167 | |
1ff0c00d RK |
1168 | /* If REGNO is in the equivalence table already but is of the |
1169 | wrong mode for that equivalence, don't do anything here. */ | |
1170 | ||
1bb98cec DM |
1171 | qty_valid = REGNO_QTY_VALID_P (regno); |
1172 | if (qty_valid) | |
1173 | { | |
1174 | struct qty_table_elem *ent = &qty_table[REG_QTY (regno)]; | |
1ff0c00d | 1175 | |
1bb98cec DM |
1176 | if (ent->mode != GET_MODE (x)) |
1177 | return 0; | |
1178 | } | |
1179 | ||
1180 | if (modified || ! qty_valid) | |
7afe21cc RK |
1181 | { |
1182 | if (classp) | |
1183 | for (classp = classp->first_same_value; | |
1184 | classp != 0; | |
1185 | classp = classp->next_same_value) | |
f8cfc6aa | 1186 | if (REG_P (classp->exp) |
7afe21cc RK |
1187 | && GET_MODE (classp->exp) == GET_MODE (x)) |
1188 | { | |
1189 | make_regs_eqv (regno, REGNO (classp->exp)); | |
1190 | return 1; | |
1191 | } | |
1192 | ||
d9f20424 R |
1193 | /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger |
1194 | than REG_IN_TABLE to find out if there was only a single preceding | |
1195 | invalidation - for the SUBREG - or another one, which would be | |
1196 | for the full register. However, if we find here that REG_TICK | |
1197 | indicates that the register is invalid, it means that it has | |
1198 | been invalidated in a separate operation. The SUBREG might be used | |
1199 | now (then this is a recursive call), or we might use the full REG | |
1200 | now and a SUBREG of it later. So bump up REG_TICK so that | |
1201 | mention_regs will do the right thing. */ | |
1202 | if (! modified | |
1203 | && REG_IN_TABLE (regno) >= 0 | |
1204 | && REG_TICK (regno) == REG_IN_TABLE (regno) + 1) | |
1205 | REG_TICK (regno)++; | |
1bb98cec | 1206 | make_new_qty (regno, GET_MODE (x)); |
7afe21cc RK |
1207 | return 1; |
1208 | } | |
cdf4112f TG |
1209 | |
1210 | return 0; | |
7afe21cc | 1211 | } |
c610adec RK |
1212 | |
1213 | /* If X is a SUBREG, we will likely be inserting the inner register in the | |
1214 | table. If that register doesn't have an assigned quantity number at | |
1215 | this point but does later, the insertion that we will be doing now will | |
1216 | not be accessible because its hash code will have changed. So assign | |
1217 | a quantity number now. */ | |
1218 | ||
f8cfc6aa | 1219 | else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x)) |
c610adec RK |
1220 | && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x)))) |
1221 | { | |
9714cf43 | 1222 | insert_regs (SUBREG_REG (x), NULL, 0); |
34c73909 | 1223 | mention_regs (x); |
c610adec RK |
1224 | return 1; |
1225 | } | |
7afe21cc RK |
1226 | else |
1227 | return mention_regs (x); | |
1228 | } | |
1229 | \f | |
1230 | /* Look in or update the hash table. */ | |
1231 | ||
7afe21cc RK |
1232 | /* Remove table element ELT from use in the table. |
1233 | HASH is its hash code, made using the HASH macro. | |
1234 | It's an argument because often that is known in advance | |
1235 | and we save much time not recomputing it. */ | |
1236 | ||
1237 | static void | |
7080f735 | 1238 | remove_from_table (struct table_elt *elt, unsigned int hash) |
7afe21cc RK |
1239 | { |
1240 | if (elt == 0) | |
1241 | return; | |
1242 | ||
1243 | /* Mark this element as removed. See cse_insn. */ | |
1244 | elt->first_same_value = 0; | |
1245 | ||
1246 | /* Remove the table element from its equivalence class. */ | |
278a83b2 | 1247 | |
7afe21cc | 1248 | { |
b3694847 SS |
1249 | struct table_elt *prev = elt->prev_same_value; |
1250 | struct table_elt *next = elt->next_same_value; | |
7afe21cc | 1251 | |
278a83b2 KH |
1252 | if (next) |
1253 | next->prev_same_value = prev; | |
7afe21cc RK |
1254 | |
1255 | if (prev) | |
1256 | prev->next_same_value = next; | |
1257 | else | |
1258 | { | |
b3694847 | 1259 | struct table_elt *newfirst = next; |
7afe21cc RK |
1260 | while (next) |
1261 | { | |
1262 | next->first_same_value = newfirst; | |
1263 | next = next->next_same_value; | |
1264 | } | |
1265 | } | |
1266 | } | |
1267 | ||
1268 | /* Remove the table element from its hash bucket. */ | |
1269 | ||
1270 | { | |
b3694847 SS |
1271 | struct table_elt *prev = elt->prev_same_hash; |
1272 | struct table_elt *next = elt->next_same_hash; | |
7afe21cc | 1273 | |
278a83b2 KH |
1274 | if (next) |
1275 | next->prev_same_hash = prev; | |
7afe21cc RK |
1276 | |
1277 | if (prev) | |
1278 | prev->next_same_hash = next; | |
1279 | else if (table[hash] == elt) | |
1280 | table[hash] = next; | |
1281 | else | |
1282 | { | |
1283 | /* This entry is not in the proper hash bucket. This can happen | |
1284 | when two classes were merged by `merge_equiv_classes'. Search | |
1285 | for the hash bucket that it heads. This happens only very | |
1286 | rarely, so the cost is acceptable. */ | |
9b1549b8 | 1287 | for (hash = 0; hash < HASH_SIZE; hash++) |
7afe21cc RK |
1288 | if (table[hash] == elt) |
1289 | table[hash] = next; | |
1290 | } | |
1291 | } | |
1292 | ||
1293 | /* Remove the table element from its related-value circular chain. */ | |
1294 | ||
1295 | if (elt->related_value != 0 && elt->related_value != elt) | |
1296 | { | |
b3694847 | 1297 | struct table_elt *p = elt->related_value; |
770ae6cc | 1298 | |
7afe21cc RK |
1299 | while (p->related_value != elt) |
1300 | p = p->related_value; | |
1301 | p->related_value = elt->related_value; | |
1302 | if (p->related_value == p) | |
1303 | p->related_value = 0; | |
1304 | } | |
1305 | ||
9b1549b8 DM |
1306 | /* Now add it to the free element chain. */ |
1307 | elt->next_same_hash = free_element_chain; | |
1308 | free_element_chain = elt; | |
7afe21cc RK |
1309 | } |
1310 | ||
1311 | /* Look up X in the hash table and return its table element, | |
1312 | or 0 if X is not in the table. | |
1313 | ||
1314 | MODE is the machine-mode of X, or if X is an integer constant | |
1315 | with VOIDmode then MODE is the mode with which X will be used. | |
1316 | ||
1317 | Here we are satisfied to find an expression whose tree structure | |
1318 | looks like X. */ | |
1319 | ||
1320 | static struct table_elt * | |
7080f735 | 1321 | lookup (rtx x, unsigned int hash, enum machine_mode mode) |
7afe21cc | 1322 | { |
b3694847 | 1323 | struct table_elt *p; |
7afe21cc RK |
1324 | |
1325 | for (p = table[hash]; p; p = p->next_same_hash) | |
f8cfc6aa JQ |
1326 | if (mode == p->mode && ((x == p->exp && REG_P (x)) |
1327 | || exp_equiv_p (x, p->exp, !REG_P (x), 0))) | |
7afe21cc RK |
1328 | return p; |
1329 | ||
1330 | return 0; | |
1331 | } | |
1332 | ||
1333 | /* Like `lookup' but don't care whether the table element uses invalid regs. | |
1334 | Also ignore discrepancies in the machine mode of a register. */ | |
1335 | ||
1336 | static struct table_elt * | |
7080f735 | 1337 | lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode) |
7afe21cc | 1338 | { |
b3694847 | 1339 | struct table_elt *p; |
7afe21cc | 1340 | |
f8cfc6aa | 1341 | if (REG_P (x)) |
7afe21cc | 1342 | { |
770ae6cc RK |
1343 | unsigned int regno = REGNO (x); |
1344 | ||
7afe21cc RK |
1345 | /* Don't check the machine mode when comparing registers; |
1346 | invalidating (REG:SI 0) also invalidates (REG:DF 0). */ | |
1347 | for (p = table[hash]; p; p = p->next_same_hash) | |
f8cfc6aa | 1348 | if (REG_P (p->exp) |
7afe21cc RK |
1349 | && REGNO (p->exp) == regno) |
1350 | return p; | |
1351 | } | |
1352 | else | |
1353 | { | |
1354 | for (p = table[hash]; p; p = p->next_same_hash) | |
1355 | if (mode == p->mode && (x == p->exp || exp_equiv_p (x, p->exp, 0, 0))) | |
1356 | return p; | |
1357 | } | |
1358 | ||
1359 | return 0; | |
1360 | } | |
1361 | ||
1362 | /* Look for an expression equivalent to X and with code CODE. | |
1363 | If one is found, return that expression. */ | |
1364 | ||
1365 | static rtx | |
7080f735 | 1366 | lookup_as_function (rtx x, enum rtx_code code) |
7afe21cc | 1367 | { |
b3694847 | 1368 | struct table_elt *p |
770ae6cc RK |
1369 | = lookup (x, safe_hash (x, VOIDmode) & HASH_MASK, GET_MODE (x)); |
1370 | ||
34c73909 R |
1371 | /* If we are looking for a CONST_INT, the mode doesn't really matter, as |
1372 | long as we are narrowing. So if we looked in vain for a mode narrower | |
1373 | than word_mode before, look for word_mode now. */ | |
1374 | if (p == 0 && code == CONST_INT | |
1375 | && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode)) | |
1376 | { | |
1377 | x = copy_rtx (x); | |
1378 | PUT_MODE (x, word_mode); | |
9b1549b8 | 1379 | p = lookup (x, safe_hash (x, VOIDmode) & HASH_MASK, word_mode); |
34c73909 R |
1380 | } |
1381 | ||
7afe21cc RK |
1382 | if (p == 0) |
1383 | return 0; | |
1384 | ||
1385 | for (p = p->first_same_value; p; p = p->next_same_value) | |
770ae6cc RK |
1386 | if (GET_CODE (p->exp) == code |
1387 | /* Make sure this is a valid entry in the table. */ | |
1388 | && exp_equiv_p (p->exp, p->exp, 1, 0)) | |
1389 | return p->exp; | |
278a83b2 | 1390 | |
7afe21cc RK |
1391 | return 0; |
1392 | } | |
1393 | ||
1394 | /* Insert X in the hash table, assuming HASH is its hash code | |
1395 | and CLASSP is an element of the class it should go in | |
1396 | (or 0 if a new class should be made). | |
1397 | It is inserted at the proper position to keep the class in | |
1398 | the order cheapest first. | |
1399 | ||
1400 | MODE is the machine-mode of X, or if X is an integer constant | |
1401 | with VOIDmode then MODE is the mode with which X will be used. | |
1402 | ||
1403 | For elements of equal cheapness, the most recent one | |
1404 | goes in front, except that the first element in the list | |
1405 | remains first unless a cheaper element is added. The order of | |
1406 | pseudo-registers does not matter, as canon_reg will be called to | |
830a38ee | 1407 | find the cheapest when a register is retrieved from the table. |
7afe21cc RK |
1408 | |
1409 | The in_memory field in the hash table element is set to 0. | |
1410 | The caller must set it nonzero if appropriate. | |
1411 | ||
1412 | You should call insert_regs (X, CLASSP, MODIFY) before calling here, | |
1413 | and if insert_regs returns a nonzero value | |
1414 | you must then recompute its hash code before calling here. | |
1415 | ||
1416 | If necessary, update table showing constant values of quantities. */ | |
1417 | ||
630c79be | 1418 | #define CHEAPER(X, Y) \ |
56ae04af | 1419 | (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0) |
7afe21cc RK |
1420 | |
1421 | static struct table_elt * | |
7080f735 | 1422 | insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode) |
7afe21cc | 1423 | { |
b3694847 | 1424 | struct table_elt *elt; |
7afe21cc RK |
1425 | |
1426 | /* If X is a register and we haven't made a quantity for it, | |
1427 | something is wrong. */ | |
f8cfc6aa | 1428 | if (REG_P (x) && ! REGNO_QTY_VALID_P (REGNO (x))) |
7afe21cc RK |
1429 | abort (); |
1430 | ||
1431 | /* If X is a hard register, show it is being put in the table. */ | |
f8cfc6aa | 1432 | if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER) |
7afe21cc | 1433 | { |
770ae6cc | 1434 | unsigned int regno = REGNO (x); |
66fd46b6 | 1435 | unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)]; |
770ae6cc | 1436 | unsigned int i; |
7afe21cc RK |
1437 | |
1438 | for (i = regno; i < endregno; i++) | |
770ae6cc | 1439 | SET_HARD_REG_BIT (hard_regs_in_table, i); |
7afe21cc RK |
1440 | } |
1441 | ||
7afe21cc RK |
1442 | /* Put an element for X into the right hash bucket. */ |
1443 | ||
9b1549b8 DM |
1444 | elt = free_element_chain; |
1445 | if (elt) | |
770ae6cc | 1446 | free_element_chain = elt->next_same_hash; |
9b1549b8 DM |
1447 | else |
1448 | { | |
1449 | n_elements_made++; | |
703ad42b | 1450 | elt = xmalloc (sizeof (struct table_elt)); |
9b1549b8 DM |
1451 | } |
1452 | ||
7afe21cc | 1453 | elt->exp = x; |
db048faf | 1454 | elt->canon_exp = NULL_RTX; |
7afe21cc | 1455 | elt->cost = COST (x); |
630c79be | 1456 | elt->regcost = approx_reg_cost (x); |
7afe21cc RK |
1457 | elt->next_same_value = 0; |
1458 | elt->prev_same_value = 0; | |
1459 | elt->next_same_hash = table[hash]; | |
1460 | elt->prev_same_hash = 0; | |
1461 | elt->related_value = 0; | |
1462 | elt->in_memory = 0; | |
1463 | elt->mode = mode; | |
1464 | elt->is_const = (CONSTANT_P (x) | |
1465 | /* GNU C++ takes advantage of this for `this' | |
1466 | (and other const values). */ | |
f8cfc6aa | 1467 | || (REG_P (x) |
2adc7f12 | 1468 | && RTX_UNCHANGING_P (x) |
7afe21cc | 1469 | && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
4977bab6 | 1470 | || fixed_base_plus_p (x)); |
7afe21cc RK |
1471 | |
1472 | if (table[hash]) | |
1473 | table[hash]->prev_same_hash = elt; | |
1474 | table[hash] = elt; | |
1475 | ||
1476 | /* Put it into the proper value-class. */ | |
1477 | if (classp) | |
1478 | { | |
1479 | classp = classp->first_same_value; | |
1480 | if (CHEAPER (elt, classp)) | |
f9da5064 | 1481 | /* Insert at the head of the class. */ |
7afe21cc | 1482 | { |
b3694847 | 1483 | struct table_elt *p; |
7afe21cc RK |
1484 | elt->next_same_value = classp; |
1485 | classp->prev_same_value = elt; | |
1486 | elt->first_same_value = elt; | |
1487 | ||
1488 | for (p = classp; p; p = p->next_same_value) | |
1489 | p->first_same_value = elt; | |
1490 | } | |
1491 | else | |
1492 | { | |
1493 | /* Insert not at head of the class. */ | |
1494 | /* Put it after the last element cheaper than X. */ | |
b3694847 | 1495 | struct table_elt *p, *next; |
770ae6cc | 1496 | |
7afe21cc RK |
1497 | for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt); |
1498 | p = next); | |
770ae6cc | 1499 | |
7afe21cc RK |
1500 | /* Put it after P and before NEXT. */ |
1501 | elt->next_same_value = next; | |
1502 | if (next) | |
1503 | next->prev_same_value = elt; | |
770ae6cc | 1504 | |
7afe21cc RK |
1505 | elt->prev_same_value = p; |
1506 | p->next_same_value = elt; | |
1507 | elt->first_same_value = classp; | |
1508 | } | |
1509 | } | |
1510 | else | |
1511 | elt->first_same_value = elt; | |
1512 | ||
1513 | /* If this is a constant being set equivalent to a register or a register | |
1514 | being set equivalent to a constant, note the constant equivalence. | |
1515 | ||
1516 | If this is a constant, it cannot be equivalent to a different constant, | |
1517 | and a constant is the only thing that can be cheaper than a register. So | |
1518 | we know the register is the head of the class (before the constant was | |
1519 | inserted). | |
1520 | ||
1521 | If this is a register that is not already known equivalent to a | |
1522 | constant, we must check the entire class. | |
1523 | ||
1524 | If this is a register that is already known equivalent to an insn, | |
1bb98cec | 1525 | update the qtys `const_insn' to show that `this_insn' is the latest |
7afe21cc RK |
1526 | insn making that quantity equivalent to the constant. */ |
1527 | ||
f8cfc6aa JQ |
1528 | if (elt->is_const && classp && REG_P (classp->exp) |
1529 | && !REG_P (x)) | |
7afe21cc | 1530 | { |
1bb98cec DM |
1531 | int exp_q = REG_QTY (REGNO (classp->exp)); |
1532 | struct qty_table_elem *exp_ent = &qty_table[exp_q]; | |
1533 | ||
4de249d9 | 1534 | exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x); |
1bb98cec | 1535 | exp_ent->const_insn = this_insn; |
7afe21cc RK |
1536 | } |
1537 | ||
f8cfc6aa | 1538 | else if (REG_P (x) |
1bb98cec DM |
1539 | && classp |
1540 | && ! qty_table[REG_QTY (REGNO (x))].const_rtx | |
f353588a | 1541 | && ! elt->is_const) |
7afe21cc | 1542 | { |
b3694847 | 1543 | struct table_elt *p; |
7afe21cc RK |
1544 | |
1545 | for (p = classp; p != 0; p = p->next_same_value) | |
1546 | { | |
f8cfc6aa | 1547 | if (p->is_const && !REG_P (p->exp)) |
7afe21cc | 1548 | { |
1bb98cec DM |
1549 | int x_q = REG_QTY (REGNO (x)); |
1550 | struct qty_table_elem *x_ent = &qty_table[x_q]; | |
1551 | ||
770ae6cc | 1552 | x_ent->const_rtx |
4de249d9 | 1553 | = gen_lowpart (GET_MODE (x), p->exp); |
1bb98cec | 1554 | x_ent->const_insn = this_insn; |
7afe21cc RK |
1555 | break; |
1556 | } | |
1557 | } | |
1558 | } | |
1559 | ||
f8cfc6aa | 1560 | else if (REG_P (x) |
1bb98cec DM |
1561 | && qty_table[REG_QTY (REGNO (x))].const_rtx |
1562 | && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode) | |
1563 | qty_table[REG_QTY (REGNO (x))].const_insn = this_insn; | |
7afe21cc RK |
1564 | |
1565 | /* If this is a constant with symbolic value, | |
1566 | and it has a term with an explicit integer value, | |
1567 | link it up with related expressions. */ | |
1568 | if (GET_CODE (x) == CONST) | |
1569 | { | |
1570 | rtx subexp = get_related_value (x); | |
2197a88a | 1571 | unsigned subhash; |
7afe21cc RK |
1572 | struct table_elt *subelt, *subelt_prev; |
1573 | ||
1574 | if (subexp != 0) | |
1575 | { | |
1576 | /* Get the integer-free subexpression in the hash table. */ | |
9b1549b8 | 1577 | subhash = safe_hash (subexp, mode) & HASH_MASK; |
7afe21cc RK |
1578 | subelt = lookup (subexp, subhash, mode); |
1579 | if (subelt == 0) | |
9714cf43 | 1580 | subelt = insert (subexp, NULL, subhash, mode); |
7afe21cc RK |
1581 | /* Initialize SUBELT's circular chain if it has none. */ |
1582 | if (subelt->related_value == 0) | |
1583 | subelt->related_value = subelt; | |
1584 | /* Find the element in the circular chain that precedes SUBELT. */ | |
1585 | subelt_prev = subelt; | |
1586 | while (subelt_prev->related_value != subelt) | |
1587 | subelt_prev = subelt_prev->related_value; | |
1588 | /* Put new ELT into SUBELT's circular chain just before SUBELT. | |
1589 | This way the element that follows SUBELT is the oldest one. */ | |
1590 | elt->related_value = subelt_prev->related_value; | |
1591 | subelt_prev->related_value = elt; | |
1592 | } | |
1593 | } | |
1594 | ||
1595 | return elt; | |
1596 | } | |
1597 | \f | |
1598 | /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from | |
1599 | CLASS2 into CLASS1. This is done when we have reached an insn which makes | |
1600 | the two classes equivalent. | |
1601 | ||
1602 | CLASS1 will be the surviving class; CLASS2 should not be used after this | |
1603 | call. | |
1604 | ||
1605 | Any invalid entries in CLASS2 will not be copied. */ | |
1606 | ||
1607 | static void | |
7080f735 | 1608 | merge_equiv_classes (struct table_elt *class1, struct table_elt *class2) |
7afe21cc RK |
1609 | { |
1610 | struct table_elt *elt, *next, *new; | |
1611 | ||
1612 | /* Ensure we start with the head of the classes. */ | |
1613 | class1 = class1->first_same_value; | |
1614 | class2 = class2->first_same_value; | |
1615 | ||
1616 | /* If they were already equal, forget it. */ | |
1617 | if (class1 == class2) | |
1618 | return; | |
1619 | ||
1620 | for (elt = class2; elt; elt = next) | |
1621 | { | |
770ae6cc | 1622 | unsigned int hash; |
7afe21cc RK |
1623 | rtx exp = elt->exp; |
1624 | enum machine_mode mode = elt->mode; | |
1625 | ||
1626 | next = elt->next_same_value; | |
1627 | ||
1628 | /* Remove old entry, make a new one in CLASS1's class. | |
1629 | Don't do this for invalid entries as we cannot find their | |
0f41302f | 1630 | hash code (it also isn't necessary). */ |
f8cfc6aa | 1631 | if (REG_P (exp) || exp_equiv_p (exp, exp, 1, 0)) |
7afe21cc | 1632 | { |
a90fc8e0 RH |
1633 | bool need_rehash = false; |
1634 | ||
7afe21cc | 1635 | hash_arg_in_memory = 0; |
7afe21cc | 1636 | hash = HASH (exp, mode); |
278a83b2 | 1637 | |
f8cfc6aa | 1638 | if (REG_P (exp)) |
a90fc8e0 RH |
1639 | { |
1640 | need_rehash = (unsigned) REG_QTY (REGNO (exp)) != REGNO (exp); | |
1641 | delete_reg_equiv (REGNO (exp)); | |
1642 | } | |
278a83b2 | 1643 | |
7afe21cc RK |
1644 | remove_from_table (elt, hash); |
1645 | ||
a90fc8e0 | 1646 | if (insert_regs (exp, class1, 0) || need_rehash) |
8ae2b8f6 JW |
1647 | { |
1648 | rehash_using_reg (exp); | |
1649 | hash = HASH (exp, mode); | |
1650 | } | |
7afe21cc RK |
1651 | new = insert (exp, class1, hash, mode); |
1652 | new->in_memory = hash_arg_in_memory; | |
7afe21cc RK |
1653 | } |
1654 | } | |
1655 | } | |
1656 | \f | |
01e752d3 JL |
1657 | /* Flush the entire hash table. */ |
1658 | ||
1659 | static void | |
7080f735 | 1660 | flush_hash_table (void) |
01e752d3 JL |
1661 | { |
1662 | int i; | |
1663 | struct table_elt *p; | |
1664 | ||
9b1549b8 | 1665 | for (i = 0; i < HASH_SIZE; i++) |
01e752d3 JL |
1666 | for (p = table[i]; p; p = table[i]) |
1667 | { | |
1668 | /* Note that invalidate can remove elements | |
1669 | after P in the current hash chain. */ | |
f8cfc6aa | 1670 | if (REG_P (p->exp)) |
01e752d3 JL |
1671 | invalidate (p->exp, p->mode); |
1672 | else | |
1673 | remove_from_table (p, i); | |
1674 | } | |
1675 | } | |
14a774a9 | 1676 | \f |
2ce6dc2f JH |
1677 | /* Function called for each rtx to check whether true dependence exist. */ |
1678 | struct check_dependence_data | |
1679 | { | |
1680 | enum machine_mode mode; | |
1681 | rtx exp; | |
9ddb66ca | 1682 | rtx addr; |
2ce6dc2f | 1683 | }; |
be8ac49a | 1684 | |
2ce6dc2f | 1685 | static int |
7080f735 | 1686 | check_dependence (rtx *x, void *data) |
2ce6dc2f JH |
1687 | { |
1688 | struct check_dependence_data *d = (struct check_dependence_data *) data; | |
3c0cb5de | 1689 | if (*x && MEM_P (*x)) |
9ddb66ca JH |
1690 | return canon_true_dependence (d->exp, d->mode, d->addr, *x, |
1691 | cse_rtx_varies_p); | |
2ce6dc2f JH |
1692 | else |
1693 | return 0; | |
1694 | } | |
1695 | \f | |
14a774a9 RK |
1696 | /* Remove from the hash table, or mark as invalid, all expressions whose |
1697 | values could be altered by storing in X. X is a register, a subreg, or | |
1698 | a memory reference with nonvarying address (because, when a memory | |
1699 | reference with a varying address is stored in, all memory references are | |
1700 | removed by invalidate_memory so specific invalidation is superfluous). | |
1701 | FULL_MODE, if not VOIDmode, indicates that this much should be | |
1702 | invalidated instead of just the amount indicated by the mode of X. This | |
1703 | is only used for bitfield stores into memory. | |
1704 | ||
1705 | A nonvarying address may be just a register or just a symbol reference, | |
1706 | or it may be either of those plus a numeric offset. */ | |
7afe21cc RK |
1707 | |
1708 | static void | |
7080f735 | 1709 | invalidate (rtx x, enum machine_mode full_mode) |
7afe21cc | 1710 | { |
b3694847 SS |
1711 | int i; |
1712 | struct table_elt *p; | |
9ddb66ca | 1713 | rtx addr; |
7afe21cc | 1714 | |
14a774a9 | 1715 | switch (GET_CODE (x)) |
7afe21cc | 1716 | { |
14a774a9 RK |
1717 | case REG: |
1718 | { | |
1719 | /* If X is a register, dependencies on its contents are recorded | |
1720 | through the qty number mechanism. Just change the qty number of | |
1721 | the register, mark it as invalid for expressions that refer to it, | |
1722 | and remove it itself. */ | |
770ae6cc RK |
1723 | unsigned int regno = REGNO (x); |
1724 | unsigned int hash = HASH (x, GET_MODE (x)); | |
7afe21cc | 1725 | |
14a774a9 RK |
1726 | /* Remove REGNO from any quantity list it might be on and indicate |
1727 | that its value might have changed. If it is a pseudo, remove its | |
1728 | entry from the hash table. | |
7afe21cc | 1729 | |
14a774a9 RK |
1730 | For a hard register, we do the first two actions above for any |
1731 | additional hard registers corresponding to X. Then, if any of these | |
1732 | registers are in the table, we must remove any REG entries that | |
1733 | overlap these registers. */ | |
7afe21cc | 1734 | |
14a774a9 RK |
1735 | delete_reg_equiv (regno); |
1736 | REG_TICK (regno)++; | |
46081bb3 | 1737 | SUBREG_TICKED (regno) = -1; |
85e4d983 | 1738 | |
14a774a9 RK |
1739 | if (regno >= FIRST_PSEUDO_REGISTER) |
1740 | { | |
1741 | /* Because a register can be referenced in more than one mode, | |
1742 | we might have to remove more than one table entry. */ | |
1743 | struct table_elt *elt; | |
85e4d983 | 1744 | |
14a774a9 RK |
1745 | while ((elt = lookup_for_remove (x, hash, GET_MODE (x)))) |
1746 | remove_from_table (elt, hash); | |
1747 | } | |
1748 | else | |
1749 | { | |
1750 | HOST_WIDE_INT in_table | |
1751 | = TEST_HARD_REG_BIT (hard_regs_in_table, regno); | |
770ae6cc | 1752 | unsigned int endregno |
66fd46b6 | 1753 | = regno + hard_regno_nregs[regno][GET_MODE (x)]; |
770ae6cc | 1754 | unsigned int tregno, tendregno, rn; |
b3694847 | 1755 | struct table_elt *p, *next; |
7afe21cc | 1756 | |
14a774a9 | 1757 | CLEAR_HARD_REG_BIT (hard_regs_in_table, regno); |
7afe21cc | 1758 | |
770ae6cc | 1759 | for (rn = regno + 1; rn < endregno; rn++) |
14a774a9 | 1760 | { |
770ae6cc RK |
1761 | in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn); |
1762 | CLEAR_HARD_REG_BIT (hard_regs_in_table, rn); | |
1763 | delete_reg_equiv (rn); | |
1764 | REG_TICK (rn)++; | |
46081bb3 | 1765 | SUBREG_TICKED (rn) = -1; |
14a774a9 | 1766 | } |
7afe21cc | 1767 | |
14a774a9 | 1768 | if (in_table) |
9b1549b8 | 1769 | for (hash = 0; hash < HASH_SIZE; hash++) |
14a774a9 RK |
1770 | for (p = table[hash]; p; p = next) |
1771 | { | |
1772 | next = p->next_same_hash; | |
7afe21cc | 1773 | |
f8cfc6aa | 1774 | if (!REG_P (p->exp) |
278a83b2 KH |
1775 | || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER) |
1776 | continue; | |
1777 | ||
14a774a9 RK |
1778 | tregno = REGNO (p->exp); |
1779 | tendregno | |
66fd46b6 | 1780 | = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)]; |
14a774a9 RK |
1781 | if (tendregno > regno && tregno < endregno) |
1782 | remove_from_table (p, hash); | |
1783 | } | |
1784 | } | |
1785 | } | |
7afe21cc | 1786 | return; |
7afe21cc | 1787 | |
14a774a9 | 1788 | case SUBREG: |
bb4034b3 | 1789 | invalidate (SUBREG_REG (x), VOIDmode); |
7afe21cc | 1790 | return; |
aac5cc16 | 1791 | |
14a774a9 | 1792 | case PARALLEL: |
278a83b2 | 1793 | for (i = XVECLEN (x, 0) - 1; i >= 0; --i) |
aac5cc16 RH |
1794 | invalidate (XVECEXP (x, 0, i), VOIDmode); |
1795 | return; | |
aac5cc16 | 1796 | |
14a774a9 RK |
1797 | case EXPR_LIST: |
1798 | /* This is part of a disjoint return value; extract the location in | |
1799 | question ignoring the offset. */ | |
aac5cc16 RH |
1800 | invalidate (XEXP (x, 0), VOIDmode); |
1801 | return; | |
7afe21cc | 1802 | |
14a774a9 | 1803 | case MEM: |
9ddb66ca | 1804 | addr = canon_rtx (get_addr (XEXP (x, 0))); |
db048faf MM |
1805 | /* Calculate the canonical version of X here so that |
1806 | true_dependence doesn't generate new RTL for X on each call. */ | |
1807 | x = canon_rtx (x); | |
1808 | ||
14a774a9 RK |
1809 | /* Remove all hash table elements that refer to overlapping pieces of |
1810 | memory. */ | |
1811 | if (full_mode == VOIDmode) | |
1812 | full_mode = GET_MODE (x); | |
bb4034b3 | 1813 | |
9b1549b8 | 1814 | for (i = 0; i < HASH_SIZE; i++) |
7afe21cc | 1815 | { |
b3694847 | 1816 | struct table_elt *next; |
14a774a9 RK |
1817 | |
1818 | for (p = table[i]; p; p = next) | |
1819 | { | |
1820 | next = p->next_same_hash; | |
db048faf MM |
1821 | if (p->in_memory) |
1822 | { | |
2ce6dc2f JH |
1823 | struct check_dependence_data d; |
1824 | ||
1825 | /* Just canonicalize the expression once; | |
1826 | otherwise each time we call invalidate | |
1827 | true_dependence will canonicalize the | |
1828 | expression again. */ | |
1829 | if (!p->canon_exp) | |
1830 | p->canon_exp = canon_rtx (p->exp); | |
1831 | d.exp = x; | |
9ddb66ca | 1832 | d.addr = addr; |
2ce6dc2f JH |
1833 | d.mode = full_mode; |
1834 | if (for_each_rtx (&p->canon_exp, check_dependence, &d)) | |
db048faf | 1835 | remove_from_table (p, i); |
db048faf | 1836 | } |
14a774a9 | 1837 | } |
7afe21cc | 1838 | } |
14a774a9 RK |
1839 | return; |
1840 | ||
1841 | default: | |
1842 | abort (); | |
7afe21cc RK |
1843 | } |
1844 | } | |
14a774a9 | 1845 | \f |
7afe21cc RK |
1846 | /* Remove all expressions that refer to register REGNO, |
1847 | since they are already invalid, and we are about to | |
1848 | mark that register valid again and don't want the old | |
1849 | expressions to reappear as valid. */ | |
1850 | ||
1851 | static void | |
7080f735 | 1852 | remove_invalid_refs (unsigned int regno) |
7afe21cc | 1853 | { |
770ae6cc RK |
1854 | unsigned int i; |
1855 | struct table_elt *p, *next; | |
7afe21cc | 1856 | |
9b1549b8 | 1857 | for (i = 0; i < HASH_SIZE; i++) |
7afe21cc RK |
1858 | for (p = table[i]; p; p = next) |
1859 | { | |
1860 | next = p->next_same_hash; | |
f8cfc6aa | 1861 | if (!REG_P (p->exp) |
68252e27 | 1862 | && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0)) |
7afe21cc RK |
1863 | remove_from_table (p, i); |
1864 | } | |
1865 | } | |
34c73909 | 1866 | |
ddef6bc7 JJ |
1867 | /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET, |
1868 | and mode MODE. */ | |
34c73909 | 1869 | static void |
7080f735 AJ |
1870 | remove_invalid_subreg_refs (unsigned int regno, unsigned int offset, |
1871 | enum machine_mode mode) | |
34c73909 | 1872 | { |
770ae6cc RK |
1873 | unsigned int i; |
1874 | struct table_elt *p, *next; | |
ddef6bc7 | 1875 | unsigned int end = offset + (GET_MODE_SIZE (mode) - 1); |
34c73909 | 1876 | |
9b1549b8 | 1877 | for (i = 0; i < HASH_SIZE; i++) |
34c73909 R |
1878 | for (p = table[i]; p; p = next) |
1879 | { | |
ddef6bc7 | 1880 | rtx exp = p->exp; |
34c73909 | 1881 | next = p->next_same_hash; |
278a83b2 | 1882 | |
f8cfc6aa | 1883 | if (!REG_P (exp) |
34c73909 | 1884 | && (GET_CODE (exp) != SUBREG |
f8cfc6aa | 1885 | || !REG_P (SUBREG_REG (exp)) |
34c73909 | 1886 | || REGNO (SUBREG_REG (exp)) != regno |
ddef6bc7 JJ |
1887 | || (((SUBREG_BYTE (exp) |
1888 | + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset) | |
1889 | && SUBREG_BYTE (exp) <= end)) | |
68252e27 | 1890 | && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0)) |
34c73909 R |
1891 | remove_from_table (p, i); |
1892 | } | |
1893 | } | |
7afe21cc RK |
1894 | \f |
1895 | /* Recompute the hash codes of any valid entries in the hash table that | |
1896 | reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG. | |
1897 | ||
1898 | This is called when we make a jump equivalence. */ | |
1899 | ||
1900 | static void | |
7080f735 | 1901 | rehash_using_reg (rtx x) |
7afe21cc | 1902 | { |
973838fd | 1903 | unsigned int i; |
7afe21cc | 1904 | struct table_elt *p, *next; |
2197a88a | 1905 | unsigned hash; |
7afe21cc RK |
1906 | |
1907 | if (GET_CODE (x) == SUBREG) | |
1908 | x = SUBREG_REG (x); | |
1909 | ||
1910 | /* If X is not a register or if the register is known not to be in any | |
1911 | valid entries in the table, we have no work to do. */ | |
1912 | ||
f8cfc6aa | 1913 | if (!REG_P (x) |
30f72379 MM |
1914 | || REG_IN_TABLE (REGNO (x)) < 0 |
1915 | || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x))) | |
7afe21cc RK |
1916 | return; |
1917 | ||
1918 | /* Scan all hash chains looking for valid entries that mention X. | |
a90fc8e0 | 1919 | If we find one and it is in the wrong hash chain, move it. */ |
7afe21cc | 1920 | |
9b1549b8 | 1921 | for (i = 0; i < HASH_SIZE; i++) |
7afe21cc RK |
1922 | for (p = table[i]; p; p = next) |
1923 | { | |
1924 | next = p->next_same_hash; | |
a90fc8e0 | 1925 | if (reg_mentioned_p (x, p->exp) |
538b78e7 | 1926 | && exp_equiv_p (p->exp, p->exp, 1, 0) |
9b1549b8 | 1927 | && i != (hash = safe_hash (p->exp, p->mode) & HASH_MASK)) |
7afe21cc RK |
1928 | { |
1929 | if (p->next_same_hash) | |
1930 | p->next_same_hash->prev_same_hash = p->prev_same_hash; | |
1931 | ||
1932 | if (p->prev_same_hash) | |
1933 | p->prev_same_hash->next_same_hash = p->next_same_hash; | |
1934 | else | |
1935 | table[i] = p->next_same_hash; | |
1936 | ||
1937 | p->next_same_hash = table[hash]; | |
1938 | p->prev_same_hash = 0; | |
1939 | if (table[hash]) | |
1940 | table[hash]->prev_same_hash = p; | |
1941 | table[hash] = p; | |
1942 | } | |
1943 | } | |
1944 | } | |
1945 | \f | |
7afe21cc RK |
1946 | /* Remove from the hash table any expression that is a call-clobbered |
1947 | register. Also update their TICK values. */ | |
1948 | ||
1949 | static void | |
7080f735 | 1950 | invalidate_for_call (void) |
7afe21cc | 1951 | { |
770ae6cc RK |
1952 | unsigned int regno, endregno; |
1953 | unsigned int i; | |
2197a88a | 1954 | unsigned hash; |
7afe21cc RK |
1955 | struct table_elt *p, *next; |
1956 | int in_table = 0; | |
1957 | ||
1958 | /* Go through all the hard registers. For each that is clobbered in | |
1959 | a CALL_INSN, remove the register from quantity chains and update | |
1960 | reg_tick if defined. Also see if any of these registers is currently | |
1961 | in the table. */ | |
1962 | ||
1963 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
1964 | if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno)) | |
1965 | { | |
1966 | delete_reg_equiv (regno); | |
30f72379 | 1967 | if (REG_TICK (regno) >= 0) |
46081bb3 SH |
1968 | { |
1969 | REG_TICK (regno)++; | |
1970 | SUBREG_TICKED (regno) = -1; | |
1971 | } | |
7afe21cc | 1972 | |
0e227018 | 1973 | in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0); |
7afe21cc RK |
1974 | } |
1975 | ||
1976 | /* In the case where we have no call-clobbered hard registers in the | |
1977 | table, we are done. Otherwise, scan the table and remove any | |
1978 | entry that overlaps a call-clobbered register. */ | |
1979 | ||
1980 | if (in_table) | |
9b1549b8 | 1981 | for (hash = 0; hash < HASH_SIZE; hash++) |
7afe21cc RK |
1982 | for (p = table[hash]; p; p = next) |
1983 | { | |
1984 | next = p->next_same_hash; | |
1985 | ||
f8cfc6aa | 1986 | if (!REG_P (p->exp) |
7afe21cc RK |
1987 | || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER) |
1988 | continue; | |
1989 | ||
1990 | regno = REGNO (p->exp); | |
66fd46b6 | 1991 | endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)]; |
7afe21cc RK |
1992 | |
1993 | for (i = regno; i < endregno; i++) | |
1994 | if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)) | |
1995 | { | |
1996 | remove_from_table (p, hash); | |
1997 | break; | |
1998 | } | |
1999 | } | |
2000 | } | |
2001 | \f | |
2002 | /* Given an expression X of type CONST, | |
2003 | and ELT which is its table entry (or 0 if it | |
2004 | is not in the hash table), | |
2005 | return an alternate expression for X as a register plus integer. | |
2006 | If none can be found, return 0. */ | |
2007 | ||
2008 | static rtx | |
7080f735 | 2009 | use_related_value (rtx x, struct table_elt *elt) |
7afe21cc | 2010 | { |
b3694847 SS |
2011 | struct table_elt *relt = 0; |
2012 | struct table_elt *p, *q; | |
906c4e36 | 2013 | HOST_WIDE_INT offset; |
7afe21cc RK |
2014 | |
2015 | /* First, is there anything related known? | |
2016 | If we have a table element, we can tell from that. | |
2017 | Otherwise, must look it up. */ | |
2018 | ||
2019 | if (elt != 0 && elt->related_value != 0) | |
2020 | relt = elt; | |
2021 | else if (elt == 0 && GET_CODE (x) == CONST) | |
2022 | { | |
2023 | rtx subexp = get_related_value (x); | |
2024 | if (subexp != 0) | |
2025 | relt = lookup (subexp, | |
9b1549b8 | 2026 | safe_hash (subexp, GET_MODE (subexp)) & HASH_MASK, |
7afe21cc RK |
2027 | GET_MODE (subexp)); |
2028 | } | |
2029 | ||
2030 | if (relt == 0) | |
2031 | return 0; | |
2032 | ||
2033 | /* Search all related table entries for one that has an | |
2034 | equivalent register. */ | |
2035 | ||
2036 | p = relt; | |
2037 | while (1) | |
2038 | { | |
2039 | /* This loop is strange in that it is executed in two different cases. | |
