]>
Commit | Line | Data |
---|---|---|
230d793d | 1 | /* Optimize by combining instructions for GNU compiler. |
c85f7c16 | 2 | Copyright (C) 1987, 88, 92-97, 1998 Free Software Foundation, Inc. |
230d793d RS |
3 | |
4 | This file is part of GNU CC. | |
5 | ||
6 | GNU CC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8 | the Free Software Foundation; either version 2, or (at your option) | |
9 | any later version. | |
10 | ||
11 | GNU CC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GNU CC; see the file COPYING. If not, write to | |
940d9d63 RK |
18 | the Free Software Foundation, 59 Temple Place - Suite 330, |
19 | Boston, MA 02111-1307, USA. */ | |
230d793d RS |
20 | |
21 | ||
22 | /* This module is essentially the "combiner" phase of the U. of Arizona | |
23 | Portable Optimizer, but redone to work on our list-structured | |
24 | representation for RTL instead of their string representation. | |
25 | ||
26 | The LOG_LINKS of each insn identify the most recent assignment | |
27 | to each REG used in the insn. It is a list of previous insns, | |
28 | each of which contains a SET for a REG that is used in this insn | |
29 | and not used or set in between. LOG_LINKs never cross basic blocks. | |
30 | They were set up by the preceding pass (lifetime analysis). | |
31 | ||
32 | We try to combine each pair of insns joined by a logical link. | |
33 | We also try to combine triples of insns A, B and C when | |
34 | C has a link back to B and B has a link back to A. | |
35 | ||
36 | LOG_LINKS does not have links for use of the CC0. They don't | |
37 | need to, because the insn that sets the CC0 is always immediately | |
38 | before the insn that tests it. So we always regard a branch | |
39 | insn as having a logical link to the preceding insn. The same is true | |
40 | for an insn explicitly using CC0. | |
41 | ||
42 | We check (with use_crosses_set_p) to avoid combining in such a way | |
43 | as to move a computation to a place where its value would be different. | |
44 | ||
45 | Combination is done by mathematically substituting the previous | |
46 | insn(s) values for the regs they set into the expressions in | |
47 | the later insns that refer to these regs. If the result is a valid insn | |
48 | for our target machine, according to the machine description, | |
49 | we install it, delete the earlier insns, and update the data flow | |
50 | information (LOG_LINKS and REG_NOTES) for what we did. | |
51 | ||
52 | There are a few exceptions where the dataflow information created by | |
53 | flow.c aren't completely updated: | |
54 | ||
55 | - reg_live_length is not updated | |
56 | - reg_n_refs is not adjusted in the rare case when a register is | |
57 | no longer required in a computation | |
58 | - there are extremely rare cases (see distribute_regnotes) when a | |
59 | REG_DEAD note is lost | |
60 | - a LOG_LINKS entry that refers to an insn with multiple SETs may be | |
61 | removed because there is no way to know which register it was | |
62 | linking | |
63 | ||
64 | To simplify substitution, we combine only when the earlier insn(s) | |
65 | consist of only a single assignment. To simplify updating afterward, | |
66 | we never combine when a subroutine call appears in the middle. | |
67 | ||
68 | Since we do not represent assignments to CC0 explicitly except when that | |
69 | is all an insn does, there is no LOG_LINKS entry in an insn that uses | |
70 | the condition code for the insn that set the condition code. | |
71 | Fortunately, these two insns must be consecutive. | |
72 | Therefore, every JUMP_INSN is taken to have an implicit logical link | |
73 | to the preceding insn. This is not quite right, since non-jumps can | |
74 | also use the condition code; but in practice such insns would not | |
75 | combine anyway. */ | |
76 | ||
230d793d | 77 | #include "config.h" |
4f90e4a0 | 78 | #ifdef __STDC__ |
04fe4385 | 79 | #include <stdarg.h> |
4f90e4a0 | 80 | #else |
04fe4385 | 81 | #include <varargs.h> |
4f90e4a0 | 82 | #endif |
dfa3449b | 83 | |
670ee920 KG |
84 | /* stdio.h must precede rtl.h for FFS. */ |
85 | #include "system.h" | |
9c3b4c8b | 86 | |
230d793d RS |
87 | #include "rtl.h" |
88 | #include "flags.h" | |
89 | #include "regs.h" | |
55310dad | 90 | #include "hard-reg-set.h" |
230d793d RS |
91 | #include "basic-block.h" |
92 | #include "insn-config.h" | |
d6f4ec51 KG |
93 | /* Include expr.h after insn-config.h so we get HAVE_conditional_move. */ |
94 | #include "expr.h" | |
230d793d RS |
95 | #include "insn-flags.h" |
96 | #include "insn-codes.h" | |
97 | #include "insn-attr.h" | |
98 | #include "recog.h" | |
99 | #include "real.h" | |
2e107e9e | 100 | #include "toplev.h" |
230d793d RS |
101 | |
102 | /* It is not safe to use ordinary gen_lowpart in combine. | |
103 | Use gen_lowpart_for_combine instead. See comments there. */ | |
104 | #define gen_lowpart dont_use_gen_lowpart_you_dummy | |
105 | ||
106 | /* Number of attempts to combine instructions in this function. */ | |
107 | ||
108 | static int combine_attempts; | |
109 | ||
110 | /* Number of attempts that got as far as substitution in this function. */ | |
111 | ||
112 | static int combine_merges; | |
113 | ||
114 | /* Number of instructions combined with added SETs in this function. */ | |
115 | ||
116 | static int combine_extras; | |
117 | ||
118 | /* Number of instructions combined in this function. */ | |
119 | ||
120 | static int combine_successes; | |
121 | ||
122 | /* Totals over entire compilation. */ | |
123 | ||
124 | static int total_attempts, total_merges, total_extras, total_successes; | |
9210df58 | 125 | |
ddd5a7c1 | 126 | /* Define a default value for REVERSIBLE_CC_MODE. |
9210df58 RK |
127 | We can never assume that a condition code mode is safe to reverse unless |
128 | the md tells us so. */ | |
129 | #ifndef REVERSIBLE_CC_MODE | |
130 | #define REVERSIBLE_CC_MODE(MODE) 0 | |
131 | #endif | |
230d793d RS |
132 | \f |
133 | /* Vector mapping INSN_UIDs to cuids. | |
5089e22e | 134 | The cuids are like uids but increase monotonically always. |
230d793d RS |
135 | Combine always uses cuids so that it can compare them. |
136 | But actually renumbering the uids, which we used to do, | |
137 | proves to be a bad idea because it makes it hard to compare | |
138 | the dumps produced by earlier passes with those from later passes. */ | |
139 | ||
140 | static int *uid_cuid; | |
4255220d | 141 | static int max_uid_cuid; |
230d793d RS |
142 | |
143 | /* Get the cuid of an insn. */ | |
144 | ||
1427d6d2 RK |
145 | #define INSN_CUID(INSN) \ |
146 | (INSN_UID (INSN) > max_uid_cuid ? insn_cuid (INSN) : uid_cuid[INSN_UID (INSN)]) | |
230d793d RS |
147 | |
148 | /* Maximum register number, which is the size of the tables below. */ | |
149 | ||
150 | static int combine_max_regno; | |
151 | ||
152 | /* Record last point of death of (hard or pseudo) register n. */ | |
153 | ||
154 | static rtx *reg_last_death; | |
155 | ||
156 | /* Record last point of modification of (hard or pseudo) register n. */ | |
157 | ||
158 | static rtx *reg_last_set; | |
159 | ||
160 | /* Record the cuid of the last insn that invalidated memory | |
161 | (anything that writes memory, and subroutine calls, but not pushes). */ | |
162 | ||
163 | static int mem_last_set; | |
164 | ||
165 | /* Record the cuid of the last CALL_INSN | |
166 | so we can tell whether a potential combination crosses any calls. */ | |
167 | ||
168 | static int last_call_cuid; | |
169 | ||
170 | /* When `subst' is called, this is the insn that is being modified | |
171 | (by combining in a previous insn). The PATTERN of this insn | |
172 | is still the old pattern partially modified and it should not be | |
173 | looked at, but this may be used to examine the successors of the insn | |
174 | to judge whether a simplification is valid. */ | |
175 | ||
176 | static rtx subst_insn; | |
177 | ||
0d9641d1 JW |
178 | /* This is an insn that belongs before subst_insn, but is not currently |
179 | on the insn chain. */ | |
180 | ||
181 | static rtx subst_prev_insn; | |
182 | ||
230d793d RS |
183 | /* This is the lowest CUID that `subst' is currently dealing with. |
184 | get_last_value will not return a value if the register was set at or | |
185 | after this CUID. If not for this mechanism, we could get confused if | |
186 | I2 or I1 in try_combine were an insn that used the old value of a register | |
187 | to obtain a new value. In that case, we might erroneously get the | |
188 | new value of the register when we wanted the old one. */ | |
189 | ||
190 | static int subst_low_cuid; | |
191 | ||
6e25d159 RK |
192 | /* This contains any hard registers that are used in newpat; reg_dead_at_p |
193 | must consider all these registers to be always live. */ | |
194 | ||
195 | static HARD_REG_SET newpat_used_regs; | |
196 | ||
abe6e52f RK |
197 | /* This is an insn to which a LOG_LINKS entry has been added. If this |
198 | insn is the earlier than I2 or I3, combine should rescan starting at | |
199 | that location. */ | |
200 | ||
201 | static rtx added_links_insn; | |
202 | ||
0d4d42c3 RK |
203 | /* Basic block number of the block in which we are performing combines. */ |
204 | static int this_basic_block; | |
230d793d RS |
205 | \f |
206 | /* The next group of arrays allows the recording of the last value assigned | |
207 | to (hard or pseudo) register n. We use this information to see if a | |
5089e22e | 208 | operation being processed is redundant given a prior operation performed |
230d793d RS |
209 | on the register. For example, an `and' with a constant is redundant if |
210 | all the zero bits are already known to be turned off. | |
211 | ||
212 | We use an approach similar to that used by cse, but change it in the | |
213 | following ways: | |
214 | ||
215 | (1) We do not want to reinitialize at each label. | |
216 | (2) It is useful, but not critical, to know the actual value assigned | |
217 | to a register. Often just its form is helpful. | |
218 | ||
219 | Therefore, we maintain the following arrays: | |
220 | ||
221 | reg_last_set_value the last value assigned | |
222 | reg_last_set_label records the value of label_tick when the | |
223 | register was assigned | |
224 | reg_last_set_table_tick records the value of label_tick when a | |
225 | value using the register is assigned | |
226 | reg_last_set_invalid set to non-zero when it is not valid | |
227 | to use the value of this register in some | |
228 | register's value | |
229 | ||
230 | To understand the usage of these tables, it is important to understand | |
231 | the distinction between the value in reg_last_set_value being valid | |
232 | and the register being validly contained in some other expression in the | |
233 | table. | |
234 | ||
235 | Entry I in reg_last_set_value is valid if it is non-zero, and either | |
236 | reg_n_sets[i] is 1 or reg_last_set_label[i] == label_tick. | |
237 | ||
238 | Register I may validly appear in any expression returned for the value | |
239 | of another register if reg_n_sets[i] is 1. It may also appear in the | |
240 | value for register J if reg_last_set_label[i] < reg_last_set_label[j] or | |
241 | reg_last_set_invalid[j] is zero. | |
242 | ||
243 | If an expression is found in the table containing a register which may | |
244 | not validly appear in an expression, the register is replaced by | |
245 | something that won't match, (clobber (const_int 0)). | |
246 | ||
247 | reg_last_set_invalid[i] is set non-zero when register I is being assigned | |
248 | to and reg_last_set_table_tick[i] == label_tick. */ | |
249 | ||
0f41302f | 250 | /* Record last value assigned to (hard or pseudo) register n. */ |
230d793d RS |
251 | |
252 | static rtx *reg_last_set_value; | |
253 | ||
254 | /* Record the value of label_tick when the value for register n is placed in | |
255 | reg_last_set_value[n]. */ | |
256 | ||
568356af | 257 | static int *reg_last_set_label; |
230d793d RS |
258 | |
259 | /* Record the value of label_tick when an expression involving register n | |
0f41302f | 260 | is placed in reg_last_set_value. */ |
230d793d | 261 | |
568356af | 262 | static int *reg_last_set_table_tick; |
230d793d RS |
263 | |
264 | /* Set non-zero if references to register n in expressions should not be | |
265 | used. */ | |
266 | ||
267 | static char *reg_last_set_invalid; | |
268 | ||
0f41302f | 269 | /* Incremented for each label. */ |
230d793d | 270 | |
568356af | 271 | static int label_tick; |
230d793d RS |
272 | |
273 | /* Some registers that are set more than once and used in more than one | |
274 | basic block are nevertheless always set in similar ways. For example, | |
275 | a QImode register may be loaded from memory in two places on a machine | |
276 | where byte loads zero extend. | |
277 | ||
951553af | 278 | We record in the following array what we know about the nonzero |
230d793d RS |
279 | bits of a register, specifically which bits are known to be zero. |
280 | ||
281 | If an entry is zero, it means that we don't know anything special. */ | |
282 | ||
55310dad | 283 | static unsigned HOST_WIDE_INT *reg_nonzero_bits; |
230d793d | 284 | |
951553af | 285 | /* Mode used to compute significance in reg_nonzero_bits. It is the largest |
5f4f0e22 | 286 | integer mode that can fit in HOST_BITS_PER_WIDE_INT. */ |
230d793d | 287 | |
951553af | 288 | static enum machine_mode nonzero_bits_mode; |
230d793d | 289 | |
d0ab8cd3 RK |
290 | /* Nonzero if we know that a register has some leading bits that are always |
291 | equal to the sign bit. */ | |
292 | ||
293 | static char *reg_sign_bit_copies; | |
294 | ||
951553af | 295 | /* Nonzero when reg_nonzero_bits and reg_sign_bit_copies can be safely used. |
1a26b032 RK |
296 | It is zero while computing them and after combine has completed. This |
297 | former test prevents propagating values based on previously set values, | |
298 | which can be incorrect if a variable is modified in a loop. */ | |
230d793d | 299 | |
951553af | 300 | static int nonzero_sign_valid; |
55310dad RK |
301 | |
302 | /* These arrays are maintained in parallel with reg_last_set_value | |
303 | and are used to store the mode in which the register was last set, | |
304 | the bits that were known to be zero when it was last set, and the | |
305 | number of sign bits copies it was known to have when it was last set. */ | |
306 | ||
307 | static enum machine_mode *reg_last_set_mode; | |
308 | static unsigned HOST_WIDE_INT *reg_last_set_nonzero_bits; | |
309 | static char *reg_last_set_sign_bit_copies; | |
230d793d RS |
310 | \f |
311 | /* Record one modification to rtl structure | |
312 | to be undone by storing old_contents into *where. | |
313 | is_int is 1 if the contents are an int. */ | |
314 | ||
315 | struct undo | |
316 | { | |
241cea85 | 317 | struct undo *next; |
230d793d | 318 | int is_int; |
f5393ab9 RS |
319 | union {rtx r; int i;} old_contents; |
320 | union {rtx *r; int *i;} where; | |
230d793d RS |
321 | }; |
322 | ||
323 | /* Record a bunch of changes to be undone, up to MAX_UNDO of them. | |
324 | num_undo says how many are currently recorded. | |
325 | ||
326 | storage is nonzero if we must undo the allocation of new storage. | |
327 | The value of storage is what to pass to obfree. | |
328 | ||
329 | other_insn is nonzero if we have modified some other insn in the process | |
241cea85 | 330 | of working on subst_insn. It must be verified too. |
230d793d | 331 | |
241cea85 RK |
332 | previous_undos is the value of undobuf.undos when we started processing |
333 | this substitution. This will prevent gen_rtx_combine from re-used a piece | |
334 | from the previous expression. Doing so can produce circular rtl | |
335 | structures. */ | |
230d793d RS |
336 | |
337 | struct undobuf | |
338 | { | |
230d793d | 339 | char *storage; |
241cea85 RK |
340 | struct undo *undos; |
341 | struct undo *frees; | |
342 | struct undo *previous_undos; | |
230d793d RS |
343 | rtx other_insn; |
344 | }; | |
345 | ||
346 | static struct undobuf undobuf; | |
347 | ||
cc876596 | 348 | /* Substitute NEWVAL, an rtx expression, into INTO, a place in some |
230d793d | 349 | insn. The substitution can be undone by undo_all. If INTO is already |
cc876596 RK |
350 | set to NEWVAL, do not record this change. Because computing NEWVAL might |
351 | also call SUBST, we have to compute it before we put anything into | |
352 | the undo table. */ | |
230d793d RS |
353 | |
354 | #define SUBST(INTO, NEWVAL) \ | |
241cea85 RK |
355 | do { rtx _new = (NEWVAL); \ |
356 | struct undo *_buf; \ | |
357 | \ | |
358 | if (undobuf.frees) \ | |
359 | _buf = undobuf.frees, undobuf.frees = _buf->next; \ | |
360 | else \ | |
361 | _buf = (struct undo *) xmalloc (sizeof (struct undo)); \ | |
362 | \ | |
363 | _buf->is_int = 0; \ | |
364 | _buf->where.r = &INTO; \ | |
365 | _buf->old_contents.r = INTO; \ | |
366 | INTO = _new; \ | |
367 | if (_buf->old_contents.r == INTO) \ | |
368 | _buf->next = undobuf.frees, undobuf.frees = _buf; \ | |
369 | else \ | |
370 | _buf->next = undobuf.undos, undobuf.undos = _buf; \ | |
230d793d RS |
371 | } while (0) |
372 | ||
241cea85 RK |
373 | /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution |
374 | for the value of a HOST_WIDE_INT value (including CONST_INT) is | |
375 | not safe. */ | |
230d793d RS |
376 | |
377 | #define SUBST_INT(INTO, NEWVAL) \ | |
241cea85 RK |
378 | do { struct undo *_buf; \ |
379 | \ | |
380 | if (undobuf.frees) \ | |
381 | _buf = undobuf.frees, undobuf.frees = _buf->next; \ | |
382 | else \ | |
383 | _buf = (struct undo *) xmalloc (sizeof (struct undo)); \ | |
384 | \ | |
385 | _buf->is_int = 1; \ | |
386 | _buf->where.i = (int *) &INTO; \ | |
387 | _buf->old_contents.i = INTO; \ | |
388 | INTO = NEWVAL; \ | |
389 | if (_buf->old_contents.i == INTO) \ | |
390 | _buf->next = undobuf.frees, undobuf.frees = _buf; \ | |
391 | else \ | |
392 | _buf->next = undobuf.undos, undobuf.undos = _buf; \ | |
230d793d RS |
393 | } while (0) |
394 | ||
395 | /* Number of times the pseudo being substituted for | |
396 | was found and replaced. */ | |
397 | ||
398 | static int n_occurrences; | |
399 | ||
c5ad722c RK |
400 | static void init_reg_last_arrays PROTO((void)); |
401 | static void setup_incoming_promotions PROTO((void)); | |
fe2db4fb RK |
402 | static void set_nonzero_bits_and_sign_copies PROTO((rtx, rtx)); |
403 | static int can_combine_p PROTO((rtx, rtx, rtx, rtx, rtx *, rtx *)); | |
e009aaf3 | 404 | static int sets_function_arg_p PROTO((rtx)); |
fe2db4fb RK |
405 | static int combinable_i3pat PROTO((rtx, rtx *, rtx, rtx, int, rtx *)); |
406 | static rtx try_combine PROTO((rtx, rtx, rtx)); | |
407 | static void undo_all PROTO((void)); | |
408 | static rtx *find_split_point PROTO((rtx *, rtx)); | |
409 | static rtx subst PROTO((rtx, rtx, rtx, int, int)); | |
8079805d RK |
410 | static rtx simplify_rtx PROTO((rtx, enum machine_mode, int, int)); |
411 | static rtx simplify_if_then_else PROTO((rtx)); | |
412 | static rtx simplify_set PROTO((rtx)); | |
413 | static rtx simplify_logical PROTO((rtx, int)); | |
fe2db4fb RK |
414 | static rtx expand_compound_operation PROTO((rtx)); |
415 | static rtx expand_field_assignment PROTO((rtx)); | |
416 | static rtx make_extraction PROTO((enum machine_mode, rtx, int, rtx, int, | |
417 | int, int, int)); | |
71923da7 | 418 | static rtx extract_left_shift PROTO((rtx, int)); |
fe2db4fb RK |
419 | static rtx make_compound_operation PROTO((rtx, enum rtx_code)); |
420 | static int get_pos_from_mask PROTO((unsigned HOST_WIDE_INT, int *)); | |
6139ff20 | 421 | static rtx force_to_mode PROTO((rtx, enum machine_mode, |
e3d616e3 | 422 | unsigned HOST_WIDE_INT, rtx, int)); |
abe6e52f | 423 | static rtx if_then_else_cond PROTO((rtx, rtx *, rtx *)); |
fe2db4fb | 424 | static rtx known_cond PROTO((rtx, enum rtx_code, rtx, rtx)); |
e11fa86f | 425 | static int rtx_equal_for_field_assignment_p PROTO((rtx, rtx)); |
fe2db4fb RK |
426 | static rtx make_field_assignment PROTO((rtx)); |
427 | static rtx apply_distributive_law PROTO((rtx)); | |
428 | static rtx simplify_and_const_int PROTO((rtx, enum machine_mode, rtx, | |
429 | unsigned HOST_WIDE_INT)); | |
430 | static unsigned HOST_WIDE_INT nonzero_bits PROTO((rtx, enum machine_mode)); | |
431 | static int num_sign_bit_copies PROTO((rtx, enum machine_mode)); | |
432 | static int merge_outer_ops PROTO((enum rtx_code *, HOST_WIDE_INT *, | |
433 | enum rtx_code, HOST_WIDE_INT, | |
434 | enum machine_mode, int *)); | |
435 | static rtx simplify_shift_const PROTO((rtx, enum rtx_code, enum machine_mode, | |
436 | rtx, int)); | |
a29ca9db | 437 | static int recog_for_combine PROTO((rtx *, rtx, rtx *, int *)); |
fe2db4fb | 438 | static rtx gen_lowpart_for_combine PROTO((enum machine_mode, rtx)); |
d18225c4 | 439 | static rtx gen_rtx_combine PVPROTO((enum rtx_code code, enum machine_mode mode, |
4f90e4a0 | 440 | ...)); |
fe2db4fb RK |
441 | static rtx gen_binary PROTO((enum rtx_code, enum machine_mode, |
442 | rtx, rtx)); | |
0c1c8ea6 RK |
443 | static rtx gen_unary PROTO((enum rtx_code, enum machine_mode, |
444 | enum machine_mode, rtx)); | |
fe2db4fb RK |
445 | static enum rtx_code simplify_comparison PROTO((enum rtx_code, rtx *, rtx *)); |
446 | static int reversible_comparison_p PROTO((rtx)); | |
447 | static void update_table_tick PROTO((rtx)); | |
448 | static void record_value_for_reg PROTO((rtx, rtx, rtx)); | |
449 | static void record_dead_and_set_regs_1 PROTO((rtx, rtx)); | |
450 | static void record_dead_and_set_regs PROTO((rtx)); | |
9a893315 | 451 | static int get_last_value_validate PROTO((rtx *, rtx, int, int)); |
fe2db4fb RK |
452 | static rtx get_last_value PROTO((rtx)); |
453 | static int use_crosses_set_p PROTO((rtx, int)); | |
454 | static void reg_dead_at_p_1 PROTO((rtx, rtx)); | |
455 | static int reg_dead_at_p PROTO((rtx, rtx)); | |
6eb12cef | 456 | static void move_deaths PROTO((rtx, rtx, int, rtx, rtx *)); |
fe2db4fb RK |
457 | static int reg_bitfield_target_p PROTO((rtx, rtx)); |
458 | static void distribute_notes PROTO((rtx, rtx, rtx, rtx, rtx, rtx)); | |
459 | static void distribute_links PROTO((rtx)); | |
6e25d159 | 460 | static void mark_used_regs_combine PROTO((rtx)); |
1427d6d2 | 461 | static int insn_cuid PROTO((rtx)); |
230d793d RS |
462 | \f |
463 | /* Main entry point for combiner. F is the first insn of the function. | |
464 | NREGS is the first unused pseudo-reg number. */ | |
465 | ||
466 | void | |
467 | combine_instructions (f, nregs) | |
468 | rtx f; | |
469 | int nregs; | |
470 | { | |
b729186a JL |
471 | register rtx insn, next; |
472 | #ifdef HAVE_cc0 | |
473 | register rtx prev; | |
474 | #endif | |
230d793d RS |
475 | register int i; |
476 | register rtx links, nextlinks; | |
477 | ||
478 | combine_attempts = 0; | |
479 | combine_merges = 0; | |
480 | combine_extras = 0; | |
481 | combine_successes = 0; | |
241cea85 | 482 | undobuf.undos = undobuf.previous_undos = 0; |
230d793d RS |
483 | |
484 | combine_max_regno = nregs; | |
485 | ||
ef026f91 RS |
486 | reg_nonzero_bits |
487 | = (unsigned HOST_WIDE_INT *) alloca (nregs * sizeof (HOST_WIDE_INT)); | |
488 | reg_sign_bit_copies = (char *) alloca (nregs * sizeof (char)); | |
489 | ||
4c9a05bc | 490 | bzero ((char *) reg_nonzero_bits, nregs * sizeof (HOST_WIDE_INT)); |
ef026f91 RS |
491 | bzero (reg_sign_bit_copies, nregs * sizeof (char)); |
492 | ||
230d793d RS |
493 | reg_last_death = (rtx *) alloca (nregs * sizeof (rtx)); |
494 | reg_last_set = (rtx *) alloca (nregs * sizeof (rtx)); | |
495 | reg_last_set_value = (rtx *) alloca (nregs * sizeof (rtx)); | |
568356af RK |
496 | reg_last_set_table_tick = (int *) alloca (nregs * sizeof (int)); |
497 | reg_last_set_label = (int *) alloca (nregs * sizeof (int)); | |
5f4f0e22 | 498 | reg_last_set_invalid = (char *) alloca (nregs * sizeof (char)); |
55310dad RK |
499 | reg_last_set_mode |
500 | = (enum machine_mode *) alloca (nregs * sizeof (enum machine_mode)); | |
501 | reg_last_set_nonzero_bits | |
502 | = (unsigned HOST_WIDE_INT *) alloca (nregs * sizeof (HOST_WIDE_INT)); | |
503 | reg_last_set_sign_bit_copies | |
504 | = (char *) alloca (nregs * sizeof (char)); | |
505 | ||
ef026f91 | 506 | init_reg_last_arrays (); |
230d793d RS |
507 | |
508 | init_recog_no_volatile (); | |
509 | ||
510 | /* Compute maximum uid value so uid_cuid can be allocated. */ | |
511 | ||
512 | for (insn = f, i = 0; insn; insn = NEXT_INSN (insn)) | |
513 | if (INSN_UID (insn) > i) | |
514 | i = INSN_UID (insn); | |
515 | ||
516 | uid_cuid = (int *) alloca ((i + 1) * sizeof (int)); | |
4255220d | 517 | max_uid_cuid = i; |
230d793d | 518 | |
951553af | 519 | nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0); |
230d793d | 520 | |
951553af | 521 | /* Don't use reg_nonzero_bits when computing it. This can cause problems |
230d793d RS |
522 | when, for example, we have j <<= 1 in a loop. */ |
523 | ||
951553af | 524 | nonzero_sign_valid = 0; |
230d793d RS |
525 | |
526 | /* Compute the mapping from uids to cuids. | |
527 | Cuids are numbers assigned to insns, like uids, | |
528 | except that cuids increase monotonically through the code. | |
529 | ||
530 | Scan all SETs and see if we can deduce anything about what | |
951553af | 531 | bits are known to be zero for some registers and how many copies |
d79f08e0 RK |
532 | of the sign bit are known to exist for those registers. |
533 | ||
534 | Also set any known values so that we can use it while searching | |
535 | for what bits are known to be set. */ | |
536 | ||
537 | label_tick = 1; | |
230d793d | 538 | |
bcd49eb7 JW |
539 | /* We need to initialize it here, because record_dead_and_set_regs may call |
540 | get_last_value. */ | |
541 | subst_prev_insn = NULL_RTX; | |
542 | ||
7988fd36 RK |
543 | setup_incoming_promotions (); |
544 | ||
230d793d RS |
545 | for (insn = f, i = 0; insn; insn = NEXT_INSN (insn)) |
546 | { | |
4255220d | 547 | uid_cuid[INSN_UID (insn)] = ++i; |
d79f08e0 RK |
548 | subst_low_cuid = i; |
549 | subst_insn = insn; | |
550 | ||
230d793d | 551 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i') |
d79f08e0 RK |
552 | { |
553 | note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies); | |
554 | record_dead_and_set_regs (insn); | |
2dab894a RK |
555 | |
556 | #ifdef AUTO_INC_DEC | |
557 | for (links = REG_NOTES (insn); links; links = XEXP (links, 1)) | |
558 | if (REG_NOTE_KIND (links) == REG_INC) | |
559 | set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX); | |
560 | #endif | |
d79f08e0 RK |
561 | } |
562 | ||
563 | if (GET_CODE (insn) == CODE_LABEL) | |
564 | label_tick++; | |
230d793d RS |
565 | } |
566 | ||
951553af | 567 | nonzero_sign_valid = 1; |
230d793d RS |
568 | |
569 | /* Now scan all the insns in forward order. */ | |
570 | ||
0d4d42c3 | 571 | this_basic_block = -1; |
230d793d RS |
572 | label_tick = 1; |
573 | last_call_cuid = 0; | |
574 | mem_last_set = 0; | |
ef026f91 | 575 | init_reg_last_arrays (); |
7988fd36 RK |
576 | setup_incoming_promotions (); |
577 | ||
230d793d RS |
578 | for (insn = f; insn; insn = next ? next : NEXT_INSN (insn)) |
579 | { | |
580 | next = 0; | |
581 | ||
0d4d42c3 | 582 | /* If INSN starts a new basic block, update our basic block number. */ |
f085c9cd | 583 | if (this_basic_block + 1 < n_basic_blocks |
0d4d42c3 RK |
584 | && basic_block_head[this_basic_block + 1] == insn) |
585 | this_basic_block++; | |
586 | ||
230d793d RS |
587 | if (GET_CODE (insn) == CODE_LABEL) |
588 | label_tick++; | |
589 | ||
0d4d42c3 | 590 | else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i') |
230d793d RS |
591 | { |
592 | /* Try this insn with each insn it links back to. */ | |
593 | ||
594 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
5f4f0e22 | 595 | if ((next = try_combine (insn, XEXP (links, 0), NULL_RTX)) != 0) |
230d793d RS |
596 | goto retry; |
597 | ||
598 | /* Try each sequence of three linked insns ending with this one. */ | |
599 | ||
600 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
601 | for (nextlinks = LOG_LINKS (XEXP (links, 0)); 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 | #ifdef HAVE_cc0 | |
608 | /* Try to combine a jump insn that uses CC0 | |
609 | with a preceding insn that sets CC0, and maybe with its | |
610 | logical predecessor as well. | |
611 | This is how we make decrement-and-branch insns. | |
612 | We need this special code because data flow connections | |
613 | via CC0 do not get entered in LOG_LINKS. */ | |
614 | ||
615 | if (GET_CODE (insn) == JUMP_INSN | |
616 | && (prev = prev_nonnote_insn (insn)) != 0 | |
617 | && GET_CODE (prev) == INSN | |
618 | && sets_cc0_p (PATTERN (prev))) | |
619 | { | |
5f4f0e22 | 620 | if ((next = try_combine (insn, prev, NULL_RTX)) != 0) |
230d793d RS |
621 | goto retry; |
622 | ||
623 | for (nextlinks = LOG_LINKS (prev); nextlinks; | |
624 | nextlinks = XEXP (nextlinks, 1)) | |
625 | if ((next = try_combine (insn, prev, | |
626 | XEXP (nextlinks, 0))) != 0) | |
627 | goto retry; | |
628 | } | |
629 | ||
630 | /* Do the same for an insn that explicitly references CC0. */ | |
631 | if (GET_CODE (insn) == INSN | |
632 | && (prev = prev_nonnote_insn (insn)) != 0 | |
633 | && GET_CODE (prev) == INSN | |
634 | && sets_cc0_p (PATTERN (prev)) | |
635 | && GET_CODE (PATTERN (insn)) == SET | |
636 | && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn)))) | |
637 | { | |
5f4f0e22 | 638 | if ((next = try_combine (insn, prev, NULL_RTX)) != 0) |
230d793d RS |
639 | goto retry; |
640 | ||
641 | for (nextlinks = LOG_LINKS (prev); nextlinks; | |
642 | nextlinks = XEXP (nextlinks, 1)) | |
643 | if ((next = try_combine (insn, prev, | |
644 | XEXP (nextlinks, 0))) != 0) | |
645 | goto retry; | |
646 | } | |
647 | ||
648 | /* Finally, see if any of the insns that this insn links to | |
649 | explicitly references CC0. If so, try this insn, that insn, | |
5089e22e | 650 | and its predecessor if it sets CC0. */ |
230d793d RS |
651 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) |
652 | if (GET_CODE (XEXP (links, 0)) == INSN | |
653 | && GET_CODE (PATTERN (XEXP (links, 0))) == SET | |
654 | && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0)))) | |
655 | && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0 | |
656 | && GET_CODE (prev) == INSN | |
657 | && sets_cc0_p (PATTERN (prev)) | |
658 | && (next = try_combine (insn, XEXP (links, 0), prev)) != 0) | |
659 | goto retry; | |
660 | #endif | |
661 | ||
662 | /* Try combining an insn with two different insns whose results it | |
663 | uses. */ | |
664 | for (links = LOG_LINKS (insn); links; links = XEXP (links, 1)) | |
665 | for (nextlinks = XEXP (links, 1); nextlinks; | |
666 | nextlinks = XEXP (nextlinks, 1)) | |
667 | if ((next = try_combine (insn, XEXP (links, 0), | |
668 | XEXP (nextlinks, 0))) != 0) | |
669 | goto retry; | |
670 | ||
671 | if (GET_CODE (insn) != NOTE) | |
672 | record_dead_and_set_regs (insn); | |
673 | ||
674 | retry: | |
675 | ; | |
676 | } | |
677 | } | |
678 | ||
679 | total_attempts += combine_attempts; | |
680 | total_merges += combine_merges; | |
681 | total_extras += combine_extras; | |
682 | total_successes += combine_successes; | |
1a26b032 | 683 | |
951553af | 684 | nonzero_sign_valid = 0; |
230d793d | 685 | } |
ef026f91 RS |
686 | |
687 | /* Wipe the reg_last_xxx arrays in preparation for another pass. */ | |
688 | ||
689 | static void | |
690 | init_reg_last_arrays () | |
691 | { | |
692 | int nregs = combine_max_regno; | |
693 | ||
4c9a05bc RK |
694 | bzero ((char *) reg_last_death, nregs * sizeof (rtx)); |
695 | bzero ((char *) reg_last_set, nregs * sizeof (rtx)); | |
696 | bzero ((char *) reg_last_set_value, nregs * sizeof (rtx)); | |
697 | bzero ((char *) reg_last_set_table_tick, nregs * sizeof (int)); | |
698 | bzero ((char *) reg_last_set_label, nregs * sizeof (int)); | |
ef026f91 | 699 | bzero (reg_last_set_invalid, nregs * sizeof (char)); |
4c9a05bc RK |
700 | bzero ((char *) reg_last_set_mode, nregs * sizeof (enum machine_mode)); |
701 | bzero ((char *) reg_last_set_nonzero_bits, nregs * sizeof (HOST_WIDE_INT)); | |
ef026f91 RS |
702 | bzero (reg_last_set_sign_bit_copies, nregs * sizeof (char)); |
703 | } | |
230d793d | 704 | \f |
7988fd36 RK |
705 | /* Set up any promoted values for incoming argument registers. */ |
706 | ||
ee791cc3 | 707 | static void |
7988fd36 RK |
708 | setup_incoming_promotions () |
709 | { | |
710 | #ifdef PROMOTE_FUNCTION_ARGS | |
711 | int regno; | |
712 | rtx reg; | |
713 | enum machine_mode mode; | |
714 | int unsignedp; | |
715 | rtx first = get_insns (); | |
716 | ||
717 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
718 | if (FUNCTION_ARG_REGNO_P (regno) | |
719 | && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0) | |
38a448ca RH |
720 | { |
721 | record_value_for_reg | |
722 | (reg, first, gen_rtx_fmt_e ((unsignedp ? ZERO_EXTEND | |
723 | : SIGN_EXTEND), | |
724 | GET_MODE (reg), | |
725 | gen_rtx_CLOBBER (mode, const0_rtx))); | |
726 | } | |
7988fd36 RK |
727 | #endif |
728 | } | |
729 | \f | |
91102d5a RK |
730 | /* Called via note_stores. If X is a pseudo that is narrower than |
731 | HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero. | |
230d793d RS |
732 | |
733 | If we are setting only a portion of X and we can't figure out what | |
734 | portion, assume all bits will be used since we don't know what will | |
d0ab8cd3 RK |
735 | be happening. |
736 | ||
737 | Similarly, set how many bits of X are known to be copies of the sign bit | |
738 | at all locations in the function. This is the smallest number implied | |
739 | by any set of X. */ | |
230d793d RS |
740 | |
741 | static void | |
951553af | 742 | set_nonzero_bits_and_sign_copies (x, set) |
230d793d RS |
743 | rtx x; |
744 | rtx set; | |
745 | { | |
d0ab8cd3 RK |
746 | int num; |
747 | ||
230d793d RS |
748 | if (GET_CODE (x) == REG |
749 | && REGNO (x) >= FIRST_PSEUDO_REGISTER | |
e8095e80 RK |
750 | /* If this register is undefined at the start of the file, we can't |
751 | say what its contents were. */ | |
8e08106d | 752 | && ! REGNO_REG_SET_P (basic_block_live_at_start[0], REGNO (x)) |
5f4f0e22 | 753 | && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT) |
230d793d | 754 | { |
2dab894a | 755 | if (set == 0 || GET_CODE (set) == CLOBBER) |
e8095e80 RK |
756 | { |
757 | reg_nonzero_bits[REGNO (x)] = GET_MODE_MASK (GET_MODE (x)); | |
88306d12 | 758 | reg_sign_bit_copies[REGNO (x)] = 1; |
e8095e80 RK |
759 | return; |
760 | } | |
230d793d RS |
761 | |
762 | /* If this is a complex assignment, see if we can convert it into a | |
5089e22e | 763 | simple assignment. */ |
230d793d | 764 | set = expand_field_assignment (set); |
d79f08e0 RK |
765 | |
766 | /* If this is a simple assignment, or we have a paradoxical SUBREG, | |
767 | set what we know about X. */ | |
768 | ||
769 | if (SET_DEST (set) == x | |
770 | || (GET_CODE (SET_DEST (set)) == SUBREG | |
705c7b3b JW |
771 | && (GET_MODE_SIZE (GET_MODE (SET_DEST (set))) |
772 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set))))) | |
d79f08e0 | 773 | && SUBREG_REG (SET_DEST (set)) == x)) |
d0ab8cd3 | 774 | { |
9afa3d54 RK |
775 | rtx src = SET_SRC (set); |
776 | ||
777 | #ifdef SHORT_IMMEDIATES_SIGN_EXTEND | |
778 | /* If X is narrower than a word and SRC is a non-negative | |
779 | constant that would appear negative in the mode of X, | |
780 | sign-extend it for use in reg_nonzero_bits because some | |
781 | machines (maybe most) will actually do the sign-extension | |
782 | and this is the conservative approach. | |
783 | ||
784 | ??? For 2.5, try to tighten up the MD files in this regard | |
785 | instead of this kludge. */ | |
786 | ||
787 | if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD | |
788 | && GET_CODE (src) == CONST_INT | |
789 | && INTVAL (src) > 0 | |
790 | && 0 != (INTVAL (src) | |
791 | & ((HOST_WIDE_INT) 1 | |
9e69be8c | 792 | << (GET_MODE_BITSIZE (GET_MODE (x)) - 1)))) |
9afa3d54 RK |
793 | src = GEN_INT (INTVAL (src) |
794 | | ((HOST_WIDE_INT) (-1) | |
795 | << GET_MODE_BITSIZE (GET_MODE (x)))); | |
796 | #endif | |
797 | ||
951553af | 798 | reg_nonzero_bits[REGNO (x)] |
9afa3d54 | 799 | |= nonzero_bits (src, nonzero_bits_mode); |
d0ab8cd3 RK |
800 | num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x)); |
801 | if (reg_sign_bit_copies[REGNO (x)] == 0 | |
802 | || reg_sign_bit_copies[REGNO (x)] > num) | |
803 | reg_sign_bit_copies[REGNO (x)] = num; | |
804 | } | |
230d793d | 805 | else |
d0ab8cd3 | 806 | { |
951553af | 807 | reg_nonzero_bits[REGNO (x)] = GET_MODE_MASK (GET_MODE (x)); |
88306d12 | 808 | reg_sign_bit_copies[REGNO (x)] = 1; |
d0ab8cd3 | 809 | } |
230d793d RS |
810 | } |
811 | } | |
812 | \f | |
813 | /* See if INSN can be combined into I3. PRED and SUCC are optionally | |
814 | insns that were previously combined into I3 or that will be combined | |
815 | into the merger of INSN and I3. | |
816 | ||
817 | Return 0 if the combination is not allowed for any reason. | |
818 | ||
819 | If the combination is allowed, *PDEST will be set to the single | |
820 | destination of INSN and *PSRC to the single source, and this function | |
821 | will return 1. */ | |
822 | ||
823 | static int | |
824 | can_combine_p (insn, i3, pred, succ, pdest, psrc) | |
825 | rtx insn; | |
826 | rtx i3; | |
827 | rtx pred, succ; | |
828 | rtx *pdest, *psrc; | |
829 | { | |
830 | int i; | |
831 | rtx set = 0, src, dest; | |
b729186a JL |
832 | rtx p; |
833 | #ifdef AUTO_INC_DEC | |
76d31c63 | 834 | rtx link; |
b729186a | 835 | #endif |
230d793d RS |
836 | int all_adjacent = (succ ? (next_active_insn (insn) == succ |
837 | && next_active_insn (succ) == i3) | |
838 | : next_active_insn (insn) == i3); | |
839 | ||
840 | /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0. | |
841 | or a PARALLEL consisting of such a SET and CLOBBERs. | |
842 | ||
843 | If INSN has CLOBBER parallel parts, ignore them for our processing. | |
844 | By definition, these happen during the execution of the insn. When it | |
845 | is merged with another insn, all bets are off. If they are, in fact, | |
846 | needed and aren't also supplied in I3, they may be added by | |
847 | recog_for_combine. Otherwise, it won't match. | |
848 | ||
849 | We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED | |
850 | note. | |
851 | ||
852 | Get the source and destination of INSN. If more than one, can't | |
853 | combine. */ | |
854 | ||
855 | if (GET_CODE (PATTERN (insn)) == SET) | |
856 | set = PATTERN (insn); | |
857 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
858 | && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET) | |
859 | { | |
860 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
861 | { | |
862 | rtx elt = XVECEXP (PATTERN (insn), 0, i); | |
863 | ||
864 | switch (GET_CODE (elt)) | |
865 | { | |
e3258cef R |
866 | /* This is important to combine floating point insns |
867 | for the SH4 port. */ | |
868 | case USE: | |
869 | /* Combining an isolated USE doesn't make sense. | |
870 | We depend here on combinable_i3_pat to reject them. */ | |
871 | /* The code below this loop only verifies that the inputs of | |
872 | the SET in INSN do not change. We call reg_set_between_p | |
873 | to verify that the REG in the USE does not change betweeen | |
874 | I3 and INSN. | |
875 | If the USE in INSN was for a pseudo register, the matching | |
876 | insn pattern will likely match any register; combining this | |
877 | with any other USE would only be safe if we knew that the | |
878 | used registers have identical values, or if there was | |
879 | something to tell them apart, e.g. different modes. For | |
880 | now, we forgo such compilcated tests and simply disallow | |
881 | combining of USES of pseudo registers with any other USE. */ | |
882 | if (GET_CODE (XEXP (elt, 0)) == REG | |
883 | && GET_CODE (PATTERN (i3)) == PARALLEL) | |
884 | { | |
885 | rtx i3pat = PATTERN (i3); | |
886 | int i = XVECLEN (i3pat, 0) - 1; | |
887 | int regno = REGNO (XEXP (elt, 0)); | |
888 | do | |
889 | { | |
890 | rtx i3elt = XVECEXP (i3pat, 0, i); | |
891 | if (GET_CODE (i3elt) == USE | |
892 | && GET_CODE (XEXP (i3elt, 0)) == REG | |
893 | && (REGNO (XEXP (i3elt, 0)) == regno | |
894 | ? reg_set_between_p (XEXP (elt, 0), | |
895 | PREV_INSN (insn), i3) | |
896 | : regno >= FIRST_PSEUDO_REGISTER)) | |
897 | return 0; | |
898 | } | |
899 | while (--i >= 0); | |
900 | } | |
901 | break; | |
902 | ||
230d793d RS |
903 | /* We can ignore CLOBBERs. */ |
904 | case CLOBBER: | |
905 | break; | |
906 | ||
907 | case SET: | |
908 | /* Ignore SETs whose result isn't used but not those that | |
909 | have side-effects. */ | |
910 | if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt)) | |
911 | && ! side_effects_p (elt)) | |
912 | break; | |
913 | ||
914 | /* If we have already found a SET, this is a second one and | |
915 | so we cannot combine with this insn. */ | |
916 | if (set) | |
917 | return 0; | |
918 | ||
919 | set = elt; | |
920 | break; | |
921 | ||
922 | default: | |
923 | /* Anything else means we can't combine. */ | |
924 | return 0; | |
925 | } | |
926 | } | |
927 | ||
928 | if (set == 0 | |
929 | /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs, | |
930 | so don't do anything with it. */ | |
931 | || GET_CODE (SET_SRC (set)) == ASM_OPERANDS) | |
932 | return 0; | |
933 | } | |
934 | else | |
935 | return 0; | |
936 | ||
937 | if (set == 0) | |
938 | return 0; | |
939 | ||
940 | set = expand_field_assignment (set); | |
941 | src = SET_SRC (set), dest = SET_DEST (set); | |
942 | ||
943 | /* Don't eliminate a store in the stack pointer. */ | |
944 | if (dest == stack_pointer_rtx | |
230d793d RS |
945 | /* If we couldn't eliminate a field assignment, we can't combine. */ |
946 | || GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == STRICT_LOW_PART | |
947 | /* Don't combine with an insn that sets a register to itself if it has | |
948 | a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */ | |
5f4f0e22 | 949 | || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX)) |
230d793d RS |
950 | /* Can't merge a function call. */ |
951 | || GET_CODE (src) == CALL | |
cd5e8f1f | 952 | /* Don't eliminate a function call argument. */ |
4dca5ec5 RK |
953 | || (GET_CODE (i3) == CALL_INSN |
954 | && (find_reg_fusage (i3, USE, dest) | |
955 | || (GET_CODE (dest) == REG | |
956 | && REGNO (dest) < FIRST_PSEUDO_REGISTER | |
957 | && global_regs[REGNO (dest)]))) | |
230d793d RS |
958 | /* Don't substitute into an incremented register. */ |
959 | || FIND_REG_INC_NOTE (i3, dest) | |
960 | || (succ && FIND_REG_INC_NOTE (succ, dest)) | |
961 | /* Don't combine the end of a libcall into anything. */ | |
5f4f0e22 | 962 | || find_reg_note (insn, REG_RETVAL, NULL_RTX) |
230d793d RS |
963 | /* Make sure that DEST is not used after SUCC but before I3. */ |
964 | || (succ && ! all_adjacent | |
965 | && reg_used_between_p (dest, succ, i3)) | |
966 | /* Make sure that the value that is to be substituted for the register | |
967 | does not use any registers whose values alter in between. However, | |
968 | If the insns are adjacent, a use can't cross a set even though we | |
969 | think it might (this can happen for a sequence of insns each setting | |
970 | the same destination; reg_last_set of that register might point to | |
d81481d3 RK |
971 | a NOTE). If INSN has a REG_EQUIV note, the register is always |
972 | equivalent to the memory so the substitution is valid even if there | |
973 | are intervening stores. Also, don't move a volatile asm or | |
974 | UNSPEC_VOLATILE across any other insns. */ | |
230d793d | 975 | || (! all_adjacent |
d81481d3 RK |
976 | && (((GET_CODE (src) != MEM |
977 | || ! find_reg_note (insn, REG_EQUIV, src)) | |
978 | && use_crosses_set_p (src, INSN_CUID (insn))) | |
a66a10c7 RS |
979 | || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src)) |
980 | || GET_CODE (src) == UNSPEC_VOLATILE)) | |
230d793d RS |
981 | /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get |
982 | better register allocation by not doing the combine. */ | |
983 | || find_reg_note (i3, REG_NO_CONFLICT, dest) | |
984 | || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest)) | |
985 | /* Don't combine across a CALL_INSN, because that would possibly | |
986 | change whether the life span of some REGs crosses calls or not, | |
987 | and it is a pain to update that information. | |
988 | Exception: if source is a constant, moving it later can't hurt. | |
989 | Accept that special case, because it helps -fforce-addr a lot. */ | |
990 | || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src))) | |
991 | return 0; | |
992 | ||
993 | /* DEST must either be a REG or CC0. */ | |
994 | if (GET_CODE (dest) == REG) | |
995 | { | |
996 | /* If register alignment is being enforced for multi-word items in all | |
997 | cases except for parameters, it is possible to have a register copy | |
998 | insn referencing a hard register that is not allowed to contain the | |
999 | mode being copied and which would not be valid as an operand of most | |
1000 | insns. Eliminate this problem by not combining with such an insn. | |
1001 | ||
1002 | Also, on some machines we don't want to extend the life of a hard | |
4d2c432d RK |
1003 | register. |
1004 | ||
1005 | This is the same test done in can_combine except that we don't test | |
1006 | if SRC is a CALL operation to permit a hard register with | |
1007 | SMALL_REGISTER_CLASSES, and that we have to take all_adjacent | |
1008 | into account. */ | |
230d793d RS |
1009 | |
1010 | if (GET_CODE (src) == REG | |
1011 | && ((REGNO (dest) < FIRST_PSEUDO_REGISTER | |
1012 | && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest))) | |
c448a43e RK |
1013 | /* Don't extend the life of a hard register unless it is |
1014 | user variable (if we have few registers) or it can't | |
1015 | fit into the desired register (meaning something special | |
ecd40809 RK |
1016 | is going on). |
1017 | Also avoid substituting a return register into I3, because | |
1018 | reload can't handle a conflict with constraints of other | |
1019 | inputs. */ | |
230d793d | 1020 | || (REGNO (src) < FIRST_PSEUDO_REGISTER |
c448a43e | 1021 | && (! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)) |
f95182a4 ILT |
1022 | || (SMALL_REGISTER_CLASSES |
1023 | && ((! all_adjacent && ! REG_USERVAR_P (src)) | |
1024 | || (FUNCTION_VALUE_REGNO_P (REGNO (src)) | |
e9a25f70 | 1025 | && ! REG_USERVAR_P (src)))))))) |
230d793d RS |
1026 | return 0; |
1027 | } | |
1028 | else if (GET_CODE (dest) != CC0) | |
1029 | return 0; | |
1030 | ||
5f96750d RS |
1031 | /* Don't substitute for a register intended as a clobberable operand. |
1032 | Similarly, don't substitute an expression containing a register that | |
1033 | will be clobbered in I3. */ | |
230d793d RS |
1034 | if (GET_CODE (PATTERN (i3)) == PARALLEL) |
1035 | for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--) | |
1036 | if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER | |
5f96750d RS |
1037 | && (reg_overlap_mentioned_p (XEXP (XVECEXP (PATTERN (i3), 0, i), 0), |
1038 | src) | |
1039 | || rtx_equal_p (XEXP (XVECEXP (PATTERN (i3), 0, i), 0), dest))) | |
230d793d RS |
1040 | return 0; |
1041 | ||
1042 | /* If INSN contains anything volatile, or is an `asm' (whether volatile | |
d276f2bb | 1043 | or not), reject, unless nothing volatile comes between it and I3 */ |
230d793d RS |
1044 | |
1045 | if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src)) | |
d276f2bb CM |
1046 | { |
1047 | /* Make sure succ doesn't contain a volatile reference. */ | |
1048 | if (succ != 0 && volatile_refs_p (PATTERN (succ))) | |
1049 | return 0; | |
1050 | ||
1051 | for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p)) | |
1052 | if (GET_RTX_CLASS (GET_CODE (p)) == 'i' | |
1053 | && p != succ && volatile_refs_p (PATTERN (p))) | |
1054 | return 0; | |
1055 | } | |
230d793d | 1056 | |
b79ee7eb RH |
1057 | /* If INSN is an asm, and DEST is a hard register, reject, since it has |
1058 | to be an explicit register variable, and was chosen for a reason. */ | |
1059 | ||
1060 | if (GET_CODE (src) == ASM_OPERANDS | |
1061 | && GET_CODE (dest) == REG && REGNO (dest) < FIRST_PSEUDO_REGISTER) | |
1062 | return 0; | |
1063 | ||
4b2cb4a2 RS |
1064 | /* If there are any volatile insns between INSN and I3, reject, because |
1065 | they might affect machine state. */ | |
1066 | ||
1067 | for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p)) | |
1068 | if (GET_RTX_CLASS (GET_CODE (p)) == 'i' | |
1069 | && p != succ && volatile_insn_p (PATTERN (p))) | |
1070 | return 0; | |
1071 | ||
230d793d RS |
1072 | /* If INSN or I2 contains an autoincrement or autodecrement, |
1073 | make sure that register is not used between there and I3, | |
1074 | and not already used in I3 either. | |
1075 | Also insist that I3 not be a jump; if it were one | |
1076 | and the incremented register were spilled, we would lose. */ | |
1077 | ||
1078 | #ifdef AUTO_INC_DEC | |
1079 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
1080 | if (REG_NOTE_KIND (link) == REG_INC | |
1081 | && (GET_CODE (i3) == JUMP_INSN | |
1082 | || reg_used_between_p (XEXP (link, 0), insn, i3) | |
1083 | || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3)))) | |
1084 | return 0; | |
1085 | #endif | |
1086 | ||
1087 | #ifdef HAVE_cc0 | |
1088 | /* Don't combine an insn that follows a CC0-setting insn. | |
1089 | An insn that uses CC0 must not be separated from the one that sets it. | |
1090 | We do, however, allow I2 to follow a CC0-setting insn if that insn | |
1091 | is passed as I1; in that case it will be deleted also. | |
1092 | We also allow combining in this case if all the insns are adjacent | |
1093 | because that would leave the two CC0 insns adjacent as well. | |
1094 | It would be more logical to test whether CC0 occurs inside I1 or I2, | |
1095 | but that would be much slower, and this ought to be equivalent. */ | |
1096 | ||
1097 | p = prev_nonnote_insn (insn); | |
1098 | if (p && p != pred && GET_CODE (p) == INSN && sets_cc0_p (PATTERN (p)) | |
1099 | && ! all_adjacent) | |
1100 | return 0; | |
1101 | #endif | |
1102 | ||
1103 | /* If we get here, we have passed all the tests and the combination is | |
1104 | to be allowed. */ | |
1105 | ||
1106 | *pdest = dest; | |
1107 | *psrc = src; | |
1108 | ||
1109 | return 1; | |
1110 | } | |
1111 | \f | |
956d6950 JL |
1112 | /* Check if PAT is an insn - or a part of it - used to set up an |
1113 | argument for a function in a hard register. */ | |
1114 | ||
