Commit | Line | Data |
---|---|---|
9bf86ebb PR |
1 | /* Optimize by combining instructions for GNU compiler. |
2 | Copyright (C) 1987, 1988, 1992, 1993 Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GNU CC. | |
5 | ||
6 | GNU CC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8 | the Free Software Foundation; either version 2, or (at your option) | |
9 | any later version. | |
10 | ||
11 | GNU CC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GNU CC; see the file COPYING. If not, write to | |
18 | the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */ | |
19 | ||
20 | ||
21 | /* This module is essentially the "combiner" phase of the U. of Arizona | |
22 | Portable Optimizer, but redone to work on our list-structured | |
23 | representation for RTL instead of their string representation. | |
24 | ||
25 | The LOG_LINKS of each insn identify the most recent assignment | |
26 | to each REG used in the insn. It is a list of previous insns, | |
27 | each of which contains a SET for a REG that is used in this insn | |
28 | and not used or set in between. LOG_LINKs never cross basic blocks. | |
29 | They were set up by the preceding pass (lifetime analysis). | |
30 | ||
31 | We try to combine each pair of insns joined by a logical link. | |
32 | We also try to combine triples of insns A, B and C when | |
33 | C has a link back to B and B has a link back to A. | |
34 | ||
35 | LOG_LINKS does not have links for use of the CC0. They don't | |
36 | need to, because the insn that sets the CC0 is always immediately | |
37 | before the insn that tests it. So we always regard a branch | |
38 | insn as having a logical link to the preceding insn. The same is true | |
39 | for an insn explicitly using CC0. | |
40 | ||
41 | We check (with use_crosses_set_p) to avoid combining in such a way | |
42 | as to move a computation to a place where its value would be different. | |
43 | ||
44 | Combination is done by mathematically substituting the previous | |
45 | insn(s) values for the regs they set into the expressions in | |
46 | the later insns that refer to these regs. If the result is a valid insn | |
47 | for our target machine, according to the machine description, | |
48 | we install it, delete the earlier insns, and update the data flow | |
49 | information (LOG_LINKS and REG_NOTES) for what we did. | |
50 | ||
51 | There are a few exceptions where the dataflow information created by | |
52 | flow.c aren't completely updated: | |
53 | ||
54 | - reg_live_length is not updated | |
55 | - reg_n_refs is not adjusted in the rare case when a register is | |
56 | no longer required in a computation | |
57 | - there are extremely rare cases (see distribute_regnotes) when a | |
58 | REG_DEAD note is lost | |
59 | - a LOG_LINKS entry that refers to an insn with multiple SETs may be | |
60 | removed because there is no way to know which register it was | |
61 | linking | |
62 | ||
63 | To simplify substitution, we combine only when the earlier insn(s) | |
64 | consist of only a single assignment. To simplify updating afterward, | |
65 | we never combine when a subroutine call appears in the middle. | |
66 | ||
67 | Since we do not represent assignments to CC0 explicitly except when that | |
68 | is all an insn does, there is no LOG_LINKS entry in an insn that uses | |
69 | the condition code for the insn that set the condition code. | |
70 | Fortunately, these two insns must be consecutive. | |
71 | Therefore, every JUMP_INSN is taken to have an implicit logical link | |
72 | to the preceding insn. This is not quite right, since non-jumps can | |
73 | also use the condition code; but in practice such insns would not | |
74 | combine anyway. */ | |
75 | ||
76 | #include "config.h" | |
77 | #include "gvarargs.h" | |
78 | #include "rtl.h" | |
79 | #include "flags.h" | |
80 | #include "regs.h" | |
81 | #include "hard-reg-set.h" | |
82 | #include "expr.h" | |
83 | #include "basic-block.h" | |
84 | #include "insn-config.h" | |
85 | #include "insn-flags.h" | |
86 | #include "insn-codes.h" | |
87 | #include "insn-attr.h" | |
88 | #include "recog.h" | |
89 | #include "real.h" | |
90 | #include <stdio.h> | |
91 | ||
92 | /* It is not safe to use ordinary gen_lowpart in combine. | |
93 | Use gen_lowpart_for_combine instead. See comments there. */ | |
94 | #define gen_lowpart dont_use_gen_lowpart_you_dummy | |
95 | ||
96 | /* If byte loads either zero- or sign- extend, define BYTE_LOADS_EXTEND | |
97 | for cases when we don't care which is true. Define LOAD_EXTEND to | |
98 | be ZERO_EXTEND or SIGN_EXTEND, depending on which was defined. */ | |
99 | ||
100 | #ifdef BYTE_LOADS_ZERO_EXTEND | |
101 | #define BYTE_LOADS_EXTEND | |
102 | #define LOAD_EXTEND ZERO_EXTEND | |
103 | #endif | |
104 | ||
105 | #ifdef BYTE_LOADS_SIGN_EXTEND | |
106 | #define BYTE_LOADS_EXTEND | |
107 | #define LOAD_EXTEND SIGN_EXTEND | |
108 | #endif | |
109 | ||
110 | /* Number of attempts to combine instructions in this function. */ | |
111 | ||
112 | static int combine_attempts; | |
113 | ||
114 | /* Number of attempts that got as far as substitution in this function. */ | |
115 | ||
116 | static int combine_merges; | |
117 | ||
118 | /* Number of instructions combined with added SETs in this function. */ | |
119 | ||
120 | static int combine_extras; | |
121 | ||
122 | /* Number of instructions combined in this function. */ | |
123 | ||
124 | static int combine_successes; | |
125 | ||
126 | /* Totals over entire compilation. */ | |
127 | ||
128 | static int total_attempts, total_merges, total_extras, total_successes; | |
129 | \f | |
130 | /* Vector mapping INSN_UIDs to cuids. | |
131 | The cuids are like uids but increase monotonically always. | |
132 | Combine always uses cuids so that it can compare them. | |
133 | But actually renumbering the uids, which we used to do, | |
134 | proves to be a bad idea because it makes it hard to compare | |
135 | the dumps produced by earlier passes with those from later passes. */ | |
136 | ||
137 | static int *uid_cuid; | |
138 | ||
139 | /* Get the cuid of an insn. */ | |
140 | ||
141 | #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)]) | |
142 | ||
143 | /* Maximum register number, which is the size of the tables below. */ | |
144 | ||
145 | static int combine_max_regno; | |
146 | ||
147 | /* Record last point of death of (hard or pseudo) register n. */ | |
148 | ||
149 | static rtx *reg_last_death; | |
150 | ||
151 | /* Record last point of modification of (hard or pseudo) register n. */ | |
152 | ||
153 | static rtx *reg_last_set; | |
154 | ||
155 | /* Record the cuid of the last insn that invalidated memory | |
156 | (anything that writes memory, and subroutine calls, but not pushes). */ | |
157 | ||
158 | static int mem_last_set; | |
159 | ||
160 | /* Record the cuid of the last CALL_INSN | |
161 | so we can tell whether a potential combination crosses any calls. */ | |
162 | ||
163 | static int last_call_cuid; | |
164 | ||
165 | /* When `subst' is called, this is the insn that is being modified | |
166 | (by combining in a previous insn). The PATTERN of this insn | |
167 | is still the old pattern partially modified and it should not be | |
168 | looked at, but this may be used to examine the successors of the insn | |
169 | to judge whether a simplification is valid. */ | |
170 | ||
171 | static rtx subst_insn; | |
172 | ||
173 | /* This is the lowest CUID that `subst' is currently dealing with. | |
174 | get_last_value will not return a value if the register was set at or | |
175 | after this CUID. If not for this mechanism, we could get confused if | |
176 | I2 or I1 in try_combine were an insn that used the old value of a register | |
177 | to obtain a new value. In that case, we might erroneously get the | |
178 | new value of the register when we wanted the old one. */ | |
179 | ||
180 | static int subst_low_cuid; | |
181 | ||
182 | /* This is the value of undobuf.num_undo when we started processing this | |
183 | substitution. This will prevent gen_rtx_combine from re-used a piece | |
184 | from the previous expression. Doing so can produce circular rtl | |
185 | structures. */ | |
186 | ||
187 | static int previous_num_undos; | |
188 | \f | |
189 | /* The next group of arrays allows the recording of the last value assigned | |
190 | to (hard or pseudo) register n. We use this information to see if a | |
191 | operation being processed is redundant given a prior operation performed | |
192 | on the register. For example, an `and' with a constant is redundant if | |
193 | all the zero bits are already known to be turned off. | |
194 | ||
195 | We use an approach similar to that used by cse, but change it in the | |
196 | following ways: | |
197 | ||
198 | (1) We do not want to reinitialize at each label. | |
199 | (2) It is useful, but not critical, to know the actual value assigned | |
200 | to a register. Often just its form is helpful. | |
201 | ||
202 | Therefore, we maintain the following arrays: | |
203 | ||
204 | reg_last_set_value the last value assigned | |
205 | reg_last_set_label records the value of label_tick when the | |
206 | register was assigned | |
207 | reg_last_set_table_tick records the value of label_tick when a | |
208 | value using the register is assigned | |
209 | reg_last_set_invalid set to non-zero when it is not valid | |
210 | to use the value of this register in some | |
211 | register's value | |
212 | ||
213 | To understand the usage of these tables, it is important to understand | |
214 | the distinction between the value in reg_last_set_value being valid | |
215 | and the register being validly contained in some other expression in the | |
216 | table. | |
217 | ||
218 | Entry I in reg_last_set_value is valid if it is non-zero, and either | |
219 | reg_n_sets[i] is 1 or reg_last_set_label[i] == label_tick. | |
220 | ||
221 | Register I may validly appear in any expression returned for the value | |
222 | of another register if reg_n_sets[i] is 1. It may also appear in the | |
223 | value for register J if reg_last_set_label[i] < reg_last_set_label[j] or | |
224 | reg_last_set_invalid[j] is zero. | |
225 | ||
226 | If an expression is found in the table containing a register which may | |
227 | not validly appear in an expression, the register is replaced by | |
228 | something that won't match, (clobber (const_int 0)). | |
229 | ||
230 | reg_last_set_invalid[i] is set non-zero when register I is being assigned | |
231 | to and reg_last_set_table_tick[i] == label_tick. */ | |
232 | ||
233 | /* Record last value assigned to (hard or pseudo) register n. */ | |
234 | ||
235 | static rtx *reg_last_set_value; | |
236 | ||
237 | /* Record the value of label_tick when the value for register n is placed in | |
238 | reg_last_set_value[n]. */ | |
239 | ||
240 | static int *reg_last_set_label; | |
241 | ||
242 | /* Record the value of label_tick when an expression involving register n | |
243 | is placed in reg_last_set_value. */ | |
244 | ||
245 | static int *reg_last_set_table_tick; | |
246 | ||
247 | /* Set non-zero if references to register n in expressions should not be | |
248 | used. */ | |
249 | ||
250 | static char *reg_last_set_invalid; | |
251 | ||
252 | /* Incremented for each label. */ | |
253 | ||
254 | static int label_tick; | |
255 | ||
256 | /* Some registers that are set more than once and used in more than one | |
257 | basic block are nevertheless always set in similar ways. For example, | |
258 | a QImode register may be loaded from memory in two places on a machine | |
259 | where byte loads zero extend. | |
260 | ||
261 | We record in the following array what we know about the nonzero | |
262 | bits of a register, specifically which bits are known to be zero. | |
263 | ||
264 | If an entry is zero, it means that we don't know anything special. */ | |
265 | ||
266 | static unsigned HOST_WIDE_INT *reg_nonzero_bits; | |
267 | ||
268 | /* Mode used to compute significance in reg_nonzero_bits. It is the largest | |
269 | integer mode that can fit in HOST_BITS_PER_WIDE_INT. */ | |
270 | ||
271 | static enum machine_mode nonzero_bits_mode; | |
272 | ||
273 | /* Nonzero if we know that a register has some leading bits that are always | |
274 | equal to the sign bit. */ | |
275 | ||
276 | static char *reg_sign_bit_copies; | |
277 | ||
278 | /* Nonzero when reg_nonzero_bits and reg_sign_bit_copies can be safely used. | |
279 | It is zero while computing them and after combine has completed. This | |
280 | former test prevents propagating values based on previously set values, | |
281 | which can be incorrect if a variable is modified in a loop. */ | |
282 | ||
283 | static int nonzero_sign_valid; | |
284 | ||
285 | /* These arrays are maintained in parallel with reg_last_set_value | |
286 | and are used to store the mode in which the register was last set, | |
287 | the bits that were known to be zero when it was last set, and the | |
288 | number of sign bits copies it was known to have when it was last set. */ | |
289 | ||
290 | static enum machine_mode *reg_last_set_mode; | |
291 | static unsigned HOST_WIDE_INT *reg_last_set_nonzero_bits; | |
292 | static char *reg_last_set_sign_bit_copies; | |
293 | \f | |
294 | /* Record one modification to rtl structure | |
295 | to be undone by storing old_contents into *where. | |
296 | is_int is 1 if the contents are an int. */ | |
297 | ||
298 | struct undo | |
299 | { | |
300 | int is_int; | |
301 | union {rtx rtx; int i;} old_contents; | |
302 | union {rtx *rtx; int *i;} where; | |
303 | }; | |
304 | ||
305 | /* Record a bunch of changes to be undone, up to MAX_UNDO of them. | |
306 | num_undo says how many are currently recorded. | |
307 | ||
308 | storage is nonzero if we must undo the allocation of new storage. | |
309 | The value of storage is what to pass to obfree. | |
310 | ||
311 | other_insn is nonzero if we have modified some other insn in the process | |
312 | of working on subst_insn. It must be verified too. */ | |
313 | ||
314 | #define MAX_UNDO 50 | |
315 | ||
316 | struct undobuf | |
317 | { | |
318 | int num_undo; | |
319 | char *storage; | |
320 | struct undo undo[MAX_UNDO]; | |
321 | rtx other_insn; | |
322 | }; | |
323 | ||
324 | static struct undobuf undobuf; | |
325 | ||
326 | /* Substitute NEWVAL, an rtx expression, into INTO, a place in some | |
327 | insn. The substitution can be undone by undo_all. If INTO is already | |
328 | set to NEWVAL, do not record this change. Because computing NEWVAL might | |
329 | also call SUBST, we have to compute it before we put anything into | |
330 | the undo table. */ | |
331 | ||
332 | #define SUBST(INTO, NEWVAL) \ | |
333 | do { rtx _new = (NEWVAL); \ | |
334 | if (undobuf.num_undo < MAX_UNDO) \ | |
335 | { \ | |
336 | undobuf.undo[undobuf.num_undo].is_int = 0; \ | |
337 | undobuf.undo[undobuf.num_undo].where.rtx = &INTO; \ | |
338 | undobuf.undo[undobuf.num_undo].old_contents.rtx = INTO; \ | |
339 | INTO = _new; \ | |
340 | if (undobuf.undo[undobuf.num_undo].old_contents.rtx != INTO) \ | |
341 | undobuf.num_undo++; \ | |
342 | } \ | |
343 | } while (0) | |
344 | ||
345 | /* Similar to SUBST, but NEWVAL is an int. INTO will normally be an XINT | |
346 | expression. | |
347 | Note that substitution for the value of a CONST_INT is not safe. */ | |
348 | ||
349 | #define SUBST_INT(INTO, NEWVAL) \ | |
350 | do { if (undobuf.num_undo < MAX_UNDO) \ | |
351 | { \ | |
352 | undobuf.undo[undobuf.num_undo].is_int = 1; \ | |
353 | undobuf.undo[undobuf.num_undo].where.i = (int *) &INTO; \ | |
354 | undobuf.undo[undobuf.num_undo].old_contents.i = INTO; \ | |
355 | INTO = NEWVAL; \ | |
356 | if (undobuf.undo[undobuf.num_undo].old_contents.i != INTO) \ | |
357 | undobuf.num_undo++; \ | |
358 | } \ | |
359 | } while (0) | |
360 | ||
361 | /* Number of times the pseudo being substituted for | |
362 | was found and replaced. */ | |
363 | ||
364 | static int n_occurrences; | |
365 | ||
366 | static void set_nonzero_bits_and_sign_copies (); | |
367 | static void setup_incoming_promotions (); | |
368 | static void move_deaths (); | |
369 | rtx remove_death (); | |
370 | static void record_value_for_reg (); | |
371 | static void record_dead_and_set_regs (); | |
372 | static int use_crosses_set_p (); | |
373 | static rtx try_combine (); | |
374 | static rtx *find_split_point (); | |
375 | static rtx subst (); | |
376 | static void undo_all (); | |
377 | static int reg_dead_at_p (); | |
378 | static rtx expand_compound_operation (); | |
379 | static rtx expand_field_assignment (); | |
380 | static rtx make_extraction (); | |
381 | static int get_pos_from_mask (); | |
382 | static rtx force_to_mode (); | |
383 | static rtx known_cond (); | |
384 | static rtx make_field_assignment (); | |
385 | static rtx make_compound_operation (); | |
386 | static rtx apply_distributive_law (); | |
387 | static rtx simplify_and_const_int (); | |
388 | static unsigned HOST_WIDE_INT nonzero_bits (); | |
389 | static int num_sign_bit_copies (); | |
390 | static int merge_outer_ops (); | |
391 | static rtx simplify_shift_const (); | |
392 | static int recog_for_combine (); | |
393 | static rtx gen_lowpart_for_combine (); | |
394 | static rtx gen_rtx_combine (); | |
395 | static rtx gen_binary (); | |
396 | static rtx gen_unary (); | |
397 | static enum rtx_code simplify_comparison (); | |
398 | static int reversible_comparison_p (); | |
399 | static int get_last_value_validate (); | |
400 | static rtx get_last_value (); | |
401 | static void distribute_notes (); | |
402 | static void distribute_links (); | |
403 | \f | |
404 | /* Main entry point for combiner. F is the first insn of the function. | |
405 | NREGS is the first unused pseudo-reg number. */ | |
406 | ||
407 | void | |
408 | combine_instructions (f, nregs) | |
409 | rtx f; | |
410 | int nregs; | |
411 | { | |
412 | register rtx insn, next, prev; | |
413 | register int i; | |
414 | register rtx links, nextlinks; | |
415 | ||
416 | combine_attempts = 0; | |
417 | combine_merges = 0; | |
418 | combine_extras = 0; | |
419 | combine_successes = 0; | |
420 | undobuf.num_undo = previous_num_undos = 0; | |
421 | ||
422 | combine_max_regno = nregs; | |
423 | ||
424 | reg_last_death = (rtx *) alloca (nregs * sizeof (rtx)); | |
425 | reg_last_set = (rtx *) alloca (nregs * sizeof (rtx)); | |
426 | reg_last_set_value = (rtx *) alloca (nregs * sizeof (rtx)); | |
427 | reg_last_set_table_tick = (int *) alloca (nregs * sizeof (int)); | |
428 | reg_last_set_label = (int *) alloca (nregs * sizeof (int)); | |
429 | reg_last_set_invalid = (char *) alloca (nregs * sizeof (char)); | |
430 | reg_last_set_mode | |
431 | = (enum machine_mode *) alloca (nregs * sizeof (enum machine_mode)); | |
432 | reg_last_set_nonzero_bits | |
433 | = (unsigned HOST_WIDE_INT *) alloca (nregs * sizeof (HOST_WIDE_INT)); | |
434 | reg_last_set_sign_bit_copies | |
435 | = (char *) alloca (nregs * sizeof (char)); | |
436 | ||
437 | reg_nonzero_bits | |
438 | = (unsigned HOST_WIDE_INT *) alloca (nregs * sizeof (HOST_WIDE_INT)); | |
439 | reg_sign_bit_copies = (char *) alloca (nregs * sizeof (char)); | |
440 | ||
441 | bzero (reg_last_death, nregs * sizeof (rtx)); | |
442 | bzero (reg_last_set, nregs * sizeof (rtx)); | |
443 | bzero (reg_last_set_value, nregs * sizeof (rtx)); | |
444 | bzero (reg_last_set_table_tick, nregs * sizeof (int)); | |
445 | bzero (reg_last_set_label, nregs * sizeof (int)); | |
446 | bzero (reg_last_set_invalid, nregs * sizeof (char)); | |
447 | bzero (reg_last_set_mode, nregs * sizeof (enum machine_mode)); | |
448 | bzero (reg_last_set_nonzero_bits, nregs * sizeof (HOST_WIDE_INT)); | |
449 | bzero (reg_last_set_sign_bit_copies, nregs * sizeof (char)); | |
450 | bzero (reg_nonzero_bits, nregs * sizeof (HOST_WIDE_INT)); | |
451 | bzero (reg_sign_bit_copies, nregs * sizeof (char)); | |
452 | ||
453 | init_recog_no_volatile (); | |
454 | ||
455 | /* Compute maximum uid value so uid_cuid can be allocated. */ | |
456 | ||
457 | for (insn = f, i = 0; insn; insn = NEXT_INSN (insn)) | |
458 | if (INSN_UID (insn) > i) | |
459 | i = INSN_UID (insn); | |
460 | ||
461 | uid_cuid = (int *) alloca ((i + 1) * sizeof (int)); | |
462 | ||
463 | nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0); | |
464 | ||
465 | /* Don't use reg_nonzero_bits when computing it. This can cause problems | |
466 | when, for example, we have j <<= 1 in a loop. */ | |
467 | ||
468 | nonzero_sign_valid = 0; | |
469 | ||
470 | /* Compute the mapping from uids to cuids. | |
471 | Cuids are numbers assigned to insns, like uids, | |
472 | except that cuids increase monotonically through the code. | |
473 | ||
474 | Scan all SETs and see if we can deduce anything about what | |
475 | bits are known to be zero for some registers and how many copies | |
476 | of the sign bit are known to exist for those registers. | |
477 | ||
478 | Also set any known values so that we can use it while searching | |
479 | for what bits are known to be set. */ | |
480 | ||
481 | label_tick = 1; | |
482 | ||
483 | setup_incoming_promotions (); | |
484 | ||
485 | for (insn = f, i = 0; insn; insn = NEXT_INSN (insn)) | |
486 | { | |
487 | INSN_CUID (insn) = ++i; | |
488 | subst_low_cuid = i; | |
489 | subst_insn = insn; | |
490 | ||
491 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i') | |
492 | { | |
493 | note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies); | |
494 | record_dead_and_set_regs (insn); | |
495 | } | |
496 | ||
497 | if (GET_CODE (insn) == CODE_LABEL) | |
498 | label_tick++; | |
499 | } | |
500 | ||
501 | nonzero_sign_valid = 1; | |
502 | ||
503 | /* Now scan all the insns in forward order. */ | |
504 | ||
505 | label_tick = 1; | |
506 | last_call_cuid = 0; | |
507 | mem_last_set = 0; | |
508 | bzero (reg_last_death, nregs * sizeof (rtx)); | |
509 | bzero (reg_last_set, nregs * sizeof (rtx)); | |
510 | bzero (reg_last_set_value, nregs * sizeof (rtx)); | |
511 | bzero (reg_last_set_table_tick, nregs * sizeof (int)); | |
512 | bzero (reg_last_set_label, nregs * sizeof (int)); | |
513 | bzero (reg_last_set_invalid, nregs * sizeof (char)); | |
514 | ||
515 | setup_incoming_promotions (); | |
516 | ||
517 | for (insn = f; insn; insn = next ? next : NEXT_INSN (insn)) | |
518 | { | |
519 | next = 0; | |
520 | ||
521 | if (GET_CODE (insn) == CODE_LABEL) | |
522 | label_tick++; | |
523 | ||
524 | else if (GET_CODE (insn) == INSN | |
525 | || GET_CODE (insn) == CALL_INSN | |
526 | || GET_CODE (insn) == JUMP_INSN) | |
527 | { | |
528 | /* Try this insn with each insn it links back to. */ | |
529 | ||
530 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
531 | if ((next = try_combine (insn, XEXP (links, 0), NULL_RTX)) != 0) | |
532 | goto retry; | |
533 | ||
534 | /* Try each sequence of three linked insns ending with this one. */ | |
535 | ||
536 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
537 | for (nextlinks = LOG_LINKS (XEXP (links, 0)); nextlinks; | |
538 | nextlinks = XEXP (nextlinks, 1)) | |
539 | if ((next = try_combine (insn, XEXP (links, 0), | |
540 | XEXP (nextlinks, 0))) != 0) | |
541 | goto retry; | |
542 | ||
543 | #ifdef HAVE_cc0 | |
544 | /* Try to combine a jump insn that uses CC0 | |
545 | with a preceding insn that sets CC0, and maybe with its | |
546 | logical predecessor as well. | |
547 | This is how we make decrement-and-branch insns. | |
548 | We need this special code because data flow connections | |
549 | via CC0 do not get entered in LOG_LINKS. */ | |
550 | ||
551 | if (GET_CODE (insn) == JUMP_INSN | |
552 | && (prev = prev_nonnote_insn (insn)) != 0 | |
553 | && GET_CODE (prev) == INSN | |
554 | && sets_cc0_p (PATTERN (prev))) | |
555 | { | |
556 | if ((next = try_combine (insn, prev, NULL_RTX)) != 0) | |
557 | goto retry; | |
558 | ||
559 | for (nextlinks = LOG_LINKS (prev); nextlinks; | |
560 | nextlinks = XEXP (nextlinks, 1)) | |
561 | if ((next = try_combine (insn, prev, | |
562 | XEXP (nextlinks, 0))) != 0) | |
563 | goto retry; | |
564 | } | |
565 | ||
566 | /* Do the same for an insn that explicitly references CC0. */ | |
567 | if (GET_CODE (insn) == INSN | |
568 | && (prev = prev_nonnote_insn (insn)) != 0 | |
569 | && GET_CODE (prev) == INSN | |
570 | && sets_cc0_p (PATTERN (prev)) | |
571 | && GET_CODE (PATTERN (insn)) == SET | |
572 | && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn)))) | |
573 | { | |
574 | if ((next = try_combine (insn, prev, NULL_RTX)) != 0) | |
575 | goto retry; | |
576 | ||
577 | for (nextlinks = LOG_LINKS (prev); nextlinks; | |
578 | nextlinks = XEXP (nextlinks, 1)) | |
579 | if ((next = try_combine (insn, prev, | |
580 | XEXP (nextlinks, 0))) != 0) | |
581 | goto retry; | |
582 | } | |
583 | ||
584 | /* Finally, see if any of the insns that this insn links to | |
585 | explicitly references CC0. If so, try this insn, that insn, | |
586 | and its predecessor if it sets CC0. */ | |
587 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
588 | if (GET_CODE (XEXP (links, 0)) == INSN | |
589 | && GET_CODE (PATTERN (XEXP (links, 0))) == SET | |
590 | && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0)))) | |
591 | && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0 | |
592 | && GET_CODE (prev) == INSN | |
593 | && sets_cc0_p (PATTERN (prev)) | |
594 | && (next = try_combine (insn, XEXP (links, 0), prev)) != 0) | |
595 | goto retry; | |
596 | #endif | |
597 | ||
598 | /* Try combining an insn with two different insns whose results it | |
599 | uses. */ | |
600 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
601 | for (nextlinks = XEXP (links, 1); nextlinks; | |
602 | nextlinks = XEXP (nextlinks, 1)) | |
603 | if ((next = try_combine (insn, XEXP (links, 0), | |
604 | XEXP (nextlinks, 0))) != 0) | |
605 | goto retry; | |
606 | ||
607 | if (GET_CODE (insn) != NOTE) | |
608 | record_dead_and_set_regs (insn); | |
609 | ||
610 | retry: | |
611 | ; | |
612 | } | |
613 | } | |
614 | ||
615 | total_attempts += combine_attempts; | |
616 | total_merges += combine_merges; | |
617 | total_extras += combine_extras; | |
618 | total_successes += combine_successes; | |
619 | ||
620 | nonzero_sign_valid = 0; | |
621 | } | |
622 | \f | |
623 | /* Set up any promoted values for incoming argument registers. */ | |
624 | ||
625 | static void | |
626 | setup_incoming_promotions () | |
627 | { | |
628 | #ifdef PROMOTE_FUNCTION_ARGS | |
629 | int regno; | |
630 | rtx reg; | |
631 | enum machine_mode mode; | |
632 | int unsignedp; | |
633 | rtx first = get_insns (); | |
634 | ||
635 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
636 | if (FUNCTION_ARG_REGNO_P (regno) | |
637 | && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0) | |
638 | record_value_for_reg (reg, first, | |
639 | gen_rtx (unsignedp ? ZERO_EXTEND : SIGN_EXTEND, | |
640 | GET_MODE (reg), | |
641 | gen_rtx (CLOBBER, mode, const0_rtx))); | |
642 | #endif | |
643 | } | |
644 | \f | |
645 | /* Called via note_stores. If X is a pseudo that is used in more than | |
646 | one basic block, is narrower that HOST_BITS_PER_WIDE_INT, and is being | |
647 | set, record what bits are known zero. If we are clobbering X, | |
648 | ignore this "set" because the clobbered value won't be used. | |
649 | ||
650 | If we are setting only a portion of X and we can't figure out what | |
651 | portion, assume all bits will be used since we don't know what will | |
652 | be happening. | |
653 | ||
654 | Similarly, set how many bits of X are known to be copies of the sign bit | |
655 | at all locations in the function. This is the smallest number implied | |
656 | by any set of X. */ | |
657 | ||
658 | static void | |
659 | set_nonzero_bits_and_sign_copies (x, set) | |
660 | rtx x; | |
661 | rtx set; | |
662 | { | |
663 | int num; | |
664 | ||
665 | if (GET_CODE (x) == REG | |
666 | && REGNO (x) >= FIRST_PSEUDO_REGISTER | |
667 | && reg_n_sets[REGNO (x)] > 1 | |
668 | && reg_basic_block[REGNO (x)] < 0 | |
669 | /* If this register is undefined at the start of the file, we can't | |
670 | say what its contents were. */ | |
671 | && ! (basic_block_live_at_start[0][REGNO (x) / REGSET_ELT_BITS] | |
672 | & ((REGSET_ELT_TYPE) 1 << (REGNO (x) % REGSET_ELT_BITS))) | |
673 | && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT) | |
674 | { | |
675 | if (GET_CODE (set) == CLOBBER) | |
676 | { | |
677 | reg_nonzero_bits[REGNO (x)] = GET_MODE_MASK (GET_MODE (x)); | |
678 | reg_sign_bit_copies[REGNO (x)] = 0; | |
679 | return; | |
680 | } | |
681 | ||
682 | /* If this is a complex assignment, see if we can convert it into a | |
683 | simple assignment. */ | |
684 | set = expand_field_assignment (set); | |
685 | ||
686 | /* If this is a simple assignment, or we have a paradoxical SUBREG, | |
687 | set what we know about X. */ | |
688 | ||
689 | if (SET_DEST (set) == x | |
690 | || (GET_CODE (SET_DEST (set)) == SUBREG | |
691 | && (GET_MODE_SIZE (GET_MODE (SET_DEST (set))) | |
692 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set))))) | |
693 | && SUBREG_REG (SET_DEST (set)) == x)) | |
694 | { | |
695 | rtx src = SET_SRC (set); | |
696 | ||
697 | #ifdef SHORT_IMMEDIATES_SIGN_EXTEND | |
698 | /* If X is narrower than a word and SRC is a non-negative | |
699 | constant that would appear negative in the mode of X, | |
700 | sign-extend it for use in reg_nonzero_bits because some | |
701 | machines (maybe most) will actually do the sign-extension | |
702 | and this is the conservative approach. | |
703 | ||
704 | ??? For 2.5, try to tighten up the MD files in this regard | |
705 | instead of this kludge. */ | |
706 | ||
707 | if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD | |
708 | && GET_CODE (src) == CONST_INT | |
709 | && INTVAL (src) > 0 | |
710 | && 0 != (INTVAL (src) | |
711 | & ((HOST_WIDE_INT) 1 | |
712 | << GET_MODE_BITSIZE (GET_MODE (x))))) | |
713 | src = GEN_INT (INTVAL (src) | |
714 | | ((HOST_WIDE_INT) (-1) | |
715 | << GET_MODE_BITSIZE (GET_MODE (x)))); | |
716 | #endif | |
717 | ||
718 | reg_nonzero_bits[REGNO (x)] | |
719 | |= nonzero_bits (src, nonzero_bits_mode); | |
720 | num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x)); | |
721 | if (reg_sign_bit_copies[REGNO (x)] == 0 | |
722 | || reg_sign_bit_copies[REGNO (x)] > num) | |
723 | reg_sign_bit_copies[REGNO (x)] = num; | |
724 | } | |
725 | else | |
726 | { | |
727 | reg_nonzero_bits[REGNO (x)] = GET_MODE_MASK (GET_MODE (x)); | |
728 | reg_sign_bit_copies[REGNO (x)] = 0; | |
729 | } | |
730 | } | |
731 | } | |
732 | \f | |
733 | /* See if INSN can be combined into I3. PRED and SUCC are optionally | |
734 | insns that were previously combined into I3 or that will be combined | |
735 | into the merger of INSN and I3. | |
736 | ||
737 | Return 0 if the combination is not allowed for any reason. | |
738 | ||
739 | If the combination is allowed, *PDEST will be set to the single | |
740 | destination of INSN and *PSRC to the single source, and this function | |
741 | will return 1. */ | |
742 | ||
743 | static int | |
744 | can_combine_p (insn, i3, pred, succ, pdest, psrc) | |
745 | rtx insn; | |
746 | rtx i3; | |
747 | rtx pred, succ; | |
748 | rtx *pdest, *psrc; | |
749 | { | |
750 | int i; | |
751 | rtx set = 0, src, dest; | |
752 | rtx p, link; | |
753 | int all_adjacent = (succ ? (next_active_insn (insn) == succ | |
754 | && next_active_insn (succ) == i3) | |
755 | : next_active_insn (insn) == i3); | |
756 | ||
757 | /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0. | |
758 | or a PARALLEL consisting of such a SET and CLOBBERs. | |
759 | ||
760 | If INSN has CLOBBER parallel parts, ignore them for our processing. | |
761 | By definition, these happen during the execution of the insn. When it | |
762 | is merged with another insn, all bets are off. If they are, in fact, | |
763 | needed and aren't also supplied in I3, they may be added by | |
764 | recog_for_combine. Otherwise, it won't match. | |
765 | ||
766 | We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED | |
767 | note. | |
768 | ||
769 | Get the source and destination of INSN. If more than one, can't | |
770 | combine. */ | |
771 | ||
772 | if (GET_CODE (PATTERN (insn)) == SET) | |
773 | set = PATTERN (insn); | |
774 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
775 | && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET) | |
776 | { | |
777 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
778 | { | |
779 | rtx elt = XVECEXP (PATTERN (insn), 0, i); | |
780 | ||
781 | switch (GET_CODE (elt)) | |
782 | { | |
783 | /* We can ignore CLOBBERs. */ | |
784 | case CLOBBER: | |
785 | break; | |
786 | ||
787 | case SET: | |
788 | /* Ignore SETs whose result isn't used but not those that | |
789 | have side-effects. */ | |
790 | if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt)) | |
791 | && ! side_effects_p (elt)) | |
792 | break; | |
793 | ||
794 | /* If we have already found a SET, this is a second one and | |
795 | so we cannot combine with this insn. */ | |
796 | if (set) | |
797 | return 0; | |
798 | ||
799 | set = elt; | |
800 | break; | |
801 | ||
802 | default: | |
803 | /* Anything else means we can't combine. */ | |
804 | return 0; | |
805 | } | |
806 | } | |
807 | ||
808 | if (set == 0 | |
809 | /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs, | |
810 | so don't do anything with it. */ | |
811 | || GET_CODE (SET_SRC (set)) == ASM_OPERANDS) | |
812 | return 0; | |
813 | } | |
814 | else | |
815 | return 0; | |
816 | ||
817 | if (set == 0) | |
818 | return 0; | |
819 | ||
820 | set = expand_field_assignment (set); | |
821 | src = SET_SRC (set), dest = SET_DEST (set); | |
822 | ||
823 | /* Don't eliminate a store in the stack pointer. */ | |
824 | if (dest == stack_pointer_rtx | |
825 | /* Don't install a subreg involving two modes not tieable. | |
826 | It can worsen register allocation, and can even make invalid reload | |
827 | insns, since the reg inside may need to be copied from in the | |
828 | outside mode, and that may be invalid if it is an fp reg copied in | |
829 | integer mode. As a special exception, we can allow this if | |
830 | I3 is simply copying DEST, a REG, to CC0. */ | |
831 | || (GET_CODE (src) == SUBREG | |
832 | && ! MODES_TIEABLE_P (GET_MODE (src), GET_MODE (SUBREG_REG (src))) | |
833 | #ifdef HAVE_cc0 | |
834 | && ! (GET_CODE (i3) == INSN && GET_CODE (PATTERN (i3)) == SET | |
835 | && SET_DEST (PATTERN (i3)) == cc0_rtx | |
836 | && GET_CODE (dest) == REG && dest == SET_SRC (PATTERN (i3))) | |
837 | #endif | |
838 | ) | |
839 | /* If we couldn't eliminate a field assignment, we can't combine. */ | |
840 | || GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == STRICT_LOW_PART | |
841 | /* Don't combine with an insn that sets a register to itself if it has | |
842 | a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */ | |
843 | || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX)) | |
844 | /* Can't merge a function call. */ | |
845 | || GET_CODE (src) == CALL | |
846 | /* Don't substitute into an incremented register. */ | |
847 | || FIND_REG_INC_NOTE (i3, dest) | |
848 | || (succ && FIND_REG_INC_NOTE (succ, dest)) | |
849 | /* Don't combine the end of a libcall into anything. */ | |
850 | || find_reg_note (insn, REG_RETVAL, NULL_RTX) | |
851 | /* Make sure that DEST is not used after SUCC but before I3. */ | |
852 | || (succ && ! all_adjacent | |
853 | && reg_used_between_p (dest, succ, i3)) | |
854 | /* Make sure that the value that is to be substituted for the register | |
855 | does not use any registers whose values alter in between. However, | |
856 | If the insns are adjacent, a use can't cross a set even though we | |
857 | think it might (this can happen for a sequence of insns each setting | |
858 | the same destination; reg_last_set of that register might point to | |
859 | a NOTE). Also, don't move a volatile asm or UNSPEC_VOLATILE across | |
860 | any other insns. */ | |
861 | || (! all_adjacent | |
862 | && (use_crosses_set_p (src, INSN_CUID (insn)) | |
863 | || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src)) | |
864 | || GET_CODE (src) == UNSPEC_VOLATILE)) | |
865 | /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get | |
866 | better register allocation by not doing the combine. */ | |
867 | || find_reg_note (i3, REG_NO_CONFLICT, dest) | |
868 | || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest)) | |
869 | /* Don't combine across a CALL_INSN, because that would possibly | |
870 | change whether the life span of some REGs crosses calls or not, | |
871 | and it is a pain to update that information. | |
872 | Exception: if source is a constant, moving it later can't hurt. | |
873 | Accept that special case, because it helps -fforce-addr a lot. */ | |
874 | || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src))) | |
875 | return 0; | |
876 | ||
877 | /* DEST must either be a REG or CC0. */ | |
878 | if (GET_CODE (dest) == REG) | |
879 | { | |
880 | /* If register alignment is being enforced for multi-word items in all | |
881 | cases except for parameters, it is possible to have a register copy | |
882 | insn referencing a hard register that is not allowed to contain the | |
883 | mode being copied and which would not be valid as an operand of most | |
884 | insns. Eliminate this problem by not combining with such an insn. | |
885 | ||
886 | Also, on some machines we don't want to extend the life of a hard | |
887 | register. */ | |
888 | ||
889 | if (GET_CODE (src) == REG | |
890 | && ((REGNO (dest) < FIRST_PSEUDO_REGISTER | |
891 | && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest))) | |
892 | #ifdef SMALL_REGISTER_CLASSES | |
893 | /* Don't extend the life of a hard register. */ | |
894 | || REGNO (src) < FIRST_PSEUDO_REGISTER | |
895 | #else | |
896 | || (REGNO (src) < FIRST_PSEUDO_REGISTER | |
897 | && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src))) | |
898 | #endif | |
899 | )) | |
900 | return 0; | |
901 | } | |
902 | else if (GET_CODE (dest) != CC0) | |
903 | return 0; | |
904 | ||
905 | /* Don't substitute for a register intended as a clobberable operand. | |
906 | Similarly, don't substitute an expression containing a register that | |
907 | will be clobbered in I3. */ | |
908 | if (GET_CODE (PATTERN (i3)) == PARALLEL) | |
909 | for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--) | |
910 | if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER | |
911 | && (reg_overlap_mentioned_p (XEXP (XVECEXP (PATTERN (i3), 0, i), 0), | |
912 | src) | |
913 | || rtx_equal_p (XEXP (XVECEXP (PATTERN (i3), 0, i), 0), dest))) | |
914 | return 0; | |
915 | ||
916 | /* If INSN contains anything volatile, or is an `asm' (whether volatile | |
917 | or not), reject, unless nothing volatile comes between it and I3, | |
918 | with the exception of SUCC. */ | |
919 | ||
920 | if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src)) | |
921 | for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p)) | |
922 | if (GET_RTX_CLASS (GET_CODE (p)) == 'i' | |
923 | && p != succ && volatile_refs_p (PATTERN (p))) | |
924 | return 0; | |
925 | ||
926 | /* If INSN or I2 contains an autoincrement or autodecrement, | |
927 | make sure that register is not used between there and I3, | |
928 | and not already used in I3 either. | |
929 | Also insist that I3 not be a jump; if it were one | |
930 | and the incremented register were spilled, we would lose. */ | |
931 | ||
932 | #ifdef AUTO_INC_DEC | |
933 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
934 | if (REG_NOTE_KIND (link) == REG_INC | |
935 | && (GET_CODE (i3) == JUMP_INSN | |
936 | || reg_used_between_p (XEXP (link, 0), insn, i3) | |
937 | || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3)))) | |
938 | return 0; | |
939 | #endif | |
940 | ||
941 | #ifdef HAVE_cc0 | |
942 | /* Don't combine an insn that follows a CC0-setting insn. | |
943 | An insn that uses CC0 must not be separated from the one that sets it. | |
944 | We do, however, allow I2 to follow a CC0-setting insn if that insn | |
945 | is passed as I1; in that case it will be deleted also. | |
946 | We also allow combining in this case if all the insns are adjacent | |
947 | because that would leave the two CC0 insns adjacent as well. | |
948 | It would be more logical to test whether CC0 occurs inside I1 or I2, | |
949 | but that would be much slower, and this ought to be equivalent. */ | |
950 | ||
951 | p = prev_nonnote_insn (insn); | |
952 | if (p && p != pred && GET_CODE (p) == INSN && sets_cc0_p (PATTERN (p)) | |
953 | && ! all_adjacent) | |
954 | return 0; | |
955 | #endif | |
956 | ||
957 | /* If we get here, we have passed all the tests and the combination is | |
958 | to be allowed. */ | |
959 | ||
960 | *pdest = dest; | |
961 | *psrc = src; | |
962 | ||
963 | return 1; | |
964 | } | |
965 | \f | |
966 | /* LOC is the location within I3 that contains its pattern or the component | |
967 | of a PARALLEL of the pattern. We validate that it is valid for combining. | |
968 | ||
969 | One problem is if I3 modifies its output, as opposed to replacing it | |
970 | entirely, we can't allow the output to contain I2DEST or I1DEST as doing | |
971 | so would produce an insn that is not equivalent to the original insns. | |
972 | ||
973 | Consider: | |
974 | ||
975 | (set (reg:DI 101) (reg:DI 100)) | |
976 | (set (subreg:SI (reg:DI 101) 0) <foo>) | |
977 | ||
978 | This is NOT equivalent to: | |
979 | ||
980 | (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>) | |
981 | (set (reg:DI 101) (reg:DI 100))]) | |
982 | ||
983 | Not only does this modify 100 (in which case it might still be valid | |
984 | if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100. | |
985 | ||
986 | We can also run into a problem if I2 sets a register that I1 | |
987 | uses and I1 gets directly substituted into I3 (not via I2). In that | |
988 | case, we would be getting the wrong value of I2DEST into I3, so we | |
989 | must reject the combination. This case occurs when I2 and I1 both | |
990 | feed into I3, rather than when I1 feeds into I2, which feeds into I3. | |
991 | If I1_NOT_IN_SRC is non-zero, it means that finding I1 in the source | |
992 | of a SET must prevent combination from occurring. | |
993 | ||
994 | On machines where SMALL_REGISTER_CLASSES is defined, we don't combine | |
995 | if the destination of a SET is a hard register. | |
996 | ||
997 | Before doing the above check, we first try to expand a field assignment | |
998 | into a set of logical operations. | |
999 | ||
1000 | If PI3_DEST_KILLED is non-zero, it is a pointer to a location in which | |
1001 | we place a register that is both set and used within I3. If more than one | |
1002 | such register is detected, we fail. | |
1003 | ||
1004 | Return 1 if the combination is valid, zero otherwise. */ | |
1005 | ||
1006 | static int | |
1007 | combinable_i3pat (i3, loc, i2dest, i1dest, i1_not_in_src, pi3dest_killed) | |
1008 | rtx i3; | |
1009 | rtx *loc; | |
1010 | rtx i2dest; | |
1011 | rtx i1dest; | |
1012 | int i1_not_in_src; | |
1013 | rtx *pi3dest_killed; | |
1014 | { | |
1015 | rtx x = *loc; | |
1016 | ||
1017 | if (GET_CODE (x) == SET) | |
1018 | { | |
1019 | rtx set = expand_field_assignment (x); | |
1020 | rtx dest = SET_DEST (set); | |
1021 | rtx src = SET_SRC (set); | |
1022 | rtx inner_dest = dest, inner_src = src; | |
1023 | ||
1024 | SUBST (*loc, set); | |
1025 | ||
1026 | while (GET_CODE (inner_dest) == STRICT_LOW_PART | |
1027 | || GET_CODE (inner_dest) == SUBREG | |
1028 | || GET_CODE (inner_dest) == ZERO_EXTRACT) | |
1029 | inner_dest = XEXP (inner_dest, 0); | |
1030 | ||
1031 | /* We probably don't need this any more now that LIMIT_RELOAD_CLASS | |
1032 | was added. */ | |
1033 | #if 0 | |
1034 | while (GET_CODE (inner_src) == STRICT_LOW_PART | |
1035 | || GET_CODE (inner_src) == SUBREG | |
1036 | || GET_CODE (inner_src) == ZERO_EXTRACT) | |
1037 | inner_src = XEXP (inner_src, 0); | |
1038 | ||
1039 | /* If it is better that two different modes keep two different pseudos, | |
1040 | avoid combining them. This avoids producing the following pattern | |
1041 | on a 386: | |
1042 | (set (subreg:SI (reg/v:QI 21) 0) | |
1043 | (lshiftrt:SI (reg/v:SI 20) | |
1044 | (const_int 24))) | |
1045 | If that were made, reload could not handle the pair of | |
1046 | reg 20/21, since it would try to get any GENERAL_REGS | |
1047 | but some of them don't handle QImode. */ | |
1048 | ||
1049 | if (rtx_equal_p (inner_src, i2dest) | |
1050 | && GET_CODE (inner_dest) == REG | |
1051 | && ! MODES_TIEABLE_P (GET_MODE (i2dest), GET_MODE (inner_dest))) | |
1052 | return 0; | |
1053 | #endif | |
1054 | ||
1055 | /* Check for the case where I3 modifies its output, as | |
1056 | discussed above. */ | |
1057 | if ((inner_dest != dest | |
1058 | && (reg_overlap_mentioned_p (i2dest, inner_dest) | |
1059 | || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest)))) | |
1060 | /* This is the same test done in can_combine_p except that we | |
1061 | allow a hard register with SMALL_REGISTER_CLASSES if SRC is a | |
1062 | CALL operation. */ | |
1063 | || (GET_CODE (inner_dest) == REG | |
1064 | && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER | |
1065 | #ifdef SMALL_REGISTER_CLASSES | |
1066 | && GET_CODE (src) != CALL | |
1067 | #else | |
1068 | && ! HARD_REGNO_MODE_OK (REGNO (inner_dest), | |
1069 | GET_MODE (inner_dest)) | |
1070 | #endif | |
1071 | ) | |
1072 | ||
1073 | || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src))) | |
1074 | return 0; | |
1075 | ||
1076 | /* If DEST is used in I3, it is being killed in this insn, | |
1077 | so record that for later. | |
1078 | Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the | |
1079 | STACK_POINTER_REGNUM, since these are always considered to be | |
1080 | live. Similarly for ARG_POINTER_REGNUM if it is fixed. */ | |
1081 | if (pi3dest_killed && GET_CODE (dest) == REG | |
1082 | && reg_referenced_p (dest, PATTERN (i3)) | |
1083 | && REGNO (dest) != FRAME_POINTER_REGNUM | |
1084 | #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM | |
1085 | && (REGNO (dest) != ARG_POINTER_REGNUM | |
1086 | || ! fixed_regs [REGNO (dest)]) | |
1087 | #endif | |
1088 | && REGNO (dest) != STACK_POINTER_REGNUM) | |
1089 | { | |
1090 | if (*pi3dest_killed) | |
1091 | return 0; | |
1092 | ||
1093 | *pi3dest_killed = dest; | |
1094 | } | |
1095 | } | |
1096 | ||
1097 | else if (GET_CODE (x) == PARALLEL) | |
1098 | { | |
1099 | int i; | |
1100 | ||
1101 | for (i = 0; i < XVECLEN (x, 0); i++) | |
1102 | if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest, | |
1103 | i1_not_in_src, pi3dest_killed)) | |
1104 | return 0; | |
1105 | } | |
1106 | ||
1107 | return 1; | |
1108 | } | |
1109 | \f | |
1110 | /* Try to combine the insns I1 and I2 into I3. | |
1111 | Here I1 and I2 appear earlier than I3. | |
1112 | I1 can be zero; then we combine just I2 into I3. | |
1113 | ||
1114 | It we are combining three insns and the resulting insn is not recognized, | |
1115 | try splitting it into two insns. If that happens, I2 and I3 are retained | |
1116 | and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2 | |
1117 | are pseudo-deleted. | |
1118 | ||
1119 | If we created two insns, return I2; otherwise return I3. | |
1120 | Return 0 if the combination does not work. Then nothing is changed. */ | |
1121 | ||
1122 | static rtx | |
1123 | try_combine (i3, i2, i1) | |
1124 | register rtx i3, i2, i1; | |
1125 | { | |
1126 | /* New patterns for I3 and I3, respectively. */ | |
1127 | rtx newpat, newi2pat = 0; | |
1128 | /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */ | |
1129 | int added_sets_1, added_sets_2; | |
1130 | /* Total number of SETs to put into I3. */ | |
1131 | int total_sets; | |
1132 | /* Nonzero is I2's body now appears in I3. */ | |
1133 | int i2_is_used; | |
1134 | /* INSN_CODEs for new I3, new I2, and user of condition code. */ | |
1135 | int insn_code_number, i2_code_number, other_code_number; | |
1136 | /* Contains I3 if the destination of I3 is used in its source, which means | |
1137 | that the old life of I3 is being killed. If that usage is placed into | |
1138 | I2 and not in I3, a REG_DEAD note must be made. */ | |
1139 | rtx i3dest_killed = 0; | |
1140 | /* SET_DEST and SET_SRC of I2 and I1. */ | |
1141 | rtx i2dest, i2src, i1dest = 0, i1src = 0; | |
1142 | /* PATTERN (I2), or a copy of it in certain cases. */ | |
1143 | rtx i2pat; | |
1144 | /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */ | |
1145 | int i2dest_in_i2src, i1dest_in_i1src = 0, i2dest_in_i1src = 0; | |
1146 | int i1_feeds_i3 = 0; | |
1147 | /* Notes that must be added to REG_NOTES in I3 and I2. */ | |
1148 | rtx new_i3_notes, new_i2_notes; | |
1149 | ||
1150 | int maxreg; | |
1151 | rtx temp; | |
1152 | register rtx link; | |
1153 | int i; | |
1154 | ||
1155 | /* If any of I1, I2, and I3 isn't really an insn, we can't do anything. | |
1156 | This can occur when flow deletes an insn that it has merged into an | |
1157 | auto-increment address. We also can't do anything if I3 has a | |
1158 | REG_LIBCALL note since we don't want to disrupt the contiguity of a | |
1159 | libcall. */ | |
1160 | ||
1161 | if (GET_RTX_CLASS (GET_CODE (i3)) != 'i' | |
1162 | || GET_RTX_CLASS (GET_CODE (i2)) != 'i' | |
1163 | || (i1 && GET_RTX_CLASS (GET_CODE (i1)) != 'i') | |
1164 | || find_reg_note (i3, REG_LIBCALL, NULL_RTX)) | |
1165 | return 0; | |
1166 | ||
1167 | combine_attempts++; | |
1168 | ||
1169 | undobuf.num_undo = previous_num_undos = 0; | |
1170 | undobuf.other_insn = 0; | |
1171 | ||
1172 | /* Save the current high-water-mark so we can free storage if we didn't | |
1173 | accept this combination. */ | |
1174 | undobuf.storage = (char *) oballoc (0); | |
1175 | ||
1176 | /* If I1 and I2 both feed I3, they can be in any order. To simplify the | |
1177 | code below, set I1 to be the earlier of the two insns. */ | |
1178 | if (i1 && INSN_CUID (i1) > INSN_CUID (i2)) | |
1179 | temp = i1, i1 = i2, i2 = temp; | |
1180 | ||
1181 | /* First check for one important special-case that the code below will | |
1182 | not handle. Namely, the case where I1 is zero, I2 has multiple sets, | |
1183 | and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case, | |
1184 | we may be able to replace that destination with the destination of I3. | |
1185 | This occurs in the common code where we compute both a quotient and | |
1186 | remainder into a structure, in which case we want to do the computation | |
1187 | directly into the structure to avoid register-register copies. | |
1188 | ||
1189 | We make very conservative checks below and only try to handle the | |
1190 | most common cases of this. For example, we only handle the case | |
1191 | where I2 and I3 are adjacent to avoid making difficult register | |
1192 | usage tests. */ | |
1193 | ||
1194 | if (i1 == 0 && GET_CODE (i3) == INSN && GET_CODE (PATTERN (i3)) == SET | |
1195 | && GET_CODE (SET_SRC (PATTERN (i3))) == REG | |
1196 | && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER | |
1197 | #ifdef SMALL_REGISTER_CLASSES | |
1198 | && (GET_CODE (SET_DEST (PATTERN (i3))) != REG | |