2040 | The first is when X is already in the table. Then it is searching | |
2041 | the RELATED_VALUE list of X's class (RELT). The second case is when | |
2042 | X is not in the table. Then RELT points to a class for the related | |
2043 | value. | |
2044 | ||
2045 | Ensure that, whatever case we are in, that we ignore classes that have | |
2046 | the same value as X. */ | |
2047 | ||
2048 | if (rtx_equal_p (x, p->exp)) | |
2049 | q = 0; | |
2050 | else | |
2051 | for (q = p->first_same_value; q; q = q->next_same_value) | |
f8cfc6aa | 2052 | if (REG_P (q->exp)) |
7afe21cc RK |
2053 | break; |
2054 | ||
2055 | if (q) | |
2056 | break; | |
2057 | ||
2058 | p = p->related_value; | |
2059 | ||
2060 | /* We went all the way around, so there is nothing to be found. | |
2061 | Alternatively, perhaps RELT was in the table for some other reason | |
2062 | and it has no related values recorded. */ | |
2063 | if (p == relt || p == 0) | |
2064 | break; | |
2065 | } | |
2066 | ||
2067 | if (q == 0) | |
2068 | return 0; | |
2069 | ||
2070 | offset = (get_integer_term (x) - get_integer_term (p->exp)); | |
2071 | /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */ | |
2072 | return plus_constant (q->exp, offset); | |
2073 | } | |
2074 | \f | |
6462bb43 AO |
2075 | /* Hash a string. Just add its bytes up. */ |
2076 | static inline unsigned | |
7080f735 | 2077 | canon_hash_string (const char *ps) |
6462bb43 AO |
2078 | { |
2079 | unsigned hash = 0; | |
68252e27 KH |
2080 | const unsigned char *p = (const unsigned char *) ps; |
2081 | ||
6462bb43 AO |
2082 | if (p) |
2083 | while (*p) | |
2084 | hash += *p++; | |
2085 | ||
2086 | return hash; | |
2087 | } | |
2088 | ||
7afe21cc RK |
2089 | /* Hash an rtx. We are careful to make sure the value is never negative. |
2090 | Equivalent registers hash identically. | |
2091 | MODE is used in hashing for CONST_INTs only; | |
2092 | otherwise the mode of X is used. | |
2093 | ||
2094 | Store 1 in do_not_record if any subexpression is volatile. | |
2095 | ||
2096 | Store 1 in hash_arg_in_memory if X contains a MEM rtx | |
2097 | which does not have the RTX_UNCHANGING_P bit set. | |
7afe21cc RK |
2098 | |
2099 | Note that cse_insn knows that the hash code of a MEM expression | |
2100 | is just (int) MEM plus the hash code of the address. */ | |
2101 | ||
2197a88a | 2102 | static unsigned |
7080f735 | 2103 | canon_hash (rtx x, enum machine_mode mode) |
7afe21cc | 2104 | { |
b3694847 SS |
2105 | int i, j; |
2106 | unsigned hash = 0; | |
2107 | enum rtx_code code; | |
2108 | const char *fmt; | |
7afe21cc RK |
2109 | |
2110 | /* repeat is used to turn tail-recursion into iteration. */ | |
2111 | repeat: | |
2112 | if (x == 0) | |
2113 | return hash; | |
2114 | ||
2115 | code = GET_CODE (x); | |
2116 | switch (code) | |
2117 | { | |
2118 | case REG: | |
2119 | { | |
770ae6cc | 2120 | unsigned int regno = REGNO (x); |
1441374b | 2121 | bool record; |
7afe21cc RK |
2122 | |
2123 | /* On some machines, we can't record any non-fixed hard register, | |
2124 | because extending its life will cause reload problems. We | |
1441374b | 2125 | consider ap, fp, sp, gp to be fixed for this purpose. |
9a794e50 RH |
2126 | |
2127 | We also consider CCmode registers to be fixed for this purpose; | |
2128 | failure to do so leads to failure to simplify 0<100 type of | |
2129 | conditionals. | |
2130 | ||
68252e27 | 2131 | On all machines, we can't record any global registers. |
8bf4dfc2 GK |
2132 | Nor should we record any register that is in a small |
2133 | class, as defined by CLASS_LIKELY_SPILLED_P. */ | |
7afe21cc | 2134 | |
1441374b RH |
2135 | if (regno >= FIRST_PSEUDO_REGISTER) |
2136 | record = true; | |
2137 | else if (x == frame_pointer_rtx | |
2138 | || x == hard_frame_pointer_rtx | |
2139 | || x == arg_pointer_rtx | |
2140 | || x == stack_pointer_rtx | |
2141 | || x == pic_offset_table_rtx) | |
2142 | record = true; | |
2143 | else if (global_regs[regno]) | |
2144 | record = false; | |
2145 | else if (fixed_regs[regno]) | |
2146 | record = true; | |
2147 | else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC) | |
2148 | record = true; | |
2149 | else if (SMALL_REGISTER_CLASSES) | |
2150 | record = false; | |
2151 | else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno))) | |
2152 | record = false; | |
2153 | else | |
2154 | record = true; | |
7080f735 | 2155 | |
1441374b | 2156 | if (!record) |
7afe21cc RK |
2157 | { |
2158 | do_not_record = 1; | |
2159 | return 0; | |
2160 | } | |
770ae6cc | 2161 | |
30f72379 | 2162 | hash += ((unsigned) REG << 7) + (unsigned) REG_QTY (regno); |
2197a88a | 2163 | return hash; |
7afe21cc RK |
2164 | } |
2165 | ||
34c73909 R |
2166 | /* We handle SUBREG of a REG specially because the underlying |
2167 | reg changes its hash value with every value change; we don't | |
2168 | want to have to forget unrelated subregs when one subreg changes. */ | |
2169 | case SUBREG: | |
2170 | { | |
f8cfc6aa | 2171 | if (REG_P (SUBREG_REG (x))) |
34c73909 R |
2172 | { |
2173 | hash += (((unsigned) SUBREG << 7) | |
ddef6bc7 JJ |
2174 | + REGNO (SUBREG_REG (x)) |
2175 | + (SUBREG_BYTE (x) / UNITS_PER_WORD)); | |
34c73909 R |
2176 | return hash; |
2177 | } | |
2178 | break; | |
2179 | } | |
2180 | ||
7afe21cc | 2181 | case CONST_INT: |
2197a88a RK |
2182 | { |
2183 | unsigned HOST_WIDE_INT tem = INTVAL (x); | |
2184 | hash += ((unsigned) CONST_INT << 7) + (unsigned) mode + tem; | |
2185 | return hash; | |
2186 | } | |
7afe21cc RK |
2187 | |
2188 | case CONST_DOUBLE: | |
2189 | /* This is like the general case, except that it only counts | |
2190 | the integers representing the constant. */ | |
2197a88a | 2191 | hash += (unsigned) code + (unsigned) GET_MODE (x); |
969c8517 | 2192 | if (GET_MODE (x) != VOIDmode) |
46b33600 | 2193 | hash += real_hash (CONST_DOUBLE_REAL_VALUE (x)); |
969c8517 RK |
2194 | else |
2195 | hash += ((unsigned) CONST_DOUBLE_LOW (x) | |
2196 | + (unsigned) CONST_DOUBLE_HIGH (x)); | |
7afe21cc RK |
2197 | return hash; |
2198 | ||
69ef87e2 AH |
2199 | case CONST_VECTOR: |
2200 | { | |
2201 | int units; | |
2202 | rtx elt; | |
2203 | ||
2204 | units = CONST_VECTOR_NUNITS (x); | |
2205 | ||
2206 | for (i = 0; i < units; ++i) | |
2207 | { | |
2208 | elt = CONST_VECTOR_ELT (x, i); | |
2209 | hash += canon_hash (elt, GET_MODE (elt)); | |
2210 | } | |
2211 | ||
2212 | return hash; | |
2213 | } | |
2214 | ||
7afe21cc RK |
2215 | /* Assume there is only one rtx object for any given label. */ |
2216 | case LABEL_REF: | |
ddc356e8 | 2217 | hash += ((unsigned) LABEL_REF << 7) + (unsigned long) XEXP (x, 0); |
2197a88a | 2218 | return hash; |
7afe21cc RK |
2219 | |
2220 | case SYMBOL_REF: | |
ddc356e8 | 2221 | hash += ((unsigned) SYMBOL_REF << 7) + (unsigned long) XSTR (x, 0); |
2197a88a | 2222 | return hash; |
7afe21cc RK |
2223 | |
2224 | case MEM: | |
14a774a9 RK |
2225 | /* We don't record if marked volatile or if BLKmode since we don't |
2226 | know the size of the move. */ | |
2227 | if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode) | |
7afe21cc RK |
2228 | { |
2229 | do_not_record = 1; | |
2230 | return 0; | |
2231 | } | |
4977bab6 ZW |
2232 | if (! RTX_UNCHANGING_P (x) || fixed_base_plus_p (XEXP (x, 0))) |
2233 | hash_arg_in_memory = 1; | |
2234 | ||
7afe21cc RK |
2235 | /* Now that we have already found this special case, |
2236 | might as well speed it up as much as possible. */ | |
2197a88a | 2237 | hash += (unsigned) MEM; |
7afe21cc RK |
2238 | x = XEXP (x, 0); |
2239 | goto repeat; | |
2240 | ||
bb07060a JW |
2241 | case USE: |
2242 | /* A USE that mentions non-volatile memory needs special | |
2243 | handling since the MEM may be BLKmode which normally | |
2244 | prevents an entry from being made. Pure calls are | |
2245 | marked by a USE which mentions BLKmode memory. */ | |
3c0cb5de | 2246 | if (MEM_P (XEXP (x, 0)) |
bb07060a JW |
2247 | && ! MEM_VOLATILE_P (XEXP (x, 0))) |
2248 | { | |
68252e27 | 2249 | hash += (unsigned) USE; |
bb07060a JW |
2250 | x = XEXP (x, 0); |
2251 | ||
4977bab6 | 2252 | if (! RTX_UNCHANGING_P (x) || fixed_base_plus_p (XEXP (x, 0))) |
bb07060a JW |
2253 | hash_arg_in_memory = 1; |
2254 | ||
2255 | /* Now that we have already found this special case, | |
2256 | might as well speed it up as much as possible. */ | |
2257 | hash += (unsigned) MEM; | |
2258 | x = XEXP (x, 0); | |
2259 | goto repeat; | |
2260 | } | |
2261 | break; | |
2262 | ||
7afe21cc RK |
2263 | case PRE_DEC: |
2264 | case PRE_INC: | |
2265 | case POST_DEC: | |
2266 | case POST_INC: | |
4b983fdc RH |
2267 | case PRE_MODIFY: |
2268 | case POST_MODIFY: | |
7afe21cc RK |
2269 | case PC: |
2270 | case CC0: | |
2271 | case CALL: | |
2272 | case UNSPEC_VOLATILE: | |
2273 | do_not_record = 1; | |
2274 | return 0; | |
2275 | ||
2276 | case ASM_OPERANDS: | |
2277 | if (MEM_VOLATILE_P (x)) | |
2278 | { | |
2279 | do_not_record = 1; | |
2280 | return 0; | |
2281 | } | |
6462bb43 AO |
2282 | else |
2283 | { | |
2284 | /* We don't want to take the filename and line into account. */ | |
2285 | hash += (unsigned) code + (unsigned) GET_MODE (x) | |
2286 | + canon_hash_string (ASM_OPERANDS_TEMPLATE (x)) | |
2287 | + canon_hash_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x)) | |
2288 | + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x); | |
2289 | ||
2290 | if (ASM_OPERANDS_INPUT_LENGTH (x)) | |
2291 | { | |
2292 | for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++) | |
2293 | { | |
2294 | hash += (canon_hash (ASM_OPERANDS_INPUT (x, i), | |
2295 | GET_MODE (ASM_OPERANDS_INPUT (x, i))) | |
2296 | + canon_hash_string (ASM_OPERANDS_INPUT_CONSTRAINT | |
2297 | (x, i))); | |
2298 | } | |
2299 | ||
2300 | hash += canon_hash_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0)); | |
2301 | x = ASM_OPERANDS_INPUT (x, 0); | |
2302 | mode = GET_MODE (x); | |
2303 | goto repeat; | |
2304 | } | |
2305 | ||
2306 | return hash; | |
2307 | } | |
e9a25f70 | 2308 | break; |
278a83b2 | 2309 | |
e9a25f70 JL |
2310 | default: |
2311 | break; | |
7afe21cc RK |
2312 | } |
2313 | ||
2314 | i = GET_RTX_LENGTH (code) - 1; | |
2197a88a | 2315 | hash += (unsigned) code + (unsigned) GET_MODE (x); |
7afe21cc RK |
2316 | fmt = GET_RTX_FORMAT (code); |
2317 | for (; i >= 0; i--) | |
2318 | { | |
2319 | if (fmt[i] == 'e') | |
2320 | { | |
2321 | rtx tem = XEXP (x, i); | |
7afe21cc RK |
2322 | |
2323 | /* If we are about to do the last recursive call | |
2324 | needed at this level, change it into iteration. | |
2325 | This function is called enough to be worth it. */ | |
2326 | if (i == 0) | |
2327 | { | |
2328 | x = tem; | |
2329 | goto repeat; | |
2330 | } | |
2331 | hash += canon_hash (tem, 0); | |
2332 | } | |
2333 | else if (fmt[i] == 'E') | |
2334 | for (j = 0; j < XVECLEN (x, i); j++) | |
2335 | hash += canon_hash (XVECEXP (x, i, j), 0); | |
2336 | else if (fmt[i] == 's') | |
6462bb43 | 2337 | hash += canon_hash_string (XSTR (x, i)); |
7afe21cc RK |
2338 | else if (fmt[i] == 'i') |
2339 | { | |
b3694847 | 2340 | unsigned tem = XINT (x, i); |
2197a88a | 2341 | hash += tem; |
7afe21cc | 2342 | } |
8f985ec4 | 2343 | else if (fmt[i] == '0' || fmt[i] == 't') |
ddc356e8 KH |
2344 | /* Unused. */ |
2345 | ; | |
7afe21cc RK |
2346 | else |
2347 | abort (); | |
2348 | } | |
2349 | return hash; | |
2350 | } | |
2351 | ||
2352 | /* Like canon_hash but with no side effects. */ | |
2353 | ||
2197a88a | 2354 | static unsigned |
7080f735 | 2355 | safe_hash (rtx x, enum machine_mode mode) |
7afe21cc RK |
2356 | { |
2357 | int save_do_not_record = do_not_record; | |
2358 | int save_hash_arg_in_memory = hash_arg_in_memory; | |
2197a88a | 2359 | unsigned hash = canon_hash (x, mode); |
7afe21cc | 2360 | hash_arg_in_memory = save_hash_arg_in_memory; |
7afe21cc RK |
2361 | do_not_record = save_do_not_record; |
2362 | return hash; | |
2363 | } | |
2364 | \f | |
2365 | /* Return 1 iff X and Y would canonicalize into the same thing, | |
2366 | without actually constructing the canonicalization of either one. | |
2367 | If VALIDATE is nonzero, | |
2368 | we assume X is an expression being processed from the rtl | |
2369 | and Y was found in the hash table. We check register refs | |
2370 | in Y for being marked as valid. | |
2371 | ||
2372 | If EQUAL_VALUES is nonzero, we allow a register to match a constant value | |
2373 | that is known to be in the register. Ordinarily, we don't allow them | |
2374 | to match, because letting them match would cause unpredictable results | |
2375 | in all the places that search a hash table chain for an equivalent | |
2376 | for a given value. A possible equivalent that has different structure | |
2377 | has its hash code computed from different data. Whether the hash code | |
38e01259 | 2378 | is the same as that of the given value is pure luck. */ |
7afe21cc RK |
2379 | |
2380 | static int | |
7080f735 | 2381 | exp_equiv_p (rtx x, rtx y, int validate, int equal_values) |
7afe21cc | 2382 | { |
b3694847 SS |
2383 | int i, j; |
2384 | enum rtx_code code; | |
2385 | const char *fmt; | |
7afe21cc RK |
2386 | |
2387 | /* Note: it is incorrect to assume an expression is equivalent to itself | |
2388 | if VALIDATE is nonzero. */ | |
2389 | if (x == y && !validate) | |
2390 | return 1; | |
2391 | if (x == 0 || y == 0) | |
2392 | return x == y; | |
2393 | ||
2394 | code = GET_CODE (x); | |
2395 | if (code != GET_CODE (y)) | |
2396 | { | |
2397 | if (!equal_values) | |
2398 | return 0; | |
2399 | ||
2400 | /* If X is a constant and Y is a register or vice versa, they may be | |
2401 | equivalent. We only have to validate if Y is a register. */ | |
f8cfc6aa | 2402 | if (CONSTANT_P (x) && REG_P (y) |
1bb98cec DM |
2403 | && REGNO_QTY_VALID_P (REGNO (y))) |
2404 | { | |
2405 | int y_q = REG_QTY (REGNO (y)); | |
2406 | struct qty_table_elem *y_ent = &qty_table[y_q]; | |
2407 | ||
2408 | if (GET_MODE (y) == y_ent->mode | |
2409 | && rtx_equal_p (x, y_ent->const_rtx) | |
2410 | && (! validate || REG_IN_TABLE (REGNO (y)) == REG_TICK (REGNO (y)))) | |
2411 | return 1; | |
2412 | } | |
7afe21cc RK |
2413 | |
2414 | if (CONSTANT_P (y) && code == REG | |
1bb98cec DM |
2415 | && REGNO_QTY_VALID_P (REGNO (x))) |
2416 | { | |
2417 | int x_q = REG_QTY (REGNO (x)); | |
2418 | struct qty_table_elem *x_ent = &qty_table[x_q]; | |
2419 | ||
2420 | if (GET_MODE (x) == x_ent->mode | |
2421 | && rtx_equal_p (y, x_ent->const_rtx)) | |
2422 | return 1; | |
2423 | } | |
7afe21cc RK |
2424 | |
2425 | return 0; | |
2426 | } | |
2427 | ||
2428 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */ | |
2429 | if (GET_MODE (x) != GET_MODE (y)) | |
2430 | return 0; | |
2431 | ||
2432 | switch (code) | |
2433 | { | |
2434 | case PC: | |
2435 | case CC0: | |
7afe21cc | 2436 | case CONST_INT: |
c13e8210 | 2437 | return x == y; |
7afe21cc RK |
2438 | |
2439 | case LABEL_REF: | |
7afe21cc RK |
2440 | return XEXP (x, 0) == XEXP (y, 0); |
2441 | ||
f54d4924 RK |
2442 | case SYMBOL_REF: |
2443 | return XSTR (x, 0) == XSTR (y, 0); | |
2444 | ||
7afe21cc RK |
2445 | case REG: |
2446 | { | |
770ae6cc RK |
2447 | unsigned int regno = REGNO (y); |
2448 | unsigned int endregno | |
7afe21cc | 2449 | = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1 |
66fd46b6 | 2450 | : hard_regno_nregs[regno][GET_MODE (y)]); |
770ae6cc | 2451 | unsigned int i; |
7afe21cc RK |
2452 | |
2453 | /* If the quantities are not the same, the expressions are not | |
2454 | equivalent. If there are and we are not to validate, they | |
2455 | are equivalent. Otherwise, ensure all regs are up-to-date. */ | |
2456 | ||
30f72379 | 2457 | if (REG_QTY (REGNO (x)) != REG_QTY (regno)) |
7afe21cc RK |
2458 | return 0; |
2459 | ||
2460 | if (! validate) | |
2461 | return 1; | |
2462 | ||
2463 | for (i = regno; i < endregno; i++) | |
30f72379 | 2464 | if (REG_IN_TABLE (i) != REG_TICK (i)) |
7afe21cc RK |
2465 | return 0; |
2466 | ||
2467 | return 1; | |
2468 | } | |
2469 | ||
2470 | /* For commutative operations, check both orders. */ | |
2471 | case PLUS: | |
2472 | case MULT: | |
2473 | case AND: | |
2474 | case IOR: | |
2475 | case XOR: | |
2476 | case NE: | |
2477 | case EQ: | |
2478 | return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0), validate, equal_values) | |
2479 | && exp_equiv_p (XEXP (x, 1), XEXP (y, 1), | |
2480 | validate, equal_values)) | |
2481 | || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1), | |
2482 | validate, equal_values) | |
2483 | && exp_equiv_p (XEXP (x, 1), XEXP (y, 0), | |
2484 | validate, equal_values))); | |
278a83b2 | 2485 | |
6462bb43 AO |
2486 | case ASM_OPERANDS: |
2487 | /* We don't use the generic code below because we want to | |
2488 | disregard filename and line numbers. */ | |
2489 | ||
2490 | /* A volatile asm isn't equivalent to any other. */ | |
2491 | if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y)) | |
2492 | return 0; | |
2493 | ||
2494 | if (GET_MODE (x) != GET_MODE (y) | |
2495 | || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y)) | |
2496 | || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x), | |
2497 | ASM_OPERANDS_OUTPUT_CONSTRAINT (y)) | |
2498 | || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y) | |
2499 | || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y)) | |
2500 | return 0; | |
2501 | ||
2502 | if (ASM_OPERANDS_INPUT_LENGTH (x)) | |
2503 | { | |
2504 | for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) | |
2505 | if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i), | |
2506 | ASM_OPERANDS_INPUT (y, i), | |
2507 | validate, equal_values) | |
2508 | || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i), | |
2509 | ASM_OPERANDS_INPUT_CONSTRAINT (y, i))) | |
2510 | return 0; | |
2511 | } | |
2512 | ||
2513 | return 1; | |
2514 | ||
e9a25f70 JL |
2515 | default: |
2516 | break; | |
7afe21cc RK |
2517 | } |
2518 | ||
2519 | /* Compare the elements. If any pair of corresponding elements | |
2520 | fail to match, return 0 for the whole things. */ | |
2521 | ||
2522 | fmt = GET_RTX_FORMAT (code); | |
2523 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2524 | { | |
906c4e36 | 2525 | switch (fmt[i]) |
7afe21cc | 2526 | { |
906c4e36 | 2527 | case 'e': |
7afe21cc RK |
2528 | if (! exp_equiv_p (XEXP (x, i), XEXP (y, i), validate, equal_values)) |
2529 | return 0; | |
906c4e36 RK |
2530 | break; |
2531 | ||
2532 | case 'E': | |
7afe21cc RK |
2533 | if (XVECLEN (x, i) != XVECLEN (y, i)) |
2534 | return 0; | |
2535 | for (j = 0; j < XVECLEN (x, i); j++) | |
2536 | if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j), | |
2537 | validate, equal_values)) | |
2538 | return 0; | |
906c4e36 RK |
2539 | break; |
2540 | ||
2541 | case 's': | |
7afe21cc RK |
2542 | if (strcmp (XSTR (x, i), XSTR (y, i))) |
2543 | return 0; | |
906c4e36 RK |
2544 | break; |
2545 | ||
2546 | case 'i': | |
7afe21cc RK |
2547 | if (XINT (x, i) != XINT (y, i)) |
2548 | return 0; | |
906c4e36 RK |
2549 | break; |
2550 | ||
2551 | case 'w': | |
2552 | if (XWINT (x, i) != XWINT (y, i)) | |
2553 | return 0; | |
278a83b2 | 2554 | break; |
906c4e36 RK |
2555 | |
2556 | case '0': | |
8f985ec4 | 2557 | case 't': |
906c4e36 RK |
2558 | break; |
2559 | ||
2560 | default: | |
2561 | abort (); | |
7afe21cc | 2562 | } |
278a83b2 | 2563 | } |
906c4e36 | 2564 | |
7afe21cc RK |
2565 | return 1; |
2566 | } | |
2567 | \f | |
9ae8ffe7 JL |
2568 | /* Return 1 if X has a value that can vary even between two |
2569 | executions of the program. 0 means X can be compared reliably | |
2570 | against certain constants or near-constants. */ | |
7afe21cc RK |
2571 | |
2572 | static int | |
7080f735 | 2573 | cse_rtx_varies_p (rtx x, int from_alias) |
7afe21cc RK |
2574 | { |
2575 | /* We need not check for X and the equivalence class being of the same | |
2576 | mode because if X is equivalent to a constant in some mode, it | |
2577 | doesn't vary in any mode. */ | |
2578 | ||
f8cfc6aa | 2579 | if (REG_P (x) |
1bb98cec DM |
2580 | && REGNO_QTY_VALID_P (REGNO (x))) |
2581 | { | |
2582 | int x_q = REG_QTY (REGNO (x)); | |
2583 | struct qty_table_elem *x_ent = &qty_table[x_q]; | |
2584 | ||
2585 | if (GET_MODE (x) == x_ent->mode | |
2586 | && x_ent->const_rtx != NULL_RTX) | |
2587 | return 0; | |
2588 | } | |
7afe21cc | 2589 | |
9ae8ffe7 JL |
2590 | if (GET_CODE (x) == PLUS |
2591 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
f8cfc6aa | 2592 | && REG_P (XEXP (x, 0)) |
1bb98cec DM |
2593 | && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))) |
2594 | { | |
2595 | int x0_q = REG_QTY (REGNO (XEXP (x, 0))); | |
2596 | struct qty_table_elem *x0_ent = &qty_table[x0_q]; | |
2597 | ||
2598 | if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode) | |
2599 | && x0_ent->const_rtx != NULL_RTX) | |
2600 | return 0; | |
2601 | } | |
7afe21cc | 2602 | |
9c6b0bae RK |
2603 | /* This can happen as the result of virtual register instantiation, if |
2604 | the initial constant is too large to be a valid address. This gives | |
2605 | us a three instruction sequence, load large offset into a register, | |
2606 | load fp minus a constant into a register, then a MEM which is the | |
2607 | sum of the two `constant' registers. */ | |
9ae8ffe7 | 2608 | if (GET_CODE (x) == PLUS |
f8cfc6aa JQ |
2609 | && REG_P (XEXP (x, 0)) |
2610 | && REG_P (XEXP (x, 1)) | |
9ae8ffe7 | 2611 | && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))) |
1bb98cec DM |
2612 | && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1)))) |
2613 | { | |
2614 | int x0_q = REG_QTY (REGNO (XEXP (x, 0))); | |
2615 | int x1_q = REG_QTY (REGNO (XEXP (x, 1))); | |
2616 | struct qty_table_elem *x0_ent = &qty_table[x0_q]; | |
2617 | struct qty_table_elem *x1_ent = &qty_table[x1_q]; | |
2618 | ||
2619 | if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode) | |
2620 | && x0_ent->const_rtx != NULL_RTX | |
2621 | && (GET_MODE (XEXP (x, 1)) == x1_ent->mode) | |
2622 | && x1_ent->const_rtx != NULL_RTX) | |
2623 | return 0; | |
2624 | } | |
9c6b0bae | 2625 | |
2be28ee2 | 2626 | return rtx_varies_p (x, from_alias); |
7afe21cc RK |
2627 | } |
2628 | \f | |
2629 | /* Canonicalize an expression: | |
2630 | replace each register reference inside it | |
2631 | with the "oldest" equivalent register. | |
2632 | ||
da7d8304 | 2633 | If INSN is nonzero and we are replacing a pseudo with a hard register |
7722328e | 2634 | or vice versa, validate_change is used to ensure that INSN remains valid |
da7d8304 | 2635 | after we make our substitution. The calls are made with IN_GROUP nonzero |
7722328e RK |
2636 | so apply_change_group must be called upon the outermost return from this |
2637 | function (unless INSN is zero). The result of apply_change_group can | |
2638 | generally be discarded since the changes we are making are optional. */ | |
7afe21cc RK |
2639 | |
2640 | static rtx | |
7080f735 | 2641 | canon_reg (rtx x, rtx insn) |
7afe21cc | 2642 | { |
b3694847 SS |
2643 | int i; |
2644 | enum rtx_code code; | |
2645 | const char *fmt; | |
7afe21cc RK |
2646 | |
2647 | if (x == 0) | |
2648 | return x; | |
2649 | ||
2650 | code = GET_CODE (x); | |
2651 | switch (code) | |
2652 | { | |
2653 | case PC: | |
2654 | case CC0: | |
2655 | case CONST: | |
2656 | case CONST_INT: | |
2657 | case CONST_DOUBLE: | |
69ef87e2 | 2658 | case CONST_VECTOR: |
7afe21cc RK |
2659 | case SYMBOL_REF: |
2660 | case LABEL_REF: | |
2661 | case ADDR_VEC: | |
2662 | case ADDR_DIFF_VEC: | |
2663 | return x; | |
2664 | ||
2665 | case REG: | |
2666 | { | |
b3694847 SS |
2667 | int first; |
2668 | int q; | |
2669 | struct qty_table_elem *ent; | |
7afe21cc RK |
2670 | |
2671 | /* Never replace a hard reg, because hard regs can appear | |
2672 | in more than one machine mode, and we must preserve the mode | |
2673 | of each occurrence. Also, some hard regs appear in | |
2674 | MEMs that are shared and mustn't be altered. Don't try to | |
2675 | replace any reg that maps to a reg of class NO_REGS. */ | |
2676 | if (REGNO (x) < FIRST_PSEUDO_REGISTER | |
2677 | || ! REGNO_QTY_VALID_P (REGNO (x))) | |
2678 | return x; | |
2679 | ||
278a83b2 | 2680 | q = REG_QTY (REGNO (x)); |
1bb98cec DM |
2681 | ent = &qty_table[q]; |
2682 | first = ent->first_reg; | |
7afe21cc RK |
2683 | return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first] |
2684 | : REGNO_REG_CLASS (first) == NO_REGS ? x | |
1bb98cec | 2685 | : gen_rtx_REG (ent->mode, first)); |
7afe21cc | 2686 | } |
278a83b2 | 2687 | |
e9a25f70 JL |
2688 | default: |
2689 | break; | |
7afe21cc RK |
2690 | } |
2691 | ||
2692 | fmt = GET_RTX_FORMAT (code); | |
2693 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2694 | { | |
b3694847 | 2695 | int j; |
7afe21cc RK |
2696 | |
2697 | if (fmt[i] == 'e') | |
2698 | { | |
2699 | rtx new = canon_reg (XEXP (x, i), insn); | |
58873255 | 2700 | int insn_code; |
7afe21cc RK |
2701 | |
2702 | /* If replacing pseudo with hard reg or vice versa, ensure the | |
178c39f6 | 2703 | insn remains valid. Likewise if the insn has MATCH_DUPs. */ |
aee9dc31 | 2704 | if (insn != 0 && new != 0 |
f8cfc6aa | 2705 | && REG_P (new) && REG_P (XEXP (x, i)) |
178c39f6 RK |
2706 | && (((REGNO (new) < FIRST_PSEUDO_REGISTER) |
2707 | != (REGNO (XEXP (x, i)) < FIRST_PSEUDO_REGISTER)) | |
58873255 | 2708 | || (insn_code = recog_memoized (insn)) < 0 |
a995e389 | 2709 | || insn_data[insn_code].n_dups > 0)) |
77fa0940 | 2710 | validate_change (insn, &XEXP (x, i), new, 1); |
7afe21cc RK |
2711 | else |
2712 | XEXP (x, i) = new; | |
2713 | } | |
2714 | else if (fmt[i] == 'E') | |
2715 | for (j = 0; j < XVECLEN (x, i); j++) | |
2716 | XVECEXP (x, i, j) = canon_reg (XVECEXP (x, i, j), insn); | |
2717 | } | |
2718 | ||
2719 | return x; | |
2720 | } | |
2721 | \f | |
a2cabb29 | 2722 | /* LOC is a location within INSN that is an operand address (the contents of |
7afe21cc RK |
2723 | a MEM). Find the best equivalent address to use that is valid for this |
2724 | insn. | |
2725 | ||
2726 | On most CISC machines, complicated address modes are costly, and rtx_cost | |
2727 | is a good approximation for that cost. However, most RISC machines have | |
2728 | only a few (usually only one) memory reference formats. If an address is | |
2729 | valid at all, it is often just as cheap as any other address. Hence, for | |
e37135f7 RH |
2730 | RISC machines, we use `address_cost' to compare the costs of various |
2731 | addresses. For two addresses of equal cost, choose the one with the | |
2732 | highest `rtx_cost' value as that has the potential of eliminating the | |
2733 | most insns. For equal costs, we choose the first in the equivalence | |
2734 | class. Note that we ignore the fact that pseudo registers are cheaper than | |
2735 | hard registers here because we would also prefer the pseudo registers. */ | |
7afe21cc | 2736 | |
6cd4575e | 2737 | static void |
7080f735 | 2738 | find_best_addr (rtx insn, rtx *loc, enum machine_mode mode) |
7afe21cc | 2739 | { |
7a87758d | 2740 | struct table_elt *elt; |
7afe21cc | 2741 | rtx addr = *loc; |
7a87758d | 2742 | struct table_elt *p; |
7afe21cc RK |
2743 | int found_better = 1; |
2744 | int save_do_not_record = do_not_record; | |
2745 | int save_hash_arg_in_memory = hash_arg_in_memory; | |
7afe21cc RK |
2746 | int addr_volatile; |
2747 | int regno; | |
2197a88a | 2748 | unsigned hash; |
7afe21cc RK |
2749 | |
2750 | /* Do not try to replace constant addresses or addresses of local and | |
2751 | argument slots. These MEM expressions are made only once and inserted | |
2752 | in many instructions, as well as being used to control symbol table | |
2753 | output. It is not safe to clobber them. | |
2754 | ||
2755 | There are some uncommon cases where the address is already in a register | |
2756 | for some reason, but we cannot take advantage of that because we have | |
2757 | no easy way to unshare the MEM. In addition, looking up all stack | |
2758 | addresses is costly. */ | |
2759 | if ((GET_CODE (addr) == PLUS | |
f8cfc6aa | 2760 | && REG_P (XEXP (addr, 0)) |
7afe21cc RK |
2761 | && GET_CODE (XEXP (addr, 1)) == CONST_INT |
2762 | && (regno = REGNO (XEXP (addr, 0)), | |
8bc169f2 DE |
2763 | regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM |
2764 | || regno == ARG_POINTER_REGNUM)) | |
f8cfc6aa | 2765 | || (REG_P (addr) |
8bc169f2 DE |
2766 | && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM |
2767 | || regno == HARD_FRAME_POINTER_REGNUM | |
2768 | || regno == ARG_POINTER_REGNUM)) | |
7afe21cc RK |
2769 | || CONSTANT_ADDRESS_P (addr)) |
2770 | return; | |
2771 | ||
2772 | /* If this address is not simply a register, try to fold it. This will | |
2773 | sometimes simplify the expression. Many simplifications | |
2774 | will not be valid, but some, usually applying the associative rule, will | |
2775 | be valid and produce better code. */ | |
f8cfc6aa | 2776 | if (!REG_P (addr)) |
8c87f107 RK |
2777 | { |
2778 | rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX); | |
630c79be BS |
2779 | int addr_folded_cost = address_cost (folded, mode); |
2780 | int addr_cost = address_cost (addr, mode); | |
2781 | ||
2782 | if ((addr_folded_cost < addr_cost | |
2783 | || (addr_folded_cost == addr_cost | |
2784 | /* ??? The rtx_cost comparison is left over from an older | |
2785 | version of this code. It is probably no longer helpful. */ | |
2786 | && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM) | |
2787 | || approx_reg_cost (folded) < approx_reg_cost (addr)))) | |
8c87f107 RK |
2788 | && validate_change (insn, loc, folded, 0)) |
2789 | addr = folded; | |
2790 | } | |
278a83b2 | 2791 | |
42495ca0 RK |
2792 | /* If this address is not in the hash table, we can't look for equivalences |
2793 | of the whole address. Also, ignore if volatile. */ | |
2794 | ||
7afe21cc | 2795 | do_not_record = 0; |
2197a88a | 2796 | hash = HASH (addr, Pmode); |
7afe21cc RK |
2797 | addr_volatile = do_not_record; |
2798 | do_not_record = save_do_not_record; | |
2799 | hash_arg_in_memory = save_hash_arg_in_memory; | |
7afe21cc RK |
2800 | |
2801 | if (addr_volatile) | |
2802 | return; | |
2803 | ||
2197a88a | 2804 | elt = lookup (addr, hash, Pmode); |
7afe21cc | 2805 | |
42495ca0 RK |
2806 | if (elt) |
2807 | { | |
2808 | /* We need to find the best (under the criteria documented above) entry | |
2809 | in the class that is valid. We use the `flag' field to indicate | |
2810 | choices that were invalid and iterate until we can't find a better | |
2811 | one that hasn't already been tried. */ | |
7afe21cc | 2812 | |
42495ca0 RK |
2813 | for (p = elt->first_same_value; p; p = p->next_same_value) |
2814 | p->flag = 0; | |
7afe21cc | 2815 | |
42495ca0 RK |
2816 | while (found_better) |
2817 | { | |
01329426 | 2818 | int best_addr_cost = address_cost (*loc, mode); |
42495ca0 | 2819 | int best_rtx_cost = (elt->cost + 1) >> 1; |
01329426 | 2820 | int exp_cost; |
278a83b2 | 2821 | struct table_elt *best_elt = elt; |
42495ca0 RK |
2822 | |
2823 | found_better = 0; | |
2824 | for (p = elt->first_same_value; p; p = p->next_same_value) | |
2f541799 | 2825 | if (! p->flag) |
42495ca0 | 2826 | { |
f8cfc6aa | 2827 | if ((REG_P (p->exp) |
2f541799 | 2828 | || exp_equiv_p (p->exp, p->exp, 1, 0)) |
01329426 JH |
2829 | && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost |
2830 | || (exp_cost == best_addr_cost | |
05bd3d41 | 2831 | && ((p->cost + 1) >> 1) > best_rtx_cost))) |
2f541799 MM |
2832 | { |
2833 | found_better = 1; | |
01329426 | 2834 | best_addr_cost = exp_cost; |
2f541799 MM |
2835 | best_rtx_cost = (p->cost + 1) >> 1; |
2836 | best_elt = p; | |
2837 | } | |
42495ca0 | 2838 | } |
7afe21cc | 2839 | |
42495ca0 RK |
2840 | if (found_better) |
2841 | { | |
2842 | if (validate_change (insn, loc, | |
906c4e36 RK |
2843 | canon_reg (copy_rtx (best_elt->exp), |
2844 | NULL_RTX), 0)) | |
42495ca0 RK |
2845 | return; |
2846 | else | |
2847 | best_elt->flag = 1; | |
2848 | } | |
2849 | } | |
2850 | } | |
7afe21cc | 2851 | |
42495ca0 RK |
2852 | /* If the address is a binary operation with the first operand a register |
2853 | and the second a constant, do the same as above, but looking for | |
2854 | equivalences of the register. Then try to simplify before checking for | |
2855 | the best address to use. This catches a few cases: First is when we | |
2856 | have REG+const and the register is another REG+const. We can often merge | |
2857 | the constants and eliminate one insn and one register. It may also be | |
2858 | that a machine has a cheap REG+REG+const. Finally, this improves the | |
2859 | code on the Alpha for unaligned byte stores. */ | |
2860 | ||
2861 | if (flag_expensive_optimizations | |
ec8e098d | 2862 | && ARITHMETIC_P (*loc) |
f8cfc6aa | 2863 | && REG_P (XEXP (*loc, 0))) |
7afe21cc | 2864 | { |
7b9c108f | 2865 | rtx op1 = XEXP (*loc, 1); |
42495ca0 RK |
2866 | |
2867 | do_not_record = 0; | |
2197a88a | 2868 | hash = HASH (XEXP (*loc, 0), Pmode); |
42495ca0 RK |
2869 | do_not_record = save_do_not_record; |
2870 | hash_arg_in_memory = save_hash_arg_in_memory; | |
42495ca0 | 2871 | |
2197a88a | 2872 | elt = lookup (XEXP (*loc, 0), hash, Pmode); |
42495ca0 RK |
2873 | if (elt == 0) |
2874 | return; | |
2875 | ||
2876 | /* We need to find the best (under the criteria documented above) entry | |
2877 | in the class that is valid. We use the `flag' field to indicate | |
2878 | choices that were invalid and iterate until we can't find a better | |
2879 | one that hasn't already been tried. */ | |
7afe21cc | 2880 | |
7afe21cc | 2881 | for (p = elt->first_same_value; p; p = p->next_same_value) |
42495ca0 | 2882 | p->flag = 0; |
7afe21cc | 2883 | |
42495ca0 | 2884 | while (found_better) |
7afe21cc | 2885 | { |
01329426 | 2886 | int best_addr_cost = address_cost (*loc, mode); |
42495ca0 | 2887 | int best_rtx_cost = (COST (*loc) + 1) >> 1; |
278a83b2 | 2888 | struct table_elt *best_elt = elt; |
42495ca0 | 2889 | rtx best_rtx = *loc; |
f6516aee JW |
2890 | int count; |
2891 | ||
2892 | /* This is at worst case an O(n^2) algorithm, so limit our search | |
2893 | to the first 32 elements on the list. This avoids trouble | |
2894 | compiling code with very long basic blocks that can easily | |
0cedb36c JL |
2895 | call simplify_gen_binary so many times that we run out of |
2896 | memory. */ | |
96b0e481 | 2897 | |
0cedb36c JL |
2898 | found_better = 0; |
2899 | for (p = elt->first_same_value, count = 0; | |
2900 | p && count < 32; | |
2901 | p = p->next_same_value, count++) | |
2902 | if (! p->flag | |
f8cfc6aa | 2903 | && (REG_P (p->exp) |
0cedb36c JL |
2904 | || exp_equiv_p (p->exp, p->exp, 1, 0))) |
2905 | { | |
2906 | rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode, | |
7b9c108f | 2907 | p->exp, op1); |
01329426 JH |
2908 | int new_cost; |
2909 | new_cost = address_cost (new, mode); | |
96b0e481 | 2910 | |
01329426 JH |
2911 | if (new_cost < best_addr_cost |
2912 | || (new_cost == best_addr_cost | |
2913 | && (COST (new) + 1) >> 1 > best_rtx_cost)) | |
0cedb36c JL |
2914 | { |
2915 | found_better = 1; | |
01329426 | 2916 | best_addr_cost = new_cost; |
0cedb36c JL |
2917 | best_rtx_cost = (COST (new) + 1) >> 1; |
2918 | best_elt = p; | |
2919 | best_rtx = new; | |
2920 | } | |
2921 | } | |
96b0e481 | 2922 | |
0cedb36c JL |