1115 | static int | |
1116 | sets_function_arg_p (pat) | |
1117 | rtx pat; | |
1118 | { | |
1119 | int i; | |
1120 | rtx inner_dest; | |
1121 | ||
1122 | switch (GET_CODE (pat)) | |
1123 | { | |
1124 | case INSN: | |
1125 | return sets_function_arg_p (PATTERN (pat)); | |
1126 | ||
1127 | case PARALLEL: | |
1128 | for (i = XVECLEN (pat, 0); --i >= 0;) | |
1129 | if (sets_function_arg_p (XVECEXP (pat, 0, i))) | |
1130 | return 1; | |
1131 | ||
1132 | break; | |
1133 | ||
1134 | case SET: | |
1135 | inner_dest = SET_DEST (pat); | |
1136 | while (GET_CODE (inner_dest) == STRICT_LOW_PART | |
1137 | || GET_CODE (inner_dest) == SUBREG | |
1138 | || GET_CODE (inner_dest) == ZERO_EXTRACT) | |
1139 | inner_dest = XEXP (inner_dest, 0); | |
1140 | ||
1141 | return (GET_CODE (inner_dest) == REG | |
1142 | && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER | |
1143 | && FUNCTION_ARG_REGNO_P (REGNO (inner_dest))); | |
1d300e19 KG |
1144 | |
1145 | default: | |
1146 | break; | |
956d6950 JL |
1147 | } |
1148 | ||
1149 | return 0; | |
1150 | } | |
1151 | ||
230d793d RS |
1152 | /* LOC is the location within I3 that contains its pattern or the component |
1153 | of a PARALLEL of the pattern. We validate that it is valid for combining. | |
1154 | ||
1155 | One problem is if I3 modifies its output, as opposed to replacing it | |
1156 | entirely, we can't allow the output to contain I2DEST or I1DEST as doing | |
1157 | so would produce an insn that is not equivalent to the original insns. | |
1158 | ||
1159 | Consider: | |
1160 | ||
1161 | (set (reg:DI 101) (reg:DI 100)) | |
1162 | (set (subreg:SI (reg:DI 101) 0) <foo>) | |
1163 | ||
1164 | This is NOT equivalent to: | |
1165 | ||
1166 | (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>) | |
1167 | (set (reg:DI 101) (reg:DI 100))]) | |
1168 | ||
1169 | Not only does this modify 100 (in which case it might still be valid | |
1170 | if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100. | |
1171 | ||
1172 | We can also run into a problem if I2 sets a register that I1 | |
1173 | uses and I1 gets directly substituted into I3 (not via I2). In that | |
1174 | case, we would be getting the wrong value of I2DEST into I3, so we | |
1175 | must reject the combination. This case occurs when I2 and I1 both | |
1176 | feed into I3, rather than when I1 feeds into I2, which feeds into I3. | |
1177 | If I1_NOT_IN_SRC is non-zero, it means that finding I1 in the source | |
1178 | of a SET must prevent combination from occurring. | |
1179 | ||
e9a25f70 | 1180 | On machines where SMALL_REGISTER_CLASSES is non-zero, we don't combine |
c448a43e RK |
1181 | if the destination of a SET is a hard register that isn't a user |
1182 | variable. | |
230d793d RS |
1183 | |
1184 | Before doing the above check, we first try to expand a field assignment | |
1185 | into a set of logical operations. | |
1186 | ||
1187 | If PI3_DEST_KILLED is non-zero, it is a pointer to a location in which | |
1188 | we place a register that is both set and used within I3. If more than one | |
1189 | such register is detected, we fail. | |
1190 | ||
1191 | Return 1 if the combination is valid, zero otherwise. */ | |
1192 | ||
1193 | static int | |
1194 | combinable_i3pat (i3, loc, i2dest, i1dest, i1_not_in_src, pi3dest_killed) | |
1195 | rtx i3; | |
1196 | rtx *loc; | |
1197 | rtx i2dest; | |
1198 | rtx i1dest; | |
1199 | int i1_not_in_src; | |
1200 | rtx *pi3dest_killed; | |
1201 | { | |
1202 | rtx x = *loc; | |
1203 | ||
1204 | if (GET_CODE (x) == SET) | |
1205 | { | |
1206 | rtx set = expand_field_assignment (x); | |
1207 | rtx dest = SET_DEST (set); | |
1208 | rtx src = SET_SRC (set); | |
29a82058 JL |
1209 | rtx inner_dest = dest; |
1210 | ||
1211 | #if 0 | |
1212 | rtx inner_src = src; | |
1213 | #endif | |
230d793d RS |
1214 | |
1215 | SUBST (*loc, set); | |
1216 | ||
1217 | while (GET_CODE (inner_dest) == STRICT_LOW_PART | |
1218 | || GET_CODE (inner_dest) == SUBREG | |
1219 | || GET_CODE (inner_dest) == ZERO_EXTRACT) | |
1220 | inner_dest = XEXP (inner_dest, 0); | |
1221 | ||
1222 | /* We probably don't need this any more now that LIMIT_RELOAD_CLASS | |
1223 | was added. */ | |
1224 | #if 0 | |
1225 | while (GET_CODE (inner_src) == STRICT_LOW_PART | |
1226 | || GET_CODE (inner_src) == SUBREG | |
1227 | || GET_CODE (inner_src) == ZERO_EXTRACT) | |
1228 | inner_src = XEXP (inner_src, 0); | |
1229 | ||
1230 | /* If it is better that two different modes keep two different pseudos, | |
1231 | avoid combining them. This avoids producing the following pattern | |
1232 | on a 386: | |
1233 | (set (subreg:SI (reg/v:QI 21) 0) | |
1234 | (lshiftrt:SI (reg/v:SI 20) | |
1235 | (const_int 24))) | |
1236 | If that were made, reload could not handle the pair of | |
1237 | reg 20/21, since it would try to get any GENERAL_REGS | |
1238 | but some of them don't handle QImode. */ | |
1239 | ||
1240 | if (rtx_equal_p (inner_src, i2dest) | |
1241 | && GET_CODE (inner_dest) == REG | |
1242 | && ! MODES_TIEABLE_P (GET_MODE (i2dest), GET_MODE (inner_dest))) | |
1243 | return 0; | |
1244 | #endif | |
1245 | ||
1246 | /* Check for the case where I3 modifies its output, as | |
1247 | discussed above. */ | |
1248 | if ((inner_dest != dest | |
1249 | && (reg_overlap_mentioned_p (i2dest, inner_dest) | |
1250 | || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest)))) | |
956d6950 | 1251 | |
3f508eca RK |
1252 | /* This is the same test done in can_combine_p except that we |
1253 | allow a hard register with SMALL_REGISTER_CLASSES if SRC is a | |
956d6950 JL |
1254 | CALL operation. Moreover, we can't test all_adjacent; we don't |
1255 | have to, since this instruction will stay in place, thus we are | |
1256 | not considering increasing the lifetime of INNER_DEST. | |
1257 | ||
1258 | Also, if this insn sets a function argument, combining it with | |
1259 | something that might need a spill could clobber a previous | |
1260 | function argument; the all_adjacent test in can_combine_p also | |
1261 | checks this; here, we do a more specific test for this case. */ | |
1262 | ||
230d793d | 1263 | || (GET_CODE (inner_dest) == REG |
dfbe1b2f | 1264 | && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER |
c448a43e RK |
1265 | && (! HARD_REGNO_MODE_OK (REGNO (inner_dest), |
1266 | GET_MODE (inner_dest)) | |
e9a25f70 JL |
1267 | || (SMALL_REGISTER_CLASSES && GET_CODE (src) != CALL |
1268 | && ! REG_USERVAR_P (inner_dest) | |
956d6950 JL |
1269 | && (FUNCTION_VALUE_REGNO_P (REGNO (inner_dest)) |
1270 | || (FUNCTION_ARG_REGNO_P (REGNO (inner_dest)) | |
1271 | && i3 != 0 | |
1272 | && sets_function_arg_p (prev_nonnote_insn (i3))))))) | |
230d793d RS |
1273 | || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src))) |
1274 | return 0; | |
1275 | ||
1276 | /* If DEST is used in I3, it is being killed in this insn, | |
36a9c2e9 JL |
1277 | so record that for later. |
1278 | Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the | |
1279 | STACK_POINTER_REGNUM, since these are always considered to be | |
1280 | live. Similarly for ARG_POINTER_REGNUM if it is fixed. */ | |
230d793d | 1281 | if (pi3dest_killed && GET_CODE (dest) == REG |
36a9c2e9 JL |
1282 | && reg_referenced_p (dest, PATTERN (i3)) |
1283 | && REGNO (dest) != FRAME_POINTER_REGNUM | |
6d7096b0 DE |
1284 | #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM |
1285 | && REGNO (dest) != HARD_FRAME_POINTER_REGNUM | |
1286 | #endif | |
36a9c2e9 JL |
1287 | #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM |
1288 | && (REGNO (dest) != ARG_POINTER_REGNUM | |
1289 | || ! fixed_regs [REGNO (dest)]) | |
1290 | #endif | |
1291 | && REGNO (dest) != STACK_POINTER_REGNUM) | |
230d793d RS |
1292 | { |
1293 | if (*pi3dest_killed) | |
1294 | return 0; | |
1295 | ||
1296 | *pi3dest_killed = dest; | |
1297 | } | |
1298 | } | |
1299 | ||
1300 | else if (GET_CODE (x) == PARALLEL) | |
1301 | { | |
1302 | int i; | |
1303 | ||
1304 | for (i = 0; i < XVECLEN (x, 0); i++) | |
1305 | if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest, | |
1306 | i1_not_in_src, pi3dest_killed)) | |
1307 | return 0; | |
1308 | } | |
1309 | ||
1310 | return 1; | |
1311 | } | |
1312 | \f | |
1313 | /* Try to combine the insns I1 and I2 into I3. | |
1314 | Here I1 and I2 appear earlier than I3. | |
1315 | I1 can be zero; then we combine just I2 into I3. | |
1316 | ||
1317 | It we are combining three insns and the resulting insn is not recognized, | |
1318 | try splitting it into two insns. If that happens, I2 and I3 are retained | |
1319 | and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2 | |
1320 | are pseudo-deleted. | |
1321 | ||
abe6e52f RK |
1322 | Return 0 if the combination does not work. Then nothing is changed. |
1323 | If we did the combination, return the insn at which combine should | |
1324 | resume scanning. */ | |
230d793d RS |
1325 | |
1326 | static rtx | |
1327 | try_combine (i3, i2, i1) | |
1328 | register rtx i3, i2, i1; | |
1329 | { | |
1330 | /* New patterns for I3 and I3, respectively. */ | |
1331 | rtx newpat, newi2pat = 0; | |
1332 | /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */ | |
1333 | int added_sets_1, added_sets_2; | |
1334 | /* Total number of SETs to put into I3. */ | |
1335 | int total_sets; | |
1336 | /* Nonzero is I2's body now appears in I3. */ | |
1337 | int i2_is_used; | |
1338 | /* INSN_CODEs for new I3, new I2, and user of condition code. */ | |
1339 | int insn_code_number, i2_code_number, other_code_number; | |
1340 | /* Contains I3 if the destination of I3 is used in its source, which means | |
1341 | that the old life of I3 is being killed. If that usage is placed into | |
1342 | I2 and not in I3, a REG_DEAD note must be made. */ | |
1343 | rtx i3dest_killed = 0; | |
1344 | /* SET_DEST and SET_SRC of I2 and I1. */ | |
1345 | rtx i2dest, i2src, i1dest = 0, i1src = 0; | |
1346 | /* PATTERN (I2), or a copy of it in certain cases. */ | |
1347 | rtx i2pat; | |
1348 | /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */ | |
c4e861e8 | 1349 | int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0; |
230d793d RS |
1350 | int i1_feeds_i3 = 0; |
1351 | /* Notes that must be added to REG_NOTES in I3 and I2. */ | |
1352 | rtx new_i3_notes, new_i2_notes; | |
176c9e6b JW |
1353 | /* Notes that we substituted I3 into I2 instead of the normal case. */ |
1354 | int i3_subst_into_i2 = 0; | |
df7d75de RK |
1355 | /* Notes that I1, I2 or I3 is a MULT operation. */ |
1356 | int have_mult = 0; | |
a29ca9db RK |
1357 | /* Number of clobbers of SCRATCH we had to add. */ |
1358 | int i3_scratches = 0, i2_scratches = 0, other_scratches = 0; | |
230d793d RS |
1359 | |
1360 | int maxreg; | |
1361 | rtx temp; | |
1362 | register rtx link; | |
1363 | int i; | |
1364 | ||
1365 | /* If any of I1, I2, and I3 isn't really an insn, we can't do anything. | |
1366 | This can occur when flow deletes an insn that it has merged into an | |
1367 | auto-increment address. We also can't do anything if I3 has a | |
1368 | REG_LIBCALL note since we don't want to disrupt the contiguity of a | |
1369 | libcall. */ | |
1370 | ||
1371 | if (GET_RTX_CLASS (GET_CODE (i3)) != 'i' | |
1372 | || GET_RTX_CLASS (GET_CODE (i2)) != 'i' | |
1373 | || (i1 && GET_RTX_CLASS (GET_CODE (i1)) != 'i') | |
5f4f0e22 | 1374 | || find_reg_note (i3, REG_LIBCALL, NULL_RTX)) |
230d793d RS |
1375 | return 0; |
1376 | ||
1377 | combine_attempts++; | |
1378 | ||
241cea85 | 1379 | undobuf.undos = undobuf.previous_undos = 0; |
230d793d RS |
1380 | undobuf.other_insn = 0; |
1381 | ||
1382 | /* Save the current high-water-mark so we can free storage if we didn't | |
1383 | accept this combination. */ | |
1384 | undobuf.storage = (char *) oballoc (0); | |
1385 | ||
6e25d159 RK |
1386 | /* Reset the hard register usage information. */ |
1387 | CLEAR_HARD_REG_SET (newpat_used_regs); | |
1388 | ||
230d793d RS |
1389 | /* If I1 and I2 both feed I3, they can be in any order. To simplify the |
1390 | code below, set I1 to be the earlier of the two insns. */ | |
1391 | if (i1 && INSN_CUID (i1) > INSN_CUID (i2)) | |
1392 | temp = i1, i1 = i2, i2 = temp; | |
1393 | ||
abe6e52f | 1394 | added_links_insn = 0; |
137e889e | 1395 | |
230d793d RS |
1396 | /* First check for one important special-case that the code below will |
1397 | not handle. Namely, the case where I1 is zero, I2 has multiple sets, | |
1398 | and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case, | |
1399 | we may be able to replace that destination with the destination of I3. | |
1400 | This occurs in the common code where we compute both a quotient and | |
1401 | remainder into a structure, in which case we want to do the computation | |
1402 | directly into the structure to avoid register-register copies. | |
1403 | ||
1404 | We make very conservative checks below and only try to handle the | |
1405 | most common cases of this. For example, we only handle the case | |
1406 | where I2 and I3 are adjacent to avoid making difficult register | |
1407 | usage tests. */ | |
1408 | ||
1409 | if (i1 == 0 && GET_CODE (i3) == INSN && GET_CODE (PATTERN (i3)) == SET | |
1410 | && GET_CODE (SET_SRC (PATTERN (i3))) == REG | |
1411 | && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER | |
f95182a4 | 1412 | && (! SMALL_REGISTER_CLASSES |
e9a25f70 JL |
1413 | || (GET_CODE (SET_DEST (PATTERN (i3))) != REG |
1414 | || REGNO (SET_DEST (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER | |
1415 | || REG_USERVAR_P (SET_DEST (PATTERN (i3))))) | |
230d793d RS |
1416 | && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3))) |
1417 | && GET_CODE (PATTERN (i2)) == PARALLEL | |
1418 | && ! side_effects_p (SET_DEST (PATTERN (i3))) | |
5089e22e RS |
1419 | /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code |
1420 | below would need to check what is inside (and reg_overlap_mentioned_p | |
1421 | doesn't support those codes anyway). Don't allow those destinations; | |
1422 | the resulting insn isn't likely to be recognized anyway. */ | |
1423 | && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT | |
1424 | && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART | |
230d793d RS |
1425 | && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)), |
1426 | SET_DEST (PATTERN (i3))) | |
1427 | && next_real_insn (i2) == i3) | |
5089e22e RS |
1428 | { |
1429 | rtx p2 = PATTERN (i2); | |
1430 | ||
1431 | /* Make sure that the destination of I3, | |
1432 | which we are going to substitute into one output of I2, | |
1433 | is not used within another output of I2. We must avoid making this: | |
1434 | (parallel [(set (mem (reg 69)) ...) | |
1435 | (set (reg 69) ...)]) | |
1436 | which is not well-defined as to order of actions. | |
1437 | (Besides, reload can't handle output reloads for this.) | |
1438 | ||
1439 | The problem can also happen if the dest of I3 is a memory ref, | |
1440 | if another dest in I2 is an indirect memory ref. */ | |
1441 | for (i = 0; i < XVECLEN (p2, 0); i++) | |
7ca919b7 RK |
1442 | if ((GET_CODE (XVECEXP (p2, 0, i)) == SET |
1443 | || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER) | |
5089e22e RS |
1444 | && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)), |
1445 | SET_DEST (XVECEXP (p2, 0, i)))) | |
1446 | break; | |
230d793d | 1447 | |
5089e22e RS |
1448 | if (i == XVECLEN (p2, 0)) |
1449 | for (i = 0; i < XVECLEN (p2, 0); i++) | |
1450 | if (SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3))) | |
1451 | { | |
1452 | combine_merges++; | |
230d793d | 1453 | |
5089e22e RS |
1454 | subst_insn = i3; |
1455 | subst_low_cuid = INSN_CUID (i2); | |
230d793d | 1456 | |
c4e861e8 | 1457 | added_sets_2 = added_sets_1 = 0; |
5089e22e | 1458 | i2dest = SET_SRC (PATTERN (i3)); |
230d793d | 1459 | |
5089e22e RS |
1460 | /* Replace the dest in I2 with our dest and make the resulting |
1461 | insn the new pattern for I3. Then skip to where we | |
1462 | validate the pattern. Everything was set up above. */ | |
1463 | SUBST (SET_DEST (XVECEXP (p2, 0, i)), | |
1464 | SET_DEST (PATTERN (i3))); | |
1465 | ||
1466 | newpat = p2; | |
176c9e6b | 1467 | i3_subst_into_i2 = 1; |
5089e22e RS |
1468 | goto validate_replacement; |
1469 | } | |
1470 | } | |
230d793d RS |
1471 | |
1472 | #ifndef HAVE_cc0 | |
1473 | /* If we have no I1 and I2 looks like: | |
1474 | (parallel [(set (reg:CC X) (compare:CC OP (const_int 0))) | |
1475 | (set Y OP)]) | |
1476 | make up a dummy I1 that is | |
1477 | (set Y OP) | |
1478 | and change I2 to be | |
1479 | (set (reg:CC X) (compare:CC Y (const_int 0))) | |
1480 | ||
1481 | (We can ignore any trailing CLOBBERs.) | |
1482 | ||
1483 | This undoes a previous combination and allows us to match a branch-and- | |
1484 | decrement insn. */ | |
1485 | ||
1486 | if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL | |
1487 | && XVECLEN (PATTERN (i2), 0) >= 2 | |
1488 | && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET | |
1489 | && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)))) | |
1490 | == MODE_CC) | |
1491 | && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE | |
1492 | && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx | |
1493 | && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET | |
1494 | && GET_CODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 1))) == REG | |
1495 | && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0), | |
1496 | SET_SRC (XVECEXP (PATTERN (i2), 0, 1)))) | |
1497 | { | |
1498 | for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--) | |
1499 | if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER) | |
1500 | break; | |
1501 | ||
1502 | if (i == 1) | |
1503 | { | |
1504 | /* We make I1 with the same INSN_UID as I2. This gives it | |
1505 | the same INSN_CUID for value tracking. Our fake I1 will | |
1506 | never appear in the insn stream so giving it the same INSN_UID | |
1507 | as I2 will not cause a problem. */ | |
1508 | ||
0d9641d1 | 1509 | subst_prev_insn = i1 |
38a448ca RH |
1510 | = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2, |
1511 | XVECEXP (PATTERN (i2), 0, 1), -1, NULL_RTX, | |
1512 | NULL_RTX); | |
230d793d RS |
1513 | |
1514 | SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0)); | |
1515 | SUBST (XEXP (SET_SRC (PATTERN (i2)), 0), | |
1516 | SET_DEST (PATTERN (i1))); | |
1517 | } | |
1518 | } | |
1519 | #endif | |
1520 | ||
1521 | /* Verify that I2 and I1 are valid for combining. */ | |
5f4f0e22 CH |
1522 | if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src) |
1523 | || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src))) | |
230d793d RS |
1524 | { |
1525 | undo_all (); | |
1526 | return 0; | |
1527 | } | |
1528 | ||
1529 | /* Record whether I2DEST is used in I2SRC and similarly for the other | |
1530 | cases. Knowing this will help in register status updating below. */ | |
1531 | i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src); | |
1532 | i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src); | |
1533 | i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src); | |
1534 | ||
916f14f1 | 1535 | /* See if I1 directly feeds into I3. It does if I1DEST is not used |
230d793d RS |
1536 | in I2SRC. */ |
1537 | i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src); | |
1538 | ||
1539 | /* Ensure that I3's pattern can be the destination of combines. */ | |
1540 | if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest, | |
1541 | i1 && i2dest_in_i1src && i1_feeds_i3, | |
1542 | &i3dest_killed)) | |
1543 | { | |
1544 | undo_all (); | |
1545 | return 0; | |
1546 | } | |
1547 | ||
df7d75de RK |
1548 | /* See if any of the insns is a MULT operation. Unless one is, we will |
1549 | reject a combination that is, since it must be slower. Be conservative | |
1550 | here. */ | |
1551 | if (GET_CODE (i2src) == MULT | |
1552 | || (i1 != 0 && GET_CODE (i1src) == MULT) | |
1553 | || (GET_CODE (PATTERN (i3)) == SET | |
1554 | && GET_CODE (SET_SRC (PATTERN (i3))) == MULT)) | |
1555 | have_mult = 1; | |
1556 | ||
230d793d RS |
1557 | /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd. |
1558 | We used to do this EXCEPT in one case: I3 has a post-inc in an | |
1559 | output operand. However, that exception can give rise to insns like | |
1560 | mov r3,(r3)+ | |
1561 | which is a famous insn on the PDP-11 where the value of r3 used as the | |
5089e22e | 1562 | source was model-dependent. Avoid this sort of thing. */ |
230d793d RS |
1563 | |
1564 | #if 0 | |
1565 | if (!(GET_CODE (PATTERN (i3)) == SET | |
1566 | && GET_CODE (SET_SRC (PATTERN (i3))) == REG | |
1567 | && GET_CODE (SET_DEST (PATTERN (i3))) == MEM | |
1568 | && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC | |
1569 | || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC))) | |
1570 | /* It's not the exception. */ | |
1571 | #endif | |
1572 | #ifdef AUTO_INC_DEC | |
1573 | for (link = REG_NOTES (i3); link; link = XEXP (link, 1)) | |
1574 | if (REG_NOTE_KIND (link) == REG_INC | |
1575 | && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2)) | |
1576 | || (i1 != 0 | |
1577 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1))))) | |
1578 | { | |
1579 | undo_all (); | |
1580 | return 0; | |
1581 | } | |
1582 | #endif | |
1583 | ||
1584 | /* See if the SETs in I1 or I2 need to be kept around in the merged | |
1585 | instruction: whenever the value set there is still needed past I3. | |
1586 | For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3. | |
1587 | ||
1588 | For the SET in I1, we have two cases: If I1 and I2 independently | |
1589 | feed into I3, the set in I1 needs to be kept around if I1DEST dies | |
1590 | or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set | |
1591 | in I1 needs to be kept around unless I1DEST dies or is set in either | |
1592 | I2 or I3. We can distinguish these cases by seeing if I2SRC mentions | |
1593 | I1DEST. If so, we know I1 feeds into I2. */ | |
1594 | ||
1595 | added_sets_2 = ! dead_or_set_p (i3, i2dest); | |
1596 | ||
1597 | added_sets_1 | |
1598 | = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest) | |
1599 | : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest))); | |
1600 | ||
1601 | /* If the set in I2 needs to be kept around, we must make a copy of | |
1602 | PATTERN (I2), so that when we substitute I1SRC for I1DEST in | |
5089e22e | 1603 | PATTERN (I2), we are only substituting for the original I1DEST, not into |
230d793d RS |
1604 | an already-substituted copy. This also prevents making self-referential |
1605 | rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to | |
1606 | I2DEST. */ | |
1607 | ||
1608 | i2pat = (GET_CODE (PATTERN (i2)) == PARALLEL | |
38a448ca | 1609 | ? gen_rtx_SET (VOIDmode, i2dest, i2src) |
230d793d RS |
1610 | : PATTERN (i2)); |
1611 | ||
1612 | if (added_sets_2) | |
1613 | i2pat = copy_rtx (i2pat); | |
1614 | ||
1615 | combine_merges++; | |
1616 | ||
1617 | /* Substitute in the latest insn for the regs set by the earlier ones. */ | |
1618 | ||
1619 | maxreg = max_reg_num (); | |
1620 | ||
1621 | subst_insn = i3; | |
230d793d RS |
1622 | |
1623 | /* It is possible that the source of I2 or I1 may be performing an | |
1624 | unneeded operation, such as a ZERO_EXTEND of something that is known | |
1625 | to have the high part zero. Handle that case by letting subst look at | |
1626 | the innermost one of them. | |
1627 | ||
1628 | Another way to do this would be to have a function that tries to | |
1629 | simplify a single insn instead of merging two or more insns. We don't | |
1630 | do this because of the potential of infinite loops and because | |
1631 | of the potential extra memory required. However, doing it the way | |
1632 | we are is a bit of a kludge and doesn't catch all cases. | |
1633 | ||
1634 | But only do this if -fexpensive-optimizations since it slows things down | |
1635 | and doesn't usually win. */ | |
1636 | ||
1637 | if (flag_expensive_optimizations) | |
1638 | { | |
1639 | /* Pass pc_rtx so no substitutions are done, just simplifications. | |
1640 | The cases that we are interested in here do not involve the few | |
1641 | cases were is_replaced is checked. */ | |
1642 | if (i1) | |
d0ab8cd3 RK |
1643 | { |
1644 | subst_low_cuid = INSN_CUID (i1); | |
1645 | i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0); | |
1646 | } | |
230d793d | 1647 | else |
d0ab8cd3 RK |
1648 | { |
1649 | subst_low_cuid = INSN_CUID (i2); | |
1650 | i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0); | |
1651 | } | |
230d793d | 1652 | |
241cea85 | 1653 | undobuf.previous_undos = undobuf.undos; |
230d793d RS |
1654 | } |
1655 | ||
1656 | #ifndef HAVE_cc0 | |
1657 | /* Many machines that don't use CC0 have insns that can both perform an | |
1658 | arithmetic operation and set the condition code. These operations will | |
1659 | be represented as a PARALLEL with the first element of the vector | |
1660 | being a COMPARE of an arithmetic operation with the constant zero. | |
1661 | The second element of the vector will set some pseudo to the result | |
1662 | of the same arithmetic operation. If we simplify the COMPARE, we won't | |
1663 | match such a pattern and so will generate an extra insn. Here we test | |
1664 | for this case, where both the comparison and the operation result are | |
1665 | needed, and make the PARALLEL by just replacing I2DEST in I3SRC with | |
1666 | I2SRC. Later we will make the PARALLEL that contains I2. */ | |
1667 | ||
1668 | if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET | |
1669 | && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE | |
1670 | && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx | |
1671 | && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest)) | |
1672 | { | |
081f5e7e | 1673 | #ifdef EXTRA_CC_MODES |
230d793d RS |
1674 | rtx *cc_use; |
1675 | enum machine_mode compare_mode; | |
081f5e7e | 1676 | #endif |
230d793d RS |
1677 | |
1678 | newpat = PATTERN (i3); | |
1679 | SUBST (XEXP (SET_SRC (newpat), 0), i2src); | |
1680 | ||
1681 | i2_is_used = 1; | |
1682 | ||
1683 | #ifdef EXTRA_CC_MODES | |
1684 | /* See if a COMPARE with the operand we substituted in should be done | |
1685 | with the mode that is currently being used. If not, do the same | |
1686 | processing we do in `subst' for a SET; namely, if the destination | |
1687 | is used only once, try to replace it with a register of the proper | |
1688 | mode and also replace the COMPARE. */ | |
1689 | if (undobuf.other_insn == 0 | |
1690 | && (cc_use = find_single_use (SET_DEST (newpat), i3, | |
1691 | &undobuf.other_insn)) | |
77fa0940 RK |
1692 | && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use), |
1693 | i2src, const0_rtx)) | |
230d793d RS |
1694 | != GET_MODE (SET_DEST (newpat)))) |
1695 | { | |
1696 | int regno = REGNO (SET_DEST (newpat)); | |
38a448ca | 1697 | rtx new_dest = gen_rtx_REG (compare_mode, regno); |
230d793d RS |
1698 | |
1699 | if (regno < FIRST_PSEUDO_REGISTER | |
b1f21e0a | 1700 | || (REG_N_SETS (regno) == 1 && ! added_sets_2 |
230d793d RS |
1701 | && ! REG_USERVAR_P (SET_DEST (newpat)))) |
1702 | { | |
1703 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1704 | SUBST (regno_reg_rtx[regno], new_dest); | |
1705 | ||
1706 | SUBST (SET_DEST (newpat), new_dest); | |
1707 | SUBST (XEXP (*cc_use, 0), new_dest); | |
1708 | SUBST (SET_SRC (newpat), | |
1709 | gen_rtx_combine (COMPARE, compare_mode, | |
1710 | i2src, const0_rtx)); | |
1711 | } | |
1712 | else | |
1713 | undobuf.other_insn = 0; | |
1714 | } | |
1715 | #endif | |
1716 | } | |
1717 | else | |
1718 | #endif | |
1719 | { | |
1720 | n_occurrences = 0; /* `subst' counts here */ | |
1721 | ||
1722 | /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we | |
1723 | need to make a unique copy of I2SRC each time we substitute it | |
1724 | to avoid self-referential rtl. */ | |
1725 | ||
d0ab8cd3 | 1726 | subst_low_cuid = INSN_CUID (i2); |
230d793d RS |
1727 | newpat = subst (PATTERN (i3), i2dest, i2src, 0, |
1728 | ! i1_feeds_i3 && i1dest_in_i1src); | |
241cea85 | 1729 | undobuf.previous_undos = undobuf.undos; |
230d793d RS |
1730 | |
1731 | /* Record whether i2's body now appears within i3's body. */ | |
1732 | i2_is_used = n_occurrences; | |
1733 | } | |
1734 | ||
1735 | /* If we already got a failure, don't try to do more. Otherwise, | |
1736 | try to substitute in I1 if we have it. */ | |
1737 | ||
1738 | if (i1 && GET_CODE (newpat) != CLOBBER) | |
1739 | { | |
1740 | /* Before we can do this substitution, we must redo the test done | |
1741 | above (see detailed comments there) that ensures that I1DEST | |
0f41302f | 1742 | isn't mentioned in any SETs in NEWPAT that are field assignments. */ |
230d793d | 1743 | |
5f4f0e22 CH |
1744 | if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX, |
1745 | 0, NULL_PTR)) | |
230d793d RS |
1746 | { |
1747 | undo_all (); | |
1748 | return 0; | |
1749 | } | |
1750 | ||
1751 | n_occurrences = 0; | |
d0ab8cd3 | 1752 | subst_low_cuid = INSN_CUID (i1); |
230d793d | 1753 | newpat = subst (newpat, i1dest, i1src, 0, 0); |
241cea85 | 1754 | undobuf.previous_undos = undobuf.undos; |
230d793d RS |
1755 | } |
1756 | ||
916f14f1 RK |
1757 | /* Fail if an autoincrement side-effect has been duplicated. Be careful |
1758 | to count all the ways that I2SRC and I1SRC can be used. */ | |
5f4f0e22 | 1759 | if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0 |
916f14f1 | 1760 | && i2_is_used + added_sets_2 > 1) |
5f4f0e22 | 1761 | || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0 |
916f14f1 RK |
1762 | && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3) |
1763 | > 1)) | |
230d793d RS |
1764 | /* Fail if we tried to make a new register (we used to abort, but there's |
1765 | really no reason to). */ | |
1766 | || max_reg_num () != maxreg | |
1767 | /* Fail if we couldn't do something and have a CLOBBER. */ | |
df7d75de RK |
1768 | || GET_CODE (newpat) == CLOBBER |
1769 | /* Fail if this new pattern is a MULT and we didn't have one before | |
1770 | at the outer level. */ | |
1771 | || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT | |
1772 | && ! have_mult)) | |
230d793d RS |
1773 | { |
1774 | undo_all (); | |
1775 | return 0; | |
1776 | } | |
1777 | ||
1778 | /* If the actions of the earlier insns must be kept | |
1779 | in addition to substituting them into the latest one, | |
1780 | we must make a new PARALLEL for the latest insn | |
1781 | to hold additional the SETs. */ | |
1782 | ||
1783 | if (added_sets_1 || added_sets_2) | |
1784 | { | |
1785 | combine_extras++; | |
1786 | ||
1787 | if (GET_CODE (newpat) == PARALLEL) | |
1788 | { | |
1789 | rtvec old = XVEC (newpat, 0); | |
1790 | total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2; | |
38a448ca | 1791 | newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets)); |
59888de2 | 1792 | bcopy ((char *) &old->elem[0], (char *) XVEC (newpat, 0)->elem, |
230d793d RS |
1793 | sizeof (old->elem[0]) * old->num_elem); |
1794 | } | |
1795 | else | |
1796 | { | |
1797 | rtx old = newpat; | |
1798 | total_sets = 1 + added_sets_1 + added_sets_2; | |
38a448ca | 1799 | newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets)); |
230d793d RS |
1800 | XVECEXP (newpat, 0, 0) = old; |
1801 | } | |
1802 | ||
1803 | if (added_sets_1) | |
1804 | XVECEXP (newpat, 0, --total_sets) | |
1805 | = (GET_CODE (PATTERN (i1)) == PARALLEL | |
38a448ca | 1806 | ? gen_rtx_SET (VOIDmode, i1dest, i1src) : PATTERN (i1)); |
230d793d RS |
1807 | |
1808 | if (added_sets_2) | |
1809 | { | |
1810 | /* If there is no I1, use I2's body as is. We used to also not do | |
1811 | the subst call below if I2 was substituted into I3, | |
1812 | but that could lose a simplification. */ | |
1813 | if (i1 == 0) | |
1814 | XVECEXP (newpat, 0, --total_sets) = i2pat; | |
1815 | else | |
1816 | /* See comment where i2pat is assigned. */ | |
1817 | XVECEXP (newpat, 0, --total_sets) | |
1818 | = subst (i2pat, i1dest, i1src, 0, 0); | |
1819 | } | |
1820 | } | |
1821 | ||
1822 | /* We come here when we are replacing a destination in I2 with the | |
1823 | destination of I3. */ | |
1824 | validate_replacement: | |
1825 | ||
6e25d159 RK |
1826 | /* Note which hard regs this insn has as inputs. */ |
1827 | mark_used_regs_combine (newpat); | |
1828 | ||
230d793d | 1829 | /* Is the result of combination a valid instruction? */ |
a29ca9db RK |
1830 | insn_code_number |
1831 | = recog_for_combine (&newpat, i3, &new_i3_notes, &i3_scratches); | |
230d793d RS |
1832 | |
1833 | /* If the result isn't valid, see if it is a PARALLEL of two SETs where | |
1834 | the second SET's destination is a register that is unused. In that case, | |
1835 | we just need the first SET. This can occur when simplifying a divmod | |
1836 | insn. We *must* test for this case here because the code below that | |
1837 | splits two independent SETs doesn't handle this case correctly when it | |
1838 | updates the register status. Also check the case where the first | |
1839 | SET's destination is unused. That would not cause incorrect code, but | |
1840 | does cause an unneeded insn to remain. */ | |
1841 | ||
1842 | if (insn_code_number < 0 && GET_CODE (newpat) == PARALLEL | |
1843 | && XVECLEN (newpat, 0) == 2 | |
1844 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
1845 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
1846 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == REG | |
1847 | && find_reg_note (i3, REG_UNUSED, SET_DEST (XVECEXP (newpat, 0, 1))) | |
1848 | && ! side_effects_p (SET_SRC (XVECEXP (newpat, 0, 1))) | |
1849 | && asm_noperands (newpat) < 0) | |
1850 | { | |
1851 | newpat = XVECEXP (newpat, 0, 0); | |
a29ca9db RK |
1852 | insn_code_number |
1853 | = recog_for_combine (&newpat, i3, &new_i3_notes, &i3_scratches); | |
230d793d RS |
1854 | } |
1855 | ||
1856 | else if (insn_code_number < 0 && GET_CODE (newpat) == PARALLEL | |
1857 | && XVECLEN (newpat, 0) == 2 | |
1858 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
1859 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
1860 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) == REG | |
1861 | && find_reg_note (i3, REG_UNUSED, SET_DEST (XVECEXP (newpat, 0, 0))) | |
1862 | && ! side_effects_p (SET_SRC (XVECEXP (newpat, 0, 0))) | |
1863 | && asm_noperands (newpat) < 0) | |
1864 | { | |
1865 | newpat = XVECEXP (newpat, 0, 1); | |
a29ca9db RK |
1866 | insn_code_number |
1867 | = recog_for_combine (&newpat, i3, &new_i3_notes, &i3_scratches); | |
230d793d RS |
1868 | } |
1869 | ||
1870 | /* If we were combining three insns and the result is a simple SET | |
1871 | with no ASM_OPERANDS that wasn't recognized, try to split it into two | |
916f14f1 RK |
1872 | insns. There are two ways to do this. It can be split using a |
1873 | machine-specific method (like when you have an addition of a large | |
1874 | constant) or by combine in the function find_split_point. */ | |
1875 | ||
230d793d RS |
1876 | if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET |
1877 | && asm_noperands (newpat) < 0) | |
1878 | { | |
916f14f1 | 1879 | rtx m_split, *split; |
42495ca0 | 1880 | rtx ni2dest = i2dest; |
916f14f1 RK |
1881 | |
1882 | /* See if the MD file can split NEWPAT. If it can't, see if letting it | |
42495ca0 RK |
1883 | use I2DEST as a scratch register will help. In the latter case, |
1884 | convert I2DEST to the mode of the source of NEWPAT if we can. */ | |
916f14f1 RK |
1885 | |
1886 | m_split = split_insns (newpat, i3); | |
a70c61d9 JW |
1887 | |
1888 | /* We can only use I2DEST as a scratch reg if it doesn't overlap any | |
1889 | inputs of NEWPAT. */ | |
1890 | ||
1891 | /* ??? If I2DEST is not safe, and I1DEST exists, then it would be | |
1892 | possible to try that as a scratch reg. This would require adding | |
1893 | more code to make it work though. */ | |
1894 | ||
1895 | if (m_split == 0 && ! reg_overlap_mentioned_p (ni2dest, newpat)) | |
42495ca0 RK |
1896 | { |
1897 | /* If I2DEST is a hard register or the only use of a pseudo, | |
1898 | we can change its mode. */ | |
1899 | if (GET_MODE (SET_DEST (newpat)) != GET_MODE (i2dest) | |
02f4ada4 | 1900 | && GET_MODE (SET_DEST (newpat)) != VOIDmode |
60654f77 | 1901 | && GET_CODE (i2dest) == REG |
42495ca0 | 1902 | && (REGNO (i2dest) < FIRST_PSEUDO_REGISTER |
b1f21e0a | 1903 | || (REG_N_SETS (REGNO (i2dest)) == 1 && ! added_sets_2 |
42495ca0 | 1904 | && ! REG_USERVAR_P (i2dest)))) |
38a448ca | 1905 | ni2dest = gen_rtx_REG (GET_MODE (SET_DEST (newpat)), |
42495ca0 RK |
1906 | REGNO (i2dest)); |
1907 | ||
38a448ca RH |
1908 | m_split = split_insns |
1909 | (gen_rtx_PARALLEL (VOIDmode, | |
1910 | gen_rtvec (2, newpat, | |
1911 | gen_rtx_CLOBBER (VOIDmode, | |
1912 | ni2dest))), | |
1913 | i3); | |
42495ca0 | 1914 | } |
916f14f1 RK |
1915 | |
1916 | if (m_split && GET_CODE (m_split) == SEQUENCE | |
3f508eca RK |
1917 | && XVECLEN (m_split, 0) == 2 |
1918 | && (next_real_insn (i2) == i3 | |
1919 | || ! use_crosses_set_p (PATTERN (XVECEXP (m_split, 0, 0)), | |
1920 | INSN_CUID (i2)))) | |
916f14f1 | 1921 | { |
1a26b032 | 1922 | rtx i2set, i3set; |
d0ab8cd3 | 1923 | rtx newi3pat = PATTERN (XVECEXP (m_split, 0, 1)); |
916f14f1 | 1924 | newi2pat = PATTERN (XVECEXP (m_split, 0, 0)); |
916f14f1 | 1925 | |
e4ba89be RK |
1926 | i3set = single_set (XVECEXP (m_split, 0, 1)); |
1927 | i2set = single_set (XVECEXP (m_split, 0, 0)); | |
1a26b032 | 1928 | |
42495ca0 RK |
1929 | /* In case we changed the mode of I2DEST, replace it in the |
1930 | pseudo-register table here. We can't do it above in case this | |
1931 | code doesn't get executed and we do a split the other way. */ | |
1932 | ||
1933 | if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER) | |
1934 | SUBST (regno_reg_rtx[REGNO (i2dest)], ni2dest); | |
1935 | ||
a29ca9db RK |
1936 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes, |
1937 | &i2_scratches); | |
1a26b032 RK |
1938 | |
1939 | /* If I2 or I3 has multiple SETs, we won't know how to track | |
9cc96794 RK |
1940 | register status, so don't use these insns. If I2's destination |
1941 | is used between I2 and I3, we also can't use these insns. */ | |
1a26b032 | 1942 | |
9cc96794 RK |
1943 | if (i2_code_number >= 0 && i2set && i3set |
1944 | && (next_real_insn (i2) == i3 | |
1945 | || ! reg_used_between_p (SET_DEST (i2set), i2, i3))) | |
a29ca9db RK |
1946 | insn_code_number = recog_for_combine (&newi3pat, i3, &new_i3_notes, |
1947 | &i3_scratches); | |
d0ab8cd3 RK |
1948 | if (insn_code_number >= 0) |
1949 | newpat = newi3pat; | |
1950 | ||
c767f54b | 1951 | /* It is possible that both insns now set the destination of I3. |
22609cbf | 1952 | If so, we must show an extra use of it. */ |
c767f54b | 1953 | |
393de53f RK |
1954 | if (insn_code_number >= 0) |
1955 | { | |
1956 | rtx new_i3_dest = SET_DEST (i3set); | |
1957 | rtx new_i2_dest = SET_DEST (i2set); | |
1958 | ||
1959 | while (GET_CODE (new_i3_dest) == ZERO_EXTRACT | |
1960 | || GET_CODE (new_i3_dest) == STRICT_LOW_PART | |
1961 | || GET_CODE (new_i3_dest) == SUBREG) | |
1962 | new_i3_dest = XEXP (new_i3_dest, 0); | |
1963 | ||
d4096689 RK |
1964 | while (GET_CODE (new_i2_dest) == ZERO_EXTRACT |
1965 | || GET_CODE (new_i2_dest) == STRICT_LOW_PART | |
1966 | || GET_CODE (new_i2_dest) == SUBREG) | |
1967 | new_i2_dest = XEXP (new_i2_dest, 0); | |
1968 | ||
393de53f RK |
1969 | if (GET_CODE (new_i3_dest) == REG |
1970 | && GET_CODE (new_i2_dest) == REG | |
1971 | && REGNO (new_i3_dest) == REGNO (new_i2_dest)) | |
b1f21e0a | 1972 | REG_N_SETS (REGNO (new_i2_dest))++; |
393de53f | 1973 | } |
916f14f1 | 1974 | } |
230d793d RS |
1975 | |
1976 | /* If we can split it and use I2DEST, go ahead and see if that | |
1977 | helps things be recognized. Verify that none of the registers | |
1978 | are set between I2 and I3. */ | |
d0ab8cd3 | 1979 | if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0 |
230d793d RS |
1980 | #ifdef HAVE_cc0 |
1981 | && GET_CODE (i2dest) == REG | |
1982 | #endif | |
1983 | /* We need I2DEST in the proper mode. If it is a hard register | |
1984 | or the only use of a pseudo, we can change its mode. */ | |
1985 | && (GET_MODE (*split) == GET_MODE (i2dest) | |
1986 | || GET_MODE (*split) == VOIDmode | |
1987 | || REGNO (i2dest) < FIRST_PSEUDO_REGISTER | |
b1f21e0a | 1988 | || (REG_N_SETS (REGNO (i2dest)) == 1 && ! added_sets_2 |
230d793d RS |
1989 | && ! REG_USERVAR_P (i2dest))) |
1990 | && (next_real_insn (i2) == i3 | |
1991 | || ! use_crosses_set_p (*split, INSN_CUID (i2))) | |
1992 | /* We can't overwrite I2DEST if its value is still used by | |
1993 | NEWPAT. */ | |
1994 | && ! reg_referenced_p (i2dest, newpat)) | |
1995 | { | |
1996 | rtx newdest = i2dest; | |
df7d75de RK |
1997 | enum rtx_code split_code = GET_CODE (*split); |
1998 | enum machine_mode split_mode = GET_MODE (*split); | |
230d793d RS |
1999 | |
2000 | /* Get NEWDEST as a register in the proper mode. We have already | |
2001 | validated that we can do this. */ | |
df7d75de | 2002 | if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode) |
230d793d | 2003 | { |
38a448ca | 2004 | newdest = gen_rtx_REG (split_mode, REGNO (i2dest)); |
230d793d RS |
2005 | |
2006 | if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER) | |
2007 | SUBST (regno_reg_rtx[REGNO (i2dest)], newdest); | |
2008 | } | |
2009 | ||
2010 | /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to | |
2011 | an ASHIFT. This can occur if it was inside a PLUS and hence | |
2012 | appeared to be a memory address. This is a kludge. */ | |
df7d75de | 2013 | if (split_code == MULT |
230d793d RS |
2014 | && GET_CODE (XEXP (*split, 1)) == CONST_INT |
2015 | && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0) | |
1dc8a823 JW |
2016 | { |
2017 | SUBST (*split, gen_rtx_combine (ASHIFT, split_mode, | |
2018 | XEXP (*split, 0), GEN_INT (i))); | |
2019 | /* Update split_code because we may not have a multiply | |
2020 | anymore. */ | |
2021 | split_code = GET_CODE (*split); | |
2022 | } | |
230d793d RS |
2023 | |
2024 | #ifdef INSN_SCHEDULING | |
2025 | /* If *SPLIT is a paradoxical SUBREG, when we split it, it should | |
2026 | be written as a ZERO_EXTEND. */ | |
df7d75de RK |
2027 | if (split_code == SUBREG && GET_CODE (SUBREG_REG (*split)) == MEM) |
2028 | SUBST (*split, gen_rtx_combine (ZERO_EXTEND, split_mode, | |
230d793d RS |
2029 | XEXP (*split, 0))); |
2030 | #endif | |
2031 | ||
2032 | newi2pat = gen_rtx_combine (SET, VOIDmode, newdest, *split); | |
2033 | SUBST (*split, newdest); | |
a29ca9db RK |
2034 | i2_code_number |
2035 | = recog_for_combine (&newi2pat, i2, &new_i2_notes, &i2_scratches); | |
df7d75de RK |
2036 | |
2037 | /* If the split point was a MULT and we didn't have one before, | |
2038 | don't use one now. */ | |
2039 | if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult)) | |
a29ca9db RK |
2040 | insn_code_number |
2041 | = recog_for_combine (&newpat, i3, &new_i3_notes, &i3_scratches); | |
230d793d RS |
2042 | } |
2043 | } | |
2044 | ||
2045 | /* Check for a case where we loaded from memory in a narrow mode and | |
2046 | then sign extended it, but we need both registers. In that case, | |
2047 | we have a PARALLEL with both loads from the same memory location. | |
2048 | We can split this into a load from memory followed by a register-register | |
2049 | copy. This saves at least one insn, more if register allocation can | |
f0343c74 RK |
2050 | eliminate the copy. |
2051 | ||
2052 | We cannot do this if the destination of the second assignment is | |
2053 | a register that we have already assumed is zero-extended. Similarly | |
2054 | for a SUBREG of such a register. */ | |
230d793d RS |
2055 | |
2056 | else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0 | |
2057 | && GET_CODE (newpat) == PARALLEL | |
2058 | && XVECLEN (newpat, 0) == 2 | |
2059 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
2060 | && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND | |
2061 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
2062 | && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)), | |
2063 | XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0)) | |
2064 | && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)), | |
2065 | INSN_CUID (i2)) | |
2066 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT | |
2067 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART | |
f0343c74 RK |
2068 | && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)), |
2069 | (GET_CODE (temp) == REG | |
2070 | && reg_nonzero_bits[REGNO (temp)] != 0 | |
2071 | && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD | |
2072 | && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT | |
2073 | && (reg_nonzero_bits[REGNO (temp)] | |
2074 | != GET_MODE_MASK (word_mode)))) | |
2075 | && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG | |
2076 | && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))), | |
2077 | (GET_CODE (temp) == REG | |
2078 | && reg_nonzero_bits[REGNO (temp)] != 0 | |
2079 | && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD | |
2080 | && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT | |
2081 | && (reg_nonzero_bits[REGNO (temp)] | |
2082 | != GET_MODE_MASK (word_mode))))) | |
230d793d RS |
2083 | && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)), |
2084 | SET_SRC (XVECEXP (newpat, 0, 1))) | |
2085 | && ! find_reg_note (i3, REG_UNUSED, | |
2086 | SET_DEST (XVECEXP (newpat, 0, 0)))) | |
2087 | { | |
472fbdd1 RK |
2088 | rtx ni2dest; |
2089 | ||
230d793d | 2090 | newi2pat = XVECEXP (newpat, 0, 0); |
472fbdd1 | 2091 | ni2dest = SET_DEST (XVECEXP (newpat, 0, 0)); |
230d793d RS |
2092 | newpat = XVECEXP (newpat, 0, 1); |
2093 | SUBST (SET_SRC (newpat), | |
472fbdd1 | 2094 | gen_lowpart_for_combine (GET_MODE (SET_SRC (newpat)), ni2dest)); |
a29ca9db RK |
2095 | i2_code_number |
2096 | = recog_for_combine (&newi2pat, i2, &new_i2_notes, &i2_scratches); | |
2097 | ||
230d793d | 2098 | if (i2_code_number >= 0) |
a29ca9db RK |
2099 | insn_code_number |
2100 | = recog_for_combine (&newpat, i3, &new_i3_notes, &i3_scratches); | |
5089e22e RS |
2101 | |
2102 | if (insn_code_number >= 0) | |
2103 | { | |
2104 | rtx insn; | |
2105 | rtx link; | |
2106 | ||
2107 | /* If we will be able to accept this, we have made a change to the | |
2108 | destination of I3. This can invalidate a LOG_LINKS pointing | |
2109 | to I3. No other part of combine.c makes such a transformation. | |
2110 | ||
2111 | The new I3 will have a destination that was previously the | |
2112 | destination of I1 or I2 and which was used in i2 or I3. Call | |
2113 | distribute_links to make a LOG_LINK from the next use of | |
2114 | that destination. */ | |
2115 | ||
2116 | PATTERN (i3) = newpat; | |
38a448ca | 2117 | distribute_links (gen_rtx_INSN_LIST (VOIDmode, i3, NULL_RTX)); |
5089e22e RS |
2118 | |
2119 | /* I3 now uses what used to be its destination and which is | |
2120 | now I2's destination. That means we need a LOG_LINK from | |
2121 | I3 to I2. But we used to have one, so we still will. | |
2122 | ||
2123 | However, some later insn might be using I2's dest and have | |
2124 | a LOG_LINK pointing at I3. We must remove this link. | |
2125 | The simplest way to remove the link is to point it at I1, | |
2126 | which we know will be a NOTE. */ | |
2127 | ||
2128 | for (insn = NEXT_INSN (i3); | |
0d4d42c3 RK |
2129 | insn && (this_basic_block == n_basic_blocks - 1 |
2130 | || insn != basic_block_head[this_basic_block + 1]); | |
5089e22e RS |
2131 | insn = NEXT_INSN (insn)) |
2132 | { | |
2133 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
472fbdd1 | 2134 | && reg_referenced_p (ni2dest, PATTERN (insn))) |
5089e22e RS |
2135 | { |
2136 | for (link = LOG_LINKS (insn); link; | |
2137 | link = XEXP (link, 1)) | |
2138 | if (XEXP (link, 0) == i3) | |
2139 | XEXP (link, 0) = i1; | |
2140 | ||
2141 | break; | |
2142 | } | |
2143 | } | |
2144 | } | |
230d793d RS |
2145 | } |
2146 | ||
2147 | /* Similarly, check for a case where we have a PARALLEL of two independent | |
2148 | SETs but we started with three insns. In this case, we can do the sets | |
2149 | as two separate insns. This case occurs when some SET allows two | |
2150 | other insns to combine, but the destination of that SET is still live. */ | |
2151 | ||
2152 | else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0 | |
2153 | && GET_CODE (newpat) == PARALLEL | |
2154 | && XVECLEN (newpat, 0) == 2 | |
2155 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET | |
2156 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT | |
2157 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART | |
2158 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET | |
2159 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT | |
2160 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART | |
2161 | && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)), | |
2162 | INSN_CUID (i2)) | |
2163 | /* Don't pass sets with (USE (MEM ...)) dests to the following. */ | |
2164 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != USE | |
2165 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != USE | |
2166 | && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)), | |
2167 | XVECEXP (newpat, 0, 0)) | |
2168 | && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)), | |
2169 | XVECEXP (newpat, 0, 1))) | |
2170 | { | |
e9a25f70 JL |
2171 | /* Normally, it doesn't matter which of the two is done first, |
2172 | but it does if one references cc0. In that case, it has to | |
2173 | be first. */ | |
2174 | #ifdef HAVE_cc0 | |
2175 | if (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0))) | |
2176 | { | |
2177 | newi2pat = XVECEXP (newpat, 0, 0); | |
2178 | newpat = XVECEXP (newpat, 0, 1); | |
2179 | } | |
2180 | else | |
2181 | #endif | |
2182 | { | |
2183 | newi2pat = XVECEXP (newpat, 0, 1); | |
2184 | newpat = XVECEXP (newpat, 0, 0); | |
2185 | } | |
230d793d | 2186 | |
a29ca9db RK |
2187 | i2_code_number |
2188 | = recog_for_combine (&newi2pat, i2, &new_i2_notes, &i2_scratches); | |
2189 | ||
230d793d | 2190 | if (i2_code_number >= 0) |
a29ca9db RK |
2191 | insn_code_number |
2192 | = recog_for_combine (&newpat, i3, &new_i3_notes, &i3_scratches); | |
230d793d RS |
2193 | } |
2194 | ||
2195 | /* If it still isn't recognized, fail and change things back the way they | |
2196 | were. */ | |
2197 | if ((insn_code_number < 0 | |
2198 | /* Is the result a reasonable ASM_OPERANDS? */ | |
2199 | && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2))) | |
2200 | { | |
2201 | undo_all (); | |
2202 | return 0; | |
2203 | } | |
2204 | ||