1199 | || REGNO (SET_DEST (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER) | |
1200 | #endif | |
1201 | && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3))) | |
1202 | && GET_CODE (PATTERN (i2)) == PARALLEL | |
1203 | && ! side_effects_p (SET_DEST (PATTERN (i3))) | |
1204 | /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code | |
1205 | below would need to check what is inside (and reg_overlap_mentioned_p | |
1206 | doesn't support those codes anyway). Don't allow those destinations; | |
1207 | the resulting insn isn't likely to be recognized anyway. */ | |
1208 | && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT | |
1209 | && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART | |
1210 | && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)), | |
1211 | SET_DEST (PATTERN (i3))) | |
1212 | && next_real_insn (i2) == i3) | |
1213 | { | |
1214 | rtx p2 = PATTERN (i2); | |
1215 | ||
1216 | /* Make sure that the destination of I3, | |
1217 | which we are going to substitute into one output of I2, | |
1218 | is not used within another output of I2. We must avoid making this: | |
1219 | (parallel [(set (mem (reg 69)) ...) | |
1220 | (set (reg 69) ...)]) | |
1221 | which is not well-defined as to order of actions. | |
1222 | (Besides, reload can't handle output reloads for this.) | |
1223 | ||
1224 | The problem can also happen if the dest of I3 is a memory ref, | |
1225 | if another dest in I2 is an indirect memory ref. */ | |
1226 | for (i = 0; i < XVECLEN (p2, 0); i++) | |
1227 | if (GET_CODE (XVECEXP (p2, 0, i)) == SET | |
1228 | && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)), | |
1229 | SET_DEST (XVECEXP (p2, 0, i)))) | |
1230 | break; | |
1231 | ||
1232 | if (i == XVECLEN (p2, 0)) | |
1233 | for (i = 0; i < XVECLEN (p2, 0); i++) | |
1234 | if (SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3))) | |
1235 | { | |
1236 | combine_merges++; | |
1237 | ||
1238 | subst_insn = i3; | |
1239 | subst_low_cuid = INSN_CUID (i2); | |
1240 | ||
1241 | added_sets_2 = 0; | |
1242 | i2dest = SET_SRC (PATTERN (i3)); | |
1243 | ||
1244 | /* Replace the dest in I2 with our dest and make the resulting | |
1245 | insn the new pattern for I3. Then skip to where we | |
1246 | validate the pattern. Everything was set up above. */ | |
1247 | SUBST (SET_DEST (XVECEXP (p2, 0, i)), | |
1248 | SET_DEST (PATTERN (i3))); | |
1249 | ||
1250 | newpat = p2; | |
1251 | goto validate_replacement; | |
1252 | } | |
1253 | } | |
1254 | ||
1255 | #ifndef HAVE_cc0 | |
1256 | /* If we have no I1 and I2 looks like: | |
1257 | (parallel [(set (reg:CC X) (compare:CC OP (const_int 0))) | |
1258 | (set Y OP)]) | |
1259 | make up a dummy I1 that is | |
1260 | (set Y OP) | |
1261 | and change I2 to be | |
1262 | (set (reg:CC X) (compare:CC Y (const_int 0))) | |
1263 | ||
1264 | (We can ignore any trailing CLOBBERs.) | |
1265 | ||
1266 | This undoes a previous combination and allows us to match a branch-and- | |
1267 | decrement insn. */ | |
1268 | ||
1269 | if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL | |
1270 | && XVECLEN (PATTERN (i2), 0) >= 2 | |
1271 | && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET | |
1272 | && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)))) | |
1273 | == MODE_CC) | |
1274 | && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE | |
1275 | && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx | |
1276 | && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET | |
1277 | && GET_CODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 1))) == REG | |
1278 | && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0), | |
1279 | SET_SRC (XVECEXP (PATTERN (i2), 0, 1)))) | |
1280 | { | |
1281 | for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--) | |
1282 | if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER) | |
1283 | break; | |
1284 | ||
1285 | if (i == 1) | |
1286 | { | |
1287 | /* We make I1 with the same INSN_UID as I2. This gives it | |
1288 | the same INSN_CUID for value tracking. Our fake I1 will | |
1289 | never appear in the insn stream so giving it the same INSN_UID | |
1290 | as I2 will not cause a problem. */ | |
1291 | ||
1292 | i1 = gen_rtx (INSN, VOIDmode, INSN_UID (i2), 0, i2, | |
1293 | XVECEXP (PATTERN (i2), 0, 1), -1, 0, 0); | |
1294 | ||
1295 | SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0)); | |
1296 | SUBST (XEXP (SET_SRC (PATTERN (i2)), 0), | |
1297 | SET_DEST (PATTERN (i1))); | |
1298 | } | |
1299 | } | |
1300 | #endif | |
1301 | ||
1302 | /* Verify that I2 and I1 are valid for combining. */ | |
1303 | if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src) | |
1304 | || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src))) | |
1305 | { | |
1306 | undo_all (); | |
1307 | return 0; | |
1308 | } | |
1309 | ||
1310 | /* Record whether I2DEST is used in I2SRC and similarly for the other | |
1311 | cases. Knowing this will help in register status updating below. */ | |
1312 | i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src); | |
1313 | i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src); | |
1314 | i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src); | |
1315 | ||
1316 | /* See if I1 directly feeds into I3. It does if I1DEST is not used | |
1317 | in I2SRC. */ | |
1318 | i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src); | |
1319 | ||
1320 | /* Ensure that I3's pattern can be the destination of combines. */ | |
1321 | if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest, | |
1322 | i1 && i2dest_in_i1src && i1_feeds_i3, | |
1323 | &i3dest_killed)) | |
1324 | { | |
1325 | undo_all (); | |
1326 | return 0; | |
1327 | } | |
1328 | ||
1329 | /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd. | |
1330 | We used to do this EXCEPT in one case: I3 has a post-inc in an | |
1331 | output operand. However, that exception can give rise to insns like | |
1332 | mov r3,(r3)+ | |
1333 | which is a famous insn on the PDP-11 where the value of r3 used as the | |
1334 | source was model-dependent. Avoid this sort of thing. */ | |
1335 | ||
1336 | #if 0 | |
1337 | if (!(GET_CODE (PATTERN (i3)) == SET | |
1338 | && GET_CODE (SET_SRC (PATTERN (i3))) == REG | |
1339 | && GET_CODE (SET_DEST (PATTERN (i3))) == MEM | |
1340 | && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC | |
1341 | || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC))) | |
1342 | /* It's not the exception. */ | |
1343 | #endif | |
1344 | #ifdef AUTO_INC_DEC | |
1345 | for (link = REG_NOTES (i3); link; link = XEXP (link, 1)) | |
1346 | if (REG_NOTE_KIND (link) == REG_INC | |
1347 | && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2)) | |
1348 | || (i1 != 0 | |
1349 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1))))) | |
1350 | { | |
1351 | undo_all (); | |
1352 | return 0; | |
1353 | } | |
1354 | #endif | |
1355 | ||
1356 | /* See if the SETs in I1 or I2 need to be kept around in the merged | |
1357 | instruction: whenever the value set there is still needed past I3. | |
1358 | For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3. | |
1359 | ||
1360 | For the SET in I1, we have two cases: If I1 and I2 independently | |
1361 | feed into I3, the set in I1 needs to be kept around if I1DEST dies | |
1362 | or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set | |
1363 | in I1 needs to be kept around unless I1DEST dies or is set in either | |
1364 | I2 or I3. We can distinguish these cases by seeing if I2SRC mentions | |
1365 | I1DEST. If so, we know I1 feeds into I2. */ | |
1366 | ||
1367 | added_sets_2 = ! dead_or_set_p (i3, i2dest); | |
1368 | ||
1369 | added_sets_1 | |
1370 | = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest) | |
1371 | : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest))); | |
1372 | ||
1373 | /* If the set in I2 needs to be kept around, we must make a copy of | |
1374 | PATTERN (I2), so that when we substitute I1SRC for I1DEST in | |
1375 | PATTERN (I2), we are only substituting for the original I1DEST, not into | |
1376 | an already-substituted copy. This also prevents making self-referential | |
1377 | rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to | |
1378 | I2DEST. */ | |
1379 | ||
1380 | i2pat = (GET_CODE (PATTERN (i2)) == PARALLEL | |
1381 | ? gen_rtx (SET, VOIDmode, i2dest, i2src) | |
1382 | : PATTERN (i2)); | |
1383 | ||
1384 | if (added_sets_2) | |
1385 | i2pat = copy_rtx (i2pat); | |
1386 | ||
1387 | combine_merges++; | |
1388 | ||
1389 | /* Substitute in the latest insn for the regs set by the earlier ones. */ | |
1390 | ||
1391 | maxreg = max_reg_num (); | |
1392 | ||
1393 | subst_insn = i3; | |
1394 | ||
1395 | /* It is possible that the source of I2 or I1 may be performing an | |
1396 | unneeded operation, such as a ZERO_EXTEND of something that is known | |
1397 | to have the high part zero. Handle that case by letting subst look at | |
1398 | the innermost one of them. | |
1399 | ||
1400 | Another way to do this would be to have a function that tries to | |
1401 | simplify a single insn instead of merging two or more insns. We don't | |
1402 | do this because of the potential of infinite loops and because | |
1403 | of the potential extra memory required. However, doing it the way | |
1404 | we are is a bit of a kludge and doesn't catch all cases. | |
1405 | ||
1406 | But only do this if -fexpensive-optimizations since it slows things down | |
1407 | and doesn't usually win. */ | |
1408 | ||
1409 | if (flag_expensive_optimizations) | |
1410 | { | |
1411 | /* Pass pc_rtx so no substitutions are done, just simplifications. | |
1412 | The cases that we are interested in here do not involve the few | |
1413 | cases were is_replaced is checked. */ | |
1414 | if (i1) | |
1415 | { | |
1416 | subst_low_cuid = INSN_CUID (i1); | |
1417 | i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0); | |
1418 | } | |
1419 | else | |
1420 | { | |
1421 | subst_low_cuid = INSN_CUID (i2); | |
1422 | i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0); | |
1423 | } | |
1424 | ||
1425 | previous_num_undos = undobuf.num_undo; | |
1426 | } | |
1427 | ||
1428 | #ifndef HAVE_cc0 | |
1429 | /* Many machines that don't use CC0 have insns that can both perform an | |
1430 | arithmetic operation and set the condition code. These operations will | |
1431 | be represented as a PARALLEL with the first element of the vector | |
1432 | being a COMPARE of an arithmetic operation with the constant zero. | |
1433 | The second element of the vector will set some pseudo to the result | |
1434 | of the same arithmetic operation. If we simplify the COMPARE, we won't | |
1435 | match such a pattern and so will generate an extra insn. Here we test | |
1436 | for this case, where both the comparison and the operation result are | |
1437 | needed, and make the PARALLEL by just replacing I2DEST in I3SRC with | |
1438 | I2SRC. Later we will make the PARALLEL that contains I2. */ | |
1439 | ||
1440 | if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET | |
1441 | && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE | |
1442 | && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx | |
1443 | && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest)) | |
1444 | { | |
1445 | rtx *cc_use; | |
1446 | enum machine_mode compare_mode; | |
1447 | ||
1448 | newpat = PATTERN (i3); | |
1449 | SUBST (XEXP (SET_SRC (newpat), 0), i2src); | |
1450 | ||
1451 | i2_is_used = 1; | |
1452 | ||
1453 | #ifdef EXTRA_CC_MODES | |
1454 | /* See if a COMPARE with the operand we substituted in should be done | |
1455 | with the mode that is currently being used. If not, do the same | |
1456 | processing we do in `subst' for a SET; namely, if the destination | |
1457 | is used only once, try to replace it with a register of the proper | |
1458 | mode and also replace the COMPARE. */ | |
1459 | if (undobuf.other_insn == 0 | |
1460 | && (cc_use = find_single_use (SET_DEST (newpat), i3, | |
1461 | &undobuf.other_insn)) | |
1462 | && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use), | |
1463 | i2src, const0_rtx)) | |
1464 | != GET_MODE (SET_DEST (newpat)))) | |
1465 | { | |
1466 | int regno = REGNO (SET_DEST (newpat)); | |
1467 | rtx new_dest = gen_rtx (REG, compare_mode, regno); | |
1468 | ||
1469 | if (regno < FIRST_PSEUDO_REGISTER | |
1470 | || (reg_n_sets[regno] == 1 && ! added_sets_2 | |
1471 | && ! REG_USERVAR_P (SET_DEST (newpat)))) | |
1472 | { | |
1473 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1474 | SUBST (regno_reg_rtx[regno], new_dest); | |
1475 | ||
1476 | SUBST (SET_DEST (newpat), new_dest); | |
1477 | SUBST (XEXP (*cc_use, 0), new_dest); | |
1478 | SUBST (SET_SRC (newpat), | |
1479 | gen_rtx_combine (COMPARE, compare_mode, | |
1480 | i2src, const0_rtx)); | |
1481 | } | |
1482 | else | |
1483 | undobuf.other_insn = 0; | |
1484 | } | |
1485 | #endif | |
1486 | } | |
1487 | else | |
1488 | #endif | |
1489 | { | |
1490 | n_occurrences = 0; /* `subst' counts here */ | |
1491 | ||
1492 | /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we | |
1493 | need to make a unique copy of I2SRC each time we substitute it | |
1494 | to avoid self-referential rtl. */ | |
1495 | ||
1496 | subst_low_cuid = INSN_CUID (i2); | |
1497 | newpat = subst (PATTERN (i3), i2dest, i2src, 0, | |
1498 | ! i1_feeds_i3 && i1dest_in_i1src); | |
1499 | previous_num_undos = undobuf.num_undo; | |
1500 | ||
1501 | /* Record whether i2's body now appears within i3's body. */ | |
1502 | i2_is_used = n_occurrences; | |
1503 | } | |
1504 | ||
1505 | /* If we already got a failure, don't try to do more. Otherwise, | |
1506 | try to substitute in I1 if we have it. */ | |
1507 | ||
1508 | if (i1 && GET_CODE (newpat) != CLOBBER) | |
1509 | { | |
1510 | /* Before we can do this substitution, we must redo the test done | |
1511 | above (see detailed comments there) that ensures that I1DEST | |
1512 | isn't mentioned in any SETs in NEWPAT that are field assignments. */ | |
1513 | ||
1514 | if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX, | |
1515 | 0, NULL_PTR)) | |
1516 | { | |
1517 | undo_all (); | |
1518 | return 0; | |
1519 | } | |
1520 | ||
1521 | n_occurrences = 0; | |
1522 | subst_low_cuid = INSN_CUID (i1); | |
1523 | newpat = subst (newpat, i1dest, i1src, 0, 0); | |
1524 | previous_num_undos = undobuf.num_undo; | |
1525 | } | |
1526 | ||
1527 | /* Fail if an autoincrement side-effect has been duplicated. Be careful | |
1528 | to count all the ways that I2SRC and I1SRC can be used. */ | |
1529 | if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0 | |
1530 | && i2_is_used + added_sets_2 > 1) | |
1531 | || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0 | |
1532 | && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3) | |
1533 | > 1)) | |
1534 | /* Fail if we tried to make a new register (we used to abort, but there's | |
1535 | really no reason to). */ | |
1536 | || max_reg_num () != maxreg | |
1537 | /* Fail if we couldn't do something and have a CLOBBER. */ | |
1538 | || GET_CODE (newpat) == CLOBBER) | |
1539 | { | |
1540 | undo_all (); | |
1541 | return 0; | |
1542 | } | |
1543 | ||
1544 | /* If the actions of the earlier insns must be kept | |
1545 | in addition to substituting them into the latest one, | |
1546 | we must make a new PARALLEL for the latest insn | |
1547 | to hold additional the SETs. */ | |
1548 | ||
1549 | if (added_sets_1 || added_sets_2) | |
1550 | { | |
1551 | combine_extras++; | |
1552 | ||
1553 | if (GET_CODE (newpat) == PARALLEL) | |
1554 | { | |
1555 | rtvec old = XVEC (newpat, 0); | |
1556 | total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2; | |
1557 | newpat = gen_rtx (PARALLEL, VOIDmode, rtvec_alloc (total_sets)); | |
1558 | bcopy (&old->elem[0], &XVECEXP (newpat, 0, 0), | |
1559 | sizeof (old->elem[0]) * old->num_elem); | |
1560 | } | |
1561 | else | |
1562 | { | |
1563 | rtx old = newpat; | |
1564 | total_sets = 1 + added_sets_1 + added_sets_2; | |
1565 | newpat = gen_rtx (PARALLEL, VOIDmode, rtvec_alloc (total_sets)); | |
1566 | XVECEXP (newpat, 0, 0) = old; | |
1567 | } | |
1568 | ||
1569 | if (added_sets_1) | |
1570 | XVECEXP (newpat, 0, --total_sets) | |
1571 | = (GET_CODE (PATTERN (i1)) == PARALLEL | |
1572 | ? gen_rtx (SET, VOIDmode, i1dest, i1src) : PATTERN (i1)); | |
1573 | ||
1574 | if (added_sets_2) | |
1575 | { | |
1576 | /* If there is no I1, use I2's body as is. We used to also not do | |
1577 | the subst call below if I2 was substituted into I3, | |
1578 | but that could lose a simplification. */ | |
1579 | if (i1 == 0) | |
1580 | XVECEXP (newpat, 0, --total_sets) = i2pat; | |
1581 | else | |
1582 | /* See comment where i2pat is assigned. */ | |
1583 | XVECEXP (newpat, 0, --total_sets) | |
1584 | = subst (i2pat, i1dest, i1src, 0, 0); | |
1585 | } | |
1586 | } | |
1587 | ||
1588 | /* We come here when we are replacing a destination in I2 with the | |
1589 | destination of I3. */ | |
1590 | validate_replacement: | |
1591 | ||
1592 | /* Is the result of combination a valid instruction? */ | |
1593 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); | |
1594 | ||
1595 | /* If the result isn't valid, see if it is a PARALLEL of two SETs where | |
1596 | the second SET's destination is a register that is unused. In that case, | |
1597 | we just need the first SET. This can occur when simplifying a divmod | |
1598 | insn. We *must* test for this case here because the code below that | |
1599 | splits two independent SETs doesn't handle this case correctly when it | |
1600 | updates the register status. Also check the case where the first | |
1601 | SET's destination is unused. That would not cause incorrect code, but | |
1602 | does cause an unneeded insn to remain. */ | |
1603 | ||
1604 | if (insn_code_number < 0 && GET_CODE (newpat) == PARALLEL | |
1605 | && XVECLEN (newpat, 0) == 2 | |
1606 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
1607 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
1608 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == REG | |
1609 | && find_reg_note (i3, REG_UNUSED, SET_DEST (XVECEXP (newpat, 0, 1))) | |
1610 | && ! side_effects_p (SET_SRC (XVECEXP (newpat, 0, 1))) | |
1611 | && asm_noperands (newpat) < 0) | |
1612 | { | |
1613 | newpat = XVECEXP (newpat, 0, 0); | |
1614 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); | |
1615 | } | |
1616 | ||
1617 | else if (insn_code_number < 0 && GET_CODE (newpat) == PARALLEL | |
1618 | && XVECLEN (newpat, 0) == 2 | |
1619 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
1620 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
1621 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) == REG | |
1622 | && find_reg_note (i3, REG_UNUSED, SET_DEST (XVECEXP (newpat, 0, 0))) | |
1623 | && ! side_effects_p (SET_SRC (XVECEXP (newpat, 0, 0))) | |
1624 | && asm_noperands (newpat) < 0) | |
1625 | { | |
1626 | newpat = XVECEXP (newpat, 0, 1); | |
1627 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); | |
1628 | } | |
1629 | ||
1630 | /* See if this is an XOR. If so, perhaps the problem is that the | |
1631 | constant is out of range. Replace it with a complemented XOR with | |
1632 | a complemented constant; it might be in range. */ | |
1633 | ||
1634 | else if (insn_code_number < 0 && GET_CODE (newpat) == SET | |
1635 | && GET_CODE (SET_SRC (newpat)) == XOR | |
1636 | && GET_CODE (XEXP (SET_SRC (newpat), 1)) == CONST_INT | |
1637 | && ((temp = simplify_unary_operation (NOT, | |
1638 | GET_MODE (SET_SRC (newpat)), | |
1639 | XEXP (SET_SRC (newpat), 1), | |
1640 | GET_MODE (SET_SRC (newpat)))) | |
1641 | != 0)) | |
1642 | { | |
1643 | enum machine_mode i_mode = GET_MODE (SET_SRC (newpat)); | |
1644 | rtx pat | |
1645 | = gen_rtx_combine (SET, VOIDmode, SET_DEST (newpat), | |
1646 | gen_unary (NOT, i_mode, | |
1647 | gen_binary (XOR, i_mode, | |
1648 | XEXP (SET_SRC (newpat), 0), | |
1649 | temp))); | |
1650 | ||
1651 | insn_code_number = recog_for_combine (&pat, i3, &new_i3_notes); | |
1652 | if (insn_code_number >= 0) | |
1653 | newpat = pat; | |
1654 | } | |
1655 | ||
1656 | /* If we were combining three insns and the result is a simple SET | |
1657 | with no ASM_OPERANDS that wasn't recognized, try to split it into two | |
1658 | insns. There are two ways to do this. It can be split using a | |
1659 | machine-specific method (like when you have an addition of a large | |
1660 | constant) or by combine in the function find_split_point. */ | |
1661 | ||
1662 | if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET | |
1663 | && asm_noperands (newpat) < 0) | |
1664 | { | |
1665 | rtx m_split, *split; | |
1666 | rtx ni2dest = i2dest; | |
1667 | ||
1668 | /* See if the MD file can split NEWPAT. If it can't, see if letting it | |
1669 | use I2DEST as a scratch register will help. In the latter case, | |
1670 | convert I2DEST to the mode of the source of NEWPAT if we can. */ | |
1671 | ||
1672 | m_split = split_insns (newpat, i3); | |
1673 | ||
1674 | /* We can only use I2DEST as a scratch reg if it doesn't overlap any | |
1675 | inputs of NEWPAT. */ | |
1676 | ||
1677 | /* ??? If I2DEST is not safe, and I1DEST exists, then it would be | |
1678 | possible to try that as a scratch reg. This would require adding | |
1679 | more code to make it work though. */ | |
1680 | ||
1681 | if (m_split == 0 && ! reg_overlap_mentioned_p (ni2dest, newpat)) | |
1682 | { | |
1683 | /* If I2DEST is a hard register or the only use of a pseudo, | |
1684 | we can change its mode. */ | |
1685 | if (GET_MODE (SET_DEST (newpat)) != GET_MODE (i2dest) | |
1686 | && GET_MODE (SET_DEST (newpat)) != VOIDmode | |
1687 | && GET_CODE (i2dest) == REG | |
1688 | && (REGNO (i2dest) < FIRST_PSEUDO_REGISTER | |
1689 | || (reg_n_sets[REGNO (i2dest)] == 1 && ! added_sets_2 | |
1690 | && ! REG_USERVAR_P (i2dest)))) | |
1691 | ni2dest = gen_rtx (REG, GET_MODE (SET_DEST (newpat)), | |
1692 | REGNO (i2dest)); | |
1693 | ||
1694 | m_split = split_insns (gen_rtx (PARALLEL, VOIDmode, | |
1695 | gen_rtvec (2, newpat, | |
1696 | gen_rtx (CLOBBER, | |
1697 | VOIDmode, | |
1698 | ni2dest))), | |
1699 | i3); | |
1700 | } | |
1701 | ||
1702 | if (m_split && GET_CODE (m_split) == SEQUENCE | |
1703 | && XVECLEN (m_split, 0) == 2 | |
1704 | && (next_real_insn (i2) == i3 | |
1705 | || ! use_crosses_set_p (PATTERN (XVECEXP (m_split, 0, 0)), | |
1706 | INSN_CUID (i2)))) | |
1707 | { | |
1708 | rtx i2set, i3set; | |
1709 | rtx newi3pat = PATTERN (XVECEXP (m_split, 0, 1)); | |
1710 | newi2pat = PATTERN (XVECEXP (m_split, 0, 0)); | |
1711 | ||
1712 | i3set = single_set (XVECEXP (m_split, 0, 1)); | |
1713 | i2set = single_set (XVECEXP (m_split, 0, 0)); | |
1714 | ||
1715 | /* In case we changed the mode of I2DEST, replace it in the | |
1716 | pseudo-register table here. We can't do it above in case this | |
1717 | code doesn't get executed and we do a split the other way. */ | |
1718 | ||
1719 | if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER) | |
1720 | SUBST (regno_reg_rtx[REGNO (i2dest)], ni2dest); | |
1721 | ||
1722 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); | |
1723 | ||
1724 | /* If I2 or I3 has multiple SETs, we won't know how to track | |
1725 | register status, so don't use these insns. */ | |
1726 | ||
1727 | if (i2_code_number >= 0 && i2set && i3set) | |
1728 | insn_code_number = recog_for_combine (&newi3pat, i3, | |
1729 | &new_i3_notes); | |
1730 | ||
1731 | if (insn_code_number >= 0) | |
1732 | newpat = newi3pat; | |
1733 | ||
1734 | /* It is possible that both insns now set the destination of I3. | |
1735 | If so, we must show an extra use of it. */ | |
1736 | ||
1737 | if (insn_code_number >= 0 && GET_CODE (SET_DEST (i3set)) == REG | |
1738 | && GET_CODE (SET_DEST (i2set)) == REG | |
1739 | && REGNO (SET_DEST (i3set)) == REGNO (SET_DEST (i2set))) | |
1740 | reg_n_sets[REGNO (SET_DEST (i2set))]++; | |
1741 | } | |
1742 | ||
1743 | /* If we can split it and use I2DEST, go ahead and see if that | |
1744 | helps things be recognized. Verify that none of the registers | |
1745 | are set between I2 and I3. */ | |
1746 | if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0 | |
1747 | #ifdef HAVE_cc0 | |
1748 | && GET_CODE (i2dest) == REG | |
1749 | #endif | |
1750 | /* We need I2DEST in the proper mode. If it is a hard register | |
1751 | or the only use of a pseudo, we can change its mode. */ | |
1752 | && (GET_MODE (*split) == GET_MODE (i2dest) | |
1753 | || GET_MODE (*split) == VOIDmode | |
1754 | || REGNO (i2dest) < FIRST_PSEUDO_REGISTER | |
1755 | || (reg_n_sets[REGNO (i2dest)] == 1 && ! added_sets_2 | |
1756 | && ! REG_USERVAR_P (i2dest))) | |
1757 | && (next_real_insn (i2) == i3 | |
1758 | || ! use_crosses_set_p (*split, INSN_CUID (i2))) | |
1759 | /* We can't overwrite I2DEST if its value is still used by | |
1760 | NEWPAT. */ | |
1761 | && ! reg_referenced_p (i2dest, newpat)) | |
1762 | { | |
1763 | rtx newdest = i2dest; | |
1764 | ||
1765 | /* Get NEWDEST as a register in the proper mode. We have already | |
1766 | validated that we can do this. */ | |
1767 | if (GET_MODE (i2dest) != GET_MODE (*split) | |
1768 | && GET_MODE (*split) != VOIDmode) | |
1769 | { | |
1770 | newdest = gen_rtx (REG, GET_MODE (*split), REGNO (i2dest)); | |
1771 | ||
1772 | if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER) | |
1773 | SUBST (regno_reg_rtx[REGNO (i2dest)], newdest); | |
1774 | } | |
1775 | ||
1776 | /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to | |
1777 | an ASHIFT. This can occur if it was inside a PLUS and hence | |
1778 | appeared to be a memory address. This is a kludge. */ | |
1779 | if (GET_CODE (*split) == MULT | |
1780 | && GET_CODE (XEXP (*split, 1)) == CONST_INT | |
1781 | && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0) | |
1782 | SUBST (*split, gen_rtx_combine (ASHIFT, GET_MODE (*split), | |
1783 | XEXP (*split, 0), GEN_INT (i))); | |
1784 | ||
1785 | #ifdef INSN_SCHEDULING | |
1786 | /* If *SPLIT is a paradoxical SUBREG, when we split it, it should | |
1787 | be written as a ZERO_EXTEND. */ | |
1788 | if (GET_CODE (*split) == SUBREG | |
1789 | && GET_CODE (SUBREG_REG (*split)) == MEM) | |
1790 | SUBST (*split, gen_rtx_combine (ZERO_EXTEND, GET_MODE (*split), | |
1791 | XEXP (*split, 0))); | |
1792 | #endif | |
1793 | ||
1794 | newi2pat = gen_rtx_combine (SET, VOIDmode, newdest, *split); | |
1795 | SUBST (*split, newdest); | |
1796 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); | |
1797 | if (i2_code_number >= 0) | |
1798 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); | |
1799 | } | |
1800 | } | |
1801 | ||
1802 | /* Check for a case where we loaded from memory in a narrow mode and | |
1803 | then sign extended it, but we need both registers. In that case, | |
1804 | we have a PARALLEL with both loads from the same memory location. | |
1805 | We can split this into a load from memory followed by a register-register | |
1806 | copy. This saves at least one insn, more if register allocation can | |
1807 | eliminate the copy. */ | |
1808 | ||
1809 | else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0 | |
1810 | && GET_CODE (newpat) == PARALLEL | |
1811 | && XVECLEN (newpat, 0) == 2 | |
1812 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
1813 | && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND | |
1814 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
1815 | && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)), | |
1816 | XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0)) | |
1817 | && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)), | |
1818 | INSN_CUID (i2)) | |
1819 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT | |
1820 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART | |
1821 | && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)), | |
1822 | SET_SRC (XVECEXP (newpat, 0, 1))) | |
1823 | && ! find_reg_note (i3, REG_UNUSED, | |
1824 | SET_DEST (XVECEXP (newpat, 0, 0)))) | |
1825 | { | |
1826 | rtx ni2dest; | |
1827 | ||
1828 | newi2pat = XVECEXP (newpat, 0, 0); | |
1829 | ni2dest = SET_DEST (XVECEXP (newpat, 0, 0)); | |
1830 | newpat = XVECEXP (newpat, 0, 1); | |
1831 | SUBST (SET_SRC (newpat), | |
1832 | gen_lowpart_for_combine (GET_MODE (SET_SRC (newpat)), ni2dest)); | |
1833 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); | |
1834 | if (i2_code_number >= 0) | |
1835 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); | |
1836 | ||
1837 | if (insn_code_number >= 0) | |
1838 | { | |
1839 | rtx insn; | |
1840 | rtx link; | |
1841 | ||
1842 | /* If we will be able to accept this, we have made a change to the | |
1843 | destination of I3. This can invalidate a LOG_LINKS pointing | |
1844 | to I3. No other part of combine.c makes such a transformation. | |
1845 | ||
1846 | The new I3 will have a destination that was previously the | |
1847 | destination of I1 or I2 and which was used in i2 or I3. Call | |
1848 | distribute_links to make a LOG_LINK from the next use of | |
1849 | that destination. */ | |
1850 | ||
1851 | PATTERN (i3) = newpat; | |
1852 | distribute_links (gen_rtx (INSN_LIST, VOIDmode, i3, NULL_RTX)); | |
1853 | ||
1854 | /* I3 now uses what used to be its destination and which is | |
1855 | now I2's destination. That means we need a LOG_LINK from | |
1856 | I3 to I2. But we used to have one, so we still will. | |
1857 | ||
1858 | However, some later insn might be using I2's dest and have | |
1859 | a LOG_LINK pointing at I3. We must remove this link. | |
1860 | The simplest way to remove the link is to point it at I1, | |
1861 | which we know will be a NOTE. */ | |
1862 | ||
1863 | for (insn = NEXT_INSN (i3); | |
1864 | insn && GET_CODE (insn) != CODE_LABEL | |
1865 | && GET_CODE (PREV_INSN (insn)) != JUMP_INSN; | |
1866 | insn = NEXT_INSN (insn)) | |
1867 | { | |
1868 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
1869 | && reg_referenced_p (ni2dest, PATTERN (insn))) | |
1870 | { | |
1871 | for (link = LOG_LINKS (insn); link; | |
1872 | link = XEXP (link, 1)) | |
1873 | if (XEXP (link, 0) == i3) | |
1874 | XEXP (link, 0) = i1; | |
1875 | ||
1876 | break; | |
1877 | } | |
1878 | } | |
1879 | } | |
1880 | } | |
1881 | ||
1882 | /* Similarly, check for a case where we have a PARALLEL of two independent | |
1883 | SETs but we started with three insns. In this case, we can do the sets | |
1884 | as two separate insns. This case occurs when some SET allows two | |
1885 | other insns to combine, but the destination of that SET is still live. */ | |
1886 | ||
1887 | else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0 | |
1888 | && GET_CODE (newpat) == PARALLEL | |
1889 | && XVECLEN (newpat, 0) == 2 | |
1890 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
1891 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT | |
1892 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART | |
1893 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
1894 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT | |
1895 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART | |
1896 | && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)), | |
1897 | INSN_CUID (i2)) | |
1898 | /* Don't pass sets with (USE (MEM ...)) dests to the following. */ | |
1899 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != USE | |
1900 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != USE | |
1901 | && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)), | |
1902 | XVECEXP (newpat, 0, 0)) | |
1903 | && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)), | |
1904 | XVECEXP (newpat, 0, 1))) | |
1905 | { | |
1906 | newi2pat = XVECEXP (newpat, 0, 1); | |
1907 | newpat = XVECEXP (newpat, 0, 0); | |
1908 | ||
1909 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); | |
1910 | if (i2_code_number >= 0) | |
1911 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); | |
1912 | } | |
1913 | ||
1914 | /* If it still isn't recognized, fail and change things back the way they | |
1915 | were. */ | |
1916 | if ((insn_code_number < 0 | |
1917 | /* Is the result a reasonable ASM_OPERANDS? */ | |
1918 | && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2))) | |
1919 | { | |
1920 | undo_all (); | |
1921 | return 0; | |
1922 | } | |
1923 | ||
1924 | /* If we had to change another insn, make sure it is valid also. */ | |
1925 | if (undobuf.other_insn) | |
1926 | { | |
1927 | rtx other_notes = REG_NOTES (undobuf.other_insn); | |
1928 | rtx other_pat = PATTERN (undobuf.other_insn); | |
1929 | rtx new_other_notes; | |
1930 | rtx note, next; | |
1931 | ||
1932 | other_code_number = recog_for_combine (&other_pat, undobuf.other_insn, | |
1933 | &new_other_notes); | |
1934 | ||
1935 | if (other_code_number < 0 && ! check_asm_operands (other_pat)) | |
1936 | { | |
1937 | undo_all (); | |
1938 | return 0; | |
1939 | } | |
1940 | ||
1941 | PATTERN (undobuf.other_insn) = other_pat; | |
1942 | ||
1943 | /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they | |
1944 | are still valid. Then add any non-duplicate notes added by | |
1945 | recog_for_combine. */ | |
1946 | for (note = REG_NOTES (undobuf.other_insn); note; note = next) | |
1947 | { | |
1948 | next = XEXP (note, 1); | |
1949 | ||
1950 | if (REG_NOTE_KIND (note) == REG_UNUSED | |
1951 | && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn))) | |
1952 | { | |
1953 | if (GET_CODE (XEXP (note, 0)) == REG) | |
1954 | reg_n_deaths[REGNO (XEXP (note, 0))]--; | |
1955 | ||
1956 | remove_note (undobuf.other_insn, note); | |
1957 | } | |
1958 | } | |
1959 | ||
1960 | for (note = new_other_notes; note; note = XEXP (note, 1)) | |
1961 | if (GET_CODE (XEXP (note, 0)) == REG) | |
1962 | reg_n_deaths[REGNO (XEXP (note, 0))]++; | |
1963 | ||
1964 | distribute_notes (new_other_notes, undobuf.other_insn, | |
1965 | undobuf.other_insn, NULL_RTX, NULL_RTX, NULL_RTX); | |
1966 | } | |
1967 | ||
1968 | /* We now know that we can do this combination. Merge the insns and | |
1969 | update the status of registers and LOG_LINKS. */ | |
1970 | ||
1971 | { | |
1972 | rtx i3notes, i2notes, i1notes = 0; | |
1973 | rtx i3links, i2links, i1links = 0; | |
1974 | rtx midnotes = 0; | |
1975 | int all_adjacent = (next_real_insn (i2) == i3 | |
1976 | && (i1 == 0 || next_real_insn (i1) == i2)); | |
1977 | register int regno; | |
1978 | /* Compute which registers we expect to eliminate. */ | |
1979 | rtx elim_i2 = (newi2pat || i2dest_in_i2src || i2dest_in_i1src | |
1980 | ? 0 : i2dest); | |
1981 | rtx elim_i1 = i1 == 0 || i1dest_in_i1src ? 0 : i1dest; | |
1982 | ||
1983 | /* Get the old REG_NOTES and LOG_LINKS from all our insns and | |
1984 | clear them. */ | |
1985 | i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3); | |
1986 | i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2); | |
1987 | if (i1) | |
1988 | i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1); | |
1989 | ||
1990 | /* Ensure that we do not have something that should not be shared but | |
1991 | occurs multiple times in the new insns. Check this by first | |
1992 | resetting all the `used' flags and then copying anything is shared. */ | |
1993 | ||
1994 | reset_used_flags (i3notes); | |
1995 | reset_used_flags (i2notes); | |
1996 | reset_used_flags (i1notes); | |
1997 | reset_used_flags (newpat); | |
1998 | reset_used_flags (newi2pat); | |
1999 | if (undobuf.other_insn) | |
2000 | reset_used_flags (PATTERN (undobuf.other_insn)); | |
2001 | ||
2002 | i3notes = copy_rtx_if_shared (i3notes); | |
2003 | i2notes = copy_rtx_if_shared (i2notes); | |
2004 | i1notes = copy_rtx_if_shared (i1notes); | |
2005 | newpat = copy_rtx_if_shared (newpat); | |
2006 | newi2pat = copy_rtx_if_shared (newi2pat); | |
2007 | if (undobuf.other_insn) | |
2008 | reset_used_flags (PATTERN (undobuf.other_insn)); | |
2009 | ||
2010 | INSN_CODE (i3) = insn_code_number; | |
2011 | PATTERN (i3) = newpat; | |
2012 | if (undobuf.other_insn) | |
2013 | INSN_CODE (undobuf.other_insn) = other_code_number; | |
2014 | ||
2015 | /* We had one special case above where I2 had more than one set and | |
2016 | we replaced a destination of one of those sets with the destination | |
2017 | of I3. In that case, we have to update LOG_LINKS of insns later | |
2018 | in this basic block. Note that this (expensive) case is rare. */ | |
2019 | ||
2020 | if (GET_CODE (PATTERN (i2)) == PARALLEL) | |
2021 | for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++) | |
2022 | if (GET_CODE (SET_DEST (XVECEXP (PATTERN (i2), 0, i))) == REG | |
2023 | && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest | |
2024 | && ! find_reg_note (i2, REG_UNUSED, | |
2025 | SET_DEST (XVECEXP (PATTERN (i2), 0, i)))) | |
2026 | { | |
2027 | register rtx insn; | |
2028 | ||
2029 | for (insn = NEXT_INSN (i2); insn; insn = NEXT_INSN (insn)) | |
2030 | { | |
2031 | if (insn != i3 && GET_RTX_CLASS (GET_CODE (insn)) == 'i') | |
2032 | for (link = LOG_LINKS (insn); link; link = XEXP (link, 1)) | |
2033 | if (XEXP (link, 0) == i2) | |
2034 | XEXP (link, 0) = i3; | |
2035 | ||
2036 | if (GET_CODE (insn) == CODE_LABEL | |
2037 | || GET_CODE (insn) == JUMP_INSN) | |
2038 | break; | |
2039 | } | |
2040 | } | |
2041 | ||
2042 | LOG_LINKS (i3) = 0; | |
2043 | REG_NOTES (i3) = 0; | |
2044 | LOG_LINKS (i2) = 0; | |
2045 | REG_NOTES (i2) = 0; | |
2046 | ||
2047 | if (newi2pat) | |
2048 | { | |
2049 | INSN_CODE (i2) = i2_code_number; | |
2050 | PATTERN (i2) = newi2pat; | |
2051 | } | |
2052 | else | |
2053 | { | |
2054 | PUT_CODE (i2, NOTE); | |
2055 | NOTE_LINE_NUMBER (i2) = NOTE_INSN_DELETED; | |
2056 | NOTE_SOURCE_FILE (i2) = 0; | |
2057 | } | |
2058 | ||
2059 | if (i1) | |
2060 | { | |
2061 | LOG_LINKS (i1) = 0; | |
2062 | REG_NOTES (i1) = 0; | |
2063 | PUT_CODE (i1, NOTE); | |
2064 | NOTE_LINE_NUMBER (i1) = NOTE_INSN_DELETED; | |
2065 | NOTE_SOURCE_FILE (i1) = 0; | |
2066 | } | |
2067 | ||
2068 | /* Get death notes for everything that is now used in either I3 or | |
2069 | I2 and used to die in a previous insn. */ | |
2070 | ||
2071 | move_deaths (newpat, i1 ? INSN_CUID (i1) : INSN_CUID (i2), i3, &midnotes); | |
2072 | if (newi2pat) | |
2073 | move_deaths (newi2pat, INSN_CUID (i1), i2, &midnotes); | |
2074 | ||
2075 | /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */ | |
2076 | if (i3notes) | |
2077 | distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX, | |
2078 | elim_i2, elim_i1); | |
2079 | if (i2notes) | |
2080 | distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX, | |
2081 | elim_i2, elim_i1); | |
2082 | if (i1notes) | |
2083 | distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX, | |
2084 | elim_i2, elim_i1); | |
2085 | if (midnotes) | |
2086 | distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, | |
2087 | elim_i2, elim_i1); | |
2088 | ||
2089 | /* Distribute any notes added to I2 or I3 by recog_for_combine. We | |
2090 | know these are REG_UNUSED and want them to go to the desired insn, | |
2091 | so we always pass it as i3. We have not counted the notes in | |
2092 | reg_n_deaths yet, so we need to do so now. */ | |
2093 | ||
2094 | if (newi2pat && new_i2_notes) | |
2095 | { | |
2096 | for (temp = new_i2_notes; temp; temp = XEXP (temp, 1)) | |
2097 | if (GET_CODE (XEXP (temp, 0)) == REG) | |
2098 | reg_n_deaths[REGNO (XEXP (temp, 0))]++; | |
2099 | ||
2100 | distribute_notes (new_i2_notes, i2, i2, NULL_RTX, NULL_RTX, NULL_RTX); | |
2101 | } | |
2102 | ||
2103 | if (new_i3_notes) | |
2104 | { | |
2105 | for (temp = new_i3_notes; temp; temp = XEXP (temp, 1)) | |
2106 | if (GET_CODE (XEXP (temp, 0)) == REG) | |
2107 | reg_n_deaths[REGNO (XEXP (temp, 0))]++; | |
2108 | ||
2109 | distribute_notes (new_i3_notes, i3, i3, NULL_RTX, NULL_RTX, NULL_RTX); | |
2110 | } | |
2111 | ||
2112 | /* If I3DEST was used in I3SRC, it really died in I3. We may need to | |
2113 | put a REG_DEAD note for it somewhere. Similarly for I2 and I1. | |
2114 | Show an additional death due to the REG_DEAD note we make here. If | |
2115 | we discard it in distribute_notes, we will decrement it again. */ | |
2116 | ||
2117 | if (i3dest_killed) | |
2118 | { | |
2119 | if (GET_CODE (i3dest_killed) == REG) | |
2120 | reg_n_deaths[REGNO (i3dest_killed)]++; | |
2121 | ||
2122 | distribute_notes (gen_rtx (EXPR_LIST, REG_DEAD, i3dest_killed, | |
2123 | NULL_RTX), | |
2124 | NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, | |
2125 | NULL_RTX, NULL_RTX); | |
2126 | } | |
2127 | ||
2128 | /* For I2 and I1, we have to be careful. If NEWI2PAT exists and sets | |
2129 | I2DEST or I1DEST, the death must be somewhere before I2, not I3. If | |
2130 | we passed I3 in that case, it might delete I2. */ | |
2131 | ||
2132 | if (i2dest_in_i2src) | |
2133 | { | |
2134 | if (GET_CODE (i2dest) == REG) | |
2135 | reg_n_deaths[REGNO (i2dest)]++; | |
2136 | ||
2137 | if (newi2pat && reg_set_p (i2dest, newi2pat)) | |
2138 | distribute_notes (gen_rtx (EXPR_LIST, REG_DEAD, i2dest, NULL_RTX), | |
2139 | NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX); | |
2140 | else | |
2141 | distribute_notes (gen_rtx (EXPR_LIST, REG_DEAD, i2dest, NULL_RTX), | |
2142 | NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, | |
2143 | NULL_RTX, NULL_RTX); | |
2144 | } | |
2145 | ||
2146 | if (i1dest_in_i1src) | |
2147 | { | |
2148 | if (GET_CODE (i1dest) == REG) | |
2149 | reg_n_deaths[REGNO (i1dest)]++; | |
2150 | ||
2151 | if (newi2pat && reg_set_p (i1dest, newi2pat)) | |
2152 | distribute_notes (gen_rtx (EXPR_LIST, REG_DEAD, i1dest, NULL_RTX), | |
2153 | NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX); | |
2154 | else | |
2155 | distribute_notes (gen_rtx (EXPR_LIST, REG_DEAD, i1dest, NULL_RTX), | |
2156 | NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, | |
2157 | NULL_RTX, NULL_RTX); | |
2158 | } | |
2159 | ||
2160 | distribute_links (i3links); | |
2161 | distribute_links (i2links); | |
2162 | distribute_links (i1links); | |
2163 | ||
2164 | if (GET_CODE (i2dest) == REG) | |
2165 | { | |
2166 | rtx link; | |
2167 | rtx i2_insn = 0, i2_val = 0, set; | |
2168 | ||
2169 | /* The insn that used to set this register doesn't exist, and | |
2170 | this life of the register may not exist either. See if one of | |
2171 | I3's links points to an insn that sets I2DEST. If it does, | |
2172 | that is now the last known value for I2DEST. If we don't update | |
2173 | this and I2 set the register to a value that depended on its old | |
2174 | contents, we will get confused. If this insn is used, thing | |
2175 | will be set correctly in combine_instructions. */ | |
2176 | ||
2177 | for (link = LOG_LINKS (i3); link; link = XEXP (link, 1)) | |
2178 | if ((set = single_set (XEXP (link, 0))) != 0 | |
2179 | && rtx_equal_p (i2dest, SET_DEST (set))) | |
2180 | i2_insn = XEXP (link, 0), i2_val = SET_SRC (set); | |
2181 | ||
2182 | record_value_for_reg (i2dest, i2_insn, i2_val); | |
2183 | ||
2184 | /* If the reg formerly set in I2 died only once and that was in I3, | |
2185 | zero its use count so it won't make `reload' do any work. */ | |
2186 | if (! added_sets_2 && newi2pat == 0) | |
2187 | { | |
2188 | regno = REGNO (i2dest); | |
2189 | reg_n_sets[regno]--; | |
2190 | if (reg_n_sets[regno] == 0 | |
2191 | && ! (basic_block_live_at_start[0][regno / REGSET_ELT_BITS] | |
2192 | & ((REGSET_ELT_TYPE) 1 << (regno % REGSET_ELT_BITS)))) | |
2193 | reg_n_refs[regno] = 0; | |
2194 | } | |
2195 | } | |
2196 | ||
2197 | if (i1 && GET_CODE (i1dest) == REG) | |
2198 | { | |
2199 | rtx link; | |
2200 | rtx i1_insn = 0, i1_val = 0, set; | |
2201 | ||
2202 | for (link = LOG_LINKS (i3); link; link = XEXP (link, 1)) | |
2203 | if ((set = single_set (XEXP (link, 0))) != 0 | |
2204 | && rtx_equal_p (i1dest, SET_DEST (set))) | |
2205 | i1_insn = XEXP (link, 0), i1_val = SET_SRC (set); | |
2206 | ||
2207 | record_value_for_reg (i1dest, i1_insn, i1_val); | |
2208 | ||
2209 | regno = REGNO (i1dest); | |
2210 | if (! added_sets_1) | |
2211 | { | |
2212 | reg_n_sets[regno]--; | |
2213 | if (reg_n_sets[regno] == 0 | |
2214 | && ! (basic_block_live_at_start[0][regno / REGSET_ELT_BITS] | |
2215 | & ((REGSET_ELT_TYPE) 1 << (regno % REGSET_ELT_BITS)))) | |
2216 | reg_n_refs[regno] = 0; | |
2217 | } | |
2218 | } | |
2219 | ||
2220 | /* Update reg_nonzero_bits et al for any changes that may have been made | |
2221 | to this insn. */ | |
2222 | ||
2223 | note_stores (newpat, set_nonzero_bits_and_sign_copies); | |
2224 | if (newi2pat) | |
2225 | note_stores (newi2pat, set_nonzero_bits_and_sign_copies); | |
2226 | ||
2227 | /* If I3 is now an unconditional jump, ensure that it has a | |
2228 | BARRIER following it since it may have initially been a | |
2229 | conditional jump. It may also be the last nonnote insn. */ | |
2230 | ||
2231 | if ((GET_CODE (newpat) == RETURN || simplejump_p (i3)) | |
2232 | && ((temp = next_nonnote_insn (i3)) == NULL_RTX | |
2233 | || GET_CODE (temp) != BARRIER)) | |
2234 | emit_barrier_after (i3); | |
2235 | } | |
2236 | ||
2237 | combine_successes++; | |
2238 | ||
2239 | return newi2pat ? i2 : i3; | |
2240 | } | |
2241 | \f | |
2242 | /* Undo all the modifications recorded in undobuf. */ | |
2243 | ||
2244 | static void | |
2245 | undo_all () | |
2246 | { | |
2247 | register int i; | |
2248 | if (undobuf.num_undo > MAX_UNDO) | |
2249 | undobuf.num_undo = MAX_UNDO; | |
2250 | for (i = undobuf.num_undo - 1; i >= 0; i--) | |
2251 | { | |
2252 | if (undobuf.undo[i].is_int) | |
2253 | *undobuf.undo[i].where.i = undobuf.undo[i].old_contents.i; | |
2254 | else | |
2255 | *undobuf.undo[i].where.rtx = undobuf.undo[i].old_contents.rtx; | |
2256 | ||
2257 | } | |
2258 | ||
2259 | obfree (undobuf.storage); | |
2260 | undobuf.num_undo = 0; | |
2261 | } | |
2262 | \f | |
2263 | /* Find the innermost point within the rtx at LOC, possibly LOC itself, | |
2264 | where we have an arithmetic expression and return that point. LOC will | |
2265 | be inside INSN. | |
2266 | ||
2267 | try_combine will call this function to see if an insn can be split into | |
2268 | two insns. */ | |
2269 | ||
2270 | static rtx * | |
2271 | find_split_point (loc, insn) | |
2272 | rtx *loc; | |
2273 | rtx insn; | |
2274 | { | |
2275 | rtx x = *loc; | |
2276 | enum rtx_code code = GET_CODE (x); | |
2277 | rtx *split; | |
2278 | int len = 0, pos, unsignedp; | |
2279 | rtx inner; | |
2280 | ||
2281 | /* First special-case some codes. */ | |
2282 | switch (code) | |
2283 | { | |
2284 | case SUBREG: | |
2285 | #ifdef INSN_SCHEDULING | |
2286 | /* If we are making a paradoxical SUBREG invalid, it becomes a split | |
2287 | point. */ | |
2288 | if (GET_CODE (SUBREG_REG (x)) == MEM) | |
2289 | return loc; | |
2290 | #endif | |
2291 | return find_split_point (&SUBREG_REG (x), insn); | |
2292 | ||
2293 | case MEM: | |
2294 | #ifdef HAVE_lo_sum | |
2295 | /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it | |
2296 | using LO_SUM and HIGH. */ | |
2297 | if (GET_CODE (XEXP (x, 0)) == CONST | |
2298 | || GET_CODE (XEXP (x, 0)) == SYMBOL_REF) | |
2299 | { | |
2300 | SUBST (XEXP (x, 0), | |
2301 | gen_rtx_combine (LO_SUM, Pmode, | |
2302 | gen_rtx_combine (HIGH, Pmode, XEXP (x, 0)), | |
2303 | XEXP (x, 0))); | |
2304 | return &XEXP (XEXP (x, 0), 0); | |
2305 | } | |
2306 | #endif | |
2307 | ||
2308 | /* If we have a PLUS whose second operand is a constant and the | |
2309 | address is not valid, perhaps will can split it up using | |
2310 | the machine-specific way to split large constants. We use | |
2311 | the first psuedo-reg (one of the virtual regs) as a placeholder; | |
2312 | it will not remain in the result. */ | |
2313 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
2314 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
2315 | && ! memory_address_p (GET_MODE (x), XEXP (x, 0))) | |
2316 | { | |
2317 | rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER]; | |
2318 | rtx seq = split_insns (gen_rtx (SET, VOIDmode, reg, XEXP (x, 0)), | |
2319 | subst_insn); | |
2320 | ||
2321 | /* This should have produced two insns, each of which sets our | |
2322 | placeholder. If the source of the second is a valid address, | |
2323 | we can make put both sources together and make a split point | |
2324 | in the middle. */ | |
2325 | ||
2326 | if (seq && XVECLEN (seq, 0) == 2 | |
2327 | && GET_CODE (XVECEXP (seq, 0, 0)) == INSN | |
2328 | && GET_CODE (PATTERN (XVECEXP (seq, 0, 0))) == SET | |
2329 | && SET_DEST (PATTERN (XVECEXP (seq, 0, 0))) == reg | |
2330 | && ! reg_mentioned_p (reg, | |
2331 | SET_SRC (PATTERN (XVECEXP (seq, 0, 0)))) | |
2332 | && GET_CODE (XVECEXP (seq, 0, 1)) == INSN | |
2333 | && GET_CODE (PATTERN (XVECEXP (seq, 0, 1))) == SET | |
2334 | && SET_DEST (PATTERN (XVECEXP (seq, 0, 1))) == reg | |
2335 | && memory_address_p (GET_MODE (x), | |
2336 | SET_SRC (PATTERN (XVECEXP (seq, 0, 1))))) | |
2337 | { | |
2338 | rtx src1 = SET_SRC (PATTERN (XVECEXP (seq, 0, 0))); | |
2339 | rtx src2 = SET_SRC (PATTERN (XVECEXP (seq, 0, 1))); | |
2340 | ||
2341 | /* Replace the placeholder in SRC2 with SRC1. If we can | |
2342 | find where in SRC2 it was placed, that can become our | |
2343 | split point and we can replace this address with SRC2. | |
2344 | Just try two obvious places. */ | |
2345 | ||
2346 | src2 = replace_rtx (src2, reg, src1); | |
2347 | split = 0; | |
2348 | if (XEXP (src2, 0) == src1) | |
2349 | split = &XEXP (src2, 0); | |
2350 | else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e' | |
2351 | && XEXP (XEXP (src2, 0), 0) == src1) | |
2352 | split = &XEXP (XEXP (src2, 0), 0); | |
2353 | ||
2354 | if (split) | |
2355 | { | |
2356 | SUBST (XEXP (x, 0), src2); | |
2357 | return split; | |
2358 | } | |
2359 | } | |
2360 | ||
2361 | /* If that didn't work, perhaps the first operand is complex and | |
2362 | needs to be computed separately, so make a split point there. | |
2363 | This will occur on machines that just support REG + CONST | |
2364 | and have a constant moved through some previous computation. */ | |
2365 | ||
2366 | else if (GET_RTX_CLASS (GET_CODE (XEXP (XEXP (x, 0), 0))) != 'o' | |
2367 | && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG | |
2368 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (XEXP (x, 0), 0)))) | |
2369 | == 'o'))) | |
2370 | return &XEXP (XEXP (x, 0), 0); | |
2371 | } | |
2372 | break; | |
2373 | ||
2374 | case SET: | |
2375 | #ifdef HAVE_cc0 | |
2376 | /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a | |
2377 | ZERO_EXTRACT, the most likely reason why this doesn't match is that | |
2378 | we need to put the operand into a register. So split at that | |
2379 | point. */ | |
2380 | ||
2381 | if (SET_DEST (x) == cc0_rtx | |
2382 | && GET_CODE (SET_SRC (x)) != COMPARE | |
2383 | && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT | |
2384 | && GET_RTX_CLASS (GET_CODE (SET_SRC (x))) != 'o' | |
2385 | && ! (GET_CODE (SET_SRC (x)) == SUBREG | |
2386 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (SET_SRC (x)))) == 'o')) | |
2387 | return &SET_SRC (x); | |
2388 | #endif | |
2389 | ||
2390 | /* See if we can split SET_SRC as it stands. */ | |
2391 | split = find_split_point (&SET_SRC (x), insn); | |
2392 | if (split && split != &SET_SRC (x)) | |
2393 | return split; | |
2394 | ||