2923 | if (found_better) |
2924 | { | |
2925 | if (validate_change (insn, loc, | |
2926 | canon_reg (copy_rtx (best_rtx), | |
2927 | NULL_RTX), 0)) | |
2928 | return; | |
2929 | else | |
2930 | best_elt->flag = 1; | |
2931 | } | |
2932 | } | |
2933 | } | |
96b0e481 RK |
2934 | } |
2935 | \f | |
bca05d20 RK |
2936 | /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison |
2937 | operation (EQ, NE, GT, etc.), follow it back through the hash table and | |
2938 | what values are being compared. | |
1a87eea2 | 2939 | |
bca05d20 RK |
2940 | *PARG1 and *PARG2 are updated to contain the rtx representing the values |
2941 | actually being compared. For example, if *PARG1 was (cc0) and *PARG2 | |
2942 | was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were | |
2943 | compared to produce cc0. | |
a432f20d | 2944 | |
bca05d20 RK |
2945 | The return value is the comparison operator and is either the code of |
2946 | A or the code corresponding to the inverse of the comparison. */ | |
7afe21cc | 2947 | |
0cedb36c | 2948 | static enum rtx_code |
7080f735 AJ |
2949 | find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2, |
2950 | enum machine_mode *pmode1, enum machine_mode *pmode2) | |
7afe21cc | 2951 | { |
0cedb36c | 2952 | rtx arg1, arg2; |
1a87eea2 | 2953 | |
0cedb36c | 2954 | arg1 = *parg1, arg2 = *parg2; |
7afe21cc | 2955 | |
0cedb36c | 2956 | /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */ |
7afe21cc | 2957 | |
0cedb36c | 2958 | while (arg2 == CONST0_RTX (GET_MODE (arg1))) |
a432f20d | 2959 | { |
da7d8304 | 2960 | /* Set nonzero when we find something of interest. */ |
0cedb36c JL |
2961 | rtx x = 0; |
2962 | int reverse_code = 0; | |
2963 | struct table_elt *p = 0; | |
6076248a | 2964 | |
0cedb36c JL |
2965 | /* If arg1 is a COMPARE, extract the comparison arguments from it. |
2966 | On machines with CC0, this is the only case that can occur, since | |
2967 | fold_rtx will return the COMPARE or item being compared with zero | |
2968 | when given CC0. */ | |
6076248a | 2969 | |
0cedb36c JL |
2970 | if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx) |
2971 | x = arg1; | |
6076248a | 2972 | |
0cedb36c JL |
2973 | /* If ARG1 is a comparison operator and CODE is testing for |
2974 | STORE_FLAG_VALUE, get the inner arguments. */ | |
a432f20d | 2975 | |
ec8e098d | 2976 | else if (COMPARISON_P (arg1)) |
7afe21cc | 2977 | { |
efdc7e19 RH |
2978 | #ifdef FLOAT_STORE_FLAG_VALUE |
2979 | REAL_VALUE_TYPE fsfv; | |
2980 | #endif | |
2981 | ||
0cedb36c JL |
2982 | if (code == NE |
2983 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT | |
2984 | && code == LT && STORE_FLAG_VALUE == -1) | |
2985 | #ifdef FLOAT_STORE_FLAG_VALUE | |
2986 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT | |
efdc7e19 RH |
2987 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
2988 | REAL_VALUE_NEGATIVE (fsfv))) | |
7afe21cc | 2989 | #endif |
a432f20d | 2990 | ) |
0cedb36c JL |
2991 | x = arg1; |
2992 | else if (code == EQ | |
2993 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT | |
2994 | && code == GE && STORE_FLAG_VALUE == -1) | |
2995 | #ifdef FLOAT_STORE_FLAG_VALUE | |
2996 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT | |
efdc7e19 RH |
2997 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
2998 | REAL_VALUE_NEGATIVE (fsfv))) | |
0cedb36c JL |
2999 | #endif |
3000 | ) | |
3001 | x = arg1, reverse_code = 1; | |
7afe21cc RK |
3002 | } |
3003 | ||
0cedb36c | 3004 | /* ??? We could also check for |
7afe21cc | 3005 | |
0cedb36c | 3006 | (ne (and (eq (...) (const_int 1))) (const_int 0)) |
7afe21cc | 3007 | |
0cedb36c | 3008 | and related forms, but let's wait until we see them occurring. */ |
7afe21cc | 3009 | |
0cedb36c JL |
3010 | if (x == 0) |
3011 | /* Look up ARG1 in the hash table and see if it has an equivalence | |
3012 | that lets us see what is being compared. */ | |
9b1549b8 | 3013 | p = lookup (arg1, safe_hash (arg1, GET_MODE (arg1)) & HASH_MASK, |
0cedb36c | 3014 | GET_MODE (arg1)); |
278a83b2 | 3015 | if (p) |
8b03b984 R |
3016 | { |
3017 | p = p->first_same_value; | |
3018 | ||
3019 | /* If what we compare is already known to be constant, that is as | |
3020 | good as it gets. | |
3021 | We need to break the loop in this case, because otherwise we | |
3022 | can have an infinite loop when looking at a reg that is known | |
3023 | to be a constant which is the same as a comparison of a reg | |
3024 | against zero which appears later in the insn stream, which in | |
3025 | turn is constant and the same as the comparison of the first reg | |
3026 | against zero... */ | |
3027 | if (p->is_const) | |
3028 | break; | |
3029 | } | |
7afe21cc | 3030 | |
0cedb36c | 3031 | for (; p; p = p->next_same_value) |
7afe21cc | 3032 | { |
0cedb36c | 3033 | enum machine_mode inner_mode = GET_MODE (p->exp); |
efdc7e19 RH |
3034 | #ifdef FLOAT_STORE_FLAG_VALUE |
3035 | REAL_VALUE_TYPE fsfv; | |
3036 | #endif | |
7afe21cc | 3037 | |
0cedb36c JL |
3038 | /* If the entry isn't valid, skip it. */ |
3039 | if (! exp_equiv_p (p->exp, p->exp, 1, 0)) | |
3040 | continue; | |
f76b9db2 | 3041 | |
bca05d20 RK |
3042 | if (GET_CODE (p->exp) == COMPARE |
3043 | /* Another possibility is that this machine has a compare insn | |
3044 | that includes the comparison code. In that case, ARG1 would | |
3045 | be equivalent to a comparison operation that would set ARG1 to | |
3046 | either STORE_FLAG_VALUE or zero. If this is an NE operation, | |
3047 | ORIG_CODE is the actual comparison being done; if it is an EQ, | |
3048 | we must reverse ORIG_CODE. On machine with a negative value | |
3049 | for STORE_FLAG_VALUE, also look at LT and GE operations. */ | |
3050 | || ((code == NE | |
3051 | || (code == LT | |
3052 | && GET_MODE_CLASS (inner_mode) == MODE_INT | |
3053 | && (GET_MODE_BITSIZE (inner_mode) | |
3054 | <= HOST_BITS_PER_WIDE_INT) | |
3055 | && (STORE_FLAG_VALUE | |
3056 | & ((HOST_WIDE_INT) 1 | |
3057 | << (GET_MODE_BITSIZE (inner_mode) - 1)))) | |
0cedb36c | 3058 | #ifdef FLOAT_STORE_FLAG_VALUE |
bca05d20 RK |
3059 | || (code == LT |
3060 | && GET_MODE_CLASS (inner_mode) == MODE_FLOAT | |
efdc7e19 RH |
3061 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
3062 | REAL_VALUE_NEGATIVE (fsfv))) | |
0cedb36c | 3063 | #endif |
bca05d20 | 3064 | ) |
ec8e098d | 3065 | && COMPARISON_P (p->exp))) |
7afe21cc | 3066 | { |
0cedb36c JL |
3067 | x = p->exp; |
3068 | break; | |
3069 | } | |
3070 | else if ((code == EQ | |
3071 | || (code == GE | |
3072 | && GET_MODE_CLASS (inner_mode) == MODE_INT | |
3073 | && (GET_MODE_BITSIZE (inner_mode) | |
3074 | <= HOST_BITS_PER_WIDE_INT) | |
3075 | && (STORE_FLAG_VALUE | |
3076 | & ((HOST_WIDE_INT) 1 | |
3077 | << (GET_MODE_BITSIZE (inner_mode) - 1)))) | |
3078 | #ifdef FLOAT_STORE_FLAG_VALUE | |
3079 | || (code == GE | |
3080 | && GET_MODE_CLASS (inner_mode) == MODE_FLOAT | |
efdc7e19 RH |
3081 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
3082 | REAL_VALUE_NEGATIVE (fsfv))) | |
0cedb36c JL |
3083 | #endif |
3084 | ) | |
ec8e098d | 3085 | && COMPARISON_P (p->exp)) |
0cedb36c JL |
3086 | { |
3087 | reverse_code = 1; | |
3088 | x = p->exp; | |
3089 | break; | |
7afe21cc RK |
3090 | } |
3091 | ||
4977bab6 ZW |
3092 | /* If this non-trapping address, e.g. fp + constant, the |
3093 | equivalent is a better operand since it may let us predict | |
3094 | the value of the comparison. */ | |
3095 | else if (!rtx_addr_can_trap_p (p->exp)) | |
0cedb36c JL |
3096 | { |
3097 | arg1 = p->exp; | |
3098 | continue; | |
3099 | } | |
7afe21cc | 3100 | } |
7afe21cc | 3101 | |
0cedb36c JL |
3102 | /* If we didn't find a useful equivalence for ARG1, we are done. |
3103 | Otherwise, set up for the next iteration. */ | |
3104 | if (x == 0) | |
3105 | break; | |
7afe21cc | 3106 | |
78192b09 RH |
3107 | /* If we need to reverse the comparison, make sure that that is |
3108 | possible -- we can't necessarily infer the value of GE from LT | |
3109 | with floating-point operands. */ | |
0cedb36c | 3110 | if (reverse_code) |
261efdef JH |
3111 | { |
3112 | enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX); | |
3113 | if (reversed == UNKNOWN) | |
3114 | break; | |
68252e27 KH |
3115 | else |
3116 | code = reversed; | |
261efdef | 3117 | } |
ec8e098d | 3118 | else if (COMPARISON_P (x)) |
261efdef JH |
3119 | code = GET_CODE (x); |
3120 | arg1 = XEXP (x, 0), arg2 = XEXP (x, 1); | |
7afe21cc RK |
3121 | } |
3122 | ||
0cedb36c JL |
3123 | /* Return our results. Return the modes from before fold_rtx |
3124 | because fold_rtx might produce const_int, and then it's too late. */ | |
3125 | *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2); | |
3126 | *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0); | |
3127 | ||
3128 | return code; | |
7afe21cc RK |
3129 | } |
3130 | \f | |
3131 | /* If X is a nontrivial arithmetic operation on an argument | |
3132 | for which a constant value can be determined, return | |
3133 | the result of operating on that value, as a constant. | |
3134 | Otherwise, return X, possibly with one or more operands | |
3135 | modified by recursive calls to this function. | |
3136 | ||
e7bb59fa RK |
3137 | If X is a register whose contents are known, we do NOT |
3138 | return those contents here. equiv_constant is called to | |
3139 | perform that task. | |
7afe21cc RK |
3140 | |
3141 | INSN is the insn that we may be modifying. If it is 0, make a copy | |
3142 | of X before modifying it. */ | |
3143 | ||
3144 | static rtx | |
7080f735 | 3145 | fold_rtx (rtx x, rtx insn) |
7afe21cc | 3146 | { |
b3694847 SS |
3147 | enum rtx_code code; |
3148 | enum machine_mode mode; | |
3149 | const char *fmt; | |
3150 | int i; | |
7afe21cc RK |
3151 | rtx new = 0; |
3152 | int copied = 0; | |
3153 | int must_swap = 0; | |
3154 | ||
3155 | /* Folded equivalents of first two operands of X. */ | |
3156 | rtx folded_arg0; | |
3157 | rtx folded_arg1; | |
3158 | ||
3159 | /* Constant equivalents of first three operands of X; | |
3160 | 0 when no such equivalent is known. */ | |
3161 | rtx const_arg0; | |
3162 | rtx const_arg1; | |
3163 | rtx const_arg2; | |
3164 | ||
3165 | /* The mode of the first operand of X. We need this for sign and zero | |
3166 | extends. */ | |
3167 | enum machine_mode mode_arg0; | |
3168 | ||
3169 | if (x == 0) | |
3170 | return x; | |
3171 | ||
3172 | mode = GET_MODE (x); | |
3173 | code = GET_CODE (x); | |
3174 | switch (code) | |
3175 | { | |
3176 | case CONST: | |
3177 | case CONST_INT: | |
3178 | case CONST_DOUBLE: | |
69ef87e2 | 3179 | case CONST_VECTOR: |
7afe21cc RK |
3180 | case SYMBOL_REF: |
3181 | case LABEL_REF: | |
3182 | case REG: | |
3183 | /* No use simplifying an EXPR_LIST | |
3184 | since they are used only for lists of args | |
3185 | in a function call's REG_EQUAL note. */ | |
3186 | case EXPR_LIST: | |
3187 | return x; | |
3188 | ||
3189 | #ifdef HAVE_cc0 | |
3190 | case CC0: | |
3191 | return prev_insn_cc0; | |
3192 | #endif | |
3193 | ||
3194 | case PC: | |
3195 | /* If the next insn is a CODE_LABEL followed by a jump table, | |
3196 | PC's value is a LABEL_REF pointing to that label. That | |
8aeea6e6 | 3197 | lets us fold switch statements on the VAX. */ |
e1233a7d RH |
3198 | { |
3199 | rtx next; | |
7c2aa9d7 | 3200 | if (insn && tablejump_p (insn, &next, NULL)) |
e1233a7d RH |
3201 | return gen_rtx_LABEL_REF (Pmode, next); |
3202 | } | |
7afe21cc RK |
3203 | break; |
3204 | ||
3205 | case SUBREG: | |
c610adec RK |
3206 | /* See if we previously assigned a constant value to this SUBREG. */ |
3207 | if ((new = lookup_as_function (x, CONST_INT)) != 0 | |
3208 | || (new = lookup_as_function (x, CONST_DOUBLE)) != 0) | |
7afe21cc RK |
3209 | return new; |
3210 | ||
4b980e20 RK |
3211 | /* If this is a paradoxical SUBREG, we have no idea what value the |
3212 | extra bits would have. However, if the operand is equivalent | |
3213 | to a SUBREG whose operand is the same as our mode, and all the | |
3214 | modes are within a word, we can just use the inner operand | |
31c85c78 RK |
3215 | because these SUBREGs just say how to treat the register. |
3216 | ||
3217 | Similarly if we find an integer constant. */ | |
4b980e20 | 3218 | |
e5f6a288 | 3219 | if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) |
4b980e20 RK |
3220 | { |
3221 | enum machine_mode imode = GET_MODE (SUBREG_REG (x)); | |
3222 | struct table_elt *elt; | |
3223 | ||
3224 | if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD | |
3225 | && GET_MODE_SIZE (imode) <= UNITS_PER_WORD | |
3226 | && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode), | |
3227 | imode)) != 0) | |
ddc356e8 | 3228 | for (elt = elt->first_same_value; elt; elt = elt->next_same_value) |
31c85c78 RK |
3229 | { |
3230 | if (CONSTANT_P (elt->exp) | |
3231 | && GET_MODE (elt->exp) == VOIDmode) | |
3232 | return elt->exp; | |
3233 | ||
4b980e20 RK |
3234 | if (GET_CODE (elt->exp) == SUBREG |
3235 | && GET_MODE (SUBREG_REG (elt->exp)) == mode | |
906c4e36 | 3236 | && exp_equiv_p (elt->exp, elt->exp, 1, 0)) |
4b980e20 | 3237 | return copy_rtx (SUBREG_REG (elt->exp)); |
1bb98cec | 3238 | } |
4b980e20 RK |
3239 | |
3240 | return x; | |
3241 | } | |
e5f6a288 | 3242 | |
7afe21cc RK |
3243 | /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG. |
3244 | We might be able to if the SUBREG is extracting a single word in an | |
3245 | integral mode or extracting the low part. */ | |
3246 | ||
3247 | folded_arg0 = fold_rtx (SUBREG_REG (x), insn); | |
3248 | const_arg0 = equiv_constant (folded_arg0); | |
3249 | if (const_arg0) | |
3250 | folded_arg0 = const_arg0; | |
3251 | ||
3252 | if (folded_arg0 != SUBREG_REG (x)) | |
3253 | { | |
949c5d62 JH |
3254 | new = simplify_subreg (mode, folded_arg0, |
3255 | GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x)); | |
7afe21cc RK |
3256 | if (new) |
3257 | return new; | |
3258 | } | |
e5f6a288 | 3259 | |
f8cfc6aa | 3260 | if (REG_P (folded_arg0) |
4c442790 | 3261 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0))) |
e5f6a288 RK |
3262 | { |
3263 | struct table_elt *elt; | |
3264 | ||
3265 | /* We can use HASH here since we know that canon_hash won't be | |
3266 | called. */ | |
3267 | elt = lookup (folded_arg0, | |
3268 | HASH (folded_arg0, GET_MODE (folded_arg0)), | |
3269 | GET_MODE (folded_arg0)); | |
3270 | ||
3271 | if (elt) | |
3272 | elt = elt->first_same_value; | |
3273 | ||
4c442790 PB |
3274 | if (subreg_lowpart_p (x)) |
3275 | /* If this is a narrowing SUBREG and our operand is a REG, see | |
3276 | if we can find an equivalence for REG that is an arithmetic | |
3277 | operation in a wider mode where both operands are paradoxical | |
3278 | SUBREGs from objects of our result mode. In that case, we | |
3279 | couldn-t report an equivalent value for that operation, since we | |
3280 | don't know what the extra bits will be. But we can find an | |
3281 | equivalence for this SUBREG by folding that operation in the | |
3282 | narrow mode. This allows us to fold arithmetic in narrow modes | |
3283 | when the machine only supports word-sized arithmetic. | |
3284 | ||
3285 | Also look for a case where we have a SUBREG whose operand | |
3286 | is the same as our result. If both modes are smaller | |
3287 | than a word, we are simply interpreting a register in | |
3288 | different modes and we can use the inner value. */ | |
3289 | ||
3290 | for (; elt; elt = elt->next_same_value) | |
3291 | { | |
3292 | enum rtx_code eltcode = GET_CODE (elt->exp); | |
3293 | ||
3294 | /* Just check for unary and binary operations. */ | |
ec8e098d PB |
3295 | if (UNARY_P (elt->exp) |
3296 | && eltcode != SIGN_EXTEND | |
3297 | && eltcode != ZERO_EXTEND | |
4c442790 PB |
3298 | && GET_CODE (XEXP (elt->exp, 0)) == SUBREG |
3299 | && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode | |
3300 | && (GET_MODE_CLASS (mode) | |
3301 | == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0))))) | |
3302 | { | |
3303 | rtx op0 = SUBREG_REG (XEXP (elt->exp, 0)); | |
e5f6a288 | 3304 | |
f8cfc6aa | 3305 | if (!REG_P (op0) && ! CONSTANT_P (op0)) |
4c442790 | 3306 | op0 = fold_rtx (op0, NULL_RTX); |
e5f6a288 | 3307 | |
e5f6a288 | 3308 | op0 = equiv_constant (op0); |
4c442790 PB |
3309 | if (op0) |
3310 | new = simplify_unary_operation (GET_CODE (elt->exp), mode, | |
3311 | op0, mode); | |
3312 | } | |
ec8e098d | 3313 | else if (ARITHMETIC_P (elt->exp) |
4c442790 PB |
3314 | && eltcode != DIV && eltcode != MOD |
3315 | && eltcode != UDIV && eltcode != UMOD | |
3316 | && eltcode != ASHIFTRT && eltcode != LSHIFTRT | |
3317 | && eltcode != ROTATE && eltcode != ROTATERT | |
3318 | && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG | |
3319 | && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) | |
3320 | == mode)) | |
3321 | || CONSTANT_P (XEXP (elt->exp, 0))) | |
3322 | && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG | |
3323 | && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1))) | |
3324 | == mode)) | |
3325 | || CONSTANT_P (XEXP (elt->exp, 1)))) | |
3326 | { | |
3327 | rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0)); | |
3328 | rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1)); | |
3329 | ||
f8cfc6aa | 3330 | if (op0 && !REG_P (op0) && ! CONSTANT_P (op0)) |
4c442790 PB |
3331 | op0 = fold_rtx (op0, NULL_RTX); |
3332 | ||
3333 | if (op0) | |
3334 | op0 = equiv_constant (op0); | |
3335 | ||
f8cfc6aa | 3336 | if (op1 && !REG_P (op1) && ! CONSTANT_P (op1)) |
4c442790 PB |
3337 | op1 = fold_rtx (op1, NULL_RTX); |
3338 | ||
3339 | if (op1) | |
3340 | op1 = equiv_constant (op1); | |
3341 | ||
3342 | /* If we are looking for the low SImode part of | |
3343 | (ashift:DI c (const_int 32)), it doesn't work | |
3344 | to compute that in SImode, because a 32-bit shift | |
3345 | in SImode is unpredictable. We know the value is 0. */ | |
3346 | if (op0 && op1 | |
3347 | && GET_CODE (elt->exp) == ASHIFT | |
3348 | && GET_CODE (op1) == CONST_INT | |
3349 | && INTVAL (op1) >= GET_MODE_BITSIZE (mode)) | |
3350 | { | |
3351 | if (INTVAL (op1) | |
3352 | < GET_MODE_BITSIZE (GET_MODE (elt->exp))) | |
3353 | /* If the count fits in the inner mode's width, | |
3354 | but exceeds the outer mode's width, | |
3355 | the value will get truncated to 0 | |
3356 | by the subreg. */ | |
3357 | new = CONST0_RTX (mode); | |
3358 | else | |
3359 | /* If the count exceeds even the inner mode's width, | |
76fb0b60 | 3360 | don't fold this expression. */ |
4c442790 PB |
3361 | new = 0; |
3362 | } | |
3363 | else if (op0 && op1) | |
3364 | new = simplify_binary_operation (GET_CODE (elt->exp), mode, op0, op1); | |
3365 | } | |
e5f6a288 | 3366 | |
4c442790 PB |
3367 | else if (GET_CODE (elt->exp) == SUBREG |
3368 | && GET_MODE (SUBREG_REG (elt->exp)) == mode | |
3369 | && (GET_MODE_SIZE (GET_MODE (folded_arg0)) | |
3370 | <= UNITS_PER_WORD) | |
3371 | && exp_equiv_p (elt->exp, elt->exp, 1, 0)) | |
3372 | new = copy_rtx (SUBREG_REG (elt->exp)); | |
4b980e20 | 3373 | |
4c442790 PB |
3374 | if (new) |
3375 | return new; | |
3376 | } | |
3377 | else | |
3378 | /* A SUBREG resulting from a zero extension may fold to zero if | |
3379 | it extracts higher bits than the ZERO_EXTEND's source bits. | |
3380 | FIXME: if combine tried to, er, combine these instructions, | |
3381 | this transformation may be moved to simplify_subreg. */ | |
3382 | for (; elt; elt = elt->next_same_value) | |
3383 | { | |
3384 | if (GET_CODE (elt->exp) == ZERO_EXTEND | |
3385 | && subreg_lsb (x) | |
3386 | >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0)))) | |
3387 | return CONST0_RTX (mode); | |
3388 | } | |
e5f6a288 RK |
3389 | } |
3390 | ||
7afe21cc RK |
3391 | return x; |
3392 | ||
3393 | case NOT: | |
3394 | case NEG: | |
3395 | /* If we have (NOT Y), see if Y is known to be (NOT Z). | |
3396 | If so, (NOT Y) simplifies to Z. Similarly for NEG. */ | |
3397 | new = lookup_as_function (XEXP (x, 0), code); | |
3398 | if (new) | |
3399 | return fold_rtx (copy_rtx (XEXP (new, 0)), insn); | |
3400 | break; | |
13c9910f | 3401 | |
7afe21cc RK |
3402 | case MEM: |
3403 | /* If we are not actually processing an insn, don't try to find the | |
3404 | best address. Not only don't we care, but we could modify the | |
3405 | MEM in an invalid way since we have no insn to validate against. */ | |
3406 | if (insn != 0) | |
01329426 | 3407 | find_best_addr (insn, &XEXP (x, 0), GET_MODE (x)); |
7afe21cc RK |
3408 | |
3409 | { | |
3410 | /* Even if we don't fold in the insn itself, | |
3411 | we can safely do so here, in hopes of getting a constant. */ | |
906c4e36 | 3412 | rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX); |
7afe21cc | 3413 | rtx base = 0; |
906c4e36 | 3414 | HOST_WIDE_INT offset = 0; |
7afe21cc | 3415 | |
f8cfc6aa | 3416 | if (REG_P (addr) |
1bb98cec DM |
3417 | && REGNO_QTY_VALID_P (REGNO (addr))) |
3418 | { | |
3419 | int addr_q = REG_QTY (REGNO (addr)); | |
3420 | struct qty_table_elem *addr_ent = &qty_table[addr_q]; | |
3421 | ||
3422 | if (GET_MODE (addr) == addr_ent->mode | |
3423 | && addr_ent->const_rtx != NULL_RTX) | |
3424 | addr = addr_ent->const_rtx; | |
3425 | } | |
7afe21cc RK |
3426 | |
3427 | /* If address is constant, split it into a base and integer offset. */ | |
3428 | if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF) | |
3429 | base = addr; | |
3430 | else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS | |
3431 | && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT) | |
3432 | { | |
3433 | base = XEXP (XEXP (addr, 0), 0); | |
3434 | offset = INTVAL (XEXP (XEXP (addr, 0), 1)); | |
3435 | } | |
3436 | else if (GET_CODE (addr) == LO_SUM | |
3437 | && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF) | |
3438 | base = XEXP (addr, 1); | |
3439 | ||
3440 | /* If this is a constant pool reference, we can fold it into its | |
3441 | constant to allow better value tracking. */ | |
3442 | if (base && GET_CODE (base) == SYMBOL_REF | |
3443 | && CONSTANT_POOL_ADDRESS_P (base)) | |
3444 | { | |
3445 | rtx constant = get_pool_constant (base); | |
3446 | enum machine_mode const_mode = get_pool_mode (base); | |
3447 | rtx new; | |
3448 | ||
3449 | if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT) | |
dd0ba281 RS |
3450 | { |
3451 | constant_pool_entries_cost = COST (constant); | |
3452 | constant_pool_entries_regcost = approx_reg_cost (constant); | |
3453 | } | |
7afe21cc RK |
3454 | |
3455 | /* If we are loading the full constant, we have an equivalence. */ | |
3456 | if (offset == 0 && mode == const_mode) | |
3457 | return constant; | |
3458 | ||
9faa82d8 | 3459 | /* If this actually isn't a constant (weird!), we can't do |
7afe21cc RK |
3460 | anything. Otherwise, handle the two most common cases: |
3461 | extracting a word from a multi-word constant, and extracting | |
3462 | the low-order bits. Other cases don't seem common enough to | |
3463 | worry about. */ | |
3464 | if (! CONSTANT_P (constant)) | |
3465 | return x; | |
3466 | ||
3467 | if (GET_MODE_CLASS (mode) == MODE_INT | |
3468 | && GET_MODE_SIZE (mode) == UNITS_PER_WORD | |
3469 | && offset % UNITS_PER_WORD == 0 | |
3470 | && (new = operand_subword (constant, | |
3471 | offset / UNITS_PER_WORD, | |
3472 | 0, const_mode)) != 0) | |
3473 | return new; | |
3474 | ||
3475 | if (((BYTES_BIG_ENDIAN | |
3476 | && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1) | |
3477 | || (! BYTES_BIG_ENDIAN && offset == 0)) | |
4de249d9 | 3478 | && (new = gen_lowpart (mode, constant)) != 0) |
7afe21cc RK |
3479 | return new; |
3480 | } | |
3481 | ||
3482 | /* If this is a reference to a label at a known position in a jump | |
3483 | table, we also know its value. */ | |
3484 | if (base && GET_CODE (base) == LABEL_REF) | |
3485 | { | |
3486 | rtx label = XEXP (base, 0); | |
3487 | rtx table_insn = NEXT_INSN (label); | |
278a83b2 | 3488 | |
4b4bf941 | 3489 | if (table_insn && JUMP_P (table_insn) |
7afe21cc RK |
3490 | && GET_CODE (PATTERN (table_insn)) == ADDR_VEC) |
3491 | { | |
3492 | rtx table = PATTERN (table_insn); | |
3493 | ||
3494 | if (offset >= 0 | |
3495 | && (offset / GET_MODE_SIZE (GET_MODE (table)) | |
3496 | < XVECLEN (table, 0))) | |
3497 | return XVECEXP (table, 0, | |
3498 | offset / GET_MODE_SIZE (GET_MODE (table))); | |
3499 | } | |
4b4bf941 | 3500 | if (table_insn && JUMP_P (table_insn) |
7afe21cc RK |
3501 | && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC) |
3502 | { | |
3503 | rtx table = PATTERN (table_insn); | |
3504 | ||
3505 | if (offset >= 0 | |
3506 | && (offset / GET_MODE_SIZE (GET_MODE (table)) | |
3507 | < XVECLEN (table, 1))) | |
3508 | { | |
3509 | offset /= GET_MODE_SIZE (GET_MODE (table)); | |
38a448ca RH |
3510 | new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset), |
3511 | XEXP (table, 0)); | |
7afe21cc RK |
3512 | |
3513 | if (GET_MODE (table) != Pmode) | |
38a448ca | 3514 | new = gen_rtx_TRUNCATE (GET_MODE (table), new); |
7afe21cc | 3515 | |
278a83b2 | 3516 | /* Indicate this is a constant. This isn't a |
67a37737 RK |
3517 | valid form of CONST, but it will only be used |
3518 | to fold the next insns and then discarded, so | |
ac7ef8d5 FS |
3519 | it should be safe. |
3520 | ||
3521 | Note this expression must be explicitly discarded, | |
3522 | by cse_insn, else it may end up in a REG_EQUAL note | |
3523 | and "escape" to cause problems elsewhere. */ | |
38a448ca | 3524 | return gen_rtx_CONST (GET_MODE (new), new); |
7afe21cc RK |
3525 | } |
3526 | } | |
3527 | } | |
3528 | ||
3529 | return x; | |
3530 | } | |
9255709c | 3531 | |
a5e5cf67 RH |
3532 | #ifdef NO_FUNCTION_CSE |
3533 | case CALL: | |
3534 | if (CONSTANT_P (XEXP (XEXP (x, 0), 0))) | |
3535 | return x; | |
3536 | break; | |
3537 | #endif | |
3538 | ||
9255709c | 3539 | case ASM_OPERANDS: |
6462bb43 AO |
3540 | for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) |
3541 | validate_change (insn, &ASM_OPERANDS_INPUT (x, i), | |
3542 | fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0); | |
9255709c | 3543 | break; |
278a83b2 | 3544 | |
e9a25f70 JL |
3545 | default: |
3546 | break; | |
7afe21cc RK |
3547 | } |
3548 | ||
3549 | const_arg0 = 0; | |
3550 | const_arg1 = 0; | |
3551 | const_arg2 = 0; | |
3552 | mode_arg0 = VOIDmode; | |
3553 | ||
3554 | /* Try folding our operands. | |
3555 | Then see which ones have constant values known. */ | |
3556 | ||
3557 | fmt = GET_RTX_FORMAT (code); | |
3558 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3559 | if (fmt[i] == 'e') | |
3560 | { | |
3561 | rtx arg = XEXP (x, i); | |
3562 | rtx folded_arg = arg, const_arg = 0; | |
3563 | enum machine_mode mode_arg = GET_MODE (arg); | |
3564 | rtx cheap_arg, expensive_arg; | |
3565 | rtx replacements[2]; | |
3566 | int j; | |
5b437e0f | 3567 | int old_cost = COST_IN (XEXP (x, i), code); |
7afe21cc RK |
3568 | |
3569 | /* Most arguments are cheap, so handle them specially. */ | |
3570 | switch (GET_CODE (arg)) | |
3571 | { | |
3572 | case REG: | |
3573 | /* This is the same as calling equiv_constant; it is duplicated | |
3574 | here for speed. */ | |
1bb98cec DM |
3575 | if (REGNO_QTY_VALID_P (REGNO (arg))) |
3576 | { | |
3577 | int arg_q = REG_QTY (REGNO (arg)); | |
3578 | struct qty_table_elem *arg_ent = &qty_table[arg_q]; | |
3579 | ||
3580 | if (arg_ent->const_rtx != NULL_RTX | |
f8cfc6aa | 3581 | && !REG_P (arg_ent->const_rtx) |
1bb98cec DM |
3582 | && GET_CODE (arg_ent->const_rtx) != PLUS) |
3583 | const_arg | |
4de249d9 | 3584 | = gen_lowpart (GET_MODE (arg), |
1bb98cec DM |
3585 | arg_ent->const_rtx); |
3586 | } | |
7afe21cc RK |
3587 | break; |
3588 | ||
3589 | case CONST: | |
3590 | case CONST_INT: | |
3591 | case SYMBOL_REF: | |
3592 | case LABEL_REF: | |
3593 | case CONST_DOUBLE: | |
69ef87e2 | 3594 | case CONST_VECTOR: |
7afe21cc RK |
3595 | const_arg = arg; |
3596 | break; | |
3597 | ||
3598 | #ifdef HAVE_cc0 | |
3599 | case CC0: | |
3600 | folded_arg = prev_insn_cc0; | |
3601 | mode_arg = prev_insn_cc0_mode; | |
3602 | const_arg = equiv_constant (folded_arg); | |
3603 | break; | |
3604 | #endif | |
3605 | ||
3606 | default: | |
3607 | folded_arg = fold_rtx (arg, insn); | |
3608 | const_arg = equiv_constant (folded_arg); | |
3609 | } | |
3610 | ||
3611 | /* For the first three operands, see if the operand | |
3612 | is constant or equivalent to a constant. */ | |
3613 | switch (i) | |
3614 | { | |
3615 | case 0: | |
3616 | folded_arg0 = folded_arg; | |
3617 | const_arg0 = const_arg; | |
3618 | mode_arg0 = mode_arg; | |
3619 | break; | |
3620 | case 1: | |
3621 | folded_arg1 = folded_arg; | |
3622 | const_arg1 = const_arg; | |
3623 | break; | |
3624 | case 2: | |
3625 | const_arg2 = const_arg; | |
3626 | break; | |
3627 | } | |
3628 | ||
3629 | /* Pick the least expensive of the folded argument and an | |
3630 | equivalent constant argument. */ | |
3631 | if (const_arg == 0 || const_arg == folded_arg | |
f2fa288f | 3632 | || COST_IN (const_arg, code) > COST_IN (folded_arg, code)) |
7afe21cc RK |
3633 | cheap_arg = folded_arg, expensive_arg = const_arg; |
3634 | else | |
3635 | cheap_arg = const_arg, expensive_arg = folded_arg; | |
3636 | ||
3637 | /* Try to replace the operand with the cheapest of the two | |
3638 | possibilities. If it doesn't work and this is either of the first | |
3639 | two operands of a commutative operation, try swapping them. | |
3640 | If THAT fails, try the more expensive, provided it is cheaper | |
3641 | than what is already there. */ | |
3642 | ||
3643 | if (cheap_arg == XEXP (x, i)) | |
3644 | continue; | |
3645 | ||
3646 | if (insn == 0 && ! copied) | |
3647 | { | |
3648 | x = copy_rtx (x); | |
3649 | copied = 1; | |
3650 | } | |
3651 | ||
f2fa288f RH |
3652 | /* Order the replacements from cheapest to most expensive. */ |
3653 | replacements[0] = cheap_arg; | |
3654 | replacements[1] = expensive_arg; | |
3655 | ||
68252e27 | 3656 | for (j = 0; j < 2 && replacements[j]; j++) |
7afe21cc | 3657 | { |
f2fa288f RH |
3658 | int new_cost = COST_IN (replacements[j], code); |
3659 | ||
3660 | /* Stop if what existed before was cheaper. Prefer constants | |
3661 | in the case of a tie. */ | |
3662 | if (new_cost > old_cost | |
3663 | || (new_cost == old_cost && CONSTANT_P (XEXP (x, i)))) | |
3664 | break; | |
3665 | ||
8cce3d04 RS |
3666 | /* It's not safe to substitute the operand of a conversion |
3667 | operator with a constant, as the conversion's identity | |
3668 | depends upon the mode of it's operand. This optimization | |
3669 | is handled by the call to simplify_unary_operation. */ | |
3670 | if (GET_RTX_CLASS (code) == RTX_UNARY | |
3671 | && GET_MODE (replacements[j]) != mode_arg0 | |
3672 | && (code == ZERO_EXTEND | |
3673 | || code == SIGN_EXTEND | |
3674 | || code == TRUNCATE | |
3675 | || code == FLOAT_TRUNCATE | |
3676 | || code == FLOAT_EXTEND | |
3677 | || code == FLOAT | |
3678 | || code == FIX | |
3679 | || code == UNSIGNED_FLOAT | |
3680 | || code == UNSIGNED_FIX)) | |
3681 | continue; | |
3682 | ||
7afe21cc RK |
3683 | if (validate_change (insn, &XEXP (x, i), replacements[j], 0)) |
3684 | break; | |
3685 | ||
ec8e098d PB |
3686 | if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE |
3687 | || GET_RTX_CLASS (code) == RTX_COMM_ARITH) | |
7afe21cc RK |
3688 | { |
3689 | validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1); | |
3690 | validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1); | |
3691 | ||
3692 | if (apply_change_group ()) | |
3693 | { | |
3694 | /* Swap them back to be invalid so that this loop can | |
3695 | continue and flag them to be swapped back later. */ | |
3696 | rtx tem; | |
3697 | ||
3698 | tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1); | |
3699 | XEXP (x, 1) = tem; | |
3700 | must_swap = 1; | |
3701 | break; | |
3702 | } | |
3703 | } | |
3704 | } | |
3705 | } | |
3706 | ||
2d8b0f3a JL |
3707 | else |
3708 | { | |
3709 | if (fmt[i] == 'E') | |
3710 | /* Don't try to fold inside of a vector of expressions. | |
3711 | Doing nothing is harmless. */ | |
e49a1d2e | 3712 | {;} |
2d8b0f3a | 3713 | } |
7afe21cc RK |
3714 | |
3715 | /* If a commutative operation, place a constant integer as the second | |
3716 | operand unless the first operand is also a constant integer. Otherwise, | |
3717 | place any constant second unless the first operand is also a constant. */ | |
3718 | ||
ec8e098d | 3719 | if (COMMUTATIVE_P (x)) |
7afe21cc | 3720 | { |
c715abdd RS |
3721 | if (must_swap |
3722 | || swap_commutative_operands_p (const_arg0 ? const_arg0 | |
3723 | : XEXP (x, 0), | |
3724 | const_arg1 ? const_arg1 | |
3725 | : XEXP (x, 1))) | |
7afe21cc | 3726 | { |
b3694847 | 3727 | rtx tem = XEXP (x, 0); |
7afe21cc RK |
3728 | |
3729 | if (insn == 0 && ! copied) | |
3730 | { | |
3731 | x = copy_rtx (x); | |
3732 | copied = 1; | |
3733 | } | |
3734 | ||
3735 | validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1); | |
3736 | validate_change (insn, &XEXP (x, 1), tem, 1); | |
3737 | if (apply_change_group ()) | |
3738 | { | |
3739 | tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem; | |
3740 | tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem; | |
3741 | } | |
3742 | } | |
3743 | } | |
3744 | ||
3745 | /* If X is an arithmetic operation, see if we can simplify it. */ | |
3746 | ||
3747 | switch (GET_RTX_CLASS (code)) | |
3748 | { | |
ec8e098d | 3749 | case RTX_UNARY: |
67a37737 RK |
3750 | { |
3751 | int is_const = 0; | |
3752 | ||
3753 | /* We can't simplify extension ops unless we know the | |
3754 | original mode. */ | |
3755 | if ((code == ZERO_EXTEND || code == SIGN_EXTEND) | |
3756 | && mode_arg0 == VOIDmode) | |
3757 | break; | |
3758 | ||
3759 | /* If we had a CONST, strip it off and put it back later if we | |
3760 | fold. */ | |
3761 | if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST) | |
3762 | is_const = 1, const_arg0 = XEXP (const_arg0, 0); | |
3763 | ||
3764 | new = simplify_unary_operation (code, mode, | |
3765 | const_arg0 ? const_arg0 : folded_arg0, | |
3766 | mode_arg0); | |
3767 | if (new != 0 && is_const) | |
38a448ca | 3768 | new = gen_rtx_CONST (mode, new); |
67a37737 | 3769 | } |
7afe21cc | 3770 | break; |
278a83b2 | 3771 | |
ec8e098d PB |
3772 | case RTX_COMPARE: |
3773 | case RTX_COMM_COMPARE: | |
7afe21cc RK |
3774 | /* See what items are actually being compared and set FOLDED_ARG[01] |
3775 | to those values and CODE to the actual comparison code. If any are | |
3776 | constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't | |
3777 | do anything if both operands are already known to be constant. */ | |
3778 | ||
3779 | if (const_arg0 == 0 || const_arg1 == 0) | |
3780 | { | |
3781 | struct table_elt *p0, *p1; | |
d6edb99e | 3782 | rtx true_rtx = const_true_rtx, false_rtx = const0_rtx; |
13c9910f | 3783 | enum machine_mode mode_arg1; |
c610adec RK |
3784 | |
3785 | #ifdef FLOAT_STORE_FLAG_VALUE | |
c7c955ee | 3786 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
c610adec | 3787 | { |
d6edb99e | 3788 | true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE |
68252e27 | 3789 | (FLOAT_STORE_FLAG_VALUE (mode), mode)); |
d6edb99e | 3790 | false_rtx = CONST0_RTX (mode); |
c610adec RK |
3791 | } |
3792 | #endif | |
7afe21cc | 3793 | |
13c9910f RS |