2205 | /* If we had to change another insn, make sure it is valid also. */ | |
2206 | if (undobuf.other_insn) | |
2207 | { | |
230d793d RS |
2208 | rtx other_pat = PATTERN (undobuf.other_insn); |
2209 | rtx new_other_notes; | |
2210 | rtx note, next; | |
2211 | ||
6e25d159 RK |
2212 | CLEAR_HARD_REG_SET (newpat_used_regs); |
2213 | ||
a29ca9db RK |
2214 | other_code_number |
2215 | = recog_for_combine (&other_pat, undobuf.other_insn, | |
2216 | &new_other_notes, &other_scratches); | |
230d793d RS |
2217 | |
2218 | if (other_code_number < 0 && ! check_asm_operands (other_pat)) | |
2219 | { | |
2220 | undo_all (); | |
2221 | return 0; | |
2222 | } | |
2223 | ||
2224 | PATTERN (undobuf.other_insn) = other_pat; | |
2225 | ||
2226 | /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they | |
2227 | are still valid. Then add any non-duplicate notes added by | |
2228 | recog_for_combine. */ | |
2229 | for (note = REG_NOTES (undobuf.other_insn); note; note = next) | |
2230 | { | |
2231 | next = XEXP (note, 1); | |
2232 | ||
2233 | if (REG_NOTE_KIND (note) == REG_UNUSED | |
2234 | && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn))) | |
1a26b032 RK |
2235 | { |
2236 | if (GET_CODE (XEXP (note, 0)) == REG) | |
b1f21e0a | 2237 | REG_N_DEATHS (REGNO (XEXP (note, 0)))--; |
1a26b032 RK |
2238 | |
2239 | remove_note (undobuf.other_insn, note); | |
2240 | } | |
230d793d RS |
2241 | } |
2242 | ||
1a26b032 RK |
2243 | for (note = new_other_notes; note; note = XEXP (note, 1)) |
2244 | if (GET_CODE (XEXP (note, 0)) == REG) | |
b1f21e0a | 2245 | REG_N_DEATHS (REGNO (XEXP (note, 0)))++; |
1a26b032 | 2246 | |
230d793d | 2247 | distribute_notes (new_other_notes, undobuf.other_insn, |
5f4f0e22 | 2248 | undobuf.other_insn, NULL_RTX, NULL_RTX, NULL_RTX); |
230d793d RS |
2249 | } |
2250 | ||
2251 | /* We now know that we can do this combination. Merge the insns and | |
2252 | update the status of registers and LOG_LINKS. */ | |
2253 | ||
2254 | { | |
2255 | rtx i3notes, i2notes, i1notes = 0; | |
2256 | rtx i3links, i2links, i1links = 0; | |
2257 | rtx midnotes = 0; | |
230d793d | 2258 | register int regno; |
ff3467a9 JW |
2259 | /* Compute which registers we expect to eliminate. newi2pat may be setting |
2260 | either i3dest or i2dest, so we must check it. Also, i1dest may be the | |
2261 | same as i3dest, in which case newi2pat may be setting i1dest. */ | |
2262 | rtx elim_i2 = ((newi2pat && reg_set_p (i2dest, newi2pat)) | |
2263 | || i2dest_in_i2src || i2dest_in_i1src | |
230d793d | 2264 | ? 0 : i2dest); |
ff3467a9 JW |
2265 | rtx elim_i1 = (i1 == 0 || i1dest_in_i1src |
2266 | || (newi2pat && reg_set_p (i1dest, newi2pat)) | |
2267 | ? 0 : i1dest); | |
230d793d RS |
2268 | |
2269 | /* Get the old REG_NOTES and LOG_LINKS from all our insns and | |
2270 | clear them. */ | |
2271 | i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3); | |
2272 | i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2); | |
2273 | if (i1) | |
2274 | i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1); | |
2275 | ||
2276 | /* Ensure that we do not have something that should not be shared but | |
2277 | occurs multiple times in the new insns. Check this by first | |
5089e22e | 2278 | resetting all the `used' flags and then copying anything is shared. */ |
230d793d RS |
2279 | |
2280 | reset_used_flags (i3notes); | |
2281 | reset_used_flags (i2notes); | |
2282 | reset_used_flags (i1notes); | |
2283 | reset_used_flags (newpat); | |
2284 | reset_used_flags (newi2pat); | |
2285 | if (undobuf.other_insn) | |
2286 | reset_used_flags (PATTERN (undobuf.other_insn)); | |
2287 | ||
2288 | i3notes = copy_rtx_if_shared (i3notes); | |
2289 | i2notes = copy_rtx_if_shared (i2notes); | |
2290 | i1notes = copy_rtx_if_shared (i1notes); | |
2291 | newpat = copy_rtx_if_shared (newpat); | |
2292 | newi2pat = copy_rtx_if_shared (newi2pat); | |
2293 | if (undobuf.other_insn) | |
2294 | reset_used_flags (PATTERN (undobuf.other_insn)); | |
2295 | ||
2296 | INSN_CODE (i3) = insn_code_number; | |
2297 | PATTERN (i3) = newpat; | |
2298 | if (undobuf.other_insn) | |
2299 | INSN_CODE (undobuf.other_insn) = other_code_number; | |
2300 | ||
2301 | /* We had one special case above where I2 had more than one set and | |
2302 | we replaced a destination of one of those sets with the destination | |
2303 | of I3. In that case, we have to update LOG_LINKS of insns later | |
176c9e6b JW |
2304 | in this basic block. Note that this (expensive) case is rare. |
2305 | ||
2306 | Also, in this case, we must pretend that all REG_NOTEs for I2 | |
2307 | actually came from I3, so that REG_UNUSED notes from I2 will be | |
2308 | properly handled. */ | |
2309 | ||
2310 | if (i3_subst_into_i2) | |
2311 | { | |
2312 | for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++) | |
2313 | if (GET_CODE (SET_DEST (XVECEXP (PATTERN (i2), 0, i))) == REG | |
2314 | && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest | |
2315 | && ! find_reg_note (i2, REG_UNUSED, | |
2316 | SET_DEST (XVECEXP (PATTERN (i2), 0, i)))) | |
2317 | for (temp = NEXT_INSN (i2); | |
2318 | temp && (this_basic_block == n_basic_blocks - 1 | |
2319 | || basic_block_head[this_basic_block] != temp); | |
2320 | temp = NEXT_INSN (temp)) | |
2321 | if (temp != i3 && GET_RTX_CLASS (GET_CODE (temp)) == 'i') | |
2322 | for (link = LOG_LINKS (temp); link; link = XEXP (link, 1)) | |
2323 | if (XEXP (link, 0) == i2) | |
2324 | XEXP (link, 0) = i3; | |
2325 | ||
2326 | if (i3notes) | |
2327 | { | |
2328 | rtx link = i3notes; | |
2329 | while (XEXP (link, 1)) | |
2330 | link = XEXP (link, 1); | |
2331 | XEXP (link, 1) = i2notes; | |
2332 | } | |
2333 | else | |
2334 | i3notes = i2notes; | |
2335 | i2notes = 0; | |
2336 | } | |
230d793d RS |
2337 | |
2338 | LOG_LINKS (i3) = 0; | |
2339 | REG_NOTES (i3) = 0; | |
2340 | LOG_LINKS (i2) = 0; | |
2341 | REG_NOTES (i2) = 0; | |
2342 | ||
2343 | if (newi2pat) | |
2344 | { | |
2345 | INSN_CODE (i2) = i2_code_number; | |
2346 | PATTERN (i2) = newi2pat; | |
2347 | } | |
2348 | else | |
2349 | { | |
2350 | PUT_CODE (i2, NOTE); | |
2351 | NOTE_LINE_NUMBER (i2) = NOTE_INSN_DELETED; | |
2352 | NOTE_SOURCE_FILE (i2) = 0; | |
2353 | } | |
2354 | ||
2355 | if (i1) | |
2356 | { | |
2357 | LOG_LINKS (i1) = 0; | |
2358 | REG_NOTES (i1) = 0; | |
2359 | PUT_CODE (i1, NOTE); | |
2360 | NOTE_LINE_NUMBER (i1) = NOTE_INSN_DELETED; | |
2361 | NOTE_SOURCE_FILE (i1) = 0; | |
2362 | } | |
2363 | ||
2364 | /* Get death notes for everything that is now used in either I3 or | |
6eb12cef RK |
2365 | I2 and used to die in a previous insn. If we built two new |
2366 | patterns, move from I1 to I2 then I2 to I3 so that we get the | |
2367 | proper movement on registers that I2 modifies. */ | |
230d793d | 2368 | |
230d793d | 2369 | if (newi2pat) |
6eb12cef RK |
2370 | { |
2371 | move_deaths (newi2pat, NULL_RTX, INSN_CUID (i1), i2, &midnotes); | |
2372 | move_deaths (newpat, newi2pat, INSN_CUID (i1), i3, &midnotes); | |
2373 | } | |
2374 | else | |
2375 | move_deaths (newpat, NULL_RTX, i1 ? INSN_CUID (i1) : INSN_CUID (i2), | |
2376 | i3, &midnotes); | |
230d793d RS |
2377 | |
2378 | /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */ | |
2379 | if (i3notes) | |
5f4f0e22 CH |
2380 | distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX, |
2381 | elim_i2, elim_i1); | |
230d793d | 2382 | if (i2notes) |
5f4f0e22 CH |
2383 | distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX, |
2384 | elim_i2, elim_i1); | |
230d793d | 2385 | if (i1notes) |
5f4f0e22 CH |
2386 | distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX, |
2387 | elim_i2, elim_i1); | |
230d793d | 2388 | if (midnotes) |
5f4f0e22 CH |
2389 | distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, |
2390 | elim_i2, elim_i1); | |
230d793d RS |
2391 | |
2392 | /* Distribute any notes added to I2 or I3 by recog_for_combine. We | |
2393 | know these are REG_UNUSED and want them to go to the desired insn, | |
1a26b032 RK |
2394 | so we always pass it as i3. We have not counted the notes in |
2395 | reg_n_deaths yet, so we need to do so now. */ | |
2396 | ||
230d793d | 2397 | if (newi2pat && new_i2_notes) |
1a26b032 RK |
2398 | { |
2399 | for (temp = new_i2_notes; temp; temp = XEXP (temp, 1)) | |
2400 | if (GET_CODE (XEXP (temp, 0)) == REG) | |
b1f21e0a | 2401 | REG_N_DEATHS (REGNO (XEXP (temp, 0)))++; |
1a26b032 RK |
2402 | |
2403 | distribute_notes (new_i2_notes, i2, i2, NULL_RTX, NULL_RTX, NULL_RTX); | |
2404 | } | |
2405 | ||
230d793d | 2406 | if (new_i3_notes) |
1a26b032 RK |
2407 | { |
2408 | for (temp = new_i3_notes; temp; temp = XEXP (temp, 1)) | |
2409 | if (GET_CODE (XEXP (temp, 0)) == REG) | |
b1f21e0a | 2410 | REG_N_DEATHS (REGNO (XEXP (temp, 0)))++; |
1a26b032 RK |
2411 | |
2412 | distribute_notes (new_i3_notes, i3, i3, NULL_RTX, NULL_RTX, NULL_RTX); | |
2413 | } | |
230d793d RS |
2414 | |
2415 | /* If I3DEST was used in I3SRC, it really died in I3. We may need to | |
e9a25f70 JL |
2416 | put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets |
2417 | I3DEST, the death must be somewhere before I2, not I3. If we passed I3 | |
2418 | in that case, it might delete I2. Similarly for I2 and I1. | |
1a26b032 RK |
2419 | Show an additional death due to the REG_DEAD note we make here. If |
2420 | we discard it in distribute_notes, we will decrement it again. */ | |
d0ab8cd3 | 2421 | |
230d793d | 2422 | if (i3dest_killed) |
1a26b032 RK |
2423 | { |
2424 | if (GET_CODE (i3dest_killed) == REG) | |
b1f21e0a | 2425 | REG_N_DEATHS (REGNO (i3dest_killed))++; |
1a26b032 | 2426 | |
e9a25f70 | 2427 | if (newi2pat && reg_set_p (i3dest_killed, newi2pat)) |
38a448ca RH |
2428 | distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed, |
2429 | NULL_RTX), | |
ff3467a9 | 2430 | NULL_RTX, i2, NULL_RTX, elim_i2, elim_i1); |
e9a25f70 | 2431 | else |
38a448ca RH |
2432 | distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed, |
2433 | NULL_RTX), | |
e9a25f70 | 2434 | NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, |
ff3467a9 | 2435 | elim_i2, elim_i1); |
1a26b032 | 2436 | } |
58c8c593 | 2437 | |
230d793d | 2438 | if (i2dest_in_i2src) |
58c8c593 | 2439 | { |
1a26b032 | 2440 | if (GET_CODE (i2dest) == REG) |
b1f21e0a | 2441 | REG_N_DEATHS (REGNO (i2dest))++; |
1a26b032 | 2442 | |
58c8c593 | 2443 | if (newi2pat && reg_set_p (i2dest, newi2pat)) |
38a448ca | 2444 | distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX), |
58c8c593 RK |
2445 | NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX); |
2446 | else | |
38a448ca | 2447 | distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX), |
58c8c593 RK |
2448 | NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, |
2449 | NULL_RTX, NULL_RTX); | |
2450 | } | |
2451 | ||
230d793d | 2452 | if (i1dest_in_i1src) |
58c8c593 | 2453 | { |
1a26b032 | 2454 | if (GET_CODE (i1dest) == REG) |
b1f21e0a | 2455 | REG_N_DEATHS (REGNO (i1dest))++; |
1a26b032 | 2456 | |
58c8c593 | 2457 | if (newi2pat && reg_set_p (i1dest, newi2pat)) |
38a448ca | 2458 | distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX), |
58c8c593 RK |
2459 | NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX); |
2460 | else | |
38a448ca | 2461 | distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX), |
58c8c593 RK |
2462 | NULL_RTX, i3, newi2pat ? i2 : NULL_RTX, |
2463 | NULL_RTX, NULL_RTX); | |
2464 | } | |
230d793d RS |
2465 | |
2466 | distribute_links (i3links); | |
2467 | distribute_links (i2links); | |
2468 | distribute_links (i1links); | |
2469 | ||
2470 | if (GET_CODE (i2dest) == REG) | |
2471 | { | |
d0ab8cd3 RK |
2472 | rtx link; |
2473 | rtx i2_insn = 0, i2_val = 0, set; | |
2474 | ||
2475 | /* The insn that used to set this register doesn't exist, and | |
2476 | this life of the register may not exist either. See if one of | |
2477 | I3's links points to an insn that sets I2DEST. If it does, | |
2478 | that is now the last known value for I2DEST. If we don't update | |
2479 | this and I2 set the register to a value that depended on its old | |
230d793d RS |
2480 | contents, we will get confused. If this insn is used, thing |
2481 | will be set correctly in combine_instructions. */ | |
d0ab8cd3 RK |
2482 | |
2483 | for (link = LOG_LINKS (i3); link; link = XEXP (link, 1)) | |
2484 | if ((set = single_set (XEXP (link, 0))) != 0 | |
2485 | && rtx_equal_p (i2dest, SET_DEST (set))) | |
2486 | i2_insn = XEXP (link, 0), i2_val = SET_SRC (set); | |
2487 | ||
2488 | record_value_for_reg (i2dest, i2_insn, i2_val); | |
230d793d RS |
2489 | |
2490 | /* If the reg formerly set in I2 died only once and that was in I3, | |
2491 | zero its use count so it won't make `reload' do any work. */ | |
538fe8cd ILT |
2492 | if (! added_sets_2 |
2493 | && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat)) | |
2494 | && ! i2dest_in_i2src) | |
230d793d RS |
2495 | { |
2496 | regno = REGNO (i2dest); | |
b1f21e0a MM |
2497 | REG_N_SETS (regno)--; |
2498 | if (REG_N_SETS (regno) == 0 | |
8e08106d | 2499 | && ! REGNO_REG_SET_P (basic_block_live_at_start[0], regno)) |
b1f21e0a | 2500 | REG_N_REFS (regno) = 0; |
230d793d RS |
2501 | } |
2502 | } | |
2503 | ||
2504 | if (i1 && GET_CODE (i1dest) == REG) | |
2505 | { | |
d0ab8cd3 RK |
2506 | rtx link; |
2507 | rtx i1_insn = 0, i1_val = 0, set; | |
2508 | ||
2509 | for (link = LOG_LINKS (i3); link; link = XEXP (link, 1)) | |
2510 | if ((set = single_set (XEXP (link, 0))) != 0 | |
2511 | && rtx_equal_p (i1dest, SET_DEST (set))) | |
2512 | i1_insn = XEXP (link, 0), i1_val = SET_SRC (set); | |
2513 | ||
2514 | record_value_for_reg (i1dest, i1_insn, i1_val); | |
2515 | ||
230d793d | 2516 | regno = REGNO (i1dest); |
5af91171 | 2517 | if (! added_sets_1 && ! i1dest_in_i1src) |
230d793d | 2518 | { |
b1f21e0a MM |
2519 | REG_N_SETS (regno)--; |
2520 | if (REG_N_SETS (regno) == 0 | |
8e08106d | 2521 | && ! REGNO_REG_SET_P (basic_block_live_at_start[0], regno)) |
b1f21e0a | 2522 | REG_N_REFS (regno) = 0; |
230d793d RS |
2523 | } |
2524 | } | |
2525 | ||
951553af | 2526 | /* Update reg_nonzero_bits et al for any changes that may have been made |
22609cbf RK |
2527 | to this insn. */ |
2528 | ||
951553af | 2529 | note_stores (newpat, set_nonzero_bits_and_sign_copies); |
22609cbf | 2530 | if (newi2pat) |
951553af | 2531 | note_stores (newi2pat, set_nonzero_bits_and_sign_copies); |
22609cbf | 2532 | |
a29ca9db RK |
2533 | /* If we added any (clobber (scratch)), add them to the max for a |
2534 | block. This is a very pessimistic calculation, since we might | |
2535 | have had them already and this might not be the worst block, but | |
2536 | it's not worth doing any better. */ | |
2537 | max_scratch += i3_scratches + i2_scratches + other_scratches; | |
2538 | ||
230d793d RS |
2539 | /* If I3 is now an unconditional jump, ensure that it has a |
2540 | BARRIER following it since it may have initially been a | |
381ee8af | 2541 | conditional jump. It may also be the last nonnote insn. */ |
230d793d RS |
2542 | |
2543 | if ((GET_CODE (newpat) == RETURN || simplejump_p (i3)) | |
381ee8af TW |
2544 | && ((temp = next_nonnote_insn (i3)) == NULL_RTX |
2545 | || GET_CODE (temp) != BARRIER)) | |
230d793d RS |
2546 | emit_barrier_after (i3); |
2547 | } | |
2548 | ||
2549 | combine_successes++; | |
2550 | ||
bcd49eb7 JW |
2551 | /* Clear this here, so that subsequent get_last_value calls are not |
2552 | affected. */ | |
2553 | subst_prev_insn = NULL_RTX; | |
2554 | ||
abe6e52f RK |
2555 | if (added_links_insn |
2556 | && (newi2pat == 0 || INSN_CUID (added_links_insn) < INSN_CUID (i2)) | |
2557 | && INSN_CUID (added_links_insn) < INSN_CUID (i3)) | |
2558 | return added_links_insn; | |
2559 | else | |
2560 | return newi2pat ? i2 : i3; | |
230d793d RS |
2561 | } |
2562 | \f | |
2563 | /* Undo all the modifications recorded in undobuf. */ | |
2564 | ||
2565 | static void | |
2566 | undo_all () | |
2567 | { | |
241cea85 RK |
2568 | struct undo *undo, *next; |
2569 | ||
2570 | for (undo = undobuf.undos; undo; undo = next) | |
7c046e4e | 2571 | { |
241cea85 RK |
2572 | next = undo->next; |
2573 | if (undo->is_int) | |
2574 | *undo->where.i = undo->old_contents.i; | |
7c046e4e | 2575 | else |
241cea85 RK |
2576 | *undo->where.r = undo->old_contents.r; |
2577 | ||
2578 | undo->next = undobuf.frees; | |
2579 | undobuf.frees = undo; | |
7c046e4e | 2580 | } |
230d793d RS |
2581 | |
2582 | obfree (undobuf.storage); | |
845fc875 | 2583 | undobuf.undos = undobuf.previous_undos = 0; |
bcd49eb7 JW |
2584 | |
2585 | /* Clear this here, so that subsequent get_last_value calls are not | |
2586 | affected. */ | |
2587 | subst_prev_insn = NULL_RTX; | |
230d793d RS |
2588 | } |
2589 | \f | |
2590 | /* Find the innermost point within the rtx at LOC, possibly LOC itself, | |
d0ab8cd3 RK |
2591 | where we have an arithmetic expression and return that point. LOC will |
2592 | be inside INSN. | |
230d793d RS |
2593 | |
2594 | try_combine will call this function to see if an insn can be split into | |
2595 | two insns. */ | |
2596 | ||
2597 | static rtx * | |
d0ab8cd3 | 2598 | find_split_point (loc, insn) |
230d793d | 2599 | rtx *loc; |
d0ab8cd3 | 2600 | rtx insn; |
230d793d RS |
2601 | { |
2602 | rtx x = *loc; | |
2603 | enum rtx_code code = GET_CODE (x); | |
2604 | rtx *split; | |
2605 | int len = 0, pos, unsignedp; | |
2606 | rtx inner; | |
2607 | ||
2608 | /* First special-case some codes. */ | |
2609 | switch (code) | |
2610 | { | |
2611 | case SUBREG: | |
2612 | #ifdef INSN_SCHEDULING | |
2613 | /* If we are making a paradoxical SUBREG invalid, it becomes a split | |
2614 | point. */ | |
2615 | if (GET_CODE (SUBREG_REG (x)) == MEM) | |
2616 | return loc; | |
2617 | #endif | |
d0ab8cd3 | 2618 | return find_split_point (&SUBREG_REG (x), insn); |
230d793d | 2619 | |
230d793d | 2620 | case MEM: |
916f14f1 | 2621 | #ifdef HAVE_lo_sum |
230d793d RS |
2622 | /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it |
2623 | using LO_SUM and HIGH. */ | |
2624 | if (GET_CODE (XEXP (x, 0)) == CONST | |
2625 | || GET_CODE (XEXP (x, 0)) == SYMBOL_REF) | |
2626 | { | |
2627 | SUBST (XEXP (x, 0), | |
2628 | gen_rtx_combine (LO_SUM, Pmode, | |
2629 | gen_rtx_combine (HIGH, Pmode, XEXP (x, 0)), | |
2630 | XEXP (x, 0))); | |
2631 | return &XEXP (XEXP (x, 0), 0); | |
2632 | } | |
230d793d RS |
2633 | #endif |
2634 | ||
916f14f1 RK |
2635 | /* If we have a PLUS whose second operand is a constant and the |
2636 | address is not valid, perhaps will can split it up using | |
2637 | the machine-specific way to split large constants. We use | |
ddd5a7c1 | 2638 | the first pseudo-reg (one of the virtual regs) as a placeholder; |
916f14f1 RK |
2639 | it will not remain in the result. */ |
2640 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
2641 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
2642 | && ! memory_address_p (GET_MODE (x), XEXP (x, 0))) | |
2643 | { | |
2644 | rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER]; | |
38a448ca | 2645 | rtx seq = split_insns (gen_rtx_SET (VOIDmode, reg, XEXP (x, 0)), |
916f14f1 RK |
2646 | subst_insn); |
2647 | ||
2648 | /* This should have produced two insns, each of which sets our | |
2649 | placeholder. If the source of the second is a valid address, | |
2650 | we can make put both sources together and make a split point | |
2651 | in the middle. */ | |
2652 | ||
2653 | if (seq && XVECLEN (seq, 0) == 2 | |
2654 | && GET_CODE (XVECEXP (seq, 0, 0)) == INSN | |
2655 | && GET_CODE (PATTERN (XVECEXP (seq, 0, 0))) == SET | |
2656 | && SET_DEST (PATTERN (XVECEXP (seq, 0, 0))) == reg | |
2657 | && ! reg_mentioned_p (reg, | |
2658 | SET_SRC (PATTERN (XVECEXP (seq, 0, 0)))) | |
2659 | && GET_CODE (XVECEXP (seq, 0, 1)) == INSN | |
2660 | && GET_CODE (PATTERN (XVECEXP (seq, 0, 1))) == SET | |
2661 | && SET_DEST (PATTERN (XVECEXP (seq, 0, 1))) == reg | |
2662 | && memory_address_p (GET_MODE (x), | |
2663 | SET_SRC (PATTERN (XVECEXP (seq, 0, 1))))) | |
2664 | { | |
2665 | rtx src1 = SET_SRC (PATTERN (XVECEXP (seq, 0, 0))); | |
2666 | rtx src2 = SET_SRC (PATTERN (XVECEXP (seq, 0, 1))); | |
2667 | ||
2668 | /* Replace the placeholder in SRC2 with SRC1. If we can | |
2669 | find where in SRC2 it was placed, that can become our | |
2670 | split point and we can replace this address with SRC2. | |
2671 | Just try two obvious places. */ | |
2672 | ||
2673 | src2 = replace_rtx (src2, reg, src1); | |
2674 | split = 0; | |
2675 | if (XEXP (src2, 0) == src1) | |
2676 | split = &XEXP (src2, 0); | |
2677 | else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e' | |
2678 | && XEXP (XEXP (src2, 0), 0) == src1) | |
2679 | split = &XEXP (XEXP (src2, 0), 0); | |
2680 | ||
2681 | if (split) | |
2682 | { | |
2683 | SUBST (XEXP (x, 0), src2); | |
2684 | return split; | |
2685 | } | |
2686 | } | |
1a26b032 RK |
2687 | |
2688 | /* If that didn't work, perhaps the first operand is complex and | |
2689 | needs to be computed separately, so make a split point there. | |
2690 | This will occur on machines that just support REG + CONST | |
2691 | and have a constant moved through some previous computation. */ | |
2692 | ||
2693 | else if (GET_RTX_CLASS (GET_CODE (XEXP (XEXP (x, 0), 0))) != 'o' | |
2694 | && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG | |
2695 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (XEXP (x, 0), 0)))) | |
2696 | == 'o'))) | |
2697 | return &XEXP (XEXP (x, 0), 0); | |
916f14f1 RK |
2698 | } |
2699 | break; | |
2700 | ||
230d793d RS |
2701 | case SET: |
2702 | #ifdef HAVE_cc0 | |
2703 | /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a | |
2704 | ZERO_EXTRACT, the most likely reason why this doesn't match is that | |
2705 | we need to put the operand into a register. So split at that | |
2706 | point. */ | |
2707 | ||
2708 | if (SET_DEST (x) == cc0_rtx | |
2709 | && GET_CODE (SET_SRC (x)) != COMPARE | |
2710 | && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT | |
2711 | && GET_RTX_CLASS (GET_CODE (SET_SRC (x))) != 'o' | |
2712 | && ! (GET_CODE (SET_SRC (x)) == SUBREG | |
2713 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (SET_SRC (x)))) == 'o')) | |
2714 | return &SET_SRC (x); | |
2715 | #endif | |
2716 | ||
2717 | /* See if we can split SET_SRC as it stands. */ | |
d0ab8cd3 | 2718 | split = find_split_point (&SET_SRC (x), insn); |
230d793d RS |
2719 | if (split && split != &SET_SRC (x)) |
2720 | return split; | |
2721 | ||
041d7180 JL |
2722 | /* See if we can split SET_DEST as it stands. */ |
2723 | split = find_split_point (&SET_DEST (x), insn); | |
2724 | if (split && split != &SET_DEST (x)) | |
2725 | return split; | |
2726 | ||
230d793d RS |
2727 | /* See if this is a bitfield assignment with everything constant. If |
2728 | so, this is an IOR of an AND, so split it into that. */ | |
2729 | if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT | |
2730 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))) | |
5f4f0e22 | 2731 | <= HOST_BITS_PER_WIDE_INT) |
230d793d RS |
2732 | && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT |
2733 | && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT | |
2734 | && GET_CODE (SET_SRC (x)) == CONST_INT | |
2735 | && ((INTVAL (XEXP (SET_DEST (x), 1)) | |
2736 | + INTVAL (XEXP (SET_DEST (x), 2))) | |
2737 | <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))) | |
2738 | && ! side_effects_p (XEXP (SET_DEST (x), 0))) | |
2739 | { | |
2740 | int pos = INTVAL (XEXP (SET_DEST (x), 2)); | |
2741 | int len = INTVAL (XEXP (SET_DEST (x), 1)); | |
2742 | int src = INTVAL (SET_SRC (x)); | |
2743 | rtx dest = XEXP (SET_DEST (x), 0); | |
2744 | enum machine_mode mode = GET_MODE (dest); | |
5f4f0e22 | 2745 | unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1; |
230d793d | 2746 | |
f76b9db2 ILT |
2747 | if (BITS_BIG_ENDIAN) |
2748 | pos = GET_MODE_BITSIZE (mode) - len - pos; | |
230d793d RS |
2749 | |
2750 | if (src == mask) | |
2751 | SUBST (SET_SRC (x), | |
5f4f0e22 | 2752 | gen_binary (IOR, mode, dest, GEN_INT (src << pos))); |
230d793d RS |
2753 | else |
2754 | SUBST (SET_SRC (x), | |
2755 | gen_binary (IOR, mode, | |
2756 | gen_binary (AND, mode, dest, | |
5f4f0e22 CH |
2757 | GEN_INT (~ (mask << pos) |
2758 | & GET_MODE_MASK (mode))), | |
2759 | GEN_INT (src << pos))); | |
230d793d RS |
2760 | |
2761 | SUBST (SET_DEST (x), dest); | |
2762 | ||
d0ab8cd3 | 2763 | split = find_split_point (&SET_SRC (x), insn); |
230d793d RS |
2764 | if (split && split != &SET_SRC (x)) |
2765 | return split; | |
2766 | } | |
2767 | ||
2768 | /* Otherwise, see if this is an operation that we can split into two. | |
2769 | If so, try to split that. */ | |
2770 | code = GET_CODE (SET_SRC (x)); | |
2771 | ||
2772 | switch (code) | |
2773 | { | |
d0ab8cd3 RK |
2774 | case AND: |
2775 | /* If we are AND'ing with a large constant that is only a single | |
2776 | bit and the result is only being used in a context where we | |
2777 | need to know if it is zero or non-zero, replace it with a bit | |
2778 | extraction. This will avoid the large constant, which might | |
2779 | have taken more than one insn to make. If the constant were | |
2780 | not a valid argument to the AND but took only one insn to make, | |
2781 | this is no worse, but if it took more than one insn, it will | |
2782 | be better. */ | |
2783 | ||
2784 | if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT | |
2785 | && GET_CODE (XEXP (SET_SRC (x), 0)) == REG | |
2786 | && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7 | |
2787 | && GET_CODE (SET_DEST (x)) == REG | |
2788 | && (split = find_single_use (SET_DEST (x), insn, NULL_PTR)) != 0 | |
2789 | && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE) | |
2790 | && XEXP (*split, 0) == SET_DEST (x) | |
2791 | && XEXP (*split, 1) == const0_rtx) | |
2792 | { | |
76184def DE |
2793 | rtx extraction = make_extraction (GET_MODE (SET_DEST (x)), |
2794 | XEXP (SET_SRC (x), 0), | |
2795 | pos, NULL_RTX, 1, 1, 0, 0); | |
2796 | if (extraction != 0) | |
2797 | { | |
2798 | SUBST (SET_SRC (x), extraction); | |
2799 | return find_split_point (loc, insn); | |
2800 | } | |
d0ab8cd3 RK |
2801 | } |
2802 | break; | |
2803 | ||
1a6ec070 RK |
2804 | case NE: |
2805 | /* if STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X | |
2806 | is known to be on, this can be converted into a NEG of a shift. */ | |
2807 | if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx | |
2808 | && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0)) | |
4eb2cb10 | 2809 | && 1 <= (pos = exact_log2 |
1a6ec070 RK |
2810 | (nonzero_bits (XEXP (SET_SRC (x), 0), |
2811 | GET_MODE (XEXP (SET_SRC (x), 0)))))) | |
2812 | { | |
2813 | enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0)); | |
2814 | ||
2815 | SUBST (SET_SRC (x), | |
2816 | gen_rtx_combine (NEG, mode, | |
2817 | gen_rtx_combine (LSHIFTRT, mode, | |
2818 | XEXP (SET_SRC (x), 0), | |
4eb2cb10 | 2819 | GEN_INT (pos)))); |
1a6ec070 RK |
2820 | |
2821 | split = find_split_point (&SET_SRC (x), insn); | |
2822 | if (split && split != &SET_SRC (x)) | |
2823 | return split; | |
2824 | } | |
2825 | break; | |
2826 | ||
230d793d RS |
2827 | case SIGN_EXTEND: |
2828 | inner = XEXP (SET_SRC (x), 0); | |
101c1a3d JL |
2829 | |
2830 | /* We can't optimize if either mode is a partial integer | |
2831 | mode as we don't know how many bits are significant | |
2832 | in those modes. */ | |
2833 | if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT | |
2834 | || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT) | |
2835 | break; | |
2836 | ||
230d793d RS |
2837 | pos = 0; |
2838 | len = GET_MODE_BITSIZE (GET_MODE (inner)); | |
2839 | unsignedp = 0; | |
2840 | break; | |
2841 | ||
2842 | case SIGN_EXTRACT: | |
2843 | case ZERO_EXTRACT: | |
2844 | if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT | |
2845 | && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT) | |
2846 | { | |
2847 | inner = XEXP (SET_SRC (x), 0); | |
2848 | len = INTVAL (XEXP (SET_SRC (x), 1)); | |
2849 | pos = INTVAL (XEXP (SET_SRC (x), 2)); | |
2850 | ||
f76b9db2 ILT |
2851 | if (BITS_BIG_ENDIAN) |
2852 | pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos; | |
230d793d RS |
2853 | unsignedp = (code == ZERO_EXTRACT); |
2854 | } | |
2855 | break; | |
e9a25f70 JL |
2856 | |
2857 | default: | |
2858 | break; | |
230d793d RS |
2859 | } |
2860 | ||
2861 | if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner))) | |
2862 | { | |
2863 | enum machine_mode mode = GET_MODE (SET_SRC (x)); | |
2864 | ||
d0ab8cd3 RK |
2865 | /* For unsigned, we have a choice of a shift followed by an |
2866 | AND or two shifts. Use two shifts for field sizes where the | |
2867 | constant might be too large. We assume here that we can | |
2868 | always at least get 8-bit constants in an AND insn, which is | |
2869 | true for every current RISC. */ | |
2870 | ||
2871 | if (unsignedp && len <= 8) | |
230d793d RS |
2872 | { |
2873 | SUBST (SET_SRC (x), | |
2874 | gen_rtx_combine | |
2875 | (AND, mode, | |
2876 | gen_rtx_combine (LSHIFTRT, mode, | |
2877 | gen_lowpart_for_combine (mode, inner), | |
5f4f0e22 CH |
2878 | GEN_INT (pos)), |
2879 | GEN_INT (((HOST_WIDE_INT) 1 << len) - 1))); | |
230d793d | 2880 | |
d0ab8cd3 | 2881 | split = find_split_point (&SET_SRC (x), insn); |
230d793d RS |
2882 | if (split && split != &SET_SRC (x)) |
2883 | return split; | |
2884 | } | |
2885 | else | |
2886 | { | |
2887 | SUBST (SET_SRC (x), | |
2888 | gen_rtx_combine | |
d0ab8cd3 | 2889 | (unsignedp ? LSHIFTRT : ASHIFTRT, mode, |
230d793d RS |
2890 | gen_rtx_combine (ASHIFT, mode, |
2891 | gen_lowpart_for_combine (mode, inner), | |
5f4f0e22 CH |
2892 | GEN_INT (GET_MODE_BITSIZE (mode) |
2893 | - len - pos)), | |
2894 | GEN_INT (GET_MODE_BITSIZE (mode) - len))); | |
230d793d | 2895 | |
d0ab8cd3 | 2896 | split = find_split_point (&SET_SRC (x), insn); |
230d793d RS |
2897 | if (split && split != &SET_SRC (x)) |
2898 | return split; | |
2899 | } | |
2900 | } | |
2901 | ||
2902 | /* See if this is a simple operation with a constant as the second | |
2903 | operand. It might be that this constant is out of range and hence | |
2904 | could be used as a split point. */ | |
2905 | if ((GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '2' | |
2906 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == 'c' | |
2907 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '<') | |
2908 | && CONSTANT_P (XEXP (SET_SRC (x), 1)) | |
2909 | && (GET_RTX_CLASS (GET_CODE (XEXP (SET_SRC (x), 0))) == 'o' | |
2910 | || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG | |
2911 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (SET_SRC (x), 0)))) | |
2912 | == 'o')))) | |
2913 | return &XEXP (SET_SRC (x), 1); | |
2914 | ||
2915 | /* Finally, see if this is a simple operation with its first operand | |
2916 | not in a register. The operation might require this operand in a | |
2917 | register, so return it as a split point. We can always do this | |
2918 | because if the first operand were another operation, we would have | |
2919 | already found it as a split point. */ | |
2920 | if ((GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '2' | |
2921 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == 'c' | |
2922 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '<' | |
2923 | || GET_RTX_CLASS (GET_CODE (SET_SRC (x))) == '1') | |
2924 | && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode)) | |
2925 | return &XEXP (SET_SRC (x), 0); | |
2926 | ||
2927 | return 0; | |
2928 | ||
2929 | case AND: | |
2930 | case IOR: | |
2931 | /* We write NOR as (and (not A) (not B)), but if we don't have a NOR, | |
2932 | it is better to write this as (not (ior A B)) so we can split it. | |
2933 | Similarly for IOR. */ | |
2934 | if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT) | |
2935 | { | |
2936 | SUBST (*loc, | |
2937 | gen_rtx_combine (NOT, GET_MODE (x), | |
2938 | gen_rtx_combine (code == IOR ? AND : IOR, | |
2939 | GET_MODE (x), | |
2940 | XEXP (XEXP (x, 0), 0), | |
2941 | XEXP (XEXP (x, 1), 0)))); | |
d0ab8cd3 | 2942 | return find_split_point (loc, insn); |
230d793d RS |
2943 | } |
2944 | ||
2945 | /* Many RISC machines have a large set of logical insns. If the | |
2946 | second operand is a NOT, put it first so we will try to split the | |
2947 | other operand first. */ | |
2948 | if (GET_CODE (XEXP (x, 1)) == NOT) | |
2949 | { | |
2950 | rtx tem = XEXP (x, 0); | |
2951 | SUBST (XEXP (x, 0), XEXP (x, 1)); | |
2952 | SUBST (XEXP (x, 1), tem); | |
2953 | } | |
2954 | break; | |
e9a25f70 JL |
2955 | |
2956 | default: | |
2957 | break; | |
230d793d RS |
2958 | } |
2959 | ||
2960 | /* Otherwise, select our actions depending on our rtx class. */ | |
2961 | switch (GET_RTX_CLASS (code)) | |
2962 | { | |
2963 | case 'b': /* This is ZERO_EXTRACT and SIGN_EXTRACT. */ | |
2964 | case '3': | |
d0ab8cd3 | 2965 | split = find_split_point (&XEXP (x, 2), insn); |
230d793d RS |
2966 | if (split) |
2967 | return split; | |
0f41302f | 2968 | /* ... fall through ... */ |
230d793d RS |
2969 | case '2': |
2970 | case 'c': | |
2971 | case '<': | |
d0ab8cd3 | 2972 | split = find_split_point (&XEXP (x, 1), insn); |
230d793d RS |
2973 | if (split) |
2974 | return split; | |
0f41302f | 2975 | /* ... fall through ... */ |
230d793d RS |
2976 | case '1': |
2977 | /* Some machines have (and (shift ...) ...) insns. If X is not | |
2978 | an AND, but XEXP (X, 0) is, use it as our split point. */ | |
2979 | if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND) | |
2980 | return &XEXP (x, 0); | |
2981 | ||
d0ab8cd3 | 2982 | split = find_split_point (&XEXP (x, 0), insn); |
230d793d RS |
2983 | if (split) |
2984 | return split; | |
2985 | return loc; | |
2986 | } | |
2987 | ||
2988 | /* Otherwise, we don't have a split point. */ | |
2989 | return 0; | |
2990 | } | |
2991 | \f | |
2992 | /* Throughout X, replace FROM with TO, and return the result. | |
2993 | The result is TO if X is FROM; | |
2994 | otherwise the result is X, but its contents may have been modified. | |
2995 | If they were modified, a record was made in undobuf so that | |
2996 | undo_all will (among other things) return X to its original state. | |
2997 | ||
2998 | If the number of changes necessary is too much to record to undo, | |
2999 | the excess changes are not made, so the result is invalid. | |
3000 | The changes already made can still be undone. | |
3001 | undobuf.num_undo is incremented for such changes, so by testing that | |
3002 | the caller can tell whether the result is valid. | |
3003 | ||
3004 | `n_occurrences' is incremented each time FROM is replaced. | |
3005 | ||
3006 | IN_DEST is non-zero if we are processing the SET_DEST of a SET. | |
3007 | ||
5089e22e | 3008 | UNIQUE_COPY is non-zero if each substitution must be unique. We do this |
230d793d RS |
3009 | by copying if `n_occurrences' is non-zero. */ |
3010 | ||
3011 | static rtx | |
3012 | subst (x, from, to, in_dest, unique_copy) | |
3013 | register rtx x, from, to; | |
3014 | int in_dest; | |
3015 | int unique_copy; | |
3016 | { | |
f24ad0e4 | 3017 | register enum rtx_code code = GET_CODE (x); |
230d793d | 3018 | enum machine_mode op0_mode = VOIDmode; |
8079805d RK |
3019 | register char *fmt; |
3020 | register int len, i; | |
3021 | rtx new; | |
230d793d RS |
3022 | |
3023 | /* Two expressions are equal if they are identical copies of a shared | |
3024 | RTX or if they are both registers with the same register number | |
3025 | and mode. */ | |
3026 | ||
3027 | #define COMBINE_RTX_EQUAL_P(X,Y) \ | |
3028 | ((X) == (Y) \ | |
3029 | || (GET_CODE (X) == REG && GET_CODE (Y) == REG \ | |
3030 | && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y))) | |
3031 | ||
3032 | if (! in_dest && COMBINE_RTX_EQUAL_P (x, from)) | |
3033 | { | |
3034 | n_occurrences++; | |
3035 | return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to); | |
3036 | } | |
3037 | ||
3038 | /* If X and FROM are the same register but different modes, they will | |
3039 | not have been seen as equal above. However, flow.c will make a | |
3040 | LOG_LINKS entry for that case. If we do nothing, we will try to | |
3041 | rerecognize our original insn and, when it succeeds, we will | |
3042 | delete the feeding insn, which is incorrect. | |
3043 | ||
3044 | So force this insn not to match in this (rare) case. */ | |
3045 | if (! in_dest && code == REG && GET_CODE (from) == REG | |
3046 | && REGNO (x) == REGNO (from)) | |
38a448ca | 3047 | return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
230d793d RS |
3048 | |
3049 | /* If this is an object, we are done unless it is a MEM or LO_SUM, both | |
3050 | of which may contain things that can be combined. */ | |
3051 | if (code != MEM && code != LO_SUM && GET_RTX_CLASS (code) == 'o') | |
3052 | return x; | |
3053 | ||
3054 | /* It is possible to have a subexpression appear twice in the insn. | |
3055 | Suppose that FROM is a register that appears within TO. | |
3056 | Then, after that subexpression has been scanned once by `subst', | |
3057 | the second time it is scanned, TO may be found. If we were | |
3058 | to scan TO here, we would find FROM within it and create a | |
3059 | self-referent rtl structure which is completely wrong. */ | |
3060 | if (COMBINE_RTX_EQUAL_P (x, to)) | |
3061 | return to; | |
3062 | ||
3063 | len = GET_RTX_LENGTH (code); | |
3064 | fmt = GET_RTX_FORMAT (code); | |
3065 | ||
3066 | /* We don't need to process a SET_DEST that is a register, CC0, or PC, so | |
3067 | set up to skip this common case. All other cases where we want to | |
3068 | suppress replacing something inside a SET_SRC are handled via the | |
3069 | IN_DEST operand. */ | |
3070 | if (code == SET | |
3071 | && (GET_CODE (SET_DEST (x)) == REG | |
3072 | || GET_CODE (SET_DEST (x)) == CC0 | |
3073 | || GET_CODE (SET_DEST (x)) == PC)) | |
3074 | fmt = "ie"; | |
3075 | ||
0f41302f MS |
3076 | /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a |
3077 | constant. */ | |
230d793d RS |
3078 | if (fmt[0] == 'e') |
3079 | op0_mode = GET_MODE (XEXP (x, 0)); | |
3080 | ||
3081 | for (i = 0; i < len; i++) | |
3082 | { | |
3083 | if (fmt[i] == 'E') | |
3084 | { | |
3085 | register int j; | |
3086 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3087 | { | |
230d793d RS |
3088 | if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from)) |
3089 | { | |
3090 | new = (unique_copy && n_occurrences ? copy_rtx (to) : to); | |
3091 | n_occurrences++; | |
3092 | } | |
3093 | else | |
3094 | { | |
3095 | new = subst (XVECEXP (x, i, j), from, to, 0, unique_copy); | |
3096 | ||
3097 | /* If this substitution failed, this whole thing fails. */ | |
3098 | if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx) | |
3099 | return new; | |
3100 | } | |
3101 | ||
3102 | SUBST (XVECEXP (x, i, j), new); | |
3103 | } | |
3104 | } | |
3105 | else if (fmt[i] == 'e') | |
3106 | { | |
230d793d RS |
3107 | if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from)) |
3108 | { | |
42301240 RK |
3109 | /* In general, don't install a subreg involving two modes not |
3110 | tieable. It can worsen register allocation, and can even | |
3111 | make invalid reload insns, since the reg inside may need to | |
3112 | be copied from in the outside mode, and that may be invalid | |
3113 | if it is an fp reg copied in integer mode. | |
3114 | ||
3115 | We allow two exceptions to this: It is valid if it is inside | |
3116 | another SUBREG and the mode of that SUBREG and the mode of | |
3117 | the inside of TO is tieable and it is valid if X is a SET | |
3118 | that copies FROM to CC0. */ | |
3119 | if (GET_CODE (to) == SUBREG | |
3120 | && ! MODES_TIEABLE_P (GET_MODE (to), | |
3121 | GET_MODE (SUBREG_REG (to))) | |
3122 | && ! (code == SUBREG | |
8079805d RK |
3123 | && MODES_TIEABLE_P (GET_MODE (x), |
3124 | GET_MODE (SUBREG_REG (to)))) | |
42301240 RK |
3125 | #ifdef HAVE_cc0 |
3126 | && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx) | |
3127 | #endif | |
3128 | ) | |
38a448ca | 3129 | return gen_rtx_CLOBBER (VOIDmode, const0_rtx); |
42301240 | 3130 | |
230d793d RS |
3131 | new = (unique_copy && n_occurrences ? copy_rtx (to) : to); |
3132 | n_occurrences++; | |
3133 | } | |
3134 | else | |
3135 | /* If we are in a SET_DEST, suppress most cases unless we | |
3136 | have gone inside a MEM, in which case we want to | |
3137 | simplify the address. We assume here that things that | |
3138 | are actually part of the destination have their inner | |
3139 | parts in the first expression. This is true for SUBREG, | |
3140 | STRICT_LOW_PART, and ZERO_EXTRACT, which are the only | |
3141 | things aside from REG and MEM that should appear in a | |
3142 | SET_DEST. */ | |
3143 | new = subst (XEXP (x, i), from, to, | |
3144 | (((in_dest | |
3145 | && (code == SUBREG || code == STRICT_LOW_PART | |
3146 | || code == ZERO_EXTRACT)) | |
3147 | || code == SET) | |
3148 | && i == 0), unique_copy); | |
3149 | ||
3150 | /* If we found that we will have to reject this combination, | |
3151 | indicate that by returning the CLOBBER ourselves, rather than | |
3152 | an expression containing it. This will speed things up as | |
3153 | well as prevent accidents where two CLOBBERs are considered | |
3154 | to be equal, thus producing an incorrect simplification. */ | |
3155 | ||
3156 | if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx) | |
3157 | return new; | |
3158 | ||
3159 | SUBST (XEXP (x, i), new); | |
3160 | } | |
3161 | } | |
3162 | ||
8079805d RK |
3163 | /* Try to simplify X. If the simplification changed the code, it is likely |
3164 | that further simplification will help, so loop, but limit the number | |
3165 | of repetitions that will be performed. */ | |
3166 | ||
3167 | for (i = 0; i < 4; i++) | |
3168 | { | |
3169 | /* If X is sufficiently simple, don't bother trying to do anything | |
3170 | with it. */ | |
3171 | if (code != CONST_INT && code != REG && code != CLOBBER) | |
3172 | x = simplify_rtx (x, op0_mode, i == 3, in_dest); | |
d0ab8cd3 | 3173 | |
8079805d RK |
3174 | if (GET_CODE (x) == code) |
3175 | break; | |
d0ab8cd3 | 3176 | |
8079805d | 3177 | code = GET_CODE (x); |
eeb43d32 | 3178 | |
8079805d RK |
3179 | /* We no longer know the original mode of operand 0 since we |
3180 | have changed the form of X) */ | |
3181 | op0_mode = VOIDmode; | |
3182 | } | |
eeb43d32 | 3183 | |
8079805d RK |
3184 | return x; |
3185 | } | |
3186 | \f | |
3187 | /* Simplify X, a piece of RTL. We just operate on the expression at the | |
3188 | outer level; call `subst' to simplify recursively. Return the new | |
3189 | expression. | |
3190 | ||
3191 | OP0_MODE is the original mode of XEXP (x, 0); LAST is nonzero if this | |
3192 | will be the iteration even if an expression with a code different from | |
3193 | X is returned; IN_DEST is nonzero if we are inside a SET_DEST. */ | |
eeb43d32 | 3194 | |
8079805d RK |
3195 | static rtx |
3196 | simplify_rtx (x, op0_mode, last, in_dest) | |
3197 | rtx x; | |
3198 | enum machine_mode op0_mode; | |
3199 | int last; | |
3200 | int in_dest; | |
3201 | { | |
3202 | enum rtx_code code = GET_CODE (x); | |
3203 | enum machine_mode mode = GET_MODE (x); | |
3204 | rtx temp; | |
3205 | int i; | |
d0ab8cd3 | 3206 | |
230d793d RS |
3207 | /* If this is a commutative operation, put a constant last and a complex |
3208 | expression first. We don't need to do this for comparisons here. */ | |
3209 | if (GET_RTX_CLASS (code) == 'c' | |
3210 | && ((CONSTANT_P (XEXP (x, 0)) && GET_CODE (XEXP (x, 1)) != CONST_INT) | |
3211 | || (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == 'o' | |
3212 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) != 'o') | |
3213 | || (GET_CODE (XEXP (x, 0)) == SUBREG | |
3214 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 0)))) == 'o' | |
3215 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) != 'o'))) | |
3216 | { | |
3217 | temp = XEXP (x, 0); | |
3218 | SUBST (XEXP (x, 0), XEXP (x, 1)); | |
3219 | SUBST (XEXP (x, 1), temp); | |
3220 | } | |
3221 | ||
22609cbf RK |
3222 | /* If this is a PLUS, MINUS, or MULT, and the first operand is the |
3223 | sign extension of a PLUS with a constant, reverse the order of the sign | |
3224 | extension and the addition. Note that this not the same as the original | |
3225 | code, but overflow is undefined for signed values. Also note that the | |
3226 | PLUS will have been partially moved "inside" the sign-extension, so that | |
3227 | the first operand of X will really look like: | |
3228 | (ashiftrt (plus (ashift A C4) C5) C4). | |
3229 | We convert this to | |
3230 | (plus (ashiftrt (ashift A C4) C2) C4) | |
3231 | and replace the first operand of X with that expression. Later parts | |
3232 | of this function may simplify the expression further. | |
3233 | ||
3234 | For example, if we start with (mult (sign_extend (plus A C1)) C2), | |
3235 | we swap the SIGN_EXTEND and PLUS. Later code will apply the | |
3236 | distributive law to produce (plus (mult (sign_extend X) C1) C3). | |
3237 | ||
3238 | We do this to simplify address expressions. */ | |
3239 | ||
3240 | if ((code == PLUS || code == MINUS || code == MULT) | |
3241 | && GET_CODE (XEXP (x, 0)) == ASHIFTRT | |
3242 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == PLUS | |
3243 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ASHIFT | |
3244 | && GET_CODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1)) == CONST_INT | |
3245 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3246 | && XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1) == XEXP (XEXP (x, 0), 1) | |
3247 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
3248 | && (temp = simplify_binary_operation (ASHIFTRT, mode, | |
3249 | XEXP (XEXP (XEXP (x, 0), 0), 1), | |
3250 | XEXP (XEXP (x, 0), 1))) != 0) | |
3251 | { | |
3252 | rtx new | |
3253 | = simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
3254 | XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), | |
3255 | INTVAL (XEXP (XEXP (x, 0), 1))); | |
3256 | ||
3257 | new = simplify_shift_const (NULL_RTX, ASHIFTRT, mode, new, | |
3258 | INTVAL (XEXP (XEXP (x, 0), 1))); | |
3259 | ||
3260 | SUBST (XEXP (x, 0), gen_binary (PLUS, mode, new, temp)); | |
3261 | } | |
3262 | ||
d0ab8cd3 RK |
3263 | /* If this is a simple operation applied to an IF_THEN_ELSE, try |
3264 | applying it to the arms of the IF_THEN_ELSE. This often simplifies | |
abe6e52f RK |
3265 | things. Check for cases where both arms are testing the same |
3266 | condition. | |
3267 | ||
3268 | Don't do anything if all operands are very simple. */ | |
3269 | ||
3270 | if (((GET_RTX_CLASS (code) == '2' || GET_RTX_CLASS (code) == 'c' | |
3271 | || GET_RTX_CLASS (code) == '<') | |
3272 | && ((GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) != 'o' | |
3273 | && ! (GET_CODE (XEXP (x, 0)) == SUBREG | |
3274 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 0)))) | |
3275 | == 'o'))) | |
3276 | || (GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) != 'o' | |
3277 | && ! (GET_CODE (XEXP (x, 1)) == SUBREG | |
3278 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 1)))) | |
3279 | == 'o'))))) | |
3280 | || (GET_RTX_CLASS (code) == '1' | |
3281 | && ((GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) != 'o' | |
3282 | && ! (GET_CODE (XEXP (x, 0)) == SUBREG | |
3283 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 0)))) | |
3284 | == 'o')))))) | |
d0ab8cd3 | 3285 | { |
abe6e52f RK |
3286 | rtx cond, true, false; |
3287 | ||
3288 | cond = if_then_else_cond (x, &true, &false); | |
0802d516 RK |
3289 | if (cond != 0 |
3290 | /* If everything is a comparison, what we have is highly unlikely | |
3291 | to be simpler, so don't use it. */ | |
3292 | && ! (GET_RTX_CLASS (code) == '<' | |
3293 | && (GET_RTX_CLASS (GET_CODE (true)) == '<' | |
3294 | || GET_RTX_CLASS (GET_CODE (false)) == '<'))) | |
abe6e52f RK |
3295 | { |
3296 | rtx cop1 = const0_rtx; | |
3297 | enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1); | |
3298 | ||
15448afc RK |
3299 | if (cond_code == NE && GET_RTX_CLASS (GET_CODE (cond)) == '<') |
3300 | return x; | |
3301 | ||
9210df58 RK |
3302 | /* Simplify the alternative arms; this may collapse the true and |
3303 | false arms to store-flag values. */ | |
3304 | true = subst (true, pc_rtx, pc_rtx, 0, 0); | |
3305 | false = subst (false, pc_rtx, pc_rtx, 0, 0); | |
3306 | ||
3307 | /* Restarting if we generate a store-flag expression will cause | |
3308 | us to loop. Just drop through in this case. */ | |
3309 | ||
abe6e52f RK |
3310 | /* If the result values are STORE_FLAG_VALUE and zero, we can |
3311 | just make the comparison operation. */ | |
3312 | if (true == const_true_rtx && false == const0_rtx) | |
3313 | x = gen_binary (cond_code, mode, cond, cop1); | |
3314 | else if (true == const0_rtx && false == const_true_rtx) | |
3315 | x = gen_binary (reverse_condition (cond_code), mode, cond, cop1); | |
3316 | ||
3317 | /* Likewise, we can make the negate of a comparison operation | |
3318 | if the result values are - STORE_FLAG_VALUE and zero. */ | |
3319 | else if (GET_CODE (true) == CONST_INT | |
3320 | && INTVAL (true) == - STORE_FLAG_VALUE | |
3321 | && false == const0_rtx) | |
0c1c8ea6 | 3322 | x = gen_unary (NEG, mode, mode, |
abe6e52f RK |
3323 | gen_binary (cond_code, mode, cond, cop1)); |
3324 | else if (GET_CODE (false) == CONST_INT | |
3325 | && INTVAL (false) == - STORE_FLAG_VALUE | |
3326 | && true == const0_rtx) | |
0c1c8ea6 | 3327 | x = gen_unary (NEG, mode, mode, |
abe6e52f RK |
3328 | gen_binary (reverse_condition (cond_code), |
3329 | mode, cond, cop1)); | |
3330 | else | |
38a448ca RH |
3331 | return gen_rtx_IF_THEN_ELSE (mode, |
3332 | gen_binary (cond_code, VOIDmode, | |
3333 | cond, cop1), | |
3334 | true, false); | |
5109d49f | 3335 | |
9210df58 RK |
3336 | code = GET_CODE (x); |
3337 | op0_mode = VOIDmode; | |
abe6e52f | 3338 | } |
d0ab8cd3 RK |
3339 | } |
3340 | ||
230d793d RS |
3341 | /* Try to fold this expression in case we have constants that weren't |
3342 | present before. */ | |
3343 | temp = 0; | |
3344 | switch (GET_RTX_CLASS (code)) | |
3345 | { | |
3346 | case '1': | |
3347 | temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode); | |
3348 | break; | |
3349 | case '<': | |
3350 | temp = simplify_relational_operation (code, op0_mode, | |
3351 | XEXP (x, 0), XEXP (x, 1)); | |
77fa0940 RK |
3352 | #ifdef FLOAT_STORE_FLAG_VALUE |
3353 | if (temp != 0 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
3354 | temp = ((temp == const0_rtx) ? CONST0_RTX (GET_MODE (x)) | |
3355 | : immed_real_const_1 (FLOAT_STORE_FLAG_VALUE, GET_MODE (x))); | |
3356 | #endif | |