2395 | /* See if this is a bitfield assignment with everything constant. If | |
2396 | so, this is an IOR of an AND, so split it into that. */ | |
2397 | if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT | |
2398 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))) | |
2399 | <= HOST_BITS_PER_WIDE_INT) | |
2400 | && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT | |
2401 | && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT | |
2402 | && GET_CODE (SET_SRC (x)) == CONST_INT | |
2403 | && ((INTVAL (XEXP (SET_DEST (x), 1)) | |
2404 | + INTVAL (XEXP (SET_DEST (x), 2))) | |
2405 | <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))) | |
2406 | && ! side_effects_p (XEXP (SET_DEST (x), 0))) | |
2407 | { | |
2408 | int pos = INTVAL (XEXP (SET_DEST (x), 2)); | |
2409 | int len = INTVAL (XEXP (SET_DEST (x), 1)); | |
2410 | int src = INTVAL (SET_SRC (x)); | |
2411 | rtx dest = XEXP (SET_DEST (x), 0); | |
2412 | enum machine_mode mode = GET_MODE (dest); | |
2413 | unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1; | |
2414 | ||
2415 | #if BITS_BIG_ENDIAN | |
2416 | pos = GET_MODE_BITSIZE (mode) - len - pos; | |
2417 | #endif | |
2418 | ||
2419 | if (src == mask) | |
2420 | SUBST (SET_SRC (x), | |
2421 | gen_binary (IOR, mode, dest, GEN_INT (src << pos))); | |
2422 | else | |
2423 | SUBST (SET_SRC (x), | |
2424 | gen_binary (IOR, mode, | |
2425 | gen_binary (AND, mode, dest, | |
2426 | GEN_INT (~ (mask << pos) | |
2427 | & GET_MODE_MASK (mode))), | |
2428 | GEN_INT (src << pos))); | |
2429 | ||
2430 | SUBST (SET_DEST (x), dest); | |
2431 | ||
2432 | split = find_split_point (&SET_SRC (x), insn); | |
2433 | if (split && split != &SET_SRC (x)) | |
2434 | return split; | |
2435 | } | |
2436 | ||
2437 | /* Otherwise, see if this is an operation that we can split into two. | |
2438 | If so, try to split that. */ | |
2439 | code = GET_CODE (SET_SRC (x)); | |
2440 | ||
2441 | switch (code) | |
2442 | { | |
2443 | case AND: | |
2444 | /* If we are AND'ing with a large constant that is only a single | |
2445 | bit and the result is only being used in a context where we | |
2446 | need to know if it is zero or non-zero, replace it with a bit | |
2447 | extraction. This will avoid the large constant, which might | |
2448 | have taken more than one insn to make. If the constant were | |
2449 | not a valid argument to the AND but took only one insn to make, | |
2450 | this is no worse, but if it took more than one insn, it will | |
2451 | be better. */ | |
2452 | ||
2453 | if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT | |
2454 | && GET_CODE (XEXP (SET_SRC (x), 0)) == REG | |
2455 | && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7 | |
2456 | && GET_CODE (SET_DEST (x)) == REG | |
2457 | && (split = find_single_use (SET_DEST (x), insn, NULL_PTR)) != 0 | |
2458 | && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE) | |
2459 | && XEXP (*split, 0) == SET_DEST (x) | |
2460 | && XEXP (*split, 1) == const0_rtx) | |
2461 | { | |
2462 | SUBST (SET_SRC (x), | |
2463 | make_extraction (GET_MODE (SET_DEST (x)), | |
2464 | XEXP (SET_SRC (x), 0), | |
2465 | pos, NULL_RTX, 1, 1, 0, 0)); | |
2466 | return find_split_point (loc, insn); | |
2467 | } | |
2468 | break; | |
2469 | ||
2470 | case SIGN_EXTEND: | |
2471 | inner = XEXP (SET_SRC (x), 0); | |
2472 | pos = 0; | |
2473 | len = GET_MODE_BITSIZE (GET_MODE (inner)); | |
2474 | unsignedp = 0; | |
2475 | break; | |
2476 | ||
2477 | case SIGN_EXTRACT: | |
2478 | case ZERO_EXTRACT: | |
2479 | if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT | |
2480 | && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT) | |
2481 | { | |
2482 | inner = XEXP (SET_SRC (x), 0); | |
2483 | len = INTVAL (XEXP (SET_SRC (x), 1)); | |
2484 | pos = INTVAL (XEXP (SET_SRC (x), 2)); | |
2485 | ||
2486 | #if BITS_BIG_ENDIAN | |
2487 | pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos; | |
2488 | #endif | |
2489 | unsignedp = (code == ZERO_EXTRACT); | |
2490 | } | |
2491 | break; | |
2492 | } | |
2493 | ||
2494 | if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner))) | |
2495 | { | |
2496 | enum machine_mode mode = GET_MODE (SET_SRC (x)); | |
2497 | ||
2498 | /* For unsigned, we have a choice of a shift followed by an | |
2499 | AND or two shifts. Use two shifts for field sizes where the | |
2500 | constant might be too large. We assume here that we can | |
2501 | always at least get 8-bit constants in an AND insn, which is | |
2502 | true for every current RISC. */ | |
2503 | ||
2504 | if (unsignedp && len <= 8) | |
2505 | { | |
2506 | SUBST (SET_SRC (x), | |
2507 | gen_rtx_combine | |
2508 | (AND, mode, | |
2509 | gen_rtx_combine (LSHIFTRT, mode, | |
2510 | gen_lowpart_for_combine (mode, inner), | |
2511 | GEN_INT (pos)), | |
2512 | GEN_INT (((HOST_WIDE_INT) 1 << len) - 1))); | |
2513 | ||
2514 | split = find_split_point (&SET_SRC (x), insn); | |
2515 | if (split && split != &SET_SRC (x)) | |
2516 | return split; | |
2517 | } | |
2518 | else | |
2519 | { | |
2520 | SUBST (SET_SRC (x), | |
2521 | gen_rtx_combine | |
2522 | (unsignedp ? LSHIFTRT : ASHIFTRT, mode, | |
2523 | gen_rtx_combine (ASHIFT, mode, | |
2524 | gen_lowpart_for_combine (mode, inner), | |
2525 | GEN_INT (GET_MODE_BITSIZE (mode) | |
2526 | - len - pos)), | |
2527 | GEN_INT (GET_MODE_BITSIZE (mode) - len))); | |
2528 | ||
2529 | split = find_split_point (&SET_SRC (x), insn); | |
2530 | if (split && split != &SET_SRC (x)) | |
2531 | return split; | |
2532 | } | |
2533 | } | |
2534 | ||
2535 | /* See if this is a simple operation with a constant as the second | |
2536 | operand. It might be that this constant is out of range and hence | |
2537 | could be used as a split point. */ | |
2538 | if ((GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '2' | |
2539 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == 'c' | |
2540 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '<') | |
2541 | && CONSTANT_P (XEXP (SET_SRC (x), 1)) | |
2542 | && (GET_RTX_CLASS (GET_CODE (XEXP (SET_SRC (x), 0))) == 'o' | |
2543 | || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG | |
2544 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (SET_SRC (x), 0)))) | |
2545 | == 'o')))) | |
2546 | return &XEXP (SET_SRC (x), 1); | |
2547 | ||
2548 | /* Finally, see if this is a simple operation with its first operand | |
2549 | not in a register. The operation might require this operand in a | |
2550 | register, so return it as a split point. We can always do this | |
2551 | because if the first operand were another operation, we would have | |
2552 | already found it as a split point. */ | |
2553 | if ((GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '2' | |
2554 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == 'c' | |
2555 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '<' | |
2556 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '1') | |
2557 | && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode)) | |
2558 | return &XEXP (SET_SRC (x), 0); | |
2559 | ||
2560 | return 0; | |
2561 | ||
2562 | case AND: | |
2563 | case IOR: | |
2564 | /* We write NOR as (and (not A) (not B)), but if we don't have a NOR, | |
2565 | it is better to write this as (not (ior A B)) so we can split it. | |
2566 | Similarly for IOR. */ | |
2567 | if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT) | |
2568 | { | |
2569 | SUBST (*loc, | |
2570 | gen_rtx_combine (NOT, GET_MODE (x), | |
2571 | gen_rtx_combine (code == IOR ? AND : IOR, | |
2572 | GET_MODE (x), | |
2573 | XEXP (XEXP (x, 0), 0), | |
2574 | XEXP (XEXP (x, 1), 0)))); | |
2575 | return find_split_point (loc, insn); | |
2576 | } | |
2577 | ||
2578 | /* Many RISC machines have a large set of logical insns. If the | |
2579 | second operand is a NOT, put it first so we will try to split the | |
2580 | other operand first. */ | |
2581 | if (GET_CODE (XEXP (x, 1)) == NOT) | |
2582 | { | |
2583 | rtx tem = XEXP (x, 0); | |
2584 | SUBST (XEXP (x, 0), XEXP (x, 1)); | |
2585 | SUBST (XEXP (x, 1), tem); | |
2586 | } | |
2587 | break; | |
2588 | } | |
2589 | ||
2590 | /* Otherwise, select our actions depending on our rtx class. */ | |
2591 | switch (GET_RTX_CLASS (code)) | |
2592 | { | |
2593 | case 'b': /* This is ZERO_EXTRACT and SIGN_EXTRACT. */ | |
2594 | case '3': | |
2595 | split = find_split_point (&XEXP (x, 2), insn); | |
2596 | if (split) | |
2597 | return split; | |
2598 | /* ... fall through ... */ | |
2599 | case '2': | |
2600 | case 'c': | |
2601 | case '<': | |
2602 | split = find_split_point (&XEXP (x, 1), insn); | |
2603 | if (split) | |
2604 | return split; | |
2605 | /* ... fall through ... */ | |
2606 | case '1': | |
2607 | /* Some machines have (and (shift ...) ...) insns. If X is not | |
2608 | an AND, but XEXP (X, 0) is, use it as our split point. */ | |
2609 | if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND) | |
2610 | return &XEXP (x, 0); | |
2611 | ||
2612 | split = find_split_point (&XEXP (x, 0), insn); | |
2613 | if (split) | |
2614 | return split; | |
2615 | return loc; | |
2616 | } | |
2617 | ||
2618 | /* Otherwise, we don't have a split point. */ | |
2619 | return 0; | |
2620 | } | |
2621 | \f | |
2622 | /* Throughout X, replace FROM with TO, and return the result. | |
2623 | The result is TO if X is FROM; | |
2624 | otherwise the result is X, but its contents may have been modified. | |
2625 | If they were modified, a record was made in undobuf so that | |
2626 | undo_all will (among other things) return X to its original state. | |
2627 | ||
2628 | If the number of changes necessary is too much to record to undo, | |
2629 | the excess changes are not made, so the result is invalid. | |
2630 | The changes already made can still be undone. | |
2631 | undobuf.num_undo is incremented for such changes, so by testing that | |
2632 | the caller can tell whether the result is valid. | |
2633 | ||
2634 | `n_occurrences' is incremented each time FROM is replaced. | |
2635 | ||
2636 | IN_DEST is non-zero if we are processing the SET_DEST of a SET. | |
2637 | ||
2638 | UNIQUE_COPY is non-zero if each substitution must be unique. We do this | |
2639 | by copying if `n_occurrences' is non-zero. */ | |
2640 | ||
2641 | static rtx | |
2642 | subst (x, from, to, in_dest, unique_copy) | |
2643 | register rtx x, from, to; | |
2644 | int in_dest; | |
2645 | int unique_copy; | |
2646 | { | |
2647 | register char *fmt; | |
2648 | register int len, i; | |
2649 | register enum rtx_code code = GET_CODE (x), orig_code = code; | |
2650 | rtx temp; | |
2651 | enum machine_mode mode = GET_MODE (x); | |
2652 | enum machine_mode op0_mode = VOIDmode; | |
2653 | rtx other_insn; | |
2654 | rtx *cc_use; | |
2655 | int n_restarts = 0; | |
2656 | ||
2657 | /* FAKE_EXTEND_SAFE_P (MODE, FROM) is 1 if (subreg:MODE FROM 0) is a safe | |
2658 | replacement for (zero_extend:MODE FROM) or (sign_extend:MODE FROM). | |
2659 | If it is 0, that cannot be done. We can now do this for any MEM | |
2660 | because (SUBREG (MEM...)) is guaranteed to cause the MEM to be reloaded. | |
2661 | If not for that, MEM's would very rarely be safe. */ | |
2662 | ||
2663 | /* Reject MODEs bigger than a word, because we might not be able | |
2664 | to reference a two-register group starting with an arbitrary register | |
2665 | (and currently gen_lowpart might crash for a SUBREG). */ | |
2666 | ||
2667 | #define FAKE_EXTEND_SAFE_P(MODE, FROM) \ | |
2668 | (GET_MODE_SIZE (MODE) <= UNITS_PER_WORD) | |
2669 | ||
2670 | /* Two expressions are equal if they are identical copies of a shared | |
2671 | RTX or if they are both registers with the same register number | |
2672 | and mode. */ | |
2673 | ||
2674 | #define COMBINE_RTX_EQUAL_P(X,Y) \ | |
2675 | ((X) == (Y) \ | |
2676 | || (GET_CODE (X) == REG && GET_CODE (Y) == REG \ | |
2677 | && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y))) | |
2678 | ||
2679 | if (! in_dest && COMBINE_RTX_EQUAL_P (x, from)) | |
2680 | { | |
2681 | n_occurrences++; | |
2682 | return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to); | |
2683 | } | |
2684 | ||
2685 | /* If X and FROM are the same register but different modes, they will | |
2686 | not have been seen as equal above. However, flow.c will make a | |
2687 | LOG_LINKS entry for that case. If we do nothing, we will try to | |
2688 | rerecognize our original insn and, when it succeeds, we will | |
2689 | delete the feeding insn, which is incorrect. | |
2690 | ||
2691 | So force this insn not to match in this (rare) case. */ | |
2692 | if (! in_dest && code == REG && GET_CODE (from) == REG | |
2693 | && REGNO (x) == REGNO (from)) | |
2694 | return gen_rtx (CLOBBER, GET_MODE (x), const0_rtx); | |
2695 | ||
2696 | /* If this is an object, we are done unless it is a MEM or LO_SUM, both | |
2697 | of which may contain things that can be combined. */ | |
2698 | if (code != MEM && code != LO_SUM && GET_RTX_CLASS (code) == 'o') | |
2699 | return x; | |
2700 | ||
2701 | /* It is possible to have a subexpression appear twice in the insn. | |
2702 | Suppose that FROM is a register that appears within TO. | |
2703 | Then, after that subexpression has been scanned once by `subst', | |
2704 | the second time it is scanned, TO may be found. If we were | |
2705 | to scan TO here, we would find FROM within it and create a | |
2706 | self-referent rtl structure which is completely wrong. */ | |
2707 | if (COMBINE_RTX_EQUAL_P (x, to)) | |
2708 | return to; | |
2709 | ||
2710 | len = GET_RTX_LENGTH (code); | |
2711 | fmt = GET_RTX_FORMAT (code); | |
2712 | ||
2713 | /* We don't need to process a SET_DEST that is a register, CC0, or PC, so | |
2714 | set up to skip this common case. All other cases where we want to | |
2715 | suppress replacing something inside a SET_SRC are handled via the | |
2716 | IN_DEST operand. */ | |
2717 | if (code == SET | |
2718 | && (GET_CODE (SET_DEST (x)) == REG | |
2719 | || GET_CODE (SET_DEST (x)) == CC0 | |
2720 | || GET_CODE (SET_DEST (x)) == PC)) | |
2721 | fmt = "ie"; | |
2722 | ||
2723 | /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a constant. */ | |
2724 | if (fmt[0] == 'e') | |
2725 | op0_mode = GET_MODE (XEXP (x, 0)); | |
2726 | ||
2727 | for (i = 0; i < len; i++) | |
2728 | { | |
2729 | if (fmt[i] == 'E') | |
2730 | { | |
2731 | register int j; | |
2732 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
2733 | { | |
2734 | register rtx new; | |
2735 | if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from)) | |
2736 | { | |
2737 | new = (unique_copy && n_occurrences ? copy_rtx (to) : to); | |
2738 | n_occurrences++; | |
2739 | } | |
2740 | else | |
2741 | { | |
2742 | new = subst (XVECEXP (x, i, j), from, to, 0, unique_copy); | |
2743 | ||
2744 | /* If this substitution failed, this whole thing fails. */ | |
2745 | if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx) | |
2746 | return new; | |
2747 | } | |
2748 | ||
2749 | SUBST (XVECEXP (x, i, j), new); | |
2750 | } | |
2751 | } | |
2752 | else if (fmt[i] == 'e') | |
2753 | { | |
2754 | register rtx new; | |
2755 | ||
2756 | if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from)) | |
2757 | { | |
2758 | new = (unique_copy && n_occurrences ? copy_rtx (to) : to); | |
2759 | n_occurrences++; | |
2760 | } | |
2761 | else | |
2762 | /* If we are in a SET_DEST, suppress most cases unless we | |
2763 | have gone inside a MEM, in which case we want to | |
2764 | simplify the address. We assume here that things that | |
2765 | are actually part of the destination have their inner | |
2766 | parts in the first expression. This is true for SUBREG, | |
2767 | STRICT_LOW_PART, and ZERO_EXTRACT, which are the only | |
2768 | things aside from REG and MEM that should appear in a | |
2769 | SET_DEST. */ | |
2770 | new = subst (XEXP (x, i), from, to, | |
2771 | (((in_dest | |
2772 | && (code == SUBREG || code == STRICT_LOW_PART | |
2773 | || code == ZERO_EXTRACT)) | |
2774 | || code == SET) | |
2775 | && i == 0), unique_copy); | |
2776 | ||
2777 | /* If we found that we will have to reject this combination, | |
2778 | indicate that by returning the CLOBBER ourselves, rather than | |
2779 | an expression containing it. This will speed things up as | |
2780 | well as prevent accidents where two CLOBBERs are considered | |
2781 | to be equal, thus producing an incorrect simplification. */ | |
2782 | ||
2783 | if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx) | |
2784 | return new; | |
2785 | ||
2786 | SUBST (XEXP (x, i), new); | |
2787 | } | |
2788 | } | |
2789 | ||
2790 | /* We come back to here if we have replaced the expression with one of | |
2791 | a different code and it is likely that further simplification will be | |
2792 | possible. */ | |
2793 | ||
2794 | restart: | |
2795 | ||
2796 | /* If we have restarted more than 4 times, we are probably looping, so | |
2797 | give up. */ | |
2798 | if (++n_restarts > 4) | |
2799 | return x; | |
2800 | ||
2801 | /* If we are restarting at all, it means that we no longer know the | |
2802 | original mode of operand 0 (since we have probably changed the | |
2803 | form of X). */ | |
2804 | ||
2805 | if (n_restarts > 1) | |
2806 | op0_mode = VOIDmode; | |
2807 | ||
2808 | code = GET_CODE (x); | |
2809 | ||
2810 | /* If this is a commutative operation, put a constant last and a complex | |
2811 | expression first. We don't need to do this for comparisons here. */ | |
2812 | if (GET_RTX_CLASS (code) == 'c' | |
2813 | && ((CONSTANT_P (XEXP (x, 0)) && GET_CODE (XEXP (x, 1)) != CONST_INT) | |
2814 | || (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == 'o' | |
2815 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) != 'o') | |
2816 | || (GET_CODE (XEXP (x, 0)) == SUBREG | |
2817 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 0)))) == 'o' | |
2818 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) != 'o'))) | |
2819 | { | |
2820 | temp = XEXP (x, 0); | |
2821 | SUBST (XEXP (x, 0), XEXP (x, 1)); | |
2822 | SUBST (XEXP (x, 1), temp); | |
2823 | } | |
2824 | ||
2825 | /* If this is a PLUS, MINUS, or MULT, and the first operand is the | |
2826 | sign extension of a PLUS with a constant, reverse the order of the sign | |
2827 | extension and the addition. Note that this not the same as the original | |
2828 | code, but overflow is undefined for signed values. Also note that the | |
2829 | PLUS will have been partially moved "inside" the sign-extension, so that | |
2830 | the first operand of X will really look like: | |
2831 | (ashiftrt (plus (ashift A C4) C5) C4). | |
2832 | We convert this to | |
2833 | (plus (ashiftrt (ashift A C4) C2) C4) | |
2834 | and replace the first operand of X with that expression. Later parts | |
2835 | of this function may simplify the expression further. | |
2836 | ||
2837 | For example, if we start with (mult (sign_extend (plus A C1)) C2), | |
2838 | we swap the SIGN_EXTEND and PLUS. Later code will apply the | |
2839 | distributive law to produce (plus (mult (sign_extend X) C1) C3). | |
2840 | ||
2841 | We do this to simplify address expressions. */ | |
2842 | ||
2843 | if ((code == PLUS || code == MINUS || code == MULT) | |
2844 | && GET_CODE (XEXP (x, 0)) == ASHIFTRT | |
2845 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == PLUS | |
2846 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ASHIFT | |
2847 | && GET_CODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1)) == CONST_INT | |
2848 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
2849 | && XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1) == XEXP (XEXP (x, 0), 1) | |
2850 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
2851 | && (temp = simplify_binary_operation (ASHIFTRT, mode, | |
2852 | XEXP (XEXP (XEXP (x, 0), 0), 1), | |
2853 | XEXP (XEXP (x, 0), 1))) != 0) | |
2854 | { | |
2855 | rtx new | |
2856 | = simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
2857 | XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), | |
2858 | INTVAL (XEXP (XEXP (x, 0), 1))); | |
2859 | ||
2860 | new = simplify_shift_const (NULL_RTX, ASHIFTRT, mode, new, | |
2861 | INTVAL (XEXP (XEXP (x, 0), 1))); | |
2862 | ||
2863 | SUBST (XEXP (x, 0), gen_binary (PLUS, mode, new, temp)); | |
2864 | } | |
2865 | ||
2866 | /* If this is a simple operation applied to an IF_THEN_ELSE, try | |
2867 | applying it to the arms of the IF_THEN_ELSE. This often simplifies | |
2868 | things. Don't deal with operations that change modes here. */ | |
2869 | ||
2870 | if ((GET_RTX_CLASS (code) == '2' || GET_RTX_CLASS (code) == 'c') | |
2871 | && GET_CODE (XEXP (x, 0)) == IF_THEN_ELSE) | |
2872 | { | |
2873 | /* Don't do this by using SUBST inside X since we might be messing | |
2874 | up a shared expression. */ | |
2875 | rtx cond = XEXP (XEXP (x, 0), 0); | |
2876 | rtx t_arm = subst (gen_binary (code, mode, XEXP (XEXP (x, 0), 1), | |
2877 | XEXP (x, 1)), | |
2878 | pc_rtx, pc_rtx, 0, 0); | |
2879 | rtx f_arm = subst (gen_binary (code, mode, XEXP (XEXP (x, 0), 2), | |
2880 | XEXP (x, 1)), | |
2881 | pc_rtx, pc_rtx, 0, 0); | |
2882 | ||
2883 | ||
2884 | x = gen_rtx (IF_THEN_ELSE, mode, cond, t_arm, f_arm); | |
2885 | goto restart; | |
2886 | } | |
2887 | ||
2888 | else if (GET_RTX_CLASS (code) == '1' | |
2889 | && GET_CODE (XEXP (x, 0)) == IF_THEN_ELSE | |
2890 | && GET_MODE (XEXP (x, 0)) == mode) | |
2891 | { | |
2892 | rtx cond = XEXP (XEXP (x, 0), 0); | |
2893 | rtx t_arm = subst (gen_unary (code, mode, XEXP (XEXP (x, 0), 1)), | |
2894 | pc_rtx, pc_rtx, 0, 0); | |
2895 | rtx f_arm = subst (gen_unary (code, mode, XEXP (XEXP (x, 0), 2)), | |
2896 | pc_rtx, pc_rtx, 0, 0); | |
2897 | ||
2898 | x = gen_rtx_combine (IF_THEN_ELSE, mode, cond, t_arm, f_arm); | |
2899 | goto restart; | |
2900 | } | |
2901 | ||
2902 | /* Try to fold this expression in case we have constants that weren't | |
2903 | present before. */ | |
2904 | temp = 0; | |
2905 | switch (GET_RTX_CLASS (code)) | |
2906 | { | |
2907 | case '1': | |
2908 | temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode); | |
2909 | break; | |
2910 | case '<': | |
2911 | temp = simplify_relational_operation (code, op0_mode, | |
2912 | XEXP (x, 0), XEXP (x, 1)); | |
2913 | #ifdef FLOAT_STORE_FLAG_VALUE | |
2914 | if (temp != 0 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
2915 | temp = ((temp == const0_rtx) ? CONST0_RTX (GET_MODE (x)) | |
2916 | : immed_real_const_1 (FLOAT_STORE_FLAG_VALUE, GET_MODE (x))); | |
2917 | #endif | |
2918 | break; | |
2919 | case 'c': | |
2920 | case '2': | |
2921 | temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1)); | |
2922 | break; | |
2923 | case 'b': | |
2924 | case '3': | |
2925 | temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0), | |
2926 | XEXP (x, 1), XEXP (x, 2)); | |
2927 | break; | |
2928 | } | |
2929 | ||
2930 | if (temp) | |
2931 | x = temp, code = GET_CODE (temp); | |
2932 | ||
2933 | /* First see if we can apply the inverse distributive law. */ | |
2934 | if (code == PLUS || code == MINUS || code == IOR || code == XOR) | |
2935 | { | |
2936 | x = apply_distributive_law (x); | |
2937 | code = GET_CODE (x); | |
2938 | } | |
2939 | ||
2940 | /* If CODE is an associative operation not otherwise handled, see if we | |
2941 | can associate some operands. This can win if they are constants or | |
2942 | if they are logically related (i.e. (a & b) & a. */ | |
2943 | if ((code == PLUS || code == MINUS | |
2944 | || code == MULT || code == AND || code == IOR || code == XOR | |
2945 | || code == DIV || code == UDIV | |
2946 | || code == SMAX || code == SMIN || code == UMAX || code == UMIN) | |
2947 | && GET_MODE_CLASS (mode) == MODE_INT) | |
2948 | { | |
2949 | if (GET_CODE (XEXP (x, 0)) == code) | |
2950 | { | |
2951 | rtx other = XEXP (XEXP (x, 0), 0); | |
2952 | rtx inner_op0 = XEXP (XEXP (x, 0), 1); | |
2953 | rtx inner_op1 = XEXP (x, 1); | |
2954 | rtx inner; | |
2955 | ||
2956 | /* Make sure we pass the constant operand if any as the second | |
2957 | one if this is a commutative operation. */ | |
2958 | if (CONSTANT_P (inner_op0) && GET_RTX_CLASS (code) == 'c') | |
2959 | { | |
2960 | rtx tem = inner_op0; | |
2961 | inner_op0 = inner_op1; | |
2962 | inner_op1 = tem; | |
2963 | } | |
2964 | inner = simplify_binary_operation (code == MINUS ? PLUS | |
2965 | : code == DIV ? MULT | |
2966 | : code == UDIV ? MULT | |
2967 | : code, | |
2968 | mode, inner_op0, inner_op1); | |
2969 | ||
2970 | /* For commutative operations, try the other pair if that one | |
2971 | didn't simplify. */ | |
2972 | if (inner == 0 && GET_RTX_CLASS (code) == 'c') | |
2973 | { | |
2974 | other = XEXP (XEXP (x, 0), 1); | |
2975 | inner = simplify_binary_operation (code, mode, | |
2976 | XEXP (XEXP (x, 0), 0), | |
2977 | XEXP (x, 1)); | |
2978 | } | |
2979 | ||
2980 | if (inner) | |
2981 | { | |
2982 | x = gen_binary (code, mode, other, inner); | |
2983 | goto restart; | |
2984 | ||
2985 | } | |
2986 | } | |
2987 | } | |
2988 | ||
2989 | /* A little bit of algebraic simplification here. */ | |
2990 | switch (code) | |
2991 | { | |
2992 | case MEM: | |
2993 | /* Ensure that our address has any ASHIFTs converted to MULT in case | |
2994 | address-recognizing predicates are called later. */ | |
2995 | temp = make_compound_operation (XEXP (x, 0), MEM); | |
2996 | SUBST (XEXP (x, 0), temp); | |
2997 | break; | |
2998 | ||
2999 | case SUBREG: | |
3000 | /* (subreg:A (mem:B X) N) becomes a modified MEM unless the SUBREG | |
3001 | is paradoxical. If we can't do that safely, then it becomes | |
3002 | something nonsensical so that this combination won't take place. */ | |
3003 | ||
3004 | if (GET_CODE (SUBREG_REG (x)) == MEM | |
3005 | && (GET_MODE_SIZE (mode) | |
3006 | <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))) | |
3007 | { | |
3008 | rtx inner = SUBREG_REG (x); | |
3009 | int endian_offset = 0; | |
3010 | /* Don't change the mode of the MEM | |
3011 | if that would change the meaning of the address. */ | |
3012 | if (MEM_VOLATILE_P (SUBREG_REG (x)) | |
3013 | || mode_dependent_address_p (XEXP (inner, 0))) | |
3014 | return gen_rtx (CLOBBER, mode, const0_rtx); | |
3015 | ||
3016 | #if BYTES_BIG_ENDIAN | |
3017 | if (GET_MODE_SIZE (mode) < UNITS_PER_WORD) | |
3018 | endian_offset += UNITS_PER_WORD - GET_MODE_SIZE (mode); | |
3019 | if (GET_MODE_SIZE (GET_MODE (inner)) < UNITS_PER_WORD) | |
3020 | endian_offset -= UNITS_PER_WORD - GET_MODE_SIZE (GET_MODE (inner)); | |
3021 | #endif | |
3022 | /* Note if the plus_constant doesn't make a valid address | |
3023 | then this combination won't be accepted. */ | |
3024 | x = gen_rtx (MEM, mode, | |
3025 | plus_constant (XEXP (inner, 0), | |
3026 | (SUBREG_WORD (x) * UNITS_PER_WORD | |
3027 | + endian_offset))); | |
3028 | MEM_VOLATILE_P (x) = MEM_VOLATILE_P (inner); | |
3029 | RTX_UNCHANGING_P (x) = RTX_UNCHANGING_P (inner); | |
3030 | MEM_IN_STRUCT_P (x) = MEM_IN_STRUCT_P (inner); | |
3031 | return x; | |
3032 | } | |
3033 | ||
3034 | /* If we are in a SET_DEST, these other cases can't apply. */ | |
3035 | if (in_dest) | |
3036 | return x; | |
3037 | ||
3038 | /* Changing mode twice with SUBREG => just change it once, | |
3039 | or not at all if changing back to starting mode. */ | |
3040 | if (GET_CODE (SUBREG_REG (x)) == SUBREG) | |
3041 | { | |
3042 | if (mode == GET_MODE (SUBREG_REG (SUBREG_REG (x))) | |
3043 | && SUBREG_WORD (x) == 0 && SUBREG_WORD (SUBREG_REG (x)) == 0) | |
3044 | return SUBREG_REG (SUBREG_REG (x)); | |
3045 | ||
3046 | SUBST_INT (SUBREG_WORD (x), | |
3047 | SUBREG_WORD (x) + SUBREG_WORD (SUBREG_REG (x))); | |
3048 | SUBST (SUBREG_REG (x), SUBREG_REG (SUBREG_REG (x))); | |
3049 | } | |
3050 | ||
3051 | /* SUBREG of a hard register => just change the register number | |
3052 | and/or mode. If the hard register is not valid in that mode, | |
3053 | suppress this combination. If the hard register is the stack, | |
3054 | frame, or argument pointer, leave this as a SUBREG. */ | |
3055 | ||
3056 | if (GET_CODE (SUBREG_REG (x)) == REG | |
3057 | && REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER | |
3058 | && REGNO (SUBREG_REG (x)) != FRAME_POINTER_REGNUM | |
3059 | #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM | |
3060 | && REGNO (SUBREG_REG (x)) != ARG_POINTER_REGNUM | |
3061 | #endif | |
3062 | && REGNO (SUBREG_REG (x)) != STACK_POINTER_REGNUM) | |
3063 | { | |
3064 | if (HARD_REGNO_MODE_OK (REGNO (SUBREG_REG (x)) + SUBREG_WORD (x), | |
3065 | mode)) | |
3066 | return gen_rtx (REG, mode, | |
3067 | REGNO (SUBREG_REG (x)) + SUBREG_WORD (x)); | |
3068 | else | |
3069 | return gen_rtx (CLOBBER, mode, const0_rtx); | |
3070 | } | |
3071 | ||
3072 | /* For a constant, try to pick up the part we want. Handle a full | |
3073 | word and low-order part. Only do this if we are narrowing | |
3074 | the constant; if it is being widened, we have no idea what | |
3075 | the extra bits will have been set to. */ | |
3076 | ||
3077 | if (CONSTANT_P (SUBREG_REG (x)) && op0_mode != VOIDmode | |
3078 | && GET_MODE_SIZE (mode) == UNITS_PER_WORD | |
3079 | && GET_MODE_SIZE (op0_mode) < UNITS_PER_WORD | |
3080 | && GET_MODE_CLASS (mode) == MODE_INT) | |
3081 | { | |
3082 | temp = operand_subword (SUBREG_REG (x), SUBREG_WORD (x), | |
3083 | 0, op0_mode); | |
3084 | if (temp) | |
3085 | return temp; | |
3086 | } | |
3087 | ||
3088 | /* If we want a subreg of a constant, at offset 0, | |
3089 | take the low bits. On a little-endian machine, that's | |
3090 | always valid. On a big-endian machine, it's valid | |
3091 | only if the constant's mode fits in one word. */ | |
3092 | if (CONSTANT_P (SUBREG_REG (x)) && subreg_lowpart_p (x) | |
3093 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (op0_mode) | |
3094 | #if WORDS_BIG_ENDIAN | |
3095 | && GET_MODE_BITSIZE (op0_mode) <= BITS_PER_WORD | |
3096 | #endif | |
3097 | ) | |
3098 | return gen_lowpart_for_combine (mode, SUBREG_REG (x)); | |
3099 | ||
3100 | /* If we are narrowing the object, we need to see if we can simplify | |
3101 | the expression for the object knowing that we only need the | |
3102 | low-order bits. */ | |
3103 | ||
3104 | if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))) | |
3105 | && subreg_lowpart_p (x)) | |
3106 | return force_to_mode (SUBREG_REG (x), mode, GET_MODE_BITSIZE (mode), | |
3107 | NULL_RTX); | |
3108 | break; | |
3109 | ||
3110 | case NOT: | |
3111 | /* (not (plus X -1)) can become (neg X). */ | |
3112 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3113 | && XEXP (XEXP (x, 0), 1) == constm1_rtx) | |
3114 | { | |
3115 | x = gen_rtx_combine (NEG, mode, XEXP (XEXP (x, 0), 0)); | |
3116 | goto restart; | |
3117 | } | |
3118 | ||
3119 | /* Similarly, (not (neg X)) is (plus X -1). */ | |
3120 | if (GET_CODE (XEXP (x, 0)) == NEG) | |
3121 | { | |
3122 | x = gen_rtx_combine (PLUS, mode, XEXP (XEXP (x, 0), 0), constm1_rtx); | |
3123 | goto restart; | |
3124 | } | |
3125 | ||
3126 | /* (not (xor X C)) for C constant is (xor X D) with D = ~ C. */ | |
3127 | if (GET_CODE (XEXP (x, 0)) == XOR | |
3128 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3129 | && (temp = simplify_unary_operation (NOT, mode, | |
3130 | XEXP (XEXP (x, 0), 1), | |
3131 | mode)) != 0) | |
3132 | { | |
3133 | SUBST (XEXP (XEXP (x, 0), 1), temp); | |
3134 | return XEXP (x, 0); | |
3135 | } | |
3136 | ||
3137 | /* (not (ashift 1 X)) is (rotate ~1 X). We used to do this for operands | |
3138 | other than 1, but that is not valid. We could do a similar | |
3139 | simplification for (not (lshiftrt C X)) where C is just the sign bit, | |
3140 | but this doesn't seem common enough to bother with. */ | |
3141 | if (GET_CODE (XEXP (x, 0)) == ASHIFT | |
3142 | && XEXP (XEXP (x, 0), 0) == const1_rtx) | |
3143 | { | |
3144 | x = gen_rtx (ROTATE, mode, gen_unary (NOT, mode, const1_rtx), | |
3145 | XEXP (XEXP (x, 0), 1)); | |
3146 | goto restart; | |
3147 | } | |
3148 | ||
3149 | if (GET_CODE (XEXP (x, 0)) == SUBREG | |
3150 | && subreg_lowpart_p (XEXP (x, 0)) | |
3151 | && (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) | |
3152 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (x, 0))))) | |
3153 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == ASHIFT | |
3154 | && XEXP (SUBREG_REG (XEXP (x, 0)), 0) == const1_rtx) | |
3155 | { | |
3156 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (XEXP (x, 0))); | |
3157 | ||
3158 | x = gen_rtx (ROTATE, inner_mode, | |
3159 | gen_unary (NOT, inner_mode, const1_rtx), | |
3160 | XEXP (SUBREG_REG (XEXP (x, 0)), 1)); | |
3161 | x = gen_lowpart_for_combine (mode, x); | |
3162 | goto restart; | |
3163 | } | |
3164 | ||
3165 | #if STORE_FLAG_VALUE == -1 | |
3166 | /* (not (comparison foo bar)) can be done by reversing the comparison | |
3167 | code if valid. */ | |
3168 | if (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3169 | && reversible_comparison_p (XEXP (x, 0))) | |
3170 | return gen_rtx_combine (reverse_condition (GET_CODE (XEXP (x, 0))), | |
3171 | mode, XEXP (XEXP (x, 0), 0), | |
3172 | XEXP (XEXP (x, 0), 1)); | |
3173 | ||
3174 | /* (ashiftrt foo C) where C is the number of bits in FOO minus 1 | |
3175 | is (lt foo (const_int 0)), so we can perform the above | |
3176 | simplification. */ | |
3177 | ||
3178 | if (XEXP (x, 1) == const1_rtx | |
3179 | && GET_CODE (XEXP (x, 0)) == ASHIFTRT | |
3180 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3181 | && INTVAL (XEXP (XEXP (x, 0), 1)) == GET_MODE_BITSIZE (mode) - 1) | |
3182 | return gen_rtx_combine (GE, mode, XEXP (XEXP (x, 0), 0), const0_rtx); | |
3183 | #endif | |
3184 | ||
3185 | /* Apply De Morgan's laws to reduce number of patterns for machines | |
3186 | with negating logical insns (and-not, nand, etc.). If result has | |
3187 | only one NOT, put it first, since that is how the patterns are | |
3188 | coded. */ | |
3189 | ||
3190 | if (GET_CODE (XEXP (x, 0)) == IOR || GET_CODE (XEXP (x, 0)) == AND) | |
3191 | { | |
3192 | rtx in1 = XEXP (XEXP (x, 0), 0), in2 = XEXP (XEXP (x, 0), 1); | |
3193 | ||
3194 | if (GET_CODE (in1) == NOT) | |
3195 | in1 = XEXP (in1, 0); | |
3196 | else | |
3197 | in1 = gen_rtx_combine (NOT, GET_MODE (in1), in1); | |
3198 | ||
3199 | if (GET_CODE (in2) == NOT) | |
3200 | in2 = XEXP (in2, 0); | |
3201 | else if (GET_CODE (in2) == CONST_INT | |
3202 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
3203 | in2 = GEN_INT (GET_MODE_MASK (mode) & ~ INTVAL (in2)); | |
3204 | else | |
3205 | in2 = gen_rtx_combine (NOT, GET_MODE (in2), in2); | |
3206 | ||
3207 | if (GET_CODE (in2) == NOT) | |
3208 | { | |
3209 | rtx tem = in2; | |
3210 | in2 = in1; in1 = tem; | |
3211 | } | |
3212 | ||
3213 | x = gen_rtx_combine (GET_CODE (XEXP (x, 0)) == IOR ? AND : IOR, | |
3214 | mode, in1, in2); | |
3215 | goto restart; | |
3216 | } | |
3217 | break; | |
3218 | ||
3219 | case NEG: | |
3220 | /* (neg (plus X 1)) can become (not X). */ | |
3221 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3222 | && XEXP (XEXP (x, 0), 1) == const1_rtx) | |
3223 | { | |
3224 | x = gen_rtx_combine (NOT, mode, XEXP (XEXP (x, 0), 0)); | |
3225 | goto restart; | |
3226 | } | |
3227 | ||
3228 | /* Similarly, (neg (not X)) is (plus X 1). */ | |
3229 | if (GET_CODE (XEXP (x, 0)) == NOT) | |
3230 | { | |
3231 | x = gen_rtx_combine (PLUS, mode, XEXP (XEXP (x, 0), 0), const1_rtx); | |
3232 | goto restart; | |
3233 | } | |
3234 | ||
3235 | /* (neg (minus X Y)) can become (minus Y X). */ | |
3236 | if (GET_CODE (XEXP (x, 0)) == MINUS | |
3237 | && (GET_MODE_CLASS (mode) != MODE_FLOAT | |
3238 | /* x-y != -(y-x) with IEEE floating point. */ | |
3239 | || TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT)) | |
3240 | { | |
3241 | x = gen_binary (MINUS, mode, XEXP (XEXP (x, 0), 1), | |
3242 | XEXP (XEXP (x, 0), 0)); | |
3243 | goto restart; | |
3244 | } | |
3245 | ||
3246 | /* (neg (xor A 1)) is (plus A -1) if A is known to be either 0 or 1. */ | |
3247 | if (GET_CODE (XEXP (x, 0)) == XOR && XEXP (XEXP (x, 0), 1) == const1_rtx | |
3248 | && nonzero_bits (XEXP (XEXP (x, 0), 0), mode) == 1) | |
3249 | { | |
3250 | x = gen_binary (PLUS, mode, XEXP (XEXP (x, 0), 0), constm1_rtx); | |
3251 | goto restart; | |
3252 | } | |
3253 | ||
3254 | /* NEG commutes with ASHIFT since it is multiplication. Only do this | |
3255 | if we can then eliminate the NEG (e.g., | |
3256 | if the operand is a constant). */ | |
3257 | ||
3258 | if (GET_CODE (XEXP (x, 0)) == ASHIFT) | |
3259 | { | |
3260 | temp = simplify_unary_operation (NEG, mode, | |
3261 | XEXP (XEXP (x, 0), 0), mode); | |
3262 | if (temp) | |
3263 | { | |
3264 | SUBST (XEXP (XEXP (x, 0), 0), temp); | |
3265 | return XEXP (x, 0); | |
3266 | } | |
3267 | } | |
3268 | ||
3269 | temp = expand_compound_operation (XEXP (x, 0)); | |
3270 | ||
3271 | /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be | |
3272 | replaced by (lshiftrt X C). This will convert | |
3273 | (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */ | |
3274 | ||
3275 | if (GET_CODE (temp) == ASHIFTRT | |
3276 | && GET_CODE (XEXP (temp, 1)) == CONST_INT | |
3277 | && INTVAL (XEXP (temp, 1)) == GET_MODE_BITSIZE (mode) - 1) | |
3278 | { | |
3279 | x = simplify_shift_const (temp, LSHIFTRT, mode, XEXP (temp, 0), | |
3280 | INTVAL (XEXP (temp, 1))); | |
3281 | goto restart; | |
3282 | } | |
3283 | ||
3284 | /* If X has only a single bit that might be nonzero, say, bit I, convert | |
3285 | (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of | |
3286 | MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to | |
3287 | (sign_extract X 1 Y). But only do this if TEMP isn't a register | |
3288 | or a SUBREG of one since we'd be making the expression more | |
3289 | complex if it was just a register. */ | |
3290 | ||
3291 | if (GET_CODE (temp) != REG | |
3292 | && ! (GET_CODE (temp) == SUBREG | |
3293 | && GET_CODE (SUBREG_REG (temp)) == REG) | |
3294 | && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0) | |
3295 | { | |
3296 | rtx temp1 = simplify_shift_const | |
3297 | (NULL_RTX, ASHIFTRT, mode, | |
3298 | simplify_shift_const (NULL_RTX, ASHIFT, mode, temp, | |
3299 | GET_MODE_BITSIZE (mode) - 1 - i), | |
3300 | GET_MODE_BITSIZE (mode) - 1 - i); | |
3301 | ||
3302 | /* If all we did was surround TEMP with the two shifts, we | |
3303 | haven't improved anything, so don't use it. Otherwise, | |
3304 | we are better off with TEMP1. */ | |
3305 | if (GET_CODE (temp1) != ASHIFTRT | |
3306 | || GET_CODE (XEXP (temp1, 0)) != ASHIFT | |
3307 | || XEXP (XEXP (temp1, 0), 0) != temp) | |
3308 | { | |
3309 | x = temp1; | |
3310 | goto restart; | |
3311 | } | |
3312 | } | |
3313 | break; | |
3314 | ||
3315 | case FLOAT_TRUNCATE: | |
3316 | /* (float_truncate:SF (float_extend:DF foo:SF)) = foo:SF. */ | |
3317 | if (GET_CODE (XEXP (x, 0)) == FLOAT_EXTEND | |
3318 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode) | |
3319 | return XEXP (XEXP (x, 0), 0); | |
3320 | break; | |
3321 | ||
3322 | #ifdef HAVE_cc0 | |
3323 | case COMPARE: | |
3324 | /* Convert (compare FOO (const_int 0)) to FOO unless we aren't | |
3325 | using cc0, in which case we want to leave it as a COMPARE | |
3326 | so we can distinguish it from a register-register-copy. */ | |
3327 | if (XEXP (x, 1) == const0_rtx) | |
3328 | return XEXP (x, 0); | |
3329 | ||
3330 | /* In IEEE floating point, x-0 is not the same as x. */ | |
3331 | if ((TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
3332 | || GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) == MODE_INT) | |
3333 | && XEXP (x, 1) == CONST0_RTX (GET_MODE (XEXP (x, 0)))) | |
3334 | return XEXP (x, 0); | |
3335 | break; | |
3336 | #endif | |
3337 | ||
3338 | case CONST: | |
3339 | /* (const (const X)) can become (const X). Do it this way rather than | |
3340 | returning the inner CONST since CONST can be shared with a | |
3341 | REG_EQUAL note. */ | |
3342 | if (GET_CODE (XEXP (x, 0)) == CONST) | |
3343 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
3344 | break; | |
3345 | ||
3346 | #ifdef HAVE_lo_sum | |
3347 | case LO_SUM: | |
3348 | /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we | |
3349 | can add in an offset. find_split_point will split this address up | |
3350 | again if it doesn't match. */ | |
3351 | if (GET_CODE (XEXP (x, 0)) == HIGH | |
3352 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1))) | |
3353 | return XEXP (x, 1); | |
3354 | break; | |
3355 | #endif | |
3356 | ||
3357 | case PLUS: | |
3358 | /* If we have (plus (plus (A const) B)), associate it so that CONST is | |
3359 | outermost. That's because that's the way indexed addresses are | |
3360 | supposed to appear. This code used to check many more cases, but | |
3361 | they are now checked elsewhere. */ | |
3362 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3363 | && CONSTANT_ADDRESS_P (XEXP (XEXP (x, 0), 1))) | |
3364 | return gen_binary (PLUS, mode, | |
3365 | gen_binary (PLUS, mode, XEXP (XEXP (x, 0), 0), | |
3366 | XEXP (x, 1)), | |
3367 | XEXP (XEXP (x, 0), 1)); | |
3368 | ||
3369 | /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>) | |
3370 | when c is (const_int (pow2 + 1) / 2) is a sign extension of a | |
3371 | bit-field and can be replaced by either a sign_extend or a | |
3372 | sign_extract. The `and' may be a zero_extend. */ | |
3373 | if (GET_CODE (XEXP (x, 0)) == XOR | |
3374 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
3375 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3376 | && INTVAL (XEXP (x, 1)) == - INTVAL (XEXP (XEXP (x, 0), 1)) | |
3377 | && (i = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) >= 0 | |
3378 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
3379 | && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND | |
3380 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
3381 | && (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1)) | |
3382 | == ((HOST_WIDE_INT) 1 << (i + 1)) - 1)) | |
3383 | || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND | |
3384 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0))) | |
3385 | == i + 1)))) | |
3386 | { | |
3387 | x = simplify_shift_const | |
3388 | (NULL_RTX, ASHIFTRT, mode, | |
3389 | simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
3390 | XEXP (XEXP (XEXP (x, 0), 0), 0), | |
3391 | GET_MODE_BITSIZE (mode) - (i + 1)), | |
3392 | GET_MODE_BITSIZE (mode) - (i + 1)); | |
3393 | goto restart; | |
3394 | } | |
3395 | ||
3396 | /* If only the low-order bit of X is possible nonzero, (plus x -1) | |
3397 | can become (ashiftrt (ashift (xor x 1) C) C) where C is | |
3398 | the bitsize of the mode - 1. This allows simplification of | |
3399 | "a = (b & 8) == 0;" */ | |
3400 | if (XEXP (x, 1) == constm1_rtx | |
3401 | && GET_CODE (XEXP (x, 0)) != REG | |
3402 | && ! (GET_CODE (XEXP (x,0)) == SUBREG | |
3403 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG) | |
3404 | && nonzero_bits (XEXP (x, 0), mode) == 1) | |
3405 | { | |
3406 | x = simplify_shift_const | |
3407 | (NULL_RTX, ASHIFTRT, mode, | |
3408 | simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
3409 | gen_rtx_combine (XOR, mode, | |
3410 | XEXP (x, 0), const1_rtx), | |
3411 | GET_MODE_BITSIZE (mode) - 1), | |
3412 | GET_MODE_BITSIZE (mode) - 1); | |
3413 | goto restart; | |
3414 | } | |
3415 | ||
3416 | /* If we are adding two things that have no bits in common, convert | |
3417 | the addition into an IOR. This will often be further simplified, | |
3418 | for example in cases like ((a & 1) + (a & 2)), which can | |
3419 | become a & 3. */ | |
3420 | ||
3421 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
3422 | && (nonzero_bits (XEXP (x, 0), mode) | |
3423 | & nonzero_bits (XEXP (x, 1), mode)) == 0) | |
3424 | { | |
3425 | x = gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1)); | |
3426 | goto restart; | |
3427 | } | |
3428 | break; | |
3429 | ||
3430 | case MINUS: | |
3431 | /* (minus <foo> (and <foo> (const_int -pow2))) becomes | |
3432 | (and <foo> (const_int pow2-1)) */ | |
3433 | if (GET_CODE (XEXP (x, 1)) == AND | |
3434 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT | |
3435 | && exact_log2 (- INTVAL (XEXP (XEXP (x, 1), 1))) >= 0 | |
3436 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) | |
3437 | { | |
3438 | x = simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0), | |
3439 | - INTVAL (XEXP (XEXP (x, 1), 1)) - 1); | |
3440 | goto restart; | |
3441 | } | |
3442 | break; | |
3443 | ||
3444 | case MULT: | |
3445 | /* If we have (mult (plus A B) C), apply the distributive law and then | |
3446 | the inverse distributive law to see if things simplify. This | |
3447 | occurs mostly in addresses, often when unrolling loops. */ | |
3448 | ||
3449 | if (GET_CODE (XEXP (x, 0)) == PLUS) | |
3450 | { | |
3451 | x = apply_distributive_law | |
3452 | (gen_binary (PLUS, mode, | |
3453 | gen_binary (MULT, mode, | |
3454 | XEXP (XEXP (x, 0), 0), XEXP (x, 1)), | |
3455 | gen_binary (MULT, mode, | |
3456 | XEXP (XEXP (x, 0), 1), XEXP (x, 1)))); | |
3457 | ||
3458 | if (GET_CODE (x) != MULT) | |
3459 | goto restart; | |
3460 | } | |
3461 | ||
3462 | /* If this is multiplication by a power of two and its first operand is | |
3463 | a shift, treat the multiply as a shift to allow the shifts to | |
3464 | possibly combine. */ | |
3465 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
3466 | && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0 | |
3467 | && (GET_CODE (XEXP (x, 0)) == ASHIFT | |
3468 | || GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
3469 | || GET_CODE (XEXP (x, 0)) == ASHIFTRT | |
3470 | || GET_CODE (XEXP (x, 0)) == ROTATE | |
3471 | || GET_CODE (XEXP (x, 0)) == ROTATERT)) | |
3472 | { | |
3473 | x = simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0), i); | |
3474 | goto restart; | |
3475 | } | |
3476 | ||
3477 | /* Convert (mult (ashift (const_int 1) A) B) to (ashift B A). */ | |
3478 | if (GET_CODE (XEXP (x, 0)) == ASHIFT | |
3479 | && XEXP (XEXP (x, 0), 0) == const1_rtx) | |
3480 | return gen_rtx_combine (ASHIFT, mode, XEXP (x, 1), | |
3481 | XEXP (XEXP (x, 0), 1)); | |
3482 | break; | |
3483 | ||
3484 | case UDIV: | |
3485 | /* If this is a divide by a power of two, treat it as a shift if | |
3486 | its first operand is a shift. */ | |
3487 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
3488 | && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0 | |
3489 | && (GET_CODE (XEXP (x, 0)) == ASHIFT | |
3490 | || GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
3491 | || GET_CODE (XEXP (x, 0)) == ASHIFTRT | |
3492 | || GET_CODE (XEXP (x, 0)) == ROTATE | |
3493 | || GET_CODE (XEXP (x, 0)) == ROTATERT)) | |
3494 | { | |
3495 | x = simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i); | |
3496 | goto restart; | |
3497 | } | |
3498 | break; | |
3499 | ||
3500 | case EQ: case NE: | |
3501 | case GT: case GTU: case GE: case GEU: | |
3502 | case LT: case LTU: case LE: case LEU: | |
3503 | /* If the first operand is a condition code, we can't do anything | |
3504 | with it. */ | |
3505 | if (GET_CODE (XEXP (x, 0)) == COMPARE | |
3506 | || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC | |
3507 | #ifdef HAVE_cc0 | |
3508 | && XEXP (x, 0) != cc0_rtx | |
3509 | #endif | |
3510 | )) | |
3511 | { | |
3512 | rtx op0 = XEXP (x, 0); | |
3513 | rtx op1 = XEXP (x, 1); | |
3514 | enum rtx_code new_code; | |
3515 | ||
3516 | if (GET_CODE (op0) == COMPARE) | |
3517 | op1 = XEXP (op0, 1), op0 = XEXP (op0, 0); | |
3518 | ||
3519 | /* Simplify our comparison, if possible. */ | |
3520 | new_code = simplify_comparison (code, &op0, &op1); | |
3521 | ||
3522 | #if STORE_FLAG_VALUE == 1 | |
3523 | /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X | |
3524 | if only the low-order bit is possibly nonzero in X (such as when | |
3525 | X is a ZERO_EXTRACT of one bit. Similarly, we can convert | |
3526 | EQ to (xor X 1). Remove any ZERO_EXTRACT we made when thinking | |
3527 | this was a comparison. It may now be simpler to use, e.g., an | |
3528 | AND. If a ZERO_EXTRACT is indeed appropriate, it will | |
3529 | be placed back by the call to make_compound_operation in the | |
3530 | SET case. */ | |
3531 | if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
3532 | && op1 == const0_rtx | |
3533 | && nonzero_bits (op0, GET_MODE (op0)) == 1) | |
3534 | return gen_lowpart_for_combine (mode, | |
3535 | expand_compound_operation (op0)); | |
3536 | else if (new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT | |
3537 | && op1 == const0_rtx | |
3538 | && nonzero_bits (op0, GET_MODE (op0)) == 1) | |
3539 | { | |
3540 | op0 = expand_compound_operation (op0); | |
3541 | ||
3542 | x = gen_rtx_combine (XOR, mode, | |
3543 | gen_lowpart_for_combine (mode, op0), | |
3544 | const1_rtx); | |
3545 | goto restart; | |
3546 | } | |
3547 | #endif | |
3548 | ||
3549 | #if STORE_FLAG_VALUE == -1 | |
3550 | /* If STORE_FLAG_VALUE is -1, we can convert (ne x 0) | |
3551 | to (neg x) if only the low-order bit of X can be nonzero. | |
3552 | This converts (ne (zero_extract X 1 Y) 0) to | |
3553 | (sign_extract X 1 Y). */ | |
3554 | if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
3555 | && op1 == const0_rtx | |
3556 | && nonzero_bits (op0, GET_MODE (op0)) == 1) | |
3557 | { | |
3558 | op0 = expand_compound_operation (op0); | |
3559 | x = gen_rtx_combine (NEG, mode, | |
3560 | gen_lowpart_for_combine (mode, op0)); | |
3561 | goto restart; | |
3562 | } | |
3563 | #endif | |
3564 | ||
3565 | /* If STORE_FLAG_VALUE says to just test the sign bit and X has just | |
3566 | one bit that might be nonzero, we can convert (ne x 0) to | |
3567 | (ashift x c) where C puts the bit in the sign bit. Remove any | |
3568 | AND with STORE_FLAG_VALUE when we are done, since we are only | |
3569 | going to test the sign bit. */ | |
3570 | if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
3571 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
3572 | && (STORE_FLAG_VALUE | |
3573 | == (HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1)) | |
3574 | && op1 == const0_rtx | |
3575 | && mode == GET_MODE (op0) | |
3576 | && (i = exact_log2 (nonzero_bits (op0, GET_MODE (op0)))) >= 0) | |
3577 | { | |
3578 | x = simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
3579 | expand_compound_operation (op0), | |
3580 | GET_MODE_BITSIZE (mode) - 1 - i); | |
3581 | if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx) | |
3582 | return XEXP (x, 0); | |
3583 | else | |
3584 | return x; | |
3585 | } | |
3586 | ||
3587 | /* If the code changed, return a whole new comparison. */ | |
3588 | if (new_code != code) | |
3589 | return gen_rtx_combine (new_code, mode, op0, op1); | |
3590 | ||
3591 | /* Otherwise, keep this operation, but maybe change its operands. | |
3592 | This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */ | |
3593 | SUBST (XEXP (x, 0), op0); | |
3594 | SUBST (XEXP (x, 1), op1); | |
3595 | } | |
3596 | break; | |
3597 | ||
3598 | case IF_THEN_ELSE: | |
3599 | /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register | |
3600 | used in it is being compared against certain values. Get the | |
3601 | true and false comparisons and see if that says anything about the | |
3602 | value of each arm. */ | |
3603 | ||
3604 | if (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3605 | && reversible_comparison_p (XEXP (x, 0)) | |