3794 | code = find_comparison_args (code, &folded_arg0, &folded_arg1, |
3795 | &mode_arg0, &mode_arg1); | |
7afe21cc RK |
3796 | const_arg0 = equiv_constant (folded_arg0); |
3797 | const_arg1 = equiv_constant (folded_arg1); | |
3798 | ||
13c9910f RS |
3799 | /* If the mode is VOIDmode or a MODE_CC mode, we don't know |
3800 | what kinds of things are being compared, so we can't do | |
3801 | anything with this comparison. */ | |
7afe21cc RK |
3802 | |
3803 | if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC) | |
3804 | break; | |
3805 | ||
0f41302f MS |
3806 | /* If we do not now have two constants being compared, see |
3807 | if we can nevertheless deduce some things about the | |
3808 | comparison. */ | |
7afe21cc RK |
3809 | if (const_arg0 == 0 || const_arg1 == 0) |
3810 | { | |
4977bab6 ZW |
3811 | /* Some addresses are known to be nonzero. We don't know |
3812 | their sign, but equality comparisons are known. */ | |
7afe21cc | 3813 | if (const_arg1 == const0_rtx |
4977bab6 | 3814 | && nonzero_address_p (folded_arg0)) |
7afe21cc RK |
3815 | { |
3816 | if (code == EQ) | |
d6edb99e | 3817 | return false_rtx; |
7afe21cc | 3818 | else if (code == NE) |
d6edb99e | 3819 | return true_rtx; |
7afe21cc RK |
3820 | } |
3821 | ||
fd13313f JH |
3822 | /* See if the two operands are the same. */ |
3823 | ||
3824 | if (folded_arg0 == folded_arg1 | |
f8cfc6aa JQ |
3825 | || (REG_P (folded_arg0) |
3826 | && REG_P (folded_arg1) | |
fd13313f JH |
3827 | && (REG_QTY (REGNO (folded_arg0)) |
3828 | == REG_QTY (REGNO (folded_arg1)))) | |
3829 | || ((p0 = lookup (folded_arg0, | |
3830 | (safe_hash (folded_arg0, mode_arg0) | |
3831 | & HASH_MASK), mode_arg0)) | |
3832 | && (p1 = lookup (folded_arg1, | |
3833 | (safe_hash (folded_arg1, mode_arg0) | |
3834 | & HASH_MASK), mode_arg0)) | |
3835 | && p0->first_same_value == p1->first_same_value)) | |
3836 | { | |
71925bc0 RS |
3837 | /* Sadly two equal NaNs are not equivalent. */ |
3838 | if (!HONOR_NANS (mode_arg0)) | |
3839 | return ((code == EQ || code == LE || code == GE | |
3840 | || code == LEU || code == GEU || code == UNEQ | |
3841 | || code == UNLE || code == UNGE | |
3842 | || code == ORDERED) | |
3843 | ? true_rtx : false_rtx); | |
3844 | /* Take care for the FP compares we can resolve. */ | |
3845 | if (code == UNEQ || code == UNLE || code == UNGE) | |
3846 | return true_rtx; | |
3847 | if (code == LTGT || code == LT || code == GT) | |
3848 | return false_rtx; | |
fd13313f | 3849 | } |
7afe21cc RK |
3850 | |
3851 | /* If FOLDED_ARG0 is a register, see if the comparison we are | |
3852 | doing now is either the same as we did before or the reverse | |
3853 | (we only check the reverse if not floating-point). */ | |
f8cfc6aa | 3854 | else if (REG_P (folded_arg0)) |
7afe21cc | 3855 | { |
30f72379 | 3856 | int qty = REG_QTY (REGNO (folded_arg0)); |
7afe21cc | 3857 | |
1bb98cec DM |
3858 | if (REGNO_QTY_VALID_P (REGNO (folded_arg0))) |
3859 | { | |
3860 | struct qty_table_elem *ent = &qty_table[qty]; | |
3861 | ||
3862 | if ((comparison_dominates_p (ent->comparison_code, code) | |
1eb8759b RH |
3863 | || (! FLOAT_MODE_P (mode_arg0) |
3864 | && comparison_dominates_p (ent->comparison_code, | |
3865 | reverse_condition (code)))) | |
1bb98cec DM |
3866 | && (rtx_equal_p (ent->comparison_const, folded_arg1) |
3867 | || (const_arg1 | |
3868 | && rtx_equal_p (ent->comparison_const, | |
3869 | const_arg1)) | |
f8cfc6aa | 3870 | || (REG_P (folded_arg1) |
1bb98cec DM |
3871 | && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty)))) |
3872 | return (comparison_dominates_p (ent->comparison_code, code) | |
d6edb99e | 3873 | ? true_rtx : false_rtx); |
1bb98cec | 3874 | } |
7afe21cc RK |
3875 | } |
3876 | } | |
3877 | } | |
3878 | ||
3879 | /* If we are comparing against zero, see if the first operand is | |
3880 | equivalent to an IOR with a constant. If so, we may be able to | |
3881 | determine the result of this comparison. */ | |
3882 | ||
3883 | if (const_arg1 == const0_rtx) | |
3884 | { | |
3885 | rtx y = lookup_as_function (folded_arg0, IOR); | |
3886 | rtx inner_const; | |
3887 | ||
3888 | if (y != 0 | |
3889 | && (inner_const = equiv_constant (XEXP (y, 1))) != 0 | |
3890 | && GET_CODE (inner_const) == CONST_INT | |
3891 | && INTVAL (inner_const) != 0) | |
3892 | { | |
3893 | int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1; | |
906c4e36 RK |
3894 | int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum |
3895 | && (INTVAL (inner_const) | |
3896 | & ((HOST_WIDE_INT) 1 << sign_bitnum))); | |
d6edb99e | 3897 | rtx true_rtx = const_true_rtx, false_rtx = const0_rtx; |
c610adec RK |
3898 | |
3899 | #ifdef FLOAT_STORE_FLAG_VALUE | |
c7c955ee | 3900 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
c610adec | 3901 | { |
d6edb99e | 3902 | true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE |
12530dbe | 3903 | (FLOAT_STORE_FLAG_VALUE (mode), mode)); |
d6edb99e | 3904 | false_rtx = CONST0_RTX (mode); |
c610adec RK |
3905 | } |
3906 | #endif | |
7afe21cc RK |
3907 | |
3908 | switch (code) | |
3909 | { | |
3910 | case EQ: | |
d6edb99e | 3911 | return false_rtx; |
7afe21cc | 3912 | case NE: |
d6edb99e | 3913 | return true_rtx; |
7afe21cc RK |
3914 | case LT: case LE: |
3915 | if (has_sign) | |
d6edb99e | 3916 | return true_rtx; |
7afe21cc RK |
3917 | break; |
3918 | case GT: case GE: | |
3919 | if (has_sign) | |
d6edb99e | 3920 | return false_rtx; |
7afe21cc | 3921 | break; |
e9a25f70 JL |
3922 | default: |
3923 | break; | |
7afe21cc RK |
3924 | } |
3925 | } | |
3926 | } | |
3927 | ||
c6fb08ad PB |
3928 | { |
3929 | rtx op0 = const_arg0 ? const_arg0 : folded_arg0; | |
3930 | rtx op1 = const_arg1 ? const_arg1 : folded_arg1; | |
3931 | new = simplify_relational_operation (code, mode, mode_arg0, op0, op1); | |
3932 | } | |
7afe21cc RK |
3933 | break; |
3934 | ||
ec8e098d PB |
3935 | case RTX_BIN_ARITH: |
3936 | case RTX_COMM_ARITH: | |
7afe21cc RK |
3937 | switch (code) |
3938 | { | |
3939 | case PLUS: | |
3940 | /* If the second operand is a LABEL_REF, see if the first is a MINUS | |
3941 | with that LABEL_REF as its second operand. If so, the result is | |
3942 | the first operand of that MINUS. This handles switches with an | |
3943 | ADDR_DIFF_VEC table. */ | |
3944 | if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF) | |
3945 | { | |
e650cbda RK |
3946 | rtx y |
3947 | = GET_CODE (folded_arg0) == MINUS ? folded_arg0 | |
ddc356e8 | 3948 | : lookup_as_function (folded_arg0, MINUS); |
7afe21cc RK |
3949 | |
3950 | if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF | |
3951 | && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0)) | |
3952 | return XEXP (y, 0); | |
67a37737 RK |
3953 | |
3954 | /* Now try for a CONST of a MINUS like the above. */ | |
e650cbda RK |
3955 | if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0 |
3956 | : lookup_as_function (folded_arg0, CONST))) != 0 | |
67a37737 RK |
3957 | && GET_CODE (XEXP (y, 0)) == MINUS |
3958 | && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF | |
ddc356e8 | 3959 | && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0)) |
67a37737 | 3960 | return XEXP (XEXP (y, 0), 0); |
7afe21cc | 3961 | } |
c2cc0778 | 3962 | |
e650cbda RK |
3963 | /* Likewise if the operands are in the other order. */ |
3964 | if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF) | |
3965 | { | |
3966 | rtx y | |
3967 | = GET_CODE (folded_arg1) == MINUS ? folded_arg1 | |
ddc356e8 | 3968 | : lookup_as_function (folded_arg1, MINUS); |
e650cbda RK |
3969 | |
3970 | if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF | |
3971 | && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0)) | |
3972 | return XEXP (y, 0); | |
3973 | ||
3974 | /* Now try for a CONST of a MINUS like the above. */ | |
3975 | if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1 | |
3976 | : lookup_as_function (folded_arg1, CONST))) != 0 | |
3977 | && GET_CODE (XEXP (y, 0)) == MINUS | |
3978 | && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF | |
ddc356e8 | 3979 | && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0)) |
e650cbda RK |
3980 | return XEXP (XEXP (y, 0), 0); |
3981 | } | |
3982 | ||
c2cc0778 RK |
3983 | /* If second operand is a register equivalent to a negative |
3984 | CONST_INT, see if we can find a register equivalent to the | |
3985 | positive constant. Make a MINUS if so. Don't do this for | |
5d595063 | 3986 | a non-negative constant since we might then alternate between |
a1f300c0 | 3987 | choosing positive and negative constants. Having the positive |
5d595063 RK |
3988 | constant previously-used is the more common case. Be sure |
3989 | the resulting constant is non-negative; if const_arg1 were | |
3990 | the smallest negative number this would overflow: depending | |
3991 | on the mode, this would either just be the same value (and | |
3992 | hence not save anything) or be incorrect. */ | |
3993 | if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT | |
3994 | && INTVAL (const_arg1) < 0 | |
4741f6ad JL |
3995 | /* This used to test |
3996 | ||
ddc356e8 | 3997 | -INTVAL (const_arg1) >= 0 |
4741f6ad JL |
3998 | |
3999 | But The Sun V5.0 compilers mis-compiled that test. So | |
4000 | instead we test for the problematic value in a more direct | |
4001 | manner and hope the Sun compilers get it correct. */ | |
5c45a8ac KG |
4002 | && INTVAL (const_arg1) != |
4003 | ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1)) | |
f8cfc6aa | 4004 | && REG_P (folded_arg1)) |
c2cc0778 | 4005 | { |
ddc356e8 | 4006 | rtx new_const = GEN_INT (-INTVAL (const_arg1)); |
c2cc0778 | 4007 | struct table_elt *p |
9b1549b8 | 4008 | = lookup (new_const, safe_hash (new_const, mode) & HASH_MASK, |
c2cc0778 RK |
4009 | mode); |
4010 | ||
4011 | if (p) | |
4012 | for (p = p->first_same_value; p; p = p->next_same_value) | |
f8cfc6aa | 4013 | if (REG_P (p->exp)) |
0cedb36c JL |
4014 | return simplify_gen_binary (MINUS, mode, folded_arg0, |
4015 | canon_reg (p->exp, NULL_RTX)); | |
c2cc0778 | 4016 | } |
13c9910f RS |
4017 | goto from_plus; |
4018 | ||
4019 | case MINUS: | |
4020 | /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2). | |
4021 | If so, produce (PLUS Z C2-C). */ | |
4022 | if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT) | |
4023 | { | |
4024 | rtx y = lookup_as_function (XEXP (x, 0), PLUS); | |
4025 | if (y && GET_CODE (XEXP (y, 1)) == CONST_INT) | |
f3becefd RK |
4026 | return fold_rtx (plus_constant (copy_rtx (y), |
4027 | -INTVAL (const_arg1)), | |
a3b5c94a | 4028 | NULL_RTX); |
13c9910f | 4029 | } |
7afe21cc | 4030 | |
ddc356e8 | 4031 | /* Fall through. */ |
7afe21cc | 4032 | |
13c9910f | 4033 | from_plus: |
7afe21cc RK |
4034 | case SMIN: case SMAX: case UMIN: case UMAX: |
4035 | case IOR: case AND: case XOR: | |
f930bfd0 | 4036 | case MULT: |
7afe21cc RK |
4037 | case ASHIFT: case LSHIFTRT: case ASHIFTRT: |
4038 | /* If we have (<op> <reg> <const_int>) for an associative OP and REG | |
4039 | is known to be of similar form, we may be able to replace the | |
4040 | operation with a combined operation. This may eliminate the | |
4041 | intermediate operation if every use is simplified in this way. | |
4042 | Note that the similar optimization done by combine.c only works | |
4043 | if the intermediate operation's result has only one reference. */ | |
4044 | ||
f8cfc6aa | 4045 | if (REG_P (folded_arg0) |
7afe21cc RK |
4046 | && const_arg1 && GET_CODE (const_arg1) == CONST_INT) |
4047 | { | |
4048 | int is_shift | |
4049 | = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT); | |
4050 | rtx y = lookup_as_function (folded_arg0, code); | |
4051 | rtx inner_const; | |
4052 | enum rtx_code associate_code; | |
4053 | rtx new_const; | |
4054 | ||
4055 | if (y == 0 | |
4056 | || 0 == (inner_const | |
4057 | = equiv_constant (fold_rtx (XEXP (y, 1), 0))) | |
4058 | || GET_CODE (inner_const) != CONST_INT | |
4059 | /* If we have compiled a statement like | |
4060 | "if (x == (x & mask1))", and now are looking at | |
4061 | "x & mask2", we will have a case where the first operand | |
4062 | of Y is the same as our first operand. Unless we detect | |
4063 | this case, an infinite loop will result. */ | |
4064 | || XEXP (y, 0) == folded_arg0) | |
4065 | break; | |
4066 | ||
4067 | /* Don't associate these operations if they are a PLUS with the | |
4068 | same constant and it is a power of two. These might be doable | |
4069 | with a pre- or post-increment. Similarly for two subtracts of | |
4070 | identical powers of two with post decrement. */ | |
4071 | ||
213d5fbc | 4072 | if (code == PLUS && const_arg1 == inner_const |
940da324 JL |
4073 | && ((HAVE_PRE_INCREMENT |
4074 | && exact_log2 (INTVAL (const_arg1)) >= 0) | |
4075 | || (HAVE_POST_INCREMENT | |
4076 | && exact_log2 (INTVAL (const_arg1)) >= 0) | |
4077 | || (HAVE_PRE_DECREMENT | |
4078 | && exact_log2 (- INTVAL (const_arg1)) >= 0) | |
4079 | || (HAVE_POST_DECREMENT | |
4080 | && exact_log2 (- INTVAL (const_arg1)) >= 0))) | |
7afe21cc RK |
4081 | break; |
4082 | ||
4083 | /* Compute the code used to compose the constants. For example, | |
f930bfd0 | 4084 | A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */ |
7afe21cc | 4085 | |
f930bfd0 | 4086 | associate_code = (is_shift || code == MINUS ? PLUS : code); |
7afe21cc RK |
4087 | |
4088 | new_const = simplify_binary_operation (associate_code, mode, | |
4089 | const_arg1, inner_const); | |
4090 | ||
4091 | if (new_const == 0) | |
4092 | break; | |
4093 | ||
4094 | /* If we are associating shift operations, don't let this | |
4908e508 RS |
4095 | produce a shift of the size of the object or larger. |
4096 | This could occur when we follow a sign-extend by a right | |
4097 | shift on a machine that does a sign-extend as a pair | |
4098 | of shifts. */ | |
7afe21cc RK |
4099 | |
4100 | if (is_shift && GET_CODE (new_const) == CONST_INT | |
4908e508 RS |
4101 | && INTVAL (new_const) >= GET_MODE_BITSIZE (mode)) |
4102 | { | |
4103 | /* As an exception, we can turn an ASHIFTRT of this | |
4104 | form into a shift of the number of bits - 1. */ | |
4105 | if (code == ASHIFTRT) | |
4106 | new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1); | |
4107 | else | |
4108 | break; | |
4109 | } | |
7afe21cc RK |
4110 | |
4111 | y = copy_rtx (XEXP (y, 0)); | |
4112 | ||
4113 | /* If Y contains our first operand (the most common way this | |
4114 | can happen is if Y is a MEM), we would do into an infinite | |
4115 | loop if we tried to fold it. So don't in that case. */ | |
4116 | ||
4117 | if (! reg_mentioned_p (folded_arg0, y)) | |
4118 | y = fold_rtx (y, insn); | |
4119 | ||
0cedb36c | 4120 | return simplify_gen_binary (code, mode, y, new_const); |
7afe21cc | 4121 | } |
e9a25f70 JL |
4122 | break; |
4123 | ||
f930bfd0 JW |
4124 | case DIV: case UDIV: |
4125 | /* ??? The associative optimization performed immediately above is | |
4126 | also possible for DIV and UDIV using associate_code of MULT. | |
4127 | However, we would need extra code to verify that the | |
4128 | multiplication does not overflow, that is, there is no overflow | |
4129 | in the calculation of new_const. */ | |
4130 | break; | |
4131 | ||
e9a25f70 JL |
4132 | default: |
4133 | break; | |
7afe21cc RK |
4134 | } |
4135 | ||
4136 | new = simplify_binary_operation (code, mode, | |
4137 | const_arg0 ? const_arg0 : folded_arg0, | |
4138 | const_arg1 ? const_arg1 : folded_arg1); | |
4139 | break; | |
4140 | ||
ec8e098d | 4141 | case RTX_OBJ: |
7afe21cc RK |
4142 | /* (lo_sum (high X) X) is simply X. */ |
4143 | if (code == LO_SUM && const_arg0 != 0 | |
4144 | && GET_CODE (const_arg0) == HIGH | |
4145 | && rtx_equal_p (XEXP (const_arg0, 0), const_arg1)) | |
4146 | return const_arg1; | |
4147 | break; | |
4148 | ||
ec8e098d PB |
4149 | case RTX_TERNARY: |
4150 | case RTX_BITFIELD_OPS: | |
7afe21cc RK |
4151 | new = simplify_ternary_operation (code, mode, mode_arg0, |
4152 | const_arg0 ? const_arg0 : folded_arg0, | |
4153 | const_arg1 ? const_arg1 : folded_arg1, | |
4154 | const_arg2 ? const_arg2 : XEXP (x, 2)); | |
4155 | break; | |
ee5332b8 | 4156 | |
ec8e098d PB |
4157 | default: |
4158 | break; | |
7afe21cc RK |
4159 | } |
4160 | ||
4161 | return new ? new : x; | |
4162 | } | |
4163 | \f | |
4164 | /* Return a constant value currently equivalent to X. | |
4165 | Return 0 if we don't know one. */ | |
4166 | ||
4167 | static rtx | |
7080f735 | 4168 | equiv_constant (rtx x) |
7afe21cc | 4169 | { |
f8cfc6aa | 4170 | if (REG_P (x) |
1bb98cec DM |
4171 | && REGNO_QTY_VALID_P (REGNO (x))) |
4172 | { | |
4173 | int x_q = REG_QTY (REGNO (x)); | |
4174 | struct qty_table_elem *x_ent = &qty_table[x_q]; | |
4175 | ||
4176 | if (x_ent->const_rtx) | |
4de249d9 | 4177 | x = gen_lowpart (GET_MODE (x), x_ent->const_rtx); |
1bb98cec | 4178 | } |
7afe21cc | 4179 | |
2ce5e1b4 | 4180 | if (x == 0 || CONSTANT_P (x)) |
7afe21cc RK |
4181 | return x; |
4182 | ||
fc3ffe83 RK |
4183 | /* If X is a MEM, try to fold it outside the context of any insn to see if |
4184 | it might be equivalent to a constant. That handles the case where it | |
4185 | is a constant-pool reference. Then try to look it up in the hash table | |
4186 | in case it is something whose value we have seen before. */ | |
4187 | ||
3c0cb5de | 4188 | if (MEM_P (x)) |
fc3ffe83 RK |
4189 | { |
4190 | struct table_elt *elt; | |
4191 | ||
906c4e36 | 4192 | x = fold_rtx (x, NULL_RTX); |
fc3ffe83 RK |
4193 | if (CONSTANT_P (x)) |
4194 | return x; | |
4195 | ||
9b1549b8 | 4196 | elt = lookup (x, safe_hash (x, GET_MODE (x)) & HASH_MASK, GET_MODE (x)); |
fc3ffe83 RK |
4197 | if (elt == 0) |
4198 | return 0; | |
4199 | ||
4200 | for (elt = elt->first_same_value; elt; elt = elt->next_same_value) | |
4201 | if (elt->is_const && CONSTANT_P (elt->exp)) | |
4202 | return elt->exp; | |
4203 | } | |
4204 | ||
7afe21cc RK |
4205 | return 0; |
4206 | } | |
4207 | \f | |
4208 | /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point | |
4209 | number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the | |
4210 | least-significant part of X. | |
278a83b2 | 4211 | MODE specifies how big a part of X to return. |
7afe21cc RK |
4212 | |
4213 | If the requested operation cannot be done, 0 is returned. | |
4214 | ||
4de249d9 | 4215 | This is similar to gen_lowpart_general in emit-rtl.c. */ |
7afe21cc RK |
4216 | |
4217 | rtx | |
7080f735 | 4218 | gen_lowpart_if_possible (enum machine_mode mode, rtx x) |
7afe21cc RK |
4219 | { |
4220 | rtx result = gen_lowpart_common (mode, x); | |
4221 | ||
4222 | if (result) | |
4223 | return result; | |
3c0cb5de | 4224 | else if (MEM_P (x)) |
7afe21cc RK |
4225 | { |
4226 | /* This is the only other case we handle. */ | |
b3694847 | 4227 | int offset = 0; |
7afe21cc RK |
4228 | rtx new; |
4229 | ||
f76b9db2 ILT |
4230 | if (WORDS_BIG_ENDIAN) |
4231 | offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD) | |
4232 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)); | |
4233 | if (BYTES_BIG_ENDIAN) | |
f1ec5147 RK |
4234 | /* Adjust the address so that the address-after-the-data is |
4235 | unchanged. */ | |
4236 | offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode)) | |
4237 | - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x)))); | |
4238 | ||
4239 | new = adjust_address_nv (x, mode, offset); | |
7afe21cc RK |
4240 | if (! memory_address_p (mode, XEXP (new, 0))) |
4241 | return 0; | |
f1ec5147 | 4242 | |
7afe21cc RK |
4243 | return new; |
4244 | } | |
4245 | else | |
4246 | return 0; | |
4247 | } | |
4248 | \f | |
6de9cd9a | 4249 | /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken" |
7afe21cc RK |
4250 | branch. It will be zero if not. |
4251 | ||
4252 | In certain cases, this can cause us to add an equivalence. For example, | |
278a83b2 | 4253 | if we are following the taken case of |
7080f735 | 4254 | if (i == 2) |
7afe21cc RK |
4255 | we can add the fact that `i' and '2' are now equivalent. |
4256 | ||
4257 | In any case, we can record that this comparison was passed. If the same | |
4258 | comparison is seen later, we will know its value. */ | |
4259 | ||
4260 | static void | |
7080f735 | 4261 | record_jump_equiv (rtx insn, int taken) |
7afe21cc RK |
4262 | { |
4263 | int cond_known_true; | |
4264 | rtx op0, op1; | |
7f1c097d | 4265 | rtx set; |
13c9910f | 4266 | enum machine_mode mode, mode0, mode1; |
7afe21cc RK |
4267 | int reversed_nonequality = 0; |
4268 | enum rtx_code code; | |
4269 | ||
4270 | /* Ensure this is the right kind of insn. */ | |
7f1c097d | 4271 | if (! any_condjump_p (insn)) |
7afe21cc | 4272 | return; |
7f1c097d | 4273 | set = pc_set (insn); |
7afe21cc RK |
4274 | |
4275 | /* See if this jump condition is known true or false. */ | |
4276 | if (taken) | |
7f1c097d | 4277 | cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx); |
7afe21cc | 4278 | else |
7f1c097d | 4279 | cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx); |
7afe21cc RK |
4280 | |
4281 | /* Get the type of comparison being done and the operands being compared. | |
4282 | If we had to reverse a non-equality condition, record that fact so we | |
4283 | know that it isn't valid for floating-point. */ | |
7f1c097d JH |
4284 | code = GET_CODE (XEXP (SET_SRC (set), 0)); |
4285 | op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn); | |
4286 | op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn); | |
7afe21cc | 4287 | |
13c9910f | 4288 | code = find_comparison_args (code, &op0, &op1, &mode0, &mode1); |
7afe21cc RK |
4289 | if (! cond_known_true) |
4290 | { | |
261efdef | 4291 | code = reversed_comparison_code_parts (code, op0, op1, insn); |
1eb8759b RH |
4292 | |
4293 | /* Don't remember if we can't find the inverse. */ | |
4294 | if (code == UNKNOWN) | |
4295 | return; | |
7afe21cc RK |
4296 | } |
4297 | ||
4298 | /* The mode is the mode of the non-constant. */ | |
13c9910f RS |
4299 | mode = mode0; |
4300 | if (mode1 != VOIDmode) | |
4301 | mode = mode1; | |
7afe21cc RK |
4302 | |
4303 | record_jump_cond (code, mode, op0, op1, reversed_nonequality); | |
4304 | } | |
4305 | ||
4306 | /* We know that comparison CODE applied to OP0 and OP1 in MODE is true. | |
4307 | REVERSED_NONEQUALITY is nonzero if CODE had to be swapped. | |
4308 | Make any useful entries we can with that information. Called from | |
4309 | above function and called recursively. */ | |
4310 | ||
4311 | static void | |
7080f735 AJ |
4312 | record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0, |
4313 | rtx op1, int reversed_nonequality) | |
7afe21cc | 4314 | { |
2197a88a | 4315 | unsigned op0_hash, op1_hash; |
e428d738 | 4316 | int op0_in_memory, op1_in_memory; |
7afe21cc RK |
4317 | struct table_elt *op0_elt, *op1_elt; |
4318 | ||
4319 | /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG, | |
4320 | we know that they are also equal in the smaller mode (this is also | |
4321 | true for all smaller modes whether or not there is a SUBREG, but | |
ac7ef8d5 | 4322 | is not worth testing for with no SUBREG). */ |
7afe21cc | 4323 | |
2e794ee8 | 4324 | /* Note that GET_MODE (op0) may not equal MODE. */ |
7afe21cc | 4325 | if (code == EQ && GET_CODE (op0) == SUBREG |
2e794ee8 RS |
4326 | && (GET_MODE_SIZE (GET_MODE (op0)) |
4327 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))) | |
7afe21cc RK |
4328 | { |
4329 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0)); | |
4de249d9 | 4330 | rtx tem = gen_lowpart (inner_mode, op1); |
7afe21cc RK |
4331 | |
4332 | record_jump_cond (code, mode, SUBREG_REG (op0), | |
38a448ca | 4333 | tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0), |
7afe21cc RK |
4334 | reversed_nonequality); |
4335 | } | |
4336 | ||
4337 | if (code == EQ && GET_CODE (op1) == SUBREG | |
2e794ee8 RS |
4338 | && (GET_MODE_SIZE (GET_MODE (op1)) |
4339 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1))))) | |
7afe21cc RK |
4340 | { |
4341 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1)); | |
4de249d9 | 4342 | rtx tem = gen_lowpart (inner_mode, op0); |
7afe21cc RK |
4343 | |
4344 | record_jump_cond (code, mode, SUBREG_REG (op1), | |
38a448ca | 4345 | tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0), |
7afe21cc RK |
4346 | reversed_nonequality); |
4347 | } | |
4348 | ||
278a83b2 | 4349 | /* Similarly, if this is an NE comparison, and either is a SUBREG |
7afe21cc RK |
4350 | making a smaller mode, we know the whole thing is also NE. */ |
4351 | ||
2e794ee8 RS |
4352 | /* Note that GET_MODE (op0) may not equal MODE; |
4353 | if we test MODE instead, we can get an infinite recursion | |
4354 | alternating between two modes each wider than MODE. */ | |
4355 | ||
7afe21cc RK |
4356 | if (code == NE && GET_CODE (op0) == SUBREG |
4357 | && subreg_lowpart_p (op0) | |
2e794ee8 RS |
4358 | && (GET_MODE_SIZE (GET_MODE (op0)) |
4359 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))) | |
7afe21cc RK |
4360 | { |
4361 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0)); | |
4de249d9 | 4362 | rtx tem = gen_lowpart (inner_mode, op1); |
7afe21cc RK |
4363 | |
4364 | record_jump_cond (code, mode, SUBREG_REG (op0), | |
38a448ca | 4365 | tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0), |
7afe21cc RK |
4366 | reversed_nonequality); |
4367 | } | |
4368 | ||
4369 | if (code == NE && GET_CODE (op1) == SUBREG | |
4370 | && subreg_lowpart_p (op1) | |
2e794ee8 RS |
4371 | && (GET_MODE_SIZE (GET_MODE (op1)) |
4372 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1))))) | |
7afe21cc RK |
4373 | { |
4374 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1)); | |
4de249d9 | 4375 | rtx tem = gen_lowpart (inner_mode, op0); |
7afe21cc RK |
4376 | |
4377 | record_jump_cond (code, mode, SUBREG_REG (op1), | |
38a448ca | 4378 | tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0), |
7afe21cc RK |
4379 | reversed_nonequality); |
4380 | } | |
4381 | ||
4382 | /* Hash both operands. */ | |
4383 | ||
4384 | do_not_record = 0; | |
4385 | hash_arg_in_memory = 0; | |
2197a88a | 4386 | op0_hash = HASH (op0, mode); |
7afe21cc | 4387 | op0_in_memory = hash_arg_in_memory; |
7afe21cc RK |
4388 | |
4389 | if (do_not_record) | |
4390 | return; | |
4391 | ||
4392 | do_not_record = 0; | |
4393 | hash_arg_in_memory = 0; | |
2197a88a | 4394 | op1_hash = HASH (op1, mode); |
7afe21cc | 4395 | op1_in_memory = hash_arg_in_memory; |
278a83b2 | 4396 | |
7afe21cc RK |
4397 | if (do_not_record) |
4398 | return; | |
4399 | ||
4400 | /* Look up both operands. */ | |
2197a88a RK |
4401 | op0_elt = lookup (op0, op0_hash, mode); |
4402 | op1_elt = lookup (op1, op1_hash, mode); | |
7afe21cc | 4403 | |
af3869c1 RK |
4404 | /* If both operands are already equivalent or if they are not in the |
4405 | table but are identical, do nothing. */ | |
4406 | if ((op0_elt != 0 && op1_elt != 0 | |
4407 | && op0_elt->first_same_value == op1_elt->first_same_value) | |
4408 | || op0 == op1 || rtx_equal_p (op0, op1)) | |
4409 | return; | |
4410 | ||
7afe21cc | 4411 | /* If we aren't setting two things equal all we can do is save this |
b2796a4b RK |
4412 | comparison. Similarly if this is floating-point. In the latter |
4413 | case, OP1 might be zero and both -0.0 and 0.0 are equal to it. | |
4414 | If we record the equality, we might inadvertently delete code | |
4415 | whose intent was to change -0 to +0. */ | |
4416 | ||
cbf6a543 | 4417 | if (code != EQ || FLOAT_MODE_P (GET_MODE (op0))) |
7afe21cc | 4418 | { |
1bb98cec DM |
4419 | struct qty_table_elem *ent; |
4420 | int qty; | |
4421 | ||
7afe21cc RK |
4422 | /* If we reversed a floating-point comparison, if OP0 is not a |
4423 | register, or if OP1 is neither a register or constant, we can't | |
4424 | do anything. */ | |
4425 | ||
f8cfc6aa | 4426 | if (!REG_P (op1)) |
7afe21cc RK |
4427 | op1 = equiv_constant (op1); |
4428 | ||
cbf6a543 | 4429 | if ((reversed_nonequality && FLOAT_MODE_P (mode)) |
f8cfc6aa | 4430 | || !REG_P (op0) || op1 == 0) |
7afe21cc RK |
4431 | return; |
4432 | ||
4433 | /* Put OP0 in the hash table if it isn't already. This gives it a | |
4434 | new quantity number. */ | |
4435 | if (op0_elt == 0) | |
4436 | { | |
9714cf43 | 4437 | if (insert_regs (op0, NULL, 0)) |
7afe21cc RK |
4438 | { |
4439 | rehash_using_reg (op0); | |
2197a88a | 4440 | op0_hash = HASH (op0, mode); |
2bb81c86 RK |
4441 | |
4442 | /* If OP0 is contained in OP1, this changes its hash code | |
4443 | as well. Faster to rehash than to check, except | |
4444 | for the simple case of a constant. */ | |
4445 | if (! CONSTANT_P (op1)) | |
2197a88a | 4446 | op1_hash = HASH (op1,mode); |
7afe21cc RK |
4447 | } |
4448 | ||
9714cf43 | 4449 | op0_elt = insert (op0, NULL, op0_hash, mode); |
7afe21cc | 4450 | op0_elt->in_memory = op0_in_memory; |
7afe21cc RK |
4451 | } |
4452 | ||
1bb98cec DM |
4453 | qty = REG_QTY (REGNO (op0)); |
4454 | ent = &qty_table[qty]; | |
4455 | ||
4456 | ent->comparison_code = code; | |
f8cfc6aa | 4457 | if (REG_P (op1)) |
7afe21cc | 4458 | { |
5d5ea909 | 4459 | /* Look it up again--in case op0 and op1 are the same. */ |
2197a88a | 4460 | op1_elt = lookup (op1, op1_hash, mode); |
5d5ea909 | 4461 | |
7afe21cc RK |
4462 | /* Put OP1 in the hash table so it gets a new quantity number. */ |
4463 | if (op1_elt == 0) | |
4464 | { | |
9714cf43 | 4465 | if (insert_regs (op1, NULL, 0)) |
7afe21cc RK |
4466 | { |
4467 | rehash_using_reg (op1); | |
2197a88a | 4468 | op1_hash = HASH (op1, mode); |
7afe21cc RK |
4469 | } |
4470 | ||
9714cf43 | 4471 | op1_elt = insert (op1, NULL, op1_hash, mode); |
7afe21cc | 4472 | op1_elt->in_memory = op1_in_memory; |
7afe21cc RK |
4473 | } |
4474 | ||
1bb98cec DM |
4475 | ent->comparison_const = NULL_RTX; |
4476 | ent->comparison_qty = REG_QTY (REGNO (op1)); | |
7afe21cc RK |
4477 | } |
4478 | else | |
4479 | { | |
1bb98cec DM |
4480 | ent->comparison_const = op1; |
4481 | ent->comparison_qty = -1; | |
7afe21cc RK |
4482 | } |
4483 | ||
4484 | return; | |
4485 | } | |
4486 | ||
eb5ad42a RS |
4487 | /* If either side is still missing an equivalence, make it now, |
4488 | then merge the equivalences. */ | |
7afe21cc | 4489 | |
7afe21cc RK |
4490 | if (op0_elt == 0) |
4491 | { | |
9714cf43 | 4492 | if (insert_regs (op0, NULL, 0)) |
7afe21cc RK |
4493 | { |
4494 | rehash_using_reg (op0); | |
2197a88a | 4495 | op0_hash = HASH (op0, mode); |
7afe21cc RK |
4496 | } |
4497 | ||
9714cf43 | 4498 | op0_elt = insert (op0, NULL, op0_hash, mode); |
7afe21cc | 4499 | op0_elt->in_memory = op0_in_memory; |
7afe21cc RK |
4500 | } |
4501 | ||
4502 | if (op1_elt == 0) | |
4503 | { | |
9714cf43 | 4504 | if (insert_regs (op1, NULL, 0)) |
7afe21cc RK |
4505 | { |
4506 | rehash_using_reg (op1); | |
2197a88a | 4507 | op1_hash = HASH (op1, mode); |
7afe21cc RK |
4508 | } |
4509 | ||
9714cf43 | 4510 | op1_elt = insert (op1, NULL, op1_hash, mode); |
7afe21cc | 4511 | op1_elt->in_memory = op1_in_memory; |
7afe21cc | 4512 | } |
eb5ad42a RS |
4513 | |
4514 | merge_equiv_classes (op0_elt, op1_elt); | |
4515 | last_jump_equiv_class = op0_elt; | |
7afe21cc RK |
4516 | } |
4517 | \f | |
4518 | /* CSE processing for one instruction. | |
4519 | First simplify sources and addresses of all assignments | |
4520 | in the instruction, using previously-computed equivalents values. | |
4521 | Then install the new sources and destinations in the table | |
278a83b2 | 4522 | of available values. |
7afe21cc | 4523 | |
1ed0205e VM |
4524 | If LIBCALL_INSN is nonzero, don't record any equivalence made in |
4525 | the insn. It means that INSN is inside libcall block. In this | |
ddc356e8 | 4526 | case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */ |
7afe21cc RK |
4527 | |
4528 | /* Data on one SET contained in the instruction. */ | |
4529 | ||
4530 | struct set | |
4531 | { | |
4532 | /* The SET rtx itself. */ | |
4533 | rtx rtl; | |
4534 | /* The SET_SRC of the rtx (the original value, if it is changing). */ | |
4535 | rtx src; | |
4536 | /* The hash-table element for the SET_SRC of the SET. */ | |
4537 | struct table_elt *src_elt; | |
2197a88a RK |
4538 | /* Hash value for the SET_SRC. */ |
4539 | unsigned src_hash; | |
4540 | /* Hash value for the SET_DEST. */ | |
4541 | unsigned dest_hash; | |
7afe21cc RK |
4542 | /* The SET_DEST, with SUBREG, etc., stripped. */ |
4543 | rtx inner_dest; | |
278a83b2 | 4544 | /* Nonzero if the SET_SRC is in memory. */ |
7afe21cc | 4545 | char src_in_memory; |
7afe21cc RK |
4546 | /* Nonzero if the SET_SRC contains something |
4547 | whose value cannot be predicted and understood. */ | |
4548 | char src_volatile; | |
496324d0 DN |
4549 | /* Original machine mode, in case it becomes a CONST_INT. |
4550 | The size of this field should match the size of the mode | |
4551 | field of struct rtx_def (see rtl.h). */ | |
4552 | ENUM_BITFIELD(machine_mode) mode : 8; | |
7afe21cc RK |
4553 | /* A constant equivalent for SET_SRC, if any. */ |
4554 | rtx src_const; | |
47841d1b JJ |
4555 | /* Original SET_SRC value used for libcall notes. */ |
4556 | rtx orig_src; | |
2197a88a RK |
4557 | /* Hash value of constant equivalent for SET_SRC. */ |
4558 | unsigned src_const_hash; | |
7afe21cc RK |
4559 | /* Table entry for constant equivalent for SET_SRC, if any. */ |
4560 | struct table_elt *src_const_elt; | |
4561 | }; | |
4562 | ||
4563 | static void | |
7080f735 | 4564 | cse_insn (rtx insn, rtx libcall_insn) |
7afe21cc | 4565 | { |
b3694847 SS |
4566 | rtx x = PATTERN (insn); |
4567 | int i; | |
92f9aa51 | 4568 | rtx tem; |
b3694847 | 4569 | int n_sets = 0; |
7afe21cc | 4570 | |
2d8b0f3a | 4571 | #ifdef HAVE_cc0 |
7afe21cc RK |
4572 | /* Records what this insn does to set CC0. */ |
4573 | rtx this_insn_cc0 = 0; | |
135d84b8 | 4574 | enum machine_mode this_insn_cc0_mode = VOIDmode; |
2d8b0f3a | 4575 | #endif |
7afe21cc RK |
4576 | |
4577 | rtx src_eqv = 0; | |
4578 | struct table_elt *src_eqv_elt = 0; | |
6a651371 KG |
4579 | int src_eqv_volatile = 0; |
4580 | int src_eqv_in_memory = 0; | |
6a651371 | 4581 | unsigned src_eqv_hash = 0; |
7afe21cc | 4582 | |
9714cf43 | 4583 | struct set *sets = (struct set *) 0; |
7afe21cc RK |
4584 | |
4585 | this_insn = insn; | |
7afe21cc RK |
4586 | |
4587 | /* Find all the SETs and CLOBBERs in this instruction. | |
4588 | Record all the SETs in the array `set' and count them. | |
4589 | Also determine whether there is a CLOBBER that invalidates | |
4590 | all memory references, or all references at varying addresses. */ | |
4591 | ||
4b4bf941 | 4592 | if (CALL_P (insn)) |
f1e7c95f RK |
4593 | { |
4594 | for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1)) | |
f474c6f8 AO |
4595 | { |
4596 | if (GET_CODE (XEXP (tem, 0)) == CLOBBER) | |
4597 | invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode); | |
4598 | XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn); | |
4599 | } | |
f1e7c95f RK |
4600 | } |
4601 | ||
7afe21cc RK |
4602 | if (GET_CODE (x) == SET) |
4603 | { | |
703ad42b | 4604 | sets = alloca (sizeof (struct set)); |
7afe21cc RK |
4605 | sets[0].rtl = x; |
4606 | ||
4607 | /* Ignore SETs that are unconditional jumps. | |
4608 | They never need cse processing, so this does not hurt. | |
4609 | The reason is not efficiency but rather | |
4610 | so that we can test at the end for instructions | |
4611 | that have been simplified to unconditional jumps | |