230d793d RS |
3357 | break; |
3358 | case 'c': | |
3359 | case '2': | |
3360 | temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1)); | |
3361 | break; | |
3362 | case 'b': | |
3363 | case '3': | |
3364 | temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0), | |
3365 | XEXP (x, 1), XEXP (x, 2)); | |
3366 | break; | |
3367 | } | |
3368 | ||
3369 | if (temp) | |
d0ab8cd3 | 3370 | x = temp, code = GET_CODE (temp); |
230d793d | 3371 | |
230d793d | 3372 | /* First see if we can apply the inverse distributive law. */ |
224eeff2 RK |
3373 | if (code == PLUS || code == MINUS |
3374 | || code == AND || code == IOR || code == XOR) | |
230d793d RS |
3375 | { |
3376 | x = apply_distributive_law (x); | |
3377 | code = GET_CODE (x); | |
3378 | } | |
3379 | ||
3380 | /* If CODE is an associative operation not otherwise handled, see if we | |
3381 | can associate some operands. This can win if they are constants or | |
3382 | if they are logically related (i.e. (a & b) & a. */ | |
3383 | if ((code == PLUS || code == MINUS | |
3384 | || code == MULT || code == AND || code == IOR || code == XOR | |
3385 | || code == DIV || code == UDIV | |
3386 | || code == SMAX || code == SMIN || code == UMAX || code == UMIN) | |
3ad2180a | 3387 | && INTEGRAL_MODE_P (mode)) |
230d793d RS |
3388 | { |
3389 | if (GET_CODE (XEXP (x, 0)) == code) | |
3390 | { | |
3391 | rtx other = XEXP (XEXP (x, 0), 0); | |
3392 | rtx inner_op0 = XEXP (XEXP (x, 0), 1); | |
3393 | rtx inner_op1 = XEXP (x, 1); | |
3394 | rtx inner; | |
3395 | ||
3396 | /* Make sure we pass the constant operand if any as the second | |
3397 | one if this is a commutative operation. */ | |
3398 | if (CONSTANT_P (inner_op0) && GET_RTX_CLASS (code) == 'c') | |
3399 | { | |
3400 | rtx tem = inner_op0; | |
3401 | inner_op0 = inner_op1; | |
3402 | inner_op1 = tem; | |
3403 | } | |
3404 | inner = simplify_binary_operation (code == MINUS ? PLUS | |
3405 | : code == DIV ? MULT | |
3406 | : code == UDIV ? MULT | |
3407 | : code, | |
3408 | mode, inner_op0, inner_op1); | |
3409 | ||
3410 | /* For commutative operations, try the other pair if that one | |
3411 | didn't simplify. */ | |
3412 | if (inner == 0 && GET_RTX_CLASS (code) == 'c') | |
3413 | { | |
3414 | other = XEXP (XEXP (x, 0), 1); | |
3415 | inner = simplify_binary_operation (code, mode, | |
3416 | XEXP (XEXP (x, 0), 0), | |
3417 | XEXP (x, 1)); | |
3418 | } | |
3419 | ||
3420 | if (inner) | |
8079805d | 3421 | return gen_binary (code, mode, other, inner); |
230d793d RS |
3422 | } |
3423 | } | |
3424 | ||
3425 | /* A little bit of algebraic simplification here. */ | |
3426 | switch (code) | |
3427 | { | |
3428 | case MEM: | |
3429 | /* Ensure that our address has any ASHIFTs converted to MULT in case | |
3430 | address-recognizing predicates are called later. */ | |
3431 | temp = make_compound_operation (XEXP (x, 0), MEM); | |
3432 | SUBST (XEXP (x, 0), temp); | |
3433 | break; | |
3434 | ||
3435 | case SUBREG: | |
3436 | /* (subreg:A (mem:B X) N) becomes a modified MEM unless the SUBREG | |
3437 | is paradoxical. If we can't do that safely, then it becomes | |
3438 | something nonsensical so that this combination won't take place. */ | |
3439 | ||
3440 | if (GET_CODE (SUBREG_REG (x)) == MEM | |
3441 | && (GET_MODE_SIZE (mode) | |
3442 | <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))) | |
3443 | { | |
3444 | rtx inner = SUBREG_REG (x); | |
3445 | int endian_offset = 0; | |
3446 | /* Don't change the mode of the MEM | |
3447 | if that would change the meaning of the address. */ | |
3448 | if (MEM_VOLATILE_P (SUBREG_REG (x)) | |
3449 | || mode_dependent_address_p (XEXP (inner, 0))) | |
38a448ca | 3450 | return gen_rtx_CLOBBER (mode, const0_rtx); |
230d793d | 3451 | |
f76b9db2 ILT |
3452 | if (BYTES_BIG_ENDIAN) |
3453 | { | |
3454 | if (GET_MODE_SIZE (mode) < UNITS_PER_WORD) | |
3455 | endian_offset += UNITS_PER_WORD - GET_MODE_SIZE (mode); | |
3456 | if (GET_MODE_SIZE (GET_MODE (inner)) < UNITS_PER_WORD) | |
3457 | endian_offset -= (UNITS_PER_WORD | |
3458 | - GET_MODE_SIZE (GET_MODE (inner))); | |
3459 | } | |
230d793d RS |
3460 | /* Note if the plus_constant doesn't make a valid address |
3461 | then this combination won't be accepted. */ | |
38a448ca RH |
3462 | x = gen_rtx_MEM (mode, |
3463 | plus_constant (XEXP (inner, 0), | |
3464 | (SUBREG_WORD (x) * UNITS_PER_WORD | |
3465 | + endian_offset))); | |
230d793d RS |
3466 | MEM_VOLATILE_P (x) = MEM_VOLATILE_P (inner); |
3467 | RTX_UNCHANGING_P (x) = RTX_UNCHANGING_P (inner); | |
3468 | MEM_IN_STRUCT_P (x) = MEM_IN_STRUCT_P (inner); | |
3469 | return x; | |
3470 | } | |
3471 | ||
3472 | /* If we are in a SET_DEST, these other cases can't apply. */ | |
3473 | if (in_dest) | |
3474 | return x; | |
3475 | ||
3476 | /* Changing mode twice with SUBREG => just change it once, | |
3477 | or not at all if changing back to starting mode. */ | |
3478 | if (GET_CODE (SUBREG_REG (x)) == SUBREG) | |
3479 | { | |
3480 | if (mode == GET_MODE (SUBREG_REG (SUBREG_REG (x))) | |
3481 | && SUBREG_WORD (x) == 0 && SUBREG_WORD (SUBREG_REG (x)) == 0) | |
3482 | return SUBREG_REG (SUBREG_REG (x)); | |
3483 | ||
3484 | SUBST_INT (SUBREG_WORD (x), | |
3485 | SUBREG_WORD (x) + SUBREG_WORD (SUBREG_REG (x))); | |
3486 | SUBST (SUBREG_REG (x), SUBREG_REG (SUBREG_REG (x))); | |
3487 | } | |
3488 | ||
3489 | /* SUBREG of a hard register => just change the register number | |
3490 | and/or mode. If the hard register is not valid in that mode, | |
26ecfc76 RK |
3491 | suppress this combination. If the hard register is the stack, |
3492 | frame, or argument pointer, leave this as a SUBREG. */ | |
230d793d RS |
3493 | |
3494 | if (GET_CODE (SUBREG_REG (x)) == REG | |
26ecfc76 RK |
3495 | && REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER |
3496 | && REGNO (SUBREG_REG (x)) != FRAME_POINTER_REGNUM | |
6d7096b0 DE |
3497 | #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM |
3498 | && REGNO (SUBREG_REG (x)) != HARD_FRAME_POINTER_REGNUM | |
3499 | #endif | |
26ecfc76 RK |
3500 | #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM |
3501 | && REGNO (SUBREG_REG (x)) != ARG_POINTER_REGNUM | |
3502 | #endif | |
3503 | && REGNO (SUBREG_REG (x)) != STACK_POINTER_REGNUM) | |
230d793d RS |
3504 | { |
3505 | if (HARD_REGNO_MODE_OK (REGNO (SUBREG_REG (x)) + SUBREG_WORD (x), | |
3506 | mode)) | |
38a448ca RH |
3507 | return gen_rtx_REG (mode, |
3508 | REGNO (SUBREG_REG (x)) + SUBREG_WORD (x)); | |
230d793d | 3509 | else |
38a448ca | 3510 | return gen_rtx_CLOBBER (mode, const0_rtx); |
230d793d RS |
3511 | } |
3512 | ||
3513 | /* For a constant, try to pick up the part we want. Handle a full | |
a4bde0b1 RK |
3514 | word and low-order part. Only do this if we are narrowing |
3515 | the constant; if it is being widened, we have no idea what | |
3516 | the extra bits will have been set to. */ | |
230d793d RS |
3517 | |
3518 | if (CONSTANT_P (SUBREG_REG (x)) && op0_mode != VOIDmode | |
3519 | && GET_MODE_SIZE (mode) == UNITS_PER_WORD | |
3c99d5ff | 3520 | && GET_MODE_SIZE (op0_mode) > UNITS_PER_WORD |
230d793d RS |
3521 | && GET_MODE_CLASS (mode) == MODE_INT) |
3522 | { | |
3523 | temp = operand_subword (SUBREG_REG (x), SUBREG_WORD (x), | |
5f4f0e22 | 3524 | 0, op0_mode); |
230d793d RS |
3525 | if (temp) |
3526 | return temp; | |
3527 | } | |
3528 | ||
19808e22 RS |
3529 | /* If we want a subreg of a constant, at offset 0, |
3530 | take the low bits. On a little-endian machine, that's | |
3531 | always valid. On a big-endian machine, it's valid | |
3c99d5ff | 3532 | only if the constant's mode fits in one word. Note that we |
61b1bece | 3533 | cannot use subreg_lowpart_p since SUBREG_REG may be VOIDmode. */ |
3c99d5ff RK |
3534 | if (CONSTANT_P (SUBREG_REG (x)) |
3535 | && ((GET_MODE_SIZE (op0_mode) <= UNITS_PER_WORD | |
3536 | || ! WORDS_BIG_ENDIAN) | |
3537 | ? SUBREG_WORD (x) == 0 | |
3538 | : (SUBREG_WORD (x) | |
3539 | == ((GET_MODE_SIZE (op0_mode) | |
3540 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)) | |
3541 | / UNITS_PER_WORD))) | |
f82da7d2 | 3542 | && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (op0_mode) |
f76b9db2 ILT |
3543 | && (! WORDS_BIG_ENDIAN |
3544 | || GET_MODE_BITSIZE (op0_mode) <= BITS_PER_WORD)) | |
230d793d RS |
3545 | return gen_lowpart_for_combine (mode, SUBREG_REG (x)); |
3546 | ||
b65c1b5b RK |
3547 | /* A paradoxical SUBREG of a VOIDmode constant is the same constant, |
3548 | since we are saying that the high bits don't matter. */ | |
3549 | if (CONSTANT_P (SUBREG_REG (x)) && GET_MODE (SUBREG_REG (x)) == VOIDmode | |
3550 | && GET_MODE_SIZE (mode) > GET_MODE_SIZE (op0_mode)) | |
3551 | return SUBREG_REG (x); | |
3552 | ||
87e3e0c1 RK |
3553 | /* Note that we cannot do any narrowing for non-constants since |
3554 | we might have been counting on using the fact that some bits were | |
3555 | zero. We now do this in the SET. */ | |
3556 | ||
230d793d RS |
3557 | break; |
3558 | ||
3559 | case NOT: | |
3560 | /* (not (plus X -1)) can become (neg X). */ | |
3561 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3562 | && XEXP (XEXP (x, 0), 1) == constm1_rtx) | |
8079805d | 3563 | return gen_rtx_combine (NEG, mode, XEXP (XEXP (x, 0), 0)); |
230d793d RS |
3564 | |
3565 | /* Similarly, (not (neg X)) is (plus X -1). */ | |
3566 | if (GET_CODE (XEXP (x, 0)) == NEG) | |
8079805d RK |
3567 | return gen_rtx_combine (PLUS, mode, XEXP (XEXP (x, 0), 0), |
3568 | constm1_rtx); | |
230d793d | 3569 | |
d0ab8cd3 RK |
3570 | /* (not (xor X C)) for C constant is (xor X D) with D = ~ C. */ |
3571 | if (GET_CODE (XEXP (x, 0)) == XOR | |
3572 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3573 | && (temp = simplify_unary_operation (NOT, mode, | |
3574 | XEXP (XEXP (x, 0), 1), | |
3575 | mode)) != 0) | |
787745f5 | 3576 | return gen_binary (XOR, mode, XEXP (XEXP (x, 0), 0), temp); |
d0ab8cd3 | 3577 | |
230d793d RS |
3578 | /* (not (ashift 1 X)) is (rotate ~1 X). We used to do this for operands |
3579 | other than 1, but that is not valid. We could do a similar | |
3580 | simplification for (not (lshiftrt C X)) where C is just the sign bit, | |
3581 | but this doesn't seem common enough to bother with. */ | |
3582 | if (GET_CODE (XEXP (x, 0)) == ASHIFT | |
3583 | && XEXP (XEXP (x, 0), 0) == const1_rtx) | |
38a448ca RH |
3584 | return gen_rtx_ROTATE (mode, gen_unary (NOT, mode, mode, const1_rtx), |
3585 | XEXP (XEXP (x, 0), 1)); | |
230d793d RS |
3586 | |
3587 | if (GET_CODE (XEXP (x, 0)) == SUBREG | |
3588 | && subreg_lowpart_p (XEXP (x, 0)) | |
3589 | && (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) | |
3590 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (x, 0))))) | |
3591 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == ASHIFT | |
3592 | && XEXP (SUBREG_REG (XEXP (x, 0)), 0) == const1_rtx) | |
3593 | { | |
3594 | enum machine_mode inner_mode = GET_MODE (SUBREG_REG (XEXP (x, 0))); | |
3595 | ||
38a448ca RH |
3596 | x = gen_rtx_ROTATE (inner_mode, |
3597 | gen_unary (NOT, inner_mode, inner_mode, | |
3598 | const1_rtx), | |
3599 | XEXP (SUBREG_REG (XEXP (x, 0)), 1)); | |
8079805d | 3600 | return gen_lowpart_for_combine (mode, x); |
230d793d RS |
3601 | } |
3602 | ||
0802d516 RK |
3603 | /* If STORE_FLAG_VALUE is -1, (not (comparison foo bar)) can be done by |
3604 | reversing the comparison code if valid. */ | |
3605 | if (STORE_FLAG_VALUE == -1 | |
3606 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
230d793d RS |
3607 | && reversible_comparison_p (XEXP (x, 0))) |
3608 | return gen_rtx_combine (reverse_condition (GET_CODE (XEXP (x, 0))), | |
3609 | mode, XEXP (XEXP (x, 0), 0), | |
3610 | XEXP (XEXP (x, 0), 1)); | |
500c518b RK |
3611 | |
3612 | /* (ashiftrt foo C) where C is the number of bits in FOO minus 1 | |
0802d516 RK |
3613 | is (lt foo (const_int 0)) if STORE_FLAG_VALUE is -1, so we can |
3614 | perform the above simplification. */ | |
500c518b | 3615 | |
0802d516 RK |
3616 | if (STORE_FLAG_VALUE == -1 |
3617 | && XEXP (x, 1) == const1_rtx | |
500c518b RK |
3618 | && GET_CODE (XEXP (x, 0)) == ASHIFTRT |
3619 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3620 | && INTVAL (XEXP (XEXP (x, 0), 1)) == GET_MODE_BITSIZE (mode) - 1) | |
3621 | return gen_rtx_combine (GE, mode, XEXP (XEXP (x, 0), 0), const0_rtx); | |
230d793d RS |
3622 | |
3623 | /* Apply De Morgan's laws to reduce number of patterns for machines | |
3624 | with negating logical insns (and-not, nand, etc.). If result has | |
3625 | only one NOT, put it first, since that is how the patterns are | |
3626 | coded. */ | |
3627 | ||
3628 | if (GET_CODE (XEXP (x, 0)) == IOR || GET_CODE (XEXP (x, 0)) == AND) | |
3629 | { | |
3630 | rtx in1 = XEXP (XEXP (x, 0), 0), in2 = XEXP (XEXP (x, 0), 1); | |
3631 | ||
3632 | if (GET_CODE (in1) == NOT) | |
3633 | in1 = XEXP (in1, 0); | |
3634 | else | |
3635 | in1 = gen_rtx_combine (NOT, GET_MODE (in1), in1); | |
3636 | ||
3637 | if (GET_CODE (in2) == NOT) | |
3638 | in2 = XEXP (in2, 0); | |
3639 | else if (GET_CODE (in2) == CONST_INT | |
5f4f0e22 CH |
3640 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) |
3641 | in2 = GEN_INT (GET_MODE_MASK (mode) & ~ INTVAL (in2)); | |
230d793d RS |
3642 | else |
3643 | in2 = gen_rtx_combine (NOT, GET_MODE (in2), in2); | |
3644 | ||
3645 | if (GET_CODE (in2) == NOT) | |
3646 | { | |
3647 | rtx tem = in2; | |
3648 | in2 = in1; in1 = tem; | |
3649 | } | |
3650 | ||
8079805d RK |
3651 | return gen_rtx_combine (GET_CODE (XEXP (x, 0)) == IOR ? AND : IOR, |
3652 | mode, in1, in2); | |
230d793d RS |
3653 | } |
3654 | break; | |
3655 | ||
3656 | case NEG: | |
3657 | /* (neg (plus X 1)) can become (not X). */ | |
3658 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3659 | && XEXP (XEXP (x, 0), 1) == const1_rtx) | |
8079805d | 3660 | return gen_rtx_combine (NOT, mode, XEXP (XEXP (x, 0), 0)); |
230d793d RS |
3661 | |
3662 | /* Similarly, (neg (not X)) is (plus X 1). */ | |
3663 | if (GET_CODE (XEXP (x, 0)) == NOT) | |
8079805d | 3664 | return plus_constant (XEXP (XEXP (x, 0), 0), 1); |
230d793d | 3665 | |
230d793d RS |
3666 | /* (neg (minus X Y)) can become (minus Y X). */ |
3667 | if (GET_CODE (XEXP (x, 0)) == MINUS | |
3ad2180a | 3668 | && (! FLOAT_MODE_P (mode) |
0f41302f | 3669 | /* x-y != -(y-x) with IEEE floating point. */ |
7e2a0d8e RK |
3670 | || TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT |
3671 | || flag_fast_math)) | |
8079805d RK |
3672 | return gen_binary (MINUS, mode, XEXP (XEXP (x, 0), 1), |
3673 | XEXP (XEXP (x, 0), 0)); | |
230d793d | 3674 | |
0f41302f | 3675 | /* (neg (xor A 1)) is (plus A -1) if A is known to be either 0 or 1. */ |
d0ab8cd3 | 3676 | if (GET_CODE (XEXP (x, 0)) == XOR && XEXP (XEXP (x, 0), 1) == const1_rtx |
951553af | 3677 | && nonzero_bits (XEXP (XEXP (x, 0), 0), mode) == 1) |
8079805d | 3678 | return gen_binary (PLUS, mode, XEXP (XEXP (x, 0), 0), constm1_rtx); |
d0ab8cd3 | 3679 | |
230d793d RS |
3680 | /* NEG commutes with ASHIFT since it is multiplication. Only do this |
3681 | if we can then eliminate the NEG (e.g., | |
3682 | if the operand is a constant). */ | |
3683 | ||
3684 | if (GET_CODE (XEXP (x, 0)) == ASHIFT) | |
3685 | { | |
3686 | temp = simplify_unary_operation (NEG, mode, | |
3687 | XEXP (XEXP (x, 0), 0), mode); | |
3688 | if (temp) | |
3689 | { | |
3690 | SUBST (XEXP (XEXP (x, 0), 0), temp); | |
3691 | return XEXP (x, 0); | |
3692 | } | |
3693 | } | |
3694 | ||
3695 | temp = expand_compound_operation (XEXP (x, 0)); | |
3696 | ||
3697 | /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be | |
3698 | replaced by (lshiftrt X C). This will convert | |
3699 | (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */ | |
3700 | ||
3701 | if (GET_CODE (temp) == ASHIFTRT | |
3702 | && GET_CODE (XEXP (temp, 1)) == CONST_INT | |
3703 | && INTVAL (XEXP (temp, 1)) == GET_MODE_BITSIZE (mode) - 1) | |
8079805d RK |
3704 | return simplify_shift_const (temp, LSHIFTRT, mode, XEXP (temp, 0), |
3705 | INTVAL (XEXP (temp, 1))); | |
230d793d | 3706 | |
951553af | 3707 | /* If X has only a single bit that might be nonzero, say, bit I, convert |
230d793d RS |
3708 | (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of |
3709 | MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to | |
3710 | (sign_extract X 1 Y). But only do this if TEMP isn't a register | |
3711 | or a SUBREG of one since we'd be making the expression more | |
3712 | complex if it was just a register. */ | |
3713 | ||
3714 | if (GET_CODE (temp) != REG | |
3715 | && ! (GET_CODE (temp) == SUBREG | |
3716 | && GET_CODE (SUBREG_REG (temp)) == REG) | |
951553af | 3717 | && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0) |
230d793d RS |
3718 | { |
3719 | rtx temp1 = simplify_shift_const | |
5f4f0e22 CH |
3720 | (NULL_RTX, ASHIFTRT, mode, |
3721 | simplify_shift_const (NULL_RTX, ASHIFT, mode, temp, | |
230d793d RS |
3722 | GET_MODE_BITSIZE (mode) - 1 - i), |
3723 | GET_MODE_BITSIZE (mode) - 1 - i); | |
3724 | ||
3725 | /* If all we did was surround TEMP with the two shifts, we | |
3726 | haven't improved anything, so don't use it. Otherwise, | |
3727 | we are better off with TEMP1. */ | |
3728 | if (GET_CODE (temp1) != ASHIFTRT | |
3729 | || GET_CODE (XEXP (temp1, 0)) != ASHIFT | |
3730 | || XEXP (XEXP (temp1, 0), 0) != temp) | |
8079805d | 3731 | return temp1; |
230d793d RS |
3732 | } |
3733 | break; | |
3734 | ||
2ca9ae17 | 3735 | case TRUNCATE: |
e30fb98f JL |
3736 | /* We can't handle truncation to a partial integer mode here |
3737 | because we don't know the real bitsize of the partial | |
3738 | integer mode. */ | |
3739 | if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT) | |
3740 | break; | |
3741 | ||
80608e27 JL |
3742 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
3743 | && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode), | |
3744 | GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))) | |
2ca9ae17 JW |
3745 | SUBST (XEXP (x, 0), |
3746 | force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)), | |
3747 | GET_MODE_MASK (mode), NULL_RTX, 0)); | |
0f13a422 ILT |
3748 | |
3749 | /* (truncate:SI ({sign,zero}_extend:DI foo:SI)) == foo:SI. */ | |
3750 | if ((GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3751 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3752 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode) | |
3753 | return XEXP (XEXP (x, 0), 0); | |
3754 | ||
3755 | /* (truncate:SI (OP:DI ({sign,zero}_extend:DI foo:SI))) is | |
3756 | (OP:SI foo:SI) if OP is NEG or ABS. */ | |
3757 | if ((GET_CODE (XEXP (x, 0)) == ABS | |
3758 | || GET_CODE (XEXP (x, 0)) == NEG) | |
3759 | && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SIGN_EXTEND | |
3760 | || GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND) | |
3761 | && GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == mode) | |
3762 | return gen_unary (GET_CODE (XEXP (x, 0)), mode, mode, | |
3763 | XEXP (XEXP (XEXP (x, 0), 0), 0)); | |
3764 | ||
3765 | /* (truncate:SI (subreg:DI (truncate:SI X) 0)) is | |
3766 | (truncate:SI x). */ | |
3767 | if (GET_CODE (XEXP (x, 0)) == SUBREG | |
3768 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == TRUNCATE | |
3769 | && subreg_lowpart_p (XEXP (x, 0))) | |
3770 | return SUBREG_REG (XEXP (x, 0)); | |
3771 | ||
3772 | /* If we know that the value is already truncated, we can | |
3773 | replace the TRUNCATE with a SUBREG. */ | |
9ec36da5 JL |
3774 | if (num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))) |
3775 | >= GET_MODE_BITSIZE (mode) + 1) | |
0f13a422 ILT |
3776 | return gen_lowpart_for_combine (mode, XEXP (x, 0)); |
3777 | ||
3778 | /* A truncate of a comparison can be replaced with a subreg if | |
3779 | STORE_FLAG_VALUE permits. This is like the previous test, | |
3780 | but it works even if the comparison is done in a mode larger | |
3781 | than HOST_BITS_PER_WIDE_INT. */ | |
3782 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
3783 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3784 | && ((HOST_WIDE_INT) STORE_FLAG_VALUE &~ GET_MODE_MASK (mode)) == 0) | |
3785 | return gen_lowpart_for_combine (mode, XEXP (x, 0)); | |
3786 | ||
3787 | /* Similarly, a truncate of a register whose value is a | |
3788 | comparison can be replaced with a subreg if STORE_FLAG_VALUE | |
3789 | permits. */ | |
3790 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
3791 | && ((HOST_WIDE_INT) STORE_FLAG_VALUE &~ GET_MODE_MASK (mode)) == 0 | |
3792 | && (temp = get_last_value (XEXP (x, 0))) | |
3793 | && GET_RTX_CLASS (GET_CODE (temp)) == '<') | |
3794 | return gen_lowpart_for_combine (mode, XEXP (x, 0)); | |
3795 | ||
2ca9ae17 JW |
3796 | break; |
3797 | ||
230d793d RS |
3798 | case FLOAT_TRUNCATE: |
3799 | /* (float_truncate:SF (float_extend:DF foo:SF)) = foo:SF. */ | |
3800 | if (GET_CODE (XEXP (x, 0)) == FLOAT_EXTEND | |
3801 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode) | |
3802 | return XEXP (XEXP (x, 0), 0); | |
4635f748 RK |
3803 | |
3804 | /* (float_truncate:SF (OP:DF (float_extend:DF foo:sf))) is | |
3805 | (OP:SF foo:SF) if OP is NEG or ABS. */ | |
3806 | if ((GET_CODE (XEXP (x, 0)) == ABS | |
3807 | || GET_CODE (XEXP (x, 0)) == NEG) | |
3808 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == FLOAT_EXTEND | |
3809 | && GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == mode) | |
0c1c8ea6 RK |
3810 | return gen_unary (GET_CODE (XEXP (x, 0)), mode, mode, |
3811 | XEXP (XEXP (XEXP (x, 0), 0), 0)); | |
1d12df72 RK |
3812 | |
3813 | /* (float_truncate:SF (subreg:DF (float_truncate:SF X) 0)) | |
3814 | is (float_truncate:SF x). */ | |
3815 | if (GET_CODE (XEXP (x, 0)) == SUBREG | |
3816 | && subreg_lowpart_p (XEXP (x, 0)) | |
3817 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == FLOAT_TRUNCATE) | |
3818 | return SUBREG_REG (XEXP (x, 0)); | |
230d793d RS |
3819 | break; |
3820 | ||
3821 | #ifdef HAVE_cc0 | |
3822 | case COMPARE: | |
3823 | /* Convert (compare FOO (const_int 0)) to FOO unless we aren't | |
3824 | using cc0, in which case we want to leave it as a COMPARE | |
3825 | so we can distinguish it from a register-register-copy. */ | |
3826 | if (XEXP (x, 1) == const0_rtx) | |
3827 | return XEXP (x, 0); | |
3828 | ||
3829 | /* In IEEE floating point, x-0 is not the same as x. */ | |
3830 | if ((TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
7e2a0d8e RK |
3831 | || ! FLOAT_MODE_P (GET_MODE (XEXP (x, 0))) |
3832 | || flag_fast_math) | |
230d793d RS |
3833 | && XEXP (x, 1) == CONST0_RTX (GET_MODE (XEXP (x, 0)))) |
3834 | return XEXP (x, 0); | |
3835 | break; | |
3836 | #endif | |
3837 | ||
3838 | case CONST: | |
3839 | /* (const (const X)) can become (const X). Do it this way rather than | |
3840 | returning the inner CONST since CONST can be shared with a | |
3841 | REG_EQUAL note. */ | |
3842 | if (GET_CODE (XEXP (x, 0)) == CONST) | |
3843 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
3844 | break; | |
3845 | ||
3846 | #ifdef HAVE_lo_sum | |
3847 | case LO_SUM: | |
3848 | /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we | |
3849 | can add in an offset. find_split_point will split this address up | |
3850 | again if it doesn't match. */ | |
3851 | if (GET_CODE (XEXP (x, 0)) == HIGH | |
3852 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1))) | |
3853 | return XEXP (x, 1); | |
3854 | break; | |
3855 | #endif | |
3856 | ||
3857 | case PLUS: | |
3858 | /* If we have (plus (plus (A const) B)), associate it so that CONST is | |
3859 | outermost. That's because that's the way indexed addresses are | |
3860 | supposed to appear. This code used to check many more cases, but | |
3861 | they are now checked elsewhere. */ | |
3862 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3863 | && CONSTANT_ADDRESS_P (XEXP (XEXP (x, 0), 1))) | |
3864 | return gen_binary (PLUS, mode, | |
3865 | gen_binary (PLUS, mode, XEXP (XEXP (x, 0), 0), | |
3866 | XEXP (x, 1)), | |
3867 | XEXP (XEXP (x, 0), 1)); | |
3868 | ||
3869 | /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>) | |
3870 | when c is (const_int (pow2 + 1) / 2) is a sign extension of a | |
3871 | bit-field and can be replaced by either a sign_extend or a | |
3872 | sign_extract. The `and' may be a zero_extend. */ | |
3873 | if (GET_CODE (XEXP (x, 0)) == XOR | |
3874 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
3875 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
3876 | && INTVAL (XEXP (x, 1)) == - INTVAL (XEXP (XEXP (x, 0), 1)) | |
3877 | && (i = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) >= 0 | |
5f4f0e22 | 3878 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
230d793d RS |
3879 | && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND |
3880 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
3881 | && (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1)) | |
5f4f0e22 | 3882 | == ((HOST_WIDE_INT) 1 << (i + 1)) - 1)) |
230d793d RS |
3883 | || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND |
3884 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0))) | |
3885 | == i + 1)))) | |
8079805d RK |
3886 | return simplify_shift_const |
3887 | (NULL_RTX, ASHIFTRT, mode, | |
3888 | simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
3889 | XEXP (XEXP (XEXP (x, 0), 0), 0), | |
3890 | GET_MODE_BITSIZE (mode) - (i + 1)), | |
3891 | GET_MODE_BITSIZE (mode) - (i + 1)); | |
230d793d | 3892 | |
bc0776c6 RK |
3893 | /* (plus (comparison A B) C) can become (neg (rev-comp A B)) if |
3894 | C is 1 and STORE_FLAG_VALUE is -1 or if C is -1 and STORE_FLAG_VALUE | |
3895 | is 1. This produces better code than the alternative immediately | |
3896 | below. */ | |
3897 | if (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
3898 | && reversible_comparison_p (XEXP (x, 0)) | |
3899 | && ((STORE_FLAG_VALUE == -1 && XEXP (x, 1) == const1_rtx) | |
3900 | || (STORE_FLAG_VALUE == 1 && XEXP (x, 1) == constm1_rtx))) | |
8079805d | 3901 | return |
0c1c8ea6 | 3902 | gen_unary (NEG, mode, mode, |
8079805d RK |
3903 | gen_binary (reverse_condition (GET_CODE (XEXP (x, 0))), |
3904 | mode, XEXP (XEXP (x, 0), 0), | |
3905 | XEXP (XEXP (x, 0), 1))); | |
bc0776c6 RK |
3906 | |
3907 | /* If only the low-order bit of X is possibly nonzero, (plus x -1) | |
230d793d RS |
3908 | can become (ashiftrt (ashift (xor x 1) C) C) where C is |
3909 | the bitsize of the mode - 1. This allows simplification of | |
3910 | "a = (b & 8) == 0;" */ | |
3911 | if (XEXP (x, 1) == constm1_rtx | |
3912 | && GET_CODE (XEXP (x, 0)) != REG | |
3913 | && ! (GET_CODE (XEXP (x,0)) == SUBREG | |
3914 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG) | |
951553af | 3915 | && nonzero_bits (XEXP (x, 0), mode) == 1) |
8079805d RK |
3916 | return simplify_shift_const (NULL_RTX, ASHIFTRT, mode, |
3917 | simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
3918 | gen_rtx_combine (XOR, mode, | |
3919 | XEXP (x, 0), const1_rtx), | |
3920 | GET_MODE_BITSIZE (mode) - 1), | |
3921 | GET_MODE_BITSIZE (mode) - 1); | |
02f4ada4 RK |
3922 | |
3923 | /* If we are adding two things that have no bits in common, convert | |
3924 | the addition into an IOR. This will often be further simplified, | |
3925 | for example in cases like ((a & 1) + (a & 2)), which can | |
3926 | become a & 3. */ | |
3927 | ||
ac49a949 | 3928 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
951553af RK |
3929 | && (nonzero_bits (XEXP (x, 0), mode) |
3930 | & nonzero_bits (XEXP (x, 1), mode)) == 0) | |
8079805d | 3931 | return gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1)); |
230d793d RS |
3932 | break; |
3933 | ||
3934 | case MINUS: | |
0802d516 RK |
3935 | /* If STORE_FLAG_VALUE is 1, (minus 1 (comparison foo bar)) can be done |
3936 | by reversing the comparison code if valid. */ | |
3937 | if (STORE_FLAG_VALUE == 1 | |
3938 | && XEXP (x, 0) == const1_rtx | |
5109d49f RK |
3939 | && GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) == '<' |
3940 | && reversible_comparison_p (XEXP (x, 1))) | |
3941 | return gen_binary (reverse_condition (GET_CODE (XEXP (x, 1))), | |
3942 | mode, XEXP (XEXP (x, 1), 0), | |
3943 | XEXP (XEXP (x, 1), 1)); | |
5109d49f | 3944 | |
230d793d RS |
3945 | /* (minus <foo> (and <foo> (const_int -pow2))) becomes |
3946 | (and <foo> (const_int pow2-1)) */ | |
3947 | if (GET_CODE (XEXP (x, 1)) == AND | |
3948 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT | |
3949 | && exact_log2 (- INTVAL (XEXP (XEXP (x, 1), 1))) >= 0 | |
3950 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) | |
8079805d RK |
3951 | return simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0), |
3952 | - INTVAL (XEXP (XEXP (x, 1), 1)) - 1); | |
7bef8680 RK |
3953 | |
3954 | /* Canonicalize (minus A (plus B C)) to (minus (minus A B) C) for | |
3955 | integers. */ | |
3956 | if (GET_CODE (XEXP (x, 1)) == PLUS && INTEGRAL_MODE_P (mode)) | |
8079805d RK |
3957 | return gen_binary (MINUS, mode, |
3958 | gen_binary (MINUS, mode, XEXP (x, 0), | |
3959 | XEXP (XEXP (x, 1), 0)), | |
3960 | XEXP (XEXP (x, 1), 1)); | |
230d793d RS |
3961 | break; |
3962 | ||
3963 | case MULT: | |
3964 | /* If we have (mult (plus A B) C), apply the distributive law and then | |
3965 | the inverse distributive law to see if things simplify. This | |
3966 | occurs mostly in addresses, often when unrolling loops. */ | |
3967 | ||
3968 | if (GET_CODE (XEXP (x, 0)) == PLUS) | |
3969 | { | |
3970 | x = apply_distributive_law | |
3971 | (gen_binary (PLUS, mode, | |
3972 | gen_binary (MULT, mode, | |
3973 | XEXP (XEXP (x, 0), 0), XEXP (x, 1)), | |
3974 | gen_binary (MULT, mode, | |
3975 | XEXP (XEXP (x, 0), 1), XEXP (x, 1)))); | |
3976 | ||
3977 | if (GET_CODE (x) != MULT) | |
8079805d | 3978 | return x; |
230d793d | 3979 | } |
230d793d RS |
3980 | break; |
3981 | ||
3982 | case UDIV: | |
3983 | /* If this is a divide by a power of two, treat it as a shift if | |
3984 | its first operand is a shift. */ | |
3985 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
3986 | && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0 | |
3987 | && (GET_CODE (XEXP (x, 0)) == ASHIFT | |
3988 | || GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
3989 | || GET_CODE (XEXP (x, 0)) == ASHIFTRT | |
3990 | || GET_CODE (XEXP (x, 0)) == ROTATE | |
3991 | || GET_CODE (XEXP (x, 0)) == ROTATERT)) | |
8079805d | 3992 | return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i); |
230d793d RS |
3993 | break; |
3994 | ||
3995 | case EQ: case NE: | |
3996 | case GT: case GTU: case GE: case GEU: | |
3997 | case LT: case LTU: case LE: case LEU: | |
3998 | /* If the first operand is a condition code, we can't do anything | |
3999 | with it. */ | |
4000 | if (GET_CODE (XEXP (x, 0)) == COMPARE | |
4001 | || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC | |
4002 | #ifdef HAVE_cc0 | |
4003 | && XEXP (x, 0) != cc0_rtx | |
4004 | #endif | |
4005 | )) | |
4006 | { | |
4007 | rtx op0 = XEXP (x, 0); | |
4008 | rtx op1 = XEXP (x, 1); | |
4009 | enum rtx_code new_code; | |
4010 | ||
4011 | if (GET_CODE (op0) == COMPARE) | |
4012 | op1 = XEXP (op0, 1), op0 = XEXP (op0, 0); | |
4013 | ||
4014 | /* Simplify our comparison, if possible. */ | |
4015 | new_code = simplify_comparison (code, &op0, &op1); | |
4016 | ||
230d793d | 4017 | /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X |
951553af | 4018 | if only the low-order bit is possibly nonzero in X (such as when |
5109d49f RK |
4019 | X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to |
4020 | (xor X 1) or (minus 1 X); we use the former. Finally, if X is | |
4021 | known to be either 0 or -1, NE becomes a NEG and EQ becomes | |
4022 | (plus X 1). | |
4023 | ||
4024 | Remove any ZERO_EXTRACT we made when thinking this was a | |
4025 | comparison. It may now be simpler to use, e.g., an AND. If a | |
4026 | ZERO_EXTRACT is indeed appropriate, it will be placed back by | |
4027 | the call to make_compound_operation in the SET case. */ | |
4028 | ||
0802d516 RK |
4029 | if (STORE_FLAG_VALUE == 1 |
4030 | && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
4031 | && op1 == const0_rtx && nonzero_bits (op0, mode) == 1) | |
818b11b9 RK |
4032 | return gen_lowpart_for_combine (mode, |
4033 | expand_compound_operation (op0)); | |
5109d49f | 4034 | |
0802d516 RK |
4035 | else if (STORE_FLAG_VALUE == 1 |
4036 | && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
5109d49f RK |
4037 | && op1 == const0_rtx |
4038 | && (num_sign_bit_copies (op0, mode) | |
4039 | == GET_MODE_BITSIZE (mode))) | |
4040 | { | |
4041 | op0 = expand_compound_operation (op0); | |
0c1c8ea6 | 4042 | return gen_unary (NEG, mode, mode, |
8079805d | 4043 | gen_lowpart_for_combine (mode, op0)); |
5109d49f RK |
4044 | } |
4045 | ||
0802d516 RK |
4046 | else if (STORE_FLAG_VALUE == 1 |
4047 | && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT | |
230d793d | 4048 | && op1 == const0_rtx |
5109d49f | 4049 | && nonzero_bits (op0, mode) == 1) |
818b11b9 RK |
4050 | { |
4051 | op0 = expand_compound_operation (op0); | |
8079805d RK |
4052 | return gen_binary (XOR, mode, |
4053 | gen_lowpart_for_combine (mode, op0), | |
4054 | const1_rtx); | |
5109d49f | 4055 | } |
818b11b9 | 4056 | |
0802d516 RK |
4057 | else if (STORE_FLAG_VALUE == 1 |
4058 | && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT | |
5109d49f RK |
4059 | && op1 == const0_rtx |
4060 | && (num_sign_bit_copies (op0, mode) | |
4061 | == GET_MODE_BITSIZE (mode))) | |
4062 | { | |
4063 | op0 = expand_compound_operation (op0); | |
8079805d | 4064 | return plus_constant (gen_lowpart_for_combine (mode, op0), 1); |
818b11b9 | 4065 | } |
230d793d | 4066 | |
5109d49f RK |
4067 | /* If STORE_FLAG_VALUE is -1, we have cases similar to |
4068 | those above. */ | |
0802d516 RK |
4069 | if (STORE_FLAG_VALUE == -1 |
4070 | && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
230d793d | 4071 | && op1 == const0_rtx |
5109d49f RK |
4072 | && (num_sign_bit_copies (op0, mode) |
4073 | == GET_MODE_BITSIZE (mode))) | |
4074 | return gen_lowpart_for_combine (mode, | |
4075 | expand_compound_operation (op0)); | |
4076 | ||
0802d516 RK |
4077 | else if (STORE_FLAG_VALUE == -1 |
4078 | && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT | |
5109d49f RK |
4079 | && op1 == const0_rtx |
4080 | && nonzero_bits (op0, mode) == 1) | |
4081 | { | |
4082 | op0 = expand_compound_operation (op0); | |
0c1c8ea6 | 4083 | return gen_unary (NEG, mode, mode, |
8079805d | 4084 | gen_lowpart_for_combine (mode, op0)); |
5109d49f RK |
4085 | } |
4086 | ||
0802d516 RK |
4087 | else if (STORE_FLAG_VALUE == -1 |
4088 | && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT | |
5109d49f RK |
4089 | && op1 == const0_rtx |
4090 | && (num_sign_bit_copies (op0, mode) | |
4091 | == GET_MODE_BITSIZE (mode))) | |
230d793d | 4092 | { |
818b11b9 | 4093 | op0 = expand_compound_operation (op0); |
0c1c8ea6 | 4094 | return gen_unary (NOT, mode, mode, |
8079805d | 4095 | gen_lowpart_for_combine (mode, op0)); |
5109d49f RK |
4096 | } |
4097 | ||
4098 | /* If X is 0/1, (eq X 0) is X-1. */ | |
0802d516 RK |
4099 | else if (STORE_FLAG_VALUE == -1 |
4100 | && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT | |
5109d49f RK |
4101 | && op1 == const0_rtx |
4102 | && nonzero_bits (op0, mode) == 1) | |
4103 | { | |
4104 | op0 = expand_compound_operation (op0); | |
8079805d | 4105 | return plus_constant (gen_lowpart_for_combine (mode, op0), -1); |
230d793d | 4106 | } |
230d793d RS |
4107 | |
4108 | /* If STORE_FLAG_VALUE says to just test the sign bit and X has just | |
951553af RK |
4109 | one bit that might be nonzero, we can convert (ne x 0) to |
4110 | (ashift x c) where C puts the bit in the sign bit. Remove any | |
4111 | AND with STORE_FLAG_VALUE when we are done, since we are only | |
4112 | going to test the sign bit. */ | |
3f508eca | 4113 | if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT |
5f4f0e22 | 4114 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
0802d516 | 4115 | && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode)) |
5f4f0e22 | 4116 | == (HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1)) |
230d793d RS |
4117 | && op1 == const0_rtx |
4118 | && mode == GET_MODE (op0) | |
5109d49f | 4119 | && (i = exact_log2 (nonzero_bits (op0, mode))) >= 0) |
230d793d | 4120 | { |
818b11b9 RK |
4121 | x = simplify_shift_const (NULL_RTX, ASHIFT, mode, |
4122 | expand_compound_operation (op0), | |
230d793d RS |
4123 | GET_MODE_BITSIZE (mode) - 1 - i); |
4124 | if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx) | |
4125 | return XEXP (x, 0); | |
4126 | else | |
4127 | return x; | |
4128 | } | |
4129 | ||
4130 | /* If the code changed, return a whole new comparison. */ | |
4131 | if (new_code != code) | |
4132 | return gen_rtx_combine (new_code, mode, op0, op1); | |
4133 | ||
4134 | /* Otherwise, keep this operation, but maybe change its operands. | |
4135 | This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */ | |
4136 | SUBST (XEXP (x, 0), op0); | |
4137 | SUBST (XEXP (x, 1), op1); | |
4138 | } | |
4139 | break; | |
4140 | ||
4141 | case IF_THEN_ELSE: | |
8079805d | 4142 | return simplify_if_then_else (x); |
9210df58 | 4143 | |
8079805d RK |
4144 | case ZERO_EXTRACT: |
4145 | case SIGN_EXTRACT: | |
4146 | case ZERO_EXTEND: | |
4147 | case SIGN_EXTEND: | |
0f41302f | 4148 | /* If we are processing SET_DEST, we are done. */ |
8079805d RK |
4149 | if (in_dest) |
4150 | return x; | |
d0ab8cd3 | 4151 | |
8079805d | 4152 | return expand_compound_operation (x); |
d0ab8cd3 | 4153 | |
8079805d RK |
4154 | case SET: |
4155 | return simplify_set (x); | |
1a26b032 | 4156 | |
8079805d RK |
4157 | case AND: |
4158 | case IOR: | |
4159 | case XOR: | |
4160 | return simplify_logical (x, last); | |
d0ab8cd3 | 4161 | |
b472527b | 4162 | case ABS: |
8079805d RK |
4163 | /* (abs (neg <foo>)) -> (abs <foo>) */ |
4164 | if (GET_CODE (XEXP (x, 0)) == NEG) | |
4165 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
1a26b032 | 4166 | |
b472527b JL |
4167 | /* If the mode of the operand is VOIDmode (i.e. if it is ASM_OPERANDS), |
4168 | do nothing. */ | |
4169 | if (GET_MODE (XEXP (x, 0)) == VOIDmode) | |
4170 | break; | |
f40421ce | 4171 | |
8079805d RK |
4172 | /* If operand is something known to be positive, ignore the ABS. */ |
4173 | if (GET_CODE (XEXP (x, 0)) == FFS || GET_CODE (XEXP (x, 0)) == ABS | |
4174 | || ((GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
4175 | <= HOST_BITS_PER_WIDE_INT) | |
4176 | && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0))) | |
4177 | & ((HOST_WIDE_INT) 1 | |
4178 | << (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - 1))) | |
4179 | == 0))) | |
4180 | return XEXP (x, 0); | |
1a26b032 | 4181 | |
1a26b032 | 4182 | |
8079805d RK |
4183 | /* If operand is known to be only -1 or 0, convert ABS to NEG. */ |
4184 | if (num_sign_bit_copies (XEXP (x, 0), mode) == GET_MODE_BITSIZE (mode)) | |
4185 | return gen_rtx_combine (NEG, mode, XEXP (x, 0)); | |
1a26b032 | 4186 | |
8079805d | 4187 | break; |
1a26b032 | 4188 | |
8079805d RK |
4189 | case FFS: |
4190 | /* (ffs (*_extend <X>)) = (ffs <X>) */ | |
4191 | if (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
4192 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
4193 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
4194 | break; | |
1a26b032 | 4195 | |
8079805d RK |
4196 | case FLOAT: |
4197 | /* (float (sign_extend <X>)) = (float <X>). */ | |
4198 | if (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
4199 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); | |
4200 | break; | |
1a26b032 | 4201 | |
8079805d RK |
4202 | case ASHIFT: |
4203 | case LSHIFTRT: | |
4204 | case ASHIFTRT: | |
4205 | case ROTATE: | |
4206 | case ROTATERT: | |
4207 | /* If this is a shift by a constant amount, simplify it. */ | |
4208 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4209 | return simplify_shift_const (x, code, mode, XEXP (x, 0), | |
4210 | INTVAL (XEXP (x, 1))); | |
4211 | ||
4212 | #ifdef SHIFT_COUNT_TRUNCATED | |
4213 | else if (SHIFT_COUNT_TRUNCATED && GET_CODE (XEXP (x, 1)) != REG) | |
4214 | SUBST (XEXP (x, 1), | |
4215 | force_to_mode (XEXP (x, 1), GET_MODE (x), | |
4216 | ((HOST_WIDE_INT) 1 | |
4217 | << exact_log2 (GET_MODE_BITSIZE (GET_MODE (x)))) | |
4218 | - 1, | |
4219 | NULL_RTX, 0)); | |
4220 | #endif | |
4221 | ||
4222 | break; | |
e9a25f70 JL |
4223 | |
4224 | default: | |
4225 | break; | |
8079805d RK |
4226 | } |
4227 | ||
4228 | return x; | |
4229 | } | |
4230 | \f | |
4231 | /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */ | |
5109d49f | 4232 | |
8079805d RK |
4233 | static rtx |
4234 | simplify_if_then_else (x) | |
4235 | rtx x; | |
4236 | { | |
4237 | enum machine_mode mode = GET_MODE (x); | |
4238 | rtx cond = XEXP (x, 0); | |
4239 | rtx true = XEXP (x, 1); | |
4240 | rtx false = XEXP (x, 2); | |
4241 | enum rtx_code true_code = GET_CODE (cond); | |
4242 | int comparison_p = GET_RTX_CLASS (true_code) == '<'; | |
4243 | rtx temp; | |
4244 | int i; | |
4245 | ||
0f41302f | 4246 | /* Simplify storing of the truth value. */ |
8079805d RK |
4247 | if (comparison_p && true == const_true_rtx && false == const0_rtx) |
4248 | return gen_binary (true_code, mode, XEXP (cond, 0), XEXP (cond, 1)); | |
4249 | ||
0f41302f | 4250 | /* Also when the truth value has to be reversed. */ |
8079805d RK |
4251 | if (comparison_p && reversible_comparison_p (cond) |
4252 | && true == const0_rtx && false == const_true_rtx) | |
4253 | return gen_binary (reverse_condition (true_code), | |
4254 | mode, XEXP (cond, 0), XEXP (cond, 1)); | |
4255 | ||
4256 | /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used | |
4257 | in it is being compared against certain values. Get the true and false | |
4258 | comparisons and see if that says anything about the value of each arm. */ | |
4259 | ||
4260 | if (comparison_p && reversible_comparison_p (cond) | |
4261 | && GET_CODE (XEXP (cond, 0)) == REG) | |
4262 | { | |
4263 | HOST_WIDE_INT nzb; | |
4264 | rtx from = XEXP (cond, 0); | |
4265 | enum rtx_code false_code = reverse_condition (true_code); | |
4266 | rtx true_val = XEXP (cond, 1); | |
4267 | rtx false_val = true_val; | |
4268 | int swapped = 0; | |
9210df58 | 4269 | |
8079805d | 4270 | /* If FALSE_CODE is EQ, swap the codes and arms. */ |
5109d49f | 4271 | |
8079805d | 4272 | if (false_code == EQ) |
1a26b032 | 4273 | { |
8079805d RK |
4274 | swapped = 1, true_code = EQ, false_code = NE; |
4275 | temp = true, true = false, false = temp; | |
4276 | } | |
5109d49f | 4277 | |
8079805d RK |
4278 | /* If we are comparing against zero and the expression being tested has |
4279 | only a single bit that might be nonzero, that is its value when it is | |
4280 | not equal to zero. Similarly if it is known to be -1 or 0. */ | |
4281 | ||
4282 | if (true_code == EQ && true_val == const0_rtx | |
4283 | && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0) | |
4284 | false_code = EQ, false_val = GEN_INT (nzb); | |
4285 | else if (true_code == EQ && true_val == const0_rtx | |
4286 | && (num_sign_bit_copies (from, GET_MODE (from)) | |
4287 | == GET_MODE_BITSIZE (GET_MODE (from)))) | |
4288 | false_code = EQ, false_val = constm1_rtx; | |
4289 | ||
4290 | /* Now simplify an arm if we know the value of the register in the | |
4291 | branch and it is used in the arm. Be careful due to the potential | |
4292 | of locally-shared RTL. */ | |
4293 | ||
4294 | if (reg_mentioned_p (from, true)) | |
4295 | true = subst (known_cond (copy_rtx (true), true_code, from, true_val), | |
4296 | pc_rtx, pc_rtx, 0, 0); | |
4297 | if (reg_mentioned_p (from, false)) | |
4298 | false = subst (known_cond (copy_rtx (false), false_code, | |
4299 | from, false_val), | |
4300 | pc_rtx, pc_rtx, 0, 0); | |
4301 | ||
4302 | SUBST (XEXP (x, 1), swapped ? false : true); | |
4303 | SUBST (XEXP (x, 2), swapped ? true : false); | |
4304 | ||
4305 | true = XEXP (x, 1), false = XEXP (x, 2), true_code = GET_CODE (cond); | |
4306 | } | |
5109d49f | 4307 | |
8079805d RK |
4308 | /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be |
4309 | reversed, do so to avoid needing two sets of patterns for | |
4310 | subtract-and-branch insns. Similarly if we have a constant in the true | |
4311 | arm, the false arm is the same as the first operand of the comparison, or | |
4312 | the false arm is more complicated than the true arm. */ | |
4313 | ||
4314 | if (comparison_p && reversible_comparison_p (cond) | |
4315 | && (true == pc_rtx | |
4316 | || (CONSTANT_P (true) | |
4317 | && GET_CODE (false) != CONST_INT && false != pc_rtx) | |
4318 | || true == const0_rtx | |
4319 | || (GET_RTX_CLASS (GET_CODE (true)) == 'o' | |
4320 | && GET_RTX_CLASS (GET_CODE (false)) != 'o') | |
4321 | || (GET_CODE (true) == SUBREG | |
4322 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (true))) == 'o' | |
4323 | && GET_RTX_CLASS (GET_CODE (false)) != 'o') | |
4324 | || reg_mentioned_p (true, false) | |
4325 | || rtx_equal_p (false, XEXP (cond, 0)))) | |
4326 | { | |
4327 | true_code = reverse_condition (true_code); | |
4328 | SUBST (XEXP (x, 0), | |
4329 | gen_binary (true_code, GET_MODE (cond), XEXP (cond, 0), | |
4330 | XEXP (cond, 1))); | |
5109d49f | 4331 | |
8079805d RK |
4332 | SUBST (XEXP (x, 1), false); |
4333 | SUBST (XEXP (x, 2), true); | |
1a26b032 | 4334 | |
8079805d | 4335 | temp = true, true = false, false = temp, cond = XEXP (x, 0); |
bb821298 | 4336 | |
0f41302f | 4337 | /* It is possible that the conditional has been simplified out. */ |
bb821298 RK |
4338 | true_code = GET_CODE (cond); |
4339 | comparison_p = GET_RTX_CLASS (true_code) == '<'; | |
8079805d | 4340 | } |
abe6e52f | 4341 | |
8079805d | 4342 | /* If the two arms are identical, we don't need the comparison. */ |
1a26b032 | 4343 | |
8079805d RK |
4344 | if (rtx_equal_p (true, false) && ! side_effects_p (cond)) |
4345 | return true; | |
1a26b032 | 4346 | |
5be669c7 RK |
4347 | /* Convert a == b ? b : a to "a". */ |
4348 | if (true_code == EQ && ! side_effects_p (cond) | |
4349 | && rtx_equal_p (XEXP (cond, 0), false) | |
4350 | && rtx_equal_p (XEXP (cond, 1), true)) | |
4351 | return false; | |
4352 | else if (true_code == NE && ! side_effects_p (cond) | |
4353 | && rtx_equal_p (XEXP (cond, 0), true) | |
4354 | && rtx_equal_p (XEXP (cond, 1), false)) | |
4355 | return true; | |
4356 | ||
8079805d RK |
4357 | /* Look for cases where we have (abs x) or (neg (abs X)). */ |
4358 | ||
4359 | if (GET_MODE_CLASS (mode) == MODE_INT | |
4360 | && GET_CODE (false) == NEG | |
4361 | && rtx_equal_p (true, XEXP (false, 0)) | |
4362 | && comparison_p | |
4363 | && rtx_equal_p (true, XEXP (cond, 0)) | |
4364 | && ! side_effects_p (true)) | |
4365 | switch (true_code) | |
4366 | { | |
4367 | case GT: | |
4368 | case GE: | |
0c1c8ea6 | 4369 | return gen_unary (ABS, mode, mode, true); |
8079805d RK |
4370 | case LT: |
4371 | case LE: | |
0c1c8ea6 | 4372 | return gen_unary (NEG, mode, mode, gen_unary (ABS, mode, mode, true)); |
e9a25f70 JL |
4373 | default: |
4374 | break; | |
8079805d RK |
4375 | } |
4376 | ||
4377 | /* Look for MIN or MAX. */ | |
4378 | ||
34c8be72 | 4379 | if ((! FLOAT_MODE_P (mode) || flag_fast_math) |
8079805d RK |
4380 | && comparison_p |
4381 | && rtx_equal_p (XEXP (cond, 0), true) | |
4382 | && rtx_equal_p (XEXP (cond, 1), false) | |
4383 | && ! side_effects_p (cond)) | |
4384 | switch (true_code) | |
4385 | { | |
4386 | case GE: | |
4387 | case GT: | |
4388 | return gen_binary (SMAX, mode, true, false); | |
4389 | case LE: | |
4390 | case LT: | |
4391 | return gen_binary (SMIN, mode, true, false); | |
4392 | case GEU: | |
4393 | case GTU: | |
4394 | return gen_binary (UMAX, mode, true, false); | |
4395 | case LEU: | |
4396 | case LTU: | |
4397 | return gen_binary (UMIN, mode, true, false); | |
e9a25f70 JL |
4398 | default: |
4399 | break; | |
8079805d RK |
4400 | } |
4401 | ||
8079805d RK |
4402 | /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its |
4403 | second operand is zero, this can be done as (OP Z (mult COND C2)) where | |
4404 | C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or | |
4405 | SIGN_EXTEND as long as Z is already extended (so we don't destroy it). | |
4406 | We can do this kind of thing in some cases when STORE_FLAG_VALUE is | |
0802d516 | 4407 | neither 1 or -1, but it isn't worth checking for. */ |
8079805d | 4408 | |
0802d516 RK |
4409 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
4410 | && comparison_p && mode != VOIDmode && ! side_effects_p (x)) | |
8079805d RK |
4411 | { |
4412 | rtx t = make_compound_operation (true, SET); | |
4413 | rtx f = make_compound_operation (false, SET); | |
4414 | rtx cond_op0 = XEXP (cond, 0); | |
4415 | rtx cond_op1 = XEXP (cond, 1); | |
4416 | enum rtx_code op, extend_op = NIL; | |
4417 | enum machine_mode m = mode; | |
f24ad0e4 | 4418 | rtx z = 0, c1; |
8079805d | 4419 | |
8079805d RK |
4420 | if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS |
4421 | || GET_CODE (t) == IOR || GET_CODE (t) == XOR | |
4422 | || GET_CODE (t) == ASHIFT | |
4423 | || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT) | |
4424 | && rtx_equal_p (XEXP (t, 0), f)) | |
4425 | c1 = XEXP (t, 1), op = GET_CODE (t), z = f; | |
4426 | ||
4427 | /* If an identity-zero op is commutative, check whether there | |
0f41302f | 4428 | would be a match if we swapped the operands. */ |
8079805d RK |
4429 | else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR |
4430 | || GET_CODE (t) == XOR) | |
4431 | && rtx_equal_p (XEXP (t, 1), f)) | |
4432 | c1 = XEXP (t, 0), op = GET_CODE (t), z = f; | |
4433 | else if (GET_CODE (t) == SIGN_EXTEND | |
4434 | && (GET_CODE (XEXP (t, 0)) == PLUS | |
4435 | || GET_CODE (XEXP (t, 0)) == MINUS | |
4436 | || GET_CODE (XEXP (t, 0)) == IOR | |
4437 | || GET_CODE (XEXP (t, 0)) == XOR | |
4438 | || GET_CODE (XEXP (t, 0)) == ASHIFT | |
4439 | || GET_CODE (XEXP (t, 0)) == LSHIFTRT | |
4440 | || GET_CODE (XEXP (t, 0)) == ASHIFTRT) | |
4441 | && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG | |
4442 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 0)) | |
4443 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f) | |
4444 | && (num_sign_bit_copies (f, GET_MODE (f)) | |
4445 | > (GET_MODE_BITSIZE (mode) | |
4446 | - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 0)))))) | |
4447 | { | |
4448 | c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0)); | |
4449 | extend_op = SIGN_EXTEND; | |
4450 | m = GET_MODE (XEXP (t, 0)); | |
1a26b032 | 4451 | } |
8079805d RK |
4452 | else if (GET_CODE (t) == SIGN_EXTEND |
4453 | && (GET_CODE (XEXP (t, 0)) == PLUS | |
4454 | || GET_CODE (XEXP (t, 0)) == IOR | |