3606 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == REG) | |
3607 | { | |
3608 | HOST_WIDE_INT nzb; | |
3609 | rtx from = XEXP (XEXP (x, 0), 0); | |
3610 | enum rtx_code true_code = GET_CODE (XEXP (x, 0)); | |
3611 | enum rtx_code false_code = reverse_condition (true_code); | |
3612 | rtx true_val = XEXP (XEXP (x, 0), 1); | |
3613 | rtx false_val = true_val; | |
3614 | rtx true_arm = XEXP (x, 1); | |
3615 | rtx false_arm = XEXP (x, 2); | |
3616 | int swapped = 0; | |
3617 | ||
3618 | /* If FALSE_CODE is EQ, swap the codes and arms. */ | |
3619 | ||
3620 | if (false_code == EQ) | |
3621 | { | |
3622 | swapped = 1, true_code = EQ, false_code = NE; | |
3623 | true_arm = XEXP (x, 2), false_arm = XEXP (x, 1); | |
3624 | } | |
3625 | ||
3626 | /* If we are comparing against zero and the expression being tested | |
3627 | has only a single bit that might be nonzero, that is its value | |
3628 | when it is not equal to zero. Similarly if it is known to be | |
3629 | -1 or 0. */ | |
3630 | ||
3631 | if (true_code == EQ && true_val == const0_rtx | |
3632 | && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0) | |
3633 | false_code = EQ, false_val = GEN_INT (nzb); | |
3634 | else if (true_code == EQ && true_val == const0_rtx | |
3635 | && (num_sign_bit_copies (from, GET_MODE (from)) | |
3636 | == GET_MODE_BITSIZE (GET_MODE (from)))) | |
3637 | false_code = EQ, false_val = constm1_rtx; | |
3638 | ||
3639 | /* Now simplify an arm if we know the value of the register | |
3640 | in the branch and it is used in the arm. Be carefull due to | |
3641 | the potential of locally-shared RTL. */ | |
3642 | ||
3643 | if (reg_mentioned_p (from, true_arm)) | |
3644 | true_arm = subst (known_cond (copy_rtx (true_arm), true_code, | |
3645 | from, true_val), | |
3646 | pc_rtx, pc_rtx, 0, 0); | |
3647 | if (reg_mentioned_p (from, false_arm)) | |
3648 | false_arm = subst (known_cond (copy_rtx (false_arm), false_code, | |
3649 | from, false_val), | |
3650 | pc_rtx, pc_rtx, 0, 0); | |
3651 | ||
3652 | SUBST (XEXP (x, 1), swapped ? false_arm : true_arm); | |
3653 | SUBST (XEXP (x, 2), swapped ? true_arm : false_arm); | |
3654 | } | |
3655 | ||
3656 | /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be | |
3657 | reversed, do so to avoid needing two sets of patterns for | |
3658 | subtract-and-branch insns. Similarly if we have a constant in that | |
3659 | position or if the third operand is the same as the first operand | |
3660 | of the comparison. */ | |
3661 | ||
3662 | if (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3663 | && reversible_comparison_p (XEXP (x, 0)) | |
3664 | && (XEXP (x, 1) == pc_rtx || GET_CODE (XEXP (x, 1)) == CONST_INT | |
3665 | || rtx_equal_p (XEXP (x, 2), XEXP (XEXP (x, 0), 0)))) | |
3666 | { | |
3667 | SUBST (XEXP (x, 0), | |
3668 | gen_binary (reverse_condition (GET_CODE (XEXP (x, 0))), | |
3669 | GET_MODE (XEXP (x, 0)), | |
3670 | XEXP (XEXP (x, 0), 0), XEXP (XEXP (x, 0), 1))); | |
3671 | ||
3672 | temp = XEXP (x, 1); | |
3673 | SUBST (XEXP (x, 1), XEXP (x, 2)); | |
3674 | SUBST (XEXP (x, 2), temp); | |
3675 | } | |
3676 | ||
3677 | /* If the two arms are identical, we don't need the comparison. */ | |
3678 | ||
3679 | if (rtx_equal_p (XEXP (x, 1), XEXP (x, 2)) | |
3680 | && ! side_effects_p (XEXP (x, 0))) | |
3681 | return XEXP (x, 1); | |
3682 | ||
3683 | /* Look for cases where we have (abs x) or (neg (abs X)). */ | |
3684 | ||
3685 | if (GET_MODE_CLASS (mode) == MODE_INT | |
3686 | && GET_CODE (XEXP (x, 2)) == NEG | |
3687 | && rtx_equal_p (XEXP (x, 1), XEXP (XEXP (x, 2), 0)) | |
3688 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3689 | && rtx_equal_p (XEXP (x, 1), XEXP (XEXP (x, 0), 0)) | |
3690 | && ! side_effects_p (XEXP (x, 1))) | |
3691 | switch (GET_CODE (XEXP (x, 0))) | |
3692 | { | |
3693 | case GT: | |
3694 | case GE: | |
3695 | x = gen_unary (ABS, mode, XEXP (x, 1)); | |
3696 | goto restart; | |
3697 | case LT: | |
3698 | case LE: | |
3699 | x = gen_unary (NEG, mode, gen_unary (ABS, mode, XEXP (x, 1))); | |
3700 | goto restart; | |
3701 | } | |
3702 | ||
3703 | /* Look for MIN or MAX. */ | |
3704 | ||
3705 | if (GET_MODE_CLASS (mode) == MODE_INT | |
3706 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3707 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)) | |
3708 | && rtx_equal_p (XEXP (XEXP (x, 0), 1), XEXP (x, 2)) | |
3709 | && ! side_effects_p (XEXP (x, 0))) | |
3710 | switch (GET_CODE (XEXP (x, 0))) | |
3711 | { | |
3712 | case GE: | |
3713 | case GT: | |
3714 | x = gen_binary (SMAX, mode, XEXP (x, 1), XEXP (x, 2)); | |
3715 | goto restart; | |
3716 | case LE: | |
3717 | case LT: | |
3718 | x = gen_binary (SMIN, mode, XEXP (x, 1), XEXP (x, 2)); | |
3719 | goto restart; | |
3720 | case GEU: | |
3721 | case GTU: | |
3722 | x = gen_binary (UMAX, mode, XEXP (x, 1), XEXP (x, 2)); | |
3723 | goto restart; | |
3724 | case LEU: | |
3725 | case LTU: | |
3726 | x = gen_binary (UMIN, mode, XEXP (x, 1), XEXP (x, 2)); | |
3727 | goto restart; | |
3728 | } | |
3729 | ||
3730 | /* If we have something like (if_then_else (ne A 0) (OP X C) X), | |
3731 | A is known to be either 0 or 1, and OP is an identity when its | |
3732 | second operand is zero, this can be done as (OP X (mult A C)). | |
3733 | Similarly if A is known to be 0 or -1 and also similarly if we have | |
3734 | a ZERO_EXTEND or SIGN_EXTEND as long as X is already extended (so | |
3735 | we don't destroy it). */ | |
3736 | ||
3737 | if (mode != VOIDmode | |
3738 | && (GET_CODE (XEXP (x, 0)) == EQ || GET_CODE (XEXP (x, 0)) == NE) | |
3739 | && XEXP (XEXP (x, 0), 1) == const0_rtx | |
3740 | && (nonzero_bits (XEXP (XEXP (x, 0), 0), mode) == 1 | |
3741 | || (num_sign_bit_copies (XEXP (XEXP (x, 0), 0), mode) | |
3742 | == GET_MODE_BITSIZE (mode)))) | |
3743 | { | |
3744 | rtx nz = make_compound_operation (GET_CODE (XEXP (x, 0)) == NE | |
3745 | ? XEXP (x, 1) : XEXP (x, 2)); | |
3746 | rtx z = GET_CODE (XEXP (x, 0)) == NE ? XEXP (x, 2) : XEXP (x, 1); | |
3747 | rtx dir = (nonzero_bits (XEXP (XEXP (x, 0), 0), mode) == 1 | |
3748 | ? const1_rtx : constm1_rtx); | |
3749 | rtx c = 0; | |
3750 | enum machine_mode m = mode; | |
3751 | enum rtx_code op, extend_op = 0; | |
3752 | ||
3753 | if ((GET_CODE (nz) == PLUS || GET_CODE (nz) == MINUS | |
3754 | || GET_CODE (nz) == IOR || GET_CODE (nz) == XOR | |
3755 | || GET_CODE (nz) == ASHIFT | |
3756 | || GET_CODE (nz) == LSHIFTRT || GET_CODE (nz) == ASHIFTRT) | |
3757 | && rtx_equal_p (XEXP (nz, 0), z)) | |
3758 | c = XEXP (nz, 1), op = GET_CODE (nz); | |
3759 | else if (GET_CODE (nz) == SIGN_EXTEND | |
3760 | && (GET_CODE (XEXP (nz, 0)) == PLUS | |
3761 | || GET_CODE (XEXP (nz, 0)) == MINUS | |
3762 | || GET_CODE (XEXP (nz, 0)) == IOR | |
3763 | || GET_CODE (XEXP (nz, 0)) == XOR | |
3764 | || GET_CODE (XEXP (nz, 0)) == ASHIFT | |
3765 | || GET_CODE (XEXP (nz, 0)) == LSHIFTRT | |
3766 | || GET_CODE (XEXP (nz, 0)) == ASHIFTRT) | |
3767 | && GET_CODE (XEXP (XEXP (nz, 0), 0)) == SUBREG | |
3768 | && subreg_lowpart_p (XEXP (XEXP (nz, 0), 0)) | |
3769 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (nz, 0), 0)), z) | |
3770 | && (num_sign_bit_copies (z, GET_MODE (z)) | |
3771 | >= (GET_MODE_BITSIZE (mode) | |
3772 | - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (nz, 0), 0)))))) | |
3773 | { | |
3774 | c = XEXP (XEXP (nz, 0), 1); | |
3775 | op = GET_CODE (XEXP (nz, 0)); | |
3776 | extend_op = SIGN_EXTEND; | |
3777 | m = GET_MODE (XEXP (nz, 0)); | |
3778 | } | |
3779 | else if (GET_CODE (nz) == ZERO_EXTEND | |
3780 | && (GET_CODE (XEXP (nz, 0)) == PLUS | |
3781 | || GET_CODE (XEXP (nz, 0)) == MINUS | |
3782 | || GET_CODE (XEXP (nz, 0)) == IOR | |
3783 | || GET_CODE (XEXP (nz, 0)) == XOR | |
3784 | || GET_CODE (XEXP (nz, 0)) == ASHIFT | |
3785 | || GET_CODE (XEXP (nz, 0)) == LSHIFTRT | |
3786 | || GET_CODE (XEXP (nz, 0)) == ASHIFTRT) | |
3787 | && GET_CODE (XEXP (XEXP (nz, 0), 0)) == SUBREG | |
3788 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
3789 | && subreg_lowpart_p (XEXP (XEXP (nz, 0), 0)) | |
3790 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (nz, 0), 0)), z) | |
3791 | && ((nonzero_bits (z, GET_MODE (z)) | |
3792 | & ~ GET_MODE_MASK (GET_MODE (XEXP (XEXP (nz, 0), 0)))) | |
3793 | == 0)) | |
3794 | { | |
3795 | c = XEXP (XEXP (nz, 0), 1); | |
3796 | op = GET_CODE (XEXP (nz, 0)); | |
3797 | extend_op = ZERO_EXTEND; | |
3798 | m = GET_MODE (XEXP (nz, 0)); | |
3799 | } | |
3800 | ||
3801 | if (c && ! side_effects_p (c) && ! side_effects_p (z)) | |
3802 | { | |
3803 | temp | |
3804 | = gen_binary (MULT, m, | |
3805 | gen_lowpart_for_combine (m, | |
3806 | XEXP (XEXP (x, 0), 0)), | |
3807 | gen_binary (MULT, m, c, dir)); | |
3808 | ||
3809 | temp = gen_binary (op, m, gen_lowpart_for_combine (m, z), temp); | |
3810 | ||
3811 | if (extend_op != 0) | |
3812 | temp = gen_unary (extend_op, mode, temp); | |
3813 | ||
3814 | return temp; | |
3815 | } | |
3816 | } | |
3817 | break; | |
3818 | ||
3819 | case ZERO_EXTRACT: | |
3820 | case SIGN_EXTRACT: | |
3821 | case ZERO_EXTEND: | |
3822 | case SIGN_EXTEND: | |
3823 | /* If we are processing SET_DEST, we are done. */ | |
3824 | if (in_dest) | |
3825 | return x; | |
3826 | ||
3827 | x = expand_compound_operation (x); | |
3828 | if (GET_CODE (x) != code) | |
3829 | goto restart; | |
3830 | break; | |
3831 | ||
3832 | case SET: | |
3833 | /* (set (pc) (return)) gets written as (return). */ | |
3834 | if (GET_CODE (SET_DEST (x)) == PC && GET_CODE (SET_SRC (x)) == RETURN) | |
3835 | return SET_SRC (x); | |
3836 | ||
3837 | /* Convert this into a field assignment operation, if possible. */ | |
3838 | x = make_field_assignment (x); | |
3839 | ||
3840 | /* If we are setting CC0 or if the source is a COMPARE, look for the | |
3841 | use of the comparison result and try to simplify it unless we already | |
3842 | have used undobuf.other_insn. */ | |
3843 | if ((GET_CODE (SET_SRC (x)) == COMPARE | |
3844 | #ifdef HAVE_cc0 | |
3845 | || SET_DEST (x) == cc0_rtx | |
3846 | #endif | |
3847 | ) | |
3848 | && (cc_use = find_single_use (SET_DEST (x), subst_insn, | |
3849 | &other_insn)) != 0 | |
3850 | && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn) | |
3851 | && GET_RTX_CLASS (GET_CODE (*cc_use)) == '<' | |
3852 | && XEXP (*cc_use, 0) == SET_DEST (x)) | |
3853 | { | |
3854 | enum rtx_code old_code = GET_CODE (*cc_use); | |
3855 | enum rtx_code new_code; | |
3856 | rtx op0, op1; | |
3857 | int other_changed = 0; | |
3858 | enum machine_mode compare_mode = GET_MODE (SET_DEST (x)); | |
3859 | ||
3860 | if (GET_CODE (SET_SRC (x)) == COMPARE) | |
3861 | op0 = XEXP (SET_SRC (x), 0), op1 = XEXP (SET_SRC (x), 1); | |
3862 | else | |
3863 | op0 = SET_SRC (x), op1 = const0_rtx; | |
3864 | ||
3865 | /* Simplify our comparison, if possible. */ | |
3866 | new_code = simplify_comparison (old_code, &op0, &op1); | |
3867 | ||
3868 | #ifdef EXTRA_CC_MODES | |
3869 | /* If this machine has CC modes other than CCmode, check to see | |
3870 | if we need to use a different CC mode here. */ | |
3871 | compare_mode = SELECT_CC_MODE (new_code, op0, op1); | |
3872 | #endif /* EXTRA_CC_MODES */ | |
3873 | ||
3874 | #if !defined (HAVE_cc0) && defined (EXTRA_CC_MODES) | |
3875 | /* If the mode changed, we have to change SET_DEST, the mode | |
3876 | in the compare, and the mode in the place SET_DEST is used. | |
3877 | If SET_DEST is a hard register, just build new versions with | |
3878 | the proper mode. If it is a pseudo, we lose unless it is only | |
3879 | time we set the pseudo, in which case we can safely change | |
3880 | its mode. */ | |
3881 | if (compare_mode != GET_MODE (SET_DEST (x))) | |
3882 | { | |
3883 | int regno = REGNO (SET_DEST (x)); | |
3884 | rtx new_dest = gen_rtx (REG, compare_mode, regno); | |
3885 | ||
3886 | if (regno < FIRST_PSEUDO_REGISTER | |
3887 | || (reg_n_sets[regno] == 1 | |
3888 | && ! REG_USERVAR_P (SET_DEST (x)))) | |
3889 | { | |
3890 | if (regno >= FIRST_PSEUDO_REGISTER) | |
3891 | SUBST (regno_reg_rtx[regno], new_dest); | |
3892 | ||
3893 | SUBST (SET_DEST (x), new_dest); | |
3894 | SUBST (XEXP (*cc_use, 0), new_dest); | |
3895 | other_changed = 1; | |
3896 | } | |
3897 | } | |
3898 | #endif | |
3899 | ||
3900 | /* If the code changed, we have to build a new comparison | |
3901 | in undobuf.other_insn. */ | |
3902 | if (new_code != old_code) | |
3903 | { | |
3904 | unsigned HOST_WIDE_INT mask; | |
3905 | ||
3906 | SUBST (*cc_use, gen_rtx_combine (new_code, GET_MODE (*cc_use), | |
3907 | SET_DEST (x), const0_rtx)); | |
3908 | ||
3909 | /* If the only change we made was to change an EQ into an | |
3910 | NE or vice versa, OP0 has only one bit that might be nonzero, | |
3911 | and OP1 is zero, check if changing the user of the condition | |
3912 | code will produce a valid insn. If it won't, we can keep | |
3913 | the original code in that insn by surrounding our operation | |
3914 | with an XOR. */ | |
3915 | ||
3916 | if (((old_code == NE && new_code == EQ) | |
3917 | || (old_code == EQ && new_code == NE)) | |
3918 | && ! other_changed && op1 == const0_rtx | |
3919 | && (GET_MODE_BITSIZE (GET_MODE (op0)) | |
3920 | <= HOST_BITS_PER_WIDE_INT) | |
3921 | && (exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) | |
3922 | >= 0)) | |
3923 | { | |
3924 | rtx pat = PATTERN (other_insn), note = 0; | |
3925 | ||
3926 | if ((recog_for_combine (&pat, other_insn, ¬e) < 0 | |
3927 | && ! check_asm_operands (pat))) | |
3928 | { | |
3929 | PUT_CODE (*cc_use, old_code); | |
3930 | other_insn = 0; | |
3931 | ||
3932 | op0 = gen_binary (XOR, GET_MODE (op0), op0, | |
3933 | GEN_INT (mask)); | |
3934 | } | |
3935 | } | |
3936 | ||
3937 | other_changed = 1; | |
3938 | } | |
3939 | ||
3940 | if (other_changed) | |
3941 | undobuf.other_insn = other_insn; | |
3942 | ||
3943 | #ifdef HAVE_cc0 | |
3944 | /* If we are now comparing against zero, change our source if | |
3945 | needed. If we do not use cc0, we always have a COMPARE. */ | |
3946 | if (op1 == const0_rtx && SET_DEST (x) == cc0_rtx) | |
3947 | SUBST (SET_SRC (x), op0); | |
3948 | else | |
3949 | #endif | |
3950 | ||
3951 | /* Otherwise, if we didn't previously have a COMPARE in the | |
3952 | correct mode, we need one. */ | |
3953 | if (GET_CODE (SET_SRC (x)) != COMPARE | |
3954 | || GET_MODE (SET_SRC (x)) != compare_mode) | |
3955 | SUBST (SET_SRC (x), gen_rtx_combine (COMPARE, compare_mode, | |
3956 | op0, op1)); | |
3957 | else | |
3958 | { | |
3959 | /* Otherwise, update the COMPARE if needed. */ | |
3960 | SUBST (XEXP (SET_SRC (x), 0), op0); | |
3961 | SUBST (XEXP (SET_SRC (x), 1), op1); | |
3962 | } | |
3963 | } | |
3964 | else | |
3965 | { | |
3966 | /* Get SET_SRC in a form where we have placed back any | |
3967 | compound expressions. Then do the checks below. */ | |
3968 | temp = make_compound_operation (SET_SRC (x), SET); | |
3969 | SUBST (SET_SRC (x), temp); | |
3970 | } | |
3971 | ||
3972 | /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some | |
3973 | operation, and X being a REG or (subreg (reg)), we may be able to | |
3974 | convert this to (set (subreg:m2 x) (op)). | |
3975 | ||
3976 | We can always do this if M1 is narrower than M2 because that | |
3977 | means that we only care about the low bits of the result. | |
3978 | ||
3979 | However, on most machines (those with neither BYTE_LOADS_ZERO_EXTEND | |
3980 | nor BYTES_LOADS_SIGN_EXTEND defined), we cannot perform a | |
3981 | narrower operation that requested since the high-order bits will | |
3982 | be undefined. On machine where BYTE_LOADS_*_EXTEND is defined, | |
3983 | however, this transformation is safe as long as M1 and M2 have | |
3984 | the same number of words. */ | |
3985 | ||
3986 | if (GET_CODE (SET_SRC (x)) == SUBREG | |
3987 | && subreg_lowpart_p (SET_SRC (x)) | |
3988 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (SET_SRC (x)))) != 'o' | |
3989 | && (((GET_MODE_SIZE (GET_MODE (SET_SRC (x))) + (UNITS_PER_WORD - 1)) | |
3990 | / UNITS_PER_WORD) | |
3991 | == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_SRC (x)))) | |
3992 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)) | |
3993 | #ifndef BYTE_LOADS_EXTEND | |
3994 | && (GET_MODE_SIZE (GET_MODE (SET_SRC (x))) | |
3995 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_SRC (x))))) | |
3996 | #endif | |
3997 | && (GET_CODE (SET_DEST (x)) == REG | |
3998 | || (GET_CODE (SET_DEST (x)) == SUBREG | |
3999 | && GET_CODE (SUBREG_REG (SET_DEST (x))) == REG))) | |
4000 | { | |
4001 | SUBST (SET_DEST (x), | |
4002 | gen_lowpart_for_combine (GET_MODE (SUBREG_REG (SET_SRC (x))), | |
4003 | SET_DEST (x))); | |
4004 | SUBST (SET_SRC (x), SUBREG_REG (SET_SRC (x))); | |
4005 | } | |
4006 | ||
4007 | #ifdef BYTE_LOADS_EXTEND | |
4008 | /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with | |
4009 | M wider than N, this would require a paradoxical subreg. | |
4010 | Replace the subreg with a zero_extend to avoid the reload that | |
4011 | would otherwise be required. */ | |
4012 | ||
4013 | if (GET_CODE (SET_SRC (x)) == SUBREG | |
4014 | && subreg_lowpart_p (SET_SRC (x)) | |
4015 | && SUBREG_WORD (SET_SRC (x)) == 0 | |
4016 | && (GET_MODE_SIZE (GET_MODE (SET_SRC (x))) | |
4017 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_SRC (x))))) | |
4018 | && GET_CODE (SUBREG_REG (SET_SRC (x))) == MEM) | |
4019 | SUBST (SET_SRC (x), gen_rtx_combine (LOAD_EXTEND, | |
4020 | GET_MODE (SET_SRC (x)), | |
4021 | XEXP (SET_SRC (x), 0))); | |
4022 | #endif | |
4023 | ||
4024 | #ifndef HAVE_conditional_move | |
4025 | ||
4026 | /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, | |
4027 | and we are comparing an item known to be 0 or -1 against 0, use a | |
4028 | logical operation instead. Check for one of the arms being an IOR | |
4029 | of the other arm with some value. We compute three terms to be | |
4030 | IOR'ed together. In practice, at most two will be nonzero. Then | |
4031 | we do the IOR's. */ | |
4032 | ||
4033 | if (GET_CODE (SET_DEST (x)) != PC | |
4034 | && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE | |
4035 | && (GET_CODE (XEXP (SET_SRC (x), 0)) == EQ | |
4036 | || GET_CODE (XEXP (SET_SRC (x), 0)) == NE) | |
4037 | && XEXP (XEXP (SET_SRC (x), 0), 1) == const0_rtx | |
4038 | && (num_sign_bit_copies (XEXP (XEXP (SET_SRC (x), 0), 0), | |
4039 | GET_MODE (XEXP (XEXP (SET_SRC (x), 0), 0))) | |
4040 | == GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (SET_SRC (x), 0), 0)))) | |
4041 | && ! side_effects_p (SET_SRC (x))) | |
4042 | { | |
4043 | rtx true = (GET_CODE (XEXP (SET_SRC (x), 0)) == NE | |
4044 | ? XEXP (SET_SRC (x), 1) : XEXP (SET_SRC (x), 2)); | |
4045 | rtx false = (GET_CODE (XEXP (SET_SRC (x), 0)) == NE | |
4046 | ? XEXP (SET_SRC (x), 2) : XEXP (SET_SRC (x), 1)); | |
4047 | rtx term1 = const0_rtx, term2, term3; | |
4048 | ||
4049 | if (GET_CODE (true) == IOR && rtx_equal_p (XEXP (true, 0), false)) | |
4050 | term1 = false, true = XEXP (true, 1), false = const0_rtx; | |
4051 | else if (GET_CODE (true) == IOR | |
4052 | && rtx_equal_p (XEXP (true, 1), false)) | |
4053 | term1 = false, true = XEXP (true, 0), false = const0_rtx; | |
4054 | else if (GET_CODE (false) == IOR | |
4055 | && rtx_equal_p (XEXP (false, 0), true)) | |
4056 | term1 = true, false = XEXP (false, 1), true = const0_rtx; | |
4057 | else if (GET_CODE (false) == IOR | |
4058 | && rtx_equal_p (XEXP (false, 1), true)) | |
4059 | term1 = true, false = XEXP (false, 0), true = const0_rtx; | |
4060 | ||
4061 | term2 = gen_binary (AND, GET_MODE (SET_SRC (x)), | |
4062 | XEXP (XEXP (SET_SRC (x), 0), 0), true); | |
4063 | term3 = gen_binary (AND, GET_MODE (SET_SRC (x)), | |
4064 | gen_unary (NOT, GET_MODE (SET_SRC (x)), | |
4065 | XEXP (XEXP (SET_SRC (x), 0), 0)), | |
4066 | false); | |
4067 | ||
4068 | SUBST (SET_SRC (x), | |
4069 | gen_binary (IOR, GET_MODE (SET_SRC (x)), | |
4070 | gen_binary (IOR, GET_MODE (SET_SRC (x)), | |
4071 | term1, term2), | |
4072 | term3)); | |
4073 | } | |
4074 | #endif | |
4075 | break; | |
4076 | ||
4077 | case AND: | |
4078 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4079 | { | |
4080 | x = simplify_and_const_int (x, mode, XEXP (x, 0), | |
4081 | INTVAL (XEXP (x, 1))); | |
4082 | ||
4083 | /* If we have (ior (and (X C1) C2)) and the next restart would be | |
4084 | the last, simplify this by making C1 as small as possible | |
4085 | and then exit. */ | |
4086 | if (n_restarts >= 3 && GET_CODE (x) == IOR | |
4087 | && GET_CODE (XEXP (x, 0)) == AND | |
4088 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
4089 | && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4090 | { | |
4091 | temp = gen_binary (AND, mode, XEXP (XEXP (x, 0), 0), | |
4092 | GEN_INT (INTVAL (XEXP (XEXP (x, 0), 1)) | |
4093 | & ~ INTVAL (XEXP (x, 1)))); | |
4094 | return gen_binary (IOR, mode, temp, XEXP (x, 1)); | |
4095 | } | |
4096 | ||
4097 | if (GET_CODE (x) != AND) | |
4098 | goto restart; | |
4099 | } | |
4100 | ||
4101 | /* Convert (A | B) & A to A. */ | |
4102 | if (GET_CODE (XEXP (x, 0)) == IOR | |
4103 | && (rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)) | |
4104 | || rtx_equal_p (XEXP (XEXP (x, 0), 1), XEXP (x, 1))) | |
4105 | && ! side_effects_p (XEXP (XEXP (x, 0), 0)) | |
4106 | && ! side_effects_p (XEXP (XEXP (x, 0), 1))) | |
4107 | return XEXP (x, 1); | |
4108 | ||
4109 | /* Convert (A ^ B) & A to A & (~ B) since the latter is often a single | |
4110 | insn (and may simplify more). */ | |
4111 | else if (GET_CODE (XEXP (x, 0)) == XOR | |
4112 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)) | |
4113 | && ! side_effects_p (XEXP (x, 1))) | |
4114 | { | |
4115 | x = gen_binary (AND, mode, | |
4116 | gen_unary (NOT, mode, XEXP (XEXP (x, 0), 1)), | |
4117 | XEXP (x, 1)); | |
4118 | goto restart; | |
4119 | } | |
4120 | else if (GET_CODE (XEXP (x, 0)) == XOR | |
4121 | && rtx_equal_p (XEXP (XEXP (x, 0), 1), XEXP (x, 1)) | |
4122 | && ! side_effects_p (XEXP (x, 1))) | |
4123 | { | |
4124 | x = gen_binary (AND, mode, | |
4125 | gen_unary (NOT, mode, XEXP (XEXP (x, 0), 0)), | |
4126 | XEXP (x, 1)); | |
4127 | goto restart; | |
4128 | } | |
4129 | ||
4130 | /* Similarly for (~ (A ^ B)) & A. */ | |
4131 | else if (GET_CODE (XEXP (x, 0)) == NOT | |
4132 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == XOR | |
4133 | && rtx_equal_p (XEXP (XEXP (XEXP (x, 0), 0), 0), XEXP (x, 1)) | |
4134 | && ! side_effects_p (XEXP (x, 1))) | |
4135 | { | |
4136 | x = gen_binary (AND, mode, XEXP (XEXP (XEXP (x, 0), 0), 1), | |
4137 | XEXP (x, 1)); | |
4138 | goto restart; | |
4139 | } | |
4140 | else if (GET_CODE (XEXP (x, 0)) == NOT | |
4141 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == XOR | |
4142 | && rtx_equal_p (XEXP (XEXP (XEXP (x, 0), 0), 1), XEXP (x, 1)) | |
4143 | && ! side_effects_p (XEXP (x, 1))) | |
4144 | { | |
4145 | x = gen_binary (AND, mode, XEXP (XEXP (XEXP (x, 0), 0), 0), | |
4146 | XEXP (x, 1)); | |
4147 | goto restart; | |
4148 | } | |
4149 | ||
4150 | /* If we have (and A B) with A not an object but that is known to | |
4151 | be -1 or 0, this is equivalent to the expression | |
4152 | (if_then_else (ne A (const_int 0)) B (const_int 0)) | |
4153 | We make this conversion because it may allow further | |
4154 | simplifications and then allow use of conditional move insns. | |
4155 | If the machine doesn't have condition moves, code in case SET | |
4156 | will convert the IF_THEN_ELSE back to the logical operation. | |
4157 | We build the IF_THEN_ELSE here in case further simplification | |
4158 | is possible (e.g., we can convert it to ABS). */ | |
4159 | ||
4160 | if (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) != 'o' | |
4161 | && ! (GET_CODE (XEXP (x, 0)) == SUBREG | |
4162 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 0)))) == 'o') | |
4163 | && (num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))) | |
4164 | == GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))) | |
4165 | { | |
4166 | rtx op0 = XEXP (x, 0); | |
4167 | rtx op1 = const0_rtx; | |
4168 | enum rtx_code comp_code | |
4169 | = simplify_comparison (NE, &op0, &op1); | |
4170 | ||
4171 | x = gen_rtx_combine (IF_THEN_ELSE, mode, | |
4172 | gen_binary (comp_code, VOIDmode, op0, op1), | |
4173 | XEXP (x, 1), const0_rtx); | |
4174 | goto restart; | |
4175 | } | |
4176 | ||
4177 | /* In the following group of tests (and those in case IOR below), | |
4178 | we start with some combination of logical operations and apply | |
4179 | the distributive law followed by the inverse distributive law. | |
4180 | Most of the time, this results in no change. However, if some of | |
4181 | the operands are the same or inverses of each other, simplifications | |
4182 | will result. | |
4183 | ||
4184 | For example, (and (ior A B) (not B)) can occur as the result of | |
4185 | expanding a bit field assignment. When we apply the distributive | |
4186 | law to this, we get (ior (and (A (not B))) (and (B (not B)))), | |
4187 | which then simplifies to (and (A (not B))). */ | |
4188 | ||
4189 | /* If we have (and (ior A B) C), apply the distributive law and then | |
4190 | the inverse distributive law to see if things simplify. */ | |
4191 | ||
4192 | if (GET_CODE (XEXP (x, 0)) == IOR || GET_CODE (XEXP (x, 0)) == XOR) | |
4193 | { | |
4194 | x = apply_distributive_law | |
4195 | (gen_binary (GET_CODE (XEXP (x, 0)), mode, | |
4196 | gen_binary (AND, mode, | |
4197 | XEXP (XEXP (x, 0), 0), XEXP (x, 1)), | |
4198 | gen_binary (AND, mode, | |
4199 | XEXP (XEXP (x, 0), 1), XEXP (x, 1)))); | |
4200 | if (GET_CODE (x) != AND) | |
4201 | goto restart; | |
4202 | } | |
4203 | ||
4204 | if (GET_CODE (XEXP (x, 1)) == IOR || GET_CODE (XEXP (x, 1)) == XOR) | |
4205 | { | |
4206 | x = apply_distributive_law | |
4207 | (gen_binary (GET_CODE (XEXP (x, 1)), mode, | |
4208 | gen_binary (AND, mode, | |
4209 | XEXP (XEXP (x, 1), 0), XEXP (x, 0)), | |
4210 | gen_binary (AND, mode, | |
4211 | XEXP (XEXP (x, 1), 1), XEXP (x, 0)))); | |
4212 | if (GET_CODE (x) != AND) | |
4213 | goto restart; | |
4214 | } | |
4215 | ||
4216 | /* Similarly, taking advantage of the fact that | |
4217 | (and (not A) (xor B C)) == (xor (ior A B) (ior A C)) */ | |
4218 | ||
4219 | if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == XOR) | |
4220 | { | |
4221 | x = apply_distributive_law | |
4222 | (gen_binary (XOR, mode, | |
4223 | gen_binary (IOR, mode, XEXP (XEXP (x, 0), 0), | |
4224 | XEXP (XEXP (x, 1), 0)), | |
4225 | gen_binary (IOR, mode, XEXP (XEXP (x, 0), 0), | |
4226 | XEXP (XEXP (x, 1), 1)))); | |
4227 | if (GET_CODE (x) != AND) | |
4228 | goto restart; | |
4229 | } | |
4230 | ||
4231 | else if (GET_CODE (XEXP (x, 1)) == NOT && GET_CODE (XEXP (x, 0)) == XOR) | |
4232 | { | |
4233 | x = apply_distributive_law | |
4234 | (gen_binary (XOR, mode, | |
4235 | gen_binary (IOR, mode, XEXP (XEXP (x, 1), 0), | |
4236 | XEXP (XEXP (x, 0), 0)), | |
4237 | gen_binary (IOR, mode, XEXP (XEXP (x, 1), 0), | |
4238 | XEXP (XEXP (x, 0), 1)))); | |
4239 | if (GET_CODE (x) != AND) | |
4240 | goto restart; | |
4241 | } | |
4242 | break; | |
4243 | ||
4244 | case IOR: | |
4245 | /* (ior A C) is C if all bits of A that might be nonzero are on in C. */ | |
4246 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
4247 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
4248 | && (nonzero_bits (XEXP (x, 0), mode) & ~ INTVAL (XEXP (x, 1))) == 0) | |
4249 | return XEXP (x, 1); | |
4250 | ||
4251 | /* Convert (A & B) | A to A. */ | |
4252 | if (GET_CODE (XEXP (x, 0)) == AND | |
4253 | && (rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)) | |
4254 | || rtx_equal_p (XEXP (XEXP (x, 0), 1), XEXP (x, 1))) | |
4255 | && ! side_effects_p (XEXP (XEXP (x, 0), 0)) | |
4256 | && ! side_effects_p (XEXP (XEXP (x, 0), 1))) | |
4257 | return XEXP (x, 1); | |
4258 | ||
4259 | /* If we have (ior (and A B) C), apply the distributive law and then | |
4260 | the inverse distributive law to see if things simplify. */ | |
4261 | ||
4262 | if (GET_CODE (XEXP (x, 0)) == AND) | |
4263 | { | |
4264 | x = apply_distributive_law | |
4265 | (gen_binary (AND, mode, | |
4266 | gen_binary (IOR, mode, | |
4267 | XEXP (XEXP (x, 0), 0), XEXP (x, 1)), | |
4268 | gen_binary (IOR, mode, | |
4269 | XEXP (XEXP (x, 0), 1), XEXP (x, 1)))); | |
4270 | ||
4271 | if (GET_CODE (x) != IOR) | |
4272 | goto restart; | |
4273 | } | |
4274 | ||
4275 | if (GET_CODE (XEXP (x, 1)) == AND) | |
4276 | { | |
4277 | x = apply_distributive_law | |
4278 | (gen_binary (AND, mode, | |
4279 | gen_binary (IOR, mode, | |
4280 | XEXP (XEXP (x, 1), 0), XEXP (x, 0)), | |
4281 | gen_binary (IOR, mode, | |
4282 | XEXP (XEXP (x, 1), 1), XEXP (x, 0)))); | |
4283 | ||
4284 | if (GET_CODE (x) != IOR) | |
4285 | goto restart; | |
4286 | } | |
4287 | ||
4288 | /* Convert (ior (ashift A CX) (lshiftrt A CY)) where CX+CY equals the | |
4289 | mode size to (rotate A CX). */ | |
4290 | ||
4291 | if (((GET_CODE (XEXP (x, 0)) == ASHIFT | |
4292 | && GET_CODE (XEXP (x, 1)) == LSHIFTRT) | |
4293 | || (GET_CODE (XEXP (x, 1)) == ASHIFT | |
4294 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT)) | |
4295 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (XEXP (x, 1), 0)) | |
4296 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
4297 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT | |
4298 | && (INTVAL (XEXP (XEXP (x, 0), 1)) + INTVAL (XEXP (XEXP (x, 1), 1)) | |
4299 | == GET_MODE_BITSIZE (mode))) | |
4300 | { | |
4301 | rtx shift_count; | |
4302 | ||
4303 | if (GET_CODE (XEXP (x, 0)) == ASHIFT) | |
4304 | shift_count = XEXP (XEXP (x, 0), 1); | |
4305 | else | |
4306 | shift_count = XEXP (XEXP (x, 1), 1); | |
4307 | x = gen_rtx (ROTATE, mode, XEXP (XEXP (x, 0), 0), shift_count); | |
4308 | goto restart; | |
4309 | } | |
4310 | break; | |
4311 | ||
4312 | case XOR: | |
4313 | /* Convert (XOR (NOT x) (NOT y)) to (XOR x y). | |
4314 | Also convert (XOR (NOT x) y) to (NOT (XOR x y)), similarly for | |
4315 | (NOT y). */ | |
4316 | { | |
4317 | int num_negated = 0; | |
4318 | rtx in1 = XEXP (x, 0), in2 = XEXP (x, 1); | |
4319 | ||
4320 | if (GET_CODE (in1) == NOT) | |
4321 | num_negated++, in1 = XEXP (in1, 0); | |
4322 | if (GET_CODE (in2) == NOT) | |
4323 | num_negated++, in2 = XEXP (in2, 0); | |
4324 | ||
4325 | if (num_negated == 2) | |
4326 | { | |
4327 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
4328 | SUBST (XEXP (x, 1), XEXP (XEXP (x, 1), 0)); | |
4329 | } | |
4330 | else if (num_negated == 1) | |
4331 | { | |
4332 | x = gen_unary (NOT, mode, | |
4333 | gen_binary (XOR, mode, in1, in2)); | |
4334 | goto restart; | |
4335 | } | |
4336 | } | |
4337 | ||
4338 | /* Convert (xor (and A B) B) to (and (not A) B). The latter may | |
4339 | correspond to a machine insn or result in further simplifications | |
4340 | if B is a constant. */ | |
4341 | ||
4342 | if (GET_CODE (XEXP (x, 0)) == AND | |
4343 | && rtx_equal_p (XEXP (XEXP (x, 0), 1), XEXP (x, 1)) | |
4344 | && ! side_effects_p (XEXP (x, 1))) | |
4345 | { | |
4346 | x = gen_binary (AND, mode, | |
4347 | gen_unary (NOT, mode, XEXP (XEXP (x, 0), 0)), | |
4348 | XEXP (x, 1)); | |
4349 | goto restart; | |
4350 | } | |
4351 | else if (GET_CODE (XEXP (x, 0)) == AND | |
4352 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)) | |
4353 | && ! side_effects_p (XEXP (x, 1))) | |
4354 | { | |
4355 | x = gen_binary (AND, mode, | |
4356 | gen_unary (NOT, mode, XEXP (XEXP (x, 0), 1)), | |
4357 | XEXP (x, 1)); | |
4358 | goto restart; | |
4359 | } | |
4360 | ||
4361 | ||
4362 | #if STORE_FLAG_VALUE == 1 | |
4363 | /* (xor (comparison foo bar) (const_int 1)) can become the reversed | |
4364 | comparison. */ | |
4365 | if (XEXP (x, 1) == const1_rtx | |
4366 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
4367 | && reversible_comparison_p (XEXP (x, 0))) | |
4368 | return gen_rtx_combine (reverse_condition (GET_CODE (XEXP (x, 0))), | |
4369 | mode, XEXP (XEXP (x, 0), 0), | |
4370 | XEXP (XEXP (x, 0), 1)); | |
4371 | ||
4372 | /* (lshiftrt foo C) where C is the number of bits in FOO minus 1 | |
4373 | is (lt foo (const_int 0)), so we can perform the above | |
4374 | simplification. */ | |
4375 | ||
4376 | if (XEXP (x, 1) == const1_rtx | |
4377 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
4378 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
4379 | && INTVAL (XEXP (XEXP (x, 0), 1)) == GET_MODE_BITSIZE (mode) - 1) | |
4380 | return gen_rtx_combine (GE, mode, XEXP (XEXP (x, 0), 0), const0_rtx); | |
4381 | #endif | |
4382 | ||
4383 | /* (xor (comparison foo bar) (const_int sign-bit)) | |
4384 | when STORE_FLAG_VALUE is the sign bit. */ | |
4385 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
4386 | && (STORE_FLAG_VALUE | |
4387 | == (HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1)) | |
4388 | && XEXP (x, 1) == const_true_rtx | |
4389 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
4390 | && reversible_comparison_p (XEXP (x, 0))) | |
4391 | return gen_rtx_combine (reverse_condition (GET_CODE (XEXP (x, 0))), | |
4392 | mode, XEXP (XEXP (x, 0), 0), | |
4393 | XEXP (XEXP (x, 0), 1)); | |
4394 | break; | |
4395 | ||
4396 | case ABS: | |
4397 | /* (abs (neg <foo>)) -> (abs <foo>) */ | |
4398 | if (GET_CODE (XEXP (x, 0)) == NEG) | |
4399 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
4400 | ||
4401 | /* If operand is something known to be positive, ignore the ABS. */ | |
4402 | if (GET_CODE (XEXP (x, 0)) == FFS || GET_CODE (XEXP (x, 0)) == ABS | |
4403 | || ((GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
4404 | <= HOST_BITS_PER_WIDE_INT) | |
4405 | && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0))) | |
4406 | & ((HOST_WIDE_INT) 1 | |
4407 | << (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - 1))) | |
4408 | == 0))) | |
4409 | return XEXP (x, 0); | |
4410 | ||
4411 | ||
4412 | /* If operand is known to be only -1 or 0, convert ABS to NEG. */ | |
4413 | if (num_sign_bit_copies (XEXP (x, 0), mode) == GET_MODE_BITSIZE (mode)) | |
4414 | { | |
4415 | x = gen_rtx_combine (NEG, mode, XEXP (x, 0)); | |
4416 | goto restart; | |
4417 | } | |
4418 | break; | |
4419 | ||
4420 | case FFS: | |
4421 | /* (ffs (*_extend <X>)) = (ffs <X>) */ | |
4422 | if (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
4423 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
4424 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
4425 | break; | |
4426 | ||
4427 | case FLOAT: | |
4428 | /* (float (sign_extend <X>)) = (float <X>). */ | |
4429 | if (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
4430 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
4431 | break; | |
4432 | ||
4433 | case LSHIFT: | |
4434 | case ASHIFT: | |
4435 | case LSHIFTRT: | |
4436 | case ASHIFTRT: | |
4437 | case ROTATE: | |
4438 | case ROTATERT: | |
4439 | /* If this is a shift by a constant amount, simplify it. */ | |
4440 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4441 | { | |
4442 | x = simplify_shift_const (x, code, mode, XEXP (x, 0), | |
4443 | INTVAL (XEXP (x, 1))); | |
4444 | if (GET_CODE (x) != code) | |
4445 | goto restart; | |
4446 | } | |
4447 | ||
4448 | #ifdef SHIFT_COUNT_TRUNCATED | |
4449 | else if (GET_CODE (XEXP (x, 1)) != REG) | |
4450 | SUBST (XEXP (x, 1), | |
4451 | force_to_mode (XEXP (x, 1), GET_MODE (x), | |
4452 | exact_log2 (GET_MODE_BITSIZE (GET_MODE (x))), | |
4453 | NULL_RTX)); | |
4454 | #endif | |
4455 | ||
4456 | break; | |
4457 | } | |
4458 | ||
4459 | return x; | |
4460 | } | |
4461 | \f | |
4462 | /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound | |
4463 | operations" because they can be replaced with two more basic operations. | |
4464 | ZERO_EXTEND is also considered "compound" because it can be replaced with | |
4465 | an AND operation, which is simpler, though only one operation. | |
4466 | ||
4467 | The function expand_compound_operation is called with an rtx expression | |
4468 | and will convert it to the appropriate shifts and AND operations, | |
4469 | simplifying at each stage. | |
4470 | ||
4471 | The function make_compound_operation is called to convert an expression | |
4472 | consisting of shifts and ANDs into the equivalent compound expression. | |
4473 | It is the inverse of this function, loosely speaking. */ | |
4474 | ||
4475 | static rtx | |
4476 | expand_compound_operation (x) | |
4477 | rtx x; | |
4478 | { | |
4479 | int pos = 0, len; | |
4480 | int unsignedp = 0; | |
4481 | int modewidth; | |
4482 | rtx tem; | |
4483 | ||
4484 | switch (GET_CODE (x)) | |
4485 | { | |
4486 | case ZERO_EXTEND: | |
4487 | unsignedp = 1; | |
4488 | case SIGN_EXTEND: | |
4489 | /* We can't necessarily use a const_int for a multiword mode; | |
4490 | it depends on implicitly extending the value. | |
4491 | Since we don't know the right way to extend it, | |
4492 | we can't tell whether the implicit way is right. | |
4493 | ||
4494 | Even for a mode that is no wider than a const_int, | |
4495 | we can't win, because we need to sign extend one of its bits through | |
4496 | the rest of it, and we don't know which bit. */ | |
4497 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) | |
4498 | return x; | |
4499 | ||
4500 | if (! FAKE_EXTEND_SAFE_P (GET_MODE (XEXP (x, 0)), XEXP (x, 0))) | |
4501 | return x; | |
4502 | ||
4503 | len = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))); | |
4504 | /* If the inner object has VOIDmode (the only way this can happen | |
4505 | is if it is a ASM_OPERANDS), we can't do anything since we don't | |
4506 | know how much masking to do. */ | |
4507 | if (len == 0) | |
4508 | return x; | |
4509 | ||
4510 | break; | |
4511 | ||
4512 | case ZERO_EXTRACT: | |
4513 | unsignedp = 1; | |
4514 | case SIGN_EXTRACT: | |
4515 | /* If the operand is a CLOBBER, just return it. */ | |
4516 | if (GET_CODE (XEXP (x, 0)) == CLOBBER) | |
4517 | return XEXP (x, 0); | |
4518 | ||
4519 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
4520 | || GET_CODE (XEXP (x, 2)) != CONST_INT | |
4521 | || GET_MODE (XEXP (x, 0)) == VOIDmode) | |
4522 | return x; | |
4523 | ||
4524 | len = INTVAL (XEXP (x, 1)); | |
4525 | pos = INTVAL (XEXP (x, 2)); | |
4526 | ||
4527 | /* If this goes outside the object being extracted, replace the object | |
4528 | with a (use (mem ...)) construct that only combine understands | |
4529 | and is used only for this purpose. */ | |
4530 | if (len + pos > GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))) | |
4531 | SUBST (XEXP (x, 0), gen_rtx (USE, GET_MODE (x), XEXP (x, 0))); | |
4532 | ||
4533 | #if BITS_BIG_ENDIAN | |
4534 | pos = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - len - pos; | |
4535 | #endif | |
4536 | break; | |
4537 | ||
4538 | default: | |
4539 | return x; | |
4540 | } | |
4541 | ||
4542 | /* If we reach here, we want to return a pair of shifts. The inner | |
4543 | shift is a left shift of BITSIZE - POS - LEN bits. The outer | |
4544 | shift is a right shift of BITSIZE - LEN bits. It is arithmetic or | |
4545 | logical depending on the value of UNSIGNEDP. | |
4546 | ||
4547 | If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be | |
4548 | converted into an AND of a shift. | |
4549 | ||
4550 | We must check for the case where the left shift would have a negative | |
4551 | count. This can happen in a case like (x >> 31) & 255 on machines | |
4552 | that can't shift by a constant. On those machines, we would first | |
4553 | combine the shift with the AND to produce a variable-position | |
4554 | extraction. Then the constant of 31 would be substituted in to produce | |
4555 | a such a position. */ | |
4556 | ||
4557 | modewidth = GET_MODE_BITSIZE (GET_MODE (x)); | |
4558 | if (modewidth >= pos - len) | |
4559 | tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT, | |
4560 | GET_MODE (x), | |
4561 | simplify_shift_const (NULL_RTX, ASHIFT, | |
4562 | GET_MODE (x), | |
4563 | XEXP (x, 0), | |
4564 | modewidth - pos - len), | |
4565 | modewidth - len); | |
4566 | ||
4567 | else if (unsignedp && len < HOST_BITS_PER_WIDE_INT) | |
4568 | tem = simplify_and_const_int (NULL_RTX, GET_MODE (x), | |
4569 | simplify_shift_const (NULL_RTX, LSHIFTRT, | |
4570 | GET_MODE (x), | |
4571 | XEXP (x, 0), pos), | |
4572 | ((HOST_WIDE_INT) 1 << len) - 1); | |
4573 | else | |
4574 | /* Any other cases we can't handle. */ | |
4575 | return x; | |
4576 | ||
4577 | ||
4578 | /* If we couldn't do this for some reason, return the original | |
4579 | expression. */ | |
4580 | if (GET_CODE (tem) == CLOBBER) | |
4581 | return x; | |
4582 | ||
4583 | return tem; | |
4584 | } | |
4585 | \f | |
4586 | /* X is a SET which contains an assignment of one object into | |
4587 | a part of another (such as a bit-field assignment, STRICT_LOW_PART, | |
4588 | or certain SUBREGS). If possible, convert it into a series of | |
4589 | logical operations. | |
4590 | ||
4591 | We half-heartedly support variable positions, but do not at all | |
4592 | support variable lengths. */ | |
4593 | ||
4594 | static rtx | |
4595 | expand_field_assignment (x) | |
4596 | rtx x; | |
4597 | { | |
4598 | rtx inner; | |
4599 | rtx pos; /* Always counts from low bit. */ | |
4600 | int len; | |
4601 | rtx mask; | |
4602 | enum machine_mode compute_mode; | |
4603 | ||
4604 | /* Loop until we find something we can't simplify. */ | |
4605 | while (1) | |
4606 | { | |
4607 | if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART | |
4608 | && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG) | |
4609 | { | |
4610 | inner = SUBREG_REG (XEXP (SET_DEST (x), 0)); | |
4611 | len = GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))); | |
4612 | pos = const0_rtx; | |
4613 | } | |
4614 | else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT | |
4615 | && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT) | |
4616 | { | |
4617 | inner = XEXP (SET_DEST (x), 0); | |
4618 | len = INTVAL (XEXP (SET_DEST (x), 1)); | |
4619 | pos = XEXP (SET_DEST (x), 2); | |
4620 | ||
4621 | /* If the position is constant and spans the width of INNER, | |
4622 | surround INNER with a USE to indicate this. */ | |
4623 | if (GET_CODE (pos) == CONST_INT | |
4624 | && INTVAL (pos) + len > GET_MODE_BITSIZE (GET_MODE (inner))) | |
4625 | inner = gen_rtx (USE, GET_MODE (SET_DEST (x)), inner); | |
4626 | ||
4627 | #if BITS_BIG_ENDIAN | |
4628 | if (GET_CODE (pos) == CONST_INT) | |
4629 | pos = GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) - len | |
4630 | - INTVAL (pos)); | |
4631 | else if (GET_CODE (pos) == MINUS | |
4632 | && GET_CODE (XEXP (pos, 1)) == CONST_INT | |
4633 | && (INTVAL (XEXP (pos, 1)) | |
4634 | == GET_MODE_BITSIZE (GET_MODE (inner)) - len)) | |
4635 | /* If position is ADJUST - X, new position is X. */ | |
4636 | pos = XEXP (pos, 0); | |
4637 | else | |
4638 | pos = gen_binary (MINUS, GET_MODE (pos), | |
4639 | GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) | |
4640 | - len), | |
4641 | pos); | |
4642 | #endif | |
4643 | } | |
4644 | ||
4645 | /* A SUBREG between two modes that occupy the same numbers of words | |
4646 | can be done by moving the SUBREG to the source. */ | |
4647 | else if (GET_CODE (SET_DEST (x)) == SUBREG | |
4648 | && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x))) | |
4649 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD) | |
4650 | == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x)))) | |
4651 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))) | |
4652 | { | |
4653 | x = gen_rtx (SET, VOIDmode, SUBREG_REG (SET_DEST (x)), | |
4654 | gen_lowpart_for_combine (GET_MODE (SUBREG_REG (SET_DEST (x))), | |
4655 | SET_SRC (x))); | |
4656 | continue; | |
4657 | } | |
4658 | else | |
4659 | break; | |
4660 | ||
4661 | while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner)) | |
4662 | inner = SUBREG_REG (inner); | |
4663 | ||
4664 | compute_mode = GET_MODE (inner); | |
4665 | ||
4666 | /* Compute a mask of LEN bits, if we can do this on the host machine. */ | |
4667 | if (len < HOST_BITS_PER_WIDE_INT) | |
4668 | mask = GEN_INT (((HOST_WIDE_INT) 1 << len) - 1); | |
4669 | else | |
4670 | break; | |
4671 | ||
4672 | /* Now compute the equivalent expression. Make a copy of INNER | |
4673 | for the SET_DEST in case it is a MEM into which we will substitute; | |
4674 | we don't want shared RTL in that case. */ | |
4675 | x = gen_rtx (SET, VOIDmode, copy_rtx (inner), | |
4676 | gen_binary (IOR, compute_mode, | |
4677 | gen_binary (AND, compute_mode, | |
4678 | gen_unary (NOT, compute_mode, | |
4679 | gen_binary (ASHIFT, | |
4680 | compute_mode, | |
4681 | mask, pos)), | |
4682 | inner), | |
4683 | gen_binary (ASHIFT, compute_mode, | |
4684 | gen_binary (AND, compute_mode, | |
4685 | gen_lowpart_for_combine | |
4686 | (compute_mode, | |
4687 | SET_SRC (x)), | |
4688 | mask), | |
4689 | pos))); | |
4690 | } | |
4691 | ||
4692 | return x; | |
4693 | } | |
4694 | \f | |
4695 | /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero, | |
4696 | it is an RTX that represents a variable starting position; otherwise, | |
4697 | POS is the (constant) starting bit position (counted from the LSB). | |
4698 | ||
4699 | INNER may be a USE. This will occur when we started with a bitfield | |
4700 | that went outside the boundary of the object in memory, which is | |
4701 | allowed on most machines. To isolate this case, we produce a USE | |
4702 | whose mode is wide enough and surround the MEM with it. The only | |
4703 | code that understands the USE is this routine. If it is not removed, | |
4704 | it will cause the resulting insn not to match. | |
4705 | ||
4706 | UNSIGNEDP is non-zero for an unsigned reference and zero for a | |
4707 | signed reference. | |
4708 | ||
4709 | IN_DEST is non-zero if this is a reference in the destination of a | |
4710 | SET. This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If non-zero, | |
4711 | a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will | |
4712 | be used. | |
4713 | ||
4714 | IN_COMPARE is non-zero if we are in a COMPARE. This means that a | |
4715 | ZERO_EXTRACT should be built even for bits starting at bit 0. | |
4716 | ||
4717 | MODE is the desired mode of the result (if IN_DEST == 0). */ | |
4718 | ||
4719 | static rtx | |
4720 | make_extraction (mode, inner, pos, pos_rtx, len, | |
4721 | unsignedp, in_dest, in_compare) | |
4722 | enum machine_mode mode; | |
4723 | rtx inner; | |
4724 | int pos; | |
4725 | rtx pos_rtx; | |
4726 | int len; | |
4727 | int unsignedp; | |
4728 | int in_dest, in_compare; | |
4729 | { | |
4730 | /* This mode describes the size of the storage area | |
4731 | to fetch the overall value from. Within that, we | |
4732 | ignore the POS lowest bits, etc. */ | |
4733 | enum machine_mode is_mode = GET_MODE (inner); | |
4734 | enum machine_mode inner_mode; | |
4735 | enum machine_mode wanted_mem_mode = byte_mode; | |
4736 | enum machine_mode pos_mode = word_mode; | |
4737 | enum machine_mode extraction_mode = word_mode; | |
4738 | enum machine_mode tmode = mode_for_size (len, MODE_INT, 1); | |
4739 | int spans_byte = 0; | |
4740 | rtx new = 0; | |
4741 | rtx orig_pos_rtx = pos_rtx; | |
4742 | ||
4743 | /* Get some information about INNER and get the innermost object. */ | |
4744 | if (GET_CODE (inner) == USE) | |
4745 | /* (use:SI (mem:QI foo)) stands for (mem:SI foo). */ | |
4746 | /* We don't need to adjust the position because we set up the USE | |
4747 | to pretend that it was a full-word object. */ | |
4748 | spans_byte = 1, inner = XEXP (inner, 0); | |
4749 | else if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner)) | |
4750 | { | |
4751 | /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...), | |
4752 | consider just the QI as the memory to extract from. | |
4753 | The subreg adds or removes high bits; its mode is | |
4754 | irrelevant to the meaning of this extraction, | |
4755 | since POS and LEN count from the lsb. */ | |
4756 | if (GET_CODE (SUBREG_REG (inner)) == MEM) | |
4757 | is_mode = GET_MODE (SUBREG_REG (inner)); | |
4758 | inner = SUBREG_REG (inner); | |
4759 | } | |
4760 | ||
4761 | inner_mode = GET_MODE (inner); | |
4762 | ||
4763 | if (pos_rtx && GET_CODE (pos_rtx) == CONST_INT) | |
4764 | pos = INTVAL (pos_rtx), pos_rtx = 0; | |
4765 | ||
4766 | /* See if this can be done without an extraction. We never can if the | |
4767 | width of the field is not the same as that of some integer mode. For | |
4768 | registers, we can only avoid the extraction if the position is at the | |
4769 | low-order bit and this is either not in the destination or we have the | |
4770 | appropriate STRICT_LOW_PART operation available. | |
4771 | ||
4772 | For MEM, we can avoid an extract if the field starts on an appropriate | |
4773 | boundary and we can change the mode of the memory reference. However, | |
4774 | we cannot directly access the MEM if we have a USE and the underlying | |
4775 | MEM is not TMODE. This combination means that MEM was being used in a | |
4776 | context where bits outside its mode were being referenced; that is only | |
4777 | valid in bit-field insns. */ | |
4778 | ||
4779 | if (tmode != BLKmode | |
4780 | && ! (spans_byte && inner_mode != tmode) | |
4781 | && ((pos_rtx == 0 && pos == 0 && GET_CODE (inner) != MEM | |
4782 | && (! in_dest | |
4783 | || (GET_CODE (inner) == REG | |
4784 | && (movstrict_optab->handlers[(int) tmode].insn_code | |
4785 | != CODE_FOR_nothing)))) | |
4786 | || (GET_CODE (inner) == MEM && pos_rtx == 0 | |
4787 | && (pos | |
4788 | % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode) | |
4789 | : BITS_PER_UNIT)) == 0 | |
4790 | /* We can't do this if we are widening INNER_MODE (it | |
4791 | may not be aligned, for one thing). */ | |
4792 | && GET_MODE_BITSIZE (inner_mode) >= GET_MODE_BITSIZE (tmode) | |
4793 | && (inner_mode == tmode | |
4794 | || (! mode_dependent_address_p (XEXP (inner, 0)) | |
4795 | && ! MEM_VOLATILE_P (inner)))))) | |
4796 | { | |
4797 | /* If INNER is a MEM, make a new MEM that encompasses just the desired | |
4798 | field. If the original and current mode are the same, we need not | |
4799 | adjust the offset. Otherwise, we do if bytes big endian. | |
4800 | ||
4801 | If INNER is not a MEM, get a piece consisting of the just the field | |
4802 | of interest (in this case POS must be 0). */ | |
4803 | ||
4804 | if (GET_CODE (inner) == MEM) | |
4805 | { | |
4806 | int offset; | |
4807 | /* POS counts from lsb, but make OFFSET count in memory order. */ | |
4808 | if (BYTES_BIG_ENDIAN) | |
4809 | offset = (GET_MODE_BITSIZE (is_mode) - len - pos) / BITS_PER_UNIT; | |
4810 | else | |
4811 | offset = pos / BITS_PER_UNIT; | |
4812 | ||
4813 | new = gen_rtx (MEM, tmode, plus_constant (XEXP (inner, 0), offset)); | |
4814 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (inner); | |
4815 | MEM_VOLATILE_P (new) = MEM_VOLATILE_P (inner); | |
4816 | MEM_IN_STRUCT_P (new) = MEM_IN_STRUCT_P (inner); | |
4817 | } | |
4818 | else if (GET_CODE (inner) == REG) | |
4819 | /* We can't call gen_lowpart_for_combine here since we always want | |
4820 | a SUBREG and it would sometimes return a new hard register. */ | |
4821 | new = gen_rtx (SUBREG, tmode, inner, | |
4822 | (WORDS_BIG_ENDIAN | |
4823 | && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD | |
4824 | ? ((GET_MODE_SIZE (inner_mode) - GET_MODE_SIZE (tmode)) | |
4825 | / UNITS_PER_WORD) | |
4826 | : 0)); | |
4827 | else | |
4828 | new = force_to_mode (inner, tmode, len, NULL_RTX); | |
4829 | ||
4830 | /* If this extraction is going into the destination of a SET, | |
4831 | make a STRICT_LOW_PART unless we made a MEM. */ | |
4832 | ||
4833 | if (in_dest) | |
4834 | return (GET_CODE (new) == MEM ? new | |
4835 | : (GET_CODE (new) != SUBREG | |
4836 | ? gen_rtx (CLOBBER, tmode, const0_rtx) | |
4837 | : gen_rtx_combine (STRICT_LOW_PART, VOIDmode, new))); | |
4838 | ||
4839 | /* Otherwise, sign- or zero-extend unless we already are in the | |
4840 | proper mode. */ | |
4841 | ||
4842 | return (mode == tmode ? new | |
4843 | : gen_rtx_combine (unsignedp ? ZERO_EXTEND : SIGN_EXTEND, | |
4844 | mode, new)); | |
4845 | } | |
4846 | ||
4847 | /* Unless this is a COMPARE or we have a funny memory reference, | |
4848 | don't do anything with zero-extending field extracts starting at | |
4849 | the low-order bit since they are simple AND operations. */ | |
4850 | if (pos_rtx == 0 && pos == 0 && ! in_dest | |
4851 | && ! in_compare && ! spans_byte && unsignedp) | |
4852 | return 0; | |
4853 | ||
4854 | /* Get the mode to use should INNER be a MEM, the mode for the position, | |
4855 | and the mode for the result. */ | |
4856 | #ifdef HAVE_insv | |
4857 | if (in_dest) | |
4858 | { | |
4859 | wanted_mem_mode = insn_operand_mode[(int) CODE_FOR_insv][0]; | |
4860 | pos_mode = insn_operand_mode[(int) CODE_FOR_insv][2]; | |
4861 | extraction_mode = insn_operand_mode[(int) CODE_FOR_insv][3]; | |
4862 | } | |
4863 | #endif | |
4864 | ||
4865 | #ifdef HAVE_extzv | |
4866 | if (! in_dest && unsignedp) | |
4867 | { | |
4868 | wanted_mem_mode = insn_operand_mode[(int) CODE_FOR_extzv][1]; | |
4869 | pos_mode = insn_operand_mode[(int) CODE_FOR_extzv][3]; | |
4870 | extraction_mode = insn_operand_mode[(int) CODE_FOR_extzv][0]; | |
4871 | } | |
4872 | #endif | |
4873 | ||
4874 | #ifdef HAVE_extv | |
4875 | if (! in_dest && ! unsignedp) | |
4876 | { | |
4877 | wanted_mem_mode = insn_operand_mode[(int) CODE_FOR_extv][1]; | |
4878 | pos_mode = insn_operand_mode[(int) CODE_FOR_extv][3]; | |
4879 | extraction_mode = insn_operand_mode[(int) CODE_FOR_extv][0]; | |
4880 | } | |
4881 | #endif | |
4882 | ||
4883 | /* Never narrow an object, since that might not be safe. */ | |
4884 | ||
4885 | if (mode != VOIDmode | |
4886 | && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode)) | |
4887 | extraction_mode = mode; | |
4888 | ||
4889 | if (pos_rtx && GET_MODE (pos_rtx) != VOIDmode | |
4890 | && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx))) | |
4891 | pos_mode = GET_MODE (pos_rtx); | |
4892 | ||
4893 | /* If this is not from memory or we have to change the mode of memory and | |
4894 | cannot, the desired mode is EXTRACTION_MODE. */ | |
4895 | if (GET_CODE (inner) != MEM | |
4896 | || (inner_mode != wanted_mem_mode | |
4897 | && (mode_dependent_address_p (XEXP (inner, 0)) | |
4898 | || MEM_VOLATILE_P (inner)))) | |
4899 | wanted_mem_mode = extraction_mode; | |
4900 | ||
4901 | #if BITS_BIG_ENDIAN | |
4902 | /* If position is constant, compute new position. Otherwise, build | |
4903 | subtraction. */ | |
4904 | if (pos_rtx == 0) | |
4905 | pos = (MAX (GET_MODE_BITSIZE (is_mode), GET_MODE_BITSIZE (wanted_mem_mode)) | |
4906 | - len - pos); | |
4907 | else | |
4908 | pos_rtx | |
4909 | = gen_rtx_combine (MINUS, GET_MODE (pos_rtx), | |
4910 | GEN_INT (MAX (GET_MODE_BITSIZE (is_mode), | |
4911 | GET_MODE_BITSIZE (wanted_mem_mode)) | |
4912 | - len), | |
4913 | pos_rtx); | |
4914 | #endif | |
4915 | ||
4916 | /* If INNER has a wider mode, make it smaller. If this is a constant | |
4917 | extract, try to adjust the byte to point to the byte containing | |
4918 | the value. */ | |
4919 | if (wanted_mem_mode != VOIDmode | |
4920 | && GET_MODE_SIZE (wanted_mem_mode) < GET_MODE_SIZE (is_mode) | |
4921 | && ((GET_CODE (inner) == MEM | |
4922 | && (inner_mode == wanted_mem_mode | |
4923 | || (! mode_dependent_address_p (XEXP (inner, 0)) | |
4924 | && ! MEM_VOLATILE_P (inner)))))) | |
4925 | { | |
4926 | int offset = 0; | |
4927 | ||
4928 | /* The computations below will be correct if the machine is big | |
4929 | endian in both bits and bytes or little endian in bits and bytes. | |
4930 | If it is mixed, we must adjust. */ | |
4931 | ||
4932 | /* If bytes are big endian and we had a paradoxical SUBREG, we must | |
4933 | adjust OFFSET to compensate. */ | |
4934 | #if BYTES_BIG_ENDIAN | |
4935 | if (! spans_byte | |
4936 | && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode)) | |
4937 | offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode); | |
4938 | #endif | |
4939 | ||
4940 | /* If this is a constant position, we can move to the desired byte. */ | |
4941 | if (pos_rtx == 0) | |
4942 | { | |
4943 | offset += pos / BITS_PER_UNIT; | |
4944 | pos %= GET_MODE_BITSIZE (wanted_mem_mode); | |
4945 | } | |
4946 | ||
4947 | #if BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN | |
4948 | if (! spans_byte && is_mode != wanted_mem_mode) | |
4949 | offset = (GET_MODE_SIZE (is_mode) | |
4950 | - GET_MODE_SIZE (wanted_mem_mode) - offset); | |
4951 | #endif | |
4952 | ||
4953 | if (offset != 0 || inner_mode != wanted_mem_mode) | |
4954 | { | |
4955 | rtx newmem = gen_rtx (MEM, wanted_mem_mode, | |
4956 | plus_constant (XEXP (inner, 0), offset)); | |
4957 | RTX_UNCHANGING_P (newmem) = RTX_UNCHANGING_P (inner); | |
4958 | MEM_VOLATILE_P (newmem) = MEM_VOLATILE_P (inner); | |
4959 | MEM_IN_STRUCT_P (newmem) = MEM_IN_STRUCT_P (inner); | |
4960 | inner = newmem; | |
4961 | } | |
4962 | } | |
4963 | ||
4964 | /* If INNER is not memory, we can always get it into the proper mode. */ | |
4965 | else if (GET_CODE (inner) != MEM) | |
4966 | inner = force_to_mode (inner, extraction_mode, | |
4967 | (pos < 0 ? GET_MODE_BITSIZE (extraction_mode) | |
4968 | : len + pos), | |
4969 | NULL_RTX); | |
4970 | ||
4971 | /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we | |
4972 | have to zero extend. Otherwise, we can just use a SUBREG. */ | |
4973 | if (pos_rtx != 0 | |
4974 | && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx))) | |
4975 | pos_rtx = gen_rtx_combine (ZERO_EXTEND, pos_mode, pos_rtx); | |
4976 | else if (pos_rtx != 0 | |
4977 | && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx))) | |
4978 | pos_rtx = gen_lowpart_for_combine (pos_mode, pos_rtx); | |
4979 | ||
4980 | /* Make POS_RTX unless we already have it and it is correct. If we don't | |
4981 | have a POS_RTX but we do have an ORIG_POS_RTX, the latter must | |
4982 | be a CONST_INT. */ | |
4983 | if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos) | |
4984 | pos_rtx = orig_pos_rtx; | |
4985 | ||
4986 | else if (pos_rtx == 0) | |
4987 | pos_rtx = GEN_INT (pos); | |
4988 | ||
4989 | /* Make the required operation. See if we can use existing rtx. */ | |
4990 | new = gen_rtx_combine (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT, | |
4991 | extraction_mode, inner, GEN_INT (len), pos_rtx); | |
4992 | if (! in_dest) | |
4993 | new = gen_lowpart_for_combine (mode, new); | |
4994 | ||
4995 | return new; | |
4996 | } | |
4997 | \f | |
4998 | /* Look at the expression rooted at X. Look for expressions | |
4999 | equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND. | |
5000 | Form these expressions. | |
5001 | ||
5002 | Return the new rtx, usually just X. | |
5003 | ||
5004 | Also, for machines like the Vax that don't have logical shift insns, | |
5005 | try to convert logical to arithmetic shift operations in cases where | |
5006 | they are equivalent. This undoes the canonicalizations to logical | |
5007 | shifts done elsewhere. | |
5008 | ||
5009 | We try, as much as possible, to re-use rtl expressions to save memory. | |
5010 | ||
5011 | IN_CODE says what kind of expression we are processing. Normally, it is | |
5012 | SET. In a memory address (inside a MEM, PLUS or minus, the latter two | |
5013 | being kludges), it is MEM. When processing the arguments of a comparison | |
5014 | or a COMPARE against zero, it is COMPARE. */ | |
5015 | ||
5016 | static rtx | |
5017 | make_compound_operation (x, in_code) | |
5018 | rtx x; | |
5019 | enum rtx_code in_code; | |
5020 | { | |
5021 | enum rtx_code code = GET_CODE (x); | |
5022 | enum machine_mode mode = GET_MODE (x); | |
5023 | int mode_width = GET_MODE_BITSIZE (mode); | |
5024 | enum rtx_code next_code; | |
5025 | int i, count; | |
5026 | rtx new = 0; | |
5027 | rtx tem; | |
5028 | char *fmt; | |
5029 | ||
5030 | /* Select the code to be used in recursive calls. Once we are inside an | |
5031 | address, we stay there. If we have a comparison, set to COMPARE, | |
5032 | but once inside, go back to our default of SET. */ | |
5033 | ||
5034 | next_code = (code == MEM || code == PLUS || code == MINUS ? MEM | |
5035 | : ((code == COMPARE || GET_RTX_CLASS (code) == '<') | |
5036 | && XEXP (x, 1) == const0_rtx) ? COMPARE | |
5037 | : in_code == COMPARE ? SET : in_code); | |
5038 | ||
5039 | /* Process depending on the code of this operation. If NEW is set | |
5040 | non-zero, it will be returned. */ | |
5041 | ||
5042 | switch (code) | |
5043 | { | |
5044 | case ASHIFT: | |
5045 | case LSHIFT: | |
5046 | /* Convert shifts by constants into multiplications if inside | |
5047 | an address. */ | |
5048 | if (in_code == MEM && GET_CODE (XEXP (x, 1)) == CONST_INT | |
5049 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT | |
5050 | && INTVAL (XEXP (x, 1)) >= 0) | |
5051 | { | |
5052 | new = make_compound_operation (XEXP (x, 0), next_code); | |
5053 | new = gen_rtx_combine (MULT, mode, new, | |
5054 | GEN_INT ((HOST_WIDE_INT) 1 | |
5055 | << INTVAL (XEXP (x, 1)))); | |
5056 | } | |
5057 | break; | |
5058 | ||
5059 | case AND: | |
5060 | /* If the second operand is not a constant, we can't do anything | |
5061 | with it. */ | |
5062 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
5063 | break; | |
5064 | ||
5065 | /* If the constant is a power of two minus one and the first operand | |
5066 | is a logical right shift, make an extraction. */ | |
5067 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
5068 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
5069 | { | |
5070 | new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); | |
5071 | new = make_extraction (mode, new, 0, XEXP (XEXP (x, 0), 1), i, 1, | |
5072 | 0, in_code == COMPARE); | |
5073 | } | |
5074 | ||
5075 | /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */ | |
5076 | else if (GET_CODE (XEXP (x, 0)) == SUBREG | |
5077 | && subreg_lowpart_p (XEXP (x, 0)) | |
5078 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT | |
5079 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
5080 | { | |
5081 | new = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0), | |
5082 | next_code); | |
5083 | new = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new, 0, | |
5084 | XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1, | |
5085 | 0, in_code == COMPARE); | |
5086 | } | |
5087 | ||
5088 | /* If we are have (and (rotate X C) M) and C is larger than the number | |
5089 | of bits in M, this is an extraction. */ | |
5090 | ||
5091 | else if (GET_CODE (XEXP (x, 0)) == ROTATE | |
5092 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5093 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0 | |
5094 | && i <= INTVAL (XEXP (XEXP (x, 0), 1))) | |
5095 | { | |
5096 | new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); | |
5097 | new = make_extraction (mode, new, | |
5098 | (GET_MODE_BITSIZE (mode) | |
5099 | - INTVAL (XEXP (XEXP (x, 0), 1))), | |
5100 | NULL_RTX, i, 1, 0, in_code == COMPARE); | |
5101 | } | |
5102 | ||
5103 | /* On machines without logical shifts, if the operand of the AND is | |
5104 | a logical shift and our mask turns off all the propagated sign | |
5105 | bits, we can replace the logical shift with an arithmetic shift. */ | |
5106 | else if (ashr_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing | |
5107 | && (lshr_optab->handlers[(int) mode].insn_code | |
5108 | == CODE_FOR_nothing) | |
5109 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
5110 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5111 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 | |
5112 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT | |
5113 | && mode_width <= HOST_BITS_PER_WIDE_INT) | |
5114 | { | |
5115 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); | |
5116 | ||
5117 | mask >>= INTVAL (XEXP (XEXP (x, 0), 1)); | |
5118 | if ((INTVAL (XEXP (x, 1)) & ~mask) == 0) | |
5119 | SUBST (XEXP (x, 0), | |
5120 | gen_rtx_combine (ASHIFTRT, mode, | |
5121 | make_compound_operation (XEXP (XEXP (x, 0), 0), | |
5122 | next_code), | |
5123 | XEXP (XEXP (x, 0), 1))); | |
5124 | } | |
5125 | ||
5126 | /* If the constant is one less than a power of two, this might be | |
5127 | representable by an extraction even if no shift is present. | |
5128 | If it doesn't end up being a ZERO_EXTEND, we will ignore it unless | |
5129 | we are in a COMPARE. */ | |
5130 | else if ((i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
5131 | new = make_extraction (mode, | |
5132 | make_compound_operation (XEXP (x, 0), | |
5133 | next_code), | |
5134 | 0, NULL_RTX, i, 1, 0, in_code == COMPARE); | |
5135 | ||
5136 | /* If we are in a comparison and this is an AND with a power of two, | |
5137 | convert this into the appropriate bit extract. */ | |
5138 | else if (in_code == COMPARE | |
5139 | && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0) | |
5140 | new = make_extraction (mode, | |
5141 | make_compound_operation (XEXP (x, 0), | |
5142 | next_code), | |
5143 | i, NULL_RTX, 1, 1, 0, 1); | |
5144 | ||
5145 | break; | |
5146 | ||
5147 | case LSHIFTRT: | |
5148 | /* If the sign bit is known to be zero, replace this with an | |
5149 | arithmetic shift. */ | |
5150 | if (ashr_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing | |
5151 | && lshr_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing | |
5152 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
5153 | && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0) | |
5154 | { | |
5155 | new = gen_rtx_combine (ASHIFTRT, mode, | |
5156 | make_compound_operation (XEXP (x, 0), | |
5157 | next_code), | |
5158 | XEXP (x, 1)); | |
5159 | break; | |
5160 | } | |
5161 | ||
5162 | /* ... fall through ... */ | |
5163 | ||
5164 | case ASHIFTRT: | |
5165 | /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1, | |
5166 | this is a SIGN_EXTRACT. */ | |
5167 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5168 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
5169 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5170 | && INTVAL (XEXP (x, 1)) >= INTVAL (XEXP (XEXP (x, 0), 1))) | |
5171 | { | |
5172 | new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); | |
5173 | new = make_extraction (mode, new, | |
5174 | (INTVAL (XEXP (x, 1)) | |
5175 | - INTVAL (XEXP (XEXP (x, 0), 1))), | |
5176 | NULL_RTX, mode_width - INTVAL (XEXP (x, 1)), | |
5177 | code == LSHIFTRT, 0, in_code == COMPARE); | |
5178 | } | |
5179 | ||
5180 | /* Similarly if we have (ashifrt (OP (ashift foo C1) C3) C2). In these | |
5181 | cases, we are better off returning a SIGN_EXTEND of the operation. */ | |
5182 | ||
5183 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5184 | && (GET_CODE (XEXP (x, 0)) == IOR || GET_CODE (XEXP (x, 0)) == AND | |
5185 | || GET_CODE (XEXP (x, 0)) == XOR | |
5186 | || GET_CODE (XEXP (x, 0)) == PLUS) | |
5187 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == ASHIFT | |
5188 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
5189 | && INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1)) < HOST_BITS_PER_WIDE_INT | |
5190 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5191 | && 0 == (INTVAL (XEXP (XEXP (x, 0), 1)) | |
5192 | & (((HOST_WIDE_INT) 1 | |
5193 | << (MIN (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1)), | |
5194 | INTVAL (XEXP (x, 1))) | |
5195 | - 1))))) | |
5196 | { | |
5197 | rtx c1 = XEXP (XEXP (XEXP (x, 0), 0), 1); | |
5198 | rtx c2 = XEXP (x, 1); | |
5199 | rtx c3 = XEXP (XEXP (x, 0), 1); | |
5200 | HOST_WIDE_INT newop1; | |
5201 | rtx inner = XEXP (XEXP (XEXP (x, 0), 0), 0); | |
5202 | ||
5203 | /* If C1 > C2, INNER needs to have the shift performed on it | |
5204 | for C1-C2 bits. */ | |
5205 | if (INTVAL (c1) > INTVAL (c2)) | |
5206 | { | |
5207 | inner = gen_binary (ASHIFT, mode, inner, | |
5208 | GEN_INT (INTVAL (c1) - INTVAL (c2))); | |
5209 | c1 = c2; | |
5210 | } | |
5211 | ||
5212 | newop1 = INTVAL (c3) >> INTVAL (c1); | |
5213 | new = make_compound_operation (inner, | |
5214 | GET_CODE (XEXP (x, 0)) == PLUS | |
5215 | ? MEM : GET_CODE (XEXP (x, 0))); | |
5216 | new = make_extraction (mode, | |
5217 | gen_binary (GET_CODE (XEXP (x, 0)), mode, new, | |
5218 | GEN_INT (newop1)), | |
5219 | INTVAL (c2) - INTVAL (c1), | |
5220 | NULL_RTX, mode_width - INTVAL (c2), | |
5221 | code == LSHIFTRT, 0, in_code == COMPARE); | |
5222 | } | |
5223 | ||
5224 | /* Similarly for (ashiftrt (neg (ashift FOO C1)) C2). */ | |
5225 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5226 | && GET_CODE (XEXP (x, 0)) == NEG | |
5227 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == ASHIFT | |
5228 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
5229 | && INTVAL (XEXP (x, 1)) >= INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))) | |
5230 | { | |
5231 | new = make_compound_operation (XEXP (XEXP (XEXP (x, 0), 0), 0), | |
5232 | next_code); | |
5233 | new = make_extraction (mode, | |
5234 | gen_unary (GET_CODE (XEXP (x, 0)), mode, | |
5235 | new, 0), | |
5236 | (INTVAL (XEXP (x, 1)) | |
5237 | - INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))), | |
5238 | NULL_RTX, mode_width - INTVAL (XEXP (x, 1)), | |
5239 | code == LSHIFTRT, 0, in_code == COMPARE); | |
5240 | } | |
5241 | break; | |
5242 | ||
5243 | case SUBREG: | |
5244 | /* Call ourselves recursively on the inner expression. If we are | |
5245 | narrowing the object and it has a different RTL code from | |
5246 | what it originally did, do this SUBREG as a force_to_mode. */ | |
5247 | ||
5248 | tem = make_compound_operation (SUBREG_REG (x), in_code); | |
5249 | if (GET_CODE (tem) != GET_CODE (SUBREG_REG (x)) | |
5250 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (tem)) | |
5251 | && subreg_lowpart_p (x)) | |
5252 | { | |
5253 | rtx newer = force_to_mode (tem, mode, | |
5254 | GET_MODE_BITSIZE (mode), NULL_RTX); | |
5255 | ||
5256 | /* If we have something other than a SUBREG, we might have | |
5257 | done an expansion, so rerun outselves. */ | |
5258 | if (GET_CODE (newer) != SUBREG) | |
5259 | newer = make_compound_operation (newer, in_code); | |
5260 | ||
5261 | return newer; | |
5262 | } | |
5263 | } | |
5264 | ||
5265 | if (new) | |
5266 | { | |
5267 | x = gen_lowpart_for_combine (mode, new); | |
5268 | code = GET_CODE (x); | |
5269 | } | |
5270 | ||
5271 | /* Now recursively process each operand of this operation. */ | |
5272 | fmt = GET_RTX_FORMAT (code); | |
5273 | for (i = 0; i < GET_RTX_LENGTH (code); i++) | |
5274 | if (fmt[i] == 'e') | |
5275 | { | |
5276 | new = make_compound_operation (XEXP (x, i), next_code); | |
5277 | SUBST (XEXP (x, i), new); | |
5278 | } | |
5279 | ||
5280 | return x; | |
5281 | } | |
5282 | \f | |
5283 | /* Given M see if it is a value that would select a field of bits | |
5284 | within an item, but not the entire word. Return -1 if not. | |
5285 | Otherwise, return the starting position of the field, where 0 is the | |
5286 | low-order bit. | |
5287 | ||
5288 | *PLEN is set to the length of the field. */ | |
5289 | ||
5290 | static int | |
5291 | get_pos_from_mask (m, plen) | |
5292 | unsigned HOST_WIDE_INT m; | |
5293 | int *plen; | |
5294 | { | |
5295 | /* Get the bit number of the first 1 bit from the right, -1 if none. */ | |
5296 | int pos = exact_log2 (m & - m); | |
5297 | ||
5298 | if (pos < 0) | |
5299 | return -1; | |
5300 | ||
5301 | /* Now shift off the low-order zero bits and see if we have a power of | |
5302 | two minus 1. */ | |
5303 | *plen = exact_log2 ((m >> pos) + 1); | |
5304 | ||
5305 | if (*plen <= 0) | |
5306 | return -1; | |
5307 | ||
5308 | return pos; | |
5309 | } | |
5310 | \f | |
5311 | /* Rewrite X so that it is an expression in MODE. We only care about the | |
5312 | low-order BITS bits so we can ignore AND operations that just clear | |
5313 | higher-order bits. | |
5314 | ||
5315 | Also, if REG is non-zero and X is a register equal in value to REG, | |
5316 | replace X with REG. */ | |
5317 | ||
5318 | static rtx | |
5319 | force_to_mode (x, mode, bits, reg) | |
5320 | rtx x; | |
5321 | enum machine_mode mode; | |
5322 | int bits; | |
5323 | rtx reg; | |
5324 | { | |
5325 | enum rtx_code code = GET_CODE (x); | |
5326 | enum machine_mode op_mode = mode; | |
5327 | ||
5328 | /* If X is narrower than MODE or if BITS is larger than the size of MODE, | |
5329 | just get X in the proper mode. */ | |
5330 | ||
5331 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode) | |
5332 | || bits > GET_MODE_BITSIZE (mode)) | |
5333 | return gen_lowpart_for_combine (mode, x); | |
5334 | ||
5335 | switch (code) | |
5336 | { | |
5337 | case SIGN_EXTEND: | |
5338 | case ZERO_EXTEND: | |
5339 | case ZERO_EXTRACT: | |
5340 | case SIGN_EXTRACT: | |
5341 | x = expand_compound_operation (x); | |
5342 | if (GET_CODE (x) != code) | |
5343 | return force_to_mode (x, mode, bits, reg); | |
5344 | break; | |
5345 | ||
5346 | case REG: | |
5347 | if (reg != 0 && (rtx_equal_p (get_last_value (reg), x) | |
5348 | || rtx_equal_p (reg, get_last_value (x)))) | |
5349 | x = reg; | |
5350 | break; | |
5351 | ||
5352 | case CONST_INT: | |
5353 | if (bits < HOST_BITS_PER_WIDE_INT) | |
5354 | x = GEN_INT (INTVAL (x) & (((HOST_WIDE_INT) 1 << bits) - 1)); | |
5355 | return x; | |
5356 | ||
5357 | case SUBREG: | |
5358 | /* Ignore low-order SUBREGs. */ | |
5359 | if (subreg_lowpart_p (x)) | |
5360 | return force_to_mode (SUBREG_REG (x), mode, bits, reg); | |
5361 | break; | |
5362 | ||
5363 | case AND: | |
5364 | /* If this is an AND with a constant. Otherwise, we fall through to | |
5365 | do the general binary case. */ | |
5366 | ||
5367 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5368 | { | |
5369 | HOST_WIDE_INT mask = INTVAL (XEXP (x, 1)); | |
5370 | int len = exact_log2 (mask + 1); | |
5371 | rtx op = XEXP (x, 0); | |
5372 | ||
5373 | /* If this is masking some low-order bits, we may be able to | |
5374 | impose a stricter constraint on what bits of the operand are | |
5375 | required. */ | |
5376 | ||
5377 | op = force_to_mode (op, mode, len > 0 ? MIN (len, bits) : bits, | |
5378 | reg); | |
5379 | ||
5380 | if (bits < HOST_BITS_PER_WIDE_INT) | |
5381 | mask &= ((HOST_WIDE_INT) 1 << bits) - 1; | |
5382 | ||
5383 | /* If we have no AND in MODE, use the original mode for the | |
5384 | operation. */ | |
5385 | ||
5386 | if (and_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5387 | op_mode = GET_MODE (x); | |
5388 | ||
5389 | x = simplify_and_const_int (x, op_mode, op, mask); | |
5390 | ||
5391 | /* If X is still an AND, see if it is an AND with a mask that | |
5392 | is just some low-order bits. If so, and it is BITS wide (it | |
5393 | can't be wider), we don't need it. */ | |
5394 | ||
5395 | if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT | |
5396 | && bits < HOST_BITS_PER_WIDE_INT | |
5397 | && INTVAL (XEXP (x, 1)) == ((HOST_WIDE_INT) 1 << bits) - 1) | |
5398 | x = XEXP (x, 0); | |
5399 | ||
5400 | break; | |
5401 | } | |
5402 | ||
5403 | /* ... fall through ... */ | |
5404 | ||
5405 | case PLUS: | |
5406 | case MINUS: | |
5407 | case MULT: | |
5408 | case IOR: | |
5409 | case XOR: | |
5410 | /* For most binary operations, just propagate into the operation and | |
5411 | change the mode if we have an operation of that mode. */ | |
5412 | ||
5413 | if ((code == PLUS | |
5414 | && add_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5415 | || (code == MINUS | |
5416 | && sub_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5417 | || (code == MULT && (smul_optab->handlers[(int) mode].insn_code | |
5418 | == CODE_FOR_nothing)) | |
5419 | || (code == AND | |
5420 | && and_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5421 | || (code == IOR | |
5422 | && ior_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5423 | || (code == XOR && (xor_optab->handlers[(int) mode].insn_code | |
5424 | == CODE_FOR_nothing))) | |
5425 | op_mode = GET_MODE (x); | |
5426 | ||
5427 | x = gen_binary (code, op_mode, | |
5428 | gen_lowpart_for_combine (op_mode, | |
5429 | force_to_mode (XEXP (x, 0), | |
5430 | mode, bits, | |
5431 | reg)), | |
5432 | gen_lowpart_for_combine (op_mode, | |
5433 | force_to_mode (XEXP (x, 1), | |
5434 | mode, bits, | |
5435 | reg))); | |
5436 | break; | |
5437 | ||
5438 | case ASHIFT: | |
5439 | case LSHIFT: | |
5440 | /* For left shifts, do the same, but just for the first operand. | |
5441 | However, we cannot do anything with shifts where we cannot | |
5442 | guarantee that the counts are smaller than the size of the mode | |
5443 | because such a count will have a different meaning in a | |
5444 | wider mode. | |
5445 | ||
5446 | If we can narrow the shift and know the count, we need even fewer | |
5447 | bits of the first operand. */ | |
5448 | ||
5449 | if (! (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5450 | && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode)) | |
5451 | && ! (GET_MODE (XEXP (x, 1)) != VOIDmode | |
5452 | && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1))) | |
5453 | < (unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode)))) | |
5454 | break; | |
5455 | ||
5456 | if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) < bits) | |
5457 | bits -= INTVAL (XEXP (x, 1)); | |
5458 | ||
5459 | if ((code == ASHIFT | |
5460 | && ashl_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5461 | || (code == LSHIFT && (lshl_optab->handlers[(int) mode].insn_code | |
5462 | == CODE_FOR_nothing))) | |
5463 | op_mode = GET_MODE (x); | |
5464 | ||
5465 | x = gen_binary (code, op_mode, | |
5466 | gen_lowpart_for_combine (op_mode, | |
5467 | force_to_mode (XEXP (x, 0), | |
5468 | mode, bits, | |
5469 | reg)), | |
5470 | XEXP (x, 1)); | |
5471 | break; | |
5472 | ||
5473 | case LSHIFTRT: | |
5474 | /* Here we can only do something if the shift count is a constant and | |
5475 | the count plus BITS is no larger than the width of MODE. In that | |
5476 | case, we can do the shift in MODE. */ | |
5477 | ||
5478 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5479 | && INTVAL (XEXP (x, 1)) + bits <= GET_MODE_BITSIZE (mode)) | |
5480 | { | |
5481 | rtx inner = force_to_mode (XEXP (x, 0), mode, | |
5482 | bits + INTVAL (XEXP (x, 1)), reg); | |
5483 | ||
5484 | if (lshr_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5485 | op_mode = GET_MODE (x); | |
5486 | ||
5487 | x = gen_binary (LSHIFTRT, op_mode, | |
5488 | gen_lowpart_for_combine (op_mode, inner), | |
5489 | XEXP (x, 1)); | |
5490 | } | |
5491 | break; | |
5492 | ||
5493 | case ASHIFTRT: | |
5494 | /* If this is a sign-extension operation that just affects bits | |
5495 | we don't care about, remove it. */ | |
5496 | ||
5497 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5498 | && INTVAL (XEXP (x, 1)) >= 0 | |
5499 | && INTVAL (XEXP (x, 1)) <= GET_MODE_BITSIZE (GET_MODE (x)) - bits | |
5500 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
5501 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5502 | && INTVAL (XEXP (XEXP (x, 0), 1)) == INTVAL (XEXP (x, 1))) | |
5503 | return force_to_mode (XEXP (XEXP (x, 0), 0), mode, bits, reg); | |
5504 | break; | |
5505 | ||
5506 | case NEG: | |
5507 | case NOT: | |
5508 | if ((code == NEG | |
5509 | && neg_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
5510 | || (code == NOT && (one_cmpl_optab->handlers[(int) mode].insn_code | |
5511 | == CODE_FOR_nothing))) | |
5512 | op_mode = GET_MODE (x); | |
5513 | ||
5514 | /* Handle these similarly to the way we handle most binary operations. */ | |
5515 | x = gen_unary (code, op_mode, | |
5516 | gen_lowpart_for_combine (op_mode, | |
5517 | force_to_mode (XEXP (x, 0), mode, | |
5518 | bits, reg))); | |
5519 | break; | |
5520 | ||
5521 | case IF_THEN_ELSE: | |
5522 | /* We have no way of knowing if the IF_THEN_ELSE can itself be | |
5523 | written in a narrower mode. We play it safe and do not do so. */ | |
5524 | ||
5525 | SUBST (XEXP (x, 1), | |
5526 | gen_lowpart_for_combine (GET_MODE (x), | |
5527 | force_to_mode (XEXP (x, 1), mode, | |
5528 | bits, reg))); | |
5529 | SUBST (XEXP (x, 2), | |
5530 | gen_lowpart_for_combine (GET_MODE (x), | |
5531 | force_to_mode (XEXP (x, 2), mode, | |
5532 | bits, reg))); | |
5533 | break; | |
5534 | } | |
5535 | ||
5536 | /* Ensure we return a value of the proper mode. */ | |
5537 | return gen_lowpart_for_combine (mode, x); | |
5538 | } | |
5539 | \f | |
5540 | /* Return the value of expression X given the fact that condition COND | |
5541 | is known to be true when applied to REG as its first operand and VAL | |
5542 | as its second. X is known to not be shared and so can be modified in | |
5543 | place. | |
5544 | ||
5545 | We only handle the simplest cases, and specifically those cases that | |
5546 | arise with IF_THEN_ELSE expressions. */ | |
5547 | ||
5548 | static rtx | |
5549 | known_cond (x, cond, reg, val) | |
5550 | rtx x; | |
5551 | enum rtx_code cond; | |
5552 | rtx reg, val; | |
5553 | { | |
5554 | enum rtx_code code = GET_CODE (x); | |
5555 | rtx new, temp; | |
5556 | char *fmt; | |
5557 | int i, j; | |
5558 | ||
5559 | if (side_effects_p (x)) | |
5560 | return x; | |
5561 | ||
5562 | if (cond == EQ && rtx_equal_p (x, reg)) | |
5563 | return val; | |
5564 | ||
5565 | /* If X is (abs REG) and we know something about REG's relationship | |
5566 | with zero, we may be able to simplify this. */ | |
5567 | ||
5568 | if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx) | |
5569 | switch (cond) | |
5570 | { | |
5571 | case GE: case GT: case EQ: | |
5572 | return XEXP (x, 0); | |
5573 | case LT: case LE: | |
5574 | return gen_unary (NEG, GET_MODE (XEXP (x, 0)), XEXP (x, 0)); | |
5575 | } | |
5576 | ||
5577 | /* The only other cases we handle are MIN, MAX, and comparisons if the | |
5578 | operands are the same as REG and VAL. */ | |
5579 | ||
5580 | else if (GET_RTX_CLASS (code) == '<' || GET_RTX_CLASS (code) == 'c') | |
5581 | { | |
5582 | if (rtx_equal_p (XEXP (x, 0), val)) | |
5583 | cond = swap_condition (cond), temp = val, val = reg, reg = temp; | |
5584 | ||
5585 | if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val)) | |
5586 | { | |
5587 | if (GET_RTX_CLASS (code) == '<') | |
5588 | return (comparison_dominates_p (cond, code) ? const_true_rtx | |
5589 | : (comparison_dominates_p (cond, | |
5590 | reverse_condition (code)) | |
5591 | ? const0_rtx : x)); | |
5592 | ||
5593 | else if (code == SMAX || code == SMIN | |
5594 | || code == UMIN || code == UMAX) | |
5595 | { | |
5596 | int unsignedp = (code == UMIN || code == UMAX); | |
5597 | ||
5598 | if (code == SMAX || code == UMAX) | |
5599 | cond = reverse_condition (cond); | |
5600 | ||
5601 | switch (cond) | |
5602 | { | |
5603 | case GE: case GT: | |
5604 | return unsignedp ? x : XEXP (x, 1); | |
5605 | case LE: case LT: | |
5606 | return unsignedp ? x : XEXP (x, 0); | |
5607 | case GEU: case GTU: | |
5608 | return unsignedp ? XEXP (x, 1) : x; | |
5609 | case LEU: case LTU: | |
5610 | return unsignedp ? XEXP (x, 0) : x; | |
5611 | } | |
5612 | } | |
5613 | } | |
5614 | } | |
5615 | ||
5616 | fmt = GET_RTX_FORMAT (code); | |
5617 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
5618 | { | |
5619 | if (fmt[i] == 'e') | |
5620 | SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val)); | |
5621 | else if (fmt[i] == 'E') | |
5622 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
5623 | SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j), | |
5624 | cond, reg, val)); | |
5625 | } | |
5626 | ||
5627 | return x; | |
5628 | } | |
5629 | \f | |
5630 | /* See if X, a SET operation, can be rewritten as a bit-field assignment. | |
5631 | Return that assignment if so. | |
5632 | ||
5633 | We only handle the most common cases. */ | |
5634 | ||
5635 | static rtx | |
5636 | make_field_assignment (x) | |
5637 | rtx x; | |
5638 | { | |
5639 | rtx dest = SET_DEST (x); | |
5640 | rtx src = SET_SRC (x); | |
5641 | rtx ourdest; | |
5642 | rtx assign; | |
5643 | HOST_WIDE_INT c1; | |
5644 | int pos, len; | |
5645 | rtx other; | |
5646 | enum machine_mode mode; | |
5647 | ||
5648 | /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is | |
5649 | a clear of a one-bit field. We will have changed it to | |
5650 | (and (rotate (const_int -2) POS) DEST), so check for that. Also check | |
5651 | for a SUBREG. */ | |
5652 | ||
5653 | if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE | |
5654 | && GET_CODE (XEXP (XEXP (src, 0), 0)) == CONST_INT | |
5655 | && INTVAL (XEXP (XEXP (src, 0), 0)) == -2 | |
5656 | && (rtx_equal_p (dest, XEXP (src, 1)) | |
5657 | || rtx_equal_p (dest, get_last_value (XEXP (src, 1))) | |
5658 | || rtx_equal_p (get_last_value (dest), XEXP (src, 1)))) | |
5659 | { | |
5660 | assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1), | |
5661 | 1, 1, 1, 0); | |
5662 | return gen_rtx (SET, VOIDmode, assign, const0_rtx); | |
5663 | } | |
5664 | ||
5665 | else if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG | |
5666 | && subreg_lowpart_p (XEXP (src, 0)) | |
5667 | && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0))) | |
5668 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0))))) | |
5669 | && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE | |
5670 | && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2 | |
5671 | && (rtx_equal_p (dest, XEXP (src, 1)) | |
5672 | || rtx_equal_p (dest, get_last_value (XEXP (src, 1))) | |
5673 | || rtx_equal_p (get_last_value (dest), XEXP (src, 1)))) | |
5674 | { | |
5675 | assign = make_extraction (VOIDmode, dest, 0, | |
5676 | XEXP (SUBREG_REG (XEXP (src, 0)), 1), | |
5677 | 1, 1, 1, 0); | |
5678 | return gen_rtx (SET, VOIDmode, assign, const0_rtx); | |
5679 | } | |
5680 | ||
5681 | /* If SRC is (ior (ashift (const_int 1) POS DEST)), this is a set of a | |
5682 | one-bit field. */ | |
5683 | else if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT | |
5684 | && XEXP (XEXP (src, 0), 0) == const1_rtx | |
5685 | && (rtx_equal_p (dest, XEXP (src, 1)) | |
5686 | || rtx_equal_p (dest, get_last_value (XEXP (src, 1))) | |
5687 | || rtx_equal_p (get_last_value (dest), XEXP (src, 1)))) | |
5688 | { | |
5689 | assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1), | |
5690 | 1, 1, 1, 0); | |
5691 | return gen_rtx (SET, VOIDmode, assign, const1_rtx); | |
5692 | } | |
5693 | ||
5694 | /* The other case we handle is assignments into a constant-position | |
5695 | field. They look like (ior (and DEST C1) OTHER). If C1 represents | |
5696 | a mask that has all one bits except for a group of zero bits and | |
5697 | OTHER is known to have zeros where C1 has ones, this is such an | |
5698 | assignment. Compute the position and length from C1. Shift OTHER | |
5699 | to the appropriate position, force it to the required mode, and | |
5700 | make the extraction. Check for the AND in both operands. */ | |
5701 | ||
5702 | if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == AND | |
5703 | && GET_CODE (XEXP (XEXP (src, 0), 1)) == CONST_INT | |
5704 | && (rtx_equal_p (XEXP (XEXP (src, 0), 0), dest) | |
5705 | || rtx_equal_p (XEXP (XEXP (src, 0), 0), get_last_value (dest)) | |
5706 | || rtx_equal_p (get_last_value (XEXP (XEXP (src, 0), 1)), dest))) | |
5707 | c1 = INTVAL (XEXP (XEXP (src, 0), 1)), other = XEXP (src, 1); | |
5708 | else if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 1)) == AND | |
5709 | && GET_CODE (XEXP (XEXP (src, 1), 1)) == CONST_INT | |
5710 | && (rtx_equal_p (XEXP (XEXP (src, 1), 0), dest) | |
5711 | || rtx_equal_p (XEXP (XEXP (src, 1), 0), get_last_value (dest)) | |
5712 | || rtx_equal_p (get_last_value (XEXP (XEXP (src, 1), 0)), | |
5713 | dest))) | |
5714 | c1 = INTVAL (XEXP (XEXP (src, 1), 1)), other = XEXP (src, 0); | |
5715 | else | |
5716 | return x; | |
5717 | ||
5718 | pos = get_pos_from_mask (~c1, &len); | |
5719 | if (pos < 0 || pos + len > GET_MODE_BITSIZE (GET_MODE (dest)) | |
5720 | || (GET_MODE_BITSIZE (GET_MODE (other)) <= HOST_BITS_PER_WIDE_INT | |
5721 | && (c1 & nonzero_bits (other, GET_MODE (other))) != 0)) | |
5722 | return x; | |
5723 | ||
5724 | assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0); | |
5725 | ||
5726 | /* The mode to use for the source is the mode of the assignment, or of | |
5727 | what is inside a possible STRICT_LOW_PART. */ | |
5728 | mode = (GET_CODE (assign) == STRICT_LOW_PART | |
5729 | ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign)); | |
5730 | ||
5731 | /* Shift OTHER right POS places and make it the source, restricting it | |
5732 | to the proper length and mode. */ | |
5733 | ||
5734 | src = force_to_mode (simplify_shift_const (NULL_RTX, LSHIFTRT, | |
5735 | GET_MODE (src), other, pos), | |
5736 | mode, len, dest); | |
5737 | ||
5738 | return gen_rtx_combine (SET, VOIDmode, assign, src); | |
5739 | } | |
5740 | \f | |
5741 | /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c) | |
5742 | if so. */ | |
5743 | ||
5744 | static rtx | |
5745 | apply_distributive_law (x) | |
5746 | rtx x; | |
5747 | { | |
5748 | enum rtx_code code = GET_CODE (x); | |
5749 | rtx lhs, rhs, other; | |
5750 | rtx tem; | |
5751 | enum rtx_code inner_code; | |
5752 | ||
5753 | /* Distributivity is not true for floating point. | |
5754 | It can change the value. So don't do it. | |
5755 | -- rms and moshier@world.std.com. */ | |
5756 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
5757 | return x; | |
5758 | ||
5759 | /* The outer operation can only be one of the following: */ | |
5760 | if (code != IOR && code != AND && code != XOR | |
5761 | && code != PLUS && code != MINUS) | |
5762 | return x; | |
5763 | ||
5764 | lhs = XEXP (x, 0), rhs = XEXP (x, 1); | |
5765 | ||
5766 | /* If either operand is a primitive we can't do anything, so get out fast. */ | |
5767 | if (GET_RTX_CLASS (GET_CODE (lhs)) == 'o' | |
5768 | || GET_RTX_CLASS (GET_CODE (rhs)) == 'o') | |
5769 | return x; | |
5770 | ||
5771 | lhs = expand_compound_operation (lhs); | |
5772 | rhs = expand_compound_operation (rhs); | |
5773 | inner_code = GET_CODE (lhs); | |
5774 | if (inner_code != GET_CODE (rhs)) | |
5775 | return x; | |
5776 | ||
5777 | /* See if the inner and outer operations distribute. */ | |
5778 | switch (inner_code) | |
5779 | { | |
5780 | case LSHIFTRT: | |
5781 | case ASHIFTRT: | |
5782 | case AND: | |
5783 | case IOR: | |
5784 | /* These all distribute except over PLUS. */ | |
5785 | if (code == PLUS || code == MINUS) | |
5786 | return x; | |
5787 | break; | |
5788 | ||
5789 | case MULT: | |
5790 | if (code != PLUS && code != MINUS) | |
5791 | return x; | |
5792 | break; | |
5793 | ||
5794 | case ASHIFT: | |
5795 | case LSHIFT: | |
5796 | /* These are also multiplies, so they distribute over everything. */ | |
5797 | break; | |
5798 | ||
5799 | case SUBREG: | |
5800 | /* Non-paradoxical SUBREGs distributes over all operations, provided | |
5801 | the inner modes and word numbers are the same, this is an extraction | |
5802 | of a low-order part, we don't convert an fp operation to int or | |
5803 | vice versa, and we would not be converting a single-word | |
5804 | operation into a multi-word operation. The latter test is not | |
5805 | required, but it prevents generating unneeded multi-word operations. | |
5806 | Some of the previous tests are redundant given the latter test, but | |
5807 | are retained because they are required for correctness. | |
5808 | ||
5809 | We produce the result slightly differently in this case. */ | |
5810 | ||
5811 | if (GET_MODE (SUBREG_REG (lhs)) != GET_MODE (SUBREG_REG (rhs)) | |
5812 | || SUBREG_WORD (lhs) != SUBREG_WORD (rhs) | |
5813 | || ! subreg_lowpart_p (lhs) | |
5814 | || (GET_MODE_CLASS (GET_MODE (lhs)) | |
5815 | != GET_MODE_CLASS (GET_MODE (SUBREG_REG (lhs)))) | |
5816 | || (GET_MODE_SIZE (GET_MODE (lhs)) | |
5817 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs)))) | |
5818 | || GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))) > UNITS_PER_WORD) | |
5819 | return x; | |
5820 | ||
5821 | tem = gen_binary (code, GET_MODE (SUBREG_REG (lhs)), | |
5822 | SUBREG_REG (lhs), SUBREG_REG (rhs)); | |
5823 | return gen_lowpart_for_combine (GET_MODE (x), tem); | |
5824 | ||
5825 | default: | |
5826 | return x; | |
5827 | } | |
5828 | ||
5829 | /* Set LHS and RHS to the inner operands (A and B in the example | |
5830 | above) and set OTHER to the common operand (C in the example). | |
5831 | These is only one way to do this unless the inner operation is | |
5832 | commutative. */ | |
5833 | if (GET_RTX_CLASS (inner_code) == 'c' | |
5834 | && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0))) | |
5835 | other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1); | |
5836 | else if (GET_RTX_CLASS (inner_code) == 'c' | |
5837 | && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1))) | |
5838 | other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0); | |
5839 | else if (GET_RTX_CLASS (inner_code) == 'c' | |
5840 | && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0))) | |
5841 | other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1); | |
5842 | else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1))) | |
5843 | other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0); | |
5844 | else | |
5845 | return x; | |
5846 | ||
5847 | /* Form the new inner operation, seeing if it simplifies first. */ | |
5848 | tem = gen_binary (code, GET_MODE (x), lhs, rhs); | |
5849 | ||
5850 | /* There is one exception to the general way of distributing: | |
5851 | (a ^ b) | (a ^ c) -> (~a) & (b ^ c) */ | |
5852 | if (code == XOR && inner_code == IOR) | |
5853 | { | |
5854 | inner_code = AND; | |
5855 | other = gen_unary (NOT, GET_MODE (x), other); | |
5856 | } | |
5857 | ||
5858 | /* We may be able to continuing distributing the result, so call | |
5859 | ourselves recursively on the inner operation before forming the | |
5860 | outer operation, which we return. */ | |
5861 | return gen_binary (inner_code, GET_MODE (x), | |
5862 | apply_distributive_law (tem), other); | |
5863 | } | |
5864 | \f | |
5865 | /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done | |
5866 | in MODE. | |
5867 | ||
5868 | Return an equivalent form, if different from X. Otherwise, return X. If | |
5869 | X is zero, we are to always construct the equivalent form. */ | |
5870 | ||
5871 | static rtx | |
5872 | simplify_and_const_int (x, mode, varop, constop) | |
5873 | rtx x; | |
5874 | enum machine_mode mode; | |
5875 | rtx varop; | |
5876 | unsigned HOST_WIDE_INT constop; | |
5877 | { | |
5878 | register enum machine_mode tmode; | |
5879 | register rtx temp; | |
5880 | unsigned HOST_WIDE_INT nonzero; | |
5881 | ||
5882 | /* There is a large class of optimizations based on the principle that | |
5883 | some operations produce results where certain bits are known to be zero, | |
5884 | and hence are not significant to the AND. For example, if we have just | |
5885 | done a left shift of one bit, the low-order bit is known to be zero and | |
5886 | hence an AND with a mask of ~1 would not do anything. | |
5887 | ||
5888 | At the end of the following loop, we set: | |
5889 | ||
5890 | VAROP to be the item to be AND'ed with; | |
5891 | CONSTOP to the constant value to AND it with. */ | |
5892 | ||
5893 | while (1) | |
5894 | { | |
5895 | /* If we ever encounter a mode wider than the host machine's widest | |
5896 | integer size, we can't compute the masks accurately, so give up. */ | |
5897 | if (GET_MODE_BITSIZE (GET_MODE (varop)) > HOST_BITS_PER_WIDE_INT) | |
5898 | break; | |
5899 | ||
5900 | /* Unless one of the cases below does a `continue', | |
5901 | a `break' will be executed to exit the loop. */ | |
5902 | ||
5903 | switch (GET_CODE (varop)) | |
5904 | { | |
5905 | case CLOBBER: | |
5906 | /* If VAROP is a (clobber (const_int)), return it since we know | |
5907 | we are generating something that won't match. */ | |
5908 | return varop; | |
5909 | ||
5910 | #if ! BITS_BIG_ENDIAN | |
5911 | case USE: | |
5912 | /* VAROP is a (use (mem ..)) that was made from a bit-field | |
5913 | extraction that spanned the boundary of the MEM. If we are | |
5914 | now masking so it is within that boundary, we don't need the | |
5915 | USE any more. */ | |
5916 | if ((constop & ~ GET_MODE_MASK (GET_MODE (XEXP (varop, 0)))) == 0) | |
5917 | { | |
5918 | varop = XEXP (varop, 0); | |
5919 | continue; | |
5920 | } | |
5921 | break; | |
5922 | #endif | |
5923 | ||
5924 | case SUBREG: | |
5925 | if (subreg_lowpart_p (varop) | |
5926 | /* We can ignore the effect this SUBREG if it narrows the mode | |
5927 | or, on machines where byte operations extend, if the | |
5928 | constant masks to zero all the bits the mode doesn't have. */ | |
5929 | && ((GET_MODE_SIZE (GET_MODE (varop)) | |
5930 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))) | |
5931 | #ifdef BYTE_LOADS_EXTEND | |
5932 | || (0 == (constop | |
5933 | & GET_MODE_MASK (GET_MODE (varop)) | |
5934 | & ~ GET_MODE_MASK (GET_MODE (SUBREG_REG (varop))))) | |
5935 | #endif | |
5936 | )) | |
5937 | { | |
5938 | varop = SUBREG_REG (varop); | |
5939 | continue; | |
5940 | } | |
5941 | break; | |
5942 | ||
5943 | case ZERO_EXTRACT: | |
5944 | case SIGN_EXTRACT: | |
5945 | case ZERO_EXTEND: | |
5946 | case SIGN_EXTEND: | |
5947 | /* Try to expand these into a series of shifts and then work | |
5948 | with that result. If we can't, for example, if the extract | |
5949 | isn't at a fixed position, give up. */ | |
5950 | temp = expand_compound_operation (varop); | |
5951 | if (temp != varop) | |
5952 | { | |
5953 | varop = temp; | |
5954 | continue; | |
5955 | } | |
5956 | break; | |
5957 | ||
5958 | case AND: | |
5959 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT) | |
5960 | { | |
5961 | constop &= INTVAL (XEXP (varop, 1)); | |
5962 | varop = XEXP (varop, 0); | |
5963 | continue; | |
5964 | } | |
5965 | break; | |
5966 | ||
5967 | case IOR: | |
5968 | case XOR: | |
5969 | /* If VAROP is (ior (lshiftrt FOO C1) C2), try to commute the IOR and | |
5970 | LSHIFT so we end up with an (and (lshiftrt (ior ...) ...) ...) | |
5971 | operation which may be a bitfield extraction. Ensure | |
5972 | that the constant we form is not wider than the mode of | |
5973 | VAROP. */ | |
5974 | ||
5975 | if (GET_CODE (XEXP (varop, 0)) == LSHIFTRT | |
5976 | && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT | |
5977 | && INTVAL (XEXP (XEXP (varop, 0), 1)) >= 0 | |
5978 | && INTVAL (XEXP (XEXP (varop, 0), 1)) < HOST_BITS_PER_WIDE_INT | |
5979 | && GET_CODE (XEXP (varop, 1)) == CONST_INT | |
5980 | && ((INTVAL (XEXP (XEXP (varop, 0), 1)) | |
5981 | + floor_log2 (INTVAL (XEXP (varop, 1)))) | |
5982 | < GET_MODE_BITSIZE (GET_MODE (varop))) | |
5983 | && (INTVAL (XEXP (varop, 1)) | |
5984 | & ~ nonzero_bits (XEXP (varop, 0), GET_MODE (varop)) == 0)) | |
5985 | { | |
5986 | temp = GEN_INT ((INTVAL (XEXP (varop, 1)) & constop) | |
5987 | << INTVAL (XEXP (XEXP (varop, 0), 1))); | |
5988 | temp = gen_binary (GET_CODE (varop), GET_MODE (varop), | |
5989 | XEXP (XEXP (varop, 0), 0), temp); | |
5990 | varop = gen_rtx_combine (LSHIFTRT, GET_MODE (varop), | |
5991 | temp, XEXP (varop, 1)); | |
5992 | continue; | |
5993 | } | |
5994 | ||
5995 | /* Apply the AND to both branches of the IOR or XOR, then try to | |
5996 | apply the distributive law. This may eliminate operations | |
5997 | if either branch can be simplified because of the AND. | |
5998 | It may also make some cases more complex, but those cases | |
5999 | probably won't match a pattern either with or without this. */ | |
6000 | return | |
6001 | gen_lowpart_for_combine | |
6002 | (mode, apply_distributive_law | |
6003 | (gen_rtx_combine | |
6004 | (GET_CODE (varop), GET_MODE (varop), | |
6005 | simplify_and_const_int (NULL_RTX, GET_MODE (varop), | |
6006 | XEXP (varop, 0), constop), | |
6007 | simplify_and_const_int (NULL_RTX, GET_MODE (varop), | |
6008 | XEXP (varop, 1), constop)))); | |
6009 | ||
6010 | case NOT: | |
6011 | /* (and (not FOO)) is (and (xor FOO CONST)), so if FOO is an | |
6012 | LSHIFTRT, we can do the same as above. Ensure that the constant | |
6013 | we form is not wider than the mode of VAROP. */ | |
6014 | ||
6015 | if (GET_CODE (XEXP (varop, 0)) == LSHIFTRT | |
6016 | && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT | |
6017 | && INTVAL (XEXP (XEXP (varop, 0), 1)) >= 0 | |
6018 | && (INTVAL (XEXP (XEXP (varop, 0), 1)) + floor_log2 (constop) | |
6019 | < GET_MODE_BITSIZE (GET_MODE (varop))) | |
6020 | && INTVAL (XEXP (XEXP (varop, 0), 1)) < HOST_BITS_PER_WIDE_INT) | |
6021 | { | |
6022 | temp = GEN_INT (constop << INTVAL (XEXP (XEXP (varop, 0), 1))); | |
6023 | temp = gen_binary (XOR, GET_MODE (varop), | |
6024 | XEXP (XEXP (varop, 0), 0), temp); | |
6025 | varop = gen_rtx_combine (LSHIFTRT, GET_MODE (varop), | |
6026 | temp, XEXP (XEXP (varop, 0), 1)); | |
6027 | continue; | |
6028 | } | |
6029 | break; | |
6030 | ||
6031 | case ASHIFTRT: | |
6032 | /* If we are just looking for the sign bit, we don't need this | |
6033 | shift at all, even if it has a variable count. */ | |
6034 | if (constop == ((HOST_WIDE_INT) 1 | |
6035 | << (GET_MODE_BITSIZE (GET_MODE (varop)) - 1))) | |
6036 | { | |
6037 | varop = XEXP (varop, 0); | |
6038 | continue; | |
6039 | } | |
6040 | ||
6041 | /* If this is a shift by a constant, get a mask that contains | |
6042 | those bits that are not copies of the sign bit. We then have | |
6043 | two cases: If CONSTOP only includes those bits, this can be | |
6044 | a logical shift, which may allow simplifications. If CONSTOP | |
6045 | is a single-bit field not within those bits, we are requesting | |
6046 | a copy of the sign bit and hence can shift the sign bit to | |
6047 | the appropriate location. */ | |
6048 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
6049 | && INTVAL (XEXP (varop, 1)) >= 0 | |
6050 | && INTVAL (XEXP (varop, 1)) < HOST_BITS_PER_WIDE_INT) | |
6051 | { | |
6052 | int i = -1; | |
6053 | ||
6054 | nonzero = GET_MODE_MASK (GET_MODE (varop)); | |
6055 | nonzero >>= INTVAL (XEXP (varop, 1)); | |
6056 | ||
6057 | if ((constop & ~ nonzero) == 0 | |
6058 | || (i = exact_log2 (constop)) >= 0) | |
6059 | { | |
6060 | varop = simplify_shift_const | |
6061 | (varop, LSHIFTRT, GET_MODE (varop), XEXP (varop, 0), | |
6062 | i < 0 ? INTVAL (XEXP (varop, 1)) | |
6063 | : GET_MODE_BITSIZE (GET_MODE (varop)) - 1 - i); | |
6064 | if (GET_CODE (varop) != ASHIFTRT) | |
6065 | continue; | |
6066 | } | |
6067 | } | |
6068 | ||