4612 | and not be misled by unchanged instructions | |
4613 | that were unconditional jumps to begin with. */ | |
4614 | if (SET_DEST (x) == pc_rtx | |
4615 | && GET_CODE (SET_SRC (x)) == LABEL_REF) | |
4616 | ; | |
4617 | ||
4618 | /* Don't count call-insns, (set (reg 0) (call ...)), as a set. | |
4619 | The hard function value register is used only once, to copy to | |
4620 | someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)! | |
4621 | Ensure we invalidate the destination register. On the 80386 no | |
7722328e | 4622 | other code would invalidate it since it is a fixed_reg. |
0f41302f | 4623 | We need not check the return of apply_change_group; see canon_reg. */ |
7afe21cc RK |
4624 | |
4625 | else if (GET_CODE (SET_SRC (x)) == CALL) | |
4626 | { | |
4627 | canon_reg (SET_SRC (x), insn); | |
77fa0940 | 4628 | apply_change_group (); |
7afe21cc | 4629 | fold_rtx (SET_SRC (x), insn); |
bb4034b3 | 4630 | invalidate (SET_DEST (x), VOIDmode); |
7afe21cc RK |
4631 | } |
4632 | else | |
4633 | n_sets = 1; | |
4634 | } | |
4635 | else if (GET_CODE (x) == PARALLEL) | |
4636 | { | |
b3694847 | 4637 | int lim = XVECLEN (x, 0); |
7afe21cc | 4638 | |
703ad42b | 4639 | sets = alloca (lim * sizeof (struct set)); |
7afe21cc RK |
4640 | |
4641 | /* Find all regs explicitly clobbered in this insn, | |
4642 | and ensure they are not replaced with any other regs | |
4643 | elsewhere in this insn. | |
4644 | When a reg that is clobbered is also used for input, | |
4645 | we should presume that that is for a reason, | |
4646 | and we should not substitute some other register | |
4647 | which is not supposed to be clobbered. | |
4648 | Therefore, this loop cannot be merged into the one below | |
830a38ee | 4649 | because a CALL may precede a CLOBBER and refer to the |
7afe21cc RK |
4650 | value clobbered. We must not let a canonicalization do |
4651 | anything in that case. */ | |
4652 | for (i = 0; i < lim; i++) | |
4653 | { | |
b3694847 | 4654 | rtx y = XVECEXP (x, 0, i); |
2708da92 RS |
4655 | if (GET_CODE (y) == CLOBBER) |
4656 | { | |
4657 | rtx clobbered = XEXP (y, 0); | |
4658 | ||
f8cfc6aa | 4659 | if (REG_P (clobbered) |
2708da92 | 4660 | || GET_CODE (clobbered) == SUBREG) |
bb4034b3 | 4661 | invalidate (clobbered, VOIDmode); |
2708da92 RS |
4662 | else if (GET_CODE (clobbered) == STRICT_LOW_PART |
4663 | || GET_CODE (clobbered) == ZERO_EXTRACT) | |
bb4034b3 | 4664 | invalidate (XEXP (clobbered, 0), GET_MODE (clobbered)); |
2708da92 | 4665 | } |
7afe21cc | 4666 | } |
278a83b2 | 4667 | |
7afe21cc RK |
4668 | for (i = 0; i < lim; i++) |
4669 | { | |
b3694847 | 4670 | rtx y = XVECEXP (x, 0, i); |
7afe21cc RK |
4671 | if (GET_CODE (y) == SET) |
4672 | { | |
7722328e RK |
4673 | /* As above, we ignore unconditional jumps and call-insns and |
4674 | ignore the result of apply_change_group. */ | |
7afe21cc RK |
4675 | if (GET_CODE (SET_SRC (y)) == CALL) |
4676 | { | |
4677 | canon_reg (SET_SRC (y), insn); | |
77fa0940 | 4678 | apply_change_group (); |
7afe21cc | 4679 | fold_rtx (SET_SRC (y), insn); |
bb4034b3 | 4680 | invalidate (SET_DEST (y), VOIDmode); |
7afe21cc RK |
4681 | } |
4682 | else if (SET_DEST (y) == pc_rtx | |
4683 | && GET_CODE (SET_SRC (y)) == LABEL_REF) | |
4684 | ; | |
4685 | else | |
4686 | sets[n_sets++].rtl = y; | |
4687 | } | |
4688 | else if (GET_CODE (y) == CLOBBER) | |
4689 | { | |
9ae8ffe7 | 4690 | /* If we clobber memory, canon the address. |
7afe21cc RK |
4691 | This does nothing when a register is clobbered |
4692 | because we have already invalidated the reg. */ | |
3c0cb5de | 4693 | if (MEM_P (XEXP (y, 0))) |
9ae8ffe7 | 4694 | canon_reg (XEXP (y, 0), NULL_RTX); |
7afe21cc RK |
4695 | } |
4696 | else if (GET_CODE (y) == USE | |
f8cfc6aa | 4697 | && ! (REG_P (XEXP (y, 0)) |
7afe21cc | 4698 | && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER)) |
906c4e36 | 4699 | canon_reg (y, NULL_RTX); |
7afe21cc RK |
4700 | else if (GET_CODE (y) == CALL) |
4701 | { | |
7722328e RK |
4702 | /* The result of apply_change_group can be ignored; see |
4703 | canon_reg. */ | |
7afe21cc | 4704 | canon_reg (y, insn); |
77fa0940 | 4705 | apply_change_group (); |
7afe21cc RK |
4706 | fold_rtx (y, insn); |
4707 | } | |
4708 | } | |
4709 | } | |
4710 | else if (GET_CODE (x) == CLOBBER) | |
4711 | { | |
3c0cb5de | 4712 | if (MEM_P (XEXP (x, 0))) |
9ae8ffe7 | 4713 | canon_reg (XEXP (x, 0), NULL_RTX); |
7afe21cc RK |
4714 | } |
4715 | ||
4716 | /* Canonicalize a USE of a pseudo register or memory location. */ | |
4717 | else if (GET_CODE (x) == USE | |
f8cfc6aa | 4718 | && ! (REG_P (XEXP (x, 0)) |
7afe21cc | 4719 | && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER)) |
906c4e36 | 4720 | canon_reg (XEXP (x, 0), NULL_RTX); |
7afe21cc RK |
4721 | else if (GET_CODE (x) == CALL) |
4722 | { | |
7722328e | 4723 | /* The result of apply_change_group can be ignored; see canon_reg. */ |
7afe21cc | 4724 | canon_reg (x, insn); |
77fa0940 | 4725 | apply_change_group (); |
7afe21cc RK |
4726 | fold_rtx (x, insn); |
4727 | } | |
4728 | ||
7b3ab05e JW |
4729 | /* Store the equivalent value in SRC_EQV, if different, or if the DEST |
4730 | is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV | |
4731 | is handled specially for this case, and if it isn't set, then there will | |
9faa82d8 | 4732 | be no equivalence for the destination. */ |
92f9aa51 RK |
4733 | if (n_sets == 1 && REG_NOTES (insn) != 0 |
4734 | && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0 | |
7b3ab05e JW |
4735 | && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)) |
4736 | || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART)) | |
7b668f9e JJ |
4737 | { |
4738 | src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn); | |
4739 | XEXP (tem, 0) = src_eqv; | |
4740 | } | |
7afe21cc RK |
4741 | |
4742 | /* Canonicalize sources and addresses of destinations. | |
4743 | We do this in a separate pass to avoid problems when a MATCH_DUP is | |
4744 | present in the insn pattern. In that case, we want to ensure that | |
4745 | we don't break the duplicate nature of the pattern. So we will replace | |
4746 | both operands at the same time. Otherwise, we would fail to find an | |
4747 | equivalent substitution in the loop calling validate_change below. | |
7afe21cc RK |
4748 | |
4749 | We used to suppress canonicalization of DEST if it appears in SRC, | |
77fa0940 | 4750 | but we don't do this any more. */ |
7afe21cc RK |
4751 | |
4752 | for (i = 0; i < n_sets; i++) | |
4753 | { | |
4754 | rtx dest = SET_DEST (sets[i].rtl); | |
4755 | rtx src = SET_SRC (sets[i].rtl); | |
4756 | rtx new = canon_reg (src, insn); | |
58873255 | 4757 | int insn_code; |
7afe21cc | 4758 | |
47841d1b | 4759 | sets[i].orig_src = src; |
f8cfc6aa | 4760 | if ((REG_P (new) && REG_P (src) |
77fa0940 RK |
4761 | && ((REGNO (new) < FIRST_PSEUDO_REGISTER) |
4762 | != (REGNO (src) < FIRST_PSEUDO_REGISTER))) | |
58873255 | 4763 | || (insn_code = recog_memoized (insn)) < 0 |
a995e389 | 4764 | || insn_data[insn_code].n_dups > 0) |
77fa0940 | 4765 | validate_change (insn, &SET_SRC (sets[i].rtl), new, 1); |
7afe21cc RK |
4766 | else |
4767 | SET_SRC (sets[i].rtl) = new; | |
4768 | ||
4769 | if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT) | |
4770 | { | |
4771 | validate_change (insn, &XEXP (dest, 1), | |
77fa0940 | 4772 | canon_reg (XEXP (dest, 1), insn), 1); |
7afe21cc | 4773 | validate_change (insn, &XEXP (dest, 2), |
77fa0940 | 4774 | canon_reg (XEXP (dest, 2), insn), 1); |
7afe21cc RK |
4775 | } |
4776 | ||
4777 | while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART | |
4778 | || GET_CODE (dest) == ZERO_EXTRACT | |
4779 | || GET_CODE (dest) == SIGN_EXTRACT) | |
4780 | dest = XEXP (dest, 0); | |
4781 | ||
3c0cb5de | 4782 | if (MEM_P (dest)) |
7afe21cc RK |
4783 | canon_reg (dest, insn); |
4784 | } | |
4785 | ||
77fa0940 RK |
4786 | /* Now that we have done all the replacements, we can apply the change |
4787 | group and see if they all work. Note that this will cause some | |
4788 | canonicalizations that would have worked individually not to be applied | |
4789 | because some other canonicalization didn't work, but this should not | |
278a83b2 | 4790 | occur often. |
7722328e RK |
4791 | |
4792 | The result of apply_change_group can be ignored; see canon_reg. */ | |
77fa0940 RK |
4793 | |
4794 | apply_change_group (); | |
4795 | ||
7afe21cc RK |
4796 | /* Set sets[i].src_elt to the class each source belongs to. |
4797 | Detect assignments from or to volatile things | |
4798 | and set set[i] to zero so they will be ignored | |
4799 | in the rest of this function. | |
4800 | ||
4801 | Nothing in this loop changes the hash table or the register chains. */ | |
4802 | ||
4803 | for (i = 0; i < n_sets; i++) | |
4804 | { | |
b3694847 SS |
4805 | rtx src, dest; |
4806 | rtx src_folded; | |
4807 | struct table_elt *elt = 0, *p; | |
7afe21cc RK |
4808 | enum machine_mode mode; |
4809 | rtx src_eqv_here; | |
4810 | rtx src_const = 0; | |
4811 | rtx src_related = 0; | |
4812 | struct table_elt *src_const_elt = 0; | |
99a9c946 GS |
4813 | int src_cost = MAX_COST; |
4814 | int src_eqv_cost = MAX_COST; | |
4815 | int src_folded_cost = MAX_COST; | |
4816 | int src_related_cost = MAX_COST; | |
4817 | int src_elt_cost = MAX_COST; | |
4818 | int src_regcost = MAX_COST; | |
4819 | int src_eqv_regcost = MAX_COST; | |
4820 | int src_folded_regcost = MAX_COST; | |
4821 | int src_related_regcost = MAX_COST; | |
4822 | int src_elt_regcost = MAX_COST; | |
da7d8304 | 4823 | /* Set nonzero if we need to call force_const_mem on with the |
7afe21cc RK |
4824 | contents of src_folded before using it. */ |
4825 | int src_folded_force_flag = 0; | |
4826 | ||
4827 | dest = SET_DEST (sets[i].rtl); | |
4828 | src = SET_SRC (sets[i].rtl); | |
4829 | ||
4830 | /* If SRC is a constant that has no machine mode, | |
4831 | hash it with the destination's machine mode. | |
4832 | This way we can keep different modes separate. */ | |
4833 | ||
4834 | mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src); | |
4835 | sets[i].mode = mode; | |
4836 | ||
4837 | if (src_eqv) | |
4838 | { | |
4839 | enum machine_mode eqvmode = mode; | |
4840 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
4841 | eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0))); | |
4842 | do_not_record = 0; | |
4843 | hash_arg_in_memory = 0; | |
2197a88a | 4844 | src_eqv_hash = HASH (src_eqv, eqvmode); |
7afe21cc RK |
4845 | |
4846 | /* Find the equivalence class for the equivalent expression. */ | |
4847 | ||
4848 | if (!do_not_record) | |
2197a88a | 4849 | src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode); |
7afe21cc RK |
4850 | |
4851 | src_eqv_volatile = do_not_record; | |
4852 | src_eqv_in_memory = hash_arg_in_memory; | |
7afe21cc RK |
4853 | } |
4854 | ||
4855 | /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the | |
4856 | value of the INNER register, not the destination. So it is not | |
3826a3da | 4857 | a valid substitution for the source. But save it for later. */ |
7afe21cc RK |
4858 | if (GET_CODE (dest) == STRICT_LOW_PART) |
4859 | src_eqv_here = 0; | |
4860 | else | |
4861 | src_eqv_here = src_eqv; | |
4862 | ||
4863 | /* Simplify and foldable subexpressions in SRC. Then get the fully- | |
4864 | simplified result, which may not necessarily be valid. */ | |
4865 | src_folded = fold_rtx (src, insn); | |
4866 | ||
e6a125a0 RK |
4867 | #if 0 |
4868 | /* ??? This caused bad code to be generated for the m68k port with -O2. | |
4869 | Suppose src is (CONST_INT -1), and that after truncation src_folded | |
4870 | is (CONST_INT 3). Suppose src_folded is then used for src_const. | |
4871 | At the end we will add src and src_const to the same equivalence | |
4872 | class. We now have 3 and -1 on the same equivalence class. This | |
4873 | causes later instructions to be mis-optimized. */ | |
7afe21cc RK |
4874 | /* If storing a constant in a bitfield, pre-truncate the constant |
4875 | so we will be able to record it later. */ | |
4876 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT | |
4877 | || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT) | |
4878 | { | |
4879 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
4880 | ||
4881 | if (GET_CODE (src) == CONST_INT | |
4882 | && GET_CODE (width) == CONST_INT | |
906c4e36 RK |
4883 | && INTVAL (width) < HOST_BITS_PER_WIDE_INT |
4884 | && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width)))) | |
4885 | src_folded | |
4886 | = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1 | |
4887 | << INTVAL (width)) - 1)); | |
7afe21cc | 4888 | } |
e6a125a0 | 4889 | #endif |
7afe21cc RK |
4890 | |
4891 | /* Compute SRC's hash code, and also notice if it | |
4892 | should not be recorded at all. In that case, | |
4893 | prevent any further processing of this assignment. */ | |
4894 | do_not_record = 0; | |
4895 | hash_arg_in_memory = 0; | |
7afe21cc RK |
4896 | |
4897 | sets[i].src = src; | |
2197a88a | 4898 | sets[i].src_hash = HASH (src, mode); |
7afe21cc RK |
4899 | sets[i].src_volatile = do_not_record; |
4900 | sets[i].src_in_memory = hash_arg_in_memory; | |
7afe21cc | 4901 | |
50196afa | 4902 | /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is |
43e72072 JJ |
4903 | a pseudo, do not record SRC. Using SRC as a replacement for |
4904 | anything else will be incorrect in that situation. Note that | |
4905 | this usually occurs only for stack slots, in which case all the | |
4906 | RTL would be referring to SRC, so we don't lose any optimization | |
4907 | opportunities by not having SRC in the hash table. */ | |
50196afa | 4908 | |
3c0cb5de | 4909 | if (MEM_P (src) |
43e72072 | 4910 | && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0 |
f8cfc6aa | 4911 | && REG_P (dest) |
43e72072 | 4912 | && REGNO (dest) >= FIRST_PSEUDO_REGISTER) |
50196afa RK |
4913 | sets[i].src_volatile = 1; |
4914 | ||
0dadecf6 RK |
4915 | #if 0 |
4916 | /* It is no longer clear why we used to do this, but it doesn't | |
4917 | appear to still be needed. So let's try without it since this | |
4918 | code hurts cse'ing widened ops. */ | |
9a5a17f3 | 4919 | /* If source is a paradoxical subreg (such as QI treated as an SI), |
7afe21cc RK |
4920 | treat it as volatile. It may do the work of an SI in one context |
4921 | where the extra bits are not being used, but cannot replace an SI | |
4922 | in general. */ | |
4923 | if (GET_CODE (src) == SUBREG | |
4924 | && (GET_MODE_SIZE (GET_MODE (src)) | |
4925 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))) | |
4926 | sets[i].src_volatile = 1; | |
0dadecf6 | 4927 | #endif |
7afe21cc RK |
4928 | |
4929 | /* Locate all possible equivalent forms for SRC. Try to replace | |
4930 | SRC in the insn with each cheaper equivalent. | |
4931 | ||
4932 | We have the following types of equivalents: SRC itself, a folded | |
4933 | version, a value given in a REG_EQUAL note, or a value related | |
4934 | to a constant. | |
4935 | ||
4936 | Each of these equivalents may be part of an additional class | |
4937 | of equivalents (if more than one is in the table, they must be in | |
4938 | the same class; we check for this). | |
4939 | ||
4940 | If the source is volatile, we don't do any table lookups. | |
4941 | ||
4942 | We note any constant equivalent for possible later use in a | |
4943 | REG_NOTE. */ | |
4944 | ||
4945 | if (!sets[i].src_volatile) | |
2197a88a | 4946 | elt = lookup (src, sets[i].src_hash, mode); |
7afe21cc RK |
4947 | |
4948 | sets[i].src_elt = elt; | |
4949 | ||
4950 | if (elt && src_eqv_here && src_eqv_elt) | |
278a83b2 KH |
4951 | { |
4952 | if (elt->first_same_value != src_eqv_elt->first_same_value) | |
7afe21cc RK |
4953 | { |
4954 | /* The REG_EQUAL is indicating that two formerly distinct | |
4955 | classes are now equivalent. So merge them. */ | |
4956 | merge_equiv_classes (elt, src_eqv_elt); | |
2197a88a RK |
4957 | src_eqv_hash = HASH (src_eqv, elt->mode); |
4958 | src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode); | |
7afe21cc RK |
4959 | } |
4960 | ||
278a83b2 KH |
4961 | src_eqv_here = 0; |
4962 | } | |
7afe21cc RK |
4963 | |
4964 | else if (src_eqv_elt) | |
278a83b2 | 4965 | elt = src_eqv_elt; |
7afe21cc RK |
4966 | |
4967 | /* Try to find a constant somewhere and record it in `src_const'. | |
4968 | Record its table element, if any, in `src_const_elt'. Look in | |
4969 | any known equivalences first. (If the constant is not in the | |
2197a88a | 4970 | table, also set `sets[i].src_const_hash'). */ |
7afe21cc | 4971 | if (elt) |
278a83b2 | 4972 | for (p = elt->first_same_value; p; p = p->next_same_value) |
7afe21cc RK |
4973 | if (p->is_const) |
4974 | { | |
4975 | src_const = p->exp; | |
4976 | src_const_elt = elt; | |
4977 | break; | |
4978 | } | |
4979 | ||
4980 | if (src_const == 0 | |
4981 | && (CONSTANT_P (src_folded) | |
278a83b2 | 4982 | /* Consider (minus (label_ref L1) (label_ref L2)) as |
7afe21cc RK |
4983 | "constant" here so we will record it. This allows us |
4984 | to fold switch statements when an ADDR_DIFF_VEC is used. */ | |
4985 | || (GET_CODE (src_folded) == MINUS | |
4986 | && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF | |
4987 | && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF))) | |
4988 | src_const = src_folded, src_const_elt = elt; | |
4989 | else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here)) | |
4990 | src_const = src_eqv_here, src_const_elt = src_eqv_elt; | |
4991 | ||
4992 | /* If we don't know if the constant is in the table, get its | |
4993 | hash code and look it up. */ | |
4994 | if (src_const && src_const_elt == 0) | |
4995 | { | |
2197a88a RK |
4996 | sets[i].src_const_hash = HASH (src_const, mode); |
4997 | src_const_elt = lookup (src_const, sets[i].src_const_hash, mode); | |
7afe21cc RK |
4998 | } |
4999 | ||
5000 | sets[i].src_const = src_const; | |
5001 | sets[i].src_const_elt = src_const_elt; | |
5002 | ||
5003 | /* If the constant and our source are both in the table, mark them as | |
5004 | equivalent. Otherwise, if a constant is in the table but the source | |
5005 | isn't, set ELT to it. */ | |
5006 | if (src_const_elt && elt | |
5007 | && src_const_elt->first_same_value != elt->first_same_value) | |
5008 | merge_equiv_classes (elt, src_const_elt); | |
5009 | else if (src_const_elt && elt == 0) | |
5010 | elt = src_const_elt; | |
5011 | ||
5012 | /* See if there is a register linearly related to a constant | |
5013 | equivalent of SRC. */ | |
5014 | if (src_const | |
5015 | && (GET_CODE (src_const) == CONST | |
5016 | || (src_const_elt && src_const_elt->related_value != 0))) | |
278a83b2 KH |
5017 | { |
5018 | src_related = use_related_value (src_const, src_const_elt); | |
5019 | if (src_related) | |
5020 | { | |
7afe21cc | 5021 | struct table_elt *src_related_elt |
278a83b2 | 5022 | = lookup (src_related, HASH (src_related, mode), mode); |
7afe21cc | 5023 | if (src_related_elt && elt) |
278a83b2 | 5024 | { |
7afe21cc RK |
5025 | if (elt->first_same_value |
5026 | != src_related_elt->first_same_value) | |
278a83b2 | 5027 | /* This can occur when we previously saw a CONST |
7afe21cc RK |
5028 | involving a SYMBOL_REF and then see the SYMBOL_REF |
5029 | twice. Merge the involved classes. */ | |
5030 | merge_equiv_classes (elt, src_related_elt); | |
5031 | ||
278a83b2 | 5032 | src_related = 0; |
7afe21cc | 5033 | src_related_elt = 0; |
278a83b2 KH |
5034 | } |
5035 | else if (src_related_elt && elt == 0) | |
5036 | elt = src_related_elt; | |
7afe21cc | 5037 | } |
278a83b2 | 5038 | } |
7afe21cc | 5039 | |
e4600702 RK |
5040 | /* See if we have a CONST_INT that is already in a register in a |
5041 | wider mode. */ | |
5042 | ||
5043 | if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT | |
5044 | && GET_MODE_CLASS (mode) == MODE_INT | |
5045 | && GET_MODE_BITSIZE (mode) < BITS_PER_WORD) | |
5046 | { | |
5047 | enum machine_mode wider_mode; | |
5048 | ||
5049 | for (wider_mode = GET_MODE_WIDER_MODE (mode); | |
5050 | GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD | |
5051 | && src_related == 0; | |
5052 | wider_mode = GET_MODE_WIDER_MODE (wider_mode)) | |
5053 | { | |
5054 | struct table_elt *const_elt | |
5055 | = lookup (src_const, HASH (src_const, wider_mode), wider_mode); | |
5056 | ||
5057 | if (const_elt == 0) | |
5058 | continue; | |
5059 | ||
5060 | for (const_elt = const_elt->first_same_value; | |
5061 | const_elt; const_elt = const_elt->next_same_value) | |
f8cfc6aa | 5062 | if (REG_P (const_elt->exp)) |
e4600702 | 5063 | { |
4de249d9 | 5064 | src_related = gen_lowpart (mode, |
e4600702 RK |
5065 | const_elt->exp); |
5066 | break; | |
5067 | } | |
5068 | } | |
5069 | } | |
5070 | ||
d45cf215 RS |
5071 | /* Another possibility is that we have an AND with a constant in |
5072 | a mode narrower than a word. If so, it might have been generated | |
5073 | as part of an "if" which would narrow the AND. If we already | |
5074 | have done the AND in a wider mode, we can use a SUBREG of that | |
5075 | value. */ | |
5076 | ||
5077 | if (flag_expensive_optimizations && ! src_related | |
5078 | && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT | |
5079 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD) | |
5080 | { | |
5081 | enum machine_mode tmode; | |
38a448ca | 5082 | rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1)); |
d45cf215 RS |
5083 | |
5084 | for (tmode = GET_MODE_WIDER_MODE (mode); | |
5085 | GET_MODE_SIZE (tmode) <= UNITS_PER_WORD; | |
5086 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
5087 | { | |
4de249d9 | 5088 | rtx inner = gen_lowpart (tmode, XEXP (src, 0)); |
d45cf215 RS |
5089 | struct table_elt *larger_elt; |
5090 | ||
5091 | if (inner) | |
5092 | { | |
5093 | PUT_MODE (new_and, tmode); | |
5094 | XEXP (new_and, 0) = inner; | |
5095 | larger_elt = lookup (new_and, HASH (new_and, tmode), tmode); | |
5096 | if (larger_elt == 0) | |
5097 | continue; | |
5098 | ||
5099 | for (larger_elt = larger_elt->first_same_value; | |
5100 | larger_elt; larger_elt = larger_elt->next_same_value) | |
f8cfc6aa | 5101 | if (REG_P (larger_elt->exp)) |
d45cf215 RS |
5102 | { |
5103 | src_related | |
4de249d9 | 5104 | = gen_lowpart (mode, larger_elt->exp); |
d45cf215 RS |
5105 | break; |
5106 | } | |
5107 | ||
5108 | if (src_related) | |
5109 | break; | |
5110 | } | |
5111 | } | |
5112 | } | |
7bac1be0 RK |
5113 | |
5114 | #ifdef LOAD_EXTEND_OP | |
5115 | /* See if a MEM has already been loaded with a widening operation; | |
5116 | if it has, we can use a subreg of that. Many CISC machines | |
5117 | also have such operations, but this is only likely to be | |
71cc389b | 5118 | beneficial on these machines. */ |
278a83b2 | 5119 | |
ddc356e8 | 5120 | if (flag_expensive_optimizations && src_related == 0 |
7bac1be0 RK |
5121 | && (GET_MODE_SIZE (mode) < UNITS_PER_WORD) |
5122 | && GET_MODE_CLASS (mode) == MODE_INT | |
3c0cb5de | 5123 | && MEM_P (src) && ! do_not_record |
7bac1be0 RK |
5124 | && LOAD_EXTEND_OP (mode) != NIL) |
5125 | { | |
5126 | enum machine_mode tmode; | |
278a83b2 | 5127 | |
7bac1be0 RK |
5128 | /* Set what we are trying to extend and the operation it might |
5129 | have been extended with. */ | |
5130 | PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode)); | |
5131 | XEXP (memory_extend_rtx, 0) = src; | |
278a83b2 | 5132 | |
7bac1be0 RK |
5133 | for (tmode = GET_MODE_WIDER_MODE (mode); |
5134 | GET_MODE_SIZE (tmode) <= UNITS_PER_WORD; | |
5135 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
5136 | { | |
5137 | struct table_elt *larger_elt; | |
278a83b2 | 5138 | |
7bac1be0 | 5139 | PUT_MODE (memory_extend_rtx, tmode); |
278a83b2 | 5140 | larger_elt = lookup (memory_extend_rtx, |
7bac1be0 RK |
5141 | HASH (memory_extend_rtx, tmode), tmode); |
5142 | if (larger_elt == 0) | |
5143 | continue; | |
278a83b2 | 5144 | |
7bac1be0 RK |
5145 | for (larger_elt = larger_elt->first_same_value; |
5146 | larger_elt; larger_elt = larger_elt->next_same_value) | |
f8cfc6aa | 5147 | if (REG_P (larger_elt->exp)) |
7bac1be0 | 5148 | { |
4de249d9 | 5149 | src_related = gen_lowpart (mode, |
7bac1be0 RK |
5150 | larger_elt->exp); |
5151 | break; | |
5152 | } | |
278a83b2 | 5153 | |
7bac1be0 RK |
5154 | if (src_related) |
5155 | break; | |
5156 | } | |
5157 | } | |
5158 | #endif /* LOAD_EXTEND_OP */ | |
278a83b2 | 5159 | |
7afe21cc | 5160 | if (src == src_folded) |
278a83b2 | 5161 | src_folded = 0; |
7afe21cc | 5162 | |
da7d8304 | 5163 | /* At this point, ELT, if nonzero, points to a class of expressions |
7afe21cc | 5164 | equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED, |
da7d8304 | 5165 | and SRC_RELATED, if nonzero, each contain additional equivalent |
7afe21cc RK |
5166 | expressions. Prune these latter expressions by deleting expressions |
5167 | already in the equivalence class. | |
5168 | ||
5169 | Check for an equivalent identical to the destination. If found, | |
5170 | this is the preferred equivalent since it will likely lead to | |
5171 | elimination of the insn. Indicate this by placing it in | |
5172 | `src_related'. */ | |
5173 | ||
278a83b2 KH |
5174 | if (elt) |
5175 | elt = elt->first_same_value; | |
7afe21cc | 5176 | for (p = elt; p; p = p->next_same_value) |
278a83b2 | 5177 | { |
7afe21cc RK |
5178 | enum rtx_code code = GET_CODE (p->exp); |
5179 | ||
5180 | /* If the expression is not valid, ignore it. Then we do not | |
5181 | have to check for validity below. In most cases, we can use | |
5182 | `rtx_equal_p', since canonicalization has already been done. */ | |
5183 | if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, 0)) | |
5184 | continue; | |
5185 | ||
5a03c8c4 RK |
5186 | /* Also skip paradoxical subregs, unless that's what we're |
5187 | looking for. */ | |
5188 | if (code == SUBREG | |
5189 | && (GET_MODE_SIZE (GET_MODE (p->exp)) | |
5190 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp)))) | |
5191 | && ! (src != 0 | |
5192 | && GET_CODE (src) == SUBREG | |
5193 | && GET_MODE (src) == GET_MODE (p->exp) | |
5194 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
5195 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp)))))) | |
5196 | continue; | |
5197 | ||
278a83b2 | 5198 | if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp)) |
7afe21cc | 5199 | src = 0; |
278a83b2 | 5200 | else if (src_folded && GET_CODE (src_folded) == code |
7afe21cc RK |
5201 | && rtx_equal_p (src_folded, p->exp)) |
5202 | src_folded = 0; | |
278a83b2 | 5203 | else if (src_eqv_here && GET_CODE (src_eqv_here) == code |
7afe21cc RK |
5204 | && rtx_equal_p (src_eqv_here, p->exp)) |
5205 | src_eqv_here = 0; | |
278a83b2 | 5206 | else if (src_related && GET_CODE (src_related) == code |
7afe21cc RK |
5207 | && rtx_equal_p (src_related, p->exp)) |
5208 | src_related = 0; | |
5209 | ||
5210 | /* This is the same as the destination of the insns, we want | |
5211 | to prefer it. Copy it to src_related. The code below will | |
5212 | then give it a negative cost. */ | |
5213 | if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest)) | |
5214 | src_related = dest; | |
278a83b2 | 5215 | } |
7afe21cc RK |
5216 | |
5217 | /* Find the cheapest valid equivalent, trying all the available | |
5218 | possibilities. Prefer items not in the hash table to ones | |
5219 | that are when they are equal cost. Note that we can never | |
5220 | worsen an insn as the current contents will also succeed. | |
05c33dd8 | 5221 | If we find an equivalent identical to the destination, use it as best, |
0f41302f | 5222 | since this insn will probably be eliminated in that case. */ |
7afe21cc RK |
5223 | if (src) |
5224 | { | |
5225 | if (rtx_equal_p (src, dest)) | |
f1c1dfc3 | 5226 | src_cost = src_regcost = -1; |
7afe21cc | 5227 | else |
630c79be BS |
5228 | { |
5229 | src_cost = COST (src); | |
5230 | src_regcost = approx_reg_cost (src); | |
5231 | } | |
7afe21cc RK |
5232 | } |
5233 | ||
5234 | if (src_eqv_here) | |
5235 | { | |
5236 | if (rtx_equal_p (src_eqv_here, dest)) | |
f1c1dfc3 | 5237 | src_eqv_cost = src_eqv_regcost = -1; |
7afe21cc | 5238 | else |
630c79be BS |
5239 | { |
5240 | src_eqv_cost = COST (src_eqv_here); | |
5241 | src_eqv_regcost = approx_reg_cost (src_eqv_here); | |
5242 | } | |
7afe21cc RK |
5243 | } |
5244 | ||
5245 | if (src_folded) | |
5246 | { | |
5247 | if (rtx_equal_p (src_folded, dest)) | |
f1c1dfc3 | 5248 | src_folded_cost = src_folded_regcost = -1; |
7afe21cc | 5249 | else |
630c79be BS |
5250 | { |
5251 | src_folded_cost = COST (src_folded); | |
5252 | src_folded_regcost = approx_reg_cost (src_folded); | |
5253 | } | |
7afe21cc RK |
5254 | } |
5255 | ||
5256 | if (src_related) | |
5257 | { | |
5258 | if (rtx_equal_p (src_related, dest)) | |
f1c1dfc3 | 5259 | src_related_cost = src_related_regcost = -1; |
7afe21cc | 5260 | else |
630c79be BS |
5261 | { |
5262 | src_related_cost = COST (src_related); | |
5263 | src_related_regcost = approx_reg_cost (src_related); | |
5264 | } | |
7afe21cc RK |
5265 | } |
5266 | ||
5267 | /* If this was an indirect jump insn, a known label will really be | |
5268 | cheaper even though it looks more expensive. */ | |
5269 | if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF) | |
99a9c946 | 5270 | src_folded = src_const, src_folded_cost = src_folded_regcost = -1; |
278a83b2 | 5271 | |
7afe21cc RK |
5272 | /* Terminate loop when replacement made. This must terminate since |
5273 | the current contents will be tested and will always be valid. */ | |
5274 | while (1) | |
278a83b2 KH |
5275 | { |
5276 | rtx trial; | |
7afe21cc | 5277 | |
278a83b2 | 5278 | /* Skip invalid entries. */ |
f8cfc6aa | 5279 | while (elt && !REG_P (elt->exp) |
278a83b2 KH |
5280 | && ! exp_equiv_p (elt->exp, elt->exp, 1, 0)) |
5281 | elt = elt->next_same_value; | |
5a03c8c4 RK |
5282 | |
5283 | /* A paradoxical subreg would be bad here: it'll be the right | |
5284 | size, but later may be adjusted so that the upper bits aren't | |
5285 | what we want. So reject it. */ | |
5286 | if (elt != 0 | |
5287 | && GET_CODE (elt->exp) == SUBREG | |
5288 | && (GET_MODE_SIZE (GET_MODE (elt->exp)) | |
5289 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp)))) | |
5290 | /* It is okay, though, if the rtx we're trying to match | |
5291 | will ignore any of the bits we can't predict. */ | |
5292 | && ! (src != 0 | |
5293 | && GET_CODE (src) == SUBREG | |
5294 | && GET_MODE (src) == GET_MODE (elt->exp) | |
5295 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
5296 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp)))))) | |
5297 | { | |
5298 | elt = elt->next_same_value; | |
5299 | continue; | |
5300 | } | |
278a83b2 | 5301 | |
68252e27 | 5302 | if (elt) |
630c79be BS |
5303 | { |
5304 | src_elt_cost = elt->cost; | |
5305 | src_elt_regcost = elt->regcost; | |
5306 | } | |
7afe21cc | 5307 | |
68252e27 | 5308 | /* Find cheapest and skip it for the next time. For items |
7afe21cc RK |
5309 | of equal cost, use this order: |
5310 | src_folded, src, src_eqv, src_related and hash table entry. */ | |
99a9c946 | 5311 | if (src_folded |
56ae04af KH |
5312 | && preferable (src_folded_cost, src_folded_regcost, |
5313 | src_cost, src_regcost) <= 0 | |
5314 | && preferable (src_folded_cost, src_folded_regcost, | |
5315 | src_eqv_cost, src_eqv_regcost) <= 0 | |
5316 | && preferable (src_folded_cost, src_folded_regcost, | |
5317 | src_related_cost, src_related_regcost) <= 0 | |
5318 | && preferable (src_folded_cost, src_folded_regcost, | |
5319 | src_elt_cost, src_elt_regcost) <= 0) | |
7afe21cc | 5320 | { |
f1c1dfc3 | 5321 | trial = src_folded, src_folded_cost = MAX_COST; |
7afe21cc | 5322 | if (src_folded_force_flag) |
9d8de1de EB |
5323 | { |
5324 | rtx forced = force_const_mem (mode, trial); | |
5325 | if (forced) | |
5326 | trial = forced; | |
5327 | } | |
7afe21cc | 5328 | } |
99a9c946 | 5329 | else if (src |
56ae04af KH |
5330 | && preferable (src_cost, src_regcost, |
5331 | src_eqv_cost, src_eqv_regcost) <= 0 | |
5332 | && preferable (src_cost, src_regcost, | |
5333 | src_related_cost, src_related_regcost) <= 0 | |
5334 | && preferable (src_cost, src_regcost, | |
5335 | src_elt_cost, src_elt_regcost) <= 0) | |
f1c1dfc3 | 5336 | trial = src, src_cost = MAX_COST; |
99a9c946 | 5337 | else if (src_eqv_here |
56ae04af KH |
5338 | && preferable (src_eqv_cost, src_eqv_regcost, |
5339 | src_related_cost, src_related_regcost) <= 0 | |
5340 | && preferable (src_eqv_cost, src_eqv_regcost, | |
5341 | src_elt_cost, src_elt_regcost) <= 0) | |
f1c1dfc3 | 5342 | trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST; |
99a9c946 | 5343 | else if (src_related |
56ae04af KH |
5344 | && preferable (src_related_cost, src_related_regcost, |
5345 | src_elt_cost, src_elt_regcost) <= 0) | |
68252e27 | 5346 | trial = copy_rtx (src_related), src_related_cost = MAX_COST; |
278a83b2 | 5347 | else |
7afe21cc | 5348 | { |
05c33dd8 | 5349 | trial = copy_rtx (elt->exp); |
7afe21cc | 5350 | elt = elt->next_same_value; |
f1c1dfc3 | 5351 | src_elt_cost = MAX_COST; |
7afe21cc RK |
5352 | } |
5353 | ||
5354 | /* We don't normally have an insn matching (set (pc) (pc)), so | |
5355 | check for this separately here. We will delete such an | |
5356 | insn below. | |
5357 | ||
d466c016 JL |
5358 | For other cases such as a table jump or conditional jump |
5359 | where we know the ultimate target, go ahead and replace the | |
5360 | operand. While that may not make a valid insn, we will | |
5361 | reemit the jump below (and also insert any necessary | |
5362 | barriers). */ | |
7afe21cc RK |
5363 | if (n_sets == 1 && dest == pc_rtx |
5364 | && (trial == pc_rtx | |
5365 | || (GET_CODE (trial) == LABEL_REF | |
5366 | && ! condjump_p (insn)))) | |
5367 | { | |
d466c016 | 5368 | SET_SRC (sets[i].rtl) = trial; |
602c4c0d | 5369 | cse_jumps_altered = 1; |
7afe21cc RK |
5370 | break; |
5371 | } | |
278a83b2 | 5372 | |
7afe21cc | 5373 | /* Look for a substitution that makes a valid insn. */ |
ddc356e8 | 5374 | else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0)) |
05c33dd8 | 5375 | { |
dbaff908 RS |
5376 | rtx new = canon_reg (SET_SRC (sets[i].rtl), insn); |
5377 | ||
7bd8b2a8 JL |
5378 | /* If we just made a substitution inside a libcall, then we |
5379 | need to make the same substitution in any notes attached | |
5380 | to the RETVAL insn. */ | |
1ed0205e | 5381 | if (libcall_insn |
f8cfc6aa | 5382 | && (REG_P (sets[i].orig_src) |
47841d1b | 5383 | || GET_CODE (sets[i].orig_src) == SUBREG |
3c0cb5de | 5384 | || MEM_P (sets[i].orig_src))) |
d8b7ec41 RS |
5385 | { |
5386 | rtx note = find_reg_equal_equiv_note (libcall_insn); | |
5387 | if (note != 0) | |
5388 | XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), | |
5389 | sets[i].orig_src, | |
5390 | copy_rtx (new)); | |
5391 | } | |
7bd8b2a8 | 5392 | |
7722328e RK |
5393 | /* The result of apply_change_group can be ignored; see |
5394 | canon_reg. */ | |
5395 | ||
dbaff908 | 5396 | validate_change (insn, &SET_SRC (sets[i].rtl), new, 1); |
6702af89 | 5397 | apply_change_group (); |
05c33dd8 RK |
5398 | break; |
5399 | } | |
7afe21cc | 5400 | |
278a83b2 | 5401 | /* If we previously found constant pool entries for |