4455 | || GET_CODE (XEXP (t, 0)) == XOR) | |
4456 | && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG | |
4457 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 1)) | |
4458 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f) | |
4459 | && (num_sign_bit_copies (f, GET_MODE (f)) | |
4460 | > (GET_MODE_BITSIZE (mode) | |
4461 | - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 1)))))) | |
4462 | { | |
4463 | c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0)); | |
4464 | extend_op = SIGN_EXTEND; | |
4465 | m = GET_MODE (XEXP (t, 0)); | |
4466 | } | |
4467 | else if (GET_CODE (t) == ZERO_EXTEND | |
4468 | && (GET_CODE (XEXP (t, 0)) == PLUS | |
4469 | || GET_CODE (XEXP (t, 0)) == MINUS | |
4470 | || GET_CODE (XEXP (t, 0)) == IOR | |
4471 | || GET_CODE (XEXP (t, 0)) == XOR | |
4472 | || GET_CODE (XEXP (t, 0)) == ASHIFT | |
4473 | || GET_CODE (XEXP (t, 0)) == LSHIFTRT | |
4474 | || GET_CODE (XEXP (t, 0)) == ASHIFTRT) | |
4475 | && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG | |
4476 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
4477 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 0)) | |
4478 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f) | |
4479 | && ((nonzero_bits (f, GET_MODE (f)) | |
4480 | & ~ GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 0)))) | |
4481 | == 0)) | |
4482 | { | |
4483 | c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0)); | |
4484 | extend_op = ZERO_EXTEND; | |
4485 | m = GET_MODE (XEXP (t, 0)); | |
4486 | } | |
4487 | else if (GET_CODE (t) == ZERO_EXTEND | |
4488 | && (GET_CODE (XEXP (t, 0)) == PLUS | |
4489 | || GET_CODE (XEXP (t, 0)) == IOR | |
4490 | || GET_CODE (XEXP (t, 0)) == XOR) | |
4491 | && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG | |
4492 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
4493 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 1)) | |
4494 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f) | |
4495 | && ((nonzero_bits (f, GET_MODE (f)) | |
4496 | & ~ GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 1)))) | |
4497 | == 0)) | |
4498 | { | |
4499 | c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0)); | |
4500 | extend_op = ZERO_EXTEND; | |
4501 | m = GET_MODE (XEXP (t, 0)); | |
4502 | } | |
4503 | ||
4504 | if (z) | |
4505 | { | |
4506 | temp = subst (gen_binary (true_code, m, cond_op0, cond_op1), | |
4507 | pc_rtx, pc_rtx, 0, 0); | |
4508 | temp = gen_binary (MULT, m, temp, | |
4509 | gen_binary (MULT, m, c1, const_true_rtx)); | |
4510 | temp = subst (temp, pc_rtx, pc_rtx, 0, 0); | |
4511 | temp = gen_binary (op, m, gen_lowpart_for_combine (m, z), temp); | |
4512 | ||
4513 | if (extend_op != NIL) | |
0c1c8ea6 | 4514 | temp = gen_unary (extend_op, mode, m, temp); |
8079805d RK |
4515 | |
4516 | return temp; | |
4517 | } | |
4518 | } | |
224eeff2 | 4519 | |
8079805d RK |
4520 | /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or |
4521 | 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the | |
4522 | negation of a single bit, we can convert this operation to a shift. We | |
4523 | can actually do this more generally, but it doesn't seem worth it. */ | |
4524 | ||
4525 | if (true_code == NE && XEXP (cond, 1) == const0_rtx | |
4526 | && false == const0_rtx && GET_CODE (true) == CONST_INT | |
4527 | && ((1 == nonzero_bits (XEXP (cond, 0), mode) | |
4528 | && (i = exact_log2 (INTVAL (true))) >= 0) | |
4529 | || ((num_sign_bit_copies (XEXP (cond, 0), mode) | |
4530 | == GET_MODE_BITSIZE (mode)) | |
4531 | && (i = exact_log2 (- INTVAL (true))) >= 0))) | |
4532 | return | |
4533 | simplify_shift_const (NULL_RTX, ASHIFT, mode, | |
4534 | gen_lowpart_for_combine (mode, XEXP (cond, 0)), i); | |
230d793d | 4535 | |
8079805d RK |
4536 | return x; |
4537 | } | |
4538 | \f | |
4539 | /* Simplify X, a SET expression. Return the new expression. */ | |
230d793d | 4540 | |
8079805d RK |
4541 | static rtx |
4542 | simplify_set (x) | |
4543 | rtx x; | |
4544 | { | |
4545 | rtx src = SET_SRC (x); | |
4546 | rtx dest = SET_DEST (x); | |
4547 | enum machine_mode mode | |
4548 | = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest); | |
4549 | rtx other_insn; | |
4550 | rtx *cc_use; | |
4551 | ||
4552 | /* (set (pc) (return)) gets written as (return). */ | |
4553 | if (GET_CODE (dest) == PC && GET_CODE (src) == RETURN) | |
4554 | return src; | |
230d793d | 4555 | |
87e3e0c1 RK |
4556 | /* Now that we know for sure which bits of SRC we are using, see if we can |
4557 | simplify the expression for the object knowing that we only need the | |
4558 | low-order bits. */ | |
4559 | ||
4560 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
4561 | src = force_to_mode (src, mode, GET_MODE_MASK (mode), NULL_RTX, 0); | |
4562 | ||
8079805d RK |
4563 | /* If we are setting CC0 or if the source is a COMPARE, look for the use of |
4564 | the comparison result and try to simplify it unless we already have used | |
4565 | undobuf.other_insn. */ | |
4566 | if ((GET_CODE (src) == COMPARE | |
230d793d | 4567 | #ifdef HAVE_cc0 |
8079805d | 4568 | || dest == cc0_rtx |
230d793d | 4569 | #endif |
8079805d RK |
4570 | ) |
4571 | && (cc_use = find_single_use (dest, subst_insn, &other_insn)) != 0 | |
4572 | && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn) | |
4573 | && GET_RTX_CLASS (GET_CODE (*cc_use)) == '<' | |
c0d3ac4d | 4574 | && rtx_equal_p (XEXP (*cc_use, 0), dest)) |
8079805d RK |
4575 | { |
4576 | enum rtx_code old_code = GET_CODE (*cc_use); | |
4577 | enum rtx_code new_code; | |
4578 | rtx op0, op1; | |
4579 | int other_changed = 0; | |
4580 | enum machine_mode compare_mode = GET_MODE (dest); | |
4581 | ||
4582 | if (GET_CODE (src) == COMPARE) | |
4583 | op0 = XEXP (src, 0), op1 = XEXP (src, 1); | |
4584 | else | |
4585 | op0 = src, op1 = const0_rtx; | |
230d793d | 4586 | |
8079805d RK |
4587 | /* Simplify our comparison, if possible. */ |
4588 | new_code = simplify_comparison (old_code, &op0, &op1); | |
230d793d | 4589 | |
c141a106 | 4590 | #ifdef EXTRA_CC_MODES |
8079805d RK |
4591 | /* If this machine has CC modes other than CCmode, check to see if we |
4592 | need to use a different CC mode here. */ | |
4593 | compare_mode = SELECT_CC_MODE (new_code, op0, op1); | |
c141a106 | 4594 | #endif /* EXTRA_CC_MODES */ |
230d793d | 4595 | |
c141a106 | 4596 | #if !defined (HAVE_cc0) && defined (EXTRA_CC_MODES) |
8079805d RK |
4597 | /* If the mode changed, we have to change SET_DEST, the mode in the |
4598 | compare, and the mode in the place SET_DEST is used. If SET_DEST is | |
4599 | a hard register, just build new versions with the proper mode. If it | |
4600 | is a pseudo, we lose unless it is only time we set the pseudo, in | |
4601 | which case we can safely change its mode. */ | |
4602 | if (compare_mode != GET_MODE (dest)) | |
4603 | { | |
4604 | int regno = REGNO (dest); | |
38a448ca | 4605 | rtx new_dest = gen_rtx_REG (compare_mode, regno); |
8079805d RK |
4606 | |
4607 | if (regno < FIRST_PSEUDO_REGISTER | |
b1f21e0a | 4608 | || (REG_N_SETS (regno) == 1 && ! REG_USERVAR_P (dest))) |
230d793d | 4609 | { |
8079805d RK |
4610 | if (regno >= FIRST_PSEUDO_REGISTER) |
4611 | SUBST (regno_reg_rtx[regno], new_dest); | |
230d793d | 4612 | |
8079805d RK |
4613 | SUBST (SET_DEST (x), new_dest); |
4614 | SUBST (XEXP (*cc_use, 0), new_dest); | |
4615 | other_changed = 1; | |
230d793d | 4616 | |
8079805d | 4617 | dest = new_dest; |
230d793d | 4618 | } |
8079805d | 4619 | } |
230d793d RS |
4620 | #endif |
4621 | ||
8079805d RK |
4622 | /* If the code changed, we have to build a new comparison in |
4623 | undobuf.other_insn. */ | |
4624 | if (new_code != old_code) | |
4625 | { | |
4626 | unsigned HOST_WIDE_INT mask; | |
4627 | ||
4628 | SUBST (*cc_use, gen_rtx_combine (new_code, GET_MODE (*cc_use), | |
4629 | dest, const0_rtx)); | |
4630 | ||
4631 | /* If the only change we made was to change an EQ into an NE or | |
4632 | vice versa, OP0 has only one bit that might be nonzero, and OP1 | |
4633 | is zero, check if changing the user of the condition code will | |
4634 | produce a valid insn. If it won't, we can keep the original code | |
4635 | in that insn by surrounding our operation with an XOR. */ | |
4636 | ||
4637 | if (((old_code == NE && new_code == EQ) | |
4638 | || (old_code == EQ && new_code == NE)) | |
4639 | && ! other_changed && op1 == const0_rtx | |
4640 | && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT | |
4641 | && exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) >= 0) | |
230d793d | 4642 | { |
8079805d | 4643 | rtx pat = PATTERN (other_insn), note = 0; |
a29ca9db | 4644 | int scratches; |
230d793d | 4645 | |
a29ca9db | 4646 | if ((recog_for_combine (&pat, other_insn, ¬e, &scratches) < 0 |
8079805d RK |
4647 | && ! check_asm_operands (pat))) |
4648 | { | |
4649 | PUT_CODE (*cc_use, old_code); | |
4650 | other_insn = 0; | |
230d793d | 4651 | |
8079805d | 4652 | op0 = gen_binary (XOR, GET_MODE (op0), op0, GEN_INT (mask)); |
230d793d | 4653 | } |
230d793d RS |
4654 | } |
4655 | ||
8079805d RK |
4656 | other_changed = 1; |
4657 | } | |
4658 | ||
4659 | if (other_changed) | |
4660 | undobuf.other_insn = other_insn; | |
230d793d RS |
4661 | |
4662 | #ifdef HAVE_cc0 | |
8079805d RK |
4663 | /* If we are now comparing against zero, change our source if |
4664 | needed. If we do not use cc0, we always have a COMPARE. */ | |
4665 | if (op1 == const0_rtx && dest == cc0_rtx) | |
4666 | { | |
4667 | SUBST (SET_SRC (x), op0); | |
4668 | src = op0; | |
4669 | } | |
4670 | else | |
230d793d RS |
4671 | #endif |
4672 | ||
8079805d RK |
4673 | /* Otherwise, if we didn't previously have a COMPARE in the |
4674 | correct mode, we need one. */ | |
4675 | if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode) | |
4676 | { | |
4677 | SUBST (SET_SRC (x), | |
4678 | gen_rtx_combine (COMPARE, compare_mode, op0, op1)); | |
4679 | src = SET_SRC (x); | |
230d793d RS |
4680 | } |
4681 | else | |
4682 | { | |
8079805d RK |
4683 | /* Otherwise, update the COMPARE if needed. */ |
4684 | SUBST (XEXP (src, 0), op0); | |
4685 | SUBST (XEXP (src, 1), op1); | |
230d793d | 4686 | } |
8079805d RK |
4687 | } |
4688 | else | |
4689 | { | |
4690 | /* Get SET_SRC in a form where we have placed back any | |
4691 | compound expressions. Then do the checks below. */ | |
4692 | src = make_compound_operation (src, SET); | |
4693 | SUBST (SET_SRC (x), src); | |
4694 | } | |
230d793d | 4695 | |
8079805d RK |
4696 | /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation, |
4697 | and X being a REG or (subreg (reg)), we may be able to convert this to | |
4698 | (set (subreg:m2 x) (op)). | |
df62f951 | 4699 | |
8079805d RK |
4700 | We can always do this if M1 is narrower than M2 because that means that |
4701 | we only care about the low bits of the result. | |
df62f951 | 4702 | |
8079805d | 4703 | However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot |
9ec36da5 | 4704 | perform a narrower operation than requested since the high-order bits will |
8079805d RK |
4705 | be undefined. On machine where it is defined, this transformation is safe |
4706 | as long as M1 and M2 have the same number of words. */ | |
df62f951 | 4707 | |
8079805d RK |
4708 | if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src) |
4709 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (src))) != 'o' | |
4710 | && (((GET_MODE_SIZE (GET_MODE (src)) + (UNITS_PER_WORD - 1)) | |
4711 | / UNITS_PER_WORD) | |
4712 | == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
4713 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)) | |
8baf60bb | 4714 | #ifndef WORD_REGISTER_OPERATIONS |
8079805d RK |
4715 | && (GET_MODE_SIZE (GET_MODE (src)) |
4716 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))) | |
df62f951 | 4717 | #endif |
f507a070 RK |
4718 | #ifdef CLASS_CANNOT_CHANGE_SIZE |
4719 | && ! (GET_CODE (dest) == REG && REGNO (dest) < FIRST_PSEUDO_REGISTER | |
4720 | && (TEST_HARD_REG_BIT | |
4721 | (reg_class_contents[(int) CLASS_CANNOT_CHANGE_SIZE], | |
4722 | REGNO (dest))) | |
4723 | && (GET_MODE_SIZE (GET_MODE (src)) | |
4724 | != GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))) | |
4725 | #endif | |
8079805d RK |
4726 | && (GET_CODE (dest) == REG |
4727 | || (GET_CODE (dest) == SUBREG | |
4728 | && GET_CODE (SUBREG_REG (dest)) == REG))) | |
4729 | { | |
4730 | SUBST (SET_DEST (x), | |
4731 | gen_lowpart_for_combine (GET_MODE (SUBREG_REG (src)), | |
4732 | dest)); | |
4733 | SUBST (SET_SRC (x), SUBREG_REG (src)); | |
4734 | ||
4735 | src = SET_SRC (x), dest = SET_DEST (x); | |
4736 | } | |
df62f951 | 4737 | |
8baf60bb | 4738 | #ifdef LOAD_EXTEND_OP |
8079805d RK |
4739 | /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this |
4740 | would require a paradoxical subreg. Replace the subreg with a | |
0f41302f | 4741 | zero_extend to avoid the reload that would otherwise be required. */ |
8079805d RK |
4742 | |
4743 | if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src) | |
4744 | && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))) != NIL | |
4745 | && SUBREG_WORD (src) == 0 | |
4746 | && (GET_MODE_SIZE (GET_MODE (src)) | |
4747 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))) | |
4748 | && GET_CODE (SUBREG_REG (src)) == MEM) | |
4749 | { | |
4750 | SUBST (SET_SRC (x), | |
4751 | gen_rtx_combine (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))), | |
4752 | GET_MODE (src), XEXP (src, 0))); | |
4753 | ||
4754 | src = SET_SRC (x); | |
4755 | } | |
230d793d RS |
4756 | #endif |
4757 | ||
8079805d RK |
4758 | /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we |
4759 | are comparing an item known to be 0 or -1 against 0, use a logical | |
4760 | operation instead. Check for one of the arms being an IOR of the other | |
4761 | arm with some value. We compute three terms to be IOR'ed together. In | |
4762 | practice, at most two will be nonzero. Then we do the IOR's. */ | |
4763 | ||
4764 | if (GET_CODE (dest) != PC | |
4765 | && GET_CODE (src) == IF_THEN_ELSE | |
36b8d792 | 4766 | && GET_MODE_CLASS (GET_MODE (src)) == MODE_INT |
8079805d RK |
4767 | && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE) |
4768 | && XEXP (XEXP (src, 0), 1) == const0_rtx | |
6dd49058 | 4769 | && GET_MODE (src) == GET_MODE (XEXP (XEXP (src, 0), 0)) |
ea414472 DE |
4770 | #ifdef HAVE_conditional_move |
4771 | && ! can_conditionally_move_p (GET_MODE (src)) | |
4772 | #endif | |
8079805d RK |
4773 | && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0), |
4774 | GET_MODE (XEXP (XEXP (src, 0), 0))) | |
4775 | == GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (src, 0), 0)))) | |
4776 | && ! side_effects_p (src)) | |
4777 | { | |
4778 | rtx true = (GET_CODE (XEXP (src, 0)) == NE | |
4779 | ? XEXP (src, 1) : XEXP (src, 2)); | |
4780 | rtx false = (GET_CODE (XEXP (src, 0)) == NE | |
4781 | ? XEXP (src, 2) : XEXP (src, 1)); | |
4782 | rtx term1 = const0_rtx, term2, term3; | |
4783 | ||
4784 | if (GET_CODE (true) == IOR && rtx_equal_p (XEXP (true, 0), false)) | |
4785 | term1 = false, true = XEXP (true, 1), false = const0_rtx; | |
4786 | else if (GET_CODE (true) == IOR | |
4787 | && rtx_equal_p (XEXP (true, 1), false)) | |
4788 | term1 = false, true = XEXP (true, 0), false = const0_rtx; | |
4789 | else if (GET_CODE (false) == IOR | |
4790 | && rtx_equal_p (XEXP (false, 0), true)) | |
4791 | term1 = true, false = XEXP (false, 1), true = const0_rtx; | |
4792 | else if (GET_CODE (false) == IOR | |
4793 | && rtx_equal_p (XEXP (false, 1), true)) | |
4794 | term1 = true, false = XEXP (false, 0), true = const0_rtx; | |
4795 | ||
4796 | term2 = gen_binary (AND, GET_MODE (src), XEXP (XEXP (src, 0), 0), true); | |
4797 | term3 = gen_binary (AND, GET_MODE (src), | |
0c1c8ea6 | 4798 | gen_unary (NOT, GET_MODE (src), GET_MODE (src), |
8079805d RK |
4799 | XEXP (XEXP (src, 0), 0)), |
4800 | false); | |
4801 | ||
4802 | SUBST (SET_SRC (x), | |
4803 | gen_binary (IOR, GET_MODE (src), | |
4804 | gen_binary (IOR, GET_MODE (src), term1, term2), | |
4805 | term3)); | |
4806 | ||
4807 | src = SET_SRC (x); | |
4808 | } | |
230d793d | 4809 | |
246e00f2 RK |
4810 | /* If either SRC or DEST is a CLOBBER of (const_int 0), make this |
4811 | whole thing fail. */ | |
4812 | if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx) | |
4813 | return src; | |
4814 | else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx) | |
4815 | return dest; | |
4816 | else | |
4817 | /* Convert this into a field assignment operation, if possible. */ | |
4818 | return make_field_assignment (x); | |
8079805d RK |
4819 | } |
4820 | \f | |
4821 | /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified | |
4822 | result. LAST is nonzero if this is the last retry. */ | |
4823 | ||
4824 | static rtx | |
4825 | simplify_logical (x, last) | |
4826 | rtx x; | |
4827 | int last; | |
4828 | { | |
4829 | enum machine_mode mode = GET_MODE (x); | |
4830 | rtx op0 = XEXP (x, 0); | |
4831 | rtx op1 = XEXP (x, 1); | |
4832 | ||
4833 | switch (GET_CODE (x)) | |
4834 | { | |
230d793d | 4835 | case AND: |
8079805d RK |
4836 | /* Convert (A ^ B) & A to A & (~ B) since the latter is often a single |
4837 | insn (and may simplify more). */ | |
4838 | if (GET_CODE (op0) == XOR | |
4839 | && rtx_equal_p (XEXP (op0, 0), op1) | |
4840 | && ! side_effects_p (op1)) | |
0c1c8ea6 RK |
4841 | x = gen_binary (AND, mode, |
4842 | gen_unary (NOT, mode, mode, XEXP (op0, 1)), op1); | |
8079805d RK |
4843 | |
4844 | if (GET_CODE (op0) == XOR | |
4845 | && rtx_equal_p (XEXP (op0, 1), op1) | |
4846 | && ! side_effects_p (op1)) | |
0c1c8ea6 RK |
4847 | x = gen_binary (AND, mode, |
4848 | gen_unary (NOT, mode, mode, XEXP (op0, 0)), op1); | |
8079805d RK |
4849 | |
4850 | /* Similarly for (~ (A ^ B)) & A. */ | |
4851 | if (GET_CODE (op0) == NOT | |
4852 | && GET_CODE (XEXP (op0, 0)) == XOR | |
4853 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), op1) | |
4854 | && ! side_effects_p (op1)) | |
4855 | x = gen_binary (AND, mode, XEXP (XEXP (op0, 0), 1), op1); | |
4856 | ||
4857 | if (GET_CODE (op0) == NOT | |
4858 | && GET_CODE (XEXP (op0, 0)) == XOR | |
4859 | && rtx_equal_p (XEXP (XEXP (op0, 0), 1), op1) | |
4860 | && ! side_effects_p (op1)) | |
4861 | x = gen_binary (AND, mode, XEXP (XEXP (op0, 0), 0), op1); | |
4862 | ||
4863 | if (GET_CODE (op1) == CONST_INT) | |
230d793d | 4864 | { |
8079805d | 4865 | x = simplify_and_const_int (x, mode, op0, INTVAL (op1)); |
230d793d RS |
4866 | |
4867 | /* If we have (ior (and (X C1) C2)) and the next restart would be | |
4868 | the last, simplify this by making C1 as small as possible | |
0f41302f | 4869 | and then exit. */ |
8079805d RK |
4870 | if (last |
4871 | && GET_CODE (x) == IOR && GET_CODE (op0) == AND | |
4872 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
4873 | && GET_CODE (op1) == CONST_INT) | |
4874 | return gen_binary (IOR, mode, | |
4875 | gen_binary (AND, mode, XEXP (op0, 0), | |
4876 | GEN_INT (INTVAL (XEXP (op0, 1)) | |
4877 | & ~ INTVAL (op1))), op1); | |
230d793d RS |
4878 | |
4879 | if (GET_CODE (x) != AND) | |
8079805d | 4880 | return x; |
0e32506c RK |
4881 | |
4882 | if (GET_RTX_CLASS (GET_CODE (x)) == 'c' | |
4883 | || GET_RTX_CLASS (GET_CODE (x)) == '2') | |
4884 | op0 = XEXP (x, 0), op1 = XEXP (x, 1); | |
230d793d RS |
4885 | } |
4886 | ||
4887 | /* Convert (A | B) & A to A. */ | |
8079805d RK |
4888 | if (GET_CODE (op0) == IOR |
4889 | && (rtx_equal_p (XEXP (op0, 0), op1) | |
4890 | || rtx_equal_p (XEXP (op0, 1), op1)) | |
4891 | && ! side_effects_p (XEXP (op0, 0)) | |
4892 | && ! side_effects_p (XEXP (op0, 1))) | |
4893 | return op1; | |
230d793d | 4894 | |
d0ab8cd3 | 4895 | /* In the following group of tests (and those in case IOR below), |
230d793d RS |
4896 | we start with some combination of logical operations and apply |
4897 | the distributive law followed by the inverse distributive law. | |
4898 | Most of the time, this results in no change. However, if some of | |
4899 | the operands are the same or inverses of each other, simplifications | |
4900 | will result. | |
4901 | ||
4902 | For example, (and (ior A B) (not B)) can occur as the result of | |
4903 | expanding a bit field assignment. When we apply the distributive | |
4904 | law to this, we get (ior (and (A (not B))) (and (B (not B)))), | |
8079805d | 4905 | which then simplifies to (and (A (not B))). |
230d793d | 4906 | |
8079805d | 4907 | If we have (and (ior A B) C), apply the distributive law and then |
230d793d RS |
4908 | the inverse distributive law to see if things simplify. */ |
4909 | ||
8079805d | 4910 | if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR) |
230d793d RS |
4911 | { |
4912 | x = apply_distributive_law | |
8079805d RK |
4913 | (gen_binary (GET_CODE (op0), mode, |
4914 | gen_binary (AND, mode, XEXP (op0, 0), op1), | |
4915 | gen_binary (AND, mode, XEXP (op0, 1), op1))); | |
230d793d | 4916 | if (GET_CODE (x) != AND) |
8079805d | 4917 | return x; |
230d793d RS |
4918 | } |
4919 | ||
8079805d RK |
4920 | if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR) |
4921 | return apply_distributive_law | |
4922 | (gen_binary (GET_CODE (op1), mode, | |
4923 | gen_binary (AND, mode, XEXP (op1, 0), op0), | |
4924 | gen_binary (AND, mode, XEXP (op1, 1), op0))); | |
230d793d RS |
4925 | |
4926 | /* Similarly, taking advantage of the fact that | |
4927 | (and (not A) (xor B C)) == (xor (ior A B) (ior A C)) */ | |
4928 | ||
8079805d RK |
4929 | if (GET_CODE (op0) == NOT && GET_CODE (op1) == XOR) |
4930 | return apply_distributive_law | |
4931 | (gen_binary (XOR, mode, | |
4932 | gen_binary (IOR, mode, XEXP (op0, 0), XEXP (op1, 0)), | |
4933 | gen_binary (IOR, mode, XEXP (op0, 0), XEXP (op1, 1)))); | |
230d793d | 4934 | |
8079805d RK |
4935 | else if (GET_CODE (op1) == NOT && GET_CODE (op0) == XOR) |
4936 | return apply_distributive_law | |
4937 | (gen_binary (XOR, mode, | |
4938 | gen_binary (IOR, mode, XEXP (op1, 0), XEXP (op0, 0)), | |
4939 | gen_binary (IOR, mode, XEXP (op1, 0), XEXP (op0, 1)))); | |
230d793d RS |
4940 | break; |
4941 | ||
4942 | case IOR: | |
951553af | 4943 | /* (ior A C) is C if all bits of A that might be nonzero are on in C. */ |
8079805d | 4944 | if (GET_CODE (op1) == CONST_INT |
ac49a949 | 4945 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
8079805d RK |
4946 | && (nonzero_bits (op0, mode) & ~ INTVAL (op1)) == 0) |
4947 | return op1; | |
d0ab8cd3 | 4948 | |
230d793d | 4949 | /* Convert (A & B) | A to A. */ |
8079805d RK |
4950 | if (GET_CODE (op0) == AND |
4951 | && (rtx_equal_p (XEXP (op0, 0), op1) | |
4952 | || rtx_equal_p (XEXP (op0, 1), op1)) | |
4953 | && ! side_effects_p (XEXP (op0, 0)) | |
4954 | && ! side_effects_p (XEXP (op0, 1))) | |
4955 | return op1; | |
230d793d RS |
4956 | |
4957 | /* If we have (ior (and A B) C), apply the distributive law and then | |
4958 | the inverse distributive law to see if things simplify. */ | |
4959 | ||
8079805d | 4960 | if (GET_CODE (op0) == AND) |
230d793d RS |
4961 | { |
4962 | x = apply_distributive_law | |
4963 | (gen_binary (AND, mode, | |
8079805d RK |
4964 | gen_binary (IOR, mode, XEXP (op0, 0), op1), |
4965 | gen_binary (IOR, mode, XEXP (op0, 1), op1))); | |
230d793d RS |
4966 | |
4967 | if (GET_CODE (x) != IOR) | |
8079805d | 4968 | return x; |
230d793d RS |
4969 | } |
4970 | ||
8079805d | 4971 | if (GET_CODE (op1) == AND) |
230d793d RS |
4972 | { |
4973 | x = apply_distributive_law | |
4974 | (gen_binary (AND, mode, | |
8079805d RK |
4975 | gen_binary (IOR, mode, XEXP (op1, 0), op0), |
4976 | gen_binary (IOR, mode, XEXP (op1, 1), op0))); | |
230d793d RS |
4977 | |
4978 | if (GET_CODE (x) != IOR) | |
8079805d | 4979 | return x; |
230d793d RS |
4980 | } |
4981 | ||
4982 | /* Convert (ior (ashift A CX) (lshiftrt A CY)) where CX+CY equals the | |
4983 | mode size to (rotate A CX). */ | |
4984 | ||
8079805d RK |
4985 | if (((GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT) |
4986 | || (GET_CODE (op1) == ASHIFT && GET_CODE (op0) == LSHIFTRT)) | |
4987 | && rtx_equal_p (XEXP (op0, 0), XEXP (op1, 0)) | |
4988 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
4989 | && GET_CODE (XEXP (op1, 1)) == CONST_INT | |
4990 | && (INTVAL (XEXP (op0, 1)) + INTVAL (XEXP (op1, 1)) | |
230d793d | 4991 | == GET_MODE_BITSIZE (mode))) |
38a448ca RH |
4992 | return gen_rtx_ROTATE (mode, XEXP (op0, 0), |
4993 | (GET_CODE (op0) == ASHIFT | |
4994 | ? XEXP (op0, 1) : XEXP (op1, 1))); | |
230d793d | 4995 | |
71923da7 RK |
4996 | /* If OP0 is (ashiftrt (plus ...) C), it might actually be |
4997 | a (sign_extend (plus ...)). If so, OP1 is a CONST_INT, and the PLUS | |
4998 | does not affect any of the bits in OP1, it can really be done | |
4999 | as a PLUS and we can associate. We do this by seeing if OP1 | |
5000 | can be safely shifted left C bits. */ | |
5001 | if (GET_CODE (op1) == CONST_INT && GET_CODE (op0) == ASHIFTRT | |
5002 | && GET_CODE (XEXP (op0, 0)) == PLUS | |
5003 | && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT | |
5004 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
5005 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT) | |
5006 | { | |
5007 | int count = INTVAL (XEXP (op0, 1)); | |
5008 | HOST_WIDE_INT mask = INTVAL (op1) << count; | |
5009 | ||
5010 | if (mask >> count == INTVAL (op1) | |
5011 | && (mask & nonzero_bits (XEXP (op0, 0), mode)) == 0) | |
5012 | { | |
5013 | SUBST (XEXP (XEXP (op0, 0), 1), | |
5014 | GEN_INT (INTVAL (XEXP (XEXP (op0, 0), 1)) | mask)); | |
5015 | return op0; | |
5016 | } | |
5017 | } | |
230d793d RS |
5018 | break; |
5019 | ||
5020 | case XOR: | |
5021 | /* Convert (XOR (NOT x) (NOT y)) to (XOR x y). | |
5022 | Also convert (XOR (NOT x) y) to (NOT (XOR x y)), similarly for | |
5023 | (NOT y). */ | |
5024 | { | |
5025 | int num_negated = 0; | |
230d793d | 5026 | |
8079805d RK |
5027 | if (GET_CODE (op0) == NOT) |
5028 | num_negated++, op0 = XEXP (op0, 0); | |
5029 | if (GET_CODE (op1) == NOT) | |
5030 | num_negated++, op1 = XEXP (op1, 0); | |
230d793d RS |
5031 | |
5032 | if (num_negated == 2) | |
5033 | { | |
8079805d RK |
5034 | SUBST (XEXP (x, 0), op0); |
5035 | SUBST (XEXP (x, 1), op1); | |
230d793d RS |
5036 | } |
5037 | else if (num_negated == 1) | |
0c1c8ea6 | 5038 | return gen_unary (NOT, mode, mode, gen_binary (XOR, mode, op0, op1)); |
230d793d RS |
5039 | } |
5040 | ||
5041 | /* Convert (xor (and A B) B) to (and (not A) B). The latter may | |
5042 | correspond to a machine insn or result in further simplifications | |
5043 | if B is a constant. */ | |
5044 | ||
8079805d RK |
5045 | if (GET_CODE (op0) == AND |
5046 | && rtx_equal_p (XEXP (op0, 1), op1) | |
5047 | && ! side_effects_p (op1)) | |
0c1c8ea6 RK |
5048 | return gen_binary (AND, mode, |
5049 | gen_unary (NOT, mode, mode, XEXP (op0, 0)), | |
8079805d | 5050 | op1); |
230d793d | 5051 | |
8079805d RK |
5052 | else if (GET_CODE (op0) == AND |
5053 | && rtx_equal_p (XEXP (op0, 0), op1) | |
5054 | && ! side_effects_p (op1)) | |
0c1c8ea6 RK |
5055 | return gen_binary (AND, mode, |
5056 | gen_unary (NOT, mode, mode, XEXP (op0, 1)), | |
8079805d | 5057 | op1); |
230d793d | 5058 | |
230d793d | 5059 | /* (xor (comparison foo bar) (const_int 1)) can become the reversed |
0802d516 RK |
5060 | comparison if STORE_FLAG_VALUE is 1. */ |
5061 | if (STORE_FLAG_VALUE == 1 | |
5062 | && op1 == const1_rtx | |
8079805d RK |
5063 | && GET_RTX_CLASS (GET_CODE (op0)) == '<' |
5064 | && reversible_comparison_p (op0)) | |
5065 | return gen_rtx_combine (reverse_condition (GET_CODE (op0)), | |
5066 | mode, XEXP (op0, 0), XEXP (op0, 1)); | |
500c518b RK |
5067 | |
5068 | /* (lshiftrt foo C) where C is the number of bits in FOO minus 1 | |
5069 | is (lt foo (const_int 0)), so we can perform the above | |
0802d516 | 5070 | simplification if STORE_FLAG_VALUE is 1. */ |
500c518b | 5071 | |
0802d516 RK |
5072 | if (STORE_FLAG_VALUE == 1 |
5073 | && op1 == const1_rtx | |
8079805d RK |
5074 | && GET_CODE (op0) == LSHIFTRT |
5075 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
5076 | && INTVAL (XEXP (op0, 1)) == GET_MODE_BITSIZE (mode) - 1) | |
5077 | return gen_rtx_combine (GE, mode, XEXP (op0, 0), const0_rtx); | |
230d793d RS |
5078 | |
5079 | /* (xor (comparison foo bar) (const_int sign-bit)) | |
5080 | when STORE_FLAG_VALUE is the sign bit. */ | |
5f4f0e22 | 5081 | if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
0802d516 | 5082 | && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode)) |
5f4f0e22 | 5083 | == (HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1)) |
8079805d RK |
5084 | && op1 == const_true_rtx |
5085 | && GET_RTX_CLASS (GET_CODE (op0)) == '<' | |
5086 | && reversible_comparison_p (op0)) | |
5087 | return gen_rtx_combine (reverse_condition (GET_CODE (op0)), | |
5088 | mode, XEXP (op0, 0), XEXP (op0, 1)); | |
230d793d | 5089 | break; |
e9a25f70 JL |
5090 | |
5091 | default: | |
5092 | abort (); | |
230d793d RS |
5093 | } |
5094 | ||
5095 | return x; | |
5096 | } | |
5097 | \f | |
5098 | /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound | |
5099 | operations" because they can be replaced with two more basic operations. | |
5100 | ZERO_EXTEND is also considered "compound" because it can be replaced with | |
5101 | an AND operation, which is simpler, though only one operation. | |
5102 | ||
5103 | The function expand_compound_operation is called with an rtx expression | |
5104 | and will convert it to the appropriate shifts and AND operations, | |
5105 | simplifying at each stage. | |
5106 | ||
5107 | The function make_compound_operation is called to convert an expression | |
5108 | consisting of shifts and ANDs into the equivalent compound expression. | |
5109 | It is the inverse of this function, loosely speaking. */ | |
5110 | ||
5111 | static rtx | |
5112 | expand_compound_operation (x) | |
5113 | rtx x; | |
5114 | { | |
5115 | int pos = 0, len; | |
5116 | int unsignedp = 0; | |
5117 | int modewidth; | |
5118 | rtx tem; | |
5119 | ||
5120 | switch (GET_CODE (x)) | |
5121 | { | |
5122 | case ZERO_EXTEND: | |
5123 | unsignedp = 1; | |
5124 | case SIGN_EXTEND: | |
75473182 RS |
5125 | /* We can't necessarily use a const_int for a multiword mode; |
5126 | it depends on implicitly extending the value. | |
5127 | Since we don't know the right way to extend it, | |
5128 | we can't tell whether the implicit way is right. | |
5129 | ||
5130 | Even for a mode that is no wider than a const_int, | |
5131 | we can't win, because we need to sign extend one of its bits through | |
5132 | the rest of it, and we don't know which bit. */ | |
230d793d | 5133 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) |
75473182 | 5134 | return x; |
230d793d | 5135 | |
8079805d RK |
5136 | /* Return if (subreg:MODE FROM 0) is not a safe replacement for |
5137 | (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM | |
5138 | because (SUBREG (MEM...)) is guaranteed to cause the MEM to be | |
5139 | reloaded. If not for that, MEM's would very rarely be safe. | |
5140 | ||
5141 | Reject MODEs bigger than a word, because we might not be able | |
5142 | to reference a two-register group starting with an arbitrary register | |
5143 | (and currently gen_lowpart might crash for a SUBREG). */ | |
5144 | ||
5145 | if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) > UNITS_PER_WORD) | |
230d793d RS |
5146 | return x; |
5147 | ||
5148 | len = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))); | |
5149 | /* If the inner object has VOIDmode (the only way this can happen | |
5150 | is if it is a ASM_OPERANDS), we can't do anything since we don't | |
5151 | know how much masking to do. */ | |
5152 | if (len == 0) | |
5153 | return x; | |
5154 | ||
5155 | break; | |
5156 | ||
5157 | case ZERO_EXTRACT: | |
5158 | unsignedp = 1; | |
5159 | case SIGN_EXTRACT: | |
5160 | /* If the operand is a CLOBBER, just return it. */ | |
5161 | if (GET_CODE (XEXP (x, 0)) == CLOBBER) | |
5162 | return XEXP (x, 0); | |
5163 | ||
5164 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
5165 | || GET_CODE (XEXP (x, 2)) != CONST_INT | |
5166 | || GET_MODE (XEXP (x, 0)) == VOIDmode) | |
5167 | return x; | |
5168 | ||
5169 | len = INTVAL (XEXP (x, 1)); | |
5170 | pos = INTVAL (XEXP (x, 2)); | |
5171 | ||
5172 | /* If this goes outside the object being extracted, replace the object | |
5173 | with a (use (mem ...)) construct that only combine understands | |
5174 | and is used only for this purpose. */ | |
5175 | if (len + pos > GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))) | |
38a448ca | 5176 | SUBST (XEXP (x, 0), gen_rtx_USE (GET_MODE (x), XEXP (x, 0))); |
230d793d | 5177 | |
f76b9db2 ILT |
5178 | if (BITS_BIG_ENDIAN) |
5179 | pos = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - len - pos; | |
5180 | ||
230d793d RS |
5181 | break; |
5182 | ||
5183 | default: | |
5184 | return x; | |
5185 | } | |
5186 | ||
0f13a422 ILT |
5187 | /* We can optimize some special cases of ZERO_EXTEND. */ |
5188 | if (GET_CODE (x) == ZERO_EXTEND) | |
5189 | { | |
5190 | /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we | |
5191 | know that the last value didn't have any inappropriate bits | |
5192 | set. */ | |
5193 | if (GET_CODE (XEXP (x, 0)) == TRUNCATE | |
5194 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x) | |
5195 | && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT | |
5196 | && (nonzero_bits (XEXP (XEXP (x, 0), 0), GET_MODE (x)) | |
5197 | & ~ GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0) | |
5198 | return XEXP (XEXP (x, 0), 0); | |
5199 | ||
5200 | /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */ | |
5201 | if (GET_CODE (XEXP (x, 0)) == SUBREG | |
5202 | && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x) | |
5203 | && subreg_lowpart_p (XEXP (x, 0)) | |
5204 | && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT | |
5205 | && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), GET_MODE (x)) | |
5206 | & ~ GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))) == 0) | |
5207 | return SUBREG_REG (XEXP (x, 0)); | |
5208 | ||
5209 | /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo | |
5210 | is a comparison and STORE_FLAG_VALUE permits. This is like | |
5211 | the first case, but it works even when GET_MODE (x) is larger | |
5212 | than HOST_WIDE_INT. */ | |
5213 | if (GET_CODE (XEXP (x, 0)) == TRUNCATE | |
5214 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x) | |
5215 | && GET_RTX_CLASS (GET_CODE (XEXP (XEXP (x, 0), 0))) == '<' | |
5216 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
5217 | <= HOST_BITS_PER_WIDE_INT) | |
5218 | && ((HOST_WIDE_INT) STORE_FLAG_VALUE | |
5219 | & ~ GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0) | |
5220 | return XEXP (XEXP (x, 0), 0); | |
5221 | ||
5222 | /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */ | |
5223 | if (GET_CODE (XEXP (x, 0)) == SUBREG | |
5224 | && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x) | |
5225 | && subreg_lowpart_p (XEXP (x, 0)) | |
5226 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (XEXP (x, 0)))) == '<' | |
5227 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
5228 | <= HOST_BITS_PER_WIDE_INT) | |
5229 | && ((HOST_WIDE_INT) STORE_FLAG_VALUE | |
5230 | & ~ GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0) | |
5231 | return SUBREG_REG (XEXP (x, 0)); | |
5232 | ||
5233 | /* If sign extension is cheaper than zero extension, then use it | |
5234 | if we know that no extraneous bits are set, and that the high | |
5235 | bit is not set. */ | |
5236 | if (flag_expensive_optimizations | |
5237 | && ((GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT | |
5238 | && ((nonzero_bits (XEXP (x, 0), GET_MODE (x)) | |
5239 | & ~ (((unsigned HOST_WIDE_INT) | |
5240 | GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) | |
5241 | >> 1)) | |
5242 | == 0)) | |
5243 | || (GET_RTX_CLASS (GET_CODE (XEXP (x, 0))) == '<' | |
5244 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
5245 | <= HOST_BITS_PER_WIDE_INT) | |
5246 | && (((HOST_WIDE_INT) STORE_FLAG_VALUE | |
5247 | & ~ (((unsigned HOST_WIDE_INT) | |
5248 | GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) | |
5249 | >> 1)) | |
5250 | == 0)))) | |
5251 | { | |
38a448ca | 5252 | rtx temp = gen_rtx_SIGN_EXTEND (GET_MODE (x), XEXP (x, 0)); |
0f13a422 ILT |
5253 | |
5254 | if (rtx_cost (temp, SET) < rtx_cost (x, SET)) | |
5255 | return expand_compound_operation (temp); | |
5256 | } | |
5257 | } | |
5258 | ||
230d793d RS |
5259 | /* If we reach here, we want to return a pair of shifts. The inner |
5260 | shift is a left shift of BITSIZE - POS - LEN bits. The outer | |
5261 | shift is a right shift of BITSIZE - LEN bits. It is arithmetic or | |
5262 | logical depending on the value of UNSIGNEDP. | |
5263 | ||
5264 | If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be | |
5265 | converted into an AND of a shift. | |
5266 | ||
5267 | We must check for the case where the left shift would have a negative | |
5268 | count. This can happen in a case like (x >> 31) & 255 on machines | |
5269 | that can't shift by a constant. On those machines, we would first | |
5270 | combine the shift with the AND to produce a variable-position | |
5271 | extraction. Then the constant of 31 would be substituted in to produce | |
5272 | a such a position. */ | |
5273 | ||
5274 | modewidth = GET_MODE_BITSIZE (GET_MODE (x)); | |
5275 | if (modewidth >= pos - len) | |
5f4f0e22 | 5276 | tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT, |
230d793d | 5277 | GET_MODE (x), |
5f4f0e22 CH |
5278 | simplify_shift_const (NULL_RTX, ASHIFT, |
5279 | GET_MODE (x), | |
230d793d RS |
5280 | XEXP (x, 0), |
5281 | modewidth - pos - len), | |
5282 | modewidth - len); | |
5283 | ||
5f4f0e22 CH |
5284 | else if (unsignedp && len < HOST_BITS_PER_WIDE_INT) |
5285 | tem = simplify_and_const_int (NULL_RTX, GET_MODE (x), | |
5286 | simplify_shift_const (NULL_RTX, LSHIFTRT, | |
230d793d RS |
5287 | GET_MODE (x), |
5288 | XEXP (x, 0), pos), | |
5f4f0e22 | 5289 | ((HOST_WIDE_INT) 1 << len) - 1); |
230d793d RS |
5290 | else |
5291 | /* Any other cases we can't handle. */ | |
5292 | return x; | |
5293 | ||
5294 | ||
5295 | /* If we couldn't do this for some reason, return the original | |
5296 | expression. */ | |
5297 | if (GET_CODE (tem) == CLOBBER) | |
5298 | return x; | |
5299 | ||
5300 | return tem; | |
5301 | } | |
5302 | \f | |
5303 | /* X is a SET which contains an assignment of one object into | |
5304 | a part of another (such as a bit-field assignment, STRICT_LOW_PART, | |
5305 | or certain SUBREGS). If possible, convert it into a series of | |
5306 | logical operations. | |
5307 | ||
5308 | We half-heartedly support variable positions, but do not at all | |
5309 | support variable lengths. */ | |
5310 | ||
5311 | static rtx | |
5312 | expand_field_assignment (x) | |
5313 | rtx x; | |
5314 | { | |
5315 | rtx inner; | |
0f41302f | 5316 | rtx pos; /* Always counts from low bit. */ |
230d793d RS |
5317 | int len; |
5318 | rtx mask; | |
5319 | enum machine_mode compute_mode; | |
5320 | ||
5321 | /* Loop until we find something we can't simplify. */ | |
5322 | while (1) | |
5323 | { | |
5324 | if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART | |
5325 | && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG) | |
5326 | { | |
5327 | inner = SUBREG_REG (XEXP (SET_DEST (x), 0)); | |
5328 | len = GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))); | |
4d9cfc7b | 5329 | pos = GEN_INT (BITS_PER_WORD * SUBREG_WORD (XEXP (SET_DEST (x), 0))); |
230d793d RS |
5330 | } |
5331 | else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT | |
5332 | && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT) | |
5333 | { | |
5334 | inner = XEXP (SET_DEST (x), 0); | |
5335 | len = INTVAL (XEXP (SET_DEST (x), 1)); | |
5336 | pos = XEXP (SET_DEST (x), 2); | |
5337 | ||
5338 | /* If the position is constant and spans the width of INNER, | |
5339 | surround INNER with a USE to indicate this. */ | |
5340 | if (GET_CODE (pos) == CONST_INT | |
5341 | && INTVAL (pos) + len > GET_MODE_BITSIZE (GET_MODE (inner))) | |
38a448ca | 5342 | inner = gen_rtx_USE (GET_MODE (SET_DEST (x)), inner); |
230d793d | 5343 | |
f76b9db2 ILT |
5344 | if (BITS_BIG_ENDIAN) |
5345 | { | |
5346 | if (GET_CODE (pos) == CONST_INT) | |
5347 | pos = GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) - len | |
5348 | - INTVAL (pos)); | |
5349 | else if (GET_CODE (pos) == MINUS | |
5350 | && GET_CODE (XEXP (pos, 1)) == CONST_INT | |
5351 | && (INTVAL (XEXP (pos, 1)) | |
5352 | == GET_MODE_BITSIZE (GET_MODE (inner)) - len)) | |
5353 | /* If position is ADJUST - X, new position is X. */ | |
5354 | pos = XEXP (pos, 0); | |
5355 | else | |
5356 | pos = gen_binary (MINUS, GET_MODE (pos), | |
5357 | GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) | |
5358 | - len), | |
5359 | pos); | |
5360 | } | |
230d793d RS |
5361 | } |
5362 | ||
5363 | /* A SUBREG between two modes that occupy the same numbers of words | |
5364 | can be done by moving the SUBREG to the source. */ | |
5365 | else if (GET_CODE (SET_DEST (x)) == SUBREG | |
5366 | && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x))) | |
5367 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD) | |
5368 | == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x)))) | |
5369 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))) | |
5370 | { | |
38a448ca RH |
5371 | x = gen_rtx_SET (VOIDmode, SUBREG_REG (SET_DEST (x)), |
5372 | gen_lowpart_for_combine (GET_MODE (SUBREG_REG (SET_DEST (x))), | |
5373 | SET_SRC (x))); | |
230d793d RS |
5374 | continue; |
5375 | } | |
5376 | else | |
5377 | break; | |
5378 | ||
5379 | while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner)) | |
5380 | inner = SUBREG_REG (inner); | |
5381 | ||
5382 | compute_mode = GET_MODE (inner); | |
5383 | ||
5384 | /* Compute a mask of LEN bits, if we can do this on the host machine. */ | |
5f4f0e22 CH |
5385 | if (len < HOST_BITS_PER_WIDE_INT) |
5386 | mask = GEN_INT (((HOST_WIDE_INT) 1 << len) - 1); | |
230d793d RS |
5387 | else |
5388 | break; | |
5389 | ||
5390 | /* Now compute the equivalent expression. Make a copy of INNER | |
5391 | for the SET_DEST in case it is a MEM into which we will substitute; | |
5392 | we don't want shared RTL in that case. */ | |
38a448ca RH |
5393 | x = gen_rtx_SET (VOIDmode, copy_rtx (inner), |
5394 | gen_binary (IOR, compute_mode, | |
5395 | gen_binary (AND, compute_mode, | |
5396 | gen_unary (NOT, compute_mode, | |
5397 | compute_mode, | |
5398 | gen_binary (ASHIFT, | |
5399 | compute_mode, | |
5400 | mask, pos)), | |
5401 | inner), | |
5402 | gen_binary (ASHIFT, compute_mode, | |
5403 | gen_binary (AND, compute_mode, | |
5404 | gen_lowpart_for_combine | |
5405 | (compute_mode, | |
5406 | SET_SRC (x)), | |
5407 | mask), | |
5408 | pos))); | |
230d793d RS |
5409 | } |
5410 | ||
5411 | return x; | |
5412 | } | |
5413 | \f | |
8999a12e RK |
5414 | /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero, |
5415 | it is an RTX that represents a variable starting position; otherwise, | |
5416 | POS is the (constant) starting bit position (counted from the LSB). | |
230d793d RS |
5417 | |
5418 | INNER may be a USE. This will occur when we started with a bitfield | |
5419 | that went outside the boundary of the object in memory, which is | |
5420 | allowed on most machines. To isolate this case, we produce a USE | |
5421 | whose mode is wide enough and surround the MEM with it. The only | |
5422 | code that understands the USE is this routine. If it is not removed, | |
5423 | it will cause the resulting insn not to match. | |
5424 | ||
5425 | UNSIGNEDP is non-zero for an unsigned reference and zero for a | |
5426 | signed reference. | |
5427 | ||
5428 | IN_DEST is non-zero if this is a reference in the destination of a | |
5429 | SET. This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If non-zero, | |
5430 | a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will | |
5431 | be used. | |
5432 | ||
5433 | IN_COMPARE is non-zero if we are in a COMPARE. This means that a | |
5434 | ZERO_EXTRACT should be built even for bits starting at bit 0. | |
5435 | ||
76184def DE |
5436 | MODE is the desired mode of the result (if IN_DEST == 0). |
5437 | ||
5438 | The result is an RTX for the extraction or NULL_RTX if the target | |
5439 | can't handle it. */ | |
230d793d RS |
5440 | |
5441 | static rtx | |
5442 | make_extraction (mode, inner, pos, pos_rtx, len, | |
5443 | unsignedp, in_dest, in_compare) | |
5444 | enum machine_mode mode; | |
5445 | rtx inner; | |
5446 | int pos; | |
5447 | rtx pos_rtx; | |
5448 | int len; | |
5449 | int unsignedp; | |
5450 | int in_dest, in_compare; | |
5451 | { | |
94b4b17a RS |
5452 | /* This mode describes the size of the storage area |
5453 | to fetch the overall value from. Within that, we | |
5454 | ignore the POS lowest bits, etc. */ | |
230d793d RS |
5455 | enum machine_mode is_mode = GET_MODE (inner); |
5456 | enum machine_mode inner_mode; | |
d7cd794f RK |
5457 | enum machine_mode wanted_inner_mode = byte_mode; |
5458 | enum machine_mode wanted_inner_reg_mode = word_mode; | |
230d793d RS |
5459 | enum machine_mode pos_mode = word_mode; |
5460 | enum machine_mode extraction_mode = word_mode; | |
5461 | enum machine_mode tmode = mode_for_size (len, MODE_INT, 1); | |
5462 | int spans_byte = 0; | |
5463 | rtx new = 0; | |
8999a12e | 5464 | rtx orig_pos_rtx = pos_rtx; |
6139ff20 | 5465 | int orig_pos; |
230d793d RS |
5466 | |
5467 | /* Get some information about INNER and get the innermost object. */ | |
5468 | if (GET_CODE (inner) == USE) | |
94b4b17a | 5469 | /* (use:SI (mem:QI foo)) stands for (mem:SI foo). */ |
230d793d RS |
5470 | /* We don't need to adjust the position because we set up the USE |
5471 | to pretend that it was a full-word object. */ | |
5472 | spans_byte = 1, inner = XEXP (inner, 0); | |
5473 | else if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner)) | |
94b4b17a RS |
5474 | { |
5475 | /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...), | |
5476 | consider just the QI as the memory to extract from. | |
5477 | The subreg adds or removes high bits; its mode is | |
5478 | irrelevant to the meaning of this extraction, | |
5479 | since POS and LEN count from the lsb. */ | |
5480 | if (GET_CODE (SUBREG_REG (inner)) == MEM) | |
5481 | is_mode = GET_MODE (SUBREG_REG (inner)); | |
5482 | inner = SUBREG_REG (inner); | |
5483 | } | |
230d793d RS |
5484 | |
5485 | inner_mode = GET_MODE (inner); | |
5486 | ||
5487 | if (pos_rtx && GET_CODE (pos_rtx) == CONST_INT) | |
8999a12e | 5488 | pos = INTVAL (pos_rtx), pos_rtx = 0; |
230d793d RS |
5489 | |
5490 | /* See if this can be done without an extraction. We never can if the | |
5491 | width of the field is not the same as that of some integer mode. For | |
5492 | registers, we can only avoid the extraction if the position is at the | |
5493 | low-order bit and this is either not in the destination or we have the | |
5494 | appropriate STRICT_LOW_PART operation available. | |
5495 | ||
5496 | For MEM, we can avoid an extract if the field starts on an appropriate | |
5497 | boundary and we can change the mode of the memory reference. However, | |
5498 | we cannot directly access the MEM if we have a USE and the underlying | |
5499 | MEM is not TMODE. This combination means that MEM was being used in a | |
5500 | context where bits outside its mode were being referenced; that is only | |
5501 | valid in bit-field insns. */ | |
5502 | ||
5503 | if (tmode != BLKmode | |
5504 | && ! (spans_byte && inner_mode != tmode) | |
4d9cfc7b RK |
5505 | && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0 |
5506 | && GET_CODE (inner) != MEM | |
230d793d | 5507 | && (! in_dest |
df62f951 RK |
5508 | || (GET_CODE (inner) == REG |
5509 | && (movstrict_optab->handlers[(int) tmode].insn_code | |
5510 | != CODE_FOR_nothing)))) | |
8999a12e | 5511 | || (GET_CODE (inner) == MEM && pos_rtx == 0 |
dfbe1b2f RK |
5512 | && (pos |
5513 | % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode) | |
5514 | : BITS_PER_UNIT)) == 0 | |
230d793d RS |
5515 | /* We can't do this if we are widening INNER_MODE (it |
5516 | may not be aligned, for one thing). */ | |
5517 | && GET_MODE_BITSIZE (inner_mode) >= GET_MODE_BITSIZE (tmode) | |
5518 | && (inner_mode == tmode | |
5519 | || (! mode_dependent_address_p (XEXP (inner, 0)) | |
5520 | && ! MEM_VOLATILE_P (inner)))))) | |
5521 | { | |
230d793d RS |
5522 | /* If INNER is a MEM, make a new MEM that encompasses just the desired |
5523 | field. If the original and current mode are the same, we need not | |
5524 | adjust the offset. Otherwise, we do if bytes big endian. | |
5525 | ||
4d9cfc7b RK |
5526 | If INNER is not a MEM, get a piece consisting of just the field |
5527 | of interest (in this case POS % BITS_PER_WORD must be 0). */ | |
230d793d RS |
5528 | |
5529 | if (GET_CODE (inner) == MEM) | |
5530 | { | |
94b4b17a RS |
5531 | int offset; |
5532 | /* POS counts from lsb, but make OFFSET count in memory order. */ | |
5533 | if (BYTES_BIG_ENDIAN) | |
5534 | offset = (GET_MODE_BITSIZE (is_mode) - len - pos) / BITS_PER_UNIT; | |
5535 | else | |
5536 | offset = pos / BITS_PER_UNIT; | |
230d793d | 5537 | |
38a448ca | 5538 | new = gen_rtx_MEM (tmode, plus_constant (XEXP (inner, 0), offset)); |
230d793d RS |
5539 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (inner); |
5540 | MEM_VOLATILE_P (new) = MEM_VOLATILE_P (inner); | |
5541 | MEM_IN_STRUCT_P (new) = MEM_IN_STRUCT_P (inner); | |
5542 | } | |
df62f951 | 5543 | else if (GET_CODE (inner) == REG) |
c0d3ac4d RK |
5544 | { |
5545 | /* We can't call gen_lowpart_for_combine here since we always want | |
5546 | a SUBREG and it would sometimes return a new hard register. */ | |
5547 | if (tmode != inner_mode) | |
38a448ca RH |
5548 | new = gen_rtx_SUBREG (tmode, inner, |
5549 | (WORDS_BIG_ENDIAN | |
5550 | && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD | |
5551 | ? (((GET_MODE_SIZE (inner_mode) | |
5552 | - GET_MODE_SIZE (tmode)) | |
5553 | / UNITS_PER_WORD) | |
5554 | - pos / BITS_PER_WORD) | |
5555 | : pos / BITS_PER_WORD)); | |
c0d3ac4d RK |
5556 | else |
5557 | new = inner; | |
5558 | } | |
230d793d | 5559 | else |
6139ff20 RK |
5560 | new = force_to_mode (inner, tmode, |
5561 | len >= HOST_BITS_PER_WIDE_INT | |
5562 | ? GET_MODE_MASK (tmode) | |
5563 | : ((HOST_WIDE_INT) 1 << len) - 1, | |
e3d616e3 | 5564 | NULL_RTX, 0); |
230d793d RS |
5565 | |
5566 | /* If this extraction is going into the destination of a SET, | |
5567 | make a STRICT_LOW_PART unless we made a MEM. */ | |