6069 | /* If our mask is 1, convert this to a LSHIFTRT. This can be done | |
6070 | even if the shift count isn't a constant. */ | |
6071 | if (constop == 1) | |
6072 | varop = gen_rtx_combine (LSHIFTRT, GET_MODE (varop), | |
6073 | XEXP (varop, 0), XEXP (varop, 1)); | |
6074 | break; | |
6075 | ||
6076 | case LSHIFTRT: | |
6077 | /* If we have (and (lshiftrt FOO C1) C2) where the combination of the | |
6078 | shift and AND produces only copies of the sign bit (C2 is one less | |
6079 | than a power of two), we can do this with just a shift. */ | |
6080 | ||
6081 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
6082 | && ((INTVAL (XEXP (varop, 1)) | |
6083 | + num_sign_bit_copies (XEXP (varop, 0), | |
6084 | GET_MODE (XEXP (varop, 0)))) | |
6085 | >= GET_MODE_BITSIZE (GET_MODE (varop))) | |
6086 | && exact_log2 (constop + 1) >= 0) | |
6087 | varop | |
6088 | = gen_rtx_combine (LSHIFTRT, GET_MODE (varop), XEXP (varop, 0), | |
6089 | GEN_INT (GET_MODE_BITSIZE (GET_MODE (varop)) | |
6090 | - exact_log2 (constop + 1))); | |
6091 | break; | |
6092 | ||
6093 | case NE: | |
6094 | /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is | |
6095 | included in STORE_FLAG_VALUE and FOO has no bits that might be | |
6096 | nonzero not in CONST. */ | |
6097 | if ((constop & ~ STORE_FLAG_VALUE) == 0 | |
6098 | && XEXP (varop, 0) == const0_rtx | |
6099 | && (nonzero_bits (XEXP (varop, 0), mode) & ~ constop) == 0) | |
6100 | { | |
6101 | varop = XEXP (varop, 0); | |
6102 | continue; | |
6103 | } | |
6104 | break; | |
6105 | ||
6106 | case PLUS: | |
6107 | /* In (and (plus FOO C1) M), if M is a mask that just turns off | |
6108 | low-order bits (as in an alignment operation) and FOO is already | |
6109 | aligned to that boundary, we can convert remove this AND | |
6110 | and possibly the PLUS if it is now adding zero. */ | |
6111 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
6112 | && exact_log2 (-constop) >= 0 | |
6113 | && (nonzero_bits (XEXP (varop, 0), mode) & ~ constop) == 0) | |
6114 | { | |
6115 | varop = plus_constant (XEXP (varop, 0), | |
6116 | INTVAL (XEXP (varop, 1)) & constop); | |
6117 | constop = ~0; | |
6118 | break; | |
6119 | } | |
6120 | ||
6121 | /* ... fall through ... */ | |
6122 | ||
6123 | case MINUS: | |
6124 | /* In (and (plus (and FOO M1) BAR) M2), if M1 and M2 are one | |
6125 | less than powers of two and M2 is narrower than M1, we can | |
6126 | eliminate the inner AND. This occurs when incrementing | |
6127 | bit fields. */ | |
6128 | ||
6129 | if (GET_CODE (XEXP (varop, 0)) == ZERO_EXTRACT | |
6130 | || GET_CODE (XEXP (varop, 0)) == ZERO_EXTEND) | |
6131 | SUBST (XEXP (varop, 0), | |
6132 | expand_compound_operation (XEXP (varop, 0))); | |
6133 | ||
6134 | if (GET_CODE (XEXP (varop, 0)) == AND | |
6135 | && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT | |
6136 | && exact_log2 (constop + 1) >= 0 | |
6137 | && exact_log2 (INTVAL (XEXP (XEXP (varop, 0), 1)) + 1) >= 0 | |
6138 | && (~ INTVAL (XEXP (XEXP (varop, 0), 1)) & constop) == 0) | |
6139 | SUBST (XEXP (varop, 0), XEXP (XEXP (varop, 0), 0)); | |
6140 | break; | |
6141 | } | |
6142 | ||
6143 | break; | |
6144 | } | |
6145 | ||
6146 | /* If we have reached a constant, this whole thing is constant. */ | |
6147 | if (GET_CODE (varop) == CONST_INT) | |
6148 | return GEN_INT (constop & INTVAL (varop)); | |
6149 | ||
6150 | /* See what bits may be nonzero in VAROP. Unlike the general case of | |
6151 | a call to nonzero_bits, here we don't care about bits outside | |
6152 | MODE. */ | |
6153 | ||
6154 | nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode); | |
6155 | ||
6156 | /* Turn off all bits in the constant that are known to already be zero. | |
6157 | Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS | |
6158 | which is tested below. */ | |
6159 | ||
6160 | constop &= nonzero; | |
6161 | ||
6162 | /* If we don't have any bits left, return zero. */ | |
6163 | if (constop == 0) | |
6164 | return const0_rtx; | |
6165 | ||
6166 | /* Get VAROP in MODE. Try to get a SUBREG if not. Don't make a new SUBREG | |
6167 | if we already had one (just check for the simplest cases). */ | |
6168 | if (x && GET_CODE (XEXP (x, 0)) == SUBREG | |
6169 | && GET_MODE (XEXP (x, 0)) == mode | |
6170 | && SUBREG_REG (XEXP (x, 0)) == varop) | |
6171 | varop = XEXP (x, 0); | |
6172 | else | |
6173 | varop = gen_lowpart_for_combine (mode, varop); | |
6174 | ||
6175 | /* If we can't make the SUBREG, try to return what we were given. */ | |
6176 | if (GET_CODE (varop) == CLOBBER) | |
6177 | return x ? x : varop; | |
6178 | ||
6179 | /* If we are only masking insignificant bits, return VAROP. */ | |
6180 | if (constop == nonzero) | |
6181 | x = varop; | |
6182 | ||
6183 | /* Otherwise, return an AND. See how much, if any, of X we can use. */ | |
6184 | else if (x == 0 || GET_CODE (x) != AND || GET_MODE (x) != mode) | |
6185 | x = gen_rtx_combine (AND, mode, varop, GEN_INT (constop)); | |
6186 | ||
6187 | else | |
6188 | { | |
6189 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
6190 | || INTVAL (XEXP (x, 1)) != constop) | |
6191 | SUBST (XEXP (x, 1), GEN_INT (constop)); | |
6192 | ||
6193 | SUBST (XEXP (x, 0), varop); | |
6194 | } | |
6195 | ||
6196 | return x; | |
6197 | } | |
6198 | \f | |
6199 | /* Given an expression, X, compute which bits in X can be non-zero. | |
6200 | We don't care about bits outside of those defined in MODE. | |
6201 | ||
6202 | For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is | |
6203 | a shift, AND, or zero_extract, we can do better. */ | |
6204 | ||
6205 | static unsigned HOST_WIDE_INT | |
6206 | nonzero_bits (x, mode) | |
6207 | rtx x; | |
6208 | enum machine_mode mode; | |
6209 | { | |
6210 | unsigned HOST_WIDE_INT nonzero = GET_MODE_MASK (mode); | |
6211 | unsigned HOST_WIDE_INT inner_nz; | |
6212 | enum rtx_code code; | |
6213 | int mode_width = GET_MODE_BITSIZE (mode); | |
6214 | rtx tem; | |
6215 | ||
6216 | /* If X is wider than MODE, use its mode instead. */ | |
6217 | if (GET_MODE_BITSIZE (GET_MODE (x)) > mode_width) | |
6218 | { | |
6219 | mode = GET_MODE (x); | |
6220 | nonzero = GET_MODE_MASK (mode); | |
6221 | mode_width = GET_MODE_BITSIZE (mode); | |
6222 | } | |
6223 | ||
6224 | if (mode_width > HOST_BITS_PER_WIDE_INT) | |
6225 | /* Our only callers in this case look for single bit values. So | |
6226 | just return the mode mask. Those tests will then be false. */ | |
6227 | return nonzero; | |
6228 | ||
6229 | code = GET_CODE (x); | |
6230 | switch (code) | |
6231 | { | |
6232 | case REG: | |
6233 | #ifdef STACK_BOUNDARY | |
6234 | /* If this is the stack pointer, we may know something about its | |
6235 | alignment. If PUSH_ROUNDING is defined, it is possible for the | |
6236 | stack to be momentarily aligned only to that amount, so we pick | |
6237 | the least alignment. */ | |
6238 | ||
6239 | if (x == stack_pointer_rtx) | |
6240 | { | |
6241 | int sp_alignment = STACK_BOUNDARY / BITS_PER_UNIT; | |
6242 | ||
6243 | #ifdef PUSH_ROUNDING | |
6244 | sp_alignment = MIN (PUSH_ROUNDING (1), sp_alignment); | |
6245 | #endif | |
6246 | ||
6247 | return nonzero & ~ (sp_alignment - 1); | |
6248 | } | |
6249 | #endif | |
6250 | ||
6251 | /* If X is a register whose nonzero bits value is current, use it. | |
6252 | Otherwise, if X is a register whose value we can find, use that | |
6253 | value. Otherwise, use the previously-computed global nonzero bits | |
6254 | for this register. */ | |
6255 | ||
6256 | if (reg_last_set_value[REGNO (x)] != 0 | |
6257 | && reg_last_set_mode[REGNO (x)] == mode | |
6258 | && (reg_n_sets[REGNO (x)] == 1 | |
6259 | || reg_last_set_label[REGNO (x)] == label_tick) | |
6260 | && INSN_CUID (reg_last_set[REGNO (x)]) < subst_low_cuid) | |
6261 | return reg_last_set_nonzero_bits[REGNO (x)]; | |
6262 | ||
6263 | tem = get_last_value (x); | |
6264 | ||
6265 | if (tem) | |
6266 | { | |
6267 | #ifdef SHORT_IMMEDIATES_SIGN_EXTEND | |
6268 | /* If X is narrower than MODE and TEM is a non-negative | |
6269 | constant that would appear negative in the mode of X, | |
6270 | sign-extend it for use in reg_nonzero_bits because some | |
6271 | machines (maybe most) will actually do the sign-extension | |
6272 | and this is the conservative approach. | |
6273 | ||
6274 | ??? For 2.5, try to tighten up the MD files in this regard | |
6275 | instead of this kludge. */ | |
6276 | ||
6277 | if (GET_MODE_BITSIZE (GET_MODE (x)) < mode_width | |
6278 | && GET_CODE (tem) == CONST_INT | |
6279 | && INTVAL (tem) > 0 | |
6280 | && 0 != (INTVAL (tem) | |
6281 | & ((HOST_WIDE_INT) 1 | |
6282 | << GET_MODE_BITSIZE (GET_MODE (x))))) | |
6283 | tem = GEN_INT (INTVAL (tem) | |
6284 | | ((HOST_WIDE_INT) (-1) | |
6285 | << GET_MODE_BITSIZE (GET_MODE (x)))); | |
6286 | #endif | |
6287 | return nonzero_bits (tem, mode); | |
6288 | } | |
6289 | else if (nonzero_sign_valid && reg_nonzero_bits[REGNO (x)]) | |
6290 | return reg_nonzero_bits[REGNO (x)] & nonzero; | |
6291 | else | |
6292 | return nonzero; | |
6293 | ||
6294 | case CONST_INT: | |
6295 | #ifdef SHORT_IMMEDIATES_SIGN_EXTEND | |
6296 | /* If X is negative in MODE, sign-extend the value. */ | |
6297 | if (INTVAL (x) > 0 | |
6298 | && 0 != (INTVAL (x) | |
6299 | & ((HOST_WIDE_INT) 1 << GET_MODE_BITSIZE (GET_MODE (x))))) | |
6300 | return (INTVAL (x) | |
6301 | | ((HOST_WIDE_INT) (-1) << GET_MODE_BITSIZE (GET_MODE (x)))); | |
6302 | #endif | |
6303 | ||
6304 | return INTVAL (x); | |
6305 | ||
6306 | #ifdef BYTE_LOADS_ZERO_EXTEND | |
6307 | case MEM: | |
6308 | /* In many, if not most, RISC machines, reading a byte from memory | |
6309 | zeros the rest of the register. Noticing that fact saves a lot | |
6310 | of extra zero-extends. */ | |
6311 | nonzero &= GET_MODE_MASK (GET_MODE (x)); | |
6312 | break; | |
6313 | #endif | |
6314 | ||
6315 | #if STORE_FLAG_VALUE == 1 | |
6316 | case EQ: case NE: | |
6317 | case GT: case GTU: | |
6318 | case LT: case LTU: | |
6319 | case GE: case GEU: | |
6320 | case LE: case LEU: | |
6321 | ||
6322 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
6323 | nonzero = 1; | |
6324 | ||
6325 | /* A comparison operation only sets the bits given by its mode. The | |
6326 | rest are set undefined. */ | |
6327 | if (GET_MODE_SIZE (GET_MODE (x)) < mode_width) | |
6328 | nonzero |= (GET_MODE_MASK (mode) & ~ GET_MODE_MASK (GET_MODE (x))); | |
6329 | break; | |
6330 | #endif | |
6331 | ||
6332 | case NEG: | |
6333 | if (num_sign_bit_copies (XEXP (x, 0), GET_MODE (x)) | |
6334 | == GET_MODE_BITSIZE (GET_MODE (x))) | |
6335 | nonzero = 1; | |
6336 | ||
6337 | if (GET_MODE_SIZE (GET_MODE (x)) < mode_width) | |
6338 | nonzero |= (GET_MODE_MASK (mode) & ~ GET_MODE_MASK (GET_MODE (x))); | |
6339 | break; | |
6340 | ||
6341 | case ABS: | |
6342 | if (num_sign_bit_copies (XEXP (x, 0), GET_MODE (x)) | |
6343 | == GET_MODE_BITSIZE (GET_MODE (x))) | |
6344 | nonzero = 1; | |
6345 | break; | |
6346 | ||
6347 | case TRUNCATE: | |
6348 | nonzero &= (nonzero_bits (XEXP (x, 0), mode) & GET_MODE_MASK (mode)); | |
6349 | break; | |
6350 | ||
6351 | case ZERO_EXTEND: | |
6352 | nonzero &= nonzero_bits (XEXP (x, 0), mode); | |
6353 | if (GET_MODE (XEXP (x, 0)) != VOIDmode) | |
6354 | nonzero &= GET_MODE_MASK (GET_MODE (XEXP (x, 0))); | |
6355 | break; | |
6356 | ||
6357 | case SIGN_EXTEND: | |
6358 | /* If the sign bit is known clear, this is the same as ZERO_EXTEND. | |
6359 | Otherwise, show all the bits in the outer mode but not the inner | |
6360 | may be non-zero. */ | |
6361 | inner_nz = nonzero_bits (XEXP (x, 0), mode); | |
6362 | if (GET_MODE (XEXP (x, 0)) != VOIDmode) | |
6363 | { | |
6364 | inner_nz &= GET_MODE_MASK (GET_MODE (XEXP (x, 0))); | |
6365 | if (inner_nz & | |
6366 | (((HOST_WIDE_INT) 1 | |
6367 | << (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - 1)))) | |
6368 | inner_nz |= (GET_MODE_MASK (mode) | |
6369 | & ~ GET_MODE_MASK (GET_MODE (XEXP (x, 0)))); | |
6370 | } | |
6371 | ||
6372 | nonzero &= inner_nz; | |
6373 | break; | |
6374 | ||
6375 | case AND: | |
6376 | nonzero &= (nonzero_bits (XEXP (x, 0), mode) | |
6377 | & nonzero_bits (XEXP (x, 1), mode)); | |
6378 | break; | |
6379 | ||
6380 | case XOR: case IOR: | |
6381 | case UMIN: case UMAX: case SMIN: case SMAX: | |
6382 | nonzero &= (nonzero_bits (XEXP (x, 0), mode) | |
6383 | | nonzero_bits (XEXP (x, 1), mode)); | |
6384 | break; | |
6385 | ||
6386 | case PLUS: case MINUS: | |
6387 | case MULT: | |
6388 | case DIV: case UDIV: | |
6389 | case MOD: case UMOD: | |
6390 | /* We can apply the rules of arithmetic to compute the number of | |
6391 | high- and low-order zero bits of these operations. We start by | |
6392 | computing the width (position of the highest-order non-zero bit) | |
6393 | and the number of low-order zero bits for each value. */ | |
6394 | { | |
6395 | unsigned HOST_WIDE_INT nz0 = nonzero_bits (XEXP (x, 0), mode); | |
6396 | unsigned HOST_WIDE_INT nz1 = nonzero_bits (XEXP (x, 1), mode); | |
6397 | int width0 = floor_log2 (nz0) + 1; | |
6398 | int width1 = floor_log2 (nz1) + 1; | |
6399 | int low0 = floor_log2 (nz0 & -nz0); | |
6400 | int low1 = floor_log2 (nz1 & -nz1); | |
6401 | int op0_maybe_minusp = (nz0 & ((HOST_WIDE_INT) 1 << (mode_width - 1))); | |
6402 | int op1_maybe_minusp = (nz1 & ((HOST_WIDE_INT) 1 << (mode_width - 1))); | |
6403 | int result_width = mode_width; | |
6404 | int result_low = 0; | |
6405 | ||
6406 | switch (code) | |
6407 | { | |
6408 | case PLUS: | |
6409 | result_width = MAX (width0, width1) + 1; | |
6410 | result_low = MIN (low0, low1); | |
6411 | break; | |
6412 | case MINUS: | |
6413 | result_low = MIN (low0, low1); | |
6414 | break; | |
6415 | case MULT: | |
6416 | result_width = width0 + width1; | |
6417 | result_low = low0 + low1; | |
6418 | break; | |
6419 | case DIV: | |
6420 | if (! op0_maybe_minusp && ! op1_maybe_minusp) | |
6421 | result_width = width0; | |
6422 | break; | |
6423 | case UDIV: | |
6424 | result_width = width0; | |
6425 | break; | |
6426 | case MOD: | |
6427 | if (! op0_maybe_minusp && ! op1_maybe_minusp) | |
6428 | result_width = MIN (width0, width1); | |
6429 | result_low = MIN (low0, low1); | |
6430 | break; | |
6431 | case UMOD: | |
6432 | result_width = MIN (width0, width1); | |
6433 | result_low = MIN (low0, low1); | |
6434 | break; | |
6435 | } | |
6436 | ||
6437 | if (result_width < mode_width) | |
6438 | nonzero &= ((HOST_WIDE_INT) 1 << result_width) - 1; | |
6439 | ||
6440 | if (result_low > 0) | |
6441 | nonzero &= ~ (((HOST_WIDE_INT) 1 << result_low) - 1); | |
6442 | } | |
6443 | break; | |
6444 | ||
6445 | case ZERO_EXTRACT: | |
6446 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6447 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT) | |
6448 | nonzero &= ((HOST_WIDE_INT) 1 << INTVAL (XEXP (x, 1))) - 1; | |
6449 | break; | |
6450 | ||
6451 | case SUBREG: | |
6452 | /* If this is a SUBREG formed for a promoted variable that has | |
6453 | been zero-extended, we know that at least the high-order bits | |
6454 | are zero, though others might be too. */ | |
6455 | ||
6456 | if (SUBREG_PROMOTED_VAR_P (x) && SUBREG_PROMOTED_UNSIGNED_P (x)) | |
6457 | nonzero = (GET_MODE_MASK (GET_MODE (x)) | |
6458 | & nonzero_bits (SUBREG_REG (x), GET_MODE (x))); | |
6459 | ||
6460 | /* If the inner mode is a single word for both the host and target | |
6461 | machines, we can compute this from which bits of the inner | |
6462 | object might be nonzero. */ | |
6463 | if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) <= BITS_PER_WORD | |
6464 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) | |
6465 | <= HOST_BITS_PER_WIDE_INT)) | |
6466 | { | |
6467 | nonzero &= nonzero_bits (SUBREG_REG (x), mode); | |
6468 | #ifndef BYTE_LOADS_EXTEND | |
6469 | /* On many CISC machines, accessing an object in a wider mode | |
6470 | causes the high-order bits to become undefined. So they are | |
6471 | not known to be zero. */ | |
6472 | if (GET_MODE_SIZE (GET_MODE (x)) | |
6473 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
6474 | nonzero |= (GET_MODE_MASK (GET_MODE (x)) | |
6475 | & ~ GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))); | |
6476 | #endif | |
6477 | } | |
6478 | break; | |
6479 | ||
6480 | case ASHIFTRT: | |
6481 | case LSHIFTRT: | |
6482 | case ASHIFT: | |
6483 | case LSHIFT: | |
6484 | case ROTATE: | |
6485 | /* The nonzero bits are in two classes: any bits within MODE | |
6486 | that aren't in GET_MODE (x) are always significant. The rest of the | |
6487 | nonzero bits are those that are significant in the operand of | |
6488 | the shift when shifted the appropriate number of bits. This | |
6489 | shows that high-order bits are cleared by the right shift and | |
6490 | low-order bits by left shifts. */ | |
6491 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6492 | && INTVAL (XEXP (x, 1)) >= 0 | |
6493 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT) | |
6494 | { | |
6495 | enum machine_mode inner_mode = GET_MODE (x); | |
6496 | int width = GET_MODE_BITSIZE (inner_mode); | |
6497 | int count = INTVAL (XEXP (x, 1)); | |
6498 | unsigned HOST_WIDE_INT mode_mask = GET_MODE_MASK (inner_mode); | |
6499 | unsigned HOST_WIDE_INT op_nonzero = nonzero_bits (XEXP (x, 0), mode); | |
6500 | unsigned HOST_WIDE_INT inner = op_nonzero & mode_mask; | |
6501 | unsigned HOST_WIDE_INT outer = 0; | |
6502 | ||
6503 | if (mode_width > width) | |
6504 | outer = (op_nonzero & nonzero & ~ mode_mask); | |
6505 | ||
6506 | if (code == LSHIFTRT) | |
6507 | inner >>= count; | |
6508 | else if (code == ASHIFTRT) | |
6509 | { | |
6510 | inner >>= count; | |
6511 | ||
6512 | /* If the sign bit may have been nonzero before the shift, we | |
6513 | need to mark all the places it could have been copied to | |
6514 | by the shift as possibly nonzero. */ | |
6515 | if (inner & ((HOST_WIDE_INT) 1 << (width - 1 - count))) | |
6516 | inner |= (((HOST_WIDE_INT) 1 << count) - 1) << (width - count); | |
6517 | } | |
6518 | else if (code == LSHIFT || code == ASHIFT) | |
6519 | inner <<= count; | |
6520 | else | |
6521 | inner = ((inner << (count % width) | |
6522 | | (inner >> (width - (count % width)))) & mode_mask); | |
6523 | ||
6524 | nonzero &= (outer | inner); | |
6525 | } | |
6526 | break; | |
6527 | ||
6528 | case FFS: | |
6529 | /* This is at most the number of bits in the mode. */ | |
6530 | nonzero = ((HOST_WIDE_INT) 1 << (floor_log2 (mode_width) + 1)) - 1; | |
6531 | break; | |
6532 | ||
6533 | case IF_THEN_ELSE: | |
6534 | nonzero &= (nonzero_bits (XEXP (x, 1), mode) | |
6535 | | nonzero_bits (XEXP (x, 2), mode)); | |
6536 | break; | |
6537 | } | |
6538 | ||
6539 | return nonzero; | |
6540 | } | |
6541 | \f | |
6542 | /* Return the number of bits at the high-order end of X that are known to | |
6543 | be equal to the sign bit. This number will always be between 1 and | |
6544 | the number of bits in the mode of X. MODE is the mode to be used | |
6545 | if X is VOIDmode. */ | |
6546 | ||
6547 | static int | |
6548 | num_sign_bit_copies (x, mode) | |
6549 | rtx x; | |
6550 | enum machine_mode mode; | |
6551 | { | |
6552 | enum rtx_code code = GET_CODE (x); | |
6553 | int bitwidth; | |
6554 | int num0, num1, result; | |
6555 | unsigned HOST_WIDE_INT nonzero; | |
6556 | rtx tem; | |
6557 | ||
6558 | /* If we weren't given a mode, use the mode of X. If the mode is still | |
6559 | VOIDmode, we don't know anything. */ | |
6560 | ||
6561 | if (mode == VOIDmode) | |
6562 | mode = GET_MODE (x); | |
6563 | ||
6564 | if (mode == VOIDmode) | |
6565 | return 1; | |
6566 | ||
6567 | bitwidth = GET_MODE_BITSIZE (mode); | |
6568 | ||
6569 | switch (code) | |
6570 | { | |
6571 | case REG: | |
6572 | ||
6573 | if (reg_last_set_value[REGNO (x)] != 0 | |
6574 | && reg_last_set_mode[REGNO (x)] == mode | |
6575 | && (reg_n_sets[REGNO (x)] == 1 | |
6576 | || reg_last_set_label[REGNO (x)] == label_tick) | |
6577 | && INSN_CUID (reg_last_set[REGNO (x)]) < subst_low_cuid) | |
6578 | return reg_last_set_sign_bit_copies[REGNO (x)]; | |
6579 | ||
6580 | tem = get_last_value (x); | |
6581 | if (tem != 0) | |
6582 | return num_sign_bit_copies (tem, mode); | |
6583 | ||
6584 | if (nonzero_sign_valid && reg_sign_bit_copies[REGNO (x)] != 0) | |
6585 | return reg_sign_bit_copies[REGNO (x)]; | |
6586 | break; | |
6587 | ||
6588 | #ifdef BYTE_LOADS_SIGN_EXTEND | |
6589 | case MEM: | |
6590 | /* Some RISC machines sign-extend all loads of smaller than a word. */ | |
6591 | return MAX (1, bitwidth - GET_MODE_BITSIZE (GET_MODE (x)) + 1); | |
6592 | #endif | |
6593 | ||
6594 | case CONST_INT: | |
6595 | /* If the constant is negative, take its 1's complement and remask. | |
6596 | Then see how many zero bits we have. */ | |
6597 | nonzero = INTVAL (x) & GET_MODE_MASK (mode); | |
6598 | if (bitwidth <= HOST_BITS_PER_WIDE_INT | |
6599 | && (nonzero & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) | |
6600 | nonzero = (~ nonzero) & GET_MODE_MASK (mode); | |
6601 | ||
6602 | return (nonzero == 0 ? bitwidth : bitwidth - floor_log2 (nonzero) - 1); | |
6603 | ||
6604 | case SUBREG: | |
6605 | /* If this is a SUBREG for a promoted object that is sign-extended | |
6606 | and we are looking at it in a wider mode, we know that at least the | |
6607 | high-order bits are known to be sign bit copies. */ | |
6608 | ||
6609 | if (SUBREG_PROMOTED_VAR_P (x) && ! SUBREG_PROMOTED_UNSIGNED_P (x)) | |
6610 | return MAX (bitwidth - GET_MODE_BITSIZE (GET_MODE (x)) + 1, | |
6611 | num_sign_bit_copies (SUBREG_REG (x), mode)); | |
6612 | ||
6613 | /* For a smaller object, just ignore the high bits. */ | |
6614 | if (bitwidth <= GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))) | |
6615 | { | |
6616 | num0 = num_sign_bit_copies (SUBREG_REG (x), VOIDmode); | |
6617 | return MAX (1, (num0 | |
6618 | - (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) | |
6619 | - bitwidth))); | |
6620 | } | |
6621 | ||
6622 | #ifdef BYTE_LOADS_EXTEND | |
6623 | /* For paradoxical SUBREGs, just look inside since, on machines with | |
6624 | one of these defined, we assume that operations are actually | |
6625 | performed on the full register. Note that we are passing MODE | |
6626 | to the recursive call, so the number of sign bit copies will | |
6627 | remain relative to that mode, not the inner mode. */ | |
6628 | ||
6629 | if (GET_MODE_SIZE (GET_MODE (x)) | |
6630 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
6631 | return num_sign_bit_copies (SUBREG_REG (x), mode); | |
6632 | #endif | |
6633 | ||
6634 | break; | |
6635 | ||
6636 | case SIGN_EXTRACT: | |
6637 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
6638 | return MAX (1, bitwidth - INTVAL (XEXP (x, 1))); | |
6639 | break; | |
6640 | ||
6641 | case SIGN_EXTEND: | |
6642 | return (bitwidth - GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
6643 | + num_sign_bit_copies (XEXP (x, 0), VOIDmode)); | |
6644 | ||
6645 | case TRUNCATE: | |
6646 | /* For a smaller object, just ignore the high bits. */ | |
6647 | num0 = num_sign_bit_copies (XEXP (x, 0), VOIDmode); | |
6648 | return MAX (1, (num0 - (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
6649 | - bitwidth))); | |
6650 | ||
6651 | case NOT: | |
6652 | return num_sign_bit_copies (XEXP (x, 0), mode); | |
6653 | ||
6654 | case ROTATE: case ROTATERT: | |
6655 | /* If we are rotating left by a number of bits less than the number | |
6656 | of sign bit copies, we can just subtract that amount from the | |
6657 | number. */ | |
6658 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6659 | && INTVAL (XEXP (x, 1)) >= 0 && INTVAL (XEXP (x, 1)) < bitwidth) | |
6660 | { | |
6661 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6662 | return MAX (1, num0 - (code == ROTATE ? INTVAL (XEXP (x, 1)) | |
6663 | : bitwidth - INTVAL (XEXP (x, 1)))); | |
6664 | } | |
6665 | break; | |
6666 | ||
6667 | case NEG: | |
6668 | /* In general, this subtracts one sign bit copy. But if the value | |
6669 | is known to be positive, the number of sign bit copies is the | |
6670 | same as that of the input. Finally, if the input has just one bit | |
6671 | that might be nonzero, all the bits are copies of the sign bit. */ | |
6672 | nonzero = nonzero_bits (XEXP (x, 0), mode); | |
6673 | if (nonzero == 1) | |
6674 | return bitwidth; | |
6675 | ||
6676 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6677 | if (num0 > 1 | |
6678 | && bitwidth <= HOST_BITS_PER_WIDE_INT | |
6679 | && (((HOST_WIDE_INT) 1 << (bitwidth - 1)) & nonzero)) | |
6680 | num0--; | |
6681 | ||
6682 | return num0; | |
6683 | ||
6684 | case IOR: case AND: case XOR: | |
6685 | case SMIN: case SMAX: case UMIN: case UMAX: | |
6686 | /* Logical operations will preserve the number of sign-bit copies. | |
6687 | MIN and MAX operations always return one of the operands. */ | |
6688 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6689 | num1 = num_sign_bit_copies (XEXP (x, 1), mode); | |
6690 | return MIN (num0, num1); | |
6691 | ||
6692 | case PLUS: case MINUS: | |
6693 | /* For addition and subtraction, we can have a 1-bit carry. However, | |
6694 | if we are subtracting 1 from a positive number, there will not | |
6695 | be such a carry. Furthermore, if the positive number is known to | |
6696 | be 0 or 1, we know the result is either -1 or 0. */ | |
6697 | ||
6698 | if (code == PLUS && XEXP (x, 1) == constm1_rtx | |
6699 | && bitwidth <= HOST_BITS_PER_WIDE_INT) | |
6700 | { | |
6701 | nonzero = nonzero_bits (XEXP (x, 0), mode); | |
6702 | if ((((HOST_WIDE_INT) 1 << (bitwidth - 1)) & nonzero) == 0) | |
6703 | return (nonzero == 1 || nonzero == 0 ? bitwidth | |
6704 | : bitwidth - floor_log2 (nonzero) - 1); | |
6705 | } | |
6706 | ||
6707 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6708 | num1 = num_sign_bit_copies (XEXP (x, 1), mode); | |
6709 | return MAX (1, MIN (num0, num1) - 1); | |
6710 | ||
6711 | case MULT: | |
6712 | /* The number of bits of the product is the sum of the number of | |
6713 | bits of both terms. However, unless one of the terms if known | |
6714 | to be positive, we must allow for an additional bit since negating | |
6715 | a negative number can remove one sign bit copy. */ | |
6716 | ||
6717 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6718 | num1 = num_sign_bit_copies (XEXP (x, 1), mode); | |
6719 | ||
6720 | result = bitwidth - (bitwidth - num0) - (bitwidth - num1); | |
6721 | if (result > 0 | |
6722 | && bitwidth <= HOST_BITS_PER_WIDE_INT | |
6723 | && ((nonzero_bits (XEXP (x, 0), mode) | |
6724 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) | |
6725 | && (nonzero_bits (XEXP (x, 1), mode) | |
6726 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1)) != 0)) | |
6727 | result--; | |
6728 | ||
6729 | return MAX (1, result); | |
6730 | ||
6731 | case UDIV: | |
6732 | /* The result must be <= the first operand. */ | |
6733 | return num_sign_bit_copies (XEXP (x, 0), mode); | |
6734 | ||
6735 | case UMOD: | |
6736 | /* The result must be <= the scond operand. */ | |
6737 | return num_sign_bit_copies (XEXP (x, 1), mode); | |
6738 | ||
6739 | case DIV: | |
6740 | /* Similar to unsigned division, except that we have to worry about | |
6741 | the case where the divisor is negative, in which case we have | |
6742 | to add 1. */ | |
6743 | result = num_sign_bit_copies (XEXP (x, 0), mode); | |
6744 | if (result > 1 | |
6745 | && bitwidth <= HOST_BITS_PER_WIDE_INT | |
6746 | && (nonzero_bits (XEXP (x, 1), mode) | |
6747 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) | |
6748 | result --; | |
6749 | ||
6750 | return result; | |
6751 | ||
6752 | case MOD: | |
6753 | result = num_sign_bit_copies (XEXP (x, 1), mode); | |
6754 | if (result > 1 | |
6755 | && bitwidth <= HOST_BITS_PER_WIDE_INT | |
6756 | && (nonzero_bits (XEXP (x, 1), mode) | |
6757 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) | |
6758 | result --; | |
6759 | ||
6760 | return result; | |
6761 | ||
6762 | case ASHIFTRT: | |
6763 | /* Shifts by a constant add to the number of bits equal to the | |
6764 | sign bit. */ | |
6765 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6766 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6767 | && INTVAL (XEXP (x, 1)) > 0) | |
6768 | num0 = MIN (bitwidth, num0 + INTVAL (XEXP (x, 1))); | |
6769 | ||
6770 | return num0; | |
6771 | ||
6772 | case ASHIFT: | |
6773 | case LSHIFT: | |
6774 | /* Left shifts destroy copies. */ | |
6775 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
6776 | || INTVAL (XEXP (x, 1)) < 0 | |
6777 | || INTVAL (XEXP (x, 1)) >= bitwidth) | |
6778 | return 1; | |
6779 | ||
6780 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
6781 | return MAX (1, num0 - INTVAL (XEXP (x, 1))); | |
6782 | ||
6783 | case IF_THEN_ELSE: | |
6784 | num0 = num_sign_bit_copies (XEXP (x, 1), mode); | |
6785 | num1 = num_sign_bit_copies (XEXP (x, 2), mode); | |
6786 | return MIN (num0, num1); | |
6787 | ||
6788 | #if STORE_FLAG_VALUE == -1 | |
6789 | case EQ: case NE: case GE: case GT: case LE: case LT: | |
6790 | case GEU: case GTU: case LEU: case LTU: | |
6791 | return bitwidth; | |
6792 | #endif | |
6793 | } | |
6794 | ||
6795 | /* If we haven't been able to figure it out by one of the above rules, | |
6796 | see if some of the high-order bits are known to be zero. If so, | |
6797 | count those bits and return one less than that amount. If we can't | |
6798 | safely compute the mask for this mode, always return BITWIDTH. */ | |
6799 | ||
6800 | if (bitwidth > HOST_BITS_PER_WIDE_INT) | |
6801 | return 1; | |
6802 | ||
6803 | nonzero = nonzero_bits (x, mode); | |
2a5f595d | 6804 | return (nonzero & ((HOST_WIDE_INT) 1 << (bitwidth - 1)) |
9bf86ebb PR |
6805 | ? 1 : bitwidth - floor_log2 (nonzero) - 1); |
6806 | } | |
6807 | \f | |
6808 | /* Return the number of "extended" bits there are in X, when interpreted | |
6809 | as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For | |
6810 | unsigned quantities, this is the number of high-order zero bits. | |
6811 | For signed quantities, this is the number of copies of the sign bit | |
6812 | minus 1. In both case, this function returns the number of "spare" | |
6813 | bits. For example, if two quantities for which this function returns | |
6814 | at least 1 are added, the addition is known not to overflow. | |
6815 | ||
6816 | This function will always return 0 unless called during combine, which | |
6817 | implies that it must be called from a define_split. */ | |
6818 | ||
6819 | int | |
6820 | extended_count (x, mode, unsignedp) | |
6821 | rtx x; | |
6822 | enum machine_mode mode; | |
6823 | int unsignedp; | |
6824 | { | |
6825 | if (nonzero_sign_valid == 0) | |
6826 | return 0; | |
6827 | ||
6828 | return (unsignedp | |
6829 | ? (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
6830 | && (GET_MODE_BITSIZE (mode) - 1 | |
6831 | - floor_log2 (nonzero_bits (x, mode)))) | |
6832 | : num_sign_bit_copies (x, mode) - 1); | |
6833 | } | |
6834 | \f | |
6835 | /* This function is called from `simplify_shift_const' to merge two | |
6836 | outer operations. Specifically, we have already found that we need | |
6837 | to perform operation *POP0 with constant *PCONST0 at the outermost | |
6838 | position. We would now like to also perform OP1 with constant CONST1 | |
6839 | (with *POP0 being done last). | |
6840 | ||
6841 | Return 1 if we can do the operation and update *POP0 and *PCONST0 with | |
6842 | the resulting operation. *PCOMP_P is set to 1 if we would need to | |
6843 | complement the innermost operand, otherwise it is unchanged. | |
6844 | ||
6845 | MODE is the mode in which the operation will be done. No bits outside | |
6846 | the width of this mode matter. It is assumed that the width of this mode | |
6847 | is smaller than or equal to HOST_BITS_PER_WIDE_INT. | |
6848 | ||
6849 | If *POP0 or OP1 are NIL, it means no operation is required. Only NEG, PLUS, | |
6850 | IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper | |
6851 | result is simply *PCONST0. | |
6852 | ||
6853 | If the resulting operation cannot be expressed as one operation, we | |
6854 | return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */ | |
6855 | ||
6856 | static int | |
6857 | merge_outer_ops (pop0, pconst0, op1, const1, mode, pcomp_p) | |
6858 | enum rtx_code *pop0; | |
6859 | HOST_WIDE_INT *pconst0; | |
6860 | enum rtx_code op1; | |
6861 | HOST_WIDE_INT const1; | |
6862 | enum machine_mode mode; | |
6863 | int *pcomp_p; | |
6864 | { | |
6865 | enum rtx_code op0 = *pop0; | |
6866 | HOST_WIDE_INT const0 = *pconst0; | |
6867 | ||
6868 | const0 &= GET_MODE_MASK (mode); | |
6869 | const1 &= GET_MODE_MASK (mode); | |
6870 | ||
6871 | /* If OP0 is an AND, clear unimportant bits in CONST1. */ | |
6872 | if (op0 == AND) | |
6873 | const1 &= const0; | |
6874 | ||
6875 | /* If OP0 or OP1 is NIL, this is easy. Similarly if they are the same or | |
6876 | if OP0 is SET. */ | |
6877 | ||
6878 | if (op1 == NIL || op0 == SET) | |
6879 | return 1; | |
6880 | ||
6881 | else if (op0 == NIL) | |
6882 | op0 = op1, const0 = const1; | |
6883 | ||
6884 | else if (op0 == op1) | |
6885 | { | |
6886 | switch (op0) | |
6887 | { | |
6888 | case AND: | |
6889 | const0 &= const1; | |
6890 | break; | |
6891 | case IOR: | |
6892 | const0 |= const1; | |
6893 | break; | |
6894 | case XOR: | |
6895 | const0 ^= const1; | |
6896 | break; | |
6897 | case PLUS: | |
6898 | const0 += const1; | |
6899 | break; | |
6900 | case NEG: | |
6901 | op0 = NIL; | |
6902 | break; | |
6903 | } | |
6904 | } | |
6905 | ||
6906 | /* Otherwise, if either is a PLUS or NEG, we can't do anything. */ | |
6907 | else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG) | |
6908 | return 0; | |
6909 | ||
6910 | /* If the two constants aren't the same, we can't do anything. The | |
6911 | remaining six cases can all be done. */ | |
6912 | else if (const0 != const1) | |
6913 | return 0; | |
6914 | ||
6915 | else | |
6916 | switch (op0) | |
6917 | { | |
6918 | case IOR: | |
6919 | if (op1 == AND) | |
6920 | /* (a & b) | b == b */ | |
6921 | op0 = SET; | |
6922 | else /* op1 == XOR */ | |
6923 | /* (a ^ b) | b == a | b */ | |
6924 | ; | |
6925 | break; | |
6926 | ||
6927 | case XOR: | |
6928 | if (op1 == AND) | |
6929 | /* (a & b) ^ b == (~a) & b */ | |
6930 | op0 = AND, *pcomp_p = 1; | |
6931 | else /* op1 == IOR */ | |
6932 | /* (a | b) ^ b == a & ~b */ | |
6933 | op0 = AND, *pconst0 = ~ const0; | |
6934 | break; | |
6935 | ||
6936 | case AND: | |
6937 | if (op1 == IOR) | |
6938 | /* (a | b) & b == b */ | |
6939 | op0 = SET; | |
6940 | else /* op1 == XOR */ | |
6941 | /* (a ^ b) & b) == (~a) & b */ | |
6942 | *pcomp_p = 1; | |
6943 | break; | |
6944 | } | |
6945 | ||
6946 | /* Check for NO-OP cases. */ | |
6947 | const0 &= GET_MODE_MASK (mode); | |
6948 | if (const0 == 0 | |
6949 | && (op0 == IOR || op0 == XOR || op0 == PLUS)) | |
6950 | op0 = NIL; | |
6951 | else if (const0 == 0 && op0 == AND) | |
6952 | op0 = SET; | |
6953 | else if (const0 == GET_MODE_MASK (mode) && op0 == AND) | |
6954 | op0 = NIL; | |
6955 | ||
6956 | *pop0 = op0; | |
6957 | *pconst0 = const0; | |
6958 | ||
6959 | return 1; | |
6960 | } | |
6961 | \f | |
6962 | /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift. | |
6963 | The result of the shift is RESULT_MODE. X, if non-zero, is an expression | |
6964 | that we started with. | |
6965 | ||
6966 | The shift is normally computed in the widest mode we find in VAROP, as | |
6967 | long as it isn't a different number of words than RESULT_MODE. Exceptions | |
6968 | are ASHIFTRT and ROTATE, which are always done in their original mode, */ | |
6969 | ||
6970 | static rtx | |
6971 | simplify_shift_const (x, code, result_mode, varop, count) | |
6972 | rtx x; | |
6973 | enum rtx_code code; | |
6974 | enum machine_mode result_mode; | |
6975 | rtx varop; | |
6976 | int count; | |
6977 | { | |
6978 | enum rtx_code orig_code = code; | |
6979 | int orig_count = count; | |
6980 | enum machine_mode mode = result_mode; | |
6981 | enum machine_mode shift_mode, tmode; | |
6982 | int mode_words | |
6983 | = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD; | |
6984 | /* We form (outer_op (code varop count) (outer_const)). */ | |
6985 | enum rtx_code outer_op = NIL; | |
6986 | HOST_WIDE_INT outer_const; | |
6987 | rtx const_rtx; | |
6988 | int complement_p = 0; | |
6989 | rtx new; | |
6990 | ||
6991 | /* If we were given an invalid count, don't do anything except exactly | |
6992 | what was requested. */ | |
6993 | ||
6994 | if (count < 0 || count > GET_MODE_BITSIZE (mode)) | |
6995 | { | |
6996 | if (x) | |
6997 | return x; | |
6998 | ||
6999 | return gen_rtx (code, mode, varop, GEN_INT (count)); | |
7000 | } | |
7001 | ||
7002 | /* Unless one of the branches of the `if' in this loop does a `continue', | |
7003 | we will `break' the loop after the `if'. */ | |
7004 | ||
7005 | while (count != 0) | |
7006 | { | |
7007 | /* If we have an operand of (clobber (const_int 0)), just return that | |
7008 | value. */ | |
7009 | if (GET_CODE (varop) == CLOBBER) | |
7010 | return varop; | |
7011 | ||
7012 | /* If we discovered we had to complement VAROP, leave. Making a NOT | |
7013 | here would cause an infinite loop. */ | |
7014 | if (complement_p) | |
7015 | break; | |
7016 | ||
7017 | /* Convert ROTATETRT to ROTATE. */ | |
7018 | if (code == ROTATERT) | |
7019 | code = ROTATE, count = GET_MODE_BITSIZE (result_mode) - count; | |
7020 | ||
7021 | /* Canonicalize LSHIFT to ASHIFT. */ | |
7022 | if (code == LSHIFT) | |
7023 | code = ASHIFT; | |
7024 | ||
7025 | /* We need to determine what mode we will do the shift in. If the | |
7026 | shift is a ASHIFTRT or ROTATE, we must always do it in the mode it | |
7027 | was originally done in. Otherwise, we can do it in MODE, the widest | |
7028 | mode encountered. */ | |
7029 | shift_mode = (code == ASHIFTRT || code == ROTATE ? result_mode : mode); | |
7030 | ||
7031 | /* Handle cases where the count is greater than the size of the mode | |
7032 | minus 1. For ASHIFT, use the size minus one as the count (this can | |
7033 | occur when simplifying (lshiftrt (ashiftrt ..))). For rotates, | |
7034 | take the count modulo the size. For other shifts, the result is | |
7035 | zero. | |
7036 | ||
7037 | Since these shifts are being produced by the compiler by combining | |
7038 | multiple operations, each of which are defined, we know what the | |
7039 | result is supposed to be. */ | |
7040 | ||
7041 | if (count > GET_MODE_BITSIZE (shift_mode) - 1) | |
7042 | { | |
7043 | if (code == ASHIFTRT) | |
7044 | count = GET_MODE_BITSIZE (shift_mode) - 1; | |
7045 | else if (code == ROTATE || code == ROTATERT) | |
7046 | count %= GET_MODE_BITSIZE (shift_mode); | |
7047 | else | |
7048 | { | |
7049 | /* We can't simply return zero because there may be an | |
7050 | outer op. */ | |
7051 | varop = const0_rtx; | |
7052 | count = 0; | |
7053 | break; | |
7054 | } | |
7055 | } | |
7056 | ||
7057 | /* Negative counts are invalid and should not have been made (a | |
7058 | programmer-specified negative count should have been handled | |
7059 | above). */ | |
7060 | else if (count < 0) | |
7061 | abort (); | |
7062 | ||
7063 | /* An arithmetic right shift of a quantity known to be -1 or 0 | |
7064 | is a no-op. */ | |
7065 | if (code == ASHIFTRT | |
7066 | && (num_sign_bit_copies (varop, shift_mode) | |
7067 | == GET_MODE_BITSIZE (shift_mode))) | |
7068 | { | |
7069 | count = 0; | |
7070 | break; | |
7071 | } | |
7072 | ||
7073 | /* If we are doing an arithmetic right shift and discarding all but | |
7074 | the sign bit copies, this is equivalent to doing a shift by the | |
7075 | bitsize minus one. Convert it into that shift because it will often | |
7076 | allow other simplifications. */ | |
7077 | ||
7078 | if (code == ASHIFTRT | |
7079 | && (count + num_sign_bit_copies (varop, shift_mode) | |
7080 | >= GET_MODE_BITSIZE (shift_mode))) | |
7081 | count = GET_MODE_BITSIZE (shift_mode) - 1; | |
7082 | ||
7083 | /* We simplify the tests below and elsewhere by converting | |
7084 | ASHIFTRT to LSHIFTRT if we know the sign bit is clear. | |
7085 | `make_compound_operation' will convert it to a ASHIFTRT for | |
7086 | those machines (such as Vax) that don't have a LSHIFTRT. */ | |
7087 | if (GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT | |
7088 | && code == ASHIFTRT | |
7089 | && ((nonzero_bits (varop, shift_mode) | |
7090 | & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (shift_mode) - 1))) | |
7091 | == 0)) | |
7092 | code = LSHIFTRT; | |
7093 | ||
7094 | switch (GET_CODE (varop)) | |
7095 | { | |
7096 | case SIGN_EXTEND: | |
7097 | case ZERO_EXTEND: | |
7098 | case SIGN_EXTRACT: | |
7099 | case ZERO_EXTRACT: | |
7100 | new = expand_compound_operation (varop); | |
7101 | if (new != varop) | |
7102 | { | |
7103 | varop = new; | |
7104 | continue; | |
7105 | } | |
7106 | break; | |
7107 | ||
7108 | case MEM: | |
7109 | /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH | |
7110 | minus the width of a smaller mode, we can do this with a | |
7111 | SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */ | |
7112 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
7113 | && ! mode_dependent_address_p (XEXP (varop, 0)) | |
7114 | && ! MEM_VOLATILE_P (varop) | |
7115 | && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count, | |
7116 | MODE_INT, 1)) != BLKmode) | |
7117 | { | |
7118 | #if BYTES_BIG_ENDIAN | |
7119 | new = gen_rtx (MEM, tmode, XEXP (varop, 0)); | |
7120 | #else | |
7121 | new = gen_rtx (MEM, tmode, | |
7122 | plus_constant (XEXP (varop, 0), | |
7123 | count / BITS_PER_UNIT)); | |
7124 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (varop); | |
7125 | MEM_VOLATILE_P (new) = MEM_VOLATILE_P (varop); | |
7126 | MEM_IN_STRUCT_P (new) = MEM_IN_STRUCT_P (varop); | |
7127 | #endif | |
7128 | varop = gen_rtx_combine (code == ASHIFTRT ? SIGN_EXTEND | |
7129 | : ZERO_EXTEND, mode, new); | |
7130 | count = 0; | |
7131 | continue; | |
7132 | } | |
7133 | break; | |
7134 | ||
7135 | case USE: | |
7136 | /* Similar to the case above, except that we can only do this if | |
7137 | the resulting mode is the same as that of the underlying | |
7138 | MEM and adjust the address depending on the *bits* endianness | |
7139 | because of the way that bit-field extract insns are defined. */ | |
7140 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
7141 | && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count, | |
7142 | MODE_INT, 1)) != BLKmode | |
7143 | && tmode == GET_MODE (XEXP (varop, 0))) | |
7144 | { | |
7145 | #if BITS_BIG_ENDIAN | |
7146 | new = XEXP (varop, 0); | |
7147 | #else | |
7148 | new = copy_rtx (XEXP (varop, 0)); | |
7149 | SUBST (XEXP (new, 0), | |
7150 | plus_constant (XEXP (new, 0), | |
7151 | count / BITS_PER_UNIT)); | |
7152 | #endif | |
7153 | ||
7154 | varop = gen_rtx_combine (code == ASHIFTRT ? SIGN_EXTEND | |
7155 | : ZERO_EXTEND, mode, new); | |
7156 | count = 0; | |
7157 | continue; | |
7158 | } | |
7159 | break; | |
7160 | ||
7161 | case SUBREG: | |
7162 | /* If VAROP is a SUBREG, strip it as long as the inner operand has | |
7163 | the same number of words as what we've seen so far. Then store | |
7164 | the widest mode in MODE. */ | |
7165 | if (subreg_lowpart_p (varop) | |
7166 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop))) | |
7167 | > GET_MODE_SIZE (GET_MODE (varop))) | |
7168 | && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop))) | |
7169 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD) | |
7170 | == mode_words)) | |
7171 | { | |
7172 | varop = SUBREG_REG (varop); | |
7173 | if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode)) | |
7174 | mode = GET_MODE (varop); | |
7175 | continue; | |
7176 | } | |
7177 | break; | |
7178 | ||
7179 | case MULT: | |
7180 | /* Some machines use MULT instead of ASHIFT because MULT | |
7181 | is cheaper. But it is still better on those machines to | |
7182 | merge two shifts into one. */ | |
7183 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
7184 | && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0) | |
7185 | { | |
7186 | varop = gen_binary (ASHIFT, GET_MODE (varop), XEXP (varop, 0), | |
7187 | GEN_INT (exact_log2 (INTVAL (XEXP (varop, 1)))));; | |
7188 | continue; | |
7189 | } | |
7190 | break; | |
7191 | ||
7192 | case UDIV: | |
7193 | /* Similar, for when divides are cheaper. */ | |
7194 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
7195 | && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0) | |
7196 | { | |
7197 | varop = gen_binary (LSHIFTRT, GET_MODE (varop), XEXP (varop, 0), | |
7198 | GEN_INT (exact_log2 (INTVAL (XEXP (varop, 1))))); | |
7199 | continue; | |
7200 | } | |
7201 | break; | |
7202 | ||
7203 | case ASHIFTRT: | |
7204 | /* If we are extracting just the sign bit of an arithmetic right | |
7205 | shift, that shift is not needed. */ | |
7206 | if (code == LSHIFTRT && count == GET_MODE_BITSIZE (result_mode) - 1) | |
7207 | { | |
7208 | varop = XEXP (varop, 0); | |
7209 | continue; | |
7210 | } | |
7211 | ||
7212 | /* ... fall through ... */ | |
7213 | ||
7214 | case LSHIFTRT: | |
7215 | case ASHIFT: | |
7216 | case LSHIFT: | |
7217 | case ROTATE: | |
7218 | /* Here we have two nested shifts. The result is usually the | |
7219 | AND of a new shift with a mask. We compute the result below. */ | |
7220 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
7221 | && INTVAL (XEXP (varop, 1)) >= 0 | |
7222 | && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop)) | |
7223 | && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT | |
7224 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
7225 | { | |
7226 | enum rtx_code first_code = GET_CODE (varop); | |
7227 | int first_count = INTVAL (XEXP (varop, 1)); | |
7228 | unsigned HOST_WIDE_INT mask; | |
7229 | rtx mask_rtx; | |
7230 | rtx inner; | |
7231 | ||
7232 | if (first_code == LSHIFT) | |
7233 | first_code = ASHIFT; | |
7234 | ||
7235 | /* We have one common special case. We can't do any merging if | |
7236 | the inner code is an ASHIFTRT of a smaller mode. However, if | |
7237 | we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2) | |
7238 | with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2), | |
7239 | we can convert it to | |
7240 | (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0 C2) C3) C1). | |
7241 | This simplifies certain SIGN_EXTEND operations. */ | |
7242 | if (code == ASHIFT && first_code == ASHIFTRT | |
7243 | && (GET_MODE_BITSIZE (result_mode) | |
7244 | - GET_MODE_BITSIZE (GET_MODE (varop))) == count) | |
7245 | { | |
7246 | /* C3 has the low-order C1 bits zero. */ | |
7247 | ||
7248 | mask = (GET_MODE_MASK (mode) | |
7249 | & ~ (((HOST_WIDE_INT) 1 << first_count) - 1)); | |
7250 | ||
7251 | varop = simplify_and_const_int (NULL_RTX, result_mode, | |
7252 | XEXP (varop, 0), mask); | |
7253 | varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode, | |
7254 | varop, count); | |
7255 | count = first_count; | |
7256 | code = ASHIFTRT; | |
7257 | continue; | |
7258 | } | |
7259 | ||
7260 | /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more | |
7261 | than C1 high-order bits equal to the sign bit, we can convert | |
7262 | this to either an ASHIFT or a ASHIFTRT depending on the | |
7263 | two counts. | |
7264 | ||
7265 | We cannot do this if VAROP's mode is not SHIFT_MODE. */ | |
7266 | ||
7267 | if (code == ASHIFTRT && first_code == ASHIFT | |
7268 | && GET_MODE (varop) == shift_mode | |
7269 | && (num_sign_bit_copies (XEXP (varop, 0), shift_mode) | |
7270 | > first_count)) | |
7271 | { | |
7272 | count -= first_count; | |
7273 | if (count < 0) | |
7274 | count = - count, code = ASHIFT; | |
7275 | varop = XEXP (varop, 0); | |
7276 | continue; | |
7277 | } | |
7278 | ||
7279 | /* There are some cases we can't do. If CODE is ASHIFTRT, | |
7280 | we can only do this if FIRST_CODE is also ASHIFTRT. | |
7281 | ||
7282 | We can't do the case when CODE is ROTATE and FIRST_CODE is | |
7283 | ASHIFTRT. | |
7284 | ||
7285 | If the mode of this shift is not the mode of the outer shift, | |
7286 | we can't do this if either shift is ASHIFTRT or ROTATE. | |
7287 | ||
7288 | Finally, we can't do any of these if the mode is too wide | |
7289 | unless the codes are the same. | |
7290 | ||
7291 | Handle the case where the shift codes are the same | |
7292 | first. */ | |
7293 | ||
7294 | if (code == first_code) | |
7295 | { | |
7296 | if (GET_MODE (varop) != result_mode | |
7297 | && (code == ASHIFTRT || code == ROTATE)) | |
7298 | break; | |
7299 | ||
7300 | count += first_count; | |
7301 | varop = XEXP (varop, 0); | |
7302 | continue; | |
7303 | } | |
7304 | ||
7305 | if (code == ASHIFTRT | |
7306 | || (code == ROTATE && first_code == ASHIFTRT) | |
7307 | || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT | |
7308 | || (GET_MODE (varop) != result_mode | |
7309 | && (first_code == ASHIFTRT || first_code == ROTATE | |
7310 | || code == ROTATE))) | |
7311 | break; | |
7312 | ||
7313 | /* To compute the mask to apply after the shift, shift the | |
7314 | nonzero bits of the inner shift the same way the | |
7315 | outer shift will. */ | |
7316 | ||
7317 | mask_rtx = GEN_INT (nonzero_bits (varop, GET_MODE (varop))); | |
7318 | ||
7319 | mask_rtx | |
7320 | = simplify_binary_operation (code, result_mode, mask_rtx, | |
7321 | GEN_INT (count)); | |
7322 | ||
7323 | /* Give up if we can't compute an outer operation to use. */ | |
7324 | if (mask_rtx == 0 | |
7325 | || GET_CODE (mask_rtx) != CONST_INT | |
7326 | || ! merge_outer_ops (&outer_op, &outer_const, AND, | |
7327 | INTVAL (mask_rtx), | |
7328 | result_mode, &complement_p)) | |
7329 | break; | |
7330 | ||
7331 | /* If the shifts are in the same direction, we add the | |
7332 | counts. Otherwise, we subtract them. */ | |
7333 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
7334 | == (first_code == ASHIFTRT || first_code == LSHIFTRT)) | |
7335 | count += first_count; | |
7336 | else | |
7337 | count -= first_count; | |
7338 | ||
7339 | /* If COUNT is positive, the new shift is usually CODE, | |
7340 | except for the two exceptions below, in which case it is | |
7341 | FIRST_CODE. If the count is negative, FIRST_CODE should | |
7342 | always be used */ | |
7343 | if (count > 0 | |
7344 | && ((first_code == ROTATE && code == ASHIFT) | |
7345 | || (first_code == ASHIFTRT && code == LSHIFTRT))) | |
7346 | code = first_code; | |
7347 | else if (count < 0) | |
7348 | code = first_code, count = - count; | |
7349 | ||
7350 | varop = XEXP (varop, 0); | |
7351 | continue; | |
7352 | } | |
7353 | ||
7354 | /* If we have (A << B << C) for any shift, we can convert this to | |
7355 | (A << C << B). This wins if A is a constant. Only try this if | |
7356 | B is not a constant. */ | |
7357 | ||
7358 | else if (GET_CODE (varop) == code | |
7359 | && GET_CODE (XEXP (varop, 1)) != CONST_INT | |
7360 | && 0 != (new | |
7361 | = simplify_binary_operation (code, mode, | |
7362 | XEXP (varop, 0), | |
7363 | GEN_INT (count)))) | |
7364 | { | |
7365 | varop = gen_rtx_combine (code, mode, new, XEXP (varop, 1)); | |
7366 | count = 0; | |
7367 | continue; | |
7368 | } | |
7369 | break; | |
7370 | ||
7371 | case NOT: | |
7372 | /* Make this fit the case below. */ | |