7afe21cc RK |
5402 | constants and this is a constant, try making a |
5403 | pool entry. Put it in src_folded unless we already have done | |
5404 | this since that is where it likely came from. */ | |
5405 | ||
5406 | else if (constant_pool_entries_cost | |
5407 | && CONSTANT_P (trial) | |
d51ff7cb JW |
5408 | /* Reject cases that will abort in decode_rtx_const. |
5409 | On the alpha when simplifying a switch, we get | |
5410 | (const (truncate (minus (label_ref) (label_ref)))). */ | |
1bbd065b RK |
5411 | && ! (GET_CODE (trial) == CONST |
5412 | && GET_CODE (XEXP (trial, 0)) == TRUNCATE) | |
d51ff7cb JW |
5413 | /* Likewise on IA-64, except without the truncate. */ |
5414 | && ! (GET_CODE (trial) == CONST | |
5415 | && GET_CODE (XEXP (trial, 0)) == MINUS | |
5416 | && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF | |
5417 | && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF) | |
1bbd065b | 5418 | && (src_folded == 0 |
3c0cb5de | 5419 | || (!MEM_P (src_folded) |
1bbd065b | 5420 | && ! src_folded_force_flag)) |
9ae8ffe7 JL |
5421 | && GET_MODE_CLASS (mode) != MODE_CC |
5422 | && mode != VOIDmode) | |
7afe21cc RK |
5423 | { |
5424 | src_folded_force_flag = 1; | |
5425 | src_folded = trial; | |
5426 | src_folded_cost = constant_pool_entries_cost; | |
dd0ba281 | 5427 | src_folded_regcost = constant_pool_entries_regcost; |
7afe21cc | 5428 | } |
278a83b2 | 5429 | } |
7afe21cc RK |
5430 | |
5431 | src = SET_SRC (sets[i].rtl); | |
5432 | ||
5433 | /* In general, it is good to have a SET with SET_SRC == SET_DEST. | |
5434 | However, there is an important exception: If both are registers | |
5435 | that are not the head of their equivalence class, replace SET_SRC | |
5436 | with the head of the class. If we do not do this, we will have | |
5437 | both registers live over a portion of the basic block. This way, | |
5438 | their lifetimes will likely abut instead of overlapping. */ | |
f8cfc6aa | 5439 | if (REG_P (dest) |
1bb98cec | 5440 | && REGNO_QTY_VALID_P (REGNO (dest))) |
7afe21cc | 5441 | { |
1bb98cec DM |
5442 | int dest_q = REG_QTY (REGNO (dest)); |
5443 | struct qty_table_elem *dest_ent = &qty_table[dest_q]; | |
5444 | ||
5445 | if (dest_ent->mode == GET_MODE (dest) | |
5446 | && dest_ent->first_reg != REGNO (dest) | |
f8cfc6aa | 5447 | && REG_P (src) && REGNO (src) == REGNO (dest) |
1bb98cec DM |
5448 | /* Don't do this if the original insn had a hard reg as |
5449 | SET_SRC or SET_DEST. */ | |
f8cfc6aa | 5450 | && (!REG_P (sets[i].src) |
1bb98cec | 5451 | || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER) |
f8cfc6aa | 5452 | && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER)) |
1bb98cec DM |
5453 | /* We can't call canon_reg here because it won't do anything if |
5454 | SRC is a hard register. */ | |
759bd8b7 | 5455 | { |
1bb98cec DM |
5456 | int src_q = REG_QTY (REGNO (src)); |
5457 | struct qty_table_elem *src_ent = &qty_table[src_q]; | |
5458 | int first = src_ent->first_reg; | |
5459 | rtx new_src | |
5460 | = (first >= FIRST_PSEUDO_REGISTER | |
5461 | ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first)); | |
5462 | ||
5463 | /* We must use validate-change even for this, because this | |
5464 | might be a special no-op instruction, suitable only to | |
5465 | tag notes onto. */ | |
5466 | if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0)) | |
5467 | { | |
5468 | src = new_src; | |
5469 | /* If we had a constant that is cheaper than what we are now | |
5470 | setting SRC to, use that constant. We ignored it when we | |
5471 | thought we could make this into a no-op. */ | |
5472 | if (src_const && COST (src_const) < COST (src) | |
278a83b2 KH |
5473 | && validate_change (insn, &SET_SRC (sets[i].rtl), |
5474 | src_const, 0)) | |
1bb98cec DM |
5475 | src = src_const; |
5476 | } | |
759bd8b7 | 5477 | } |
7afe21cc RK |
5478 | } |
5479 | ||
5480 | /* If we made a change, recompute SRC values. */ | |
5481 | if (src != sets[i].src) | |
278a83b2 | 5482 | { |
4eadede7 | 5483 | cse_altered = 1; |
278a83b2 KH |
5484 | do_not_record = 0; |
5485 | hash_arg_in_memory = 0; | |
7afe21cc | 5486 | sets[i].src = src; |
278a83b2 KH |
5487 | sets[i].src_hash = HASH (src, mode); |
5488 | sets[i].src_volatile = do_not_record; | |
5489 | sets[i].src_in_memory = hash_arg_in_memory; | |
5490 | sets[i].src_elt = lookup (src, sets[i].src_hash, mode); | |
5491 | } | |
7afe21cc RK |
5492 | |
5493 | /* If this is a single SET, we are setting a register, and we have an | |
5494 | equivalent constant, we want to add a REG_NOTE. We don't want | |
5495 | to write a REG_EQUAL note for a constant pseudo since verifying that | |
d45cf215 | 5496 | that pseudo hasn't been eliminated is a pain. Such a note also |
278a83b2 | 5497 | won't help anything. |
ac7ef8d5 FS |
5498 | |
5499 | Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF))) | |
5500 | which can be created for a reference to a compile time computable | |
5501 | entry in a jump table. */ | |
5502 | ||
f8cfc6aa JQ |
5503 | if (n_sets == 1 && src_const && REG_P (dest) |
5504 | && !REG_P (src_const) | |
ac7ef8d5 FS |
5505 | && ! (GET_CODE (src_const) == CONST |
5506 | && GET_CODE (XEXP (src_const, 0)) == MINUS | |
5507 | && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF | |
5508 | && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)) | |
7afe21cc | 5509 | { |
a77b7e32 RS |
5510 | /* We only want a REG_EQUAL note if src_const != src. */ |
5511 | if (! rtx_equal_p (src, src_const)) | |
5512 | { | |
5513 | /* Make sure that the rtx is not shared. */ | |
5514 | src_const = copy_rtx (src_const); | |
51e2a951 | 5515 | |
a77b7e32 RS |
5516 | /* Record the actual constant value in a REG_EQUAL note, |
5517 | making a new one if one does not already exist. */ | |
5518 | set_unique_reg_note (insn, REG_EQUAL, src_const); | |
5519 | } | |
7afe21cc RK |
5520 | } |
5521 | ||
5522 | /* Now deal with the destination. */ | |
5523 | do_not_record = 0; | |
7afe21cc RK |
5524 | |
5525 | /* Look within any SIGN_EXTRACT or ZERO_EXTRACT | |
5526 | to the MEM or REG within it. */ | |
5527 | while (GET_CODE (dest) == SIGN_EXTRACT | |
5528 | || GET_CODE (dest) == ZERO_EXTRACT | |
5529 | || GET_CODE (dest) == SUBREG | |
5530 | || GET_CODE (dest) == STRICT_LOW_PART) | |
0339ce7e | 5531 | dest = XEXP (dest, 0); |
7afe21cc RK |
5532 | |
5533 | sets[i].inner_dest = dest; | |
5534 | ||
3c0cb5de | 5535 | if (MEM_P (dest)) |
7afe21cc | 5536 | { |
9ae8ffe7 JL |
5537 | #ifdef PUSH_ROUNDING |
5538 | /* Stack pushes invalidate the stack pointer. */ | |
5539 | rtx addr = XEXP (dest, 0); | |
ec8e098d | 5540 | if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC |
9ae8ffe7 JL |
5541 | && XEXP (addr, 0) == stack_pointer_rtx) |
5542 | invalidate (stack_pointer_rtx, Pmode); | |
5543 | #endif | |
7afe21cc | 5544 | dest = fold_rtx (dest, insn); |
7afe21cc RK |
5545 | } |
5546 | ||
5547 | /* Compute the hash code of the destination now, | |
5548 | before the effects of this instruction are recorded, | |
5549 | since the register values used in the address computation | |
5550 | are those before this instruction. */ | |
2197a88a | 5551 | sets[i].dest_hash = HASH (dest, mode); |
7afe21cc RK |
5552 | |
5553 | /* Don't enter a bit-field in the hash table | |
5554 | because the value in it after the store | |
5555 | may not equal what was stored, due to truncation. */ | |
5556 | ||
5557 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT | |
5558 | || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT) | |
5559 | { | |
5560 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
5561 | ||
5562 | if (src_const != 0 && GET_CODE (src_const) == CONST_INT | |
5563 | && GET_CODE (width) == CONST_INT | |
906c4e36 RK |
5564 | && INTVAL (width) < HOST_BITS_PER_WIDE_INT |
5565 | && ! (INTVAL (src_const) | |
5566 | & ((HOST_WIDE_INT) (-1) << INTVAL (width)))) | |
7afe21cc RK |
5567 | /* Exception: if the value is constant, |
5568 | and it won't be truncated, record it. */ | |
5569 | ; | |
5570 | else | |
5571 | { | |
5572 | /* This is chosen so that the destination will be invalidated | |
5573 | but no new value will be recorded. | |
5574 | We must invalidate because sometimes constant | |
5575 | values can be recorded for bitfields. */ | |
5576 | sets[i].src_elt = 0; | |
5577 | sets[i].src_volatile = 1; | |
5578 | src_eqv = 0; | |
5579 | src_eqv_elt = 0; | |
5580 | } | |
5581 | } | |
5582 | ||
5583 | /* If only one set in a JUMP_INSN and it is now a no-op, we can delete | |
5584 | the insn. */ | |
5585 | else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx) | |
5586 | { | |
ef178af3 | 5587 | /* One less use of the label this insn used to jump to. */ |
49ce134f | 5588 | delete_insn (insn); |
7afe21cc | 5589 | cse_jumps_altered = 1; |
7afe21cc RK |
5590 | /* No more processing for this set. */ |
5591 | sets[i].rtl = 0; | |
5592 | } | |
5593 | ||
5594 | /* If this SET is now setting PC to a label, we know it used to | |
d466c016 | 5595 | be a conditional or computed branch. */ |
7afe21cc RK |
5596 | else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF) |
5597 | { | |
8fb1e50e GS |
5598 | /* Now emit a BARRIER after the unconditional jump. */ |
5599 | if (NEXT_INSN (insn) == 0 | |
4b4bf941 | 5600 | || !BARRIER_P (NEXT_INSN (insn))) |
8fb1e50e GS |
5601 | emit_barrier_after (insn); |
5602 | ||
d466c016 JL |
5603 | /* We reemit the jump in as many cases as possible just in |
5604 | case the form of an unconditional jump is significantly | |
5605 | different than a computed jump or conditional jump. | |
5606 | ||
5607 | If this insn has multiple sets, then reemitting the | |
5608 | jump is nontrivial. So instead we just force rerecognition | |
5609 | and hope for the best. */ | |
5610 | if (n_sets == 1) | |
7afe21cc | 5611 | { |
9dcb4381 | 5612 | rtx new, note; |
8fb1e50e | 5613 | |
9dcb4381 | 5614 | new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn); |
7afe21cc RK |
5615 | JUMP_LABEL (new) = XEXP (src, 0); |
5616 | LABEL_NUSES (XEXP (src, 0))++; | |
9dcb4381 RH |
5617 | |
5618 | /* Make sure to copy over REG_NON_LOCAL_GOTO. */ | |
5619 | note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0); | |
5620 | if (note) | |
5621 | { | |
5622 | XEXP (note, 1) = NULL_RTX; | |
5623 | REG_NOTES (new) = note; | |
5624 | } | |
5625 | ||
38c1593d | 5626 | delete_insn (insn); |
7afe21cc | 5627 | insn = new; |
8fb1e50e GS |
5628 | |
5629 | /* Now emit a BARRIER after the unconditional jump. */ | |
5630 | if (NEXT_INSN (insn) == 0 | |
4b4bf941 | 5631 | || !BARRIER_P (NEXT_INSN (insn))) |
8fb1e50e | 5632 | emit_barrier_after (insn); |
7afe21cc | 5633 | } |
31dcf83f | 5634 | else |
31dcf83f | 5635 | INSN_CODE (insn) = -1; |
7afe21cc | 5636 | |
8fb1e50e GS |
5637 | /* Do not bother deleting any unreachable code, |
5638 | let jump/flow do that. */ | |
7afe21cc RK |
5639 | |
5640 | cse_jumps_altered = 1; | |
5641 | sets[i].rtl = 0; | |
5642 | } | |
5643 | ||
c2a47e48 RK |
5644 | /* If destination is volatile, invalidate it and then do no further |
5645 | processing for this assignment. */ | |
7afe21cc RK |
5646 | |
5647 | else if (do_not_record) | |
c2a47e48 | 5648 | { |
f8cfc6aa | 5649 | if (REG_P (dest) || GET_CODE (dest) == SUBREG) |
bb4034b3 | 5650 | invalidate (dest, VOIDmode); |
3c0cb5de | 5651 | else if (MEM_P (dest)) |
bb07060a JW |
5652 | { |
5653 | /* Outgoing arguments for a libcall don't | |
5654 | affect any recorded expressions. */ | |
5655 | if (! libcall_insn || insn == libcall_insn) | |
5656 | invalidate (dest, VOIDmode); | |
5657 | } | |
2708da92 RS |
5658 | else if (GET_CODE (dest) == STRICT_LOW_PART |
5659 | || GET_CODE (dest) == ZERO_EXTRACT) | |
bb4034b3 | 5660 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
c2a47e48 RK |
5661 | sets[i].rtl = 0; |
5662 | } | |
7afe21cc RK |
5663 | |
5664 | if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl)) | |
2197a88a | 5665 | sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode); |
7afe21cc RK |
5666 | |
5667 | #ifdef HAVE_cc0 | |
5668 | /* If setting CC0, record what it was set to, or a constant, if it | |
5669 | is equivalent to a constant. If it is being set to a floating-point | |
5670 | value, make a COMPARE with the appropriate constant of 0. If we | |
5671 | don't do this, later code can interpret this as a test against | |
5672 | const0_rtx, which can cause problems if we try to put it into an | |
5673 | insn as a floating-point operand. */ | |
5674 | if (dest == cc0_rtx) | |
5675 | { | |
5676 | this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src; | |
5677 | this_insn_cc0_mode = mode; | |
cbf6a543 | 5678 | if (FLOAT_MODE_P (mode)) |
38a448ca RH |
5679 | this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0, |
5680 | CONST0_RTX (mode)); | |
7afe21cc RK |
5681 | } |
5682 | #endif | |
5683 | } | |
5684 | ||
5685 | /* Now enter all non-volatile source expressions in the hash table | |
5686 | if they are not already present. | |
5687 | Record their equivalence classes in src_elt. | |
5688 | This way we can insert the corresponding destinations into | |
5689 | the same classes even if the actual sources are no longer in them | |
5690 | (having been invalidated). */ | |
5691 | ||
5692 | if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile | |
5693 | && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl))) | |
5694 | { | |
b3694847 SS |
5695 | struct table_elt *elt; |
5696 | struct table_elt *classp = sets[0].src_elt; | |
7afe21cc RK |
5697 | rtx dest = SET_DEST (sets[0].rtl); |
5698 | enum machine_mode eqvmode = GET_MODE (dest); | |
5699 | ||
5700 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
5701 | { | |
5702 | eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0))); | |
5703 | classp = 0; | |
5704 | } | |
5705 | if (insert_regs (src_eqv, classp, 0)) | |
8ae2b8f6 JW |
5706 | { |
5707 | rehash_using_reg (src_eqv); | |
5708 | src_eqv_hash = HASH (src_eqv, eqvmode); | |
5709 | } | |
2197a88a | 5710 | elt = insert (src_eqv, classp, src_eqv_hash, eqvmode); |
7afe21cc | 5711 | elt->in_memory = src_eqv_in_memory; |
7afe21cc | 5712 | src_eqv_elt = elt; |
f7911249 JW |
5713 | |
5714 | /* Check to see if src_eqv_elt is the same as a set source which | |
5715 | does not yet have an elt, and if so set the elt of the set source | |
5716 | to src_eqv_elt. */ | |
5717 | for (i = 0; i < n_sets; i++) | |
26132f71 JW |
5718 | if (sets[i].rtl && sets[i].src_elt == 0 |
5719 | && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv)) | |
f7911249 | 5720 | sets[i].src_elt = src_eqv_elt; |
7afe21cc RK |
5721 | } |
5722 | ||
5723 | for (i = 0; i < n_sets; i++) | |
5724 | if (sets[i].rtl && ! sets[i].src_volatile | |
5725 | && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl))) | |
5726 | { | |
5727 | if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART) | |
5728 | { | |
5729 | /* REG_EQUAL in setting a STRICT_LOW_PART | |
5730 | gives an equivalent for the entire destination register, | |
5731 | not just for the subreg being stored in now. | |
5732 | This is a more interesting equivalence, so we arrange later | |
5733 | to treat the entire reg as the destination. */ | |
5734 | sets[i].src_elt = src_eqv_elt; | |
2197a88a | 5735 | sets[i].src_hash = src_eqv_hash; |
7afe21cc RK |
5736 | } |
5737 | else | |
5738 | { | |
5739 | /* Insert source and constant equivalent into hash table, if not | |
5740 | already present. */ | |
b3694847 SS |
5741 | struct table_elt *classp = src_eqv_elt; |
5742 | rtx src = sets[i].src; | |
5743 | rtx dest = SET_DEST (sets[i].rtl); | |
7afe21cc RK |
5744 | enum machine_mode mode |
5745 | = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src); | |
5746 | ||
1fcc57f1 AM |
5747 | /* It's possible that we have a source value known to be |
5748 | constant but don't have a REG_EQUAL note on the insn. | |
5749 | Lack of a note will mean src_eqv_elt will be NULL. This | |
5750 | can happen where we've generated a SUBREG to access a | |
5751 | CONST_INT that is already in a register in a wider mode. | |
5752 | Ensure that the source expression is put in the proper | |
5753 | constant class. */ | |
5754 | if (!classp) | |
5755 | classp = sets[i].src_const_elt; | |
5756 | ||
26132f71 | 5757 | if (sets[i].src_elt == 0) |
7afe21cc | 5758 | { |
26132f71 JW |
5759 | /* Don't put a hard register source into the table if this is |
5760 | the last insn of a libcall. In this case, we only need | |
5761 | to put src_eqv_elt in src_elt. */ | |
db4a8254 | 5762 | if (! find_reg_note (insn, REG_RETVAL, NULL_RTX)) |
8ae2b8f6 | 5763 | { |
b3694847 | 5764 | struct table_elt *elt; |
26132f71 JW |
5765 | |
5766 | /* Note that these insert_regs calls cannot remove | |
5767 | any of the src_elt's, because they would have failed to | |
5768 | match if not still valid. */ | |
5769 | if (insert_regs (src, classp, 0)) | |
5770 | { | |
5771 | rehash_using_reg (src); | |
5772 | sets[i].src_hash = HASH (src, mode); | |
5773 | } | |
5774 | elt = insert (src, classp, sets[i].src_hash, mode); | |
5775 | elt->in_memory = sets[i].src_in_memory; | |
26132f71 | 5776 | sets[i].src_elt = classp = elt; |
8ae2b8f6 | 5777 | } |
26132f71 JW |
5778 | else |
5779 | sets[i].src_elt = classp; | |
7afe21cc | 5780 | } |
7afe21cc RK |
5781 | if (sets[i].src_const && sets[i].src_const_elt == 0 |
5782 | && src != sets[i].src_const | |
5783 | && ! rtx_equal_p (sets[i].src_const, src)) | |
5784 | sets[i].src_elt = insert (sets[i].src_const, classp, | |
2197a88a | 5785 | sets[i].src_const_hash, mode); |
7afe21cc RK |
5786 | } |
5787 | } | |
5788 | else if (sets[i].src_elt == 0) | |
5789 | /* If we did not insert the source into the hash table (e.g., it was | |
5790 | volatile), note the equivalence class for the REG_EQUAL value, if any, | |
5791 | so that the destination goes into that class. */ | |
5792 | sets[i].src_elt = src_eqv_elt; | |
5793 | ||
9ae8ffe7 | 5794 | invalidate_from_clobbers (x); |
77fa0940 | 5795 | |
278a83b2 | 5796 | /* Some registers are invalidated by subroutine calls. Memory is |
77fa0940 RK |
5797 | invalidated by non-constant calls. */ |
5798 | ||
4b4bf941 | 5799 | if (CALL_P (insn)) |
7afe21cc | 5800 | { |
24a28584 | 5801 | if (! CONST_OR_PURE_CALL_P (insn)) |
9ae8ffe7 | 5802 | invalidate_memory (); |
7afe21cc RK |
5803 | invalidate_for_call (); |
5804 | } | |
5805 | ||
5806 | /* Now invalidate everything set by this instruction. | |
5807 | If a SUBREG or other funny destination is being set, | |
5808 | sets[i].rtl is still nonzero, so here we invalidate the reg | |
5809 | a part of which is being set. */ | |
5810 | ||
5811 | for (i = 0; i < n_sets; i++) | |
5812 | if (sets[i].rtl) | |
5813 | { | |
bb4034b3 JW |
5814 | /* We can't use the inner dest, because the mode associated with |
5815 | a ZERO_EXTRACT is significant. */ | |
b3694847 | 5816 | rtx dest = SET_DEST (sets[i].rtl); |
7afe21cc RK |
5817 | |
5818 | /* Needed for registers to remove the register from its | |
5819 | previous quantity's chain. | |
5820 | Needed for memory if this is a nonvarying address, unless | |
5821 | we have just done an invalidate_memory that covers even those. */ | |
f8cfc6aa | 5822 | if (REG_P (dest) || GET_CODE (dest) == SUBREG) |
bb4034b3 | 5823 | invalidate (dest, VOIDmode); |
3c0cb5de | 5824 | else if (MEM_P (dest)) |
bb07060a JW |
5825 | { |
5826 | /* Outgoing arguments for a libcall don't | |
5827 | affect any recorded expressions. */ | |
5828 | if (! libcall_insn || insn == libcall_insn) | |
5829 | invalidate (dest, VOIDmode); | |
5830 | } | |
2708da92 RS |
5831 | else if (GET_CODE (dest) == STRICT_LOW_PART |
5832 | || GET_CODE (dest) == ZERO_EXTRACT) | |
bb4034b3 | 5833 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
7afe21cc RK |
5834 | } |
5835 | ||
01e752d3 | 5836 | /* A volatile ASM invalidates everything. */ |
4b4bf941 | 5837 | if (NONJUMP_INSN_P (insn) |
01e752d3 JL |
5838 | && GET_CODE (PATTERN (insn)) == ASM_OPERANDS |
5839 | && MEM_VOLATILE_P (PATTERN (insn))) | |
5840 | flush_hash_table (); | |
5841 | ||
7afe21cc RK |
5842 | /* Make sure registers mentioned in destinations |
5843 | are safe for use in an expression to be inserted. | |
5844 | This removes from the hash table | |
5845 | any invalid entry that refers to one of these registers. | |
5846 | ||
5847 | We don't care about the return value from mention_regs because | |
5848 | we are going to hash the SET_DEST values unconditionally. */ | |
5849 | ||
5850 | for (i = 0; i < n_sets; i++) | |
34c73909 R |
5851 | { |
5852 | if (sets[i].rtl) | |
5853 | { | |
5854 | rtx x = SET_DEST (sets[i].rtl); | |
5855 | ||
f8cfc6aa | 5856 | if (!REG_P (x)) |
34c73909 R |
5857 | mention_regs (x); |
5858 | else | |
5859 | { | |
5860 | /* We used to rely on all references to a register becoming | |
5861 | inaccessible when a register changes to a new quantity, | |
5862 | since that changes the hash code. However, that is not | |
9b1549b8 | 5863 | safe, since after HASH_SIZE new quantities we get a |
34c73909 R |
5864 | hash 'collision' of a register with its own invalid |
5865 | entries. And since SUBREGs have been changed not to | |
5866 | change their hash code with the hash code of the register, | |
5867 | it wouldn't work any longer at all. So we have to check | |
5868 | for any invalid references lying around now. | |
5869 | This code is similar to the REG case in mention_regs, | |
5870 | but it knows that reg_tick has been incremented, and | |
5871 | it leaves reg_in_table as -1 . */ | |
770ae6cc RK |
5872 | unsigned int regno = REGNO (x); |
5873 | unsigned int endregno | |
34c73909 | 5874 | = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1 |
66fd46b6 | 5875 | : hard_regno_nregs[regno][GET_MODE (x)]); |
770ae6cc | 5876 | unsigned int i; |
34c73909 R |
5877 | |
5878 | for (i = regno; i < endregno; i++) | |
5879 | { | |
30f72379 | 5880 | if (REG_IN_TABLE (i) >= 0) |
34c73909 R |
5881 | { |
5882 | remove_invalid_refs (i); | |
30f72379 | 5883 | REG_IN_TABLE (i) = -1; |
34c73909 R |
5884 | } |
5885 | } | |
5886 | } | |
5887 | } | |
5888 | } | |
7afe21cc RK |
5889 | |
5890 | /* We may have just removed some of the src_elt's from the hash table. | |
5891 | So replace each one with the current head of the same class. */ | |
5892 | ||
5893 | for (i = 0; i < n_sets; i++) | |
5894 | if (sets[i].rtl) | |
5895 | { | |
5896 | if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0) | |
5897 | /* If elt was removed, find current head of same class, | |
5898 | or 0 if nothing remains of that class. */ | |
5899 | { | |
b3694847 | 5900 | struct table_elt *elt = sets[i].src_elt; |
7afe21cc RK |
5901 | |
5902 | while (elt && elt->prev_same_value) | |
5903 | elt = elt->prev_same_value; | |
5904 | ||
5905 | while (elt && elt->first_same_value == 0) | |
5906 | elt = elt->next_same_value; | |
5907 | sets[i].src_elt = elt ? elt->first_same_value : 0; | |
5908 | } | |
5909 | } | |
5910 | ||
5911 | /* Now insert the destinations into their equivalence classes. */ | |
5912 | ||
5913 | for (i = 0; i < n_sets; i++) | |
5914 | if (sets[i].rtl) | |
5915 | { | |
b3694847 | 5916 | rtx dest = SET_DEST (sets[i].rtl); |
b3694847 | 5917 | struct table_elt *elt; |
7afe21cc RK |
5918 | |
5919 | /* Don't record value if we are not supposed to risk allocating | |
5920 | floating-point values in registers that might be wider than | |
5921 | memory. */ | |
5922 | if ((flag_float_store | |
3c0cb5de | 5923 | && MEM_P (dest) |
cbf6a543 | 5924 | && FLOAT_MODE_P (GET_MODE (dest))) |
bc4ddc77 JW |
5925 | /* Don't record BLKmode values, because we don't know the |
5926 | size of it, and can't be sure that other BLKmode values | |
5927 | have the same or smaller size. */ | |
5928 | || GET_MODE (dest) == BLKmode | |
7afe21cc RK |
5929 | /* Don't record values of destinations set inside a libcall block |
5930 | since we might delete the libcall. Things should have been set | |
5931 | up so we won't want to reuse such a value, but we play it safe | |
5932 | here. */ | |
7bd8b2a8 | 5933 | || libcall_insn |
7afe21cc RK |
5934 | /* If we didn't put a REG_EQUAL value or a source into the hash |
5935 | table, there is no point is recording DEST. */ | |
1a8e9a8e RK |
5936 | || sets[i].src_elt == 0 |
5937 | /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND | |
5938 | or SIGN_EXTEND, don't record DEST since it can cause | |
5939 | some tracking to be wrong. | |
5940 | ||
5941 | ??? Think about this more later. */ | |
5942 | || (GET_CODE (dest) == SUBREG | |
5943 | && (GET_MODE_SIZE (GET_MODE (dest)) | |
5944 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))) | |
5945 | && (GET_CODE (sets[i].src) == SIGN_EXTEND | |
5946 | || GET_CODE (sets[i].src) == ZERO_EXTEND))) | |
7afe21cc RK |
5947 | continue; |
5948 | ||
5949 | /* STRICT_LOW_PART isn't part of the value BEING set, | |
5950 | and neither is the SUBREG inside it. | |
5951 | Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */ | |
5952 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
5953 | dest = SUBREG_REG (XEXP (dest, 0)); | |
5954 | ||
f8cfc6aa | 5955 | if (REG_P (dest) || GET_CODE (dest) == SUBREG) |
7afe21cc RK |
5956 | /* Registers must also be inserted into chains for quantities. */ |
5957 | if (insert_regs (dest, sets[i].src_elt, 1)) | |
8ae2b8f6 JW |
5958 | { |
5959 | /* If `insert_regs' changes something, the hash code must be | |
5960 | recalculated. */ | |
5961 | rehash_using_reg (dest); | |
5962 | sets[i].dest_hash = HASH (dest, GET_MODE (dest)); | |
5963 | } | |
7afe21cc | 5964 | |
8fff4fc1 RH |
5965 | elt = insert (dest, sets[i].src_elt, |
5966 | sets[i].dest_hash, GET_MODE (dest)); | |
9de2c71a | 5967 | |
3c0cb5de | 5968 | elt->in_memory = (MEM_P (sets[i].inner_dest) |
9ad91d71 | 5969 | && (! RTX_UNCHANGING_P (sets[i].inner_dest) |
4977bab6 | 5970 | || fixed_base_plus_p (XEXP (sets[i].inner_dest, |
9ad91d71 | 5971 | 0)))); |
c256df0b | 5972 | |
fc3ffe83 RK |
5973 | /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no |
5974 | narrower than M2, and both M1 and M2 are the same number of words, | |
5975 | we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so | |
5976 | make that equivalence as well. | |
7afe21cc | 5977 | |
4de249d9 PB |
5978 | However, BAR may have equivalences for which gen_lowpart |
5979 | will produce a simpler value than gen_lowpart applied to | |
7afe21cc | 5980 | BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all |
278a83b2 | 5981 | BAR's equivalences. If we don't get a simplified form, make |
7afe21cc RK |
5982 | the SUBREG. It will not be used in an equivalence, but will |
5983 | cause two similar assignments to be detected. | |
5984 | ||
5985 | Note the loop below will find SUBREG_REG (DEST) since we have | |
5986 | already entered SRC and DEST of the SET in the table. */ | |
5987 | ||
5988 | if (GET_CODE (dest) == SUBREG | |
6cdbaec4 RK |
5989 | && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1) |
5990 | / UNITS_PER_WORD) | |
278a83b2 | 5991 | == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD) |
7afe21cc RK |
5992 | && (GET_MODE_SIZE (GET_MODE (dest)) |
5993 | >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))) | |
5994 | && sets[i].src_elt != 0) | |
5995 | { | |
5996 | enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest)); | |
5997 | struct table_elt *elt, *classp = 0; | |
5998 | ||
5999 | for (elt = sets[i].src_elt->first_same_value; elt; | |
6000 | elt = elt->next_same_value) | |
6001 | { | |
6002 | rtx new_src = 0; | |
2197a88a | 6003 | unsigned src_hash; |
7afe21cc | 6004 | struct table_elt *src_elt; |
ff27a429 | 6005 | int byte = 0; |
7afe21cc RK |
6006 | |
6007 | /* Ignore invalid entries. */ | |
f8cfc6aa | 6008 | if (!REG_P (elt->exp) |
7afe21cc RK |
6009 | && ! exp_equiv_p (elt->exp, elt->exp, 1, 0)) |
6010 | continue; | |
6011 | ||
9beb7d20 RH |
6012 | /* We may have already been playing subreg games. If the |
6013 | mode is already correct for the destination, use it. */ | |
6014 | if (GET_MODE (elt->exp) == new_mode) | |
6015 | new_src = elt->exp; | |
6016 | else | |
6017 | { | |
6018 | /* Calculate big endian correction for the SUBREG_BYTE. | |
6019 | We have already checked that M1 (GET_MODE (dest)) | |
6020 | is not narrower than M2 (new_mode). */ | |
6021 | if (BYTES_BIG_ENDIAN) | |
6022 | byte = (GET_MODE_SIZE (GET_MODE (dest)) | |
6023 | - GET_MODE_SIZE (new_mode)); | |
6024 | ||
6025 | new_src = simplify_gen_subreg (new_mode, elt->exp, | |
6026 | GET_MODE (dest), byte); | |
6027 | } | |
6028 | ||
ff27a429 R |
6029 | /* The call to simplify_gen_subreg fails if the value |
6030 | is VOIDmode, yet we can't do any simplification, e.g. | |
6031 | for EXPR_LISTs denoting function call results. | |
6032 | It is invalid to construct a SUBREG with a VOIDmode | |
6033 | SUBREG_REG, hence a zero new_src means we can't do | |
6034 | this substitution. */ | |
6035 | if (! new_src) | |
6036 | continue; | |
7afe21cc RK |
6037 | |
6038 | src_hash = HASH (new_src, new_mode); | |
6039 | src_elt = lookup (new_src, src_hash, new_mode); | |
6040 | ||
6041 | /* Put the new source in the hash table is if isn't | |
6042 | already. */ | |
6043 | if (src_elt == 0) | |
6044 | { | |
6045 | if (insert_regs (new_src, classp, 0)) | |
8ae2b8f6 JW |
6046 | { |
6047 | rehash_using_reg (new_src); | |
6048 | src_hash = HASH (new_src, new_mode); | |
6049 | } | |
7afe21cc RK |
6050 | src_elt = insert (new_src, classp, src_hash, new_mode); |
6051 | src_elt->in_memory = elt->in_memory; | |
7afe21cc RK |
6052 | } |
6053 | else if (classp && classp != src_elt->first_same_value) | |
278a83b2 | 6054 | /* Show that two things that we've seen before are |
7afe21cc RK |
6055 | actually the same. */ |
6056 | merge_equiv_classes (src_elt, classp); | |
6057 | ||
6058 | classp = src_elt->first_same_value; | |
da932f04 JL |
6059 | /* Ignore invalid entries. */ |
6060 | while (classp | |
f8cfc6aa | 6061 | && !REG_P (classp->exp) |
da932f04 JL |
6062 | && ! exp_equiv_p (classp->exp, classp->exp, 1, 0)) |
6063 | classp = classp->next_same_value; | |
7afe21cc RK |
6064 | } |
6065 | } | |
6066 | } | |
6067 | ||
403e25d0 RK |
6068 | /* Special handling for (set REG0 REG1) where REG0 is the |
6069 | "cheapest", cheaper than REG1. After cse, REG1 will probably not | |
6070 | be used in the sequel, so (if easily done) change this insn to | |
6071 | (set REG1 REG0) and replace REG1 with REG0 in the previous insn | |
6072 | that computed their value. Then REG1 will become a dead store | |
6073 | and won't cloud the situation for later optimizations. | |
7afe21cc RK |
6074 | |
6075 | Do not make this change if REG1 is a hard register, because it will | |
6076 | then be used in the sequel and we may be changing a two-operand insn | |
6077 | into a three-operand insn. | |
6078 | ||
50270076 R |
6079 | Also do not do this if we are operating on a copy of INSN. |
6080 | ||
6081 | Also don't do this if INSN ends a libcall; this would cause an unrelated | |
6082 | register to be set in the middle of a libcall, and we then get bad code | |
6083 | if the libcall is deleted. */ | |
7afe21cc | 6084 | |
f8cfc6aa | 6085 | if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl)) |
7afe21cc | 6086 | && NEXT_INSN (PREV_INSN (insn)) == insn |
f8cfc6aa | 6087 | && REG_P (SET_SRC (sets[0].rtl)) |
7afe21cc | 6088 | && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER |
1bb98cec | 6089 | && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl)))) |
7afe21cc | 6090 | { |
1bb98cec DM |
6091 | int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl))); |
6092 | struct qty_table_elem *src_ent = &qty_table[src_q]; | |
7afe21cc | 6093 | |
1bb98cec DM |
6094 | if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl))) |
6095 | && ! find_reg_note (insn, REG_RETVAL, NULL_RTX)) | |
7afe21cc | 6096 | { |
3e25353e AH |
6097 | rtx prev = insn; |
6098 | /* Scan for the previous nonnote insn, but stop at a basic | |
6099 | block boundary. */ | |
6100 | do | |
6101 | { | |
6102 | prev = PREV_INSN (prev); | |
6103 | } | |
4b4bf941 | 6104 | while (prev && NOTE_P (prev) |
3e25353e | 6105 | && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK); |
7080f735 | 6106 | |
58ecb5e2 RS |
6107 | /* Do not swap the registers around if the previous instruction |
6108 | attaches a REG_EQUIV note to REG1. | |
6109 | ||
6110 | ??? It's not entirely clear whether we can transfer a REG_EQUIV | |
6111 | from the pseudo that originally shadowed an incoming argument | |
6112 | to another register. Some uses of REG_EQUIV might rely on it | |
6113 | being attached to REG1 rather than REG2. | |
6114 | ||
6115 | This section previously turned the REG_EQUIV into a REG_EQUAL | |
6116 | note. We cannot do that because REG_EQUIV may provide an | |
4912a07c | 6117 | uninitialized stack slot when REG_PARM_STACK_SPACE is used. */ |
58ecb5e2 | 6118 | |
4b4bf941 | 6119 | if (prev != 0 && NONJUMP_INSN_P (prev) |
403e25d0 | 6120 | && GET_CODE (PATTERN (prev)) == SET |
58ecb5e2 RS |
6121 | && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl) |
6122 | && ! find_reg_note (prev, REG_EQUIV, NULL_RTX)) | |
1bb98cec DM |
6123 | { |
6124 | rtx dest = SET_DEST (sets[0].rtl); | |
403e25d0 | 6125 | rtx src = SET_SRC (sets[0].rtl); |
58ecb5e2 | 6126 | rtx note; |
7afe21cc | 6127 | |
278a83b2 KH |
6128 | validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1); |
6129 | validate_change (insn, &SET_DEST (sets[0].rtl), src, 1); | |
6130 | validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1); | |
1bb98cec | 6131 | apply_change_group (); |
7afe21cc | 6132 | |
403e25d0 RK |
6133 | /* If INSN has a REG_EQUAL note, and this note mentions |
6134 | REG0, then we must delete it, because the value in | |
6135 | REG0 has changed. If the note's value is REG1, we must | |
6136 | also delete it because that is now this insn's dest. */ | |
1bb98cec | 6137 | note = find_reg_note (insn, REG_EQUAL, NULL_RTX); |
403e25d0 RK |
6138 | if (note != 0 |
6139 | && (reg_mentioned_p (dest, XEXP (note, 0)) | |
6140 | || rtx_equal_p (src, XEXP (note, 0)))) | |
1bb98cec DM |
6141 | remove_note (insn, note); |
6142 | } | |
7afe21cc RK |
6143 | } |
6144 | } | |
6145 | ||
6146 | /* If this is a conditional jump insn, record any known equivalences due to | |
6147 | the condition being tested. */ | |
6148 | ||
6149 | last_jump_equiv_class = 0; | |
4b4bf941 | 6150 | if (JUMP_P (insn) |
7afe21cc RK |
6151 | && n_sets == 1 && GET_CODE (x) == SET |
6152 | && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE) | |
6153 | record_jump_equiv (insn, 0); | |
6154 | ||
6155 | #ifdef HAVE_cc0 | |
6156 | /* If the previous insn set CC0 and this insn no longer references CC0, | |
6157 | delete the previous insn. Here we use the fact that nothing expects CC0 | |
6158 | to be valid over an insn, which is true until the final pass. */ | |
4b4bf941 | 6159 | if (prev_insn && NONJUMP_INSN_P (prev_insn) |
7afe21cc RK |
6160 | && (tem = single_set (prev_insn)) != 0 |
6161 | && SET_DEST (tem) == cc0_rtx | |
6162 | && ! reg_mentioned_p (cc0_rtx, x)) | |
6dee7384 | 6163 | delete_insn (prev_insn); |
7afe21cc RK |
6164 | |
6165 | prev_insn_cc0 = this_insn_cc0; | |
6166 | prev_insn_cc0_mode = this_insn_cc0_mode; | |
7afe21cc | 6167 | prev_insn = insn; |