5568 | ||
5569 | if (in_dest) | |
5570 | return (GET_CODE (new) == MEM ? new | |
77fa0940 | 5571 | : (GET_CODE (new) != SUBREG |
38a448ca | 5572 | ? gen_rtx_CLOBBER (tmode, const0_rtx) |
77fa0940 | 5573 | : gen_rtx_combine (STRICT_LOW_PART, VOIDmode, new))); |
230d793d RS |
5574 | |
5575 | /* Otherwise, sign- or zero-extend unless we already are in the | |
5576 | proper mode. */ | |
5577 | ||
5578 | return (mode == tmode ? new | |
5579 | : gen_rtx_combine (unsignedp ? ZERO_EXTEND : SIGN_EXTEND, | |
5580 | mode, new)); | |
5581 | } | |
5582 | ||
cc471082 RS |
5583 | /* Unless this is a COMPARE or we have a funny memory reference, |
5584 | don't do anything with zero-extending field extracts starting at | |
5585 | the low-order bit since they are simple AND operations. */ | |
8999a12e RK |
5586 | if (pos_rtx == 0 && pos == 0 && ! in_dest |
5587 | && ! in_compare && ! spans_byte && unsignedp) | |
230d793d RS |
5588 | return 0; |
5589 | ||
e7373556 RK |
5590 | /* Unless we are allowed to span bytes, reject this if we would be |
5591 | spanning bytes or if the position is not a constant and the length | |
5592 | is not 1. In all other cases, we would only be going outside | |
5593 | out object in cases when an original shift would have been | |
5594 | undefined. */ | |
5595 | if (! spans_byte | |
5596 | && ((pos_rtx == 0 && pos + len > GET_MODE_BITSIZE (is_mode)) | |
5597 | || (pos_rtx != 0 && len != 1))) | |
5598 | return 0; | |
5599 | ||
d7cd794f | 5600 | /* Get the mode to use should INNER not be a MEM, the mode for the position, |
230d793d RS |
5601 | and the mode for the result. */ |
5602 | #ifdef HAVE_insv | |
5603 | if (in_dest) | |
5604 | { | |
d7cd794f | 5605 | wanted_inner_reg_mode = insn_operand_mode[(int) CODE_FOR_insv][0]; |
230d793d RS |
5606 | pos_mode = insn_operand_mode[(int) CODE_FOR_insv][2]; |
5607 | extraction_mode = insn_operand_mode[(int) CODE_FOR_insv][3]; | |
5608 | } | |
5609 | #endif | |
5610 | ||
5611 | #ifdef HAVE_extzv | |
5612 | if (! in_dest && unsignedp) | |
5613 | { | |
d7cd794f | 5614 | wanted_inner_reg_mode = insn_operand_mode[(int) CODE_FOR_extzv][1]; |
230d793d RS |
5615 | pos_mode = insn_operand_mode[(int) CODE_FOR_extzv][3]; |
5616 | extraction_mode = insn_operand_mode[(int) CODE_FOR_extzv][0]; | |
5617 | } | |
5618 | #endif | |
5619 | ||
5620 | #ifdef HAVE_extv | |
5621 | if (! in_dest && ! unsignedp) | |
5622 | { | |
d7cd794f | 5623 | wanted_inner_reg_mode = insn_operand_mode[(int) CODE_FOR_extv][1]; |
230d793d RS |
5624 | pos_mode = insn_operand_mode[(int) CODE_FOR_extv][3]; |
5625 | extraction_mode = insn_operand_mode[(int) CODE_FOR_extv][0]; | |
5626 | } | |
5627 | #endif | |
5628 | ||
5629 | /* Never narrow an object, since that might not be safe. */ | |
5630 | ||
5631 | if (mode != VOIDmode | |
5632 | && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode)) | |
5633 | extraction_mode = mode; | |
5634 | ||
5635 | if (pos_rtx && GET_MODE (pos_rtx) != VOIDmode | |
5636 | && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx))) | |
5637 | pos_mode = GET_MODE (pos_rtx); | |
5638 | ||
d7cd794f RK |
5639 | /* If this is not from memory, the desired mode is wanted_inner_reg_mode; |
5640 | if we have to change the mode of memory and cannot, the desired mode is | |
5641 | EXTRACTION_MODE. */ | |
5642 | if (GET_CODE (inner) != MEM) | |
5643 | wanted_inner_mode = wanted_inner_reg_mode; | |
5644 | else if (inner_mode != wanted_inner_mode | |
5645 | && (mode_dependent_address_p (XEXP (inner, 0)) | |
5646 | || MEM_VOLATILE_P (inner))) | |
5647 | wanted_inner_mode = extraction_mode; | |
230d793d | 5648 | |
6139ff20 RK |
5649 | orig_pos = pos; |
5650 | ||
f76b9db2 ILT |
5651 | if (BITS_BIG_ENDIAN) |
5652 | { | |
cf54c2cd DE |
5653 | /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to |
5654 | BITS_BIG_ENDIAN style. If position is constant, compute new | |
5655 | position. Otherwise, build subtraction. | |
5656 | Note that POS is relative to the mode of the original argument. | |
5657 | If it's a MEM we need to recompute POS relative to that. | |
5658 | However, if we're extracting from (or inserting into) a register, | |
5659 | we want to recompute POS relative to wanted_inner_mode. */ | |
5660 | int width = (GET_CODE (inner) == MEM | |
5661 | ? GET_MODE_BITSIZE (is_mode) | |
5662 | : GET_MODE_BITSIZE (wanted_inner_mode)); | |
5663 | ||
f76b9db2 | 5664 | if (pos_rtx == 0) |
cf54c2cd | 5665 | pos = width - len - pos; |
f76b9db2 ILT |
5666 | else |
5667 | pos_rtx | |
5668 | = gen_rtx_combine (MINUS, GET_MODE (pos_rtx), | |
cf54c2cd DE |
5669 | GEN_INT (width - len), pos_rtx); |
5670 | /* POS may be less than 0 now, but we check for that below. | |
5671 | Note that it can only be less than 0 if GET_CODE (inner) != MEM. */ | |
f76b9db2 | 5672 | } |
230d793d RS |
5673 | |
5674 | /* If INNER has a wider mode, make it smaller. If this is a constant | |
5675 | extract, try to adjust the byte to point to the byte containing | |
5676 | the value. */ | |
d7cd794f RK |
5677 | if (wanted_inner_mode != VOIDmode |
5678 | && GET_MODE_SIZE (wanted_inner_mode) < GET_MODE_SIZE (is_mode) | |
230d793d | 5679 | && ((GET_CODE (inner) == MEM |
d7cd794f | 5680 | && (inner_mode == wanted_inner_mode |
230d793d RS |
5681 | || (! mode_dependent_address_p (XEXP (inner, 0)) |
5682 | && ! MEM_VOLATILE_P (inner)))))) | |
5683 | { | |
5684 | int offset = 0; | |
5685 | ||
5686 | /* The computations below will be correct if the machine is big | |
5687 | endian in both bits and bytes or little endian in bits and bytes. | |
5688 | If it is mixed, we must adjust. */ | |
5689 | ||
230d793d | 5690 | /* If bytes are big endian and we had a paradoxical SUBREG, we must |
0f41302f | 5691 | adjust OFFSET to compensate. */ |
f76b9db2 ILT |
5692 | if (BYTES_BIG_ENDIAN |
5693 | && ! spans_byte | |
230d793d RS |
5694 | && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode)) |
5695 | offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode); | |
230d793d RS |
5696 | |
5697 | /* If this is a constant position, we can move to the desired byte. */ | |
8999a12e | 5698 | if (pos_rtx == 0) |
230d793d RS |
5699 | { |
5700 | offset += pos / BITS_PER_UNIT; | |
d7cd794f | 5701 | pos %= GET_MODE_BITSIZE (wanted_inner_mode); |
230d793d RS |
5702 | } |
5703 | ||
f76b9db2 ILT |
5704 | if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN |
5705 | && ! spans_byte | |
d7cd794f | 5706 | && is_mode != wanted_inner_mode) |
c6b3f1f2 | 5707 | offset = (GET_MODE_SIZE (is_mode) |
d7cd794f | 5708 | - GET_MODE_SIZE (wanted_inner_mode) - offset); |
c6b3f1f2 | 5709 | |
d7cd794f | 5710 | if (offset != 0 || inner_mode != wanted_inner_mode) |
230d793d | 5711 | { |
38a448ca RH |
5712 | rtx newmem = gen_rtx_MEM (wanted_inner_mode, |
5713 | plus_constant (XEXP (inner, 0), offset)); | |
230d793d RS |
5714 | RTX_UNCHANGING_P (newmem) = RTX_UNCHANGING_P (inner); |
5715 | MEM_VOLATILE_P (newmem) = MEM_VOLATILE_P (inner); | |
5716 | MEM_IN_STRUCT_P (newmem) = MEM_IN_STRUCT_P (inner); | |
5717 | inner = newmem; | |
5718 | } | |
5719 | } | |
5720 | ||
9e74dc41 RK |
5721 | /* If INNER is not memory, we can always get it into the proper mode. If we |
5722 | are changing its mode, POS must be a constant and smaller than the size | |
5723 | of the new mode. */ | |
230d793d | 5724 | else if (GET_CODE (inner) != MEM) |
9e74dc41 RK |
5725 | { |
5726 | if (GET_MODE (inner) != wanted_inner_mode | |
5727 | && (pos_rtx != 0 | |
5728 | || orig_pos + len > GET_MODE_BITSIZE (wanted_inner_mode))) | |
5729 | return 0; | |
5730 | ||
5731 | inner = force_to_mode (inner, wanted_inner_mode, | |
5732 | pos_rtx | |
5733 | || len + orig_pos >= HOST_BITS_PER_WIDE_INT | |
5734 | ? GET_MODE_MASK (wanted_inner_mode) | |
5735 | : (((HOST_WIDE_INT) 1 << len) - 1) << orig_pos, | |
5736 | NULL_RTX, 0); | |
5737 | } | |
230d793d RS |
5738 | |
5739 | /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we | |
5740 | have to zero extend. Otherwise, we can just use a SUBREG. */ | |
8999a12e | 5741 | if (pos_rtx != 0 |
230d793d RS |
5742 | && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx))) |
5743 | pos_rtx = gen_rtx_combine (ZERO_EXTEND, pos_mode, pos_rtx); | |
8999a12e | 5744 | else if (pos_rtx != 0 |
230d793d RS |
5745 | && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx))) |
5746 | pos_rtx = gen_lowpart_for_combine (pos_mode, pos_rtx); | |
5747 | ||
8999a12e RK |
5748 | /* Make POS_RTX unless we already have it and it is correct. If we don't |
5749 | have a POS_RTX but we do have an ORIG_POS_RTX, the latter must | |
0f41302f | 5750 | be a CONST_INT. */ |
8999a12e RK |
5751 | if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos) |
5752 | pos_rtx = orig_pos_rtx; | |
5753 | ||
5754 | else if (pos_rtx == 0) | |
5f4f0e22 | 5755 | pos_rtx = GEN_INT (pos); |
230d793d RS |
5756 | |
5757 | /* Make the required operation. See if we can use existing rtx. */ | |
5758 | new = gen_rtx_combine (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT, | |
5f4f0e22 | 5759 | extraction_mode, inner, GEN_INT (len), pos_rtx); |
230d793d RS |
5760 | if (! in_dest) |
5761 | new = gen_lowpart_for_combine (mode, new); | |
5762 | ||
5763 | return new; | |
5764 | } | |
5765 | \f | |
71923da7 RK |
5766 | /* See if X contains an ASHIFT of COUNT or more bits that can be commuted |
5767 | with any other operations in X. Return X without that shift if so. */ | |
5768 | ||
5769 | static rtx | |
5770 | extract_left_shift (x, count) | |
5771 | rtx x; | |
5772 | int count; | |
5773 | { | |
5774 | enum rtx_code code = GET_CODE (x); | |
5775 | enum machine_mode mode = GET_MODE (x); | |
5776 | rtx tem; | |
5777 | ||
5778 | switch (code) | |
5779 | { | |
5780 | case ASHIFT: | |
5781 | /* This is the shift itself. If it is wide enough, we will return | |
5782 | either the value being shifted if the shift count is equal to | |
5783 | COUNT or a shift for the difference. */ | |
5784 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5785 | && INTVAL (XEXP (x, 1)) >= count) | |
5786 | return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0), | |
5787 | INTVAL (XEXP (x, 1)) - count); | |
5788 | break; | |
5789 | ||
5790 | case NEG: case NOT: | |
5791 | if ((tem = extract_left_shift (XEXP (x, 0), count)) != 0) | |
0c1c8ea6 | 5792 | return gen_unary (code, mode, mode, tem); |
71923da7 RK |
5793 | |
5794 | break; | |
5795 | ||
5796 | case PLUS: case IOR: case XOR: case AND: | |
5797 | /* If we can safely shift this constant and we find the inner shift, | |
5798 | make a new operation. */ | |
5799 | if (GET_CODE (XEXP (x,1)) == CONST_INT | |
b729186a | 5800 | && (INTVAL (XEXP (x, 1)) & ((((HOST_WIDE_INT) 1 << count)) - 1)) == 0 |
71923da7 RK |
5801 | && (tem = extract_left_shift (XEXP (x, 0), count)) != 0) |
5802 | return gen_binary (code, mode, tem, | |
5803 | GEN_INT (INTVAL (XEXP (x, 1)) >> count)); | |
5804 | ||
5805 | break; | |
e9a25f70 JL |
5806 | |
5807 | default: | |
5808 | break; | |
71923da7 RK |
5809 | } |
5810 | ||
5811 | return 0; | |
5812 | } | |
5813 | \f | |
230d793d RS |
5814 | /* Look at the expression rooted at X. Look for expressions |
5815 | equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND. | |
5816 | Form these expressions. | |
5817 | ||
5818 | Return the new rtx, usually just X. | |
5819 | ||
5820 | Also, for machines like the Vax that don't have logical shift insns, | |
5821 | try to convert logical to arithmetic shift operations in cases where | |
5822 | they are equivalent. This undoes the canonicalizations to logical | |
5823 | shifts done elsewhere. | |
5824 | ||
5825 | We try, as much as possible, to re-use rtl expressions to save memory. | |
5826 | ||
5827 | IN_CODE says what kind of expression we are processing. Normally, it is | |
42495ca0 RK |
5828 | SET. In a memory address (inside a MEM, PLUS or minus, the latter two |
5829 | being kludges), it is MEM. When processing the arguments of a comparison | |
230d793d RS |
5830 | or a COMPARE against zero, it is COMPARE. */ |
5831 | ||
5832 | static rtx | |
5833 | make_compound_operation (x, in_code) | |
5834 | rtx x; | |
5835 | enum rtx_code in_code; | |
5836 | { | |
5837 | enum rtx_code code = GET_CODE (x); | |
5838 | enum machine_mode mode = GET_MODE (x); | |
5839 | int mode_width = GET_MODE_BITSIZE (mode); | |
71923da7 | 5840 | rtx rhs, lhs; |
230d793d | 5841 | enum rtx_code next_code; |
f24ad0e4 | 5842 | int i; |
230d793d | 5843 | rtx new = 0; |
280f58ba | 5844 | rtx tem; |
230d793d RS |
5845 | char *fmt; |
5846 | ||
5847 | /* Select the code to be used in recursive calls. Once we are inside an | |
5848 | address, we stay there. If we have a comparison, set to COMPARE, | |
5849 | but once inside, go back to our default of SET. */ | |
5850 | ||
42495ca0 | 5851 | next_code = (code == MEM || code == PLUS || code == MINUS ? MEM |
230d793d RS |
5852 | : ((code == COMPARE || GET_RTX_CLASS (code) == '<') |
5853 | && XEXP (x, 1) == const0_rtx) ? COMPARE | |
5854 | : in_code == COMPARE ? SET : in_code); | |
5855 | ||
5856 | /* Process depending on the code of this operation. If NEW is set | |
5857 | non-zero, it will be returned. */ | |
5858 | ||
5859 | switch (code) | |
5860 | { | |
5861 | case ASHIFT: | |
230d793d RS |
5862 | /* Convert shifts by constants into multiplications if inside |
5863 | an address. */ | |
5864 | if (in_code == MEM && GET_CODE (XEXP (x, 1)) == CONST_INT | |
5f4f0e22 | 5865 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT |
230d793d | 5866 | && INTVAL (XEXP (x, 1)) >= 0) |
280f58ba RK |
5867 | { |
5868 | new = make_compound_operation (XEXP (x, 0), next_code); | |
5869 | new = gen_rtx_combine (MULT, mode, new, | |
5870 | GEN_INT ((HOST_WIDE_INT) 1 | |
5871 | << INTVAL (XEXP (x, 1)))); | |
5872 | } | |
230d793d RS |
5873 | break; |
5874 | ||
5875 | case AND: | |
5876 | /* If the second operand is not a constant, we can't do anything | |
5877 | with it. */ | |
5878 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
5879 | break; | |
5880 | ||
5881 | /* If the constant is a power of two minus one and the first operand | |
5882 | is a logical right shift, make an extraction. */ | |
5883 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
5884 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
280f58ba RK |
5885 | { |
5886 | new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); | |
5887 | new = make_extraction (mode, new, 0, XEXP (XEXP (x, 0), 1), i, 1, | |
5888 | 0, in_code == COMPARE); | |
5889 | } | |
dfbe1b2f | 5890 | |
230d793d RS |
5891 | /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */ |
5892 | else if (GET_CODE (XEXP (x, 0)) == SUBREG | |
5893 | && subreg_lowpart_p (XEXP (x, 0)) | |
5894 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT | |
5895 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
280f58ba RK |
5896 | { |
5897 | new = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0), | |
5898 | next_code); | |
2f99f437 | 5899 | new = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new, 0, |
280f58ba RK |
5900 | XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1, |
5901 | 0, in_code == COMPARE); | |
5902 | } | |
45620ed4 | 5903 | /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */ |
c2f9f64e JW |
5904 | else if ((GET_CODE (XEXP (x, 0)) == XOR |
5905 | || GET_CODE (XEXP (x, 0)) == IOR) | |
5906 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT | |
5907 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT | |
5908 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
5909 | { | |
5910 | /* Apply the distributive law, and then try to make extractions. */ | |
5911 | new = gen_rtx_combine (GET_CODE (XEXP (x, 0)), mode, | |
38a448ca RH |
5912 | gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0), |
5913 | XEXP (x, 1)), | |
5914 | gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1), | |
5915 | XEXP (x, 1))); | |
c2f9f64e JW |
5916 | new = make_compound_operation (new, in_code); |
5917 | } | |
a7c99304 RK |
5918 | |
5919 | /* If we are have (and (rotate X C) M) and C is larger than the number | |
5920 | of bits in M, this is an extraction. */ | |
5921 | ||
5922 | else if (GET_CODE (XEXP (x, 0)) == ROTATE | |
5923 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5924 | && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0 | |
5925 | && i <= INTVAL (XEXP (XEXP (x, 0), 1))) | |
280f58ba RK |
5926 | { |
5927 | new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); | |
5928 | new = make_extraction (mode, new, | |
5929 | (GET_MODE_BITSIZE (mode) | |
5930 | - INTVAL (XEXP (XEXP (x, 0), 1))), | |
5931 | NULL_RTX, i, 1, 0, in_code == COMPARE); | |
5932 | } | |
a7c99304 RK |
5933 | |
5934 | /* On machines without logical shifts, if the operand of the AND is | |
230d793d RS |
5935 | a logical shift and our mask turns off all the propagated sign |
5936 | bits, we can replace the logical shift with an arithmetic shift. */ | |
d0ab8cd3 RK |
5937 | else if (ashr_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing |
5938 | && (lshr_optab->handlers[(int) mode].insn_code | |
5939 | == CODE_FOR_nothing) | |
230d793d RS |
5940 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT |
5941 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
5942 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 | |
5f4f0e22 CH |
5943 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT |
5944 | && mode_width <= HOST_BITS_PER_WIDE_INT) | |
230d793d | 5945 | { |
5f4f0e22 | 5946 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); |
230d793d RS |
5947 | |
5948 | mask >>= INTVAL (XEXP (XEXP (x, 0), 1)); | |
5949 | if ((INTVAL (XEXP (x, 1)) & ~mask) == 0) | |
5950 | SUBST (XEXP (x, 0), | |
280f58ba RK |
5951 | gen_rtx_combine (ASHIFTRT, mode, |
5952 | make_compound_operation (XEXP (XEXP (x, 0), 0), | |
5953 | next_code), | |
230d793d RS |
5954 | XEXP (XEXP (x, 0), 1))); |
5955 | } | |
5956 | ||
5957 | /* If the constant is one less than a power of two, this might be | |
5958 | representable by an extraction even if no shift is present. | |
5959 | If it doesn't end up being a ZERO_EXTEND, we will ignore it unless | |
5960 | we are in a COMPARE. */ | |
5961 | else if ((i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0) | |
280f58ba RK |
5962 | new = make_extraction (mode, |
5963 | make_compound_operation (XEXP (x, 0), | |
5964 | next_code), | |
5965 | 0, NULL_RTX, i, 1, 0, in_code == COMPARE); | |
230d793d RS |
5966 | |
5967 | /* If we are in a comparison and this is an AND with a power of two, | |
5968 | convert this into the appropriate bit extract. */ | |
5969 | else if (in_code == COMPARE | |
5970 | && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0) | |
280f58ba RK |
5971 | new = make_extraction (mode, |
5972 | make_compound_operation (XEXP (x, 0), | |
5973 | next_code), | |
5974 | i, NULL_RTX, 1, 1, 0, 1); | |
230d793d RS |
5975 | |
5976 | break; | |
5977 | ||
5978 | case LSHIFTRT: | |
5979 | /* If the sign bit is known to be zero, replace this with an | |
5980 | arithmetic shift. */ | |
d0ab8cd3 RK |
5981 | if (ashr_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing |
5982 | && lshr_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing | |
5f4f0e22 | 5983 | && mode_width <= HOST_BITS_PER_WIDE_INT |
951553af | 5984 | && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0) |
230d793d | 5985 | { |
280f58ba RK |
5986 | new = gen_rtx_combine (ASHIFTRT, mode, |
5987 | make_compound_operation (XEXP (x, 0), | |
5988 | next_code), | |
5989 | XEXP (x, 1)); | |
230d793d RS |
5990 | break; |
5991 | } | |
5992 | ||
0f41302f | 5993 | /* ... fall through ... */ |
230d793d RS |
5994 | |
5995 | case ASHIFTRT: | |
71923da7 RK |
5996 | lhs = XEXP (x, 0); |
5997 | rhs = XEXP (x, 1); | |
5998 | ||
230d793d RS |
5999 | /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1, |
6000 | this is a SIGN_EXTRACT. */ | |
71923da7 RK |
6001 | if (GET_CODE (rhs) == CONST_INT |
6002 | && GET_CODE (lhs) == ASHIFT | |
6003 | && GET_CODE (XEXP (lhs, 1)) == CONST_INT | |
6004 | && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1))) | |
280f58ba | 6005 | { |
71923da7 | 6006 | new = make_compound_operation (XEXP (lhs, 0), next_code); |
280f58ba | 6007 | new = make_extraction (mode, new, |
71923da7 RK |
6008 | INTVAL (rhs) - INTVAL (XEXP (lhs, 1)), |
6009 | NULL_RTX, mode_width - INTVAL (rhs), | |
d0ab8cd3 RK |
6010 | code == LSHIFTRT, 0, in_code == COMPARE); |
6011 | } | |
6012 | ||
71923da7 RK |
6013 | /* See if we have operations between an ASHIFTRT and an ASHIFT. |
6014 | If so, try to merge the shifts into a SIGN_EXTEND. We could | |
6015 | also do this for some cases of SIGN_EXTRACT, but it doesn't | |
6016 | seem worth the effort; the case checked for occurs on Alpha. */ | |
6017 | ||
6018 | if (GET_RTX_CLASS (GET_CODE (lhs)) != 'o' | |
6019 | && ! (GET_CODE (lhs) == SUBREG | |
6020 | && (GET_RTX_CLASS (GET_CODE (SUBREG_REG (lhs))) == 'o')) | |
6021 | && GET_CODE (rhs) == CONST_INT | |
6022 | && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT | |
6023 | && (new = extract_left_shift (lhs, INTVAL (rhs))) != 0) | |
6024 | new = make_extraction (mode, make_compound_operation (new, next_code), | |
6025 | 0, NULL_RTX, mode_width - INTVAL (rhs), | |
6026 | code == LSHIFTRT, 0, in_code == COMPARE); | |
6027 | ||
230d793d | 6028 | break; |
280f58ba RK |
6029 | |
6030 | case SUBREG: | |
6031 | /* Call ourselves recursively on the inner expression. If we are | |
6032 | narrowing the object and it has a different RTL code from | |
6033 | what it originally did, do this SUBREG as a force_to_mode. */ | |
6034 | ||
0a5cbff6 | 6035 | tem = make_compound_operation (SUBREG_REG (x), in_code); |
280f58ba RK |
6036 | if (GET_CODE (tem) != GET_CODE (SUBREG_REG (x)) |
6037 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (tem)) | |
6038 | && subreg_lowpart_p (x)) | |
0a5cbff6 RK |
6039 | { |
6040 | rtx newer = force_to_mode (tem, mode, | |
e3d616e3 | 6041 | GET_MODE_MASK (mode), NULL_RTX, 0); |
0a5cbff6 RK |
6042 | |
6043 | /* If we have something other than a SUBREG, we might have | |
6044 | done an expansion, so rerun outselves. */ | |
6045 | if (GET_CODE (newer) != SUBREG) | |
6046 | newer = make_compound_operation (newer, in_code); | |
6047 | ||
6048 | return newer; | |
6049 | } | |
6f28d3e9 RH |
6050 | |
6051 | /* If this is a paradoxical subreg, and the new code is a sign or | |
6052 | zero extension, omit the subreg and widen the extension. If it | |
6053 | is a regular subreg, we can still get rid of the subreg by not | |
6054 | widening so much, or in fact removing the extension entirely. */ | |
6055 | if ((GET_CODE (tem) == SIGN_EXTEND | |
6056 | || GET_CODE (tem) == ZERO_EXTEND) | |
6057 | && subreg_lowpart_p (x)) | |
6058 | { | |
6059 | if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (tem)) | |
6060 | || (GET_MODE_SIZE (mode) > | |
6061 | GET_MODE_SIZE (GET_MODE (XEXP (tem, 0))))) | |
6062 | tem = gen_rtx_combine (GET_CODE (tem), mode, XEXP (tem, 0)); | |
6063 | else | |
6064 | tem = gen_lowpart_for_combine (mode, XEXP (tem, 0)); | |
6065 | return tem; | |
6066 | } | |
e9a25f70 JL |
6067 | break; |
6068 | ||
6069 | default: | |
6070 | break; | |
230d793d RS |
6071 | } |
6072 | ||
6073 | if (new) | |
6074 | { | |
df62f951 | 6075 | x = gen_lowpart_for_combine (mode, new); |
230d793d RS |
6076 | code = GET_CODE (x); |
6077 | } | |
6078 | ||
6079 | /* Now recursively process each operand of this operation. */ | |
6080 | fmt = GET_RTX_FORMAT (code); | |
6081 | for (i = 0; i < GET_RTX_LENGTH (code); i++) | |
6082 | if (fmt[i] == 'e') | |
6083 | { | |
6084 | new = make_compound_operation (XEXP (x, i), next_code); | |
6085 | SUBST (XEXP (x, i), new); | |
6086 | } | |
6087 | ||
6088 | return x; | |
6089 | } | |
6090 | \f | |
6091 | /* Given M see if it is a value that would select a field of bits | |
6092 | within an item, but not the entire word. Return -1 if not. | |
6093 | Otherwise, return the starting position of the field, where 0 is the | |
6094 | low-order bit. | |
6095 | ||
6096 | *PLEN is set to the length of the field. */ | |
6097 | ||
6098 | static int | |
6099 | get_pos_from_mask (m, plen) | |
5f4f0e22 | 6100 | unsigned HOST_WIDE_INT m; |
230d793d RS |
6101 | int *plen; |
6102 | { | |
6103 | /* Get the bit number of the first 1 bit from the right, -1 if none. */ | |
6104 | int pos = exact_log2 (m & - m); | |
6105 | ||
6106 | if (pos < 0) | |
6107 | return -1; | |
6108 | ||
6109 | /* Now shift off the low-order zero bits and see if we have a power of | |
6110 | two minus 1. */ | |
6111 | *plen = exact_log2 ((m >> pos) + 1); | |
6112 | ||
6113 | if (*plen <= 0) | |
6114 | return -1; | |
6115 | ||
6116 | return pos; | |
6117 | } | |
6118 | \f | |
6139ff20 RK |
6119 | /* See if X can be simplified knowing that we will only refer to it in |
6120 | MODE and will only refer to those bits that are nonzero in MASK. | |
6121 | If other bits are being computed or if masking operations are done | |
6122 | that select a superset of the bits in MASK, they can sometimes be | |
6123 | ignored. | |
6124 | ||
6125 | Return a possibly simplified expression, but always convert X to | |
6126 | MODE. If X is a CONST_INT, AND the CONST_INT with MASK. | |
dfbe1b2f RK |
6127 | |
6128 | Also, if REG is non-zero and X is a register equal in value to REG, | |
e3d616e3 RK |
6129 | replace X with REG. |
6130 | ||
6131 | If JUST_SELECT is nonzero, don't optimize by noticing that bits in MASK | |
6132 | are all off in X. This is used when X will be complemented, by either | |
180b8e4b | 6133 | NOT, NEG, or XOR. */ |
dfbe1b2f RK |
6134 | |
6135 | static rtx | |
e3d616e3 | 6136 | force_to_mode (x, mode, mask, reg, just_select) |
dfbe1b2f RK |
6137 | rtx x; |
6138 | enum machine_mode mode; | |
6139ff20 | 6139 | unsigned HOST_WIDE_INT mask; |
dfbe1b2f | 6140 | rtx reg; |
e3d616e3 | 6141 | int just_select; |
dfbe1b2f RK |
6142 | { |
6143 | enum rtx_code code = GET_CODE (x); | |
180b8e4b | 6144 | int next_select = just_select || code == XOR || code == NOT || code == NEG; |
ef026f91 RS |
6145 | enum machine_mode op_mode; |
6146 | unsigned HOST_WIDE_INT fuller_mask, nonzero; | |
6139ff20 RK |
6147 | rtx op0, op1, temp; |
6148 | ||
132d2040 RK |
6149 | /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the |
6150 | code below will do the wrong thing since the mode of such an | |
be3d27d6 CI |
6151 | expression is VOIDmode. |
6152 | ||
6153 | Also do nothing if X is a CLOBBER; this can happen if X was | |
6154 | the return value from a call to gen_lowpart_for_combine. */ | |
6155 | if (code == CALL || code == ASM_OPERANDS || code == CLOBBER) | |
246e00f2 RK |
6156 | return x; |
6157 | ||
6139ff20 RK |
6158 | /* We want to perform the operation is its present mode unless we know |
6159 | that the operation is valid in MODE, in which case we do the operation | |
6160 | in MODE. */ | |
1c75dfa4 RK |
6161 | op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x)) |
6162 | && code_to_optab[(int) code] != 0 | |
ef026f91 RS |
6163 | && (code_to_optab[(int) code]->handlers[(int) mode].insn_code |
6164 | != CODE_FOR_nothing)) | |
6165 | ? mode : GET_MODE (x)); | |
e3d616e3 | 6166 | |
aa988991 RS |
6167 | /* It is not valid to do a right-shift in a narrower mode |
6168 | than the one it came in with. */ | |
6169 | if ((code == LSHIFTRT || code == ASHIFTRT) | |
6170 | && GET_MODE_BITSIZE (mode) < GET_MODE_BITSIZE (GET_MODE (x))) | |
6171 | op_mode = GET_MODE (x); | |
ef026f91 RS |
6172 | |
6173 | /* Truncate MASK to fit OP_MODE. */ | |
6174 | if (op_mode) | |
6175 | mask &= GET_MODE_MASK (op_mode); | |
6139ff20 RK |
6176 | |
6177 | /* When we have an arithmetic operation, or a shift whose count we | |
6178 | do not know, we need to assume that all bit the up to the highest-order | |
6179 | bit in MASK will be needed. This is how we form such a mask. */ | |
ef026f91 RS |
6180 | if (op_mode) |
6181 | fuller_mask = (GET_MODE_BITSIZE (op_mode) >= HOST_BITS_PER_WIDE_INT | |
6182 | ? GET_MODE_MASK (op_mode) | |
6183 | : ((HOST_WIDE_INT) 1 << (floor_log2 (mask) + 1)) - 1); | |
6184 | else | |
6185 | fuller_mask = ~ (HOST_WIDE_INT) 0; | |
6186 | ||
6187 | /* Determine what bits of X are guaranteed to be (non)zero. */ | |
6188 | nonzero = nonzero_bits (x, mode); | |
6139ff20 RK |
6189 | |
6190 | /* If none of the bits in X are needed, return a zero. */ | |
e3d616e3 | 6191 | if (! just_select && (nonzero & mask) == 0) |
6139ff20 | 6192 | return const0_rtx; |
dfbe1b2f | 6193 | |
6139ff20 RK |
6194 | /* If X is a CONST_INT, return a new one. Do this here since the |
6195 | test below will fail. */ | |
6196 | if (GET_CODE (x) == CONST_INT) | |
ceb7983c RK |
6197 | { |
6198 | HOST_WIDE_INT cval = INTVAL (x) & mask; | |
6199 | int width = GET_MODE_BITSIZE (mode); | |
6200 | ||
6201 | /* If MODE is narrower that HOST_WIDE_INT and CVAL is a negative | |
6202 | number, sign extend it. */ | |
6203 | if (width > 0 && width < HOST_BITS_PER_WIDE_INT | |
6204 | && (cval & ((HOST_WIDE_INT) 1 << (width - 1))) != 0) | |
6205 | cval |= (HOST_WIDE_INT) -1 << width; | |
6206 | ||
6207 | return GEN_INT (cval); | |
6208 | } | |
dfbe1b2f | 6209 | |
180b8e4b RK |
6210 | /* If X is narrower than MODE and we want all the bits in X's mode, just |
6211 | get X in the proper mode. */ | |
6212 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode) | |
6213 | && (GET_MODE_MASK (GET_MODE (x)) & ~ mask) == 0) | |
dfbe1b2f RK |
6214 | return gen_lowpart_for_combine (mode, x); |
6215 | ||
71923da7 RK |
6216 | /* If we aren't changing the mode, X is not a SUBREG, and all zero bits in |
6217 | MASK are already known to be zero in X, we need not do anything. */ | |
6218 | if (GET_MODE (x) == mode && code != SUBREG && (~ mask & nonzero) == 0) | |
6139ff20 RK |
6219 | return x; |
6220 | ||
dfbe1b2f RK |
6221 | switch (code) |
6222 | { | |
6139ff20 RK |
6223 | case CLOBBER: |
6224 | /* If X is a (clobber (const_int)), return it since we know we are | |
0f41302f | 6225 | generating something that won't match. */ |
6139ff20 RK |
6226 | return x; |
6227 | ||
6139ff20 RK |
6228 | case USE: |
6229 | /* X is a (use (mem ..)) that was made from a bit-field extraction that | |
6230 | spanned the boundary of the MEM. If we are now masking so it is | |
6231 | within that boundary, we don't need the USE any more. */ | |
f76b9db2 ILT |
6232 | if (! BITS_BIG_ENDIAN |
6233 | && (mask & ~ GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0) | |
e3d616e3 | 6234 | return force_to_mode (XEXP (x, 0), mode, mask, reg, next_select); |
f76b9db2 | 6235 | break; |
6139ff20 | 6236 | |
dfbe1b2f RK |
6237 | case SIGN_EXTEND: |
6238 | case ZERO_EXTEND: | |
6239 | case ZERO_EXTRACT: | |
6240 | case SIGN_EXTRACT: | |
6241 | x = expand_compound_operation (x); | |
6242 | if (GET_CODE (x) != code) | |
e3d616e3 | 6243 | return force_to_mode (x, mode, mask, reg, next_select); |
dfbe1b2f RK |
6244 | break; |
6245 | ||
6246 | case REG: | |
6247 | if (reg != 0 && (rtx_equal_p (get_last_value (reg), x) | |
6248 | || rtx_equal_p (reg, get_last_value (x)))) | |
6249 | x = reg; | |
6250 | break; | |
6251 | ||
dfbe1b2f | 6252 | case SUBREG: |
6139ff20 | 6253 | if (subreg_lowpart_p (x) |
180b8e4b RK |
6254 | /* We can ignore the effect of this SUBREG if it narrows the mode or |
6255 | if the constant masks to zero all the bits the mode doesn't | |
6256 | have. */ | |
6139ff20 RK |
6257 | && ((GET_MODE_SIZE (GET_MODE (x)) |
6258 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
6139ff20 RK |
6259 | || (0 == (mask |
6260 | & GET_MODE_MASK (GET_MODE (x)) | |
180b8e4b | 6261 | & ~ GET_MODE_MASK (GET_MODE (SUBREG_REG (x))))))) |
e3d616e3 | 6262 | return force_to_mode (SUBREG_REG (x), mode, mask, reg, next_select); |
dfbe1b2f RK |
6263 | break; |
6264 | ||
6265 | case AND: | |
6139ff20 RK |
6266 | /* If this is an AND with a constant, convert it into an AND |
6267 | whose constant is the AND of that constant with MASK. If it | |
6268 | remains an AND of MASK, delete it since it is redundant. */ | |
dfbe1b2f | 6269 | |
2ca9ae17 | 6270 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) |
dfbe1b2f | 6271 | { |
6139ff20 RK |
6272 | x = simplify_and_const_int (x, op_mode, XEXP (x, 0), |
6273 | mask & INTVAL (XEXP (x, 1))); | |
dfbe1b2f RK |
6274 | |
6275 | /* If X is still an AND, see if it is an AND with a mask that | |
71923da7 RK |
6276 | is just some low-order bits. If so, and it is MASK, we don't |
6277 | need it. */ | |
dfbe1b2f RK |
6278 | |
6279 | if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT | |
6139ff20 | 6280 | && INTVAL (XEXP (x, 1)) == mask) |
dfbe1b2f | 6281 | x = XEXP (x, 0); |
d0ab8cd3 | 6282 | |
71923da7 RK |
6283 | /* If it remains an AND, try making another AND with the bits |
6284 | in the mode mask that aren't in MASK turned on. If the | |
6285 | constant in the AND is wide enough, this might make a | |
6286 | cheaper constant. */ | |
6287 | ||
6288 | if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT | |
2ca9ae17 JW |
6289 | && GET_MODE_MASK (GET_MODE (x)) != mask |
6290 | && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT) | |
71923da7 RK |
6291 | { |
6292 | HOST_WIDE_INT cval = (INTVAL (XEXP (x, 1)) | |
6293 | | (GET_MODE_MASK (GET_MODE (x)) & ~ mask)); | |
6294 | int width = GET_MODE_BITSIZE (GET_MODE (x)); | |
6295 | rtx y; | |
6296 | ||
6297 | /* If MODE is narrower that HOST_WIDE_INT and CVAL is a negative | |
6298 | number, sign extend it. */ | |
6299 | if (width > 0 && width < HOST_BITS_PER_WIDE_INT | |
6300 | && (cval & ((HOST_WIDE_INT) 1 << (width - 1))) != 0) | |
6301 | cval |= (HOST_WIDE_INT) -1 << width; | |
6302 | ||
6303 | y = gen_binary (AND, GET_MODE (x), XEXP (x, 0), GEN_INT (cval)); | |
6304 | if (rtx_cost (y, SET) < rtx_cost (x, SET)) | |
6305 | x = y; | |
6306 | } | |
6307 | ||
d0ab8cd3 | 6308 | break; |
dfbe1b2f RK |
6309 | } |
6310 | ||
6139ff20 | 6311 | goto binop; |
dfbe1b2f RK |
6312 | |
6313 | case PLUS: | |
6139ff20 RK |
6314 | /* In (and (plus FOO C1) M), if M is a mask that just turns off |
6315 | low-order bits (as in an alignment operation) and FOO is already | |
6316 | aligned to that boundary, mask C1 to that boundary as well. | |
6317 | This may eliminate that PLUS and, later, the AND. */ | |
9fa6d012 TG |
6318 | |
6319 | { | |
6320 | int width = GET_MODE_BITSIZE (mode); | |
6321 | unsigned HOST_WIDE_INT smask = mask; | |
6322 | ||
6323 | /* If MODE is narrower than HOST_WIDE_INT and mask is a negative | |
6324 | number, sign extend it. */ | |
6325 | ||
6326 | if (width < HOST_BITS_PER_WIDE_INT | |
6327 | && (smask & ((HOST_WIDE_INT) 1 << (width - 1))) != 0) | |
6328 | smask |= (HOST_WIDE_INT) -1 << width; | |
6329 | ||
6330 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
0e9ff885 DM |
6331 | && exact_log2 (- smask) >= 0) |
6332 | { | |
6333 | #ifdef STACK_BIAS | |
6334 | if (STACK_BIAS | |
6335 | && (XEXP (x, 0) == stack_pointer_rtx | |
6336 | || XEXP (x, 0) == frame_pointer_rtx)) | |
6337 | { | |
6338 | int sp_alignment = STACK_BOUNDARY / BITS_PER_UNIT; | |
6339 | unsigned HOST_WIDE_INT sp_mask = GET_MODE_MASK (mode); | |
6340 | ||
6341 | sp_mask &= ~ (sp_alignment - 1); | |
6342 | if ((sp_mask & ~ mask) == 0 | |
6343 | && ((INTVAL (XEXP (x, 1)) - STACK_BIAS) & ~ mask) != 0) | |
6344 | return force_to_mode (plus_constant (XEXP (x, 0), | |
6345 | ((INTVAL (XEXP (x, 1)) - | |
6346 | STACK_BIAS) & mask) | |
6347 | + STACK_BIAS), | |
6348 | mode, mask, reg, next_select); | |
6349 | } | |
6350 | #endif | |
6351 | if ((nonzero_bits (XEXP (x, 0), mode) & ~ mask) == 0 | |
6352 | && (INTVAL (XEXP (x, 1)) & ~ mask) != 0) | |
6353 | return force_to_mode (plus_constant (XEXP (x, 0), | |
6354 | INTVAL (XEXP (x, 1)) & mask), | |
6355 | mode, mask, reg, next_select); | |
6356 | } | |
9fa6d012 | 6357 | } |
6139ff20 | 6358 | |
0f41302f | 6359 | /* ... fall through ... */ |
6139ff20 | 6360 | |
dfbe1b2f RK |
6361 | case MINUS: |
6362 | case MULT: | |
6139ff20 RK |
6363 | /* For PLUS, MINUS and MULT, we need any bits less significant than the |
6364 | most significant bit in MASK since carries from those bits will | |
6365 | affect the bits we are interested in. */ | |
6366 | mask = fuller_mask; | |
6367 | goto binop; | |
6368 | ||
dfbe1b2f RK |
6369 | case IOR: |
6370 | case XOR: | |
6139ff20 RK |
6371 | /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and |
6372 | LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...) | |
6373 | operation which may be a bitfield extraction. Ensure that the | |
6374 | constant we form is not wider than the mode of X. */ | |
6375 | ||
6376 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
6377 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
6378 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 | |
6379 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT | |
6380 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
6381 | && ((INTVAL (XEXP (XEXP (x, 0), 1)) | |
6382 | + floor_log2 (INTVAL (XEXP (x, 1)))) | |
6383 | < GET_MODE_BITSIZE (GET_MODE (x))) | |
6384 | && (INTVAL (XEXP (x, 1)) | |
01c82bbb | 6385 | & ~ nonzero_bits (XEXP (x, 0), GET_MODE (x))) == 0) |
6139ff20 RK |
6386 | { |
6387 | temp = GEN_INT ((INTVAL (XEXP (x, 1)) & mask) | |
6388 | << INTVAL (XEXP (XEXP (x, 0), 1))); | |
6389 | temp = gen_binary (GET_CODE (x), GET_MODE (x), | |
6390 | XEXP (XEXP (x, 0), 0), temp); | |
d4d2b13f RK |
6391 | x = gen_binary (LSHIFTRT, GET_MODE (x), temp, |
6392 | XEXP (XEXP (x, 0), 1)); | |
e3d616e3 | 6393 | return force_to_mode (x, mode, mask, reg, next_select); |
6139ff20 RK |
6394 | } |
6395 | ||
6396 | binop: | |
dfbe1b2f | 6397 | /* For most binary operations, just propagate into the operation and |
6139ff20 RK |
6398 | change the mode if we have an operation of that mode. */ |
6399 | ||
e3d616e3 RK |
6400 | op0 = gen_lowpart_for_combine (op_mode, |
6401 | force_to_mode (XEXP (x, 0), mode, mask, | |
6402 | reg, next_select)); | |
6403 | op1 = gen_lowpart_for_combine (op_mode, | |
6404 | force_to_mode (XEXP (x, 1), mode, mask, | |
6405 | reg, next_select)); | |
6139ff20 | 6406 | |
2dd484ed RK |
6407 | /* If OP1 is a CONST_INT and X is an IOR or XOR, clear bits outside |
6408 | MASK since OP1 might have been sign-extended but we never want | |
6409 | to turn on extra bits, since combine might have previously relied | |
6410 | on them being off. */ | |
6411 | if (GET_CODE (op1) == CONST_INT && (code == IOR || code == XOR) | |
6412 | && (INTVAL (op1) & mask) != 0) | |
6413 | op1 = GEN_INT (INTVAL (op1) & mask); | |
6414 | ||
6139ff20 RK |
6415 | if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0) || op1 != XEXP (x, 1)) |
6416 | x = gen_binary (code, op_mode, op0, op1); | |
d0ab8cd3 | 6417 | break; |
dfbe1b2f RK |
6418 | |
6419 | case ASHIFT: | |
dfbe1b2f | 6420 | /* For left shifts, do the same, but just for the first operand. |
f6785026 RK |
6421 | However, we cannot do anything with shifts where we cannot |
6422 | guarantee that the counts are smaller than the size of the mode | |
6423 | because such a count will have a different meaning in a | |
6139ff20 | 6424 | wider mode. */ |
f6785026 RK |
6425 | |
6426 | if (! (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6139ff20 | 6427 | && INTVAL (XEXP (x, 1)) >= 0 |
f6785026 RK |
6428 | && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode)) |
6429 | && ! (GET_MODE (XEXP (x, 1)) != VOIDmode | |
6430 | && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1))) | |
adb7a1cb | 6431 | < (unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode)))) |
f6785026 RK |
6432 | break; |
6433 | ||
6139ff20 RK |
6434 | /* If the shift count is a constant and we can do arithmetic in |
6435 | the mode of the shift, refine which bits we need. Otherwise, use the | |
6436 | conservative form of the mask. */ | |
6437 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6438 | && INTVAL (XEXP (x, 1)) >= 0 | |
6439 | && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (op_mode) | |
6440 | && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT) | |
6441 | mask >>= INTVAL (XEXP (x, 1)); | |
6442 | else | |
6443 | mask = fuller_mask; | |
6444 | ||
6445 | op0 = gen_lowpart_for_combine (op_mode, | |
6446 | force_to_mode (XEXP (x, 0), op_mode, | |
e3d616e3 | 6447 | mask, reg, next_select)); |
6139ff20 RK |
6448 | |
6449 | if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0)) | |
6450 | x = gen_binary (code, op_mode, op0, XEXP (x, 1)); | |
d0ab8cd3 | 6451 | break; |
dfbe1b2f RK |
6452 | |
6453 | case LSHIFTRT: | |
1347292b JW |
6454 | /* Here we can only do something if the shift count is a constant, |
6455 | this shift constant is valid for the host, and we can do arithmetic | |
6456 | in OP_MODE. */ | |
dfbe1b2f RK |
6457 | |
6458 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
1347292b | 6459 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT |
6139ff20 | 6460 | && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT) |
d0ab8cd3 | 6461 | { |
6139ff20 RK |
6462 | rtx inner = XEXP (x, 0); |
6463 | ||
6464 | /* Select the mask of the bits we need for the shift operand. */ | |
6465 | mask <<= INTVAL (XEXP (x, 1)); | |
d0ab8cd3 | 6466 | |
6139ff20 RK |
6467 | /* We can only change the mode of the shift if we can do arithmetic |
6468 | in the mode of the shift and MASK is no wider than the width of | |
6469 | OP_MODE. */ | |
6470 | if (GET_MODE_BITSIZE (op_mode) > HOST_BITS_PER_WIDE_INT | |
6471 | || (mask & ~ GET_MODE_MASK (op_mode)) != 0) | |
d0ab8cd3 RK |
6472 | op_mode = GET_MODE (x); |
6473 | ||
e3d616e3 | 6474 | inner = force_to_mode (inner, op_mode, mask, reg, next_select); |
6139ff20 RK |
6475 | |
6476 | if (GET_MODE (x) != op_mode || inner != XEXP (x, 0)) | |
6477 | x = gen_binary (LSHIFTRT, op_mode, inner, XEXP (x, 1)); | |
d0ab8cd3 | 6478 | } |
6139ff20 RK |
6479 | |
6480 | /* If we have (and (lshiftrt FOO C1) C2) where the combination of the | |
6481 | shift and AND produces only copies of the sign bit (C2 is one less | |
6482 | than a power of two), we can do this with just a shift. */ | |
6483 | ||
6484 | if (GET_CODE (x) == LSHIFTRT | |
6485 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
6486 | && ((INTVAL (XEXP (x, 1)) | |
6487 | + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
6488 | >= GET_MODE_BITSIZE (GET_MODE (x))) | |
6489 | && exact_log2 (mask + 1) >= 0 | |
6490 | && (num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))) | |
6491 | >= exact_log2 (mask + 1))) | |
6492 | x = gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0), | |
6493 | GEN_INT (GET_MODE_BITSIZE (GET_MODE (x)) | |
6494 | - exact_log2 (mask + 1))); | |
d0ab8cd3 RK |
6495 | break; |
6496 | ||
6497 | case ASHIFTRT: | |
6139ff20 RK |
6498 | /* If we are just looking for the sign bit, we don't need this shift at |
6499 | all, even if it has a variable count. */ | |
9bf22b75 RK |
6500 | if (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT |
6501 | && (mask == ((HOST_WIDE_INT) 1 | |
6502 | << (GET_MODE_BITSIZE (GET_MODE (x)) - 1)))) | |
e3d616e3 | 6503 | return force_to_mode (XEXP (x, 0), mode, mask, reg, next_select); |
6139ff20 RK |
6504 | |
6505 | /* If this is a shift by a constant, get a mask that contains those bits | |
6506 | that are not copies of the sign bit. We then have two cases: If | |
6507 | MASK only includes those bits, this can be a logical shift, which may | |
6508 | allow simplifications. If MASK is a single-bit field not within | |
6509 | those bits, we are requesting a copy of the sign bit and hence can | |
6510 | shift the sign bit to the appropriate location. */ | |
6511 | ||
6512 | if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) >= 0 | |
6513 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT) | |
6514 | { | |
6515 | int i = -1; | |
6516 | ||
b69960ac RK |
6517 | /* If the considered data is wider then HOST_WIDE_INT, we can't |
6518 | represent a mask for all its bits in a single scalar. | |
6519 | But we only care about the lower bits, so calculate these. */ | |
6520 | ||
6a11342f | 6521 | if (GET_MODE_BITSIZE (GET_MODE (x)) > HOST_BITS_PER_WIDE_INT) |
b69960ac | 6522 | { |
0f41302f | 6523 | nonzero = ~ (HOST_WIDE_INT) 0; |
b69960ac RK |
6524 | |
6525 | /* GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1)) | |
6526 | is the number of bits a full-width mask would have set. | |
6527 | We need only shift if these are fewer than nonzero can | |
6528 | hold. If not, we must keep all bits set in nonzero. */ | |
6529 | ||
6530 | if (GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1)) | |
6531 | < HOST_BITS_PER_WIDE_INT) | |
6532 | nonzero >>= INTVAL (XEXP (x, 1)) | |
6533 | + HOST_BITS_PER_WIDE_INT | |
6534 | - GET_MODE_BITSIZE (GET_MODE (x)) ; | |
6535 | } | |
6536 | else | |
6537 | { | |
6538 | nonzero = GET_MODE_MASK (GET_MODE (x)); | |
6539 | nonzero >>= INTVAL (XEXP (x, 1)); | |
6540 | } | |
6139ff20 RK |
6541 | |
6542 | if ((mask & ~ nonzero) == 0 | |
6543 | || (i = exact_log2 (mask)) >= 0) | |
6544 | { | |
6545 | x = simplify_shift_const | |
6546 | (x, LSHIFTRT, GET_MODE (x), XEXP (x, 0), | |
6547 | i < 0 ? INTVAL (XEXP (x, 1)) | |
6548 | : GET_MODE_BITSIZE (GET_MODE (x)) - 1 - i); | |
6549 | ||
6550 | if (GET_CODE (x) != ASHIFTRT) | |
e3d616e3 | 6551 | return force_to_mode (x, mode, mask, reg, next_select); |
6139ff20 RK |
6552 | } |
6553 | } | |
6554 | ||
6555 | /* If MASK is 1, convert this to a LSHIFTRT. This can be done | |
6556 | even if the shift count isn't a constant. */ | |
6557 | if (mask == 1) | |
6558 | x = gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0), XEXP (x, 1)); | |
6559 | ||
d0ab8cd3 | 6560 | /* If this is a sign-extension operation that just affects bits |
4c002f29 RK |
6561 | we don't care about, remove it. Be sure the call above returned |
6562 | something that is still a shift. */ | |
d0ab8cd3 | 6563 | |
4c002f29 RK |
6564 | if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT) |
6565 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
d0ab8cd3 | 6566 | && INTVAL (XEXP (x, 1)) >= 0 |
6139ff20 RK |
6567 | && (INTVAL (XEXP (x, 1)) |
6568 | <= GET_MODE_BITSIZE (GET_MODE (x)) - (floor_log2 (mask) + 1)) | |
d0ab8cd3 RK |
6569 | && GET_CODE (XEXP (x, 0)) == ASHIFT |
6570 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
6571 | && INTVAL (XEXP (XEXP (x, 0), 1)) == INTVAL (XEXP (x, 1))) | |
e3d616e3 RK |
6572 | return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask, |
6573 | reg, next_select); | |
6139ff20 | 6574 | |
dfbe1b2f RK |
6575 | break; |
6576 | ||
6139ff20 RK |
6577 | case ROTATE: |
6578 | case ROTATERT: | |
6579 | /* If the shift count is constant and we can do computations | |
6580 | in the mode of X, compute where the bits we care about are. | |
6581 | Otherwise, we can't do anything. Don't change the mode of | |
6582 | the shift or propagate MODE into the shift, though. */ | |
6583 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6584 | && INTVAL (XEXP (x, 1)) >= 0) | |
6585 | { | |
6586 | temp = simplify_binary_operation (code == ROTATE ? ROTATERT : ROTATE, | |
6587 | GET_MODE (x), GEN_INT (mask), | |
6588 | XEXP (x, 1)); | |
7d171a1e | 6589 | if (temp && GET_CODE(temp) == CONST_INT) |
6139ff20 RK |
6590 | SUBST (XEXP (x, 0), |
6591 | force_to_mode (XEXP (x, 0), GET_MODE (x), | |
e3d616e3 | 6592 | INTVAL (temp), reg, next_select)); |
6139ff20 RK |
6593 | } |
6594 | break; | |
6595 | ||
dfbe1b2f | 6596 | case NEG: |
180b8e4b RK |
6597 | /* If we just want the low-order bit, the NEG isn't needed since it |
6598 | won't change the low-order bit. */ | |
6599 | if (mask == 1) | |
6600 | return force_to_mode (XEXP (x, 0), mode, mask, reg, just_select); | |
6601 | ||
6139ff20 RK |
6602 | /* We need any bits less significant than the most significant bit in |
6603 | MASK since carries from those bits will affect the bits we are | |
6604 | interested in. */ | |
6605 | mask = fuller_mask; | |
6606 | goto unop; | |
6607 | ||
dfbe1b2f | 6608 | case NOT: |
6139ff20 RK |
6609 | /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the |
6610 | same as the XOR case above. Ensure that the constant we form is not | |
6611 | wider than the mode of X. */ | |
6612 | ||
6613 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
6614 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
6615 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 | |
6616 | && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (mask) | |
6617 | < GET_MODE_BITSIZE (GET_MODE (x))) | |
6618 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT) | |
6619 | { | |
6620 | temp = GEN_INT (mask << INTVAL (XEXP (XEXP (x, 0), 1))); | |
6621 | temp = gen_binary (XOR, GET_MODE (x), XEXP (XEXP (x, 0), 0), temp); | |
6622 | x = gen_binary (LSHIFTRT, GET_MODE (x), temp, XEXP (XEXP (x, 0), 1)); | |
6623 | ||
e3d616e3 | 6624 | return force_to_mode (x, mode, mask, reg, next_select); |
6139ff20 RK |
6625 | } |
6626 | ||
f82da7d2 JW |
6627 | /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must |
6628 | use the full mask inside the NOT. */ | |
6629 | mask = fuller_mask; | |
6630 | ||
6139ff20 | 6631 | unop: |
e3d616e3 RK |
6632 | op0 = gen_lowpart_for_combine (op_mode, |
6633 | force_to_mode (XEXP (x, 0), mode, mask, | |
6634 | reg, next_select)); | |
6139ff20 | 6635 | if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0)) |
0c1c8ea6 | 6636 | x = gen_unary (code, op_mode, op_mode, op0); |
6139ff20 RK |
6637 | break; |
6638 | ||
6639 | case NE: | |
6640 | /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included | |
3aceff0d | 6641 | in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero, |
1a6ec070 | 6642 | which is equal to STORE_FLAG_VALUE. */ |
3aceff0d RK |
6643 | if ((mask & ~ STORE_FLAG_VALUE) == 0 && XEXP (x, 1) == const0_rtx |
6644 | && exact_log2 (nonzero_bits (XEXP (x, 0), mode)) >= 0 | |
1a6ec070 | 6645 | && nonzero_bits (XEXP (x, 0), mode) == STORE_FLAG_VALUE) |
e3d616e3 | 6646 | return force_to_mode (XEXP (x, 0), mode, mask, reg, next_select); |
6139ff20 | 6647 | |
d0ab8cd3 RK |
6648 | break; |
6649 | ||
6650 | case IF_THEN_ELSE: | |
6651 | /* We have no way of knowing if the IF_THEN_ELSE can itself be | |
6652 | written in a narrower mode. We play it safe and do not do so. */ | |
6653 | ||
6654 | SUBST (XEXP (x, 1), | |
6655 | gen_lowpart_for_combine (GET_MODE (x), | |
6656 | force_to_mode (XEXP (x, 1), mode, | |
e3d616e3 | 6657 | mask, reg, next_select))); |
d0ab8cd3 RK |
6658 | SUBST (XEXP (x, 2), |
6659 | gen_lowpart_for_combine (GET_MODE (x), | |
6660 | force_to_mode (XEXP (x, 2), mode, | |
e3d616e3 | 6661 | mask, reg,next_select))); |
d0ab8cd3 | 6662 | break; |
e9a25f70 JL |
6663 | |
6664 | default: | |