7373 | varop = gen_rtx_combine (XOR, mode, XEXP (varop, 0), | |
7374 | GEN_INT (GET_MODE_MASK (mode))); | |
7375 | continue; | |
7376 | ||
7377 | case IOR: | |
7378 | case AND: | |
7379 | case XOR: | |
7380 | /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C) | |
7381 | with C the size of VAROP - 1 and the shift is logical if | |
7382 | STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1, | |
7383 | we have an (le X 0) operation. If we have an arithmetic shift | |
7384 | and STORE_FLAG_VALUE is 1 or we have a logical shift with | |
7385 | STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */ | |
7386 | ||
7387 | if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS | |
7388 | && XEXP (XEXP (varop, 0), 1) == constm1_rtx | |
7389 | && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) | |
7390 | && (code == LSHIFTRT || code == ASHIFTRT) | |
7391 | && count == GET_MODE_BITSIZE (GET_MODE (varop)) - 1 | |
7392 | && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1))) | |
7393 | { | |
7394 | count = 0; | |
7395 | varop = gen_rtx_combine (LE, GET_MODE (varop), XEXP (varop, 1), | |
7396 | const0_rtx); | |
7397 | ||
7398 | if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT) | |
7399 | varop = gen_rtx_combine (NEG, GET_MODE (varop), varop); | |
7400 | ||
7401 | continue; | |
7402 | } | |
7403 | ||
7404 | /* If we have (shift (logical)), move the logical to the outside | |
7405 | to allow it to possibly combine with another logical and the | |
7406 | shift to combine with another shift. This also canonicalizes to | |
7407 | what a ZERO_EXTRACT looks like. Also, some machines have | |
7408 | (and (shift)) insns. */ | |
7409 | ||
7410 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
7411 | && (new = simplify_binary_operation (code, result_mode, | |
7412 | XEXP (varop, 1), | |
7413 | GEN_INT (count))) != 0 | |
7414 | && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop), | |
7415 | INTVAL (new), result_mode, &complement_p)) | |
7416 | { | |
7417 | varop = XEXP (varop, 0); | |
7418 | continue; | |
7419 | } | |
7420 | ||
7421 | /* If we can't do that, try to simplify the shift in each arm of the | |
7422 | logical expression, make a new logical expression, and apply | |
7423 | the inverse distributive law. */ | |
7424 | { | |
7425 | rtx lhs = simplify_shift_const (NULL_RTX, code, result_mode, | |
7426 | XEXP (varop, 0), count); | |
7427 | rtx rhs = simplify_shift_const (NULL_RTX, code, result_mode, | |
7428 | XEXP (varop, 1), count); | |
7429 | ||
7430 | varop = gen_binary (GET_CODE (varop), result_mode, lhs, rhs); | |
7431 | varop = apply_distributive_law (varop); | |
7432 | ||
7433 | count = 0; | |
7434 | } | |
7435 | break; | |
7436 | ||
7437 | case EQ: | |
7438 | /* convert (lshift (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE | |
7439 | says that the sign bit can be tested, FOO has mode MODE, C is | |
7440 | GET_MODE_BITSIZE (MODE) - 1, and FOO has only the low-order bit | |
7441 | may be nonzero. */ | |
7442 | if (code == LSHIFT | |
7443 | && XEXP (varop, 1) == const0_rtx | |
7444 | && GET_MODE (XEXP (varop, 0)) == result_mode | |
7445 | && count == GET_MODE_BITSIZE (result_mode) - 1 | |
7446 | && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT | |
7447 | && ((STORE_FLAG_VALUE | |
7448 | & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (result_mode) - 1)))) | |
7449 | && nonzero_bits (XEXP (varop, 0), result_mode) == 1 | |
7450 | && merge_outer_ops (&outer_op, &outer_const, XOR, | |
7451 | (HOST_WIDE_INT) 1, result_mode, | |
7452 | &complement_p)) | |
7453 | { | |
7454 | varop = XEXP (varop, 0); | |
7455 | count = 0; | |
7456 | continue; | |
7457 | } | |
7458 | break; | |
7459 | ||
7460 | case NEG: | |
7461 | /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less | |
7462 | than the number of bits in the mode is equivalent to A. */ | |
7463 | if (code == LSHIFTRT && count == GET_MODE_BITSIZE (result_mode) - 1 | |
7464 | && nonzero_bits (XEXP (varop, 0), result_mode) == 1) | |
7465 | { | |
7466 | varop = XEXP (varop, 0); | |
7467 | count = 0; | |
7468 | continue; | |
7469 | } | |
7470 | ||
7471 | /* NEG commutes with ASHIFT since it is multiplication. Move the | |
7472 | NEG outside to allow shifts to combine. */ | |
7473 | if (code == ASHIFT | |
7474 | && merge_outer_ops (&outer_op, &outer_const, NEG, | |
7475 | (HOST_WIDE_INT) 0, result_mode, | |
7476 | &complement_p)) | |
7477 | { | |
7478 | varop = XEXP (varop, 0); | |
7479 | continue; | |
7480 | } | |
7481 | break; | |
7482 | ||
7483 | case PLUS: | |
7484 | /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C | |
7485 | is one less than the number of bits in the mode is | |
7486 | equivalent to (xor A 1). */ | |
7487 | if (code == LSHIFTRT && count == GET_MODE_BITSIZE (result_mode) - 1 | |
7488 | && XEXP (varop, 1) == constm1_rtx | |
7489 | && nonzero_bits (XEXP (varop, 0), result_mode) == 1 | |
7490 | && merge_outer_ops (&outer_op, &outer_const, XOR, | |
7491 | (HOST_WIDE_INT) 1, result_mode, | |
7492 | &complement_p)) | |
7493 | { | |
7494 | count = 0; | |
7495 | varop = XEXP (varop, 0); | |
7496 | continue; | |
7497 | } | |
7498 | ||
7499 | /* If we have (xshiftrt (plus FOO BAR) C), and the only bits | |
7500 | that might be nonzero in BAR are those being shifted out and those | |
7501 | bits are known zero in FOO, we can replace the PLUS with FOO. | |
7502 | Similarly in the other operand order. This code occurs when | |
7503 | we are computing the size of a variable-size array. */ | |
7504 | ||
7505 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
7506 | && count < HOST_BITS_PER_WIDE_INT | |
7507 | && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0 | |
7508 | && (nonzero_bits (XEXP (varop, 1), result_mode) | |
7509 | & nonzero_bits (XEXP (varop, 0), result_mode)) == 0) | |
7510 | { | |
7511 | varop = XEXP (varop, 0); | |
7512 | continue; | |
7513 | } | |
7514 | else if ((code == ASHIFTRT || code == LSHIFTRT) | |
7515 | && count < HOST_BITS_PER_WIDE_INT | |
7516 | && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT | |
7517 | && 0 == (nonzero_bits (XEXP (varop, 0), result_mode) | |
7518 | >> count) | |
7519 | && 0 == (nonzero_bits (XEXP (varop, 0), result_mode) | |
7520 | & nonzero_bits (XEXP (varop, 1), | |
7521 | result_mode))) | |
7522 | { | |
7523 | varop = XEXP (varop, 1); | |
7524 | continue; | |
7525 | } | |
7526 | ||
7527 | /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */ | |
7528 | if (code == ASHIFT | |
7529 | && GET_CODE (XEXP (varop, 1)) == CONST_INT | |
7530 | && (new = simplify_binary_operation (ASHIFT, result_mode, | |
7531 | XEXP (varop, 1), | |
7532 | GEN_INT (count))) != 0 | |
7533 | && merge_outer_ops (&outer_op, &outer_const, PLUS, | |
7534 | INTVAL (new), result_mode, &complement_p)) | |
7535 | { | |
7536 | varop = XEXP (varop, 0); | |
7537 | continue; | |
7538 | } | |
7539 | break; | |
7540 | ||
7541 | case MINUS: | |
7542 | /* If we have (xshiftrt (minus (ashiftrt X C)) X) C) | |
7543 | with C the size of VAROP - 1 and the shift is logical if | |
7544 | STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1, | |
7545 | we have a (gt X 0) operation. If the shift is arithmetic with | |
7546 | STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1, | |
7547 | we have a (neg (gt X 0)) operation. */ | |
7548 | ||
7549 | if (GET_CODE (XEXP (varop, 0)) == ASHIFTRT | |
7550 | && count == GET_MODE_BITSIZE (GET_MODE (varop)) - 1 | |
7551 | && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) | |
7552 | && (code == LSHIFTRT || code == ASHIFTRT) | |
7553 | && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT | |
7554 | && INTVAL (XEXP (XEXP (varop, 0), 1)) == count | |
7555 | && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1))) | |
7556 | { | |
7557 | count = 0; | |
7558 | varop = gen_rtx_combine (GT, GET_MODE (varop), XEXP (varop, 1), | |
7559 | const0_rtx); | |
7560 | ||
7561 | if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT) | |
7562 | varop = gen_rtx_combine (NEG, GET_MODE (varop), varop); | |
7563 | ||
7564 | continue; | |
7565 | } | |
7566 | break; | |
7567 | } | |
7568 | ||
7569 | break; | |
7570 | } | |
7571 | ||
7572 | /* We need to determine what mode to do the shift in. If the shift is | |
7573 | a ASHIFTRT or ROTATE, we must always do it in the mode it was originally | |
7574 | done in. Otherwise, we can do it in MODE, the widest mode encountered. | |
7575 | The code we care about is that of the shift that will actually be done, | |
7576 | not the shift that was originally requested. */ | |
7577 | shift_mode = (code == ASHIFTRT || code == ROTATE ? result_mode : mode); | |
7578 | ||
7579 | /* We have now finished analyzing the shift. The result should be | |
7580 | a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If | |
7581 | OUTER_OP is non-NIL, it is an operation that needs to be applied | |
7582 | to the result of the shift. OUTER_CONST is the relevant constant, | |
7583 | but we must turn off all bits turned off in the shift. | |
7584 | ||
7585 | If we were passed a value for X, see if we can use any pieces of | |
7586 | it. If not, make new rtx. */ | |
7587 | ||
7588 | if (x && GET_RTX_CLASS (GET_CODE (x)) == '2' | |
7589 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
7590 | && INTVAL (XEXP (x, 1)) == count) | |
7591 | const_rtx = XEXP (x, 1); | |
7592 | else | |
7593 | const_rtx = GEN_INT (count); | |
7594 | ||
7595 | if (x && GET_CODE (XEXP (x, 0)) == SUBREG | |
7596 | && GET_MODE (XEXP (x, 0)) == shift_mode | |
7597 | && SUBREG_REG (XEXP (x, 0)) == varop) | |
7598 | varop = XEXP (x, 0); | |
7599 | else if (GET_MODE (varop) != shift_mode) | |
7600 | varop = gen_lowpart_for_combine (shift_mode, varop); | |
7601 | ||
7602 | /* If we can't make the SUBREG, try to return what we were given. */ | |
7603 | if (GET_CODE (varop) == CLOBBER) | |
7604 | return x ? x : varop; | |
7605 | ||
7606 | new = simplify_binary_operation (code, shift_mode, varop, const_rtx); | |
7607 | if (new != 0) | |
7608 | x = new; | |
7609 | else | |
7610 | { | |
7611 | if (x == 0 || GET_CODE (x) != code || GET_MODE (x) != shift_mode) | |
7612 | x = gen_rtx_combine (code, shift_mode, varop, const_rtx); | |
7613 | ||
7614 | SUBST (XEXP (x, 0), varop); | |
7615 | SUBST (XEXP (x, 1), const_rtx); | |
7616 | } | |
7617 | ||
7618 | /* If we were doing a LSHIFTRT in a wider mode than it was originally, | |
7619 | turn off all the bits that the shift would have turned off. */ | |
7620 | if (orig_code == LSHIFTRT && result_mode != shift_mode) | |
7621 | x = simplify_and_const_int (NULL_RTX, shift_mode, x, | |
7622 | GET_MODE_MASK (result_mode) >> orig_count); | |
7623 | ||
7624 | /* Do the remainder of the processing in RESULT_MODE. */ | |
7625 | x = gen_lowpart_for_combine (result_mode, x); | |
7626 | ||
7627 | /* If COMPLEMENT_P is set, we have to complement X before doing the outer | |
7628 | operation. */ | |
7629 | if (complement_p) | |
7630 | x = gen_unary (NOT, result_mode, x); | |
7631 | ||
7632 | if (outer_op != NIL) | |
7633 | { | |
7634 | if (GET_MODE_BITSIZE (result_mode) < HOST_BITS_PER_WIDE_INT) | |
7635 | outer_const &= GET_MODE_MASK (result_mode); | |
7636 | ||
7637 | if (outer_op == AND) | |
7638 | x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const); | |
7639 | else if (outer_op == SET) | |
7640 | /* This means that we have determined that the result is | |
7641 | equivalent to a constant. This should be rare. */ | |
7642 | x = GEN_INT (outer_const); | |
7643 | else if (GET_RTX_CLASS (outer_op) == '1') | |
7644 | x = gen_unary (outer_op, result_mode, x); | |
7645 | else | |
7646 | x = gen_binary (outer_op, result_mode, x, GEN_INT (outer_const)); | |
7647 | } | |
7648 | ||
7649 | return x; | |
7650 | } | |
7651 | \f | |
7652 | /* Like recog, but we receive the address of a pointer to a new pattern. | |
7653 | We try to match the rtx that the pointer points to. | |
7654 | If that fails, we may try to modify or replace the pattern, | |
7655 | storing the replacement into the same pointer object. | |
7656 | ||
7657 | Modifications include deletion or addition of CLOBBERs. | |
7658 | ||
7659 | PNOTES is a pointer to a location where any REG_UNUSED notes added for | |
7660 | the CLOBBERs are placed. | |
7661 | ||
7662 | The value is the final insn code from the pattern ultimately matched, | |
7663 | or -1. */ | |
7664 | ||
7665 | static int | |
7666 | recog_for_combine (pnewpat, insn, pnotes) | |
7667 | rtx *pnewpat; | |
7668 | rtx insn; | |
7669 | rtx *pnotes; | |
7670 | { | |
7671 | register rtx pat = *pnewpat; | |
7672 | int insn_code_number; | |
7673 | int num_clobbers_to_add = 0; | |
7674 | int i; | |
7675 | rtx notes = 0; | |
7676 | ||
7677 | /* Is the result of combination a valid instruction? */ | |
7678 | insn_code_number = recog (pat, insn, &num_clobbers_to_add); | |
7679 | ||
7680 | /* If it isn't, there is the possibility that we previously had an insn | |
7681 | that clobbered some register as a side effect, but the combined | |
7682 | insn doesn't need to do that. So try once more without the clobbers | |
7683 | unless this represents an ASM insn. */ | |
7684 | ||
7685 | if (insn_code_number < 0 && ! check_asm_operands (pat) | |
7686 | && GET_CODE (pat) == PARALLEL) | |
7687 | { | |
7688 | int pos; | |
7689 | ||
7690 | for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++) | |
7691 | if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER) | |
7692 | { | |
7693 | if (i != pos) | |
7694 | SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i)); | |
7695 | pos++; | |
7696 | } | |
7697 | ||
7698 | SUBST_INT (XVECLEN (pat, 0), pos); | |
7699 | ||
7700 | if (pos == 1) | |
7701 | pat = XVECEXP (pat, 0, 0); | |
7702 | ||
7703 | insn_code_number = recog (pat, insn, &num_clobbers_to_add); | |
7704 | } | |
7705 | ||
7706 | /* If we had any clobbers to add, make a new pattern than contains | |
7707 | them. Then check to make sure that all of them are dead. */ | |
7708 | if (num_clobbers_to_add) | |
7709 | { | |
7710 | rtx newpat = gen_rtx (PARALLEL, VOIDmode, | |
7711 | gen_rtvec (GET_CODE (pat) == PARALLEL | |
7712 | ? XVECLEN (pat, 0) + num_clobbers_to_add | |
7713 | : num_clobbers_to_add + 1)); | |
7714 | ||
7715 | if (GET_CODE (pat) == PARALLEL) | |
7716 | for (i = 0; i < XVECLEN (pat, 0); i++) | |
7717 | XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i); | |
7718 | else | |
7719 | XVECEXP (newpat, 0, 0) = pat; | |
7720 | ||
7721 | add_clobbers (newpat, insn_code_number); | |
7722 | ||
7723 | for (i = XVECLEN (newpat, 0) - num_clobbers_to_add; | |
7724 | i < XVECLEN (newpat, 0); i++) | |
7725 | { | |
7726 | if (GET_CODE (XEXP (XVECEXP (newpat, 0, i), 0)) == REG | |
7727 | && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn)) | |
7728 | return -1; | |
7729 | notes = gen_rtx (EXPR_LIST, REG_UNUSED, | |
7730 | XEXP (XVECEXP (newpat, 0, i), 0), notes); | |
7731 | } | |
7732 | pat = newpat; | |
7733 | } | |
7734 | ||
7735 | *pnewpat = pat; | |
7736 | *pnotes = notes; | |
7737 | ||
7738 | return insn_code_number; | |
7739 | } | |
7740 | \f | |
7741 | /* Like gen_lowpart but for use by combine. In combine it is not possible | |
7742 | to create any new pseudoregs. However, it is safe to create | |
7743 | invalid memory addresses, because combine will try to recognize | |
7744 | them and all they will do is make the combine attempt fail. | |
7745 | ||
7746 | If for some reason this cannot do its job, an rtx | |
7747 | (clobber (const_int 0)) is returned. | |
7748 | An insn containing that will not be recognized. */ | |
7749 | ||
7750 | #undef gen_lowpart | |
7751 | ||
7752 | static rtx | |
7753 | gen_lowpart_for_combine (mode, x) | |
7754 | enum machine_mode mode; | |
7755 | register rtx x; | |
7756 | { | |
7757 | rtx result; | |
7758 | ||
7759 | if (GET_MODE (x) == mode) | |
7760 | return x; | |
7761 | ||
7762 | /* We can only support MODE being wider than a word if X is a | |
7763 | constant integer or has a mode the same size. */ | |
7764 | ||
7765 | if (GET_MODE_SIZE (mode) > UNITS_PER_WORD | |
7766 | && ! ((GET_MODE (x) == VOIDmode | |
7767 | && (GET_CODE (x) == CONST_INT | |
7768 | || GET_CODE (x) == CONST_DOUBLE)) | |
7769 | || GET_MODE_SIZE (GET_MODE (x)) == GET_MODE_SIZE (mode))) | |
7770 | return gen_rtx (CLOBBER, GET_MODE (x), const0_rtx); | |
7771 | ||
7772 | /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart | |
7773 | won't know what to do. So we will strip off the SUBREG here and | |
7774 | process normally. */ | |
7775 | if (GET_CODE (x) == SUBREG && GET_CODE (SUBREG_REG (x)) == MEM) | |
7776 | { | |
7777 | x = SUBREG_REG (x); | |
7778 | if (GET_MODE (x) == mode) | |
7779 | return x; | |
7780 | } | |
7781 | ||
7782 | result = gen_lowpart_common (mode, x); | |
7783 | if (result) | |
7784 | return result; | |
7785 | ||
7786 | if (GET_CODE (x) == MEM) | |
7787 | { | |
7788 | register int offset = 0; | |
7789 | rtx new; | |
7790 | ||
7791 | /* Refuse to work on a volatile memory ref or one with a mode-dependent | |
7792 | address. */ | |
7793 | if (MEM_VOLATILE_P (x) || mode_dependent_address_p (XEXP (x, 0))) | |
7794 | return gen_rtx (CLOBBER, GET_MODE (x), const0_rtx); | |
7795 | ||
7796 | /* If we want to refer to something bigger than the original memref, | |
7797 | generate a perverse subreg instead. That will force a reload | |
7798 | of the original memref X. */ | |
7799 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)) | |
7800 | return gen_rtx (SUBREG, mode, x, 0); | |
7801 | ||
7802 | #if WORDS_BIG_ENDIAN | |
7803 | offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD) | |
7804 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)); | |
7805 | #endif | |
7806 | #if BYTES_BIG_ENDIAN | |
7807 | /* Adjust the address so that the address-after-the-data | |
7808 | is unchanged. */ | |
7809 | offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode)) | |
7810 | - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x)))); | |
7811 | #endif | |
7812 | new = gen_rtx (MEM, mode, plus_constant (XEXP (x, 0), offset)); | |
7813 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (x); | |
7814 | MEM_VOLATILE_P (new) = MEM_VOLATILE_P (x); | |
7815 | MEM_IN_STRUCT_P (new) = MEM_IN_STRUCT_P (x); | |
7816 | return new; | |
7817 | } | |
7818 | ||
7819 | /* If X is a comparison operator, rewrite it in a new mode. This | |
7820 | probably won't match, but may allow further simplifications. */ | |
7821 | else if (GET_RTX_CLASS (GET_CODE (x)) == '<') | |
7822 | return gen_rtx_combine (GET_CODE (x), mode, XEXP (x, 0), XEXP (x, 1)); | |
7823 | ||
7824 | /* If we couldn't simplify X any other way, just enclose it in a | |
7825 | SUBREG. Normally, this SUBREG won't match, but some patterns may | |
7826 | include an explicit SUBREG or we may simplify it further in combine. */ | |
7827 | else | |
7828 | { | |
7829 | int word = 0; | |
7830 | ||
7831 | if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) | |
7832 | word = ((GET_MODE_SIZE (GET_MODE (x)) | |
7833 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)) | |
7834 | / UNITS_PER_WORD); | |
7835 | return gen_rtx (SUBREG, mode, x, word); | |
7836 | } | |
7837 | } | |
7838 | \f | |
7839 | /* Make an rtx expression. This is a subset of gen_rtx and only supports | |
7840 | expressions of 1, 2, or 3 operands, each of which are rtx expressions. | |
7841 | ||
7842 | If the identical expression was previously in the insn (in the undobuf), | |
7843 | it will be returned. Only if it is not found will a new expression | |
7844 | be made. */ | |
7845 | ||
7846 | /*VARARGS2*/ | |
7847 | static rtx | |
7848 | gen_rtx_combine (va_alist) | |
7849 | va_dcl | |
7850 | { | |
7851 | va_list p; | |
7852 | enum rtx_code code; | |
7853 | enum machine_mode mode; | |
7854 | int n_args; | |
7855 | rtx args[3]; | |
7856 | int i, j; | |
7857 | char *fmt; | |
7858 | rtx rt; | |
7859 | ||
7860 | va_start (p); | |
7861 | code = va_arg (p, enum rtx_code); | |
7862 | mode = va_arg (p, enum machine_mode); | |
7863 | n_args = GET_RTX_LENGTH (code); | |
7864 | fmt = GET_RTX_FORMAT (code); | |
7865 | ||
7866 | if (n_args == 0 || n_args > 3) | |
7867 | abort (); | |
7868 | ||
7869 | /* Get each arg and verify that it is supposed to be an expression. */ | |
7870 | for (j = 0; j < n_args; j++) | |
7871 | { | |
7872 | if (*fmt++ != 'e') | |
7873 | abort (); | |
7874 | ||
7875 | args[j] = va_arg (p, rtx); | |
7876 | } | |
7877 | ||
7878 | /* See if this is in undobuf. Be sure we don't use objects that came | |
7879 | from another insn; this could produce circular rtl structures. */ | |
7880 | ||
7881 | for (i = previous_num_undos; i < undobuf.num_undo; i++) | |
7882 | if (!undobuf.undo[i].is_int | |
7883 | && GET_CODE (undobuf.undo[i].old_contents.rtx) == code | |
7884 | && GET_MODE (undobuf.undo[i].old_contents.rtx) == mode) | |
7885 | { | |
7886 | for (j = 0; j < n_args; j++) | |
7887 | if (XEXP (undobuf.undo[i].old_contents.rtx, j) != args[j]) | |
7888 | break; | |
7889 | ||
7890 | if (j == n_args) | |
7891 | return undobuf.undo[i].old_contents.rtx; | |
7892 | } | |
7893 | ||
7894 | /* Otherwise make a new rtx. We know we have 1, 2, or 3 args. | |
7895 | Use rtx_alloc instead of gen_rtx because it's faster on RISC. */ | |
7896 | rt = rtx_alloc (code); | |
7897 | PUT_MODE (rt, mode); | |
7898 | XEXP (rt, 0) = args[0]; | |
7899 | if (n_args > 1) | |
7900 | { | |
7901 | XEXP (rt, 1) = args[1]; | |
7902 | if (n_args > 2) | |
7903 | XEXP (rt, 2) = args[2]; | |
7904 | } | |
7905 | return rt; | |
7906 | } | |
7907 | ||
7908 | /* These routines make binary and unary operations by first seeing if they | |
7909 | fold; if not, a new expression is allocated. */ | |
7910 | ||
7911 | static rtx | |
7912 | gen_binary (code, mode, op0, op1) | |
7913 | enum rtx_code code; | |
7914 | enum machine_mode mode; | |
7915 | rtx op0, op1; | |
7916 | { | |
7917 | rtx result; | |
7918 | rtx tem; | |
7919 | ||
7920 | if (GET_RTX_CLASS (code) == 'c' | |
7921 | && (GET_CODE (op0) == CONST_INT | |
7922 | || (CONSTANT_P (op0) && GET_CODE (op1) != CONST_INT))) | |
7923 | tem = op0, op0 = op1, op1 = tem; | |
7924 | ||
7925 | if (GET_RTX_CLASS (code) == '<') | |
7926 | { | |
7927 | enum machine_mode op_mode = GET_MODE (op0); | |
7928 | if (op_mode == VOIDmode) | |
7929 | op_mode = GET_MODE (op1); | |
7930 | result = simplify_relational_operation (code, op_mode, op0, op1); | |
7931 | } | |
7932 | else | |
7933 | result = simplify_binary_operation (code, mode, op0, op1); | |
7934 | ||
7935 | if (result) | |
7936 | return result; | |
7937 | ||
7938 | /* Put complex operands first and constants second. */ | |
7939 | if (GET_RTX_CLASS (code) == 'c' | |
7940 | && ((CONSTANT_P (op0) && GET_CODE (op1) != CONST_INT) | |
7941 | || (GET_RTX_CLASS (GET_CODE (op0)) == 'o' | |
7942 | && GET_RTX_CLASS (GET_CODE (op1)) != 'o') | |
7943 | || (GET_CODE (op0) == SUBREG | |
7944 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (op0))) == 'o' | |
7945 | && GET_RTX_CLASS (GET_CODE (op1)) != 'o'))) | |
7946 | return gen_rtx_combine (code, mode, op1, op0); | |
7947 | ||
7948 | return gen_rtx_combine (code, mode, op0, op1); | |
7949 | } | |
7950 | ||
7951 | static rtx | |
7952 | gen_unary (code, mode, op0) | |
7953 | enum rtx_code code; | |
7954 | enum machine_mode mode; | |
7955 | rtx op0; | |
7956 | { | |
7957 | rtx result = simplify_unary_operation (code, mode, op0, mode); | |
7958 | ||
7959 | if (result) | |
7960 | return result; | |
7961 | ||
7962 | return gen_rtx_combine (code, mode, op0); | |
7963 | } | |
7964 | \f | |
7965 | /* Simplify a comparison between *POP0 and *POP1 where CODE is the | |
7966 | comparison code that will be tested. | |
7967 | ||
7968 | The result is a possibly different comparison code to use. *POP0 and | |
7969 | *POP1 may be updated. | |
7970 | ||
7971 | It is possible that we might detect that a comparison is either always | |
7972 | true or always false. However, we do not perform general constant | |
7973 | folding in combine, so this knowledge isn't useful. Such tautologies | |
7974 | should have been detected earlier. Hence we ignore all such cases. */ | |
7975 | ||
7976 | static enum rtx_code | |
7977 | simplify_comparison (code, pop0, pop1) | |
7978 | enum rtx_code code; | |
7979 | rtx *pop0; | |
7980 | rtx *pop1; | |
7981 | { | |
7982 | rtx op0 = *pop0; | |
7983 | rtx op1 = *pop1; | |
7984 | rtx tem, tem1; | |
7985 | int i; | |
7986 | enum machine_mode mode, tmode; | |
7987 | ||
7988 | /* Try a few ways of applying the same transformation to both operands. */ | |
7989 | while (1) | |
7990 | { | |
7991 | /* If both operands are the same constant shift, see if we can ignore the | |
7992 | shift. We can if the shift is a rotate or if the bits shifted out of | |
7993 | this shift are known to be zero for both inputs and if the type of | |
7994 | comparison is compatible with the shift. */ | |
7995 | if (GET_CODE (op0) == GET_CODE (op1) | |
7996 | && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT | |
7997 | && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ)) | |
7998 | || ((GET_CODE (op0) == LSHIFTRT | |
7999 | || GET_CODE (op0) == ASHIFT || GET_CODE (op0) == LSHIFT) | |
8000 | && (code != GT && code != LT && code != GE && code != LE)) | |
8001 | || (GET_CODE (op0) == ASHIFTRT | |
8002 | && (code != GTU && code != LTU | |
8003 | && code != GEU && code != GEU))) | |
8004 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8005 | && INTVAL (XEXP (op0, 1)) >= 0 | |
8006 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT | |
8007 | && XEXP (op0, 1) == XEXP (op1, 1)) | |
8008 | { | |
8009 | enum machine_mode mode = GET_MODE (op0); | |
8010 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); | |
8011 | int shift_count = INTVAL (XEXP (op0, 1)); | |
8012 | ||
8013 | if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT) | |
8014 | mask &= (mask >> shift_count) << shift_count; | |
8015 | else if (GET_CODE (op0) == ASHIFT || GET_CODE (op0) == LSHIFT) | |
8016 | mask = (mask & (mask << shift_count)) >> shift_count; | |
8017 | ||
8018 | if ((nonzero_bits (XEXP (op0, 0), mode) & ~ mask) == 0 | |
8019 | && (nonzero_bits (XEXP (op1, 0), mode) & ~ mask) == 0) | |
8020 | op0 = XEXP (op0, 0), op1 = XEXP (op1, 0); | |
8021 | else | |
8022 | break; | |
8023 | } | |
8024 | ||
8025 | /* If both operands are AND's of a paradoxical SUBREG by constant, the | |
8026 | SUBREGs are of the same mode, and, in both cases, the AND would | |
8027 | be redundant if the comparison was done in the narrower mode, | |
8028 | do the comparison in the narrower mode (e.g., we are AND'ing with 1 | |
8029 | and the operand's possibly nonzero bits are 0xffffff01; in that case | |
8030 | if we only care about QImode, we don't need the AND). This case | |
8031 | occurs if the output mode of an scc insn is not SImode and | |
8032 | STORE_FLAG_VALUE == 1 (e.g., the 386). */ | |
8033 | ||
8034 | else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND | |
8035 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8036 | && GET_CODE (XEXP (op1, 1)) == CONST_INT | |
8037 | && GET_CODE (XEXP (op0, 0)) == SUBREG | |
8038 | && GET_CODE (XEXP (op1, 0)) == SUBREG | |
8039 | && (GET_MODE_SIZE (GET_MODE (XEXP (op0, 0))) | |
8040 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (op0, 0))))) | |
8041 | && (GET_MODE (SUBREG_REG (XEXP (op0, 0))) | |
8042 | == GET_MODE (SUBREG_REG (XEXP (op1, 0)))) | |
8043 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (XEXP (op0, 0)))) | |
8044 | <= HOST_BITS_PER_WIDE_INT) | |
8045 | && (nonzero_bits (SUBREG_REG (XEXP (op0, 0)), | |
8046 | GET_MODE (SUBREG_REG (XEXP (op0, 0)))) | |
8047 | & ~ INTVAL (XEXP (op0, 1))) == 0 | |
8048 | && (nonzero_bits (SUBREG_REG (XEXP (op1, 0)), | |
8049 | GET_MODE (SUBREG_REG (XEXP (op1, 0)))) | |
8050 | & ~ INTVAL (XEXP (op1, 1))) == 0) | |
8051 | { | |
8052 | op0 = SUBREG_REG (XEXP (op0, 0)); | |
8053 | op1 = SUBREG_REG (XEXP (op1, 0)); | |
8054 | ||
8055 | /* the resulting comparison is always unsigned since we masked off | |
8056 | the original sign bit. */ | |
8057 | code = unsigned_condition (code); | |
8058 | } | |
8059 | else | |
8060 | break; | |
8061 | } | |
8062 | ||
8063 | /* If the first operand is a constant, swap the operands and adjust the | |
8064 | comparison code appropriately. */ | |
8065 | if (CONSTANT_P (op0)) | |
8066 | { | |
8067 | tem = op0, op0 = op1, op1 = tem; | |
8068 | code = swap_condition (code); | |
8069 | } | |
8070 | ||
8071 | /* We now enter a loop during which we will try to simplify the comparison. | |
8072 | For the most part, we only are concerned with comparisons with zero, | |
8073 | but some things may really be comparisons with zero but not start | |
8074 | out looking that way. */ | |
8075 | ||
8076 | while (GET_CODE (op1) == CONST_INT) | |
8077 | { | |
8078 | enum machine_mode mode = GET_MODE (op0); | |
8079 | int mode_width = GET_MODE_BITSIZE (mode); | |
8080 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); | |
8081 | int equality_comparison_p; | |
8082 | int sign_bit_comparison_p; | |
8083 | int unsigned_comparison_p; | |
8084 | HOST_WIDE_INT const_op; | |
8085 | ||
8086 | /* We only want to handle integral modes. This catches VOIDmode, | |
8087 | CCmode, and the floating-point modes. An exception is that we | |
8088 | can handle VOIDmode if OP0 is a COMPARE or a comparison | |
8089 | operation. */ | |
8090 | ||
8091 | if (GET_MODE_CLASS (mode) != MODE_INT | |
8092 | && ! (mode == VOIDmode | |
8093 | && (GET_CODE (op0) == COMPARE | |
8094 | || GET_RTX_CLASS (GET_CODE (op0)) == '<'))) | |
8095 | break; | |
8096 | ||
8097 | /* Get the constant we are comparing against and turn off all bits | |
8098 | not on in our mode. */ | |
8099 | const_op = INTVAL (op1); | |
8100 | if (mode_width <= HOST_BITS_PER_WIDE_INT) | |
8101 | const_op &= mask; | |
8102 | ||
8103 | /* If we are comparing against a constant power of two and the value | |
8104 | being compared can only have that single bit nonzero (e.g., it was | |
8105 | `and'ed with that bit), we can replace this with a comparison | |
8106 | with zero. */ | |
8107 | if (const_op | |
8108 | && (code == EQ || code == NE || code == GE || code == GEU | |
8109 | || code == LT || code == LTU) | |
8110 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8111 | && exact_log2 (const_op) >= 0 | |
8112 | && nonzero_bits (op0, mode) == const_op) | |
8113 | { | |
8114 | code = (code == EQ || code == GE || code == GEU ? NE : EQ); | |
8115 | op1 = const0_rtx, const_op = 0; | |
8116 | } | |
8117 | ||
8118 | /* Similarly, if we are comparing a value known to be either -1 or | |
8119 | 0 with -1, change it to the opposite comparison against zero. */ | |
8120 | ||
8121 | if (const_op == -1 | |
8122 | && (code == EQ || code == NE || code == GT || code == LE | |
8123 | || code == GEU || code == LTU) | |
8124 | && num_sign_bit_copies (op0, mode) == mode_width) | |
8125 | { | |
8126 | code = (code == EQ || code == LE || code == GEU ? NE : EQ); | |
8127 | op1 = const0_rtx, const_op = 0; | |
8128 | } | |
8129 | ||
8130 | /* Do some canonicalizations based on the comparison code. We prefer | |
8131 | comparisons against zero and then prefer equality comparisons. | |
8132 | If we can reduce the size of a constant, we will do that too. */ | |
8133 | ||
8134 | switch (code) | |
8135 | { | |
8136 | case LT: | |
8137 | /* < C is equivalent to <= (C - 1) */ | |
8138 | if (const_op > 0) | |
8139 | { | |
8140 | const_op -= 1; | |
8141 | op1 = GEN_INT (const_op); | |
8142 | code = LE; | |
8143 | /* ... fall through to LE case below. */ | |
8144 | } | |
8145 | else | |
8146 | break; | |
8147 | ||
8148 | case LE: | |
8149 | /* <= C is equivalent to < (C + 1); we do this for C < 0 */ | |
8150 | if (const_op < 0) | |
8151 | { | |
8152 | const_op += 1; | |
8153 | op1 = GEN_INT (const_op); | |
8154 | code = LT; | |
8155 | } | |
8156 | ||
8157 | /* If we are doing a <= 0 comparison on a value known to have | |
8158 | a zero sign bit, we can replace this with == 0. */ | |
8159 | else if (const_op == 0 | |
8160 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8161 | && (nonzero_bits (op0, mode) | |
8162 | & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0) | |
8163 | code = EQ; | |
8164 | break; | |
8165 | ||
8166 | case GE: | |
8167 | /* >= C is equivalent to > (C - 1). */ | |
8168 | if (const_op > 0) | |
8169 | { | |
8170 | const_op -= 1; | |
8171 | op1 = GEN_INT (const_op); | |
8172 | code = GT; | |
8173 | /* ... fall through to GT below. */ | |
8174 | } | |
8175 | else | |
8176 | break; | |
8177 | ||
8178 | case GT: | |
8179 | /* > C is equivalent to >= (C + 1); we do this for C < 0*/ | |
8180 | if (const_op < 0) | |
8181 | { | |
8182 | const_op += 1; | |
8183 | op1 = GEN_INT (const_op); | |
8184 | code = GE; | |
8185 | } | |
8186 | ||
8187 | /* If we are doing a > 0 comparison on a value known to have | |
8188 | a zero sign bit, we can replace this with != 0. */ | |
8189 | else if (const_op == 0 | |
8190 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8191 | && (nonzero_bits (op0, mode) | |
8192 | & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0) | |
8193 | code = NE; | |
8194 | break; | |
8195 | ||
8196 | case LTU: | |
8197 | /* < C is equivalent to <= (C - 1). */ | |
8198 | if (const_op > 0) | |
8199 | { | |
8200 | const_op -= 1; | |
8201 | op1 = GEN_INT (const_op); | |
8202 | code = LEU; | |
8203 | /* ... fall through ... */ | |
8204 | } | |
8205 | ||
8206 | /* (unsigned) < 0x80000000 is equivalent to >= 0. */ | |
8207 | else if (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)) | |
8208 | { | |
8209 | const_op = 0, op1 = const0_rtx; | |
8210 | code = GE; | |
8211 | break; | |
8212 | } | |
8213 | else | |
8214 | break; | |
8215 | ||
8216 | case LEU: | |
8217 | /* unsigned <= 0 is equivalent to == 0 */ | |
8218 | if (const_op == 0) | |
8219 | code = EQ; | |
8220 | ||
8221 | /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */ | |
8222 | else if (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1) | |
8223 | { | |
8224 | const_op = 0, op1 = const0_rtx; | |
8225 | code = GE; | |
8226 | } | |
8227 | break; | |
8228 | ||
8229 | case GEU: | |
8230 | /* >= C is equivalent to < (C - 1). */ | |
8231 | if (const_op > 1) | |
8232 | { | |
8233 | const_op -= 1; | |
8234 | op1 = GEN_INT (const_op); | |
8235 | code = GTU; | |
8236 | /* ... fall through ... */ | |
8237 | } | |
8238 | ||
8239 | /* (unsigned) >= 0x80000000 is equivalent to < 0. */ | |
8240 | else if (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)) | |
8241 | { | |
8242 | const_op = 0, op1 = const0_rtx; | |
8243 | code = LT; | |
8244 | } | |
8245 | else | |
8246 | break; | |
8247 | ||
8248 | case GTU: | |
8249 | /* unsigned > 0 is equivalent to != 0 */ | |
8250 | if (const_op == 0) | |
8251 | code = NE; | |
8252 | ||
8253 | /* (unsigned) > 0x7fffffff is equivalent to < 0. */ | |
8254 | else if (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1) | |
8255 | { | |
8256 | const_op = 0, op1 = const0_rtx; | |
8257 | code = LT; | |
8258 | } | |
8259 | break; | |
8260 | } | |
8261 | ||
8262 | /* Compute some predicates to simplify code below. */ | |
8263 | ||
8264 | equality_comparison_p = (code == EQ || code == NE); | |
8265 | sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0); | |
8266 | unsigned_comparison_p = (code == LTU || code == LEU || code == GTU | |
8267 | || code == LEU); | |
8268 | ||
8269 | /* Now try cases based on the opcode of OP0. If none of the cases | |
8270 | does a "continue", we exit this loop immediately after the | |
8271 | switch. */ | |
8272 | ||
8273 | switch (GET_CODE (op0)) | |
8274 | { | |
8275 | case ZERO_EXTRACT: | |
8276 | /* If we are extracting a single bit from a variable position in | |
8277 | a constant that has only a single bit set and are comparing it | |
8278 | with zero, we can convert this into an equality comparison | |
8279 | between the position and the location of the single bit. We can't | |
8280 | do this if bit endian and we don't have an extzv since we then | |
8281 | can't know what mode to use for the endianness adjustment. */ | |
8282 | ||
8283 | #if ! BITS_BIG_ENDIAN || defined (HAVE_extzv) | |
8284 | if (GET_CODE (XEXP (op0, 0)) == CONST_INT | |
8285 | && XEXP (op0, 1) == const1_rtx | |
8286 | && equality_comparison_p && const_op == 0 | |
8287 | && (i = exact_log2 (INTVAL (XEXP (op0, 0)))) >= 0) | |
8288 | { | |
8289 | #if BITS_BIG_ENDIAN | |
8290 | i = (GET_MODE_BITSIZE | |
8291 | (insn_operand_mode[(int) CODE_FOR_extzv][1]) - 1 - i); | |
8292 | #endif | |
8293 | ||
8294 | op0 = XEXP (op0, 2); | |
8295 | op1 = GEN_INT (i); | |
8296 | const_op = i; | |
8297 | ||
8298 | /* Result is nonzero iff shift count is equal to I. */ | |
8299 | code = reverse_condition (code); | |
8300 | continue; | |
8301 | } | |
8302 | #endif | |
8303 | ||
8304 | /* ... fall through ... */ | |
8305 | ||
8306 | case SIGN_EXTRACT: | |
8307 | tem = expand_compound_operation (op0); | |
8308 | if (tem != op0) | |
8309 | { | |
8310 | op0 = tem; | |
8311 | continue; | |
8312 | } | |
8313 | break; | |
8314 | ||
8315 | case NOT: | |
8316 | /* If testing for equality, we can take the NOT of the constant. */ | |
8317 | if (equality_comparison_p | |
8318 | && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0) | |
8319 | { | |
8320 | op0 = XEXP (op0, 0); | |
8321 | op1 = tem; | |
8322 | continue; | |
8323 | } | |
8324 | ||
8325 | /* If just looking at the sign bit, reverse the sense of the | |
8326 | comparison. */ | |
8327 | if (sign_bit_comparison_p) | |
8328 | { | |
8329 | op0 = XEXP (op0, 0); | |
8330 | code = (code == GE ? LT : GE); | |
8331 | continue; | |
8332 | } | |
8333 | break; | |
8334 | ||
8335 | case NEG: | |
8336 | /* If testing for equality, we can take the NEG of the constant. */ | |
8337 | if (equality_comparison_p | |
8338 | && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0) | |
8339 | { | |
8340 | op0 = XEXP (op0, 0); | |
8341 | op1 = tem; | |
8342 | continue; | |
8343 | } | |
8344 | ||
8345 | /* The remaining cases only apply to comparisons with zero. */ | |
8346 | if (const_op != 0) | |
8347 | break; | |
8348 | ||
8349 | /* When X is ABS or is known positive, | |
8350 | (neg X) is < 0 if and only if X != 0. */ | |
8351 | ||
8352 | if (sign_bit_comparison_p | |
8353 | && (GET_CODE (XEXP (op0, 0)) == ABS | |
8354 | || (mode_width <= HOST_BITS_PER_WIDE_INT | |
8355 | && (nonzero_bits (XEXP (op0, 0), mode) | |
8356 | & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0))) | |
8357 | { | |
8358 | op0 = XEXP (op0, 0); | |
8359 | code = (code == LT ? NE : EQ); | |
8360 | continue; | |
8361 | } | |
8362 | ||
8363 | /* If we have NEG of something whose two high-order bits are the | |
8364 | same, we know that "(-a) < 0" is equivalent to "a > 0". */ | |
8365 | if (num_sign_bit_copies (op0, mode) >= 2) | |
8366 | { | |
8367 | op0 = XEXP (op0, 0); | |
8368 | code = swap_condition (code); | |
8369 | continue; | |
8370 | } | |
8371 | break; | |
8372 | ||
8373 | case ROTATE: | |
8374 | /* If we are testing equality and our count is a constant, we | |
8375 | can perform the inverse operation on our RHS. */ | |
8376 | if (equality_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8377 | && (tem = simplify_binary_operation (ROTATERT, mode, | |
8378 | op1, XEXP (op0, 1))) != 0) | |
8379 | { | |
8380 | op0 = XEXP (op0, 0); | |
8381 | op1 = tem; | |
8382 | continue; | |
8383 | } | |
8384 | ||
8385 | /* If we are doing a < 0 or >= 0 comparison, it means we are testing | |
8386 | a particular bit. Convert it to an AND of a constant of that | |
8387 | bit. This will be converted into a ZERO_EXTRACT. */ | |
8388 | if (const_op == 0 && sign_bit_comparison_p | |
8389 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8390 | && mode_width <= HOST_BITS_PER_WIDE_INT) | |
8391 | { | |
8392 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), | |
8393 | ((HOST_WIDE_INT) 1 | |
8394 | << (mode_width - 1 | |
8395 | - INTVAL (XEXP (op0, 1))))); | |
8396 | code = (code == LT ? NE : EQ); | |
8397 | continue; | |
8398 | } | |
8399 | ||
8400 | /* ... fall through ... */ | |
8401 | ||
8402 | case ABS: | |
8403 | /* ABS is ignorable inside an equality comparison with zero. */ | |
8404 | if (const_op == 0 && equality_comparison_p) | |
8405 | { | |
8406 | op0 = XEXP (op0, 0); | |
8407 | continue; | |
8408 | } | |
8409 | break; | |
8410 | ||
8411 | ||
8412 | case SIGN_EXTEND: | |
8413 | /* Can simplify (compare (zero/sign_extend FOO) CONST) | |
8414 | to (compare FOO CONST) if CONST fits in FOO's mode and we | |
8415 | are either testing inequality or have an unsigned comparison | |
8416 | with ZERO_EXTEND or a signed comparison with SIGN_EXTEND. */ | |
8417 | if (! unsigned_comparison_p | |
8418 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0))) | |
8419 | <= HOST_BITS_PER_WIDE_INT) | |
8420 | && ((unsigned HOST_WIDE_INT) const_op | |
8421 | < (((HOST_WIDE_INT) 1 | |
8422 | << (GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0))) - 1))))) | |
8423 | { | |
8424 | op0 = XEXP (op0, 0); | |
8425 | continue; | |
8426 | } | |
8427 | break; | |
8428 | ||
8429 | case SUBREG: | |
8430 | /* Check for the case where we are comparing A - C1 with C2, | |
8431 | both constants are smaller than 1/2 the maxium positive | |
8432 | value in MODE, and the comparison is equality or unsigned. | |
8433 | In that case, if A is either zero-extended to MODE or has | |
8434 | sufficient sign bits so that the high-order bit in MODE | |
8435 | is a copy of the sign in the inner mode, we can prove that it is | |
8436 | safe to do the operation in the wider mode. This simplifies | |
8437 | many range checks. */ | |
8438 | ||
8439 | if (mode_width <= HOST_BITS_PER_WIDE_INT | |
8440 | && subreg_lowpart_p (op0) | |
8441 | && GET_CODE (SUBREG_REG (op0)) == PLUS | |
8442 | && GET_CODE (XEXP (SUBREG_REG (op0), 1)) == CONST_INT | |
8443 | && INTVAL (XEXP (SUBREG_REG (op0), 1)) < 0 | |
8444 | && (- INTVAL (XEXP (SUBREG_REG (op0), 1)) | |
8445 | < GET_MODE_MASK (mode) / 2) | |
8446 | && (unsigned HOST_WIDE_INT) const_op < GET_MODE_MASK (mode) / 2 | |
8447 | && (0 == (nonzero_bits (XEXP (SUBREG_REG (op0), 0), | |
8448 | GET_MODE (SUBREG_REG (op0))) | |
8449 | & ~ GET_MODE_MASK (mode)) | |
8450 | || (num_sign_bit_copies (XEXP (SUBREG_REG (op0), 0), | |
8451 | GET_MODE (SUBREG_REG (op0))) | |
8452 | > (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) | |
8453 | - GET_MODE_BITSIZE (mode))))) | |
8454 | { | |
8455 | op0 = SUBREG_REG (op0); | |
8456 | continue; | |
8457 | } | |
8458 | ||
8459 | /* If the inner mode is narrower and we are extracting the low part, | |
8460 | we can treat the SUBREG as if it were a ZERO_EXTEND. */ | |
8461 | if (subreg_lowpart_p (op0) | |
8462 | && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) < mode_width) | |
8463 | /* Fall through */ ; | |
8464 | else | |
8465 | break; | |
8466 | ||
8467 | /* ... fall through ... */ | |
8468 | ||
8469 | case ZERO_EXTEND: | |
8470 | if ((unsigned_comparison_p || equality_comparison_p) | |
8471 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0))) | |
8472 | <= HOST_BITS_PER_WIDE_INT) | |
8473 | && ((unsigned HOST_WIDE_INT) const_op | |
8474 | < GET_MODE_MASK (GET_MODE (XEXP (op0, 0))))) | |
8475 | { | |
8476 | op0 = XEXP (op0, 0); | |
8477 | continue; | |
8478 | } | |
8479 | break; | |
8480 | ||
8481 | case PLUS: | |
8482 | /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do | |
8483 | this for equality comparisons due to pathological cases involving | |
8484 | overflows. */ | |
8485 | if (equality_comparison_p | |
8486 | && 0 != (tem = simplify_binary_operation (MINUS, mode, | |
8487 | op1, XEXP (op0, 1)))) | |
8488 | { | |
8489 | op0 = XEXP (op0, 0); | |
8490 | op1 = tem; | |
8491 | continue; | |
8492 | } | |
8493 | ||
8494 | /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */ | |
8495 | if (const_op == 0 && XEXP (op0, 1) == constm1_rtx | |
8496 | && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p) | |
8497 | { | |
8498 | op0 = XEXP (XEXP (op0, 0), 0); | |
8499 | code = (code == LT ? EQ : NE); | |
8500 | continue; | |
8501 | } | |
8502 | break; | |
8503 | ||
8504 | case MINUS: | |
8505 | /* (eq (minus A B) C) -> (eq A (plus B C)) or | |
8506 | (eq B (minus A C)), whichever simplifies. We can only do | |
8507 | this for equality comparisons due to pathological cases involving | |
8508 | overflows. */ | |
8509 | if (equality_comparison_p | |
8510 | && 0 != (tem = simplify_binary_operation (PLUS, mode, | |
8511 | XEXP (op0, 1), op1))) | |
8512 | { | |
8513 | op0 = XEXP (op0, 0); | |
8514 | op1 = tem; | |
8515 | continue; | |
8516 | } | |
8517 | ||
8518 | if (equality_comparison_p | |
8519 | && 0 != (tem = simplify_binary_operation (MINUS, mode, | |
8520 | XEXP (op0, 0), op1))) | |
8521 | { | |
8522 | op0 = XEXP (op0, 1); | |
8523 | op1 = tem; | |
8524 | continue; | |
8525 | } | |
8526 | ||
8527 | /* The sign bit of (minus (ashiftrt X C) X), where C is the number | |
8528 | of bits in X minus 1, is one iff X > 0. */ | |
8529 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT | |
8530 | && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT | |
8531 | && INTVAL (XEXP (XEXP (op0, 0), 1)) == mode_width - 1 | |
8532 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1))) | |
8533 | { | |
8534 | op0 = XEXP (op0, 1); | |
8535 | code = (code == GE ? LE : GT); | |
8536 | continue; | |
8537 | } | |
8538 | break; | |
8539 | ||
8540 | case XOR: | |
8541 | /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification | |
8542 | if C is zero or B is a constant. */ | |
8543 | if (equality_comparison_p | |
8544 | && 0 != (tem = simplify_binary_operation (XOR, mode, | |
8545 | XEXP (op0, 1), op1))) | |
8546 | { | |
8547 | op0 = XEXP (op0, 0); | |
8548 | op1 = tem; | |
8549 | continue; | |
8550 | } | |
8551 | break; | |
8552 | ||
8553 | case EQ: case NE: | |
8554 | case LT: case LTU: case LE: case LEU: | |
8555 | case GT: case GTU: case GE: case GEU: | |
8556 | /* We can't do anything if OP0 is a condition code value, rather | |
8557 | than an actual data value. */ | |
8558 | if (const_op != 0 | |
8559 | #ifdef HAVE_cc0 | |
8560 | || XEXP (op0, 0) == cc0_rtx | |
8561 | #endif | |
8562 | || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC) | |
8563 | break; | |
8564 | ||
8565 | /* Get the two operands being compared. */ | |
8566 | if (GET_CODE (XEXP (op0, 0)) == COMPARE) | |
8567 | tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1); | |
8568 | else | |
8569 | tem = XEXP (op0, 0), tem1 = XEXP (op0, 1); | |
8570 | ||
8571 | /* Check for the cases where we simply want the result of the | |
8572 | earlier test or the opposite of that result. */ | |
8573 | if (code == NE | |
8574 | || (code == EQ && reversible_comparison_p (op0)) | |
8575 | || (GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT | |
8576 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
8577 | && (STORE_FLAG_VALUE | |
8578 | & (((HOST_WIDE_INT) 1 | |
8579 | << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1)))) | |
8580 | && (code == LT | |
8581 | || (code == GE && reversible_comparison_p (op0))))) | |
8582 | { | |
8583 | code = (code == LT || code == NE | |
8584 | ? GET_CODE (op0) : reverse_condition (GET_CODE (op0))); | |
8585 | op0 = tem, op1 = tem1; | |
8586 | continue; | |
8587 | } | |
8588 | break; | |
8589 | ||
8590 | case IOR: | |
8591 | /* The sign bit of (ior (plus X (const_int -1)) X) is non-zero | |
8592 | iff X <= 0. */ | |
8593 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS | |
8594 | && XEXP (XEXP (op0, 0), 1) == constm1_rtx | |
8595 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1))) | |
8596 | { | |
8597 | op0 = XEXP (op0, 1); | |
8598 | code = (code == GE ? GT : LE); | |
8599 | continue; | |
8600 | } | |
8601 | break; | |
8602 | ||
8603 | case AND: | |
8604 | /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This | |
8605 | will be converted to a ZERO_EXTRACT later. */ | |
8606 | if (const_op == 0 && equality_comparison_p | |
8607 | && (GET_CODE (XEXP (op0, 0)) == ASHIFT | |
8608 | || GET_CODE (XEXP (op0, 0)) == LSHIFT) | |
8609 | && XEXP (XEXP (op0, 0), 0) == const1_rtx) | |
8610 | { | |
8611 | op0 = simplify_and_const_int | |
8612 | (op0, mode, gen_rtx_combine (LSHIFTRT, mode, | |
8613 | XEXP (op0, 1), | |
8614 | XEXP (XEXP (op0, 0), 1)), | |
8615 | (HOST_WIDE_INT) 1); | |
8616 | continue; | |
8617 | } | |
8618 | ||
8619 | /* If we are comparing (and (lshiftrt X C1) C2) for equality with | |
8620 | zero and X is a comparison and C1 and C2 describe only bits set | |
8621 | in STORE_FLAG_VALUE, we can compare with X. */ | |
8622 | if (const_op == 0 && equality_comparison_p | |
8623 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8624 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8625 | && GET_CODE (XEXP (op0, 0)) == LSHIFTRT | |
8626 | && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT | |
8627 | && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0 | |
8628 | && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT) | |
8629 | { | |
8630 | mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode)) | |
8631 | << INTVAL (XEXP (XEXP (op0, 0), 1))); | |
8632 | if ((~ STORE_FLAG_VALUE & mask) == 0 | |
8633 | && (GET_RTX_CLASS (GET_CODE (XEXP (XEXP (op0, 0), 0))) == '<' | |
8634 | || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0 | |
8635 | && GET_RTX_CLASS (GET_CODE (tem)) == '<'))) | |
8636 | { | |
8637 | op0 = XEXP (XEXP (op0, 0), 0); | |
8638 | continue; | |
8639 | } | |
8640 | } | |
8641 | ||
8642 | /* If we are doing an equality comparison of an AND of a bit equal | |
8643 | to the sign bit, replace this with a LT or GE comparison of | |
8644 | the underlying value. */ | |
8645 | if (equality_comparison_p | |
8646 | && const_op == 0 | |
8647 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8648 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8649 | && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode)) | |
8650 | == (HOST_WIDE_INT) 1 << (mode_width - 1))) | |
8651 | { | |
8652 | op0 = XEXP (op0, 0); | |
8653 | code = (code == EQ ? GE : LT); | |
8654 | continue; | |
8655 | } | |
8656 | ||
8657 | /* If this AND operation is really a ZERO_EXTEND from a narrower | |
8658 | mode, the constant fits within that mode, and this is either an | |
8659 | equality or unsigned comparison, try to do this comparison in | |
8660 | the narrower mode. */ | |
8661 | if ((equality_comparison_p || unsigned_comparison_p) | |
8662 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8663 | && (i = exact_log2 ((INTVAL (XEXP (op0, 1)) | |
8664 | & GET_MODE_MASK (mode)) | |
8665 | + 1)) >= 0 | |
8666 | && const_op >> i == 0 | |
8667 | && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode) | |
8668 | { | |
8669 | op0 = gen_lowpart_for_combine (tmode, XEXP (op0, 0)); | |
8670 | continue; | |
8671 | } | |
8672 | break; | |
8673 | ||
8674 | case ASHIFT: | |
8675 | case LSHIFT: | |
8676 | /* If we have (compare (xshift FOO N) (const_int C)) and | |
8677 | the high order N bits of FOO (N+1 if an inequality comparison) | |
8678 | are known to be zero, we can do this by comparing FOO with C | |