4977bab6 | 6168 | #endif |
7afe21cc RK |
6169 | } |
6170 | \f | |
a4c6502a | 6171 | /* Remove from the hash table all expressions that reference memory. */ |
14a774a9 | 6172 | |
7afe21cc | 6173 | static void |
7080f735 | 6174 | invalidate_memory (void) |
7afe21cc | 6175 | { |
b3694847 SS |
6176 | int i; |
6177 | struct table_elt *p, *next; | |
7afe21cc | 6178 | |
9b1549b8 | 6179 | for (i = 0; i < HASH_SIZE; i++) |
9ae8ffe7 JL |
6180 | for (p = table[i]; p; p = next) |
6181 | { | |
6182 | next = p->next_same_hash; | |
6183 | if (p->in_memory) | |
6184 | remove_from_table (p, i); | |
6185 | } | |
6186 | } | |
6187 | ||
14a774a9 RK |
6188 | /* If ADDR is an address that implicitly affects the stack pointer, return |
6189 | 1 and update the register tables to show the effect. Else, return 0. */ | |
6190 | ||
9ae8ffe7 | 6191 | static int |
7080f735 | 6192 | addr_affects_sp_p (rtx addr) |
9ae8ffe7 | 6193 | { |
ec8e098d | 6194 | if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC |
f8cfc6aa | 6195 | && REG_P (XEXP (addr, 0)) |
9ae8ffe7 | 6196 | && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM) |
7afe21cc | 6197 | { |
30f72379 | 6198 | if (REG_TICK (STACK_POINTER_REGNUM) >= 0) |
46081bb3 SH |
6199 | { |
6200 | REG_TICK (STACK_POINTER_REGNUM)++; | |
6201 | /* Is it possible to use a subreg of SP? */ | |
6202 | SUBREG_TICKED (STACK_POINTER_REGNUM) = -1; | |
6203 | } | |
9ae8ffe7 JL |
6204 | |
6205 | /* This should be *very* rare. */ | |
6206 | if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM)) | |
6207 | invalidate (stack_pointer_rtx, VOIDmode); | |
14a774a9 | 6208 | |
9ae8ffe7 | 6209 | return 1; |
7afe21cc | 6210 | } |
14a774a9 | 6211 | |
9ae8ffe7 | 6212 | return 0; |
7afe21cc RK |
6213 | } |
6214 | ||
6215 | /* Perform invalidation on the basis of everything about an insn | |
6216 | except for invalidating the actual places that are SET in it. | |
6217 | This includes the places CLOBBERed, and anything that might | |
6218 | alias with something that is SET or CLOBBERed. | |
6219 | ||
7afe21cc RK |
6220 | X is the pattern of the insn. */ |
6221 | ||
6222 | static void | |
7080f735 | 6223 | invalidate_from_clobbers (rtx x) |
7afe21cc | 6224 | { |
7afe21cc RK |
6225 | if (GET_CODE (x) == CLOBBER) |
6226 | { | |
6227 | rtx ref = XEXP (x, 0); | |
9ae8ffe7 JL |
6228 | if (ref) |
6229 | { | |
f8cfc6aa | 6230 | if (REG_P (ref) || GET_CODE (ref) == SUBREG |
3c0cb5de | 6231 | || MEM_P (ref)) |
9ae8ffe7 JL |
6232 | invalidate (ref, VOIDmode); |
6233 | else if (GET_CODE (ref) == STRICT_LOW_PART | |
6234 | || GET_CODE (ref) == ZERO_EXTRACT) | |
6235 | invalidate (XEXP (ref, 0), GET_MODE (ref)); | |
6236 | } | |
7afe21cc RK |
6237 | } |
6238 | else if (GET_CODE (x) == PARALLEL) | |
6239 | { | |
b3694847 | 6240 | int i; |
7afe21cc RK |
6241 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) |
6242 | { | |
b3694847 | 6243 | rtx y = XVECEXP (x, 0, i); |
7afe21cc RK |
6244 | if (GET_CODE (y) == CLOBBER) |
6245 | { | |
6246 | rtx ref = XEXP (y, 0); | |
f8cfc6aa | 6247 | if (REG_P (ref) || GET_CODE (ref) == SUBREG |
3c0cb5de | 6248 | || MEM_P (ref)) |
9ae8ffe7 JL |
6249 | invalidate (ref, VOIDmode); |
6250 | else if (GET_CODE (ref) == STRICT_LOW_PART | |
6251 | || GET_CODE (ref) == ZERO_EXTRACT) | |
6252 | invalidate (XEXP (ref, 0), GET_MODE (ref)); | |
7afe21cc RK |
6253 | } |
6254 | } | |
6255 | } | |
6256 | } | |
6257 | \f | |
6258 | /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes | |
6259 | and replace any registers in them with either an equivalent constant | |
6260 | or the canonical form of the register. If we are inside an address, | |
6261 | only do this if the address remains valid. | |
6262 | ||
6263 | OBJECT is 0 except when within a MEM in which case it is the MEM. | |
6264 | ||
6265 | Return the replacement for X. */ | |
6266 | ||
6267 | static rtx | |
7080f735 | 6268 | cse_process_notes (rtx x, rtx object) |
7afe21cc RK |
6269 | { |
6270 | enum rtx_code code = GET_CODE (x); | |
6f7d635c | 6271 | const char *fmt = GET_RTX_FORMAT (code); |
7afe21cc RK |
6272 | int i; |
6273 | ||
6274 | switch (code) | |
6275 | { | |
6276 | case CONST_INT: | |
6277 | case CONST: | |
6278 | case SYMBOL_REF: | |
6279 | case LABEL_REF: | |
6280 | case CONST_DOUBLE: | |
69ef87e2 | 6281 | case CONST_VECTOR: |
7afe21cc RK |
6282 | case PC: |
6283 | case CC0: | |
6284 | case LO_SUM: | |
6285 | return x; | |
6286 | ||
6287 | case MEM: | |
c96208fa DC |
6288 | validate_change (x, &XEXP (x, 0), |
6289 | cse_process_notes (XEXP (x, 0), x), 0); | |
7afe21cc RK |
6290 | return x; |
6291 | ||
6292 | case EXPR_LIST: | |
6293 | case INSN_LIST: | |
6294 | if (REG_NOTE_KIND (x) == REG_EQUAL) | |
906c4e36 | 6295 | XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX); |
7afe21cc | 6296 | if (XEXP (x, 1)) |
906c4e36 | 6297 | XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX); |
7afe21cc RK |
6298 | return x; |
6299 | ||
e4890d45 RS |
6300 | case SIGN_EXTEND: |
6301 | case ZERO_EXTEND: | |
0b0ee36c | 6302 | case SUBREG: |
e4890d45 RS |
6303 | { |
6304 | rtx new = cse_process_notes (XEXP (x, 0), object); | |
6305 | /* We don't substitute VOIDmode constants into these rtx, | |
6306 | since they would impede folding. */ | |
6307 | if (GET_MODE (new) != VOIDmode) | |
6308 | validate_change (object, &XEXP (x, 0), new, 0); | |
6309 | return x; | |
6310 | } | |
6311 | ||
7afe21cc | 6312 | case REG: |
30f72379 | 6313 | i = REG_QTY (REGNO (x)); |
7afe21cc RK |
6314 | |
6315 | /* Return a constant or a constant register. */ | |
1bb98cec | 6316 | if (REGNO_QTY_VALID_P (REGNO (x))) |
7afe21cc | 6317 | { |
1bb98cec DM |
6318 | struct qty_table_elem *ent = &qty_table[i]; |
6319 | ||
6320 | if (ent->const_rtx != NULL_RTX | |
6321 | && (CONSTANT_P (ent->const_rtx) | |
f8cfc6aa | 6322 | || REG_P (ent->const_rtx))) |
1bb98cec | 6323 | { |
4de249d9 | 6324 | rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx); |
1bb98cec DM |
6325 | if (new) |
6326 | return new; | |
6327 | } | |
7afe21cc RK |
6328 | } |
6329 | ||
6330 | /* Otherwise, canonicalize this register. */ | |
906c4e36 | 6331 | return canon_reg (x, NULL_RTX); |
278a83b2 | 6332 | |
e9a25f70 JL |
6333 | default: |
6334 | break; | |
7afe21cc RK |
6335 | } |
6336 | ||
6337 | for (i = 0; i < GET_RTX_LENGTH (code); i++) | |
6338 | if (fmt[i] == 'e') | |
6339 | validate_change (object, &XEXP (x, i), | |
7fe34fdf | 6340 | cse_process_notes (XEXP (x, i), object), 0); |
7afe21cc RK |
6341 | |
6342 | return x; | |
6343 | } | |
6344 | \f | |
6345 | /* Find common subexpressions between the end test of a loop and the beginning | |
6346 | of the loop. LOOP_START is the CODE_LABEL at the start of a loop. | |
6347 | ||
6348 | Often we have a loop where an expression in the exit test is used | |
6349 | in the body of the loop. For example "while (*p) *q++ = *p++;". | |
6350 | Because of the way we duplicate the loop exit test in front of the loop, | |
6351 | however, we don't detect that common subexpression. This will be caught | |
6352 | when global cse is implemented, but this is a quite common case. | |
6353 | ||
6354 | This function handles the most common cases of these common expressions. | |
6355 | It is called after we have processed the basic block ending with the | |
6356 | NOTE_INSN_LOOP_END note that ends a loop and the previous JUMP_INSN | |
6357 | jumps to a label used only once. */ | |
6358 | ||
6359 | static void | |
7080f735 | 6360 | cse_around_loop (rtx loop_start) |
7afe21cc RK |
6361 | { |
6362 | rtx insn; | |
6363 | int i; | |
6364 | struct table_elt *p; | |
6365 | ||
6366 | /* If the jump at the end of the loop doesn't go to the start, we don't | |
6367 | do anything. */ | |
6368 | for (insn = PREV_INSN (loop_start); | |
4b4bf941 | 6369 | insn && (NOTE_P (insn) && NOTE_LINE_NUMBER (insn) >= 0); |
7afe21cc RK |
6370 | insn = PREV_INSN (insn)) |
6371 | ; | |
6372 | ||
6373 | if (insn == 0 | |
4b4bf941 | 6374 | || !NOTE_P (insn) |
7afe21cc RK |
6375 | || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG) |
6376 | return; | |
6377 | ||
6378 | /* If the last insn of the loop (the end test) was an NE comparison, | |
6379 | we will interpret it as an EQ comparison, since we fell through | |
f72aed24 | 6380 | the loop. Any equivalences resulting from that comparison are |
7afe21cc RK |
6381 | therefore not valid and must be invalidated. */ |
6382 | if (last_jump_equiv_class) | |
6383 | for (p = last_jump_equiv_class->first_same_value; p; | |
6384 | p = p->next_same_value) | |
51723711 | 6385 | { |
f8cfc6aa | 6386 | if (MEM_P (p->exp) || REG_P (p->exp) |
51723711 | 6387 | || (GET_CODE (p->exp) == SUBREG |
f8cfc6aa | 6388 | && REG_P (SUBREG_REG (p->exp)))) |
51723711 | 6389 | invalidate (p->exp, VOIDmode); |
278a83b2 KH |
6390 | else if (GET_CODE (p->exp) == STRICT_LOW_PART |
6391 | || GET_CODE (p->exp) == ZERO_EXTRACT) | |
51723711 KG |
6392 | invalidate (XEXP (p->exp, 0), GET_MODE (p->exp)); |
6393 | } | |
7afe21cc RK |
6394 | |
6395 | /* Process insns starting after LOOP_START until we hit a CALL_INSN or | |
6396 | a CODE_LABEL (we could handle a CALL_INSN, but it isn't worth it). | |
6397 | ||
6398 | The only thing we do with SET_DEST is invalidate entries, so we | |
6399 | can safely process each SET in order. It is slightly less efficient | |
556c714b JW |
6400 | to do so, but we only want to handle the most common cases. |
6401 | ||
6402 | The gen_move_insn call in cse_set_around_loop may create new pseudos. | |
6403 | These pseudos won't have valid entries in any of the tables indexed | |
6404 | by register number, such as reg_qty. We avoid out-of-range array | |
6405 | accesses by not processing any instructions created after cse started. */ | |
7afe21cc RK |
6406 | |
6407 | for (insn = NEXT_INSN (loop_start); | |
4b4bf941 | 6408 | !CALL_P (insn) && !LABEL_P (insn) |
556c714b | 6409 | && INSN_UID (insn) < max_insn_uid |
4b4bf941 | 6410 | && ! (NOTE_P (insn) |
7afe21cc RK |
6411 | && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END); |
6412 | insn = NEXT_INSN (insn)) | |
6413 | { | |
2c3c49de | 6414 | if (INSN_P (insn) |
7afe21cc RK |
6415 | && (GET_CODE (PATTERN (insn)) == SET |
6416 | || GET_CODE (PATTERN (insn)) == CLOBBER)) | |
6417 | cse_set_around_loop (PATTERN (insn), insn, loop_start); | |
2c3c49de | 6418 | else if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == PARALLEL) |
7afe21cc RK |
6419 | for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) |
6420 | if (GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == SET | |
6421 | || GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == CLOBBER) | |
6422 | cse_set_around_loop (XVECEXP (PATTERN (insn), 0, i), insn, | |
6423 | loop_start); | |
6424 | } | |
6425 | } | |
6426 | \f | |
8b3686ed RK |
6427 | /* Process one SET of an insn that was skipped. We ignore CLOBBERs |
6428 | since they are done elsewhere. This function is called via note_stores. */ | |
6429 | ||
6430 | static void | |
7080f735 | 6431 | invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED) |
8b3686ed | 6432 | { |
9ae8ffe7 JL |
6433 | enum rtx_code code = GET_CODE (dest); |
6434 | ||
6435 | if (code == MEM | |
ddc356e8 | 6436 | && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */ |
9ae8ffe7 JL |
6437 | /* There are times when an address can appear varying and be a PLUS |
6438 | during this scan when it would be a fixed address were we to know | |
6439 | the proper equivalences. So invalidate all memory if there is | |
6440 | a BLKmode or nonscalar memory reference or a reference to a | |
6441 | variable address. */ | |
6442 | && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode | |
2be28ee2 | 6443 | || cse_rtx_varies_p (XEXP (dest, 0), 0))) |
9ae8ffe7 JL |
6444 | { |
6445 | invalidate_memory (); | |
6446 | return; | |
6447 | } | |
ffcf6393 | 6448 | |
f47c02fa | 6449 | if (GET_CODE (set) == CLOBBER |
8beccec8 | 6450 | || CC0_P (dest) |
f47c02fa RK |
6451 | || dest == pc_rtx) |
6452 | return; | |
6453 | ||
9ae8ffe7 | 6454 | if (code == STRICT_LOW_PART || code == ZERO_EXTRACT) |
bb4034b3 | 6455 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
9ae8ffe7 JL |
6456 | else if (code == REG || code == SUBREG || code == MEM) |
6457 | invalidate (dest, VOIDmode); | |
8b3686ed RK |
6458 | } |
6459 | ||
6460 | /* Invalidate all insns from START up to the end of the function or the | |
6461 | next label. This called when we wish to CSE around a block that is | |
6462 | conditionally executed. */ | |
6463 | ||
6464 | static void | |
7080f735 | 6465 | invalidate_skipped_block (rtx start) |
8b3686ed RK |
6466 | { |
6467 | rtx insn; | |
8b3686ed | 6468 | |
4b4bf941 | 6469 | for (insn = start; insn && !LABEL_P (insn); |
8b3686ed RK |
6470 | insn = NEXT_INSN (insn)) |
6471 | { | |
2c3c49de | 6472 | if (! INSN_P (insn)) |
8b3686ed RK |
6473 | continue; |
6474 | ||
4b4bf941 | 6475 | if (CALL_P (insn)) |
8b3686ed | 6476 | { |
24a28584 | 6477 | if (! CONST_OR_PURE_CALL_P (insn)) |
9ae8ffe7 | 6478 | invalidate_memory (); |
8b3686ed | 6479 | invalidate_for_call (); |
8b3686ed RK |
6480 | } |
6481 | ||
97577254 | 6482 | invalidate_from_clobbers (PATTERN (insn)); |
84832317 | 6483 | note_stores (PATTERN (insn), invalidate_skipped_set, NULL); |
8b3686ed RK |
6484 | } |
6485 | } | |
6486 | \f | |
84832317 MM |
6487 | /* If modifying X will modify the value in *DATA (which is really an |
6488 | `rtx *'), indicate that fact by setting the pointed to value to | |
6489 | NULL_RTX. */ | |
7afe21cc RK |
6490 | |
6491 | static void | |
7080f735 | 6492 | cse_check_loop_start (rtx x, rtx set ATTRIBUTE_UNUSED, void *data) |
7afe21cc | 6493 | { |
84832317 MM |
6494 | rtx *cse_check_loop_start_value = (rtx *) data; |
6495 | ||
6496 | if (*cse_check_loop_start_value == NULL_RTX | |
7afe21cc RK |
6497 | || GET_CODE (x) == CC0 || GET_CODE (x) == PC) |
6498 | return; | |
6499 | ||
3c0cb5de | 6500 | if ((MEM_P (x) && MEM_P (*cse_check_loop_start_value)) |
84832317 MM |
6501 | || reg_overlap_mentioned_p (x, *cse_check_loop_start_value)) |
6502 | *cse_check_loop_start_value = NULL_RTX; | |
7afe21cc RK |
6503 | } |
6504 | ||
6505 | /* X is a SET or CLOBBER contained in INSN that was found near the start of | |
6506 | a loop that starts with the label at LOOP_START. | |
6507 | ||
6508 | If X is a SET, we see if its SET_SRC is currently in our hash table. | |
6509 | If so, we see if it has a value equal to some register used only in the | |
6510 | loop exit code (as marked by jump.c). | |
6511 | ||
6512 | If those two conditions are true, we search backwards from the start of | |
6513 | the loop to see if that same value was loaded into a register that still | |
6514 | retains its value at the start of the loop. | |
6515 | ||
6516 | If so, we insert an insn after the load to copy the destination of that | |
6517 | load into the equivalent register and (try to) replace our SET_SRC with that | |
6518 | register. | |
6519 | ||
6520 | In any event, we invalidate whatever this SET or CLOBBER modifies. */ | |
6521 | ||
6522 | static void | |
7080f735 | 6523 | cse_set_around_loop (rtx x, rtx insn, rtx loop_start) |
7afe21cc | 6524 | { |
7afe21cc | 6525 | struct table_elt *src_elt; |
7afe21cc RK |
6526 | |
6527 | /* If this is a SET, see if we can replace SET_SRC, but ignore SETs that | |
6528 | are setting PC or CC0 or whose SET_SRC is already a register. */ | |
6529 | if (GET_CODE (x) == SET | |
6530 | && GET_CODE (SET_DEST (x)) != PC && GET_CODE (SET_DEST (x)) != CC0 | |
f8cfc6aa | 6531 | && !REG_P (SET_SRC (x))) |
7afe21cc RK |
6532 | { |
6533 | src_elt = lookup (SET_SRC (x), | |
6534 | HASH (SET_SRC (x), GET_MODE (SET_DEST (x))), | |
6535 | GET_MODE (SET_DEST (x))); | |
6536 | ||
6537 | if (src_elt) | |
6538 | for (src_elt = src_elt->first_same_value; src_elt; | |
6539 | src_elt = src_elt->next_same_value) | |
f8cfc6aa | 6540 | if (REG_P (src_elt->exp) && REG_LOOP_TEST_P (src_elt->exp) |
7afe21cc RK |
6541 | && COST (src_elt->exp) < COST (SET_SRC (x))) |
6542 | { | |
6543 | rtx p, set; | |
6544 | ||
6545 | /* Look for an insn in front of LOOP_START that sets | |
6546 | something in the desired mode to SET_SRC (x) before we hit | |
6547 | a label or CALL_INSN. */ | |
6548 | ||
6549 | for (p = prev_nonnote_insn (loop_start); | |
4b4bf941 JQ |
6550 | p && !CALL_P (p) |
6551 | && !LABEL_P (p); | |
7afe21cc RK |
6552 | p = prev_nonnote_insn (p)) |
6553 | if ((set = single_set (p)) != 0 | |
f8cfc6aa | 6554 | && REG_P (SET_DEST (set)) |
7afe21cc RK |
6555 | && GET_MODE (SET_DEST (set)) == src_elt->mode |
6556 | && rtx_equal_p (SET_SRC (set), SET_SRC (x))) | |
6557 | { | |
6558 | /* We now have to ensure that nothing between P | |
6559 | and LOOP_START modified anything referenced in | |
6560 | SET_SRC (x). We know that nothing within the loop | |
6561 | can modify it, or we would have invalidated it in | |
6562 | the hash table. */ | |
6563 | rtx q; | |
84832317 | 6564 | rtx cse_check_loop_start_value = SET_SRC (x); |
7afe21cc | 6565 | for (q = p; q != loop_start; q = NEXT_INSN (q)) |
2c3c49de | 6566 | if (INSN_P (q)) |
84832317 MM |
6567 | note_stores (PATTERN (q), |
6568 | cse_check_loop_start, | |
6569 | &cse_check_loop_start_value); | |
7afe21cc RK |
6570 | |
6571 | /* If nothing was changed and we can replace our | |
6572 | SET_SRC, add an insn after P to copy its destination | |
6573 | to what we will be replacing SET_SRC with. */ | |
6574 | if (cse_check_loop_start_value | |
26771da7 JH |
6575 | && single_set (p) |
6576 | && !can_throw_internal (insn) | |
7afe21cc RK |
6577 | && validate_change (insn, &SET_SRC (x), |
6578 | src_elt->exp, 0)) | |
e89d3e6f R |
6579 | { |
6580 | /* If this creates new pseudos, this is unsafe, | |
6581 | because the regno of new pseudo is unsuitable | |
6582 | to index into reg_qty when cse_insn processes | |
6583 | the new insn. Therefore, if a new pseudo was | |
6584 | created, discard this optimization. */ | |
6585 | int nregs = max_reg_num (); | |
6586 | rtx move | |
6587 | = gen_move_insn (src_elt->exp, SET_DEST (set)); | |
6588 | if (nregs != max_reg_num ()) | |
6589 | { | |
6590 | if (! validate_change (insn, &SET_SRC (x), | |
6591 | SET_SRC (set), 0)) | |
6592 | abort (); | |
6593 | } | |
6594 | else | |
9ebfd78b | 6595 | { |
f3ea0706 R |
6596 | if (CONSTANT_P (SET_SRC (set)) |
6597 | && ! find_reg_equal_equiv_note (insn)) | |
6598 | set_unique_reg_note (insn, REG_EQUAL, | |
6599 | SET_SRC (set)); | |
9ebfd78b EB |
6600 | if (control_flow_insn_p (p)) |
6601 | /* p can cause a control flow transfer so it | |
6602 | is the last insn of a basic block. We can't | |
6603 | therefore use emit_insn_after. */ | |
6604 | emit_insn_before (move, next_nonnote_insn (p)); | |
6605 | else | |
6606 | emit_insn_after (move, p); | |
6607 | } | |
e89d3e6f | 6608 | } |
7afe21cc RK |
6609 | break; |
6610 | } | |
6611 | } | |
6612 | } | |
6613 | ||
14a774a9 RK |
6614 | /* Deal with the destination of X affecting the stack pointer. */ |
6615 | addr_affects_sp_p (SET_DEST (x)); | |
7afe21cc | 6616 | |
14a774a9 RK |
6617 | /* See comment on similar code in cse_insn for explanation of these |
6618 | tests. */ | |
f8cfc6aa | 6619 | if (REG_P (SET_DEST (x)) || GET_CODE (SET_DEST (x)) == SUBREG |
3c0cb5de | 6620 | || MEM_P (SET_DEST (x))) |
bb4034b3 | 6621 | invalidate (SET_DEST (x), VOIDmode); |
2708da92 RS |
6622 | else if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART |
6623 | || GET_CODE (SET_DEST (x)) == ZERO_EXTRACT) | |
bb4034b3 | 6624 | invalidate (XEXP (SET_DEST (x), 0), GET_MODE (SET_DEST (x))); |
7afe21cc RK |
6625 | } |
6626 | \f | |
6627 | /* Find the end of INSN's basic block and return its range, | |
6628 | the total number of SETs in all the insns of the block, the last insn of the | |
6629 | block, and the branch path. | |
6630 | ||
da7d8304 | 6631 | The branch path indicates which branches should be followed. If a nonzero |
7afe21cc RK |
6632 | path size is specified, the block should be rescanned and a different set |
6633 | of branches will be taken. The branch path is only used if | |
da7d8304 | 6634 | FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero. |
7afe21cc RK |
6635 | |
6636 | DATA is a pointer to a struct cse_basic_block_data, defined below, that is | |
6637 | used to describe the block. It is filled in with the information about | |
6638 | the current block. The incoming structure's branch path, if any, is used | |
6639 | to construct the output branch path. */ | |
6640 | ||
86caf04d | 6641 | static void |
7080f735 AJ |
6642 | cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data, |
6643 | int follow_jumps, int after_loop, int skip_blocks) | |
7afe21cc RK |
6644 | { |
6645 | rtx p = insn, q; | |
6646 | int nsets = 0; | |
6647 | int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn); | |
2c3c49de | 6648 | rtx next = INSN_P (insn) ? insn : next_real_insn (insn); |
7afe21cc RK |
6649 | int path_size = data->path_size; |
6650 | int path_entry = 0; | |
6651 | int i; | |
6652 | ||
6653 | /* Update the previous branch path, if any. If the last branch was | |
6de9cd9a DN |
6654 | previously PATH_TAKEN, mark it PATH_NOT_TAKEN. |
6655 | If it was previously PATH_NOT_TAKEN, | |
7afe21cc | 6656 | shorten the path by one and look at the previous branch. We know that |
da7d8304 | 6657 | at least one branch must have been taken if PATH_SIZE is nonzero. */ |
7afe21cc RK |
6658 | while (path_size > 0) |
6659 | { | |
6de9cd9a | 6660 | if (data->path[path_size - 1].status != PATH_NOT_TAKEN) |
7afe21cc | 6661 | { |
6de9cd9a | 6662 | data->path[path_size - 1].status = PATH_NOT_TAKEN; |
7afe21cc RK |
6663 | break; |
6664 | } | |
6665 | else | |
6666 | path_size--; | |
6667 | } | |
6668 | ||
16b702cd MM |
6669 | /* If the first instruction is marked with QImode, that means we've |
6670 | already processed this block. Our caller will look at DATA->LAST | |
6671 | to figure out where to go next. We want to return the next block | |
6672 | in the instruction stream, not some branched-to block somewhere | |
6673 | else. We accomplish this by pretending our called forbid us to | |
6674 | follow jumps, or skip blocks. */ | |
6675 | if (GET_MODE (insn) == QImode) | |
6676 | follow_jumps = skip_blocks = 0; | |
6677 | ||
7afe21cc | 6678 | /* Scan to end of this basic block. */ |
4b4bf941 | 6679 | while (p && !LABEL_P (p)) |
7afe21cc RK |
6680 | { |
6681 | /* Don't cse out the end of a loop. This makes a difference | |
6682 | only for the unusual loops that always execute at least once; | |
6683 | all other loops have labels there so we will stop in any case. | |
6684 | Cse'ing out the end of the loop is dangerous because it | |
6685 | might cause an invariant expression inside the loop | |
6686 | to be reused after the end of the loop. This would make it | |
6687 | hard to move the expression out of the loop in loop.c, | |
6688 | especially if it is one of several equivalent expressions | |
6689 | and loop.c would like to eliminate it. | |
6690 | ||
6691 | If we are running after loop.c has finished, we can ignore | |
6692 | the NOTE_INSN_LOOP_END. */ | |
6693 | ||
4b4bf941 | 6694 | if (! after_loop && NOTE_P (p) |
7afe21cc RK |
6695 | && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END) |
6696 | break; | |
6697 | ||
8aeea6e6 | 6698 | /* Don't cse over a call to setjmp; on some machines (eg VAX) |
7afe21cc RK |
6699 | the regs restored by the longjmp come from |
6700 | a later time than the setjmp. */ | |
4b4bf941 | 6701 | if (PREV_INSN (p) && CALL_P (PREV_INSN (p)) |
570a98eb | 6702 | && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL)) |
7afe21cc RK |
6703 | break; |
6704 | ||
6705 | /* A PARALLEL can have lots of SETs in it, | |
6706 | especially if it is really an ASM_OPERANDS. */ | |
2c3c49de | 6707 | if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL) |
7afe21cc | 6708 | nsets += XVECLEN (PATTERN (p), 0); |
4b4bf941 | 6709 | else if (!NOTE_P (p)) |
7afe21cc | 6710 | nsets += 1; |
278a83b2 | 6711 | |
164c8956 RK |
6712 | /* Ignore insns made by CSE; they cannot affect the boundaries of |
6713 | the basic block. */ | |
6714 | ||
6715 | if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid) | |
8b3686ed | 6716 | high_cuid = INSN_CUID (p); |
164c8956 RK |
6717 | if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid) |
6718 | low_cuid = INSN_CUID (p); | |
7afe21cc RK |
6719 | |
6720 | /* See if this insn is in our branch path. If it is and we are to | |
6721 | take it, do so. */ | |
6722 | if (path_entry < path_size && data->path[path_entry].branch == p) | |
6723 | { | |
6de9cd9a | 6724 | if (data->path[path_entry].status != PATH_NOT_TAKEN) |
7afe21cc | 6725 | p = JUMP_LABEL (p); |
278a83b2 | 6726 | |
7afe21cc RK |
6727 | /* Point to next entry in path, if any. */ |
6728 | path_entry++; | |
6729 | } | |
6730 | ||
6731 | /* If this is a conditional jump, we can follow it if -fcse-follow-jumps | |
6732 | was specified, we haven't reached our maximum path length, there are | |
6733 | insns following the target of the jump, this is the only use of the | |
8b3686ed RK |
6734 | jump label, and the target label is preceded by a BARRIER. |
6735 | ||
6736 | Alternatively, we can follow the jump if it branches around a | |
6737 | block of code and there are no other branches into the block. | |
6738 | In this case invalidate_skipped_block will be called to invalidate any | |
6739 | registers set in the block when following the jump. */ | |
6740 | ||
9bf8cfbf | 6741 | else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1 |
4b4bf941 | 6742 | && JUMP_P (p) |
278a83b2 | 6743 | && GET_CODE (PATTERN (p)) == SET |
7afe21cc | 6744 | && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE |
85c3ba60 | 6745 | && JUMP_LABEL (p) != 0 |
7afe21cc RK |
6746 | && LABEL_NUSES (JUMP_LABEL (p)) == 1 |
6747 | && NEXT_INSN (JUMP_LABEL (p)) != 0) | |
6748 | { | |
6749 | for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q)) | |
4b4bf941 | 6750 | if ((!NOTE_P (q) |
278a83b2 | 6751 | || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END |
4b4bf941 | 6752 | || (PREV_INSN (q) && CALL_P (PREV_INSN (q)) |
570a98eb | 6753 | && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL))) |
4b4bf941 | 6754 | && (!LABEL_P (q) || LABEL_NUSES (q) != 0)) |
7afe21cc RK |
6755 | break; |
6756 | ||
6757 | /* If we ran into a BARRIER, this code is an extension of the | |
6758 | basic block when the branch is taken. */ | |
4b4bf941 | 6759 | if (follow_jumps && q != 0 && BARRIER_P (q)) |
7afe21cc RK |
6760 | { |
6761 | /* Don't allow ourself to keep walking around an | |
6762 | always-executed loop. */ | |
fc3ffe83 RK |
6763 | if (next_real_insn (q) == next) |
6764 | { | |
6765 | p = NEXT_INSN (p); | |
6766 | continue; | |
6767 | } | |
7afe21cc RK |
6768 | |
6769 | /* Similarly, don't put a branch in our path more than once. */ | |
6770 | for (i = 0; i < path_entry; i++) | |
6771 | if (data->path[i].branch == p) | |
6772 | break; | |
6773 | ||
6774 | if (i != path_entry) | |
6775 | break; | |
6776 | ||
6777 | data->path[path_entry].branch = p; | |
6de9cd9a | 6778 | data->path[path_entry++].status = PATH_TAKEN; |
7afe21cc RK |
6779 | |
6780 | /* This branch now ends our path. It was possible that we | |
6781 | didn't see this branch the last time around (when the | |
6782 | insn in front of the target was a JUMP_INSN that was | |
6783 | turned into a no-op). */ | |
6784 | path_size = path_entry; | |
6785 | ||
6786 | p = JUMP_LABEL (p); | |
6787 | /* Mark block so we won't scan it again later. */ | |
6788 | PUT_MODE (NEXT_INSN (p), QImode); | |
6789 | } | |
8b3686ed | 6790 | /* Detect a branch around a block of code. */ |
4b4bf941 | 6791 | else if (skip_blocks && q != 0 && !LABEL_P (q)) |
8b3686ed | 6792 | { |
b3694847 | 6793 | rtx tmp; |
8b3686ed | 6794 | |
fc3ffe83 RK |
6795 | if (next_real_insn (q) == next) |
6796 | { | |
6797 | p = NEXT_INSN (p); | |
6798 | continue; | |
6799 | } | |
8b3686ed RK |
6800 | |
6801 | for (i = 0; i < path_entry; i++) | |
6802 | if (data->path[i].branch == p) | |
6803 | break; | |
6804 | ||
6805 | if (i != path_entry) | |
6806 | break; | |
6807 | ||
6808 | /* This is no_labels_between_p (p, q) with an added check for | |
6809 | reaching the end of a function (in case Q precedes P). */ | |
6810 | for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp)) | |
4b4bf941 | 6811 | if (LABEL_P (tmp)) |
8b3686ed | 6812 | break; |
278a83b2 | 6813 | |
8b3686ed RK |
6814 | if (tmp == q) |
6815 | { | |
6816 | data->path[path_entry].branch = p; | |
6de9cd9a | 6817 | data->path[path_entry++].status = PATH_AROUND; |
8b3686ed RK |
6818 | |
6819 | path_size = path_entry; | |
6820 | ||
6821 | p = JUMP_LABEL (p); | |
6822 | /* Mark block so we won't scan it again later. */ | |
6823 | PUT_MODE (NEXT_INSN (p), QImode); | |
6824 | } | |
6825 | } | |
7afe21cc | 6826 | } |
7afe21cc RK |
6827 | p = NEXT_INSN (p); |
6828 | } | |
6829 | ||
6830 | data->low_cuid = low_cuid; | |
6831 | data->high_cuid = high_cuid; | |
6832 | data->nsets = nsets; | |
6833 | data->last = p; | |
6834 | ||
6835 | /* If all jumps in the path are not taken, set our path length to zero | |
6836 | so a rescan won't be done. */ | |
6837 | for (i = path_size - 1; i >= 0; i--) | |
6de9cd9a | 6838 | if (data->path[i].status != PATH_NOT_TAKEN) |
7afe21cc RK |
6839 | break; |
6840 | ||
6841 | if (i == -1) | |
6842 | data->path_size = 0; | |
6843 | else | |
6844 | data->path_size = path_size; | |
6845 | ||
6846 | /* End the current branch path. */ | |
6847 | data->path[path_size].branch = 0; | |
6848 | } | |
6849 | \f | |
7afe21cc RK |
6850 | /* Perform cse on the instructions of a function. |
6851 | F is the first instruction. | |
6852 | NREGS is one plus the highest pseudo-reg number used in the instruction. | |
6853 | ||
6854 | AFTER_LOOP is 1 if this is the cse call done after loop optimization | |
6855 | (only if -frerun-cse-after-loop). | |
6856 | ||
6857 | Returns 1 if jump_optimize should be redone due to simplifications | |
6858 | in conditional jump instructions. */ | |
6859 | ||
6860 | int | |
7080f735 | 6861 | cse_main (rtx f, int nregs, int after_loop, FILE *file) |
7afe21cc RK |
6862 | { |
6863 | struct cse_basic_block_data val; | |
b3694847 SS |
6864 | rtx insn = f; |
6865 | int i; | |
7afe21cc | 6866 | |
9bf8cfbf ZD |
6867 | val.path = xmalloc (sizeof (struct branch_path) |
6868 | * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH)); | |
6869 | ||
7afe21cc | 6870 | cse_jumps_altered = 0; |
a5dfb4ee | 6871 | recorded_label_ref = 0; |
7afe21cc | 6872 | constant_pool_entries_cost = 0; |
dd0ba281 | 6873 | constant_pool_entries_regcost = 0; |
7afe21cc | 6874 | val.path_size = 0; |
2f93eea8 | 6875 | rtl_hooks = cse_rtl_hooks; |
7afe21cc RK |
6876 | |
6877 | init_recog (); | |
9ae8ffe7 | 6878 | init_alias_analysis (); |
7afe21cc RK |
6879 | |
6880 | max_reg = nregs; | |
6881 | ||
556c714b JW |
6882 | max_insn_uid = get_max_uid (); |
6883 | ||
703ad42b | 6884 | reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem)); |
7afe21cc | 6885 | |
7bac1be0 RK |
6886 | #ifdef LOAD_EXTEND_OP |
6887 | ||
6888 | /* Allocate scratch rtl here. cse_insn will fill in the memory reference | |
6889 | and change the code and mode as appropriate. */ | |
38a448ca | 6890 | memory_extend_rtx = gen_rtx_ZERO_EXTEND (VOIDmode, NULL_RTX); |
7bac1be0 RK |
6891 | #endif |
6892 | ||
1f8f4a0b MM |
6893 | /* Reset the counter indicating how many elements have been made |
6894 | thus far. */ | |
7afe21cc RK |
6895 | n_elements_made = 0; |
6896 | ||
6897 | /* Find the largest uid. */ | |
6898 | ||
164c8956 | 6899 | max_uid = get_max_uid (); |
703ad42b | 6900 | uid_cuid = xcalloc (max_uid + 1, sizeof (int)); |
7afe21cc RK |
6901 | |
6902 | /* Compute the mapping from uids to cuids. | |
6903 | CUIDs are numbers assigned to insns, like uids, | |
6904 | except that cuids increase monotonically through the code. | |
6905 | Don't assign cuids to line-number NOTEs, so that the distance in cuids | |
6906 | between two insns is not affected by -g. */ | |
6907 | ||
6908 | for (insn = f, i = 0; insn; insn = NEXT_INSN (insn)) | |
6909 | { | |
4b4bf941 | 6910 | if (!NOTE_P (insn) |
7afe21cc RK |
6911 | || NOTE_LINE_NUMBER (insn) < 0) |
6912 | INSN_CUID (insn) = ++i; | |
6913 | else | |
6914 | /* Give a line number note the same cuid as preceding insn. */ | |
6915 | INSN_CUID (insn) = i; | |
6916 | } | |
6917 | ||
1f8f4a0b | 6918 | ggc_push_context (); |
1497faf6 | 6919 | |
7afe21cc RK |
6920 | /* Loop over basic blocks. |
6921 | Compute the maximum number of qty's needed for each basic block | |
6922 | (which is 2 for each SET). */ | |
6923 | insn = f; | |
6924 | while (insn) | |
6925 | { | |
4eadede7 | 6926 | cse_altered = 0; |
8b3686ed RK |
6927 | cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps, after_loop, |
6928 | flag_cse_skip_blocks); | |
7afe21cc RK |
6929 | |
6930 | /* If this basic block was already processed or has no sets, skip it. */ | |
6931 | if (val.nsets == 0 || GET_MODE (insn) == QImode) | |
6932 | { | |
6933 | PUT_MODE (insn, VOIDmode); | |
6934 | insn = (val.last ? NEXT_INSN (val.last) : 0); | |
6935 | val.path_size = 0; | |
6936 | continue; | |
6937 | } | |
6938 | ||
6939 | cse_basic_block_start = val.low_cuid; | |
6940 | cse_basic_block_end = val.high_cuid; | |
6941 | max_qty = val.nsets * 2; | |
278a83b2 | 6942 | |
7afe21cc | 6943 | if (file) |
ab87f8c8 | 6944 | fnotice (file, ";; Processing block from %d to %d, %d sets.\n", |
7afe21cc RK |
6945 | INSN_UID (insn), val.last ? INSN_UID (val.last) : 0, |
6946 | val.nsets); | |
6947 | ||
6948 | /* Make MAX_QTY bigger to give us room to optimize | |
6949 | past the end of this basic block, if that should prove useful. */ | |
6950 | if (max_qty < 500) | |
6951 | max_qty = 500; | |
6952 | ||
6953 | max_qty += max_reg; | |
6954 | ||
6955 | /* If this basic block is being extended by following certain jumps, | |
6956 | (see `cse_end_of_basic_block'), we reprocess the code from the start. | |
6957 | Otherwise, we start after this basic block. */ | |
6958 | if (val.path_size > 0) | |
278a83b2 | 6959 | cse_basic_block (insn, val.last, val.path, 0); |
7afe21cc RK |
6960 | else |
6961 | { | |
6962 | int old_cse_jumps_altered = cse_jumps_altered; | |
6963 | rtx temp; | |
6964 | ||
6965 | /* When cse changes a conditional jump to an unconditional | |
6966 | jump, we want to reprocess the block, since it will give | |
6967 | us a new branch path to investigate. */ | |
6968 | cse_jumps_altered = 0; | |
6969 | temp = cse_basic_block (insn, val.last, val.path, ! after_loop); | |
8b3686ed RK |
6970 | if (cse_jumps_altered == 0 |
6971 | || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0)) | |
7afe21cc RK |
6972 | insn = temp; |
6973 | ||
6974 | cse_jumps_altered |= old_cse_jumps_altered; | |
6975 | } | |
6976 | ||
1f8f4a0b | 6977 | if (cse_altered) |
1497faf6 RH |
6978 | ggc_collect (); |
6979 | ||
7afe21cc RK |
6980 | #ifdef USE_C_ALLOCA |
6981 | alloca (0); | |
6982 | #endif | |
6983 | } | |
6984 | ||
1f8f4a0b | 6985 | ggc_pop_context (); |
1497faf6 | 6986 | |
7afe21cc RK |
6987 | if (max_elements_made < n_elements_made) |
6988 | max_elements_made = n_elements_made; | |