6665 | break; | |
dfbe1b2f RK |
6666 | } |
6667 | ||
d0ab8cd3 | 6668 | /* Ensure we return a value of the proper mode. */ |
dfbe1b2f RK |
6669 | return gen_lowpart_for_combine (mode, x); |
6670 | } | |
6671 | \f | |
abe6e52f RK |
6672 | /* Return nonzero if X is an expression that has one of two values depending on |
6673 | whether some other value is zero or nonzero. In that case, we return the | |
6674 | value that is being tested, *PTRUE is set to the value if the rtx being | |
6675 | returned has a nonzero value, and *PFALSE is set to the other alternative. | |
6676 | ||
6677 | If we return zero, we set *PTRUE and *PFALSE to X. */ | |
6678 | ||
6679 | static rtx | |
6680 | if_then_else_cond (x, ptrue, pfalse) | |
6681 | rtx x; | |
6682 | rtx *ptrue, *pfalse; | |
6683 | { | |
6684 | enum machine_mode mode = GET_MODE (x); | |
6685 | enum rtx_code code = GET_CODE (x); | |
6686 | int size = GET_MODE_BITSIZE (mode); | |
6687 | rtx cond0, cond1, true0, true1, false0, false1; | |
6688 | unsigned HOST_WIDE_INT nz; | |
6689 | ||
6690 | /* If this is a unary operation whose operand has one of two values, apply | |
6691 | our opcode to compute those values. */ | |
6692 | if (GET_RTX_CLASS (code) == '1' | |
6693 | && (cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0)) != 0) | |
6694 | { | |
0c1c8ea6 RK |
6695 | *ptrue = gen_unary (code, mode, GET_MODE (XEXP (x, 0)), true0); |
6696 | *pfalse = gen_unary (code, mode, GET_MODE (XEXP (x, 0)), false0); | |
abe6e52f RK |
6697 | return cond0; |
6698 | } | |
6699 | ||
3a19aabc | 6700 | /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would |
ddd5a7c1 | 6701 | make can't possibly match and would suppress other optimizations. */ |
3a19aabc RK |
6702 | else if (code == COMPARE) |
6703 | ; | |
6704 | ||
abe6e52f RK |
6705 | /* If this is a binary operation, see if either side has only one of two |
6706 | values. If either one does or if both do and they are conditional on | |
6707 | the same value, compute the new true and false values. */ | |
6708 | else if (GET_RTX_CLASS (code) == 'c' || GET_RTX_CLASS (code) == '2' | |
6709 | || GET_RTX_CLASS (code) == '<') | |
6710 | { | |
6711 | cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0); | |
6712 | cond1 = if_then_else_cond (XEXP (x, 1), &true1, &false1); | |
6713 | ||
6714 | if ((cond0 != 0 || cond1 != 0) | |
6715 | && ! (cond0 != 0 && cond1 != 0 && ! rtx_equal_p (cond0, cond1))) | |
6716 | { | |
987e845a JW |
6717 | /* If if_then_else_cond returned zero, then true/false are the |
6718 | same rtl. We must copy one of them to prevent invalid rtl | |
6719 | sharing. */ | |
6720 | if (cond0 == 0) | |
6721 | true0 = copy_rtx (true0); | |
6722 | else if (cond1 == 0) | |
6723 | true1 = copy_rtx (true1); | |
6724 | ||
abe6e52f RK |
6725 | *ptrue = gen_binary (code, mode, true0, true1); |
6726 | *pfalse = gen_binary (code, mode, false0, false1); | |
6727 | return cond0 ? cond0 : cond1; | |
6728 | } | |
9210df58 | 6729 | |
9210df58 | 6730 | /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the |
0802d516 RK |
6731 | operands is zero when the other is non-zero, and vice-versa, |
6732 | and STORE_FLAG_VALUE is 1 or -1. */ | |
9210df58 | 6733 | |
0802d516 RK |
6734 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
6735 | && (code == PLUS || code == IOR || code == XOR || code == MINUS | |
9210df58 RK |
6736 | || code == UMAX) |
6737 | && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT) | |
6738 | { | |
6739 | rtx op0 = XEXP (XEXP (x, 0), 1); | |
6740 | rtx op1 = XEXP (XEXP (x, 1), 1); | |
6741 | ||
6742 | cond0 = XEXP (XEXP (x, 0), 0); | |
6743 | cond1 = XEXP (XEXP (x, 1), 0); | |
6744 | ||
6745 | if (GET_RTX_CLASS (GET_CODE (cond0)) == '<' | |
6746 | && GET_RTX_CLASS (GET_CODE (cond1)) == '<' | |
6747 | && reversible_comparison_p (cond1) | |
6748 | && ((GET_CODE (cond0) == reverse_condition (GET_CODE (cond1)) | |
6749 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0)) | |
6750 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1))) | |
6751 | || ((swap_condition (GET_CODE (cond0)) | |
6752 | == reverse_condition (GET_CODE (cond1))) | |
6753 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1)) | |
6754 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0)))) | |
6755 | && ! side_effects_p (x)) | |
6756 | { | |
6757 | *ptrue = gen_binary (MULT, mode, op0, const_true_rtx); | |
6758 | *pfalse = gen_binary (MULT, mode, | |
6759 | (code == MINUS | |
0c1c8ea6 | 6760 | ? gen_unary (NEG, mode, mode, op1) : op1), |
9210df58 RK |
6761 | const_true_rtx); |
6762 | return cond0; | |
6763 | } | |
6764 | } | |
6765 | ||
6766 | /* Similarly for MULT, AND and UMIN, execpt that for these the result | |
6767 | is always zero. */ | |
0802d516 RK |
6768 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
6769 | && (code == MULT || code == AND || code == UMIN) | |
9210df58 RK |
6770 | && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT) |
6771 | { | |
6772 | cond0 = XEXP (XEXP (x, 0), 0); | |
6773 | cond1 = XEXP (XEXP (x, 1), 0); | |
6774 | ||
6775 | if (GET_RTX_CLASS (GET_CODE (cond0)) == '<' | |
6776 | && GET_RTX_CLASS (GET_CODE (cond1)) == '<' | |
6777 | && reversible_comparison_p (cond1) | |
6778 | && ((GET_CODE (cond0) == reverse_condition (GET_CODE (cond1)) | |
6779 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0)) | |
6780 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1))) | |
6781 | || ((swap_condition (GET_CODE (cond0)) | |
6782 | == reverse_condition (GET_CODE (cond1))) | |
6783 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1)) | |
6784 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0)))) | |
6785 | && ! side_effects_p (x)) | |
6786 | { | |
6787 | *ptrue = *pfalse = const0_rtx; | |
6788 | return cond0; | |
6789 | } | |
6790 | } | |
abe6e52f RK |
6791 | } |
6792 | ||
6793 | else if (code == IF_THEN_ELSE) | |
6794 | { | |
6795 | /* If we have IF_THEN_ELSE already, extract the condition and | |
6796 | canonicalize it if it is NE or EQ. */ | |
6797 | cond0 = XEXP (x, 0); | |
6798 | *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2); | |
6799 | if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx) | |
6800 | return XEXP (cond0, 0); | |
6801 | else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx) | |
6802 | { | |
6803 | *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1); | |
6804 | return XEXP (cond0, 0); | |
6805 | } | |
6806 | else | |
6807 | return cond0; | |
6808 | } | |
6809 | ||
6810 | /* If X is a normal SUBREG with both inner and outer modes integral, | |
6811 | we can narrow both the true and false values of the inner expression, | |
6812 | if there is a condition. */ | |
6813 | else if (code == SUBREG && GET_MODE_CLASS (mode) == MODE_INT | |
6814 | && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT | |
6815 | && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))) | |
6816 | && 0 != (cond0 = if_then_else_cond (SUBREG_REG (x), | |
6817 | &true0, &false0))) | |
6818 | { | |
00244e6b RK |
6819 | *ptrue = force_to_mode (true0, mode, GET_MODE_MASK (mode), NULL_RTX, 0); |
6820 | *pfalse | |
6821 | = force_to_mode (false0, mode, GET_MODE_MASK (mode), NULL_RTX, 0); | |
abe6e52f | 6822 | |
abe6e52f RK |
6823 | return cond0; |
6824 | } | |
6825 | ||
6826 | /* If X is a constant, this isn't special and will cause confusions | |
6827 | if we treat it as such. Likewise if it is equivalent to a constant. */ | |
6828 | else if (CONSTANT_P (x) | |
6829 | || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0))) | |
6830 | ; | |
6831 | ||
6832 | /* If X is known to be either 0 or -1, those are the true and | |
6833 | false values when testing X. */ | |
6834 | else if (num_sign_bit_copies (x, mode) == size) | |
6835 | { | |
6836 | *ptrue = constm1_rtx, *pfalse = const0_rtx; | |
6837 | return x; | |
6838 | } | |
6839 | ||
6840 | /* Likewise for 0 or a single bit. */ | |
6841 | else if (exact_log2 (nz = nonzero_bits (x, mode)) >= 0) | |
6842 | { | |
6843 | *ptrue = GEN_INT (nz), *pfalse = const0_rtx; | |
6844 | return x; | |
6845 | } | |
6846 | ||
6847 | /* Otherwise fail; show no condition with true and false values the same. */ | |
6848 | *ptrue = *pfalse = x; | |
6849 | return 0; | |
6850 | } | |
6851 | \f | |
1a26b032 RK |
6852 | /* Return the value of expression X given the fact that condition COND |
6853 | is known to be true when applied to REG as its first operand and VAL | |
6854 | as its second. X is known to not be shared and so can be modified in | |
6855 | place. | |
6856 | ||
6857 | We only handle the simplest cases, and specifically those cases that | |
6858 | arise with IF_THEN_ELSE expressions. */ | |
6859 | ||
6860 | static rtx | |
6861 | known_cond (x, cond, reg, val) | |
6862 | rtx x; | |
6863 | enum rtx_code cond; | |
6864 | rtx reg, val; | |
6865 | { | |
6866 | enum rtx_code code = GET_CODE (x); | |
f24ad0e4 | 6867 | rtx temp; |
1a26b032 RK |
6868 | char *fmt; |
6869 | int i, j; | |
6870 | ||
6871 | if (side_effects_p (x)) | |
6872 | return x; | |
6873 | ||
6874 | if (cond == EQ && rtx_equal_p (x, reg)) | |
6875 | return val; | |
6876 | ||
6877 | /* If X is (abs REG) and we know something about REG's relationship | |
6878 | with zero, we may be able to simplify this. */ | |
6879 | ||
6880 | if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx) | |
6881 | switch (cond) | |
6882 | { | |
6883 | case GE: case GT: case EQ: | |
6884 | return XEXP (x, 0); | |
6885 | case LT: case LE: | |
0c1c8ea6 RK |
6886 | return gen_unary (NEG, GET_MODE (XEXP (x, 0)), GET_MODE (XEXP (x, 0)), |
6887 | XEXP (x, 0)); | |
e9a25f70 JL |
6888 | default: |
6889 | break; | |
1a26b032 RK |
6890 | } |
6891 | ||
6892 | /* The only other cases we handle are MIN, MAX, and comparisons if the | |
6893 | operands are the same as REG and VAL. */ | |
6894 | ||
6895 | else if (GET_RTX_CLASS (code) == '<' || GET_RTX_CLASS (code) == 'c') | |
6896 | { | |
6897 | if (rtx_equal_p (XEXP (x, 0), val)) | |
6898 | cond = swap_condition (cond), temp = val, val = reg, reg = temp; | |
6899 | ||
6900 | if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val)) | |
6901 | { | |
6902 | if (GET_RTX_CLASS (code) == '<') | |
6903 | return (comparison_dominates_p (cond, code) ? const_true_rtx | |
6904 | : (comparison_dominates_p (cond, | |
6905 | reverse_condition (code)) | |
6906 | ? const0_rtx : x)); | |
6907 | ||
6908 | else if (code == SMAX || code == SMIN | |
6909 | || code == UMIN || code == UMAX) | |
6910 | { | |
6911 | int unsignedp = (code == UMIN || code == UMAX); | |
6912 | ||
6913 | if (code == SMAX || code == UMAX) | |
6914 | cond = reverse_condition (cond); | |
6915 | ||
6916 | switch (cond) | |
6917 | { | |
6918 | case GE: case GT: | |
6919 | return unsignedp ? x : XEXP (x, 1); | |
6920 | case LE: case LT: | |
6921 | return unsignedp ? x : XEXP (x, 0); | |
6922 | case GEU: case GTU: | |
6923 | return unsignedp ? XEXP (x, 1) : x; | |
6924 | case LEU: case LTU: | |
6925 | return unsignedp ? XEXP (x, 0) : x; | |
e9a25f70 JL |
6926 | default: |
6927 | break; | |
1a26b032 RK |
6928 | } |
6929 | } | |
6930 | } | |
6931 | } | |
6932 | ||
6933 | fmt = GET_RTX_FORMAT (code); | |
6934 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
6935 | { | |
6936 | if (fmt[i] == 'e') | |
6937 | SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val)); | |
6938 | else if (fmt[i] == 'E') | |
6939 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
6940 | SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j), | |
6941 | cond, reg, val)); | |
6942 | } | |
6943 | ||
6944 | return x; | |
6945 | } | |
6946 | \f | |
e11fa86f RK |
6947 | /* See if X and Y are equal for the purposes of seeing if we can rewrite an |
6948 | assignment as a field assignment. */ | |
6949 | ||
6950 | static int | |
6951 | rtx_equal_for_field_assignment_p (x, y) | |
6952 | rtx x; | |
6953 | rtx y; | |
6954 | { | |
e11fa86f RK |
6955 | if (x == y || rtx_equal_p (x, y)) |
6956 | return 1; | |
6957 | ||
6958 | if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y)) | |
6959 | return 0; | |
6960 | ||
6961 | /* Check for a paradoxical SUBREG of a MEM compared with the MEM. | |
6962 | Note that all SUBREGs of MEM are paradoxical; otherwise they | |
6963 | would have been rewritten. */ | |
6964 | if (GET_CODE (x) == MEM && GET_CODE (y) == SUBREG | |
6965 | && GET_CODE (SUBREG_REG (y)) == MEM | |
6966 | && rtx_equal_p (SUBREG_REG (y), | |
6967 | gen_lowpart_for_combine (GET_MODE (SUBREG_REG (y)), x))) | |
6968 | return 1; | |
6969 | ||
6970 | if (GET_CODE (y) == MEM && GET_CODE (x) == SUBREG | |
6971 | && GET_CODE (SUBREG_REG (x)) == MEM | |
6972 | && rtx_equal_p (SUBREG_REG (x), | |
6973 | gen_lowpart_for_combine (GET_MODE (SUBREG_REG (x)), y))) | |
6974 | return 1; | |
6975 | ||
9ec36da5 JL |
6976 | /* We used to see if get_last_value of X and Y were the same but that's |
6977 | not correct. In one direction, we'll cause the assignment to have | |
6978 | the wrong destination and in the case, we'll import a register into this | |
6979 | insn that might have already have been dead. So fail if none of the | |
6980 | above cases are true. */ | |
6981 | return 0; | |
e11fa86f RK |
6982 | } |
6983 | \f | |
230d793d RS |
6984 | /* See if X, a SET operation, can be rewritten as a bit-field assignment. |
6985 | Return that assignment if so. | |
6986 | ||
6987 | We only handle the most common cases. */ | |
6988 | ||
6989 | static rtx | |
6990 | make_field_assignment (x) | |
6991 | rtx x; | |
6992 | { | |
6993 | rtx dest = SET_DEST (x); | |
6994 | rtx src = SET_SRC (x); | |
dfbe1b2f | 6995 | rtx assign; |
e11fa86f | 6996 | rtx rhs, lhs; |
5f4f0e22 CH |
6997 | HOST_WIDE_INT c1; |
6998 | int pos, len; | |
dfbe1b2f RK |
6999 | rtx other; |
7000 | enum machine_mode mode; | |
230d793d RS |
7001 | |
7002 | /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is | |
7003 | a clear of a one-bit field. We will have changed it to | |
7004 | (and (rotate (const_int -2) POS) DEST), so check for that. Also check | |
7005 | for a SUBREG. */ | |
7006 | ||
7007 | if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE | |
7008 | && GET_CODE (XEXP (XEXP (src, 0), 0)) == CONST_INT | |
7009 | && INTVAL (XEXP (XEXP (src, 0), 0)) == -2 | |
e11fa86f | 7010 | && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1))) |
230d793d | 7011 | { |
8999a12e | 7012 | assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1), |
230d793d | 7013 | 1, 1, 1, 0); |
76184def | 7014 | if (assign != 0) |
38a448ca | 7015 | return gen_rtx_SET (VOIDmode, assign, const0_rtx); |
76184def | 7016 | return x; |
230d793d RS |
7017 | } |
7018 | ||
7019 | else if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG | |
7020 | && subreg_lowpart_p (XEXP (src, 0)) | |
7021 | && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0))) | |
7022 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0))))) | |
7023 | && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE | |
7024 | && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2 | |
e11fa86f | 7025 | && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1))) |
230d793d | 7026 | { |
8999a12e | 7027 | assign = make_extraction (VOIDmode, dest, 0, |
230d793d RS |
7028 | XEXP (SUBREG_REG (XEXP (src, 0)), 1), |
7029 | 1, 1, 1, 0); | |
76184def | 7030 | if (assign != 0) |
38a448ca | 7031 | return gen_rtx_SET (VOIDmode, assign, const0_rtx); |
76184def | 7032 | return x; |
230d793d RS |
7033 | } |
7034 | ||
9dd11dcb | 7035 | /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a |
230d793d RS |
7036 | one-bit field. */ |
7037 | else if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT | |
7038 | && XEXP (XEXP (src, 0), 0) == const1_rtx | |
e11fa86f | 7039 | && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1))) |
230d793d | 7040 | { |
8999a12e | 7041 | assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1), |
230d793d | 7042 | 1, 1, 1, 0); |
76184def | 7043 | if (assign != 0) |
38a448ca | 7044 | return gen_rtx_SET (VOIDmode, assign, const1_rtx); |
76184def | 7045 | return x; |
230d793d RS |
7046 | } |
7047 | ||
dfbe1b2f | 7048 | /* The other case we handle is assignments into a constant-position |
9dd11dcb | 7049 | field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents |
dfbe1b2f RK |
7050 | a mask that has all one bits except for a group of zero bits and |
7051 | OTHER is known to have zeros where C1 has ones, this is such an | |
7052 | assignment. Compute the position and length from C1. Shift OTHER | |
7053 | to the appropriate position, force it to the required mode, and | |
7054 | make the extraction. Check for the AND in both operands. */ | |
7055 | ||
9dd11dcb | 7056 | if (GET_CODE (src) != IOR && GET_CODE (src) != XOR) |
e11fa86f RK |
7057 | return x; |
7058 | ||
7059 | rhs = expand_compound_operation (XEXP (src, 0)); | |
7060 | lhs = expand_compound_operation (XEXP (src, 1)); | |
7061 | ||
7062 | if (GET_CODE (rhs) == AND | |
7063 | && GET_CODE (XEXP (rhs, 1)) == CONST_INT | |
7064 | && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), dest)) | |
7065 | c1 = INTVAL (XEXP (rhs, 1)), other = lhs; | |
7066 | else if (GET_CODE (lhs) == AND | |
7067 | && GET_CODE (XEXP (lhs, 1)) == CONST_INT | |
7068 | && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), dest)) | |
7069 | c1 = INTVAL (XEXP (lhs, 1)), other = rhs; | |
dfbe1b2f RK |
7070 | else |
7071 | return x; | |
230d793d | 7072 | |
e11fa86f | 7073 | pos = get_pos_from_mask ((~ c1) & GET_MODE_MASK (GET_MODE (dest)), &len); |
dfbe1b2f | 7074 | if (pos < 0 || pos + len > GET_MODE_BITSIZE (GET_MODE (dest)) |
e5e809f4 JL |
7075 | || GET_MODE_BITSIZE (GET_MODE (dest)) > HOST_BITS_PER_WIDE_INT |
7076 | || (c1 & nonzero_bits (other, GET_MODE (dest))) != 0) | |
dfbe1b2f | 7077 | return x; |
230d793d | 7078 | |
5f4f0e22 | 7079 | assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0); |
76184def DE |
7080 | if (assign == 0) |
7081 | return x; | |
230d793d | 7082 | |
dfbe1b2f RK |
7083 | /* The mode to use for the source is the mode of the assignment, or of |
7084 | what is inside a possible STRICT_LOW_PART. */ | |
7085 | mode = (GET_CODE (assign) == STRICT_LOW_PART | |
7086 | ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign)); | |
230d793d | 7087 | |
dfbe1b2f RK |
7088 | /* Shift OTHER right POS places and make it the source, restricting it |
7089 | to the proper length and mode. */ | |
230d793d | 7090 | |
5f4f0e22 CH |
7091 | src = force_to_mode (simplify_shift_const (NULL_RTX, LSHIFTRT, |
7092 | GET_MODE (src), other, pos), | |
6139ff20 RK |
7093 | mode, |
7094 | GET_MODE_BITSIZE (mode) >= HOST_BITS_PER_WIDE_INT | |
7095 | ? GET_MODE_MASK (mode) | |
7096 | : ((HOST_WIDE_INT) 1 << len) - 1, | |
e3d616e3 | 7097 | dest, 0); |
230d793d | 7098 | |
dfbe1b2f | 7099 | return gen_rtx_combine (SET, VOIDmode, assign, src); |
230d793d RS |
7100 | } |
7101 | \f | |
7102 | /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c) | |
7103 | if so. */ | |
7104 | ||
7105 | static rtx | |
7106 | apply_distributive_law (x) | |
7107 | rtx x; | |
7108 | { | |
7109 | enum rtx_code code = GET_CODE (x); | |
7110 | rtx lhs, rhs, other; | |
7111 | rtx tem; | |
7112 | enum rtx_code inner_code; | |
7113 | ||
d8a8a4da RS |
7114 | /* Distributivity is not true for floating point. |
7115 | It can change the value. So don't do it. | |
7116 | -- rms and moshier@world.std.com. */ | |
3ad2180a | 7117 | if (FLOAT_MODE_P (GET_MODE (x))) |
d8a8a4da RS |
7118 | return x; |
7119 | ||
230d793d RS |
7120 | /* The outer operation can only be one of the following: */ |
7121 | if (code != IOR && code != AND && code != XOR | |
7122 | && code != PLUS && code != MINUS) | |
7123 | return x; | |
7124 | ||
7125 | lhs = XEXP (x, 0), rhs = XEXP (x, 1); | |
7126 | ||
0f41302f MS |
7127 | /* If either operand is a primitive we can't do anything, so get out |
7128 | fast. */ | |
230d793d | 7129 | if (GET_RTX_CLASS (GET_CODE (lhs)) == 'o' |
dfbe1b2f | 7130 | || GET_RTX_CLASS (GET_CODE (rhs)) == 'o') |
230d793d RS |
7131 | return x; |
7132 | ||
7133 | lhs = expand_compound_operation (lhs); | |
7134 | rhs = expand_compound_operation (rhs); | |
7135 | inner_code = GET_CODE (lhs); | |
7136 | if (inner_code != GET_CODE (rhs)) | |
7137 | return x; | |
7138 | ||
7139 | /* See if the inner and outer operations distribute. */ | |
7140 | switch (inner_code) | |
7141 | { | |
7142 | case LSHIFTRT: | |
7143 | case ASHIFTRT: | |
7144 | case AND: | |
7145 | case IOR: | |
7146 | /* These all distribute except over PLUS. */ | |
7147 | if (code == PLUS || code == MINUS) | |
7148 | return x; | |
7149 | break; | |
7150 | ||
7151 | case MULT: | |
7152 | if (code != PLUS && code != MINUS) | |
7153 | return x; | |
7154 | break; | |
7155 | ||
7156 | case ASHIFT: | |
45620ed4 | 7157 | /* This is also a multiply, so it distributes over everything. */ |
230d793d RS |
7158 | break; |
7159 | ||
7160 | case SUBREG: | |
dfbe1b2f RK |
7161 | /* Non-paradoxical SUBREGs distributes over all operations, provided |
7162 | the inner modes and word numbers are the same, this is an extraction | |
2b4bd1bc JW |
7163 | of a low-order part, we don't convert an fp operation to int or |
7164 | vice versa, and we would not be converting a single-word | |
dfbe1b2f | 7165 | operation into a multi-word operation. The latter test is not |
2b4bd1bc | 7166 | required, but it prevents generating unneeded multi-word operations. |
dfbe1b2f RK |
7167 | Some of the previous tests are redundant given the latter test, but |
7168 | are retained because they are required for correctness. | |
7169 | ||
7170 | We produce the result slightly differently in this case. */ | |
7171 | ||
7172 | if (GET_MODE (SUBREG_REG (lhs)) != GET_MODE (SUBREG_REG (rhs)) | |
7173 | || SUBREG_WORD (lhs) != SUBREG_WORD (rhs) | |
7174 | || ! subreg_lowpart_p (lhs) | |
2b4bd1bc JW |
7175 | || (GET_MODE_CLASS (GET_MODE (lhs)) |
7176 | != GET_MODE_CLASS (GET_MODE (SUBREG_REG (lhs)))) | |
dfbe1b2f | 7177 | || (GET_MODE_SIZE (GET_MODE (lhs)) |
8af24e26 | 7178 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs)))) |
dfbe1b2f | 7179 | || GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))) > UNITS_PER_WORD) |
230d793d RS |
7180 | return x; |
7181 | ||
7182 | tem = gen_binary (code, GET_MODE (SUBREG_REG (lhs)), | |
7183 | SUBREG_REG (lhs), SUBREG_REG (rhs)); | |
7184 | return gen_lowpart_for_combine (GET_MODE (x), tem); | |
7185 | ||
7186 | default: | |
7187 | return x; | |
7188 | } | |
7189 | ||
7190 | /* Set LHS and RHS to the inner operands (A and B in the example | |
7191 | above) and set OTHER to the common operand (C in the example). | |
7192 | These is only one way to do this unless the inner operation is | |
7193 | commutative. */ | |
7194 | if (GET_RTX_CLASS (inner_code) == 'c' | |
7195 | && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0))) | |
7196 | other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1); | |
7197 | else if (GET_RTX_CLASS (inner_code) == 'c' | |
7198 | && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1))) | |
7199 | other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0); | |
7200 | else if (GET_RTX_CLASS (inner_code) == 'c' | |
7201 | && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0))) | |
7202 | other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1); | |
7203 | else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1))) | |
7204 | other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0); | |
7205 | else | |
7206 | return x; | |
7207 | ||
7208 | /* Form the new inner operation, seeing if it simplifies first. */ | |
7209 | tem = gen_binary (code, GET_MODE (x), lhs, rhs); | |
7210 | ||
7211 | /* There is one exception to the general way of distributing: | |
7212 | (a ^ b) | (a ^ c) -> (~a) & (b ^ c) */ | |
7213 | if (code == XOR && inner_code == IOR) | |
7214 | { | |
7215 | inner_code = AND; | |
0c1c8ea6 | 7216 | other = gen_unary (NOT, GET_MODE (x), GET_MODE (x), other); |
230d793d RS |
7217 | } |
7218 | ||
7219 | /* We may be able to continuing distributing the result, so call | |
7220 | ourselves recursively on the inner operation before forming the | |
7221 | outer operation, which we return. */ | |
7222 | return gen_binary (inner_code, GET_MODE (x), | |
7223 | apply_distributive_law (tem), other); | |
7224 | } | |
7225 | \f | |
7226 | /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done | |
7227 | in MODE. | |
7228 | ||
7229 | Return an equivalent form, if different from X. Otherwise, return X. If | |
7230 | X is zero, we are to always construct the equivalent form. */ | |
7231 | ||
7232 | static rtx | |
7233 | simplify_and_const_int (x, mode, varop, constop) | |
7234 | rtx x; | |
7235 | enum machine_mode mode; | |
7236 | rtx varop; | |
5f4f0e22 | 7237 | unsigned HOST_WIDE_INT constop; |
230d793d | 7238 | { |
951553af | 7239 | unsigned HOST_WIDE_INT nonzero; |
9fa6d012 | 7240 | int width = GET_MODE_BITSIZE (mode); |
42301240 | 7241 | int i; |
230d793d | 7242 | |
6139ff20 RK |
7243 | /* Simplify VAROP knowing that we will be only looking at some of the |
7244 | bits in it. */ | |
e3d616e3 | 7245 | varop = force_to_mode (varop, mode, constop, NULL_RTX, 0); |
230d793d | 7246 | |
6139ff20 RK |
7247 | /* If VAROP is a CLOBBER, we will fail so return it; if it is a |
7248 | CONST_INT, we are done. */ | |
7249 | if (GET_CODE (varop) == CLOBBER || GET_CODE (varop) == CONST_INT) | |
7250 | return varop; | |
230d793d | 7251 | |
fc06d7aa RK |
7252 | /* See what bits may be nonzero in VAROP. Unlike the general case of |
7253 | a call to nonzero_bits, here we don't care about bits outside | |
7254 | MODE. */ | |
7255 | ||
7256 | nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode); | |
230d793d | 7257 | |
9fa6d012 TG |
7258 | /* If this would be an entire word for the target, but is not for |
7259 | the host, then sign-extend on the host so that the number will look | |
7260 | the same way on the host that it would on the target. | |
7261 | ||
7262 | For example, when building a 64 bit alpha hosted 32 bit sparc | |
7263 | targeted compiler, then we want the 32 bit unsigned value -1 to be | |
7264 | represented as a 64 bit value -1, and not as 0x00000000ffffffff. | |
7265 | The later confuses the sparc backend. */ | |
7266 | ||
7267 | if (BITS_PER_WORD < HOST_BITS_PER_WIDE_INT && BITS_PER_WORD == width | |
7268 | && (nonzero & ((HOST_WIDE_INT) 1 << (width - 1)))) | |
7269 | nonzero |= ((HOST_WIDE_INT) (-1) << width); | |
7270 | ||
230d793d | 7271 | /* Turn off all bits in the constant that are known to already be zero. |
951553af | 7272 | Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS |
230d793d RS |
7273 | which is tested below. */ |
7274 | ||
951553af | 7275 | constop &= nonzero; |
230d793d RS |
7276 | |
7277 | /* If we don't have any bits left, return zero. */ | |
7278 | if (constop == 0) | |
7279 | return const0_rtx; | |
7280 | ||
42301240 RK |
7281 | /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is |
7282 | a power of two, we can replace this with a ASHIFT. */ | |
7283 | if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), mode) == 1 | |
7284 | && (i = exact_log2 (constop)) >= 0) | |
7285 | return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i); | |
7286 | ||
6139ff20 RK |
7287 | /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR |
7288 | or XOR, then try to apply the distributive law. This may eliminate | |
7289 | operations if either branch can be simplified because of the AND. | |
7290 | It may also make some cases more complex, but those cases probably | |
7291 | won't match a pattern either with or without this. */ | |
7292 | ||
7293 | if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR) | |
7294 | return | |
7295 | gen_lowpart_for_combine | |
7296 | (mode, | |
7297 | apply_distributive_law | |
7298 | (gen_binary (GET_CODE (varop), GET_MODE (varop), | |
7299 | simplify_and_const_int (NULL_RTX, GET_MODE (varop), | |
7300 | XEXP (varop, 0), constop), | |
7301 | simplify_and_const_int (NULL_RTX, GET_MODE (varop), | |
7302 | XEXP (varop, 1), constop)))); | |
7303 | ||
230d793d RS |
7304 | /* Get VAROP in MODE. Try to get a SUBREG if not. Don't make a new SUBREG |
7305 | if we already had one (just check for the simplest cases). */ | |
7306 | if (x && GET_CODE (XEXP (x, 0)) == SUBREG | |
7307 | && GET_MODE (XEXP (x, 0)) == mode | |
7308 | && SUBREG_REG (XEXP (x, 0)) == varop) | |
7309 | varop = XEXP (x, 0); | |
7310 | else | |
7311 | varop = gen_lowpart_for_combine (mode, varop); | |
7312 | ||
0f41302f | 7313 | /* If we can't make the SUBREG, try to return what we were given. */ |
230d793d RS |
7314 | if (GET_CODE (varop) == CLOBBER) |
7315 | return x ? x : varop; | |
7316 | ||
7317 | /* If we are only masking insignificant bits, return VAROP. */ | |
951553af | 7318 | if (constop == nonzero) |
230d793d RS |
7319 | x = varop; |
7320 | ||
7321 | /* Otherwise, return an AND. See how much, if any, of X we can use. */ | |
7322 | else if (x == 0 || GET_CODE (x) != AND || GET_MODE (x) != mode) | |
6139ff20 | 7323 | x = gen_binary (AND, mode, varop, GEN_INT (constop)); |
230d793d RS |
7324 | |
7325 | else | |
7326 | { | |
7327 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
7328 | || INTVAL (XEXP (x, 1)) != constop) | |
5f4f0e22 | 7329 | SUBST (XEXP (x, 1), GEN_INT (constop)); |
230d793d RS |
7330 | |
7331 | SUBST (XEXP (x, 0), varop); | |
7332 | } | |
7333 | ||
7334 | return x; | |
7335 | } | |
7336 | \f | |
b3728b0e JW |
7337 | /* We let num_sign_bit_copies recur into nonzero_bits as that is useful. |
7338 | We don't let nonzero_bits recur into num_sign_bit_copies, because that | |
7339 | is less useful. We can't allow both, because that results in exponential | |
956d6950 | 7340 | run time recursion. There is a nullstone testcase that triggered |
b3728b0e JW |
7341 | this. This macro avoids accidental uses of num_sign_bit_copies. */ |
7342 | #define num_sign_bit_copies() | |
7343 | ||
230d793d RS |
7344 | /* Given an expression, X, compute which bits in X can be non-zero. |
7345 | We don't care about bits outside of those defined in MODE. | |
7346 | ||
7347 | For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is | |
7348 | a shift, AND, or zero_extract, we can do better. */ | |
7349 | ||
5f4f0e22 | 7350 | static unsigned HOST_WIDE_INT |
951553af | 7351 | nonzero_bits (x, mode) |
230d793d RS |
7352 | rtx x; |
7353 | enum machine_mode mode; | |
7354 | { | |
951553af RK |
7355 | unsigned HOST_WIDE_INT nonzero = GET_MODE_MASK (mode); |
7356 | unsigned HOST_WIDE_INT inner_nz; | |
230d793d RS |
7357 | enum rtx_code code; |
7358 | int mode_width = GET_MODE_BITSIZE (mode); | |
7359 | rtx tem; | |
7360 | ||
1c75dfa4 RK |
7361 | /* For floating-point values, assume all bits are needed. */ |
7362 | if (FLOAT_MODE_P (GET_MODE (x)) || FLOAT_MODE_P (mode)) | |
7363 | return nonzero; | |
7364 | ||
230d793d RS |
7365 | /* If X is wider than MODE, use its mode instead. */ |
7366 | if (GET_MODE_BITSIZE (GET_MODE (x)) > mode_width) | |
7367 | { | |
7368 | mode = GET_MODE (x); | |
951553af | 7369 | nonzero = GET_MODE_MASK (mode); |
230d793d RS |
7370 | mode_width = GET_MODE_BITSIZE (mode); |
7371 | } | |
7372 | ||
5f4f0e22 | 7373 | if (mode_width > HOST_BITS_PER_WIDE_INT) |
230d793d RS |
7374 | /* Our only callers in this case look for single bit values. So |
7375 | just return the mode mask. Those tests will then be false. */ | |
951553af | 7376 | return nonzero; |
230d793d | 7377 | |
8baf60bb | 7378 | #ifndef WORD_REGISTER_OPERATIONS |
c6965c0f | 7379 | /* If MODE is wider than X, but both are a single word for both the host |
0840fd91 RK |
7380 | and target machines, we can compute this from which bits of the |
7381 | object might be nonzero in its own mode, taking into account the fact | |
7382 | that on many CISC machines, accessing an object in a wider mode | |
7383 | causes the high-order bits to become undefined. So they are | |
7384 | not known to be zero. */ | |
7385 | ||
7386 | if (GET_MODE (x) != VOIDmode && GET_MODE (x) != mode | |
7387 | && GET_MODE_BITSIZE (GET_MODE (x)) <= BITS_PER_WORD | |
7388 | && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT | |
c6965c0f | 7389 | && GET_MODE_BITSIZE (mode) > GET_MODE_BITSIZE (GET_MODE (x))) |
0840fd91 RK |
7390 | { |
7391 | nonzero &= nonzero_bits (x, GET_MODE (x)); | |
7392 | nonzero |= GET_MODE_MASK (mode) & ~ GET_MODE_MASK (GET_MODE (x)); | |
7393 | return nonzero; | |
7394 | } | |
7395 | #endif | |
7396 | ||
230d793d RS |
7397 | code = GET_CODE (x); |
7398 | switch (code) | |
7399 | { | |
7400 | case REG: | |
320dd7a7 RK |
7401 | #ifdef POINTERS_EXTEND_UNSIGNED |
7402 | /* If pointers extend unsigned and this is a pointer in Pmode, say that | |
7403 | all the bits above ptr_mode are known to be zero. */ | |
7404 | if (POINTERS_EXTEND_UNSIGNED && GET_MODE (x) == Pmode | |
7405 | && REGNO_POINTER_FLAG (REGNO (x))) | |
7406 | nonzero &= GET_MODE_MASK (ptr_mode); | |
7407 | #endif | |
7408 | ||
b0d71df9 RK |
7409 | #ifdef STACK_BOUNDARY |
7410 | /* If this is the stack pointer, we may know something about its | |
7411 | alignment. If PUSH_ROUNDING is defined, it is possible for the | |
230d793d RS |
7412 | stack to be momentarily aligned only to that amount, so we pick |
7413 | the least alignment. */ | |
7414 | ||
ee49a9c7 JW |
7415 | /* We can't check for arg_pointer_rtx here, because it is not |
7416 | guaranteed to have as much alignment as the stack pointer. | |
7417 | In particular, in the Irix6 n64 ABI, the stack has 128 bit | |
7418 | alignment but the argument pointer has only 64 bit alignment. */ | |
7419 | ||
0e9ff885 DM |
7420 | if ((x == frame_pointer_rtx |
7421 | || x == stack_pointer_rtx | |
7422 | || x == hard_frame_pointer_rtx | |
7423 | || (REGNO (x) >= FIRST_VIRTUAL_REGISTER | |
7424 | && REGNO (x) <= LAST_VIRTUAL_REGISTER)) | |
7425 | #ifdef STACK_BIAS | |
7426 | && !STACK_BIAS | |
7427 | #endif | |
7428 | ) | |
230d793d | 7429 | { |
b0d71df9 | 7430 | int sp_alignment = STACK_BOUNDARY / BITS_PER_UNIT; |
230d793d RS |
7431 | |
7432 | #ifdef PUSH_ROUNDING | |
91102d5a | 7433 | if (REGNO (x) == STACK_POINTER_REGNUM) |
b0d71df9 | 7434 | sp_alignment = MIN (PUSH_ROUNDING (1), sp_alignment); |
230d793d RS |
7435 | #endif |
7436 | ||
320dd7a7 RK |
7437 | /* We must return here, otherwise we may get a worse result from |
7438 | one of the choices below. There is nothing useful below as | |
7439 | far as the stack pointer is concerned. */ | |
b0d71df9 | 7440 | return nonzero &= ~ (sp_alignment - 1); |
230d793d | 7441 | } |
b0d71df9 | 7442 | #endif |
230d793d | 7443 | |
55310dad RK |
7444 | /* If X is a register whose nonzero bits value is current, use it. |
7445 | Otherwise, if X is a register whose value we can find, use that | |
7446 | value. Otherwise, use the previously-computed global nonzero bits | |
7447 | for this register. */ | |
7448 | ||
7449 | if (reg_last_set_value[REGNO (x)] != 0 | |
7450 | && reg_last_set_mode[REGNO (x)] == mode | |
b1f21e0a | 7451 | && (REG_N_SETS (REGNO (x)) == 1 |
55310dad RK |
7452 | || reg_last_set_label[REGNO (x)] == label_tick) |
7453 | && INSN_CUID (reg_last_set[REGNO (x)]) < subst_low_cuid) | |
7454 | return reg_last_set_nonzero_bits[REGNO (x)]; | |
230d793d RS |
7455 | |
7456 | tem = get_last_value (x); | |
9afa3d54 | 7457 | |
230d793d | 7458 | if (tem) |
9afa3d54 RK |
7459 | { |
7460 | #ifdef SHORT_IMMEDIATES_SIGN_EXTEND | |
7461 | /* If X is narrower than MODE and TEM is a non-negative | |
7462 | constant that would appear negative in the mode of X, | |
7463 | sign-extend it for use in reg_nonzero_bits because some | |
7464 | machines (maybe most) will actually do the sign-extension | |
7465 | and this is the conservative approach. | |
7466 | ||
7467 | ??? For 2.5, try to tighten up the MD files in this regard | |
7468 | instead of this kludge. */ | |
7469 | ||
7470 | if (GET_MODE_BITSIZE (GET_MODE (x)) < mode_width | |
7471 | && GET_CODE (tem) == CONST_INT | |
7472 | && INTVAL (tem) > 0 | |
7473 | && 0 != (INTVAL (tem) | |
7474 | & ((HOST_WIDE_INT) 1 | |
9e69be8c | 7475 | << (GET_MODE_BITSIZE (GET_MODE (x)) - 1)))) |
9afa3d54 RK |
7476 | tem = GEN_INT (INTVAL (tem) |
7477 | | ((HOST_WIDE_INT) (-1) | |
7478 | << GET_MODE_BITSIZE (GET_MODE (x)))); | |
7479 | #endif | |
7480 | return nonzero_bits (tem, mode); | |
7481 | } | |
951553af RK |
7482 | else if (nonzero_sign_valid && reg_nonzero_bits[REGNO (x)]) |
7483 | return reg_nonzero_bits[REGNO (x)] & nonzero; | |
230d793d | 7484 | else |
951553af | 7485 | return nonzero; |
230d793d RS |
7486 | |
7487 | case CONST_INT: | |
9afa3d54 RK |
7488 | #ifdef SHORT_IMMEDIATES_SIGN_EXTEND |
7489 | /* If X is negative in MODE, sign-extend the value. */ | |
9e69be8c RK |
7490 | if (INTVAL (x) > 0 && mode_width < BITS_PER_WORD |
7491 | && 0 != (INTVAL (x) & ((HOST_WIDE_INT) 1 << (mode_width - 1)))) | |
7492 | return (INTVAL (x) | ((HOST_WIDE_INT) (-1) << mode_width)); | |
9afa3d54 RK |
7493 | #endif |
7494 | ||
230d793d RS |
7495 | return INTVAL (x); |
7496 | ||
230d793d | 7497 | case MEM: |
8baf60bb | 7498 | #ifdef LOAD_EXTEND_OP |
230d793d RS |
7499 | /* In many, if not most, RISC machines, reading a byte from memory |
7500 | zeros the rest of the register. Noticing that fact saves a lot | |
7501 | of extra zero-extends. */ | |
8baf60bb RK |
7502 | if (LOAD_EXTEND_OP (GET_MODE (x)) == ZERO_EXTEND) |
7503 | nonzero &= GET_MODE_MASK (GET_MODE (x)); | |
230d793d | 7504 | #endif |
8baf60bb | 7505 | break; |
230d793d | 7506 | |
230d793d RS |
7507 | case EQ: case NE: |
7508 | case GT: case GTU: | |
7509 | case LT: case LTU: | |
7510 | case GE: case GEU: | |
7511 | case LE: case LEU: | |
3f508eca | 7512 | |
c6965c0f RK |
7513 | /* If this produces an integer result, we know which bits are set. |
7514 | Code here used to clear bits outside the mode of X, but that is | |
7515 | now done above. */ | |
230d793d | 7516 | |
c6965c0f RK |
7517 | if (GET_MODE_CLASS (mode) == MODE_INT |
7518 | && mode_width <= HOST_BITS_PER_WIDE_INT) | |
7519 | nonzero = STORE_FLAG_VALUE; | |
230d793d | 7520 | break; |
230d793d | 7521 | |
230d793d | 7522 | case NEG: |
b3728b0e JW |
7523 | #if 0 |
7524 | /* Disabled to avoid exponential mutual recursion between nonzero_bits | |
7525 | and num_sign_bit_copies. */ | |
d0ab8cd3 RK |
7526 | if (num_sign_bit_copies (XEXP (x, 0), GET_MODE (x)) |
7527 | == GET_MODE_BITSIZE (GET_MODE (x))) | |
951553af | 7528 | nonzero = 1; |
b3728b0e | 7529 | #endif |
230d793d RS |
7530 | |
7531 | if (GET_MODE_SIZE (GET_MODE (x)) < mode_width) | |
951553af | 7532 | nonzero |= (GET_MODE_MASK (mode) & ~ GET_MODE_MASK (GET_MODE (x))); |
230d793d | 7533 | break; |
d0ab8cd3 RK |
7534 | |
7535 | case ABS: | |
b3728b0e JW |
7536 | #if 0 |
7537 | /* Disabled to avoid exponential mutual recursion between nonzero_bits | |
7538 | and num_sign_bit_copies. */ | |
d0ab8cd3 RK |
7539 | if (num_sign_bit_copies (XEXP (x, 0), GET_MODE (x)) |
7540 | == GET_MODE_BITSIZE (GET_MODE (x))) | |
951553af | 7541 | nonzero = 1; |
b3728b0e | 7542 | #endif |
d0ab8cd3 | 7543 | break; |
230d793d RS |
7544 | |
7545 | case TRUNCATE: | |
951553af | 7546 | nonzero &= (nonzero_bits (XEXP (x, 0), mode) & GET_MODE_MASK (mode)); |
230d793d RS |
7547 | break; |
7548 | ||
7549 | case ZERO_EXTEND: | |
951553af | 7550 | nonzero &= nonzero_bits (XEXP (x, 0), mode); |
230d793d | 7551 | if (GET_MODE (XEXP (x, 0)) != VOIDmode) |
951553af | 7552 | nonzero &= GET_MODE_MASK (GET_MODE (XEXP (x, 0))); |
230d793d RS |
7553 | break; |
7554 | ||
7555 | case SIGN_EXTEND: | |
7556 | /* If the sign bit is known clear, this is the same as ZERO_EXTEND. | |
7557 | Otherwise, show all the bits in the outer mode but not the inner | |
7558 | may be non-zero. */ | |
951553af | 7559 | inner_nz = nonzero_bits (XEXP (x, 0), mode); |
230d793d RS |
7560 | if (GET_MODE (XEXP (x, 0)) != VOIDmode) |
7561 | { | |
951553af | 7562 | inner_nz &= GET_MODE_MASK (GET_MODE (XEXP (x, 0))); |
e3da301d MS |
7563 | if (inner_nz |
7564 | & (((HOST_WIDE_INT) 1 | |
7565 | << (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - 1)))) | |
951553af | 7566 | inner_nz |= (GET_MODE_MASK (mode) |
230d793d RS |
7567 | & ~ GET_MODE_MASK (GET_MODE (XEXP (x, 0)))); |
7568 | } | |
7569 | ||
951553af | 7570 | nonzero &= inner_nz; |
230d793d RS |
7571 | break; |
7572 | ||
7573 | case AND: | |
951553af RK |
7574 | nonzero &= (nonzero_bits (XEXP (x, 0), mode) |
7575 | & nonzero_bits (XEXP (x, 1), mode)); | |
230d793d RS |
7576 | break; |
7577 | ||
d0ab8cd3 RK |
7578 | case XOR: case IOR: |
7579 | case UMIN: case UMAX: case SMIN: case SMAX: | |
951553af RK |
7580 | nonzero &= (nonzero_bits (XEXP (x, 0), mode) |
7581 | | nonzero_bits (XEXP (x, 1), mode)); | |
230d793d RS |
7582 | break; |
7583 | ||
7584 | case PLUS: case MINUS: | |
7585 | case MULT: | |
7586 | case DIV: case UDIV: | |
7587 | case MOD: case UMOD: | |
7588 | /* We can apply the rules of arithmetic to compute the number of | |
7589 | high- and low-order zero bits of these operations. We start by | |
7590 | computing the width (position of the highest-order non-zero bit) | |
7591 | and the number of low-order zero bits for each value. */ | |
7592 | { | |
951553af RK |
7593 | unsigned HOST_WIDE_INT nz0 = nonzero_bits (XEXP (x, 0), mode); |
7594 | unsigned HOST_WIDE_INT nz1 = nonzero_bits (XEXP (x, 1), mode); | |
7595 | int width0 = floor_log2 (nz0) + 1; | |
7596 | int width1 = floor_log2 (nz1) + 1; | |
7597 | int low0 = floor_log2 (nz0 & -nz0); | |
7598 | int low1 = floor_log2 (nz1 & -nz1); | |
318b149c RK |
7599 | HOST_WIDE_INT op0_maybe_minusp |
7600 | = (nz0 & ((HOST_WIDE_INT) 1 << (mode_width - 1))); | |
7601 | HOST_WIDE_INT op1_maybe_minusp | |
7602 | = (nz1 & ((HOST_WIDE_INT) 1 << (mode_width - 1))); | |
230d793d RS |
7603 | int result_width = mode_width; |
7604 | int result_low = 0; | |
7605 | ||
7606 | switch (code) | |
7607 | { | |
7608 | case PLUS: | |
0e9ff885 DM |
7609 | #ifdef STACK_BIAS |
7610 | if (STACK_BIAS | |
7611 | && (XEXP (x, 0) == stack_pointer_rtx | |
7612 | || XEXP (x, 0) == frame_pointer_rtx) | |
7613 | && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
7614 | { | |
7615 | int sp_alignment = STACK_BOUNDARY / BITS_PER_UNIT; | |
7616 | ||
7617 | nz0 = (GET_MODE_MASK (mode) & ~ (sp_alignment - 1)); | |
7618 | nz1 = INTVAL (XEXP (x, 1)) - STACK_BIAS; | |
7619 | width0 = floor_log2 (nz0) + 1; | |
7620 | width1 = floor_log2 (nz1) + 1; | |
7621 | low0 = floor_log2 (nz0 & -nz0); | |
7622 | low1 = floor_log2 (nz1 & -nz1); | |
7623 | } | |
7624 | #endif | |
230d793d RS |
7625 | result_width = MAX (width0, width1) + 1; |
7626 | result_low = MIN (low0, low1); | |
7627 | break; | |
7628 | case MINUS: | |
7629 | result_low = MIN (low0, low1); | |
7630 | break; | |
7631 | case MULT: | |
7632 | result_width = width0 + width1; | |
7633 | result_low = low0 + low1; | |
7634 | break; | |
7635 | case DIV: | |
7636 | if (! op0_maybe_minusp && ! op1_maybe_minusp) | |
7637 | result_width = width0; | |
7638 | break; | |
7639 | case UDIV: | |
7640 | result_width = width0; | |
7641 | break; | |
7642 | case MOD: | |
7643 | if (! op0_maybe_minusp && ! op1_maybe_minusp) | |
7644 | result_width = MIN (width0, width1); | |
7645 | result_low = MIN (low0, low1); | |
7646 | break; | |
7647 | case UMOD: | |
7648 | result_width = MIN (width0, width1); | |
7649 | result_low = MIN (low0, low1); | |
7650 | break; | |
e9a25f70 JL |
7651 | default: |
7652 | abort (); | |
230d793d RS |
7653 | } |
7654 | ||
7655 | if (result_width < mode_width) | |
951553af | 7656 | nonzero &= ((HOST_WIDE_INT) 1 << result_width) - 1; |
230d793d RS |
7657 | |
7658 | if (result_low > 0) | |
951553af | 7659 | nonzero &= ~ (((HOST_WIDE_INT) 1 << result_low) - 1); |
230d793d RS |
7660 | } |
7661 | break; | |
7662 | ||
7663 | case ZERO_EXTRACT: | |
7664 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5f4f0e22 | 7665 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT) |
951553af | 7666 | nonzero &= ((HOST_WIDE_INT) 1 << INTVAL (XEXP (x, 1))) - 1; |
230d793d RS |
7667 | break; |
7668 | ||
7669 | case SUBREG: | |
c3c2cb37 RK |
7670 | /* If this is a SUBREG formed for a promoted variable that has |
7671 | been zero-extended, we know that at least the high-order bits | |
7672 | are zero, though others might be too. */ | |
7673 | ||
7674 | if (SUBREG_PROMOTED_VAR_P (x) && SUBREG_PROMOTED_UNSIGNED_P (x)) | |
951553af RK |
7675 | nonzero = (GET_MODE_MASK (GET_MODE (x)) |
7676 | & nonzero_bits (SUBREG_REG (x), GET_MODE (x))); | |
c3c2cb37 | 7677 | |
230d793d RS |
7678 | /* If the inner mode is a single word for both the host and target |
7679 | machines, we can compute this from which bits of the inner | |
951553af | 7680 | object might be nonzero. */ |
230d793d | 7681 | if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) <= BITS_PER_WORD |
5f4f0e22 CH |
7682 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) |
7683 | <= HOST_BITS_PER_WIDE_INT)) | |
230d793d | 7684 | { |
951553af | 7685 | nonzero &= nonzero_bits (SUBREG_REG (x), mode); |
8baf60bb | 7686 | |
b52ce03d R |
7687 | #if defined (WORD_REGISTER_OPERATIONS) && defined (LOAD_EXTEND_OP) |
7688 | /* If this is a typical RISC machine, we only have to worry | |
7689 | about the way loads are extended. */ | |
7690 | if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (x))) == SIGN_EXTEND | |
7691 | ? (nonzero | |
7692 | & (1L << (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) - 1))) | |
7693 | : LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (x))) != ZERO_EXTEND) | |
230d793d | 7694 | #endif |
b52ce03d R |
7695 | { |
7696 | /* On many CISC machines, accessing an object in a wider mode | |
7697 | causes the high-order bits to become undefined. So they are | |
7698 | not known to be zero. */ | |
7699 | if (GET_MODE_SIZE (GET_MODE (x)) | |
7700 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
7701 | nonzero |= (GET_MODE_MASK (GET_MODE (x)) | |
7702 | & ~ GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))); | |
7703 | } | |
230d793d RS |
7704 | } |
7705 | break; | |
7706 | ||
7707 | case ASHIFTRT: | |
7708 | case LSHIFTRT: | |
7709 | case ASHIFT: | |
230d793d | 7710 | case ROTATE: |
951553af | 7711 | /* The nonzero bits are in two classes: any bits within MODE |
230d793d | 7712 | that aren't in GET_MODE (x) are always significant. The rest of the |
951553af | 7713 | nonzero bits are those that are significant in the operand of |
230d793d RS |
7714 | the shift when shifted the appropriate number of bits. This |
7715 | shows that high-order bits are cleared by the right shift and | |
7716 | low-order bits by left shifts. */ | |
7717 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
7718 | && INTVAL (XEXP (x, 1)) >= 0 | |
5f4f0e22 | 7719 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT) |
230d793d RS |
7720 | { |
7721 | enum machine_mode inner_mode = GET_MODE (x); | |
7722 | int width = GET_MODE_BITSIZE (inner_mode); | |
7723 | int count = INTVAL (XEXP (x, 1)); | |
5f4f0e22 | 7724 | unsigned HOST_WIDE_INT mode_mask = GET_MODE_MASK (inner_mode); |
951553af RK |
7725 | unsigned HOST_WIDE_INT op_nonzero = nonzero_bits (XEXP (x, 0), mode); |
7726 | unsigned HOST_WIDE_INT inner = op_nonzero & mode_mask; | |
5f4f0e22 | 7727 | unsigned HOST_WIDE_INT outer = 0; |
230d793d RS |
7728 | |
7729 | if (mode_width > width) | |
951553af | 7730 | outer = (op_nonzero & nonzero & ~ mode_mask); |
230d793d RS |
7731 | |
7732 | if (code == LSHIFTRT) | |
7733 | inner >>= count; | |
7734 | else if (code == ASHIFTRT) | |
7735 | { | |
7736 | inner >>= count; | |
7737 | ||
951553af | 7738 | /* If the sign bit may have been nonzero before the shift, we |
230d793d | 7739 | need to mark all the places it could have been copied to |
951553af | 7740 | by the shift as possibly nonzero. */ |
5f4f0e22 CH |
7741 | if (inner & ((HOST_WIDE_INT) 1 << (width - 1 - count))) |
7742 | inner |= (((HOST_WIDE_INT) 1 << count) - 1) << (width - count); | |
230d793d | 7743 | } |
45620ed4 | 7744 | else if (code == ASHIFT) |
230d793d RS |
7745 | inner <<= count; |
7746 | else | |
7747 | inner = ((inner << (count % width) | |
7748 | | (inner >> (width - (count % width)))) & mode_mask); | |
7749 | ||
951553af | 7750 | nonzero &= (outer | inner); |
230d793d RS |
7751 | } |
7752 | break; | |
7753 | ||
7754 | case FFS: | |
7755 | /* This is at most the number of bits in the mode. */ | |
951553af | 7756 | nonzero = ((HOST_WIDE_INT) 1 << (floor_log2 (mode_width) + 1)) - 1; |
230d793d | 7757 | break; |
d0ab8cd3 RK |
7758 | |
7759 | case IF_THEN_ELSE: | |
951553af RK |
7760 | nonzero &= (nonzero_bits (XEXP (x, 1), mode) |
7761 | | nonzero_bits (XEXP (x, 2), mode)); | |
d0ab8cd3 | 7762 | break; |
e9a25f70 JL |
7763 | |
7764 | default: | |
7765 | break; | |
230d793d RS |
7766 | } |
7767 | ||
951553af | 7768 | return nonzero; |
230d793d | 7769 | } |
b3728b0e JW |
7770 | |
7771 | /* See the macro definition above. */ | |
7772 | #undef num_sign_bit_copies | |
230d793d | 7773 | \f |
d0ab8cd3 | 7774 | /* Return the number of bits at the high-order end of X that are known to |
5109d49f RK |
7775 | be equal to the sign bit. X will be used in mode MODE; if MODE is |
7776 | VOIDmode, X will be used in its own mode. The returned value will always | |
7777 | be between 1 and the number of bits in MODE. */ | |
d0ab8cd3 RK |
7778 | |
7779 | static int | |
7780 | num_sign_bit_copies (x, mode) | |
7781 | rtx x; | |
7782 | enum machine_mode mode; | |
7783 | { | |
7784 | enum rtx_code code = GET_CODE (x); | |
7785 | int bitwidth; | |
7786 | int num0, num1, result; | |
951553af | 7787 | unsigned HOST_WIDE_INT nonzero; |
d0ab8cd3 RK |
7788 | rtx tem; |
7789 | ||
7790 | /* If we weren't given a mode, use the mode of X. If the mode is still | |
1c75dfa4 RK |
7791 | VOIDmode, we don't know anything. Likewise if one of the modes is |
7792 | floating-point. */ | |
d0ab8cd3 RK |
7793 | |
7794 | if (mode == VOIDmode) | |
7795 | mode = GET_MODE (x); | |
7796 | ||
1c75dfa4 | 7797 | if (mode == VOIDmode || FLOAT_MODE_P (mode) || FLOAT_MODE_P (GET_MODE (x))) |
6752e8d2 | 7798 | return 1; |
d0ab8cd3 RK |
7799 | |
7800 | bitwidth = GET_MODE_BITSIZE (mode); | |
7801 | ||
0f41302f | 7802 | /* For a smaller object, just ignore the high bits. */ |
312def2e RK |
7803 | if (bitwidth < GET_MODE_BITSIZE (GET_MODE (x))) |
7804 | return MAX (1, (num_sign_bit_copies (x, GET_MODE (x)) | |
7805 | - (GET_MODE_BITSIZE (GET_MODE (x)) - bitwidth))); | |
7806 | ||
e9a25f70 JL |
7807 | if (GET_MODE (x) != VOIDmode && bitwidth > GET_MODE_BITSIZE (GET_MODE (x))) |
7808 | { | |
0c314d1a RK |
7809 | #ifndef WORD_REGISTER_OPERATIONS |