8679 | shifted right N bits so long as the low-order N bits of C are | |
8680 | zero. */ | |
8681 | if (GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8682 | && INTVAL (XEXP (op0, 1)) >= 0 | |
8683 | && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p) | |
8684 | < HOST_BITS_PER_WIDE_INT) | |
8685 | && ((const_op | |
8686 | & ((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1) == 0) | |
8687 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8688 | && (nonzero_bits (XEXP (op0, 0), mode) | |
8689 | & ~ (mask >> (INTVAL (XEXP (op0, 1)) | |
8690 | + ! equality_comparison_p))) == 0) | |
8691 | { | |
8692 | const_op >>= INTVAL (XEXP (op0, 1)); | |
8693 | op1 = GEN_INT (const_op); | |
8694 | op0 = XEXP (op0, 0); | |
8695 | continue; | |
8696 | } | |
8697 | ||
8698 | /* If we are doing a sign bit comparison, it means we are testing | |
8699 | a particular bit. Convert it to the appropriate AND. */ | |
8700 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8701 | && mode_width <= HOST_BITS_PER_WIDE_INT) | |
8702 | { | |
8703 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), | |
8704 | ((HOST_WIDE_INT) 1 | |
8705 | << (mode_width - 1 | |
8706 | - INTVAL (XEXP (op0, 1))))); | |
8707 | code = (code == LT ? NE : EQ); | |
8708 | continue; | |
8709 | } | |
8710 | ||
8711 | /* If this an equality comparison with zero and we are shifting | |
8712 | the low bit to the sign bit, we can convert this to an AND of the | |
8713 | low-order bit. */ | |
8714 | if (const_op == 0 && equality_comparison_p | |
8715 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8716 | && INTVAL (XEXP (op0, 1)) == mode_width - 1) | |
8717 | { | |
8718 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), | |
8719 | (HOST_WIDE_INT) 1); | |
8720 | continue; | |
8721 | } | |
8722 | break; | |
8723 | ||
8724 | case ASHIFTRT: | |
8725 | /* If this is an equality comparison with zero, we can do this | |
8726 | as a logical shift, which might be much simpler. */ | |
8727 | if (equality_comparison_p && const_op == 0 | |
8728 | && GET_CODE (XEXP (op0, 1)) == CONST_INT) | |
8729 | { | |
8730 | op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode, | |
8731 | XEXP (op0, 0), | |
8732 | INTVAL (XEXP (op0, 1))); | |
8733 | continue; | |
8734 | } | |
8735 | ||
8736 | /* If OP0 is a sign extension and CODE is not an unsigned comparison, | |
8737 | do the comparison in a narrower mode. */ | |
8738 | if (! unsigned_comparison_p | |
8739 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8740 | && GET_CODE (XEXP (op0, 0)) == ASHIFT | |
8741 | && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1) | |
8742 | && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)), | |
8743 | MODE_INT, 1)) != BLKmode | |
8744 | && ((unsigned HOST_WIDE_INT) const_op <= GET_MODE_MASK (tmode) | |
8745 | || ((unsigned HOST_WIDE_INT) - const_op | |
8746 | <= GET_MODE_MASK (tmode)))) | |
8747 | { | |
8748 | op0 = gen_lowpart_for_combine (tmode, XEXP (XEXP (op0, 0), 0)); | |
8749 | continue; | |
8750 | } | |
8751 | ||
8752 | /* ... fall through ... */ | |
8753 | case LSHIFTRT: | |
8754 | /* If we have (compare (xshiftrt FOO N) (const_int C)) and | |
8755 | the low order N bits of FOO are known to be zero, we can do this | |
8756 | by comparing FOO with C shifted left N bits so long as no | |
8757 | overflow occurs. */ | |
8758 | if (GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8759 | && INTVAL (XEXP (op0, 1)) >= 0 | |
8760 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT | |
8761 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
8762 | && (nonzero_bits (XEXP (op0, 0), mode) | |
8763 | & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0 | |
8764 | && (const_op == 0 | |
8765 | || (floor_log2 (const_op) + INTVAL (XEXP (op0, 1)) | |
8766 | < mode_width))) | |
8767 | { | |
8768 | const_op <<= INTVAL (XEXP (op0, 1)); | |
8769 | op1 = GEN_INT (const_op); | |
8770 | op0 = XEXP (op0, 0); | |
8771 | continue; | |
8772 | } | |
8773 | ||
8774 | /* If we are using this shift to extract just the sign bit, we | |
8775 | can replace this with an LT or GE comparison. */ | |
8776 | if (const_op == 0 | |
8777 | && (equality_comparison_p || sign_bit_comparison_p) | |
8778 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
8779 | && INTVAL (XEXP (op0, 1)) == mode_width - 1) | |
8780 | { | |
8781 | op0 = XEXP (op0, 0); | |
8782 | code = (code == NE || code == GT ? LT : GE); | |
8783 | continue; | |
8784 | } | |
8785 | break; | |
8786 | } | |
8787 | ||
8788 | break; | |
8789 | } | |
8790 | ||
8791 | /* Now make any compound operations involved in this comparison. Then, | |
8792 | check for an outmost SUBREG on OP0 that isn't doing anything or is | |
8793 | paradoxical. The latter case can only occur when it is known that the | |
8794 | "extra" bits will be zero. Therefore, it is safe to remove the SUBREG. | |
8795 | We can never remove a SUBREG for a non-equality comparison because the | |
8796 | sign bit is in a different place in the underlying object. */ | |
8797 | ||
8798 | op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET); | |
8799 | op1 = make_compound_operation (op1, SET); | |
8800 | ||
8801 | if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0) | |
8802 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
8803 | && (code == NE || code == EQ) | |
8804 | && ((GET_MODE_SIZE (GET_MODE (op0)) | |
8805 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))) | |
8806 | { | |
8807 | op0 = SUBREG_REG (op0); | |
8808 | op1 = gen_lowpart_for_combine (GET_MODE (op0), op1); | |
8809 | } | |
8810 | ||
8811 | else if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0) | |
8812 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
8813 | && (code == NE || code == EQ) | |
8814 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) | |
8815 | <= HOST_BITS_PER_WIDE_INT) | |
8816 | && (nonzero_bits (SUBREG_REG (op0), GET_MODE (SUBREG_REG (op0))) | |
8817 | & ~ GET_MODE_MASK (GET_MODE (op0))) == 0 | |
8818 | && (tem = gen_lowpart_for_combine (GET_MODE (SUBREG_REG (op0)), | |
8819 | op1), | |
8820 | (nonzero_bits (tem, GET_MODE (SUBREG_REG (op0))) | |
8821 | & ~ GET_MODE_MASK (GET_MODE (op0))) == 0)) | |
8822 | op0 = SUBREG_REG (op0), op1 = tem; | |
8823 | ||
8824 | /* We now do the opposite procedure: Some machines don't have compare | |
8825 | insns in all modes. If OP0's mode is an integer mode smaller than a | |
8826 | word and we can't do a compare in that mode, see if there is a larger | |
8827 | mode for which we can do the compare. There are a number of cases in | |
8828 | which we can use the wider mode. */ | |
8829 | ||
8830 | mode = GET_MODE (op0); | |
8831 | if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT | |
8832 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD | |
8833 | && cmp_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
8834 | for (tmode = GET_MODE_WIDER_MODE (mode); | |
8835 | (tmode != VOIDmode | |
8836 | && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT); | |
8837 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
8838 | if (cmp_optab->handlers[(int) tmode].insn_code != CODE_FOR_nothing) | |
8839 | { | |
8840 | /* If the only nonzero bits in OP0 and OP1 are those in the | |
8841 | narrower mode and this is an equality or unsigned comparison, | |
8842 | we can use the wider mode. Similarly for sign-extended | |
8843 | values and equality or signed comparisons. */ | |
8844 | if (((code == EQ || code == NE | |
8845 | || code == GEU || code == GTU || code == LEU || code == LTU) | |
8846 | && (nonzero_bits (op0, tmode) & ~ GET_MODE_MASK (mode)) == 0 | |
8847 | && (nonzero_bits (op1, tmode) & ~ GET_MODE_MASK (mode)) == 0) | |
8848 | || ((code == EQ || code == NE | |
8849 | || code == GE || code == GT || code == LE || code == LT) | |
8850 | && (num_sign_bit_copies (op0, tmode) | |
8851 | > GET_MODE_BITSIZE (tmode) - GET_MODE_BITSIZE (mode)) | |
8852 | && (num_sign_bit_copies (op1, tmode) | |
8853 | > GET_MODE_BITSIZE (tmode) - GET_MODE_BITSIZE (mode)))) | |
8854 | { | |
8855 | op0 = gen_lowpart_for_combine (tmode, op0); | |
8856 | op1 = gen_lowpart_for_combine (tmode, op1); | |
8857 | break; | |
8858 | } | |
8859 | ||
8860 | /* If this is a test for negative, we can make an explicit | |
8861 | test of the sign bit. */ | |
8862 | ||
8863 | if (op1 == const0_rtx && (code == LT || code == GE) | |
8864 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
8865 | { | |
8866 | op0 = gen_binary (AND, tmode, | |
8867 | gen_lowpart_for_combine (tmode, op0), | |
8868 | GEN_INT ((HOST_WIDE_INT) 1 | |
8869 | << (GET_MODE_BITSIZE (mode) - 1))); | |
8870 | code = (code == LT) ? NE : EQ; | |
8871 | break; | |
8872 | } | |
8873 | } | |
8874 | ||
8875 | *pop0 = op0; | |
8876 | *pop1 = op1; | |
8877 | ||
8878 | return code; | |
8879 | } | |
8880 | \f | |
8881 | /* Return 1 if we know that X, a comparison operation, is not operating | |
8882 | on a floating-point value or is EQ or NE, meaning that we can safely | |
8883 | reverse it. */ | |
8884 | ||
8885 | static int | |
8886 | reversible_comparison_p (x) | |
8887 | rtx x; | |
8888 | { | |
8889 | if (TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
8890 | || GET_CODE (x) == NE || GET_CODE (x) == EQ) | |
8891 | return 1; | |
8892 | ||
8893 | switch (GET_MODE_CLASS (GET_MODE (XEXP (x, 0)))) | |
8894 | { | |
8895 | case MODE_INT: | |
8896 | return 1; | |
8897 | ||
8898 | case MODE_CC: | |
8899 | x = get_last_value (XEXP (x, 0)); | |
8900 | return (x && GET_CODE (x) == COMPARE | |
8901 | && GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) == MODE_INT); | |
8902 | } | |
8903 | ||
8904 | return 0; | |
8905 | } | |
8906 | \f | |
8907 | /* Utility function for following routine. Called when X is part of a value | |
8908 | being stored into reg_last_set_value. Sets reg_last_set_table_tick | |
8909 | for each register mentioned. Similar to mention_regs in cse.c */ | |
8910 | ||
8911 | static void | |
8912 | update_table_tick (x) | |
8913 | rtx x; | |
8914 | { | |
8915 | register enum rtx_code code = GET_CODE (x); | |
8916 | register char *fmt = GET_RTX_FORMAT (code); | |
8917 | register int i; | |
8918 | ||
8919 | if (code == REG) | |
8920 | { | |
8921 | int regno = REGNO (x); | |
8922 | int endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
8923 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
8924 | ||
8925 | for (i = regno; i < endregno; i++) | |
8926 | reg_last_set_table_tick[i] = label_tick; | |
8927 | ||
8928 | return; | |
8929 | } | |
8930 | ||
8931 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
8932 | /* Note that we can't have an "E" in values stored; see | |
8933 | get_last_value_validate. */ | |
8934 | if (fmt[i] == 'e') | |
8935 | update_table_tick (XEXP (x, i)); | |
8936 | } | |
8937 | ||
8938 | /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we | |
8939 | are saying that the register is clobbered and we no longer know its | |
8940 | value. If INSN is zero, don't update reg_last_set; this is only permitted | |
8941 | with VALUE also zero and is used to invalidate the register. */ | |
8942 | ||
8943 | static void | |
8944 | record_value_for_reg (reg, insn, value) | |
8945 | rtx reg; | |
8946 | rtx insn; | |
8947 | rtx value; | |
8948 | { | |
8949 | int regno = REGNO (reg); | |
8950 | int endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
8951 | ? HARD_REGNO_NREGS (regno, GET_MODE (reg)) : 1); | |
8952 | int i; | |
8953 | ||
8954 | /* If VALUE contains REG and we have a previous value for REG, substitute | |
8955 | the previous value. */ | |
8956 | if (value && insn && reg_overlap_mentioned_p (reg, value)) | |
8957 | { | |
8958 | rtx tem; | |
8959 | ||
8960 | /* Set things up so get_last_value is allowed to see anything set up to | |
8961 | our insn. */ | |
8962 | subst_low_cuid = INSN_CUID (insn); | |
8963 | tem = get_last_value (reg); | |
8964 | ||
8965 | if (tem) | |
8966 | value = replace_rtx (copy_rtx (value), reg, tem); | |
8967 | } | |
8968 | ||
8969 | /* For each register modified, show we don't know its value, that | |
8970 | its value has been updated, and that we don't know the location of | |
8971 | the death of the register. */ | |
8972 | for (i = regno; i < endregno; i ++) | |
8973 | { | |
8974 | if (insn) | |
8975 | reg_last_set[i] = insn; | |
8976 | reg_last_set_value[i] = 0; | |
8977 | reg_last_death[i] = 0; | |
8978 | } | |
8979 | ||
8980 | /* Mark registers that are being referenced in this value. */ | |
8981 | if (value) | |
8982 | update_table_tick (value); | |
8983 | ||
8984 | /* Now update the status of each register being set. | |
8985 | If someone is using this register in this block, set this register | |
8986 | to invalid since we will get confused between the two lives in this | |
8987 | basic block. This makes using this register always invalid. In cse, we | |
8988 | scan the table to invalidate all entries using this register, but this | |
8989 | is too much work for us. */ | |
8990 | ||
8991 | for (i = regno; i < endregno; i++) | |
8992 | { | |
8993 | reg_last_set_label[i] = label_tick; | |
8994 | if (value && reg_last_set_table_tick[i] == label_tick) | |
8995 | reg_last_set_invalid[i] = 1; | |
8996 | else | |
8997 | reg_last_set_invalid[i] = 0; | |
8998 | } | |
8999 | ||
9000 | /* The value being assigned might refer to X (like in "x++;"). In that | |
9001 | case, we must replace it with (clobber (const_int 0)) to prevent | |
9002 | infinite loops. */ | |
9003 | if (value && ! get_last_value_validate (&value, | |
9004 | reg_last_set_label[regno], 0)) | |
9005 | { | |
9006 | value = copy_rtx (value); | |
9007 | if (! get_last_value_validate (&value, reg_last_set_label[regno], 1)) | |
9008 | value = 0; | |
9009 | } | |
9010 | ||
9011 | /* For the main register being modified, update the value, the mode, the | |
9012 | nonzero bits, and the number of sign bit copies. */ | |
9013 | ||
9014 | reg_last_set_value[regno] = value; | |
9015 | ||
9016 | if (value) | |
9017 | { | |
9018 | subst_low_cuid = INSN_CUID (insn); | |
9019 | reg_last_set_mode[regno] = GET_MODE (reg); | |
9020 | reg_last_set_nonzero_bits[regno] = nonzero_bits (value, GET_MODE (reg)); | |
9021 | reg_last_set_sign_bit_copies[regno] | |
9022 | = num_sign_bit_copies (value, GET_MODE (reg)); | |
9023 | } | |
9024 | } | |
9025 | ||
9026 | /* Used for communication between the following two routines. */ | |
9027 | static rtx record_dead_insn; | |
9028 | ||
9029 | /* Called via note_stores from record_dead_and_set_regs to handle one | |
9030 | SET or CLOBBER in an insn. */ | |
9031 | ||
9032 | static void | |
9033 | record_dead_and_set_regs_1 (dest, setter) | |
9034 | rtx dest, setter; | |
9035 | { | |
9036 | if (GET_CODE (dest) == REG) | |
9037 | { | |
9038 | /* If we are setting the whole register, we know its value. Otherwise | |
9039 | show that we don't know the value. We can handle SUBREG in | |
9040 | some cases. */ | |
9041 | if (GET_CODE (setter) == SET && dest == SET_DEST (setter)) | |
9042 | record_value_for_reg (dest, record_dead_insn, SET_SRC (setter)); | |
9043 | else if (GET_CODE (setter) == SET | |
9044 | && GET_CODE (SET_DEST (setter)) == SUBREG | |
9045 | && SUBREG_REG (SET_DEST (setter)) == dest | |
9046 | && subreg_lowpart_p (SET_DEST (setter))) | |
9047 | record_value_for_reg (dest, record_dead_insn, | |
9048 | gen_lowpart_for_combine (GET_MODE (dest), | |
9049 | SET_SRC (setter))); | |
9050 | else | |
9051 | record_value_for_reg (dest, record_dead_insn, NULL_RTX); | |
9052 | } | |
9053 | else if (GET_CODE (dest) == MEM | |
9054 | /* Ignore pushes, they clobber nothing. */ | |
9055 | && ! push_operand (dest, GET_MODE (dest))) | |
9056 | mem_last_set = INSN_CUID (record_dead_insn); | |
9057 | } | |
9058 | ||
9059 | /* Update the records of when each REG was most recently set or killed | |
9060 | for the things done by INSN. This is the last thing done in processing | |
9061 | INSN in the combiner loop. | |
9062 | ||
9063 | We update reg_last_set, reg_last_set_value, reg_last_death, and also the | |
9064 | similar information mem_last_set (which insn most recently modified memory) | |
9065 | and last_call_cuid (which insn was the most recent subroutine call). */ | |
9066 | ||
9067 | static void | |
9068 | record_dead_and_set_regs (insn) | |
9069 | rtx insn; | |
9070 | { | |
9071 | register rtx link; | |
9072 | int i; | |
9073 | ||
9074 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
9075 | { | |
9076 | if (REG_NOTE_KIND (link) == REG_DEAD | |
9077 | && GET_CODE (XEXP (link, 0)) == REG) | |
9078 | { | |
9079 | int regno = REGNO (XEXP (link, 0)); | |
9080 | int endregno | |
9081 | = regno + (regno < FIRST_PSEUDO_REGISTER | |
9082 | ? HARD_REGNO_NREGS (regno, GET_MODE (XEXP (link, 0))) | |
9083 | : 1); | |
9084 | ||
9085 | for (i = regno; i < endregno; i++) | |
9086 | reg_last_death[i] = insn; | |
9087 | } | |
9088 | else if (REG_NOTE_KIND (link) == REG_INC) | |
9089 | record_value_for_reg (XEXP (link, 0), insn, NULL_RTX); | |
9090 | } | |
9091 | ||
9092 | if (GET_CODE (insn) == CALL_INSN) | |
9093 | { | |
9094 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
9095 | if (call_used_regs[i]) | |
9096 | { | |
9097 | reg_last_set_value[i] = 0; | |
9098 | reg_last_death[i] = 0; | |
9099 | } | |
9100 | ||
9101 | last_call_cuid = mem_last_set = INSN_CUID (insn); | |
9102 | } | |
9103 | ||
9104 | record_dead_insn = insn; | |
9105 | note_stores (PATTERN (insn), record_dead_and_set_regs_1); | |
9106 | } | |
9107 | \f | |
9108 | /* Utility routine for the following function. Verify that all the registers | |
9109 | mentioned in *LOC are valid when *LOC was part of a value set when | |
9110 | label_tick == TICK. Return 0 if some are not. | |
9111 | ||
9112 | If REPLACE is non-zero, replace the invalid reference with | |
9113 | (clobber (const_int 0)) and return 1. This replacement is useful because | |
9114 | we often can get useful information about the form of a value (e.g., if | |
9115 | it was produced by a shift that always produces -1 or 0) even though | |
9116 | we don't know exactly what registers it was produced from. */ | |
9117 | ||
9118 | static int | |
9119 | get_last_value_validate (loc, tick, replace) | |
9120 | rtx *loc; | |
9121 | int tick; | |
9122 | int replace; | |
9123 | { | |
9124 | rtx x = *loc; | |
9125 | char *fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
9126 | int len = GET_RTX_LENGTH (GET_CODE (x)); | |
9127 | int i; | |
9128 | ||
9129 | if (GET_CODE (x) == REG) | |
9130 | { | |
9131 | int regno = REGNO (x); | |
9132 | int endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
9133 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
9134 | int j; | |
9135 | ||
9136 | for (j = regno; j < endregno; j++) | |
9137 | if (reg_last_set_invalid[j] | |
9138 | /* If this is a pseudo-register that was only set once, it is | |
9139 | always valid. */ | |
9140 | || (! (regno >= FIRST_PSEUDO_REGISTER && reg_n_sets[regno] == 1) | |
9141 | && reg_last_set_label[j] > tick)) | |
9142 | { | |
9143 | if (replace) | |
9144 | *loc = gen_rtx (CLOBBER, GET_MODE (x), const0_rtx); | |
9145 | return replace; | |
9146 | } | |
9147 | ||
9148 | return 1; | |
9149 | } | |
9150 | ||
9151 | for (i = 0; i < len; i++) | |
9152 | if ((fmt[i] == 'e' | |
9153 | && get_last_value_validate (&XEXP (x, i), tick, replace) == 0) | |
9154 | /* Don't bother with these. They shouldn't occur anyway. */ | |
9155 | || fmt[i] == 'E') | |
9156 | return 0; | |
9157 | ||
9158 | /* If we haven't found a reason for it to be invalid, it is valid. */ | |
9159 | return 1; | |
9160 | } | |
9161 | ||
9162 | /* Get the last value assigned to X, if known. Some registers | |
9163 | in the value may be replaced with (clobber (const_int 0)) if their value | |
9164 | is known longer known reliably. */ | |
9165 | ||
9166 | static rtx | |
9167 | get_last_value (x) | |
9168 | rtx x; | |
9169 | { | |
9170 | int regno; | |
9171 | rtx value; | |
9172 | ||
9173 | /* If this is a non-paradoxical SUBREG, get the value of its operand and | |
9174 | then convert it to the desired mode. If this is a paradoxical SUBREG, | |
9175 | we cannot predict what values the "extra" bits might have. */ | |
9176 | if (GET_CODE (x) == SUBREG | |
9177 | && subreg_lowpart_p (x) | |
9178 | && (GET_MODE_SIZE (GET_MODE (x)) | |
9179 | <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
9180 | && (value = get_last_value (SUBREG_REG (x))) != 0) | |
9181 | return gen_lowpart_for_combine (GET_MODE (x), value); | |
9182 | ||
9183 | if (GET_CODE (x) != REG) | |
9184 | return 0; | |
9185 | ||
9186 | regno = REGNO (x); | |
9187 | value = reg_last_set_value[regno]; | |
9188 | ||
9189 | /* If we don't have a value or if it isn't for this basic block, return 0. */ | |
9190 | ||
9191 | if (value == 0 | |
9192 | || (reg_n_sets[regno] != 1 | |
9193 | && reg_last_set_label[regno] != label_tick)) | |
9194 | return 0; | |
9195 | ||
9196 | /* If the value was set in a later insn that the ones we are processing, | |
9197 | we can't use it even if the register was only set once, but make a quick | |
9198 | check to see if the previous insn set it to something. This is commonly | |
9199 | the case when the same pseudo is used by repeated insns. */ | |
9200 | ||
9201 | if (INSN_CUID (reg_last_set[regno]) >= subst_low_cuid) | |
9202 | { | |
9203 | rtx insn, set; | |
9204 | ||
9205 | for (insn = prev_nonnote_insn (subst_insn); | |
9206 | insn && INSN_CUID (insn) >= subst_low_cuid; | |
9207 | insn = prev_nonnote_insn (insn)) | |
9208 | ; | |
9209 | ||
9210 | if (insn | |
9211 | && (set = single_set (insn)) != 0 | |
9212 | && rtx_equal_p (SET_DEST (set), x)) | |
9213 | { | |
9214 | value = SET_SRC (set); | |
9215 | ||
9216 | /* Make sure that VALUE doesn't reference X. Replace any | |
9217 | expliit references with a CLOBBER. If there are any remaining | |
9218 | references (rare), don't use the value. */ | |
9219 | ||
9220 | if (reg_mentioned_p (x, value)) | |
9221 | value = replace_rtx (copy_rtx (value), x, | |
9222 | gen_rtx (CLOBBER, GET_MODE (x), const0_rtx)); | |
9223 | ||
9224 | if (reg_overlap_mentioned_p (x, value)) | |
9225 | return 0; | |
9226 | } | |
9227 | else | |
9228 | return 0; | |
9229 | } | |
9230 | ||
9231 | /* If the value has all its registers valid, return it. */ | |
9232 | if (get_last_value_validate (&value, reg_last_set_label[regno], 0)) | |
9233 | return value; | |
9234 | ||
9235 | /* Otherwise, make a copy and replace any invalid register with | |
9236 | (clobber (const_int 0)). If that fails for some reason, return 0. */ | |
9237 | ||
9238 | value = copy_rtx (value); | |
9239 | if (get_last_value_validate (&value, reg_last_set_label[regno], 1)) | |
9240 | return value; | |
9241 | ||
9242 | return 0; | |
9243 | } | |
9244 | \f | |
9245 | /* Return nonzero if expression X refers to a REG or to memory | |
9246 | that is set in an instruction more recent than FROM_CUID. */ | |
9247 | ||
9248 | static int | |
9249 | use_crosses_set_p (x, from_cuid) | |
9250 | register rtx x; | |
9251 | int from_cuid; | |
9252 | { | |
9253 | register char *fmt; | |
9254 | register int i; | |
9255 | register enum rtx_code code = GET_CODE (x); | |
9256 | ||
9257 | if (code == REG) | |
9258 | { | |
9259 | register int regno = REGNO (x); | |
9260 | #ifdef PUSH_ROUNDING | |
9261 | /* Don't allow uses of the stack pointer to be moved, | |
9262 | because we don't know whether the move crosses a push insn. */ | |
9263 | if (regno == STACK_POINTER_REGNUM) | |
9264 | return 1; | |
9265 | #endif | |
9266 | return (reg_last_set[regno] | |
9267 | && INSN_CUID (reg_last_set[regno]) > from_cuid); | |
9268 | } | |
9269 | ||
9270 | if (code == MEM && mem_last_set > from_cuid) | |
9271 | return 1; | |
9272 | ||
9273 | fmt = GET_RTX_FORMAT (code); | |
9274 | ||
9275 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
9276 | { | |
9277 | if (fmt[i] == 'E') | |
9278 | { | |
9279 | register int j; | |
9280 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9281 | if (use_crosses_set_p (XVECEXP (x, i, j), from_cuid)) | |
9282 | return 1; | |
9283 | } | |
9284 | else if (fmt[i] == 'e' | |
9285 | && use_crosses_set_p (XEXP (x, i), from_cuid)) | |
9286 | return 1; | |
9287 | } | |
9288 | return 0; | |
9289 | } | |
9290 | \f | |
9291 | /* Define three variables used for communication between the following | |
9292 | routines. */ | |
9293 | ||
9294 | static int reg_dead_regno, reg_dead_endregno; | |
9295 | static int reg_dead_flag; | |
9296 | ||
9297 | /* Function called via note_stores from reg_dead_at_p. | |
9298 | ||
9299 | If DEST is within [reg_dead_rengno, reg_dead_endregno), set | |
9300 | reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */ | |
9301 | ||
9302 | static void | |
9303 | reg_dead_at_p_1 (dest, x) | |
9304 | rtx dest; | |
9305 | rtx x; | |
9306 | { | |
9307 | int regno, endregno; | |
9308 | ||
9309 | if (GET_CODE (dest) != REG) | |
9310 | return; | |
9311 | ||
9312 | regno = REGNO (dest); | |
9313 | endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
9314 | ? HARD_REGNO_NREGS (regno, GET_MODE (dest)) : 1); | |
9315 | ||
9316 | if (reg_dead_endregno > regno && reg_dead_regno < endregno) | |
9317 | reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1; | |
9318 | } | |
9319 | ||
9320 | /* Return non-zero if REG is known to be dead at INSN. | |
9321 | ||
9322 | We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER | |
9323 | referencing REG, it is dead. If we hit a SET referencing REG, it is | |
9324 | live. Otherwise, see if it is live or dead at the start of the basic | |
9325 | block we are in. */ | |
9326 | ||
9327 | static int | |
9328 | reg_dead_at_p (reg, insn) | |
9329 | rtx reg; | |
9330 | rtx insn; | |
9331 | { | |
9332 | int block, i; | |
9333 | ||
9334 | /* Set variables for reg_dead_at_p_1. */ | |
9335 | reg_dead_regno = REGNO (reg); | |
9336 | reg_dead_endregno = reg_dead_regno + (reg_dead_regno < FIRST_PSEUDO_REGISTER | |
9337 | ? HARD_REGNO_NREGS (reg_dead_regno, | |
9338 | GET_MODE (reg)) | |
9339 | : 1); | |
9340 | ||
9341 | reg_dead_flag = 0; | |
9342 | ||
9343 | /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, label, or | |
9344 | beginning of function. */ | |
9345 | for (; insn && GET_CODE (insn) != CODE_LABEL; | |
9346 | insn = prev_nonnote_insn (insn)) | |
9347 | { | |
9348 | note_stores (PATTERN (insn), reg_dead_at_p_1); | |
9349 | if (reg_dead_flag) | |
9350 | return reg_dead_flag == 1 ? 1 : 0; | |
9351 | ||
9352 | if (find_regno_note (insn, REG_DEAD, reg_dead_regno)) | |
9353 | return 1; | |
9354 | } | |
9355 | ||
9356 | /* Get the basic block number that we were in. */ | |
9357 | if (insn == 0) | |
9358 | block = 0; | |
9359 | else | |
9360 | { | |
9361 | for (block = 0; block < n_basic_blocks; block++) | |
9362 | if (insn == basic_block_head[block]) | |
9363 | break; | |
9364 | ||
9365 | if (block == n_basic_blocks) | |
9366 | return 0; | |
9367 | } | |
9368 | ||
9369 | for (i = reg_dead_regno; i < reg_dead_endregno; i++) | |
9370 | if (basic_block_live_at_start[block][i / REGSET_ELT_BITS] | |
9371 | & ((REGSET_ELT_TYPE) 1 << (i % REGSET_ELT_BITS))) | |
9372 | return 0; | |
9373 | ||
9374 | return 1; | |
9375 | } | |
9376 | \f | |
9377 | /* Remove register number REGNO from the dead registers list of INSN. | |
9378 | ||
9379 | Return the note used to record the death, if there was one. */ | |
9380 | ||
9381 | rtx | |
9382 | remove_death (regno, insn) | |
9383 | int regno; | |
9384 | rtx insn; | |
9385 | { | |
9386 | register rtx note = find_regno_note (insn, REG_DEAD, regno); | |
9387 | ||
9388 | if (note) | |
9389 | { | |
9390 | reg_n_deaths[regno]--; | |
9391 | remove_note (insn, note); | |
9392 | } | |
9393 | ||
9394 | return note; | |
9395 | } | |
9396 | ||
9397 | /* For each register (hardware or pseudo) used within expression X, if its | |
9398 | death is in an instruction with cuid between FROM_CUID (inclusive) and | |
9399 | TO_INSN (exclusive), put a REG_DEAD note for that register in the | |
9400 | list headed by PNOTES. | |
9401 | ||
9402 | This is done when X is being merged by combination into TO_INSN. These | |
9403 | notes will then be distributed as needed. */ | |
9404 | ||
9405 | static void | |
9406 | move_deaths (x, from_cuid, to_insn, pnotes) | |
9407 | rtx x; | |
9408 | int from_cuid; | |
9409 | rtx to_insn; | |
9410 | rtx *pnotes; | |
9411 | { | |
9412 | register char *fmt; | |
9413 | register int len, i; | |
9414 | register enum rtx_code code = GET_CODE (x); | |
9415 | ||
9416 | if (code == REG) | |
9417 | { | |
9418 | register int regno = REGNO (x); | |
9419 | register rtx where_dead = reg_last_death[regno]; | |
9420 | ||
9421 | if (where_dead && INSN_CUID (where_dead) >= from_cuid | |
9422 | && INSN_CUID (where_dead) < INSN_CUID (to_insn)) | |
9423 | { | |
9424 | rtx note = remove_death (regno, where_dead); | |
9425 | ||
9426 | /* It is possible for the call above to return 0. This can occur | |
9427 | when reg_last_death points to I2 or I1 that we combined with. | |
9428 | In that case make a new note. | |
9429 | ||
9430 | We must also check for the case where X is a hard register | |
9431 | and NOTE is a death note for a range of hard registers | |
9432 | including X. In that case, we must put REG_DEAD notes for | |
9433 | the remaining registers in place of NOTE. */ | |
9434 | ||
9435 | if (note != 0 && regno < FIRST_PSEUDO_REGISTER | |
9436 | && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0))) | |
9437 | != GET_MODE_SIZE (GET_MODE (x)))) | |
9438 | { | |
9439 | int deadregno = REGNO (XEXP (note, 0)); | |
9440 | int deadend | |
9441 | = (deadregno + HARD_REGNO_NREGS (deadregno, | |
9442 | GET_MODE (XEXP (note, 0)))); | |
9443 | int ourend = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
9444 | int i; | |
9445 | ||
9446 | for (i = deadregno; i < deadend; i++) | |
9447 | if (i < regno || i >= ourend) | |
9448 | REG_NOTES (where_dead) | |
9449 | = gen_rtx (EXPR_LIST, REG_DEAD, | |
9450 | gen_rtx (REG, word_mode, i), | |
9451 | REG_NOTES (where_dead)); | |
9452 | } | |
9453 | ||
9454 | if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x)) | |
9455 | { | |
9456 | XEXP (note, 1) = *pnotes; | |
9457 | *pnotes = note; | |
9458 | } | |
9459 | else | |
9460 | *pnotes = gen_rtx (EXPR_LIST, REG_DEAD, x, *pnotes); | |
9461 | ||
9462 | reg_n_deaths[regno]++; | |
9463 | } | |
9464 | ||
9465 | return; | |
9466 | } | |
9467 | ||
9468 | else if (GET_CODE (x) == SET) | |
9469 | { | |
9470 | rtx dest = SET_DEST (x); | |
9471 | ||
9472 | move_deaths (SET_SRC (x), from_cuid, to_insn, pnotes); | |
9473 | ||
9474 | /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG | |
9475 | that accesses one word of a multi-word item, some | |
9476 | piece of everything register in the expression is used by | |
9477 | this insn, so remove any old death. */ | |
9478 | ||
9479 | if (GET_CODE (dest) == ZERO_EXTRACT | |
9480 | || GET_CODE (dest) == STRICT_LOW_PART | |
9481 | || (GET_CODE (dest) == SUBREG | |
9482 | && (((GET_MODE_SIZE (GET_MODE (dest)) | |
9483 | + UNITS_PER_WORD - 1) / UNITS_PER_WORD) | |
9484 | == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) | |
9485 | + UNITS_PER_WORD - 1) / UNITS_PER_WORD)))) | |
9486 | { | |
9487 | move_deaths (dest, from_cuid, to_insn, pnotes); | |
9488 | return; | |
9489 | } | |
9490 | ||
9491 | /* If this is some other SUBREG, we know it replaces the entire | |
9492 | value, so use that as the destination. */ | |
9493 | if (GET_CODE (dest) == SUBREG) | |
9494 | dest = SUBREG_REG (dest); | |
9495 | ||
9496 | /* If this is a MEM, adjust deaths of anything used in the address. | |
9497 | For a REG (the only other possibility), the entire value is | |
9498 | being replaced so the old value is not used in this insn. */ | |
9499 | ||
9500 | if (GET_CODE (dest) == MEM) | |
9501 | move_deaths (XEXP (dest, 0), from_cuid, to_insn, pnotes); | |
9502 | return; | |
9503 | } | |
9504 | ||
9505 | else if (GET_CODE (x) == CLOBBER) | |
9506 | return; | |
9507 | ||
9508 | len = GET_RTX_LENGTH (code); | |
9509 | fmt = GET_RTX_FORMAT (code); | |
9510 | ||
9511 | for (i = 0; i < len; i++) | |
9512 | { | |
9513 | if (fmt[i] == 'E') | |
9514 | { | |
9515 | register int j; | |
9516 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9517 | move_deaths (XVECEXP (x, i, j), from_cuid, to_insn, pnotes); | |
9518 | } | |
9519 | else if (fmt[i] == 'e') | |
9520 | move_deaths (XEXP (x, i), from_cuid, to_insn, pnotes); | |
9521 | } | |
9522 | } | |
9523 | \f | |
9524 | /* Return 1 if X is the target of a bit-field assignment in BODY, the | |
9525 | pattern of an insn. X must be a REG. */ | |
9526 | ||
9527 | static int | |
9528 | reg_bitfield_target_p (x, body) | |
9529 | rtx x; | |
9530 | rtx body; | |
9531 | { | |
9532 | int i; | |
9533 | ||
9534 | if (GET_CODE (body) == SET) | |
9535 | { | |
9536 | rtx dest = SET_DEST (body); | |
9537 | rtx target; | |
9538 | int regno, tregno, endregno, endtregno; | |
9539 | ||
9540 | if (GET_CODE (dest) == ZERO_EXTRACT) | |
9541 | target = XEXP (dest, 0); | |
9542 | else if (GET_CODE (dest) == STRICT_LOW_PART) | |
9543 | target = SUBREG_REG (XEXP (dest, 0)); | |
9544 | else | |
9545 | return 0; | |
9546 | ||
9547 | if (GET_CODE (target) == SUBREG) | |
9548 | target = SUBREG_REG (target); | |
9549 | ||
9550 | if (GET_CODE (target) != REG) | |
9551 | return 0; | |
9552 | ||
9553 | tregno = REGNO (target), regno = REGNO (x); | |
9554 | if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER) | |
9555 | return target == x; | |
9556 | ||
9557 | endtregno = tregno + HARD_REGNO_NREGS (tregno, GET_MODE (target)); | |
9558 | endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
9559 | ||
9560 | return endregno > tregno && regno < endtregno; | |
9561 | } | |
9562 | ||
9563 | else if (GET_CODE (body) == PARALLEL) | |
9564 | for (i = XVECLEN (body, 0) - 1; i >= 0; i--) | |
9565 | if (reg_bitfield_target_p (x, XVECEXP (body, 0, i))) | |
9566 | return 1; | |
9567 | ||
9568 | return 0; | |
9569 | } | |
9570 | \f | |
9571 | /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them | |
9572 | as appropriate. I3 and I2 are the insns resulting from the combination | |
9573 | insns including FROM (I2 may be zero). | |
9574 | ||
9575 | ELIM_I2 and ELIM_I1 are either zero or registers that we know will | |
9576 | not need REG_DEAD notes because they are being substituted for. This | |
9577 | saves searching in the most common cases. | |
9578 | ||
9579 | Each note in the list is either ignored or placed on some insns, depending | |
9580 | on the type of note. */ | |
9581 | ||
9582 | static void | |
9583 | distribute_notes (notes, from_insn, i3, i2, elim_i2, elim_i1) | |
9584 | rtx notes; | |
9585 | rtx from_insn; | |
9586 | rtx i3, i2; | |
9587 | rtx elim_i2, elim_i1; | |
9588 | { | |
9589 | rtx note, next_note; | |
9590 | rtx tem; | |
9591 | ||
9592 | for (note = notes; note; note = next_note) | |
9593 | { | |
9594 | rtx place = 0, place2 = 0; | |
9595 | ||
9596 | /* If this NOTE references a pseudo register, ensure it references | |
9597 | the latest copy of that register. */ | |
9598 | if (XEXP (note, 0) && GET_CODE (XEXP (note, 0)) == REG | |
9599 | && REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER) | |
9600 | XEXP (note, 0) = regno_reg_rtx[REGNO (XEXP (note, 0))]; | |
9601 | ||
9602 | next_note = XEXP (note, 1); | |
9603 | switch (REG_NOTE_KIND (note)) | |
9604 | { | |
9605 | case REG_UNUSED: | |
9606 | /* If this register is set or clobbered in I3, put the note there | |
9607 | unless there is one already. */ | |
9608 | if (reg_set_p (XEXP (note, 0), PATTERN (i3))) | |
9609 | { | |
9610 | if (! (GET_CODE (XEXP (note, 0)) == REG | |
9611 | ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0))) | |
9612 | : find_reg_note (i3, REG_UNUSED, XEXP (note, 0)))) | |
9613 | place = i3; | |
9614 | } | |
9615 | /* Otherwise, if this register is used by I3, then this register | |
9616 | now dies here, so we must put a REG_DEAD note here unless there | |
9617 | is one already. */ | |
9618 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)) | |
9619 | && ! (GET_CODE (XEXP (note, 0)) == REG | |
9620 | ? find_regno_note (i3, REG_DEAD, REGNO (XEXP (note, 0))) | |
9621 | : find_reg_note (i3, REG_DEAD, XEXP (note, 0)))) | |
9622 | { | |
9623 | PUT_REG_NOTE_KIND (note, REG_DEAD); | |
9624 | place = i3; | |
9625 | } | |
9626 | break; | |
9627 | ||
9628 | case REG_EQUAL: | |
9629 | case REG_EQUIV: | |
9630 | case REG_NONNEG: | |
9631 | /* These notes say something about results of an insn. We can | |
9632 | only support them if they used to be on I3 in which case they | |
9633 | remain on I3. Otherwise they are ignored. | |
9634 | ||
9635 | If the note refers to an expression that is not a constant, we | |
9636 | must also ignore the note since we cannot tell whether the | |
9637 | equivalence is still true. It might be possible to do | |
9638 | slightly better than this (we only have a problem if I2DEST | |
9639 | or I1DEST is present in the expression), but it doesn't | |
9640 | seem worth the trouble. */ | |
9641 | ||
9642 | if (from_insn == i3 | |
9643 | && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0)))) | |
9644 | place = i3; | |
9645 | break; | |
9646 | ||
9647 | case REG_INC: | |
9648 | case REG_NO_CONFLICT: | |
9649 | case REG_LABEL: | |
9650 | /* These notes say something about how a register is used. They must | |
9651 | be present on any use of the register in I2 or I3. */ | |
9652 | if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))) | |
9653 | place = i3; | |
9654 | ||
9655 | if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2))) | |
9656 | { | |
9657 | if (place) | |
9658 | place2 = i2; | |
9659 | else | |
9660 | place = i2; | |
9661 | } | |
9662 | break; | |
9663 | ||
9664 | case REG_WAS_0: | |
9665 | /* It is too much trouble to try to see if this note is still | |
9666 | correct in all situations. It is better to simply delete it. */ | |
9667 | break; | |
9668 | ||
9669 | case REG_RETVAL: | |
9670 | /* If the insn previously containing this note still exists, | |
9671 | put it back where it was. Otherwise move it to the previous | |
9672 | insn. Adjust the corresponding REG_LIBCALL note. */ | |
9673 | if (GET_CODE (from_insn) != NOTE) | |
9674 | place = from_insn; | |
9675 | else | |
9676 | { | |
9677 | tem = find_reg_note (XEXP (note, 0), REG_LIBCALL, NULL_RTX); | |
9678 | place = prev_real_insn (from_insn); | |
9679 | if (tem && place) | |
9680 | XEXP (tem, 0) = place; | |
9681 | } | |
9682 | break; | |
9683 | ||
9684 | case REG_LIBCALL: | |
9685 | /* This is handled similarly to REG_RETVAL. */ | |
9686 | if (GET_CODE (from_insn) != NOTE) | |
9687 | place = from_insn; | |
9688 | else | |
9689 | { | |
9690 | tem = find_reg_note (XEXP (note, 0), REG_RETVAL, NULL_RTX); | |
9691 | place = next_real_insn (from_insn); | |
9692 | if (tem && place) | |
9693 | XEXP (tem, 0) = place; | |
9694 | } | |
9695 | break; | |
9696 | ||
9697 | case REG_DEAD: | |
9698 | /* If the register is used as an input in I3, it dies there. | |
9699 | Similarly for I2, if it is non-zero and adjacent to I3. | |
9700 | ||
9701 | If the register is not used as an input in either I3 or I2 | |
9702 | and it is not one of the registers we were supposed to eliminate, | |
9703 | there are two possibilities. We might have a non-adjacent I2 | |
9704 | or we might have somehow eliminated an additional register | |
9705 | from a computation. For example, we might have had A & B where | |
9706 | we discover that B will always be zero. In this case we will | |
9707 | eliminate the reference to A. | |
9708 | ||
9709 | In both cases, we must search to see if we can find a previous | |
9710 | use of A and put the death note there. */ | |
9711 | ||
9712 | if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))) | |
9713 | place = i3; | |
9714 | else if (i2 != 0 && next_nonnote_insn (i2) == i3 | |
9715 | && reg_referenced_p (XEXP (note, 0), PATTERN (i2))) | |
9716 | place = i2; | |
9717 | ||
9718 | if (XEXP (note, 0) == elim_i2 || XEXP (note, 0) == elim_i1) | |
9719 | break; | |
9720 | ||
9721 | /* If the register is used in both I2 and I3 and it dies in I3, | |
9722 | we might have added another reference to it. If reg_n_refs | |
9723 | was 2, bump it to 3. This has to be correct since the | |
9724 | register must have been set somewhere. The reason this is | |
9725 | done is because local-alloc.c treats 2 references as a | |
9726 | special case. */ | |
9727 | ||
9728 | if (place == i3 && i2 != 0 && GET_CODE (XEXP (note, 0)) == REG | |
9729 | && reg_n_refs[REGNO (XEXP (note, 0))]== 2 | |
9730 | && reg_referenced_p (XEXP (note, 0), PATTERN (i2))) | |
9731 | reg_n_refs[REGNO (XEXP (note, 0))] = 3; | |
9732 | ||
9733 | if (place == 0) | |
9734 | for (tem = prev_nonnote_insn (i3); | |
9735 | tem && (GET_CODE (tem) == INSN | |
9736 | || GET_CODE (tem) == CALL_INSN); | |
9737 | tem = prev_nonnote_insn (tem)) | |
9738 | { | |
9739 | /* If the register is being set at TEM, see if that is all | |
9740 | TEM is doing. If so, delete TEM. Otherwise, make this | |
9741 | into a REG_UNUSED note instead. */ | |
9742 | if (reg_set_p (XEXP (note, 0), PATTERN (tem))) | |
9743 | { | |
9744 | rtx set = single_set (tem); | |
9745 | ||
9746 | /* Verify that it was the set, and not a clobber that | |
9747 | modified the register. */ | |
9748 | ||
9749 | if (set != 0 && ! side_effects_p (SET_SRC (set)) | |
9750 | && rtx_equal_p (XEXP (note, 0), SET_DEST (set))) | |
9751 | { | |
9752 | /* Move the notes and links of TEM elsewhere. | |
9753 | This might delete other dead insns recursively. | |
9754 | First set the pattern to something that won't use | |
9755 | any register. */ | |
9756 | ||
9757 | PATTERN (tem) = pc_rtx; | |
9758 | ||
9759 | distribute_notes (REG_NOTES (tem), tem, tem, | |
9760 | NULL_RTX, NULL_RTX, NULL_RTX); | |
9761 | distribute_links (LOG_LINKS (tem)); | |
9762 | ||
9763 | PUT_CODE (tem, NOTE); | |
9764 | NOTE_LINE_NUMBER (tem) = NOTE_INSN_DELETED; | |
9765 | NOTE_SOURCE_FILE (tem) = 0; | |
9766 | } | |
9767 | else | |
9768 | { | |
9769 | PUT_REG_NOTE_KIND (note, REG_UNUSED); | |
9770 | ||
9771 | /* If there isn't already a REG_UNUSED note, put one | |
9772 | here. */ | |
9773 | if (! find_regno_note (tem, REG_UNUSED, | |
9774 | REGNO (XEXP (note, 0)))) | |
9775 | place = tem; | |
9776 | break; | |
9777 | } | |
9778 | } | |
9779 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem))) | |
9780 | { | |
9781 | place = tem; | |
9782 | break; | |
9783 | } | |
9784 | } | |
9785 | ||
9786 | /* If the register is set or already dead at PLACE, we needn't do | |
9787 | anything with this note if it is still a REG_DEAD note. | |
9788 | ||
9789 | Note that we cannot use just `dead_or_set_p' here since we can | |
9790 | convert an assignment to a register into a bit-field assignment. | |
9791 | Therefore, we must also omit the note if the register is the | |
9792 | target of a bitfield assignment. */ | |
9793 | ||
9794 | if (place && REG_NOTE_KIND (note) == REG_DEAD) | |
9795 | { | |
9796 | int regno = REGNO (XEXP (note, 0)); | |
9797 | ||
9798 | if (dead_or_set_p (place, XEXP (note, 0)) | |
9799 | || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place))) | |
9800 | { | |
9801 | /* Unless the register previously died in PLACE, clear | |
9802 | reg_last_death. [I no longer understand why this is | |
9803 | being done.] */ | |
9804 | if (reg_last_death[regno] != place) | |
9805 | reg_last_death[regno] = 0; | |
9806 | place = 0; | |
9807 | } | |
9808 | else | |
9809 | reg_last_death[regno] = place; | |
9810 | ||
9811 | /* If this is a death note for a hard reg that is occupying | |
9812 | multiple registers, ensure that we are still using all | |
9813 | parts of the object. If we find a piece of the object | |
9814 | that is unused, we must add a USE for that piece before | |
9815 | PLACE and put the appropriate REG_DEAD note on it. | |
9816 | ||
9817 | An alternative would be to put a REG_UNUSED for the pieces | |
9818 | on the insn that set the register, but that can't be done if | |
9819 | it is not in the same block. It is simpler, though less | |
9820 | efficient, to add the USE insns. */ | |
9821 | ||
9822 | if (place && regno < FIRST_PSEUDO_REGISTER | |
9823 | && HARD_REGNO_NREGS (regno, GET_MODE (XEXP (note, 0))) > 1) | |
9824 | { | |
9825 | int endregno | |
9826 | = regno + HARD_REGNO_NREGS (regno, | |
9827 | GET_MODE (XEXP (note, 0))); | |
9828 | int all_used = 1; | |
9829 | int i; | |
9830 | ||
9831 | for (i = regno; i < endregno; i++) | |
9832 | if (! refers_to_regno_p (i, i + 1, PATTERN (place), 0)) | |
9833 | { | |
9834 | rtx piece = gen_rtx (REG, word_mode, i); | |
9835 | rtx p; | |
9836 | ||
9837 | /* See if we already placed a USE note for this | |
9838 | register in front of PLACE. */ | |
9839 | for (p = place; | |
9840 | GET_CODE (PREV_INSN (p)) == INSN | |
9841 | && GET_CODE (PATTERN (PREV_INSN (p))) == USE; | |
9842 | p = PREV_INSN (p)) | |
9843 | if (rtx_equal_p (piece, | |
9844 | XEXP (PATTERN (PREV_INSN (p)), 0))) | |
9845 | { | |
9846 | p = 0; | |
9847 | break; | |
9848 | } | |
9849 | ||
9850 | if (p) | |
9851 | { | |
9852 | rtx use_insn | |
9853 | = emit_insn_before (gen_rtx (USE, VOIDmode, | |
9854 | piece), | |
9855 | p); | |
9856 | REG_NOTES (use_insn) | |
9857 | = gen_rtx (EXPR_LIST, REG_DEAD, piece, | |
9858 | REG_NOTES (use_insn)); | |
9859 | } | |
9860 | ||
9861 | all_used = 0; | |
9862 | } | |
9863 | ||
9864 | /* Check for the case where the register dying partially | |
9865 | overlaps the register set by this insn. */ | |
9866 | if (all_used) | |
9867 | for (i = regno; i < endregno; i++) | |
9868 | if (dead_or_set_regno_p (place, i)) | |
9869 | { | |
9870 | all_used = 0; | |
9871 | break; | |
9872 | } | |
9873 | ||
9874 | if (! all_used) | |
9875 | { | |
9876 | /* Put only REG_DEAD notes for pieces that are | |
9877 | still used and that are not already dead or set. */ | |
9878 | ||
9879 | for (i = regno; i < endregno; i++) | |
9880 | { | |
9881 | rtx piece = gen_rtx (REG, word_mode, i); | |
9882 | ||
9883 | if (reg_referenced_p (piece, PATTERN (place)) | |
9884 | && ! dead_or_set_p (place, piece) | |
9885 | && ! reg_bitfield_target_p (piece, | |
9886 | PATTERN (place))) | |
9887 | REG_NOTES (place) = gen_rtx (EXPR_LIST, REG_DEAD, | |
9888 | piece, | |
9889 | REG_NOTES (place)); | |
9890 | } | |
9891 | ||
9892 | place = 0; | |
9893 | } | |
9894 | } | |
9895 | } | |
9896 | break; | |
9897 | ||
9898 | default: | |
9899 | /* Any other notes should not be present at this point in the | |
9900 | compilation. */ | |
9901 | abort (); | |
9902 | } | |
9903 | ||
9904 | if (place) | |
9905 | { | |
9906 | XEXP (note, 1) = REG_NOTES (place); | |
9907 | REG_NOTES (place) = note; | |
9908 | } | |
9909 | else if ((REG_NOTE_KIND (note) == REG_DEAD | |
9910 | || REG_NOTE_KIND (note) == REG_UNUSED) | |
9911 | && GET_CODE (XEXP (note, 0)) == REG) | |
9912 | reg_n_deaths[REGNO (XEXP (note, 0))]--; | |
9913 | ||
9914 | if (place2) | |
9915 | { | |
9916 | if ((REG_NOTE_KIND (note) == REG_DEAD | |
9917 | || REG_NOTE_KIND (note) == REG_UNUSED) | |
9918 | && GET_CODE (XEXP (note, 0)) == REG) | |
9919 | reg_n_deaths[REGNO (XEXP (note, 0))]++; | |
9920 | ||
9921 | REG_NOTES (place2) = gen_rtx (GET_CODE (note), REG_NOTE_KIND (note), | |
9922 | XEXP (note, 0), REG_NOTES (place2)); | |
9923 | } | |
9924 | } | |
9925 | } | |
9926 | \f | |
9927 | /* Similarly to above, distribute the LOG_LINKS that used to be present on | |
9928 | I3, I2, and I1 to new locations. This is also called in one case to | |
9929 | add a link pointing at I3 when I3's destination is changed. */ | |
9930 | ||
9931 | static void | |
9932 | distribute_links (links) | |
9933 | rtx links; | |
9934 | { | |
9935 | rtx link, next_link; | |
9936 | ||
9937 | for (link = links; link; link = next_link) | |
9938 | { | |
9939 | rtx place = 0; | |
9940 | rtx insn; | |
9941 | rtx set, reg; | |
9942 | ||
9943 | next_link = XEXP (link, 1); | |
9944 | ||
9945 | /* If the insn that this link points to is a NOTE or isn't a single | |
9946 | set, ignore it. In the latter case, it isn't clear what we | |
9947 | can do other than ignore the link, since we can't tell which | |
9948 | register it was for. Such links wouldn't be used by combine | |
9949 | anyway. | |
9950 | ||
9951 | It is not possible for the destination of the target of the link to | |
9952 | have been changed by combine. The only potential of this is if we | |
9953 | replace I3, I2, and I1 by I3 and I2. But in that case the | |
9954 | destination of I2 also remains unchanged. */ | |
9955 | ||
9956 | if (GET_CODE (XEXP (link, 0)) == NOTE | |
9957 | || (set = single_set (XEXP (link, 0))) == 0) | |
9958 | continue; | |
9959 | ||
9960 | reg = SET_DEST (set); | |
9961 | while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT | |
9962 | || GET_CODE (reg) == SIGN_EXTRACT | |
9963 | || GET_CODE (reg) == STRICT_LOW_PART) | |
9964 | reg = XEXP (reg, 0); | |
9965 | ||
9966 | /* A LOG_LINK is defined as being placed on the first insn that uses | |
9967 | a register and points to the insn that sets the register. Start | |
9968 | searching at the next insn after the target of the link and stop | |
9969 | when we reach a set of the register or the end of the basic block. | |
9970 | ||
9971 | Note that this correctly handles the link that used to point from | |
9972 | I3 to I2. Also note that not much searching is typically done here | |
9973 | since most links don't point very far away. */ | |
9974 | ||
9975 | for (insn = NEXT_INSN (XEXP (link, 0)); | |
9976 | (insn && GET_CODE (insn) != CODE_LABEL | |
9977 | && GET_CODE (PREV_INSN (insn)) != JUMP_INSN); | |
9978 | insn = NEXT_INSN (insn)) | |
9979 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
9980 | && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
9981 | { | |
9982 | if (reg_referenced_p (reg, PATTERN (insn))) | |
9983 | place = insn; | |
9984 | break; | |
9985 | } | |
9986 | ||
9987 | /* If we found a place to put the link, place it there unless there | |
9988 | is already a link to the same insn as LINK at that point. */ | |
9989 | ||
9990 | if (place) | |
9991 | { | |
9992 | rtx link2; | |
9993 | ||
9994 | for (link2 = LOG_LINKS (place); link2; link2 = XEXP (link2, 1)) | |
9995 | if (XEXP (link2, 0) == XEXP (link, 0)) | |
9996 | break; | |
9997 | ||
9998 | if (link2 == 0) | |
9999 | { | |
10000 | XEXP (link, 1) = LOG_LINKS (place); | |
10001 | LOG_LINKS (place) = link; | |
10002 | } | |
10003 | } | |
10004 | } | |
10005 | } | |
10006 | \f | |
10007 | void | |
10008 | dump_combine_stats (file) | |
10009 | FILE *file; | |
10010 | { | |
10011 | fprintf | |
10012 | (file, | |
10013 | ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n", | |
10014 | combine_attempts, combine_merges, combine_extras, combine_successes); | |
10015 | } | |
10016 | ||
10017 | void | |
10018 | dump_combine_total_stats (file) | |
10019 | FILE *file; | |
10020 | { | |
10021 | fprintf | |
10022 | (file, | |
10023 | "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n", | |
10024 | total_attempts, total_merges, total_extras, total_successes); | |
10025 | } |