6989 | ||
e05e2395 MM |
6990 | /* Clean up. */ |
6991 | end_alias_analysis (); | |
75c6bd46 | 6992 | free (uid_cuid); |
1bb98cec | 6993 | free (reg_eqv_table); |
9bf8cfbf | 6994 | free (val.path); |
2f93eea8 | 6995 | rtl_hooks = general_rtl_hooks; |
e05e2395 | 6996 | |
a5dfb4ee | 6997 | return cse_jumps_altered || recorded_label_ref; |
7afe21cc RK |
6998 | } |
6999 | ||
7000 | /* Process a single basic block. FROM and TO and the limits of the basic | |
7001 | block. NEXT_BRANCH points to the branch path when following jumps or | |
7002 | a null path when not following jumps. | |
7003 | ||
da7d8304 | 7004 | AROUND_LOOP is nonzero if we are to try to cse around to the start of a |
7afe21cc RK |
7005 | loop. This is true when we are being called for the last time on a |
7006 | block and this CSE pass is before loop.c. */ | |
7007 | ||
7008 | static rtx | |
7080f735 AJ |
7009 | cse_basic_block (rtx from, rtx to, struct branch_path *next_branch, |
7010 | int around_loop) | |
7afe21cc | 7011 | { |
b3694847 | 7012 | rtx insn; |
7afe21cc | 7013 | int to_usage = 0; |
7bd8b2a8 | 7014 | rtx libcall_insn = NULL_RTX; |
e9a25f70 | 7015 | int num_insns = 0; |
26d107db | 7016 | int no_conflict = 0; |
7afe21cc | 7017 | |
1bb98cec DM |
7018 | /* This array is undefined before max_reg, so only allocate |
7019 | the space actually needed and adjust the start. */ | |
7020 | ||
703ad42b | 7021 | qty_table = xmalloc ((max_qty - max_reg) * sizeof (struct qty_table_elem)); |
1bb98cec | 7022 | qty_table -= max_reg; |
7afe21cc RK |
7023 | |
7024 | new_basic_block (); | |
7025 | ||
7026 | /* TO might be a label. If so, protect it from being deleted. */ | |
4b4bf941 | 7027 | if (to != 0 && LABEL_P (to)) |
7afe21cc RK |
7028 | ++LABEL_NUSES (to); |
7029 | ||
7030 | for (insn = from; insn != to; insn = NEXT_INSN (insn)) | |
7031 | { | |
b3694847 | 7032 | enum rtx_code code = GET_CODE (insn); |
e9a25f70 | 7033 | |
1d22a2c1 MM |
7034 | /* If we have processed 1,000 insns, flush the hash table to |
7035 | avoid extreme quadratic behavior. We must not include NOTEs | |
c13e8210 | 7036 | in the count since there may be more of them when generating |
1d22a2c1 MM |
7037 | debugging information. If we clear the table at different |
7038 | times, code generated with -g -O might be different than code | |
7039 | generated with -O but not -g. | |
e9a25f70 JL |
7040 | |
7041 | ??? This is a real kludge and needs to be done some other way. | |
7042 | Perhaps for 2.9. */ | |
1d22a2c1 | 7043 | if (code != NOTE && num_insns++ > 1000) |
e9a25f70 | 7044 | { |
01e752d3 | 7045 | flush_hash_table (); |
e9a25f70 JL |
7046 | num_insns = 0; |
7047 | } | |
7afe21cc RK |
7048 | |
7049 | /* See if this is a branch that is part of the path. If so, and it is | |
7050 | to be taken, do so. */ | |
7051 | if (next_branch->branch == insn) | |
7052 | { | |
8b3686ed | 7053 | enum taken status = next_branch++->status; |
6de9cd9a | 7054 | if (status != PATH_NOT_TAKEN) |
7afe21cc | 7055 | { |
6de9cd9a | 7056 | if (status == PATH_TAKEN) |
8b3686ed RK |
7057 | record_jump_equiv (insn, 1); |
7058 | else | |
7059 | invalidate_skipped_block (NEXT_INSN (insn)); | |
7060 | ||
7afe21cc RK |
7061 | /* Set the last insn as the jump insn; it doesn't affect cc0. |
7062 | Then follow this branch. */ | |
7063 | #ifdef HAVE_cc0 | |
7064 | prev_insn_cc0 = 0; | |
7afe21cc | 7065 | prev_insn = insn; |
4977bab6 | 7066 | #endif |
7afe21cc RK |
7067 | insn = JUMP_LABEL (insn); |
7068 | continue; | |
7069 | } | |
7070 | } | |
278a83b2 | 7071 | |
7afe21cc RK |
7072 | if (GET_MODE (insn) == QImode) |
7073 | PUT_MODE (insn, VOIDmode); | |
7074 | ||
ec8e098d | 7075 | if (GET_RTX_CLASS (code) == RTX_INSN) |
7afe21cc | 7076 | { |
7bd8b2a8 JL |
7077 | rtx p; |
7078 | ||
7afe21cc RK |
7079 | /* Process notes first so we have all notes in canonical forms when |
7080 | looking for duplicate operations. */ | |
7081 | ||
7082 | if (REG_NOTES (insn)) | |
906c4e36 | 7083 | REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX); |
7afe21cc RK |
7084 | |
7085 | /* Track when we are inside in LIBCALL block. Inside such a block, | |
7086 | we do not want to record destinations. The last insn of a | |
7087 | LIBCALL block is not considered to be part of the block, since | |
830a38ee | 7088 | its destination is the result of the block and hence should be |
7afe21cc RK |
7089 | recorded. */ |
7090 | ||
efc9bd41 RK |
7091 | if (REG_NOTES (insn) != 0) |
7092 | { | |
7093 | if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX))) | |
7094 | libcall_insn = XEXP (p, 0); | |
7095 | else if (find_reg_note (insn, REG_RETVAL, NULL_RTX)) | |
26d107db KK |
7096 | { |
7097 | /* Keep libcall_insn for the last SET insn of a no-conflict | |
7098 | block to prevent changing the destination. */ | |
7099 | if (! no_conflict) | |
7100 | libcall_insn = 0; | |
7101 | else | |
7102 | no_conflict = -1; | |
7103 | } | |
7104 | else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX)) | |
7105 | no_conflict = 1; | |
efc9bd41 | 7106 | } |
7afe21cc | 7107 | |
7bd8b2a8 | 7108 | cse_insn (insn, libcall_insn); |
f85cc4cb | 7109 | |
26d107db KK |
7110 | if (no_conflict == -1) |
7111 | { | |
7112 | libcall_insn = 0; | |
7113 | no_conflict = 0; | |
7114 | } | |
7115 | ||
be8ac49a RK |
7116 | /* If we haven't already found an insn where we added a LABEL_REF, |
7117 | check this one. */ | |
4b4bf941 | 7118 | if (NONJUMP_INSN_P (insn) && ! recorded_label_ref |
be8ac49a RK |
7119 | && for_each_rtx (&PATTERN (insn), check_for_label_ref, |
7120 | (void *) insn)) | |
f85cc4cb | 7121 | recorded_label_ref = 1; |
7afe21cc RK |
7122 | } |
7123 | ||
7124 | /* If INSN is now an unconditional jump, skip to the end of our | |
7125 | basic block by pretending that we just did the last insn in the | |
7126 | basic block. If we are jumping to the end of our block, show | |
7127 | that we can have one usage of TO. */ | |
7128 | ||
7f1c097d | 7129 | if (any_uncondjump_p (insn)) |
7afe21cc RK |
7130 | { |
7131 | if (to == 0) | |
fa0933ba JL |
7132 | { |
7133 | free (qty_table + max_reg); | |
7134 | return 0; | |
7135 | } | |
7afe21cc RK |
7136 | |
7137 | if (JUMP_LABEL (insn) == to) | |
7138 | to_usage = 1; | |
7139 | ||
6a5293dc RS |
7140 | /* Maybe TO was deleted because the jump is unconditional. |
7141 | If so, there is nothing left in this basic block. */ | |
7142 | /* ??? Perhaps it would be smarter to set TO | |
278a83b2 | 7143 | to whatever follows this insn, |
6a5293dc RS |
7144 | and pretend the basic block had always ended here. */ |
7145 | if (INSN_DELETED_P (to)) | |
7146 | break; | |
7147 | ||
7afe21cc RK |
7148 | insn = PREV_INSN (to); |
7149 | } | |
7150 | ||
7151 | /* See if it is ok to keep on going past the label | |
7152 | which used to end our basic block. Remember that we incremented | |
d45cf215 | 7153 | the count of that label, so we decrement it here. If we made |
7afe21cc RK |
7154 | a jump unconditional, TO_USAGE will be one; in that case, we don't |
7155 | want to count the use in that jump. */ | |
7156 | ||
7157 | if (to != 0 && NEXT_INSN (insn) == to | |
4b4bf941 | 7158 | && LABEL_P (to) && --LABEL_NUSES (to) == to_usage) |
7afe21cc RK |
7159 | { |
7160 | struct cse_basic_block_data val; | |
146135d6 | 7161 | rtx prev; |
7afe21cc RK |
7162 | |
7163 | insn = NEXT_INSN (to); | |
7164 | ||
146135d6 RK |
7165 | /* If TO was the last insn in the function, we are done. */ |
7166 | if (insn == 0) | |
fa0933ba JL |
7167 | { |
7168 | free (qty_table + max_reg); | |
7169 | return 0; | |
7170 | } | |
7afe21cc | 7171 | |
146135d6 RK |
7172 | /* If TO was preceded by a BARRIER we are done with this block |
7173 | because it has no continuation. */ | |
7174 | prev = prev_nonnote_insn (to); | |
4b4bf941 | 7175 | if (prev && BARRIER_P (prev)) |
fa0933ba JL |
7176 | { |
7177 | free (qty_table + max_reg); | |
7178 | return insn; | |
7179 | } | |
146135d6 RK |
7180 | |
7181 | /* Find the end of the following block. Note that we won't be | |
7182 | following branches in this case. */ | |
7afe21cc RK |
7183 | to_usage = 0; |
7184 | val.path_size = 0; | |
9bf8cfbf ZD |
7185 | val.path = xmalloc (sizeof (struct branch_path) |
7186 | * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH)); | |
8b3686ed | 7187 | cse_end_of_basic_block (insn, &val, 0, 0, 0); |
9bf8cfbf | 7188 | free (val.path); |
7afe21cc RK |
7189 | |
7190 | /* If the tables we allocated have enough space left | |
7191 | to handle all the SETs in the next basic block, | |
7192 | continue through it. Otherwise, return, | |
7193 | and that block will be scanned individually. */ | |
7194 | if (val.nsets * 2 + next_qty > max_qty) | |
7195 | break; | |
7196 | ||
7197 | cse_basic_block_start = val.low_cuid; | |
7198 | cse_basic_block_end = val.high_cuid; | |
7199 | to = val.last; | |
7200 | ||
7201 | /* Prevent TO from being deleted if it is a label. */ | |
4b4bf941 | 7202 | if (to != 0 && LABEL_P (to)) |
7afe21cc RK |
7203 | ++LABEL_NUSES (to); |
7204 | ||
7205 | /* Back up so we process the first insn in the extension. */ | |
7206 | insn = PREV_INSN (insn); | |
7207 | } | |
7208 | } | |
7209 | ||
7210 | if (next_qty > max_qty) | |
7211 | abort (); | |
7212 | ||
7213 | /* If we are running before loop.c, we stopped on a NOTE_INSN_LOOP_END, and | |
7214 | the previous insn is the only insn that branches to the head of a loop, | |
7215 | we can cse into the loop. Don't do this if we changed the jump | |
7216 | structure of a loop unless we aren't going to be following jumps. */ | |
7217 | ||
68252e27 | 7218 | insn = prev_nonnote_insn (to); |
8b3686ed RK |
7219 | if ((cse_jumps_altered == 0 |
7220 | || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0)) | |
7afe21cc | 7221 | && around_loop && to != 0 |
4b4bf941 JQ |
7222 | && NOTE_P (to) && NOTE_LINE_NUMBER (to) == NOTE_INSN_LOOP_END |
7223 | && JUMP_P (insn) | |
b5a696fb RZ |
7224 | && JUMP_LABEL (insn) != 0 |
7225 | && LABEL_NUSES (JUMP_LABEL (insn)) == 1) | |
7226 | cse_around_loop (JUMP_LABEL (insn)); | |
7afe21cc | 7227 | |
1bb98cec | 7228 | free (qty_table + max_reg); |
75c6bd46 | 7229 | |
7afe21cc RK |
7230 | return to ? NEXT_INSN (to) : 0; |
7231 | } | |
7232 | \f | |
be8ac49a | 7233 | /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which |
45c23566 | 7234 | there isn't a REG_LABEL note. Return one if so. DATA is the insn. */ |
be8ac49a RK |
7235 | |
7236 | static int | |
7080f735 | 7237 | check_for_label_ref (rtx *rtl, void *data) |
be8ac49a RK |
7238 | { |
7239 | rtx insn = (rtx) data; | |
7240 | ||
7241 | /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it, | |
7242 | we must rerun jump since it needs to place the note. If this is a | |
7243 | LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this | |
ec5c56db | 7244 | since no REG_LABEL will be added. */ |
be8ac49a | 7245 | return (GET_CODE (*rtl) == LABEL_REF |
45c23566 | 7246 | && ! LABEL_REF_NONLOCAL_P (*rtl) |
4838c5ee | 7247 | && LABEL_P (XEXP (*rtl, 0)) |
be8ac49a RK |
7248 | && INSN_UID (XEXP (*rtl, 0)) != 0 |
7249 | && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0))); | |
7250 | } | |
7251 | \f | |
7afe21cc RK |
7252 | /* Count the number of times registers are used (not set) in X. |
7253 | COUNTS is an array in which we accumulate the count, INCR is how much | |
9ab81df2 | 7254 | we count each register usage. */ |
7afe21cc RK |
7255 | |
7256 | static void | |
9ab81df2 | 7257 | count_reg_usage (rtx x, int *counts, int incr) |
7afe21cc | 7258 | { |
f1e7c95f | 7259 | enum rtx_code code; |
b17d5d7c | 7260 | rtx note; |
6f7d635c | 7261 | const char *fmt; |
7afe21cc RK |
7262 | int i, j; |
7263 | ||
f1e7c95f RK |
7264 | if (x == 0) |
7265 | return; | |
7266 | ||
7267 | switch (code = GET_CODE (x)) | |
7afe21cc RK |
7268 | { |
7269 | case REG: | |
9ab81df2 | 7270 | counts[REGNO (x)] += incr; |
7afe21cc RK |
7271 | return; |
7272 | ||
7273 | case PC: | |
7274 | case CC0: | |
7275 | case CONST: | |
7276 | case CONST_INT: | |
7277 | case CONST_DOUBLE: | |
69ef87e2 | 7278 | case CONST_VECTOR: |
7afe21cc RK |
7279 | case SYMBOL_REF: |
7280 | case LABEL_REF: | |
02e39abc JL |
7281 | return; |
7282 | ||
278a83b2 | 7283 | case CLOBBER: |
02e39abc JL |
7284 | /* If we are clobbering a MEM, mark any registers inside the address |
7285 | as being used. */ | |
3c0cb5de | 7286 | if (MEM_P (XEXP (x, 0))) |
9ab81df2 | 7287 | count_reg_usage (XEXP (XEXP (x, 0), 0), counts, incr); |
7afe21cc RK |
7288 | return; |
7289 | ||
7290 | case SET: | |
7291 | /* Unless we are setting a REG, count everything in SET_DEST. */ | |
f8cfc6aa | 7292 | if (!REG_P (SET_DEST (x))) |
9ab81df2 JDA |
7293 | count_reg_usage (SET_DEST (x), counts, incr); |
7294 | count_reg_usage (SET_SRC (x), counts, incr); | |
7afe21cc RK |
7295 | return; |
7296 | ||
f1e7c95f | 7297 | case CALL_INSN: |
9ab81df2 | 7298 | count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, incr); |
ddc356e8 | 7299 | /* Fall through. */ |
f1e7c95f | 7300 | |
7afe21cc RK |
7301 | case INSN: |
7302 | case JUMP_INSN: | |
9ab81df2 | 7303 | count_reg_usage (PATTERN (x), counts, incr); |
7afe21cc RK |
7304 | |
7305 | /* Things used in a REG_EQUAL note aren't dead since loop may try to | |
7306 | use them. */ | |
7307 | ||
b17d5d7c ZD |
7308 | note = find_reg_equal_equiv_note (x); |
7309 | if (note) | |
839844be R |
7310 | { |
7311 | rtx eqv = XEXP (note, 0); | |
7312 | ||
7313 | if (GET_CODE (eqv) == EXPR_LIST) | |
7314 | /* This REG_EQUAL note describes the result of a function call. | |
7315 | Process all the arguments. */ | |
7316 | do | |
7317 | { | |
9ab81df2 | 7318 | count_reg_usage (XEXP (eqv, 0), counts, incr); |
839844be R |
7319 | eqv = XEXP (eqv, 1); |
7320 | } | |
7321 | while (eqv && GET_CODE (eqv) == EXPR_LIST); | |
7322 | else | |
9ab81df2 | 7323 | count_reg_usage (eqv, counts, incr); |
839844be | 7324 | } |
7afe21cc RK |
7325 | return; |
7326 | ||
ee960939 OH |
7327 | case EXPR_LIST: |
7328 | if (REG_NOTE_KIND (x) == REG_EQUAL | |
7329 | || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE) | |
7330 | /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)), | |
7331 | involving registers in the address. */ | |
7332 | || GET_CODE (XEXP (x, 0)) == CLOBBER) | |
9ab81df2 | 7333 | count_reg_usage (XEXP (x, 0), counts, incr); |
ee960939 | 7334 | |
9ab81df2 | 7335 | count_reg_usage (XEXP (x, 1), counts, incr); |
ee960939 OH |
7336 | return; |
7337 | ||
a6c14a64 | 7338 | case ASM_OPERANDS: |
a6c14a64 RH |
7339 | /* Iterate over just the inputs, not the constraints as well. */ |
7340 | for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) | |
9ab81df2 | 7341 | count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, incr); |
a6c14a64 RH |
7342 | return; |
7343 | ||
7afe21cc | 7344 | case INSN_LIST: |
b17d5d7c | 7345 | abort (); |
278a83b2 | 7346 | |
e9a25f70 JL |
7347 | default: |
7348 | break; | |
7afe21cc RK |
7349 | } |
7350 | ||
7351 | fmt = GET_RTX_FORMAT (code); | |
7352 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
7353 | { | |
7354 | if (fmt[i] == 'e') | |
9ab81df2 | 7355 | count_reg_usage (XEXP (x, i), counts, incr); |
7afe21cc RK |
7356 | else if (fmt[i] == 'E') |
7357 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9ab81df2 | 7358 | count_reg_usage (XVECEXP (x, i, j), counts, incr); |
7afe21cc RK |
7359 | } |
7360 | } | |
7361 | \f | |
4793dca1 JH |
7362 | /* Return true if set is live. */ |
7363 | static bool | |
7080f735 AJ |
7364 | set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */ |
7365 | int *counts) | |
4793dca1 JH |
7366 | { |
7367 | #ifdef HAVE_cc0 | |
7368 | rtx tem; | |
7369 | #endif | |
7370 | ||
7371 | if (set_noop_p (set)) | |
7372 | ; | |
7373 | ||
7374 | #ifdef HAVE_cc0 | |
7375 | else if (GET_CODE (SET_DEST (set)) == CC0 | |
7376 | && !side_effects_p (SET_SRC (set)) | |
7377 | && ((tem = next_nonnote_insn (insn)) == 0 | |
7378 | || !INSN_P (tem) | |
7379 | || !reg_referenced_p (cc0_rtx, PATTERN (tem)))) | |
7380 | return false; | |
7381 | #endif | |
f8cfc6aa | 7382 | else if (!REG_P (SET_DEST (set)) |
4793dca1 JH |
7383 | || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER |
7384 | || counts[REGNO (SET_DEST (set))] != 0 | |
8fff4fc1 | 7385 | || side_effects_p (SET_SRC (set))) |
4793dca1 JH |
7386 | return true; |
7387 | return false; | |
7388 | } | |
7389 | ||
7390 | /* Return true if insn is live. */ | |
7391 | ||
7392 | static bool | |
7080f735 | 7393 | insn_live_p (rtx insn, int *counts) |
4793dca1 JH |
7394 | { |
7395 | int i; | |
a3745024 | 7396 | if (flag_non_call_exceptions && may_trap_p (PATTERN (insn))) |
a646f6cc AH |
7397 | return true; |
7398 | else if (GET_CODE (PATTERN (insn)) == SET) | |
0021de69 | 7399 | return set_live_p (PATTERN (insn), insn, counts); |
4793dca1 | 7400 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) |
0021de69 DB |
7401 | { |
7402 | for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) | |
7403 | { | |
7404 | rtx elt = XVECEXP (PATTERN (insn), 0, i); | |
4793dca1 | 7405 | |
0021de69 DB |
7406 | if (GET_CODE (elt) == SET) |
7407 | { | |
7408 | if (set_live_p (elt, insn, counts)) | |
7409 | return true; | |
7410 | } | |
7411 | else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE) | |
7412 | return true; | |
7413 | } | |
7414 | return false; | |
7415 | } | |
4793dca1 JH |
7416 | else |
7417 | return true; | |
7418 | } | |
7419 | ||
7420 | /* Return true if libcall is dead as a whole. */ | |
7421 | ||
7422 | static bool | |
7080f735 | 7423 | dead_libcall_p (rtx insn, int *counts) |
4793dca1 | 7424 | { |
0c19a26f RS |
7425 | rtx note, set, new; |
7426 | ||
4793dca1 JH |
7427 | /* See if there's a REG_EQUAL note on this insn and try to |
7428 | replace the source with the REG_EQUAL expression. | |
7429 | ||
7430 | We assume that insns with REG_RETVALs can only be reg->reg | |
7431 | copies at this point. */ | |
7432 | note = find_reg_note (insn, REG_EQUAL, NULL_RTX); | |
0c19a26f RS |
7433 | if (!note) |
7434 | return false; | |
7435 | ||
7436 | set = single_set (insn); | |
7437 | if (!set) | |
7438 | return false; | |
4793dca1 | 7439 | |
0c19a26f RS |
7440 | new = simplify_rtx (XEXP (note, 0)); |
7441 | if (!new) | |
7442 | new = XEXP (note, 0); | |
4793dca1 | 7443 | |
0c19a26f | 7444 | /* While changing insn, we must update the counts accordingly. */ |
9ab81df2 | 7445 | count_reg_usage (insn, counts, -1); |
1e150f2c | 7446 | |
0c19a26f RS |
7447 | if (validate_change (insn, &SET_SRC (set), new, 0)) |
7448 | { | |
9ab81df2 | 7449 | count_reg_usage (insn, counts, 1); |
0c19a26f RS |
7450 | remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX)); |
7451 | remove_note (insn, note); | |
7452 | return true; | |
7453 | } | |
7454 | ||
7455 | if (CONSTANT_P (new)) | |
7456 | { | |
7457 | new = force_const_mem (GET_MODE (SET_DEST (set)), new); | |
7458 | if (new && validate_change (insn, &SET_SRC (set), new, 0)) | |
4793dca1 | 7459 | { |
9ab81df2 | 7460 | count_reg_usage (insn, counts, 1); |
4793dca1 | 7461 | remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX)); |
1e150f2c | 7462 | remove_note (insn, note); |
4793dca1 JH |
7463 | return true; |
7464 | } | |
7465 | } | |
7080f735 | 7466 | |
9ab81df2 | 7467 | count_reg_usage (insn, counts, 1); |
4793dca1 JH |
7468 | return false; |
7469 | } | |
7470 | ||
7afe21cc RK |
7471 | /* Scan all the insns and delete any that are dead; i.e., they store a register |
7472 | that is never used or they copy a register to itself. | |
7473 | ||
c6a26dc4 JL |
7474 | This is used to remove insns made obviously dead by cse, loop or other |
7475 | optimizations. It improves the heuristics in loop since it won't try to | |
7476 | move dead invariants out of loops or make givs for dead quantities. The | |
7477 | remaining passes of the compilation are also sped up. */ | |
7afe21cc | 7478 | |
3dec4024 | 7479 | int |
7080f735 | 7480 | delete_trivially_dead_insns (rtx insns, int nreg) |
7afe21cc | 7481 | { |
4da896b2 | 7482 | int *counts; |
77fa0940 | 7483 | rtx insn, prev; |
614bb5d4 | 7484 | int in_libcall = 0, dead_libcall = 0; |
3dec4024 | 7485 | int ndead = 0, nlastdead, niterations = 0; |
7afe21cc | 7486 | |
3dec4024 | 7487 | timevar_push (TV_DELETE_TRIVIALLY_DEAD); |
7afe21cc | 7488 | /* First count the number of times each register is used. */ |
703ad42b | 7489 | counts = xcalloc (nreg, sizeof (int)); |
7afe21cc | 7490 | for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn)) |
9ab81df2 | 7491 | count_reg_usage (insn, counts, 1); |
7afe21cc | 7492 | |
3dec4024 JH |
7493 | do |
7494 | { | |
7495 | nlastdead = ndead; | |
7496 | niterations++; | |
7497 | /* Go from the last insn to the first and delete insns that only set unused | |
7498 | registers or copy a register to itself. As we delete an insn, remove | |
7499 | usage counts for registers it uses. | |
7500 | ||
7501 | The first jump optimization pass may leave a real insn as the last | |
7502 | insn in the function. We must not skip that insn or we may end | |
7503 | up deleting code that is not really dead. */ | |
7504 | insn = get_last_insn (); | |
7505 | if (! INSN_P (insn)) | |
7506 | insn = prev_real_insn (insn); | |
0cedb36c | 7507 | |
3dec4024 | 7508 | for (; insn; insn = prev) |
7afe21cc | 7509 | { |
4793dca1 | 7510 | int live_insn = 0; |
7afe21cc | 7511 | |
3dec4024 | 7512 | prev = prev_real_insn (insn); |
7afe21cc | 7513 | |
4793dca1 JH |
7514 | /* Don't delete any insns that are part of a libcall block unless |
7515 | we can delete the whole libcall block. | |
7afe21cc | 7516 | |
4793dca1 JH |
7517 | Flow or loop might get confused if we did that. Remember |
7518 | that we are scanning backwards. */ | |
7519 | if (find_reg_note (insn, REG_RETVAL, NULL_RTX)) | |
7520 | { | |
7521 | in_libcall = 1; | |
dc42616f | 7522 | live_insn = 1; |
1e150f2c | 7523 | dead_libcall = dead_libcall_p (insn, counts); |
4793dca1 JH |
7524 | } |
7525 | else if (in_libcall) | |
7526 | live_insn = ! dead_libcall; | |
7527 | else | |
7528 | live_insn = insn_live_p (insn, counts); | |
7afe21cc | 7529 | |
4793dca1 JH |
7530 | /* If this is a dead insn, delete it and show registers in it aren't |
7531 | being used. */ | |
7afe21cc | 7532 | |
4793dca1 JH |
7533 | if (! live_insn) |
7534 | { | |
9ab81df2 | 7535 | count_reg_usage (insn, counts, -1); |
3dec4024 JH |
7536 | delete_insn_and_edges (insn); |
7537 | ndead++; | |
4793dca1 | 7538 | } |
e4890d45 | 7539 | |
4793dca1 JH |
7540 | if (find_reg_note (insn, REG_LIBCALL, NULL_RTX)) |
7541 | { | |
7542 | in_libcall = 0; | |
7543 | dead_libcall = 0; | |
7544 | } | |
614bb5d4 | 7545 | } |
68252e27 KH |
7546 | } |
7547 | while (ndead != nlastdead); | |
4da896b2 | 7548 | |
c263766c RH |
7549 | if (dump_file && ndead) |
7550 | fprintf (dump_file, "Deleted %i trivially dead insns; %i iterations\n", | |
3dec4024 | 7551 | ndead, niterations); |
4da896b2 MM |
7552 | /* Clean up. */ |
7553 | free (counts); | |
3dec4024 JH |
7554 | timevar_pop (TV_DELETE_TRIVIALLY_DEAD); |
7555 | return ndead; | |
7afe21cc | 7556 | } |
e129d93a ILT |
7557 | |
7558 | /* This function is called via for_each_rtx. The argument, NEWREG, is | |
7559 | a condition code register with the desired mode. If we are looking | |
7560 | at the same register in a different mode, replace it with | |
7561 | NEWREG. */ | |
7562 | ||
7563 | static int | |
7564 | cse_change_cc_mode (rtx *loc, void *data) | |
7565 | { | |
7566 | rtx newreg = (rtx) data; | |
7567 | ||
7568 | if (*loc | |
f8cfc6aa | 7569 | && REG_P (*loc) |
e129d93a ILT |
7570 | && REGNO (*loc) == REGNO (newreg) |
7571 | && GET_MODE (*loc) != GET_MODE (newreg)) | |
7572 | { | |
7573 | *loc = newreg; | |
7574 | return -1; | |
7575 | } | |
7576 | return 0; | |
7577 | } | |
7578 | ||
7579 | /* Change the mode of any reference to the register REGNO (NEWREG) to | |
7580 | GET_MODE (NEWREG), starting at START. Stop before END. Stop at | |
2e802a6f | 7581 | any instruction which modifies NEWREG. */ |
e129d93a ILT |
7582 | |
7583 | static void | |
7584 | cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg) | |
7585 | { | |
7586 | rtx insn; | |
7587 | ||
7588 | for (insn = start; insn != end; insn = NEXT_INSN (insn)) | |
7589 | { | |
7590 | if (! INSN_P (insn)) | |
7591 | continue; | |
7592 | ||
2e802a6f | 7593 | if (reg_set_p (newreg, insn)) |
e129d93a ILT |
7594 | return; |
7595 | ||
7596 | for_each_rtx (&PATTERN (insn), cse_change_cc_mode, newreg); | |
7597 | for_each_rtx (®_NOTES (insn), cse_change_cc_mode, newreg); | |
7598 | } | |
7599 | } | |
7600 | ||
7601 | /* BB is a basic block which finishes with CC_REG as a condition code | |
7602 | register which is set to CC_SRC. Look through the successors of BB | |
7603 | to find blocks which have a single predecessor (i.e., this one), | |
7604 | and look through those blocks for an assignment to CC_REG which is | |
7605 | equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are | |
7606 | permitted to change the mode of CC_SRC to a compatible mode. This | |
7607 | returns VOIDmode if no equivalent assignments were found. | |
7608 | Otherwise it returns the mode which CC_SRC should wind up with. | |
7609 | ||
7610 | The main complexity in this function is handling the mode issues. | |
7611 | We may have more than one duplicate which we can eliminate, and we | |
7612 | try to find a mode which will work for multiple duplicates. */ | |
7613 | ||
7614 | static enum machine_mode | |
7615 | cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode) | |
7616 | { | |
7617 | bool found_equiv; | |
7618 | enum machine_mode mode; | |
7619 | unsigned int insn_count; | |
7620 | edge e; | |
7621 | rtx insns[2]; | |
7622 | enum machine_mode modes[2]; | |
7623 | rtx last_insns[2]; | |
7624 | unsigned int i; | |
7625 | rtx newreg; | |
7626 | ||
7627 | /* We expect to have two successors. Look at both before picking | |
7628 | the final mode for the comparison. If we have more successors | |
7629 | (i.e., some sort of table jump, although that seems unlikely), | |
7630 | then we require all beyond the first two to use the same | |
7631 | mode. */ | |
7632 | ||
7633 | found_equiv = false; | |
7634 | mode = GET_MODE (cc_src); | |
7635 | insn_count = 0; | |
7636 | for (e = bb->succ; e; e = e->succ_next) | |
7637 | { | |
7638 | rtx insn; | |
7639 | rtx end; | |
7640 | ||
7641 | if (e->flags & EDGE_COMPLEX) | |
7642 | continue; | |
7643 | ||
7644 | if (! e->dest->pred | |
7645 | || e->dest->pred->pred_next | |
7646 | || e->dest == EXIT_BLOCK_PTR) | |
7647 | continue; | |
7648 | ||
7649 | end = NEXT_INSN (BB_END (e->dest)); | |
7650 | for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn)) | |
7651 | { | |
7652 | rtx set; | |
7653 | ||
7654 | if (! INSN_P (insn)) | |
7655 | continue; | |
7656 | ||
7657 | /* If CC_SRC is modified, we have to stop looking for | |
7658 | something which uses it. */ | |
7659 | if (modified_in_p (cc_src, insn)) | |
7660 | break; | |
7661 | ||
7662 | /* Check whether INSN sets CC_REG to CC_SRC. */ | |
7663 | set = single_set (insn); | |
7664 | if (set | |
f8cfc6aa | 7665 | && REG_P (SET_DEST (set)) |
e129d93a ILT |
7666 | && REGNO (SET_DEST (set)) == REGNO (cc_reg)) |
7667 | { | |
7668 | bool found; | |
7669 | enum machine_mode set_mode; | |
7670 | enum machine_mode comp_mode; | |
7671 | ||
7672 | found = false; | |
7673 | set_mode = GET_MODE (SET_SRC (set)); | |
7674 | comp_mode = set_mode; | |
7675 | if (rtx_equal_p (cc_src, SET_SRC (set))) | |
7676 | found = true; | |
7677 | else if (GET_CODE (cc_src) == COMPARE | |
7678 | && GET_CODE (SET_SRC (set)) == COMPARE | |
1f44254c | 7679 | && mode != set_mode |
e129d93a ILT |
7680 | && rtx_equal_p (XEXP (cc_src, 0), |
7681 | XEXP (SET_SRC (set), 0)) | |
7682 | && rtx_equal_p (XEXP (cc_src, 1), | |
7683 | XEXP (SET_SRC (set), 1))) | |
7684 | ||
7685 | { | |
5fd9b178 | 7686 | comp_mode = targetm.cc_modes_compatible (mode, set_mode); |
e129d93a ILT |
7687 | if (comp_mode != VOIDmode |
7688 | && (can_change_mode || comp_mode == mode)) | |
7689 | found = true; | |
7690 | } | |
7691 | ||
7692 | if (found) | |
7693 | { | |
7694 | found_equiv = true; | |
1f44254c | 7695 | if (insn_count < ARRAY_SIZE (insns)) |
e129d93a ILT |
7696 | { |
7697 | insns[insn_count] = insn; | |
7698 | modes[insn_count] = set_mode; | |
7699 | last_insns[insn_count] = end; | |
7700 | ++insn_count; | |
7701 | ||
1f44254c ILT |
7702 | if (mode != comp_mode) |
7703 | { | |
7704 | if (! can_change_mode) | |
7705 | abort (); | |
7706 | mode = comp_mode; | |
7707 | PUT_MODE (cc_src, mode); | |
7708 | } | |
e129d93a ILT |
7709 | } |
7710 | else | |
7711 | { | |
7712 | if (set_mode != mode) | |
1f44254c ILT |
7713 | { |
7714 | /* We found a matching expression in the | |
7715 | wrong mode, but we don't have room to | |
7716 | store it in the array. Punt. This case | |
7717 | should be rare. */ | |
7718 | break; | |
7719 | } | |
e129d93a ILT |
7720 | /* INSN sets CC_REG to a value equal to CC_SRC |
7721 | with the right mode. We can simply delete | |
7722 | it. */ | |
7723 | delete_insn (insn); | |
7724 | } | |
7725 | ||
7726 | /* We found an instruction to delete. Keep looking, | |
7727 | in the hopes of finding a three-way jump. */ | |
7728 | continue; | |
7729 | } | |
7730 | ||
7731 | /* We found an instruction which sets the condition | |
7732 | code, so don't look any farther. */ | |
7733 | break; | |
7734 | } | |
7735 | ||
7736 | /* If INSN sets CC_REG in some other way, don't look any | |
7737 | farther. */ | |
7738 | if (reg_set_p (cc_reg, insn)) | |
7739 | break; | |
7740 | } | |
7741 | ||
7742 | /* If we fell off the bottom of the block, we can keep looking | |
7743 | through successors. We pass CAN_CHANGE_MODE as false because | |
7744 | we aren't prepared to handle compatibility between the | |
7745 | further blocks and this block. */ | |
7746 | if (insn == end) | |
7747 | { | |
1f44254c ILT |
7748 | enum machine_mode submode; |
7749 | ||
7750 | submode = cse_cc_succs (e->dest, cc_reg, cc_src, false); | |
7751 | if (submode != VOIDmode) | |
7752 | { | |
7753 | if (submode != mode) | |
7754 | abort (); | |
7755 | found_equiv = true; | |
7756 | can_change_mode = false; | |
7757 | } | |
e129d93a ILT |
7758 | } |
7759 | } | |
7760 | ||
7761 | if (! found_equiv) | |
7762 | return VOIDmode; | |
7763 | ||
7764 | /* Now INSN_COUNT is the number of instructions we found which set | |
7765 | CC_REG to a value equivalent to CC_SRC. The instructions are in | |
7766 | INSNS. The modes used by those instructions are in MODES. */ | |
7767 | ||
7768 | newreg = NULL_RTX; | |
7769 | for (i = 0; i < insn_count; ++i) | |
7770 | { | |
7771 | if (modes[i] != mode) | |
7772 | { | |
7773 | /* We need to change the mode of CC_REG in INSNS[i] and | |
7774 | subsequent instructions. */ | |
7775 | if (! newreg) | |
7776 | { | |
7777 | if (GET_MODE (cc_reg) == mode) | |
7778 | newreg = cc_reg; | |
7779 | else | |
7780 | newreg = gen_rtx_REG (mode, REGNO (cc_reg)); | |
7781 | } | |
7782 | cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i], | |
7783 | newreg); | |
7784 | } | |
7785 | ||
7786 | delete_insn (insns[i]); | |
7787 | } | |
7788 | ||
7789 | return mode; | |
7790 | } | |
7791 | ||
7792 | /* If we have a fixed condition code register (or two), walk through | |
7793 | the instructions and try to eliminate duplicate assignments. */ | |
7794 | ||
7795 | void | |
7796 | cse_condition_code_reg (void) | |
7797 | { | |
7798 | unsigned int cc_regno_1; | |
7799 | unsigned int cc_regno_2; | |
7800 | rtx cc_reg_1; | |
7801 | rtx cc_reg_2; | |
7802 | basic_block bb; | |
7803 | ||
5fd9b178 | 7804 | if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2)) |
e129d93a ILT |
7805 | return; |
7806 | ||
7807 | cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1); | |
7808 | if (cc_regno_2 != INVALID_REGNUM) | |
7809 | cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2); | |
7810 | else | |
7811 | cc_reg_2 = NULL_RTX; | |
7812 | ||
7813 | FOR_EACH_BB (bb) | |
7814 | { | |
7815 | rtx last_insn; | |
7816 | rtx cc_reg; | |
7817 | rtx insn; | |
7818 | rtx cc_src_insn; | |
7819 | rtx cc_src; | |
7820 | enum machine_mode mode; | |
1f44254c | 7821 | enum machine_mode orig_mode; |
e129d93a ILT |
7822 | |
7823 | /* Look for blocks which end with a conditional jump based on a | |
7824 | condition code register. Then look for the instruction which | |
7825 | sets the condition code register. Then look through the | |
7826 | successor blocks for instructions which set the condition | |
7827 | code register to the same value. There are other possible | |
7828 | uses of the condition code register, but these are by far the | |
7829 | most common and the ones which we are most likely to be able | |
7830 | to optimize. */ | |
7831 | ||
7832 | last_insn = BB_END (bb); | |
4b4bf941 | 7833 | if (!JUMP_P (last_insn)) |
e129d93a ILT |
7834 | continue; |
7835 | ||
7836 | if (reg_referenced_p (cc_reg_1, PATTERN (last_insn))) | |
7837 | cc_reg = cc_reg_1; | |
7838 | else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn))) | |
7839 | cc_reg = cc_reg_2; | |
7840 | else | |
7841 | continue; | |
7842 | ||
7843 | cc_src_insn = NULL_RTX; | |
7844 | cc_src = NULL_RTX; | |
7845 | for (insn = PREV_INSN (last_insn); | |
7846 | insn && insn != PREV_INSN (BB_HEAD (bb)); | |
7847 | insn = PREV_INSN (insn)) | |
7848 | { | |
7849 | rtx set; | |
7850 | ||
7851 | if (! INSN_P (insn)) | |
7852 | continue; | |
7853 | set = single_set (insn); | |
7854 | if (set | |
f8cfc6aa | 7855 | && REG_P (SET_DEST (set)) |
e129d93a ILT |
7856 | && REGNO (SET_DEST (set)) == REGNO (cc_reg)) |
7857 | { | |
7858 | cc_src_insn = insn; | |
7859 | cc_src = SET_SRC (set); | |
7860 | break; | |
7861 | } | |
7862 | else if (reg_set_p (cc_reg, insn)) | |
7863 | break; | |
7864 | } | |
7865 | ||
7866 | if (! cc_src_insn) | |
7867 | continue; | |
7868 | ||
7869 | if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn))) | |
7870 | continue; | |
7871 | ||
7872 | /* Now CC_REG is a condition code register used for a | |
7873 | conditional jump at the end of the block, and CC_SRC, in | |
7874 | CC_SRC_INSN, is the value to which that condition code | |
7875 | register is set, and CC_SRC is still meaningful at the end of | |
7876 | the basic block. */ | |
7877 | ||
1f44254c | 7878 | orig_mode = GET_MODE (cc_src); |
e129d93a | 7879 | mode = cse_cc_succs (bb, cc_reg, cc_src, true); |
1f44254c | 7880 | if (mode != VOIDmode) |
e129d93a | 7881 | { |
1f44254c ILT |
7882 | if (mode != GET_MODE (cc_src)) |
7883 | abort (); | |
7884 | if (mode != orig_mode) | |
2e802a6f KH |
7885 | { |
7886 | rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg)); | |
7887 | ||
7888 | /* Change the mode of CC_REG in CC_SRC_INSN to | |
7889 | GET_MODE (NEWREG). */ | |
7890 | for_each_rtx (&PATTERN (cc_src_insn), cse_change_cc_mode, | |
7891 | newreg); | |
7892 | for_each_rtx (®_NOTES (cc_src_insn), cse_change_cc_mode, | |
7893 | newreg); | |
7894 | ||
7895 | /* Do the same in the following insns that use the | |
7896 | current value of CC_REG within BB. */ | |
7897 | cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn), | |
7898 | NEXT_INSN (last_insn), | |
7899 | newreg); | |
7900 | } | |
e129d93a ILT |
7901 | } |
7902 | } | |
7903 | } |