7810 | /* If this machine does not do all register operations on the entire | |
7811 | register and MODE is wider than the mode of X, we can say nothing | |
7812 | at all about the high-order bits. */ | |
e9a25f70 JL |
7813 | return 1; |
7814 | #else | |
7815 | /* Likewise on machines that do, if the mode of the object is smaller | |
7816 | than a word and loads of that size don't sign extend, we can say | |
7817 | nothing about the high order bits. */ | |
7818 | if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD | |
7819 | #ifdef LOAD_EXTEND_OP | |
7820 | && LOAD_EXTEND_OP (GET_MODE (x)) != SIGN_EXTEND | |
7821 | #endif | |
7822 | ) | |
7823 | return 1; | |
0c314d1a | 7824 | #endif |
e9a25f70 | 7825 | } |
0c314d1a | 7826 | |
d0ab8cd3 RK |
7827 | switch (code) |
7828 | { | |
7829 | case REG: | |
55310dad | 7830 | |
ff0dbdd1 RK |
7831 | #ifdef POINTERS_EXTEND_UNSIGNED |
7832 | /* If pointers extend signed and this is a pointer in Pmode, say that | |
7833 | all the bits above ptr_mode are known to be sign bit copies. */ | |
7834 | if (! POINTERS_EXTEND_UNSIGNED && GET_MODE (x) == Pmode && mode == Pmode | |
7835 | && REGNO_POINTER_FLAG (REGNO (x))) | |
7836 | return GET_MODE_BITSIZE (Pmode) - GET_MODE_BITSIZE (ptr_mode) + 1; | |
7837 | #endif | |
7838 | ||
55310dad RK |
7839 | if (reg_last_set_value[REGNO (x)] != 0 |
7840 | && reg_last_set_mode[REGNO (x)] == mode | |
b1f21e0a | 7841 | && (REG_N_SETS (REGNO (x)) == 1 |
55310dad RK |
7842 | || reg_last_set_label[REGNO (x)] == label_tick) |
7843 | && INSN_CUID (reg_last_set[REGNO (x)]) < subst_low_cuid) | |
7844 | return reg_last_set_sign_bit_copies[REGNO (x)]; | |
d0ab8cd3 RK |
7845 | |
7846 | tem = get_last_value (x); | |
7847 | if (tem != 0) | |
7848 | return num_sign_bit_copies (tem, mode); | |
55310dad RK |
7849 | |
7850 | if (nonzero_sign_valid && reg_sign_bit_copies[REGNO (x)] != 0) | |
7851 | return reg_sign_bit_copies[REGNO (x)]; | |
d0ab8cd3 RK |
7852 | break; |
7853 | ||
457816e2 | 7854 | case MEM: |
8baf60bb | 7855 | #ifdef LOAD_EXTEND_OP |
457816e2 | 7856 | /* Some RISC machines sign-extend all loads of smaller than a word. */ |
8baf60bb RK |
7857 | if (LOAD_EXTEND_OP (GET_MODE (x)) == SIGN_EXTEND) |
7858 | return MAX (1, bitwidth - GET_MODE_BITSIZE (GET_MODE (x)) + 1); | |
457816e2 | 7859 | #endif |
8baf60bb | 7860 | break; |
457816e2 | 7861 | |
d0ab8cd3 RK |
7862 | case CONST_INT: |
7863 | /* If the constant is negative, take its 1's complement and remask. | |
7864 | Then see how many zero bits we have. */ | |
951553af | 7865 | nonzero = INTVAL (x) & GET_MODE_MASK (mode); |
ac49a949 | 7866 | if (bitwidth <= HOST_BITS_PER_WIDE_INT |
951553af RK |
7867 | && (nonzero & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) |
7868 | nonzero = (~ nonzero) & GET_MODE_MASK (mode); | |
d0ab8cd3 | 7869 | |
951553af | 7870 | return (nonzero == 0 ? bitwidth : bitwidth - floor_log2 (nonzero) - 1); |
d0ab8cd3 RK |
7871 | |
7872 | case SUBREG: | |
c3c2cb37 RK |
7873 | /* If this is a SUBREG for a promoted object that is sign-extended |
7874 | and we are looking at it in a wider mode, we know that at least the | |
7875 | high-order bits are known to be sign bit copies. */ | |
7876 | ||
7877 | if (SUBREG_PROMOTED_VAR_P (x) && ! SUBREG_PROMOTED_UNSIGNED_P (x)) | |
dc3e17ad RK |
7878 | return MAX (bitwidth - GET_MODE_BITSIZE (GET_MODE (x)) + 1, |
7879 | num_sign_bit_copies (SUBREG_REG (x), mode)); | |
c3c2cb37 | 7880 | |
0f41302f | 7881 | /* For a smaller object, just ignore the high bits. */ |
d0ab8cd3 RK |
7882 | if (bitwidth <= GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))) |
7883 | { | |
7884 | num0 = num_sign_bit_copies (SUBREG_REG (x), VOIDmode); | |
7885 | return MAX (1, (num0 | |
7886 | - (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) | |
7887 | - bitwidth))); | |
7888 | } | |
457816e2 | 7889 | |
8baf60bb | 7890 | #ifdef WORD_REGISTER_OPERATIONS |
2aec5b7a | 7891 | #ifdef LOAD_EXTEND_OP |
8baf60bb RK |
7892 | /* For paradoxical SUBREGs on machines where all register operations |
7893 | affect the entire register, just look inside. Note that we are | |
7894 | passing MODE to the recursive call, so the number of sign bit copies | |
7895 | will remain relative to that mode, not the inner mode. */ | |
457816e2 | 7896 | |
2aec5b7a JW |
7897 | /* This works only if loads sign extend. Otherwise, if we get a |
7898 | reload for the inner part, it may be loaded from the stack, and | |
7899 | then we lose all sign bit copies that existed before the store | |
7900 | to the stack. */ | |
7901 | ||
7902 | if ((GET_MODE_SIZE (GET_MODE (x)) | |
7903 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
7904 | && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (x))) == SIGN_EXTEND) | |
457816e2 | 7905 | return num_sign_bit_copies (SUBREG_REG (x), mode); |
2aec5b7a | 7906 | #endif |
457816e2 | 7907 | #endif |
d0ab8cd3 RK |
7908 | break; |
7909 | ||
7910 | case SIGN_EXTRACT: | |
7911 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
7912 | return MAX (1, bitwidth - INTVAL (XEXP (x, 1))); | |
7913 | break; | |
7914 | ||
7915 | case SIGN_EXTEND: | |
7916 | return (bitwidth - GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
7917 | + num_sign_bit_copies (XEXP (x, 0), VOIDmode)); | |
7918 | ||
7919 | case TRUNCATE: | |
0f41302f | 7920 | /* For a smaller object, just ignore the high bits. */ |
d0ab8cd3 RK |
7921 | num0 = num_sign_bit_copies (XEXP (x, 0), VOIDmode); |
7922 | return MAX (1, (num0 - (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) | |
7923 | - bitwidth))); | |
7924 | ||
7925 | case NOT: | |
7926 | return num_sign_bit_copies (XEXP (x, 0), mode); | |
7927 | ||
7928 | case ROTATE: case ROTATERT: | |
7929 | /* If we are rotating left by a number of bits less than the number | |
7930 | of sign bit copies, we can just subtract that amount from the | |
7931 | number. */ | |
7932 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
7933 | && INTVAL (XEXP (x, 1)) >= 0 && INTVAL (XEXP (x, 1)) < bitwidth) | |
7934 | { | |
7935 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
7936 | return MAX (1, num0 - (code == ROTATE ? INTVAL (XEXP (x, 1)) | |
7937 | : bitwidth - INTVAL (XEXP (x, 1)))); | |
7938 | } | |
7939 | break; | |
7940 | ||
7941 | case NEG: | |
7942 | /* In general, this subtracts one sign bit copy. But if the value | |
7943 | is known to be positive, the number of sign bit copies is the | |
951553af RK |
7944 | same as that of the input. Finally, if the input has just one bit |
7945 | that might be nonzero, all the bits are copies of the sign bit. */ | |
7946 | nonzero = nonzero_bits (XEXP (x, 0), mode); | |
7947 | if (nonzero == 1) | |
d0ab8cd3 RK |
7948 | return bitwidth; |
7949 | ||
7950 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
7951 | if (num0 > 1 | |
ac49a949 | 7952 | && bitwidth <= HOST_BITS_PER_WIDE_INT |
951553af | 7953 | && (((HOST_WIDE_INT) 1 << (bitwidth - 1)) & nonzero)) |
d0ab8cd3 RK |
7954 | num0--; |
7955 | ||
7956 | return num0; | |
7957 | ||
7958 | case IOR: case AND: case XOR: | |
7959 | case SMIN: case SMAX: case UMIN: case UMAX: | |
7960 | /* Logical operations will preserve the number of sign-bit copies. | |
7961 | MIN and MAX operations always return one of the operands. */ | |
7962 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
7963 | num1 = num_sign_bit_copies (XEXP (x, 1), mode); | |
7964 | return MIN (num0, num1); | |
7965 | ||
7966 | case PLUS: case MINUS: | |
7967 | /* For addition and subtraction, we can have a 1-bit carry. However, | |
7968 | if we are subtracting 1 from a positive number, there will not | |
7969 | be such a carry. Furthermore, if the positive number is known to | |
7970 | be 0 or 1, we know the result is either -1 or 0. */ | |
7971 | ||
3e3ea975 | 7972 | if (code == PLUS && XEXP (x, 1) == constm1_rtx |
9295e6af | 7973 | && bitwidth <= HOST_BITS_PER_WIDE_INT) |
d0ab8cd3 | 7974 | { |
951553af RK |
7975 | nonzero = nonzero_bits (XEXP (x, 0), mode); |
7976 | if ((((HOST_WIDE_INT) 1 << (bitwidth - 1)) & nonzero) == 0) | |
7977 | return (nonzero == 1 || nonzero == 0 ? bitwidth | |
7978 | : bitwidth - floor_log2 (nonzero) - 1); | |
d0ab8cd3 RK |
7979 | } |
7980 | ||
7981 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
7982 | num1 = num_sign_bit_copies (XEXP (x, 1), mode); | |
7983 | return MAX (1, MIN (num0, num1) - 1); | |
7984 | ||
7985 | case MULT: | |
7986 | /* The number of bits of the product is the sum of the number of | |
7987 | bits of both terms. However, unless one of the terms if known | |
7988 | to be positive, we must allow for an additional bit since negating | |
7989 | a negative number can remove one sign bit copy. */ | |
7990 | ||
7991 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
7992 | num1 = num_sign_bit_copies (XEXP (x, 1), mode); | |
7993 | ||
7994 | result = bitwidth - (bitwidth - num0) - (bitwidth - num1); | |
7995 | if (result > 0 | |
9295e6af | 7996 | && bitwidth <= HOST_BITS_PER_WIDE_INT |
951553af | 7997 | && ((nonzero_bits (XEXP (x, 0), mode) |
d0ab8cd3 | 7998 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) |
01c82bbb RK |
7999 | && ((nonzero_bits (XEXP (x, 1), mode) |
8000 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0)) | |
d0ab8cd3 RK |
8001 | result--; |
8002 | ||
8003 | return MAX (1, result); | |
8004 | ||
8005 | case UDIV: | |
8006 | /* The result must be <= the first operand. */ | |
8007 | return num_sign_bit_copies (XEXP (x, 0), mode); | |
8008 | ||
8009 | case UMOD: | |
8010 | /* The result must be <= the scond operand. */ | |
8011 | return num_sign_bit_copies (XEXP (x, 1), mode); | |
8012 | ||
8013 | case DIV: | |
8014 | /* Similar to unsigned division, except that we have to worry about | |
8015 | the case where the divisor is negative, in which case we have | |
8016 | to add 1. */ | |
8017 | result = num_sign_bit_copies (XEXP (x, 0), mode); | |
8018 | if (result > 1 | |
ac49a949 | 8019 | && bitwidth <= HOST_BITS_PER_WIDE_INT |
951553af | 8020 | && (nonzero_bits (XEXP (x, 1), mode) |
d0ab8cd3 RK |
8021 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) |
8022 | result --; | |
8023 | ||
8024 | return result; | |
8025 | ||
8026 | case MOD: | |
8027 | result = num_sign_bit_copies (XEXP (x, 1), mode); | |
8028 | if (result > 1 | |
ac49a949 | 8029 | && bitwidth <= HOST_BITS_PER_WIDE_INT |
951553af | 8030 | && (nonzero_bits (XEXP (x, 1), mode) |
d0ab8cd3 RK |
8031 | & ((HOST_WIDE_INT) 1 << (bitwidth - 1))) != 0) |
8032 | result --; | |
8033 | ||
8034 | return result; | |
8035 | ||
8036 | case ASHIFTRT: | |
8037 | /* Shifts by a constant add to the number of bits equal to the | |
8038 | sign bit. */ | |
8039 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
8040 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
8041 | && INTVAL (XEXP (x, 1)) > 0) | |
8042 | num0 = MIN (bitwidth, num0 + INTVAL (XEXP (x, 1))); | |
8043 | ||
8044 | return num0; | |
8045 | ||
8046 | case ASHIFT: | |
d0ab8cd3 RK |
8047 | /* Left shifts destroy copies. */ |
8048 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
8049 | || INTVAL (XEXP (x, 1)) < 0 | |
8050 | || INTVAL (XEXP (x, 1)) >= bitwidth) | |
8051 | return 1; | |
8052 | ||
8053 | num0 = num_sign_bit_copies (XEXP (x, 0), mode); | |
8054 | return MAX (1, num0 - INTVAL (XEXP (x, 1))); | |
8055 | ||
8056 | case IF_THEN_ELSE: | |
8057 | num0 = num_sign_bit_copies (XEXP (x, 1), mode); | |
8058 | num1 = num_sign_bit_copies (XEXP (x, 2), mode); | |
8059 | return MIN (num0, num1); | |
8060 | ||
d0ab8cd3 RK |
8061 | case EQ: case NE: case GE: case GT: case LE: case LT: |
8062 | case GEU: case GTU: case LEU: case LTU: | |
0802d516 RK |
8063 | if (STORE_FLAG_VALUE == -1) |
8064 | return bitwidth; | |
e9a25f70 JL |
8065 | break; |
8066 | ||
8067 | default: | |
8068 | break; | |
d0ab8cd3 RK |
8069 | } |
8070 | ||
8071 | /* If we haven't been able to figure it out by one of the above rules, | |
8072 | see if some of the high-order bits are known to be zero. If so, | |
ac49a949 RS |
8073 | count those bits and return one less than that amount. If we can't |
8074 | safely compute the mask for this mode, always return BITWIDTH. */ | |
8075 | ||
8076 | if (bitwidth > HOST_BITS_PER_WIDE_INT) | |
6752e8d2 | 8077 | return 1; |
d0ab8cd3 | 8078 | |
951553af | 8079 | nonzero = nonzero_bits (x, mode); |
df6f4086 | 8080 | return (nonzero & ((HOST_WIDE_INT) 1 << (bitwidth - 1)) |
951553af | 8081 | ? 1 : bitwidth - floor_log2 (nonzero) - 1); |
d0ab8cd3 RK |
8082 | } |
8083 | \f | |
1a26b032 RK |
8084 | /* Return the number of "extended" bits there are in X, when interpreted |
8085 | as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For | |
8086 | unsigned quantities, this is the number of high-order zero bits. | |
8087 | For signed quantities, this is the number of copies of the sign bit | |
8088 | minus 1. In both case, this function returns the number of "spare" | |
8089 | bits. For example, if two quantities for which this function returns | |
8090 | at least 1 are added, the addition is known not to overflow. | |
8091 | ||
8092 | This function will always return 0 unless called during combine, which | |
8093 | implies that it must be called from a define_split. */ | |
8094 | ||
8095 | int | |
8096 | extended_count (x, mode, unsignedp) | |
8097 | rtx x; | |
8098 | enum machine_mode mode; | |
8099 | int unsignedp; | |
8100 | { | |
951553af | 8101 | if (nonzero_sign_valid == 0) |
1a26b032 RK |
8102 | return 0; |
8103 | ||
8104 | return (unsignedp | |
ac49a949 RS |
8105 | ? (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT |
8106 | && (GET_MODE_BITSIZE (mode) - 1 | |
951553af | 8107 | - floor_log2 (nonzero_bits (x, mode)))) |
1a26b032 RK |
8108 | : num_sign_bit_copies (x, mode) - 1); |
8109 | } | |
8110 | \f | |
230d793d RS |
8111 | /* This function is called from `simplify_shift_const' to merge two |
8112 | outer operations. Specifically, we have already found that we need | |
8113 | to perform operation *POP0 with constant *PCONST0 at the outermost | |
8114 | position. We would now like to also perform OP1 with constant CONST1 | |
8115 | (with *POP0 being done last). | |
8116 | ||
8117 | Return 1 if we can do the operation and update *POP0 and *PCONST0 with | |
8118 | the resulting operation. *PCOMP_P is set to 1 if we would need to | |
8119 | complement the innermost operand, otherwise it is unchanged. | |
8120 | ||
8121 | MODE is the mode in which the operation will be done. No bits outside | |
8122 | the width of this mode matter. It is assumed that the width of this mode | |
5f4f0e22 | 8123 | is smaller than or equal to HOST_BITS_PER_WIDE_INT. |
230d793d RS |
8124 | |
8125 | If *POP0 or OP1 are NIL, it means no operation is required. Only NEG, PLUS, | |
8126 | IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper | |
8127 | result is simply *PCONST0. | |
8128 | ||
8129 | If the resulting operation cannot be expressed as one operation, we | |
8130 | return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */ | |
8131 | ||
8132 | static int | |
8133 | merge_outer_ops (pop0, pconst0, op1, const1, mode, pcomp_p) | |
8134 | enum rtx_code *pop0; | |
5f4f0e22 | 8135 | HOST_WIDE_INT *pconst0; |
230d793d | 8136 | enum rtx_code op1; |
5f4f0e22 | 8137 | HOST_WIDE_INT const1; |
230d793d RS |
8138 | enum machine_mode mode; |
8139 | int *pcomp_p; | |
8140 | { | |
8141 | enum rtx_code op0 = *pop0; | |
5f4f0e22 | 8142 | HOST_WIDE_INT const0 = *pconst0; |
9fa6d012 | 8143 | int width = GET_MODE_BITSIZE (mode); |
230d793d RS |
8144 | |
8145 | const0 &= GET_MODE_MASK (mode); | |
8146 | const1 &= GET_MODE_MASK (mode); | |
8147 | ||
8148 | /* If OP0 is an AND, clear unimportant bits in CONST1. */ | |
8149 | if (op0 == AND) | |
8150 | const1 &= const0; | |
8151 | ||
8152 | /* If OP0 or OP1 is NIL, this is easy. Similarly if they are the same or | |
8153 | if OP0 is SET. */ | |
8154 | ||
8155 | if (op1 == NIL || op0 == SET) | |
8156 | return 1; | |
8157 | ||
8158 | else if (op0 == NIL) | |
8159 | op0 = op1, const0 = const1; | |
8160 | ||
8161 | else if (op0 == op1) | |
8162 | { | |
8163 | switch (op0) | |
8164 | { | |
8165 | case AND: | |
8166 | const0 &= const1; | |
8167 | break; | |
8168 | case IOR: | |
8169 | const0 |= const1; | |
8170 | break; | |
8171 | case XOR: | |
8172 | const0 ^= const1; | |
8173 | break; | |
8174 | case PLUS: | |
8175 | const0 += const1; | |
8176 | break; | |
8177 | case NEG: | |
8178 | op0 = NIL; | |
8179 | break; | |
e9a25f70 JL |
8180 | default: |
8181 | break; | |
230d793d RS |
8182 | } |
8183 | } | |
8184 | ||
8185 | /* Otherwise, if either is a PLUS or NEG, we can't do anything. */ | |
8186 | else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG) | |
8187 | return 0; | |
8188 | ||
8189 | /* If the two constants aren't the same, we can't do anything. The | |
8190 | remaining six cases can all be done. */ | |
8191 | else if (const0 != const1) | |
8192 | return 0; | |
8193 | ||
8194 | else | |
8195 | switch (op0) | |
8196 | { | |
8197 | case IOR: | |
8198 | if (op1 == AND) | |
8199 | /* (a & b) | b == b */ | |
8200 | op0 = SET; | |
8201 | else /* op1 == XOR */ | |
8202 | /* (a ^ b) | b == a | b */ | |
b729186a | 8203 | {;} |
230d793d RS |
8204 | break; |
8205 | ||
8206 | case XOR: | |
8207 | if (op1 == AND) | |
8208 | /* (a & b) ^ b == (~a) & b */ | |
8209 | op0 = AND, *pcomp_p = 1; | |
8210 | else /* op1 == IOR */ | |
8211 | /* (a | b) ^ b == a & ~b */ | |
8212 | op0 = AND, *pconst0 = ~ const0; | |
8213 | break; | |
8214 | ||
8215 | case AND: | |
8216 | if (op1 == IOR) | |
8217 | /* (a | b) & b == b */ | |
8218 | op0 = SET; | |
8219 | else /* op1 == XOR */ | |
8220 | /* (a ^ b) & b) == (~a) & b */ | |
8221 | *pcomp_p = 1; | |
8222 | break; | |
e9a25f70 JL |
8223 | default: |
8224 | break; | |
230d793d RS |
8225 | } |
8226 | ||
8227 | /* Check for NO-OP cases. */ | |
8228 | const0 &= GET_MODE_MASK (mode); | |
8229 | if (const0 == 0 | |
8230 | && (op0 == IOR || op0 == XOR || op0 == PLUS)) | |
8231 | op0 = NIL; | |
8232 | else if (const0 == 0 && op0 == AND) | |
8233 | op0 = SET; | |
8234 | else if (const0 == GET_MODE_MASK (mode) && op0 == AND) | |
8235 | op0 = NIL; | |
8236 | ||
9fa6d012 TG |
8237 | /* If this would be an entire word for the target, but is not for |
8238 | the host, then sign-extend on the host so that the number will look | |
8239 | the same way on the host that it would on the target. | |
8240 | ||
8241 | For example, when building a 64 bit alpha hosted 32 bit sparc | |
8242 | targeted compiler, then we want the 32 bit unsigned value -1 to be | |
8243 | represented as a 64 bit value -1, and not as 0x00000000ffffffff. | |
8244 | The later confuses the sparc backend. */ | |
8245 | ||
8246 | if (BITS_PER_WORD < HOST_BITS_PER_WIDE_INT && BITS_PER_WORD == width | |
8247 | && (const0 & ((HOST_WIDE_INT) 1 << (width - 1)))) | |
8248 | const0 |= ((HOST_WIDE_INT) (-1) << width); | |
8249 | ||
230d793d RS |
8250 | *pop0 = op0; |
8251 | *pconst0 = const0; | |
8252 | ||
8253 | return 1; | |
8254 | } | |
8255 | \f | |
8256 | /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift. | |
8257 | The result of the shift is RESULT_MODE. X, if non-zero, is an expression | |
8258 | that we started with. | |
8259 | ||
8260 | The shift is normally computed in the widest mode we find in VAROP, as | |
8261 | long as it isn't a different number of words than RESULT_MODE. Exceptions | |
8262 | are ASHIFTRT and ROTATE, which are always done in their original mode, */ | |
8263 | ||
8264 | static rtx | |
8265 | simplify_shift_const (x, code, result_mode, varop, count) | |
8266 | rtx x; | |
8267 | enum rtx_code code; | |
8268 | enum machine_mode result_mode; | |
8269 | rtx varop; | |
8270 | int count; | |
8271 | { | |
8272 | enum rtx_code orig_code = code; | |
8273 | int orig_count = count; | |
8274 | enum machine_mode mode = result_mode; | |
8275 | enum machine_mode shift_mode, tmode; | |
8276 | int mode_words | |
8277 | = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD; | |
8278 | /* We form (outer_op (code varop count) (outer_const)). */ | |
8279 | enum rtx_code outer_op = NIL; | |
c4e861e8 | 8280 | HOST_WIDE_INT outer_const = 0; |
230d793d RS |
8281 | rtx const_rtx; |
8282 | int complement_p = 0; | |
8283 | rtx new; | |
8284 | ||
8285 | /* If we were given an invalid count, don't do anything except exactly | |
8286 | what was requested. */ | |
8287 | ||
8288 | if (count < 0 || count > GET_MODE_BITSIZE (mode)) | |
8289 | { | |
8290 | if (x) | |
8291 | return x; | |
8292 | ||
38a448ca | 8293 | return gen_rtx_fmt_ee (code, mode, varop, GEN_INT (count)); |
230d793d RS |
8294 | } |
8295 | ||
8296 | /* Unless one of the branches of the `if' in this loop does a `continue', | |
8297 | we will `break' the loop after the `if'. */ | |
8298 | ||
8299 | while (count != 0) | |
8300 | { | |
8301 | /* If we have an operand of (clobber (const_int 0)), just return that | |
8302 | value. */ | |
8303 | if (GET_CODE (varop) == CLOBBER) | |
8304 | return varop; | |
8305 | ||
8306 | /* If we discovered we had to complement VAROP, leave. Making a NOT | |
8307 | here would cause an infinite loop. */ | |
8308 | if (complement_p) | |
8309 | break; | |
8310 | ||
abc95ed3 | 8311 | /* Convert ROTATERT to ROTATE. */ |
230d793d RS |
8312 | if (code == ROTATERT) |
8313 | code = ROTATE, count = GET_MODE_BITSIZE (result_mode) - count; | |
8314 | ||
230d793d | 8315 | /* We need to determine what mode we will do the shift in. If the |
f6789c77 RK |
8316 | shift is a right shift or a ROTATE, we must always do it in the mode |
8317 | it was originally done in. Otherwise, we can do it in MODE, the | |
0f41302f | 8318 | widest mode encountered. */ |
f6789c77 RK |
8319 | shift_mode |
8320 | = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE | |
8321 | ? result_mode : mode); | |
230d793d RS |
8322 | |
8323 | /* Handle cases where the count is greater than the size of the mode | |
8324 | minus 1. For ASHIFT, use the size minus one as the count (this can | |
8325 | occur when simplifying (lshiftrt (ashiftrt ..))). For rotates, | |
8326 | take the count modulo the size. For other shifts, the result is | |
8327 | zero. | |
8328 | ||
8329 | Since these shifts are being produced by the compiler by combining | |
8330 | multiple operations, each of which are defined, we know what the | |
8331 | result is supposed to be. */ | |
8332 | ||
8333 | if (count > GET_MODE_BITSIZE (shift_mode) - 1) | |
8334 | { | |
8335 | if (code == ASHIFTRT) | |
8336 | count = GET_MODE_BITSIZE (shift_mode) - 1; | |
8337 | else if (code == ROTATE || code == ROTATERT) | |
8338 | count %= GET_MODE_BITSIZE (shift_mode); | |
8339 | else | |
8340 | { | |
8341 | /* We can't simply return zero because there may be an | |
8342 | outer op. */ | |
8343 | varop = const0_rtx; | |
8344 | count = 0; | |
8345 | break; | |
8346 | } | |
8347 | } | |
8348 | ||
8349 | /* Negative counts are invalid and should not have been made (a | |
8350 | programmer-specified negative count should have been handled | |
0f41302f | 8351 | above). */ |
230d793d RS |
8352 | else if (count < 0) |
8353 | abort (); | |
8354 | ||
312def2e RK |
8355 | /* An arithmetic right shift of a quantity known to be -1 or 0 |
8356 | is a no-op. */ | |
8357 | if (code == ASHIFTRT | |
8358 | && (num_sign_bit_copies (varop, shift_mode) | |
8359 | == GET_MODE_BITSIZE (shift_mode))) | |
d0ab8cd3 | 8360 | { |
312def2e RK |
8361 | count = 0; |
8362 | break; | |
8363 | } | |
d0ab8cd3 | 8364 | |
312def2e RK |
8365 | /* If we are doing an arithmetic right shift and discarding all but |
8366 | the sign bit copies, this is equivalent to doing a shift by the | |
8367 | bitsize minus one. Convert it into that shift because it will often | |
8368 | allow other simplifications. */ | |
500c518b | 8369 | |
312def2e RK |
8370 | if (code == ASHIFTRT |
8371 | && (count + num_sign_bit_copies (varop, shift_mode) | |
8372 | >= GET_MODE_BITSIZE (shift_mode))) | |
8373 | count = GET_MODE_BITSIZE (shift_mode) - 1; | |
500c518b | 8374 | |
230d793d RS |
8375 | /* We simplify the tests below and elsewhere by converting |
8376 | ASHIFTRT to LSHIFTRT if we know the sign bit is clear. | |
8377 | `make_compound_operation' will convert it to a ASHIFTRT for | |
8378 | those machines (such as Vax) that don't have a LSHIFTRT. */ | |
5f4f0e22 | 8379 | if (GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT |
230d793d | 8380 | && code == ASHIFTRT |
951553af | 8381 | && ((nonzero_bits (varop, shift_mode) |
5f4f0e22 CH |
8382 | & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (shift_mode) - 1))) |
8383 | == 0)) | |
230d793d RS |
8384 | code = LSHIFTRT; |
8385 | ||
8386 | switch (GET_CODE (varop)) | |
8387 | { | |
8388 | case SIGN_EXTEND: | |
8389 | case ZERO_EXTEND: | |
8390 | case SIGN_EXTRACT: | |
8391 | case ZERO_EXTRACT: | |
8392 | new = expand_compound_operation (varop); | |
8393 | if (new != varop) | |
8394 | { | |
8395 | varop = new; | |
8396 | continue; | |
8397 | } | |
8398 | break; | |
8399 | ||
8400 | case MEM: | |
8401 | /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH | |
8402 | minus the width of a smaller mode, we can do this with a | |
8403 | SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */ | |
8404 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
8405 | && ! mode_dependent_address_p (XEXP (varop, 0)) | |
8406 | && ! MEM_VOLATILE_P (varop) | |
8407 | && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count, | |
8408 | MODE_INT, 1)) != BLKmode) | |
8409 | { | |
f76b9db2 | 8410 | if (BYTES_BIG_ENDIAN) |
38a448ca | 8411 | new = gen_rtx_MEM (tmode, XEXP (varop, 0)); |
f76b9db2 | 8412 | else |
38a448ca RH |
8413 | new = gen_rtx_MEM (tmode, |
8414 | plus_constant (XEXP (varop, 0), | |
8415 | count / BITS_PER_UNIT)); | |
e24b00c8 ILT |
8416 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (varop); |
8417 | MEM_VOLATILE_P (new) = MEM_VOLATILE_P (varop); | |
8418 | MEM_IN_STRUCT_P (new) = MEM_IN_STRUCT_P (varop); | |
230d793d RS |
8419 | varop = gen_rtx_combine (code == ASHIFTRT ? SIGN_EXTEND |
8420 | : ZERO_EXTEND, mode, new); | |
8421 | count = 0; | |
8422 | continue; | |
8423 | } | |
8424 | break; | |
8425 | ||
8426 | case USE: | |
8427 | /* Similar to the case above, except that we can only do this if | |
8428 | the resulting mode is the same as that of the underlying | |
8429 | MEM and adjust the address depending on the *bits* endianness | |
8430 | because of the way that bit-field extract insns are defined. */ | |
8431 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
8432 | && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count, | |
8433 | MODE_INT, 1)) != BLKmode | |
8434 | && tmode == GET_MODE (XEXP (varop, 0))) | |
8435 | { | |
f76b9db2 ILT |
8436 | if (BITS_BIG_ENDIAN) |
8437 | new = XEXP (varop, 0); | |
8438 | else | |
8439 | { | |
8440 | new = copy_rtx (XEXP (varop, 0)); | |
8441 | SUBST (XEXP (new, 0), | |
8442 | plus_constant (XEXP (new, 0), | |
8443 | count / BITS_PER_UNIT)); | |
8444 | } | |
230d793d RS |
8445 | |
8446 | varop = gen_rtx_combine (code == ASHIFTRT ? SIGN_EXTEND | |
8447 | : ZERO_EXTEND, mode, new); | |
8448 | count = 0; | |
8449 | continue; | |
8450 | } | |
8451 | break; | |
8452 | ||
8453 | case SUBREG: | |
8454 | /* If VAROP is a SUBREG, strip it as long as the inner operand has | |
8455 | the same number of words as what we've seen so far. Then store | |
8456 | the widest mode in MODE. */ | |
f9e67232 RS |
8457 | if (subreg_lowpart_p (varop) |
8458 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop))) | |
8459 | > GET_MODE_SIZE (GET_MODE (varop))) | |
230d793d RS |
8460 | && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop))) |
8461 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD) | |
8462 | == mode_words)) | |
8463 | { | |
8464 | varop = SUBREG_REG (varop); | |
8465 | if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode)) | |
8466 | mode = GET_MODE (varop); | |
8467 | continue; | |
8468 | } | |
8469 | break; | |
8470 | ||
8471 | case MULT: | |
8472 | /* Some machines use MULT instead of ASHIFT because MULT | |
8473 | is cheaper. But it is still better on those machines to | |
8474 | merge two shifts into one. */ | |
8475 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
8476 | && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0) | |
8477 | { | |
8478 | varop = gen_binary (ASHIFT, GET_MODE (varop), XEXP (varop, 0), | |
5f4f0e22 | 8479 | GEN_INT (exact_log2 (INTVAL (XEXP (varop, 1)))));; |
230d793d RS |
8480 | continue; |
8481 | } | |
8482 | break; | |
8483 | ||
8484 | case UDIV: | |
8485 | /* Similar, for when divides are cheaper. */ | |
8486 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
8487 | && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0) | |
8488 | { | |
8489 | varop = gen_binary (LSHIFTRT, GET_MODE (varop), XEXP (varop, 0), | |
5f4f0e22 | 8490 | GEN_INT (exact_log2 (INTVAL (XEXP (varop, 1))))); |
230d793d RS |
8491 | continue; |
8492 | } | |
8493 | break; | |
8494 | ||
8495 | case ASHIFTRT: | |
8496 | /* If we are extracting just the sign bit of an arithmetic right | |
8497 | shift, that shift is not needed. */ | |
8498 | if (code == LSHIFTRT && count == GET_MODE_BITSIZE (result_mode) - 1) | |
8499 | { | |
8500 | varop = XEXP (varop, 0); | |
8501 | continue; | |
8502 | } | |
8503 | ||
0f41302f | 8504 | /* ... fall through ... */ |
230d793d RS |
8505 | |
8506 | case LSHIFTRT: | |
8507 | case ASHIFT: | |
230d793d RS |
8508 | case ROTATE: |
8509 | /* Here we have two nested shifts. The result is usually the | |
8510 | AND of a new shift with a mask. We compute the result below. */ | |
8511 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
8512 | && INTVAL (XEXP (varop, 1)) >= 0 | |
8513 | && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop)) | |
5f4f0e22 CH |
8514 | && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT |
8515 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
230d793d RS |
8516 | { |
8517 | enum rtx_code first_code = GET_CODE (varop); | |
8518 | int first_count = INTVAL (XEXP (varop, 1)); | |
5f4f0e22 | 8519 | unsigned HOST_WIDE_INT mask; |
230d793d | 8520 | rtx mask_rtx; |
230d793d | 8521 | |
230d793d RS |
8522 | /* We have one common special case. We can't do any merging if |
8523 | the inner code is an ASHIFTRT of a smaller mode. However, if | |
8524 | we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2) | |
8525 | with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2), | |
8526 | we can convert it to | |
8527 | (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0 C2) C3) C1). | |
8528 | This simplifies certain SIGN_EXTEND operations. */ | |
8529 | if (code == ASHIFT && first_code == ASHIFTRT | |
8530 | && (GET_MODE_BITSIZE (result_mode) | |
8531 | - GET_MODE_BITSIZE (GET_MODE (varop))) == count) | |
8532 | { | |
8533 | /* C3 has the low-order C1 bits zero. */ | |
8534 | ||
5f4f0e22 CH |
8535 | mask = (GET_MODE_MASK (mode) |
8536 | & ~ (((HOST_WIDE_INT) 1 << first_count) - 1)); | |
230d793d | 8537 | |
5f4f0e22 | 8538 | varop = simplify_and_const_int (NULL_RTX, result_mode, |
230d793d | 8539 | XEXP (varop, 0), mask); |
5f4f0e22 | 8540 | varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode, |
230d793d RS |
8541 | varop, count); |
8542 | count = first_count; | |
8543 | code = ASHIFTRT; | |
8544 | continue; | |
8545 | } | |
8546 | ||
d0ab8cd3 RK |
8547 | /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more |
8548 | than C1 high-order bits equal to the sign bit, we can convert | |
8549 | this to either an ASHIFT or a ASHIFTRT depending on the | |
8550 | two counts. | |
230d793d RS |
8551 | |
8552 | We cannot do this if VAROP's mode is not SHIFT_MODE. */ | |
8553 | ||
8554 | if (code == ASHIFTRT && first_code == ASHIFT | |
8555 | && GET_MODE (varop) == shift_mode | |
d0ab8cd3 RK |
8556 | && (num_sign_bit_copies (XEXP (varop, 0), shift_mode) |
8557 | > first_count)) | |
230d793d | 8558 | { |
d0ab8cd3 RK |
8559 | count -= first_count; |
8560 | if (count < 0) | |
8561 | count = - count, code = ASHIFT; | |
8562 | varop = XEXP (varop, 0); | |
8563 | continue; | |
230d793d RS |
8564 | } |
8565 | ||
8566 | /* There are some cases we can't do. If CODE is ASHIFTRT, | |
8567 | we can only do this if FIRST_CODE is also ASHIFTRT. | |
8568 | ||
8569 | We can't do the case when CODE is ROTATE and FIRST_CODE is | |
8570 | ASHIFTRT. | |
8571 | ||
8572 | If the mode of this shift is not the mode of the outer shift, | |
bdaae9a0 | 8573 | we can't do this if either shift is a right shift or ROTATE. |
230d793d RS |
8574 | |
8575 | Finally, we can't do any of these if the mode is too wide | |
8576 | unless the codes are the same. | |
8577 | ||
8578 | Handle the case where the shift codes are the same | |
8579 | first. */ | |
8580 | ||
8581 | if (code == first_code) | |
8582 | { | |
8583 | if (GET_MODE (varop) != result_mode | |
bdaae9a0 RK |
8584 | && (code == ASHIFTRT || code == LSHIFTRT |
8585 | || code == ROTATE)) | |
230d793d RS |
8586 | break; |
8587 | ||
8588 | count += first_count; | |
8589 | varop = XEXP (varop, 0); | |
8590 | continue; | |
8591 | } | |
8592 | ||
8593 | if (code == ASHIFTRT | |
8594 | || (code == ROTATE && first_code == ASHIFTRT) | |
5f4f0e22 | 8595 | || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT |
230d793d | 8596 | || (GET_MODE (varop) != result_mode |
bdaae9a0 RK |
8597 | && (first_code == ASHIFTRT || first_code == LSHIFTRT |
8598 | || first_code == ROTATE | |
230d793d RS |
8599 | || code == ROTATE))) |
8600 | break; | |
8601 | ||
8602 | /* To compute the mask to apply after the shift, shift the | |
951553af | 8603 | nonzero bits of the inner shift the same way the |
230d793d RS |
8604 | outer shift will. */ |
8605 | ||
951553af | 8606 | mask_rtx = GEN_INT (nonzero_bits (varop, GET_MODE (varop))); |
230d793d RS |
8607 | |
8608 | mask_rtx | |
8609 | = simplify_binary_operation (code, result_mode, mask_rtx, | |
5f4f0e22 | 8610 | GEN_INT (count)); |
230d793d RS |
8611 | |
8612 | /* Give up if we can't compute an outer operation to use. */ | |
8613 | if (mask_rtx == 0 | |
8614 | || GET_CODE (mask_rtx) != CONST_INT | |
8615 | || ! merge_outer_ops (&outer_op, &outer_const, AND, | |
8616 | INTVAL (mask_rtx), | |
8617 | result_mode, &complement_p)) | |
8618 | break; | |
8619 | ||
8620 | /* If the shifts are in the same direction, we add the | |
8621 | counts. Otherwise, we subtract them. */ | |
8622 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
8623 | == (first_code == ASHIFTRT || first_code == LSHIFTRT)) | |
8624 | count += first_count; | |
8625 | else | |
8626 | count -= first_count; | |
8627 | ||
8628 | /* If COUNT is positive, the new shift is usually CODE, | |
8629 | except for the two exceptions below, in which case it is | |
8630 | FIRST_CODE. If the count is negative, FIRST_CODE should | |
8631 | always be used */ | |
8632 | if (count > 0 | |
8633 | && ((first_code == ROTATE && code == ASHIFT) | |
8634 | || (first_code == ASHIFTRT && code == LSHIFTRT))) | |
8635 | code = first_code; | |
8636 | else if (count < 0) | |
8637 | code = first_code, count = - count; | |
8638 | ||
8639 | varop = XEXP (varop, 0); | |
8640 | continue; | |
8641 | } | |
8642 | ||
8643 | /* If we have (A << B << C) for any shift, we can convert this to | |
8644 | (A << C << B). This wins if A is a constant. Only try this if | |
8645 | B is not a constant. */ | |
8646 | ||
8647 | else if (GET_CODE (varop) == code | |
8648 | && GET_CODE (XEXP (varop, 1)) != CONST_INT | |
8649 | && 0 != (new | |
8650 | = simplify_binary_operation (code, mode, | |
8651 | XEXP (varop, 0), | |
5f4f0e22 | 8652 | GEN_INT (count)))) |
230d793d RS |
8653 | { |
8654 | varop = gen_rtx_combine (code, mode, new, XEXP (varop, 1)); | |
8655 | count = 0; | |
8656 | continue; | |
8657 | } | |
8658 | break; | |
8659 | ||
8660 | case NOT: | |
8661 | /* Make this fit the case below. */ | |
8662 | varop = gen_rtx_combine (XOR, mode, XEXP (varop, 0), | |
5f4f0e22 | 8663 | GEN_INT (GET_MODE_MASK (mode))); |
230d793d RS |
8664 | continue; |
8665 | ||
8666 | case IOR: | |
8667 | case AND: | |
8668 | case XOR: | |
8669 | /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C) | |
8670 | with C the size of VAROP - 1 and the shift is logical if | |
8671 | STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1, | |
8672 | we have an (le X 0) operation. If we have an arithmetic shift | |
8673 | and STORE_FLAG_VALUE is 1 or we have a logical shift with | |
8674 | STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */ | |
8675 | ||
8676 | if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS | |
8677 | && XEXP (XEXP (varop, 0), 1) == constm1_rtx | |
8678 | && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) | |
8679 | && (code == LSHIFTRT || code == ASHIFTRT) | |
8680 | && count == GET_MODE_BITSIZE (GET_MODE (varop)) - 1 | |
8681 | && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1))) | |
8682 | { | |
8683 | count = 0; | |
8684 | varop = gen_rtx_combine (LE, GET_MODE (varop), XEXP (varop, 1), | |
8685 | const0_rtx); | |
8686 | ||
8687 | if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT) | |
8688 | varop = gen_rtx_combine (NEG, GET_MODE (varop), varop); | |
8689 | ||
8690 | continue; | |
8691 | } | |
8692 | ||
8693 | /* If we have (shift (logical)), move the logical to the outside | |
8694 | to allow it to possibly combine with another logical and the | |
8695 | shift to combine with another shift. This also canonicalizes to | |
8696 | what a ZERO_EXTRACT looks like. Also, some machines have | |
8697 | (and (shift)) insns. */ | |
8698 | ||
8699 | if (GET_CODE (XEXP (varop, 1)) == CONST_INT | |
8700 | && (new = simplify_binary_operation (code, result_mode, | |
8701 | XEXP (varop, 1), | |
5f4f0e22 | 8702 | GEN_INT (count))) != 0 |
7d171a1e | 8703 | && GET_CODE(new) == CONST_INT |
230d793d RS |
8704 | && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop), |
8705 | INTVAL (new), result_mode, &complement_p)) | |
8706 | { | |
8707 | varop = XEXP (varop, 0); | |
8708 | continue; | |
8709 | } | |
8710 | ||
8711 | /* If we can't do that, try to simplify the shift in each arm of the | |
8712 | logical expression, make a new logical expression, and apply | |
8713 | the inverse distributive law. */ | |
8714 | { | |
00d4ca1c | 8715 | rtx lhs = simplify_shift_const (NULL_RTX, code, shift_mode, |
230d793d | 8716 | XEXP (varop, 0), count); |
00d4ca1c | 8717 | rtx rhs = simplify_shift_const (NULL_RTX, code, shift_mode, |
230d793d RS |
8718 | XEXP (varop, 1), count); |
8719 | ||
21a64bf1 | 8720 | varop = gen_binary (GET_CODE (varop), shift_mode, lhs, rhs); |
230d793d RS |
8721 | varop = apply_distributive_law (varop); |
8722 | ||
8723 | count = 0; | |
8724 | } | |
8725 | break; | |
8726 | ||
8727 | case EQ: | |
45620ed4 | 8728 | /* convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE |
230d793d | 8729 | says that the sign bit can be tested, FOO has mode MODE, C is |
45620ed4 RK |
8730 | GET_MODE_BITSIZE (MODE) - 1, and FOO has only its low-order bit |
8731 | that may be nonzero. */ | |
8732 | if (code == LSHIFTRT | |
230d793d RS |
8733 | && XEXP (varop, 1) == const0_rtx |
8734 | && GET_MODE (XEXP (varop, 0)) == result_mode | |
8735 | && count == GET_MODE_BITSIZE (result_mode) - 1 | |
5f4f0e22 | 8736 | && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT |
230d793d | 8737 | && ((STORE_FLAG_VALUE |
5f4f0e22 | 8738 | & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (result_mode) - 1)))) |
951553af | 8739 | && nonzero_bits (XEXP (varop, 0), result_mode) == 1 |
5f4f0e22 CH |
8740 | && merge_outer_ops (&outer_op, &outer_const, XOR, |
8741 | (HOST_WIDE_INT) 1, result_mode, | |
8742 | &complement_p)) | |
230d793d RS |
8743 | { |
8744 | varop = XEXP (varop, 0); | |
8745 | count = 0; | |
8746 | continue; | |
8747 | } | |
8748 | break; | |
8749 | ||
8750 | case NEG: | |
d0ab8cd3 RK |
8751 | /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less |
8752 | than the number of bits in the mode is equivalent to A. */ | |
8753 | if (code == LSHIFTRT && count == GET_MODE_BITSIZE (result_mode) - 1 | |
951553af | 8754 | && nonzero_bits (XEXP (varop, 0), result_mode) == 1) |
230d793d | 8755 | { |
d0ab8cd3 | 8756 | varop = XEXP (varop, 0); |
230d793d RS |
8757 | count = 0; |
8758 | continue; | |
8759 | } | |
8760 | ||
8761 | /* NEG commutes with ASHIFT since it is multiplication. Move the | |
8762 | NEG outside to allow shifts to combine. */ | |
8763 | if (code == ASHIFT | |
5f4f0e22 CH |
8764 | && merge_outer_ops (&outer_op, &outer_const, NEG, |
8765 | (HOST_WIDE_INT) 0, result_mode, | |
8766 | &complement_p)) | |
230d793d RS |
8767 | { |
8768 | varop = XEXP (varop, 0); | |
8769 | continue; | |
8770 | } | |
8771 | break; | |
8772 | ||
8773 | case PLUS: | |
d0ab8cd3 RK |
8774 | /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C |
8775 | is one less than the number of bits in the mode is | |
8776 | equivalent to (xor A 1). */ | |
230d793d RS |
8777 | if (code == LSHIFTRT && count == GET_MODE_BITSIZE (result_mode) - 1 |
8778 | && XEXP (varop, 1) == constm1_rtx | |
951553af | 8779 | && nonzero_bits (XEXP (varop, 0), result_mode) == 1 |
5f4f0e22 CH |
8780 | && merge_outer_ops (&outer_op, &outer_const, XOR, |
8781 | (HOST_WIDE_INT) 1, result_mode, | |
8782 | &complement_p)) | |
230d793d RS |
8783 | { |
8784 | count = 0; | |
8785 | varop = XEXP (varop, 0); | |
8786 | continue; | |
8787 | } | |
8788 | ||
3f508eca | 8789 | /* If we have (xshiftrt (plus FOO BAR) C), and the only bits |
951553af | 8790 | that might be nonzero in BAR are those being shifted out and those |
3f508eca RK |
8791 | bits are known zero in FOO, we can replace the PLUS with FOO. |
8792 | Similarly in the other operand order. This code occurs when | |
8793 | we are computing the size of a variable-size array. */ | |
8794 | ||
8795 | if ((code == ASHIFTRT || code == LSHIFTRT) | |
5f4f0e22 | 8796 | && count < HOST_BITS_PER_WIDE_INT |
951553af RK |
8797 | && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0 |
8798 | && (nonzero_bits (XEXP (varop, 1), result_mode) | |
8799 | & nonzero_bits (XEXP (varop, 0), result_mode)) == 0) | |
3f508eca RK |
8800 | { |
8801 | varop = XEXP (varop, 0); | |
8802 | continue; | |
8803 | } | |
8804 | else if ((code == ASHIFTRT || code == LSHIFTRT) | |
5f4f0e22 | 8805 | && count < HOST_BITS_PER_WIDE_INT |
ac49a949 | 8806 | && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT |
951553af | 8807 | && 0 == (nonzero_bits (XEXP (varop, 0), result_mode) |
3f508eca | 8808 | >> count) |
951553af RK |
8809 | && 0 == (nonzero_bits (XEXP (varop, 0), result_mode) |
8810 | & nonzero_bits (XEXP (varop, 1), | |
3f508eca RK |
8811 | result_mode))) |
8812 | { | |
8813 | varop = XEXP (varop, 1); | |
8814 | continue; | |
8815 | } | |
8816 | ||
230d793d RS |
8817 | /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */ |
8818 | if (code == ASHIFT | |
8819 | && GET_CODE (XEXP (varop, 1)) == CONST_INT | |
8820 | && (new = simplify_binary_operation (ASHIFT, result_mode, | |
8821 | XEXP (varop, 1), | |
5f4f0e22 | 8822 | GEN_INT (count))) != 0 |
7d171a1e | 8823 | && GET_CODE(new) == CONST_INT |
230d793d RS |
8824 | && merge_outer_ops (&outer_op, &outer_const, PLUS, |
8825 | INTVAL (new), result_mode, &complement_p)) | |
8826 | { | |
8827 | varop = XEXP (varop, 0); | |
8828 | continue; | |
8829 | } | |
8830 | break; | |
8831 | ||
8832 | case MINUS: | |
8833 | /* If we have (xshiftrt (minus (ashiftrt X C)) X) C) | |
8834 | with C the size of VAROP - 1 and the shift is logical if | |
8835 | STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1, | |
8836 | we have a (gt X 0) operation. If the shift is arithmetic with | |
8837 | STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1, | |
8838 | we have a (neg (gt X 0)) operation. */ | |
8839 | ||
0802d516 RK |
8840 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
8841 | && GET_CODE (XEXP (varop, 0)) == ASHIFTRT | |
230d793d | 8842 | && count == GET_MODE_BITSIZE (GET_MODE (varop)) - 1 |
230d793d RS |
8843 | && (code == LSHIFTRT || code == ASHIFTRT) |
8844 | && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT | |
8845 | && INTVAL (XEXP (XEXP (varop, 0), 1)) == count | |
8846 | && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1))) | |
8847 | { | |
8848 | count = 0; | |
8849 | varop = gen_rtx_combine (GT, GET_MODE (varop), XEXP (varop, 1), | |
8850 | const0_rtx); | |
8851 | ||
8852 | if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT) | |
8853 | varop = gen_rtx_combine (NEG, GET_MODE (varop), varop); | |
8854 | ||
8855 | continue; | |
8856 | } | |
8857 | break; | |
6e0ef100 JC |
8858 | |
8859 | case TRUNCATE: | |
8860 | /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt)) | |
8861 | if the truncate does not affect the value. */ | |
8862 | if (code == LSHIFTRT | |
8863 | && GET_CODE (XEXP (varop, 0)) == LSHIFTRT | |
8864 | && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT | |
8865 | && (INTVAL (XEXP (XEXP (varop, 0), 1)) | |
b577a8ff JL |
8866 | >= (GET_MODE_BITSIZE (GET_MODE (XEXP (varop, 0))) |
8867 | - GET_MODE_BITSIZE (GET_MODE (varop))))) | |
6e0ef100 JC |
8868 | { |
8869 | rtx varop_inner = XEXP (varop, 0); | |
8870 | ||
8871 | varop_inner = gen_rtx_combine (LSHIFTRT, | |
8872 | GET_MODE (varop_inner), | |
8873 | XEXP (varop_inner, 0), | |
8874 | GEN_INT (count + INTVAL (XEXP (varop_inner, 1)))); | |
8875 | varop = gen_rtx_combine (TRUNCATE, GET_MODE (varop), | |
8876 | varop_inner); | |
8877 | count = 0; | |
8878 | continue; | |
8879 | } | |
8880 | break; | |
e9a25f70 JL |
8881 | |
8882 | default: | |
8883 | break; | |
230d793d RS |
8884 | } |
8885 | ||
8886 | break; | |
8887 | } | |
8888 | ||
8889 | /* We need to determine what mode to do the shift in. If the shift is | |
f6789c77 RK |
8890 | a right shift or ROTATE, we must always do it in the mode it was |
8891 | originally done in. Otherwise, we can do it in MODE, the widest mode | |
8892 | encountered. The code we care about is that of the shift that will | |
8893 | actually be done, not the shift that was originally requested. */ | |
8894 | shift_mode | |
8895 | = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE | |
8896 | ? result_mode : mode); | |
230d793d RS |
8897 | |
8898 | /* We have now finished analyzing the shift. The result should be | |
8899 | a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If | |
8900 | OUTER_OP is non-NIL, it is an operation that needs to be applied | |
8901 | to the result of the shift. OUTER_CONST is the relevant constant, | |
8902 | but we must turn off all bits turned off in the shift. | |
8903 | ||
8904 | If we were passed a value for X, see if we can use any pieces of | |
8905 | it. If not, make new rtx. */ | |
8906 | ||
8907 | if (x && GET_RTX_CLASS (GET_CODE (x)) == '2' | |
8908 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
8909 | && INTVAL (XEXP (x, 1)) == count) | |
8910 | const_rtx = XEXP (x, 1); | |
8911 | else | |
5f4f0e22 | 8912 | const_rtx = GEN_INT (count); |
230d793d RS |
8913 | |
8914 | if (x && GET_CODE (XEXP (x, 0)) == SUBREG | |
8915 | && GET_MODE (XEXP (x, 0)) == shift_mode | |
8916 | && SUBREG_REG (XEXP (x, 0)) == varop) | |
8917 | varop = XEXP (x, 0); | |
8918 | else if (GET_MODE (varop) != shift_mode) | |
8919 | varop = gen_lowpart_for_combine (shift_mode, varop); | |
8920 | ||
0f41302f | 8921 | /* If we can't make the SUBREG, try to return what we were given. */ |
230d793d RS |
8922 | if (GET_CODE (varop) == CLOBBER) |
8923 | return x ? x : varop; | |
8924 | ||
8925 | new = simplify_binary_operation (code, shift_mode, varop, const_rtx); | |
8926 | if (new != 0) | |
8927 | x = new; | |
8928 | else | |
8929 | { | |
8930 | if (x == 0 || GET_CODE (x) != code || GET_MODE (x) != shift_mode) | |
8931 | x = gen_rtx_combine (code, shift_mode, varop, const_rtx); | |
8932 | ||
8933 | SUBST (XEXP (x, 0), varop); | |
8934 | SUBST (XEXP (x, 1), const_rtx); | |
8935 | } | |
8936 | ||
224eeff2 RK |
8937 | /* If we have an outer operation and we just made a shift, it is |
8938 | possible that we could have simplified the shift were it not | |
8939 | for the outer operation. So try to do the simplification | |
8940 | recursively. */ | |
8941 | ||
8942 | if (outer_op != NIL && GET_CODE (x) == code | |
8943 | && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
8944 | x = simplify_shift_const (x, code, shift_mode, XEXP (x, 0), | |
8945 | INTVAL (XEXP (x, 1))); | |
8946 | ||
230d793d RS |
8947 | /* If we were doing a LSHIFTRT in a wider mode than it was originally, |
8948 | turn off all the bits that the shift would have turned off. */ | |
8949 | if (orig_code == LSHIFTRT && result_mode != shift_mode) | |
5f4f0e22 | 8950 | x = simplify_and_const_int (NULL_RTX, shift_mode, x, |
230d793d RS |
8951 | GET_MODE_MASK (result_mode) >> orig_count); |
8952 | ||
8953 | /* Do the remainder of the processing in RESULT_MODE. */ | |
8954 | x = gen_lowpart_for_combine (result_mode, x); | |
8955 | ||
8956 | /* If COMPLEMENT_P is set, we have to complement X before doing the outer | |
8957 | operation. */ | |
8958 | if (complement_p) | |
0c1c8ea6 | 8959 | x = gen_unary (NOT, result_mode, result_mode, x); |
230d793d RS |
8960 | |
8961 | if (outer_op != NIL) | |
8962 | { | |
5f4f0e22 | 8963 | if (GET_MODE_BITSIZE (result_mode) < HOST_BITS_PER_WIDE_INT) |
9fa6d012 TG |
8964 | { |
8965 | int width = GET_MODE_BITSIZE (result_mode); | |
8966 | ||
8967 | outer_const &= GET_MODE_MASK (result_mode); | |
8968 | ||
8969 | /* If this would be an entire word for the target, but is not for | |
8970 | the host, then sign-extend on the host so that the number will | |
8971 | look the same way on the host that it would on the target. | |
8972 | ||
8973 | For example, when building a 64 bit alpha hosted 32 bit sparc | |
8974 | targeted compiler, then we want the 32 bit unsigned value -1 to be | |
8975 | represented as a 64 bit value -1, and not as 0x00000000ffffffff. | |
8976 | The later confuses the sparc backend. */ | |
8977 | ||
8978 | if (BITS_PER_WORD < HOST_BITS_PER_WIDE_INT && BITS_PER_WORD == width | |
8979 | && (outer_const & ((HOST_WIDE_INT) 1 << (width - 1)))) | |
8980 | outer_const |= ((HOST_WIDE_INT) (-1) << width); | |
8981 | } | |
230d793d RS |
8982 | |
8983 | if (outer_op == AND) | |
5f4f0e22 | 8984 | x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const); |
230d793d RS |
8985 | else if (outer_op == SET) |
8986 | /* This means that we have determined that the result is | |
8987 | equivalent to a constant. This should be rare. */ | |
5f4f0e22 | 8988 | x = GEN_INT (outer_const); |
230d793d | 8989 | else if (GET_RTX_CLASS (outer_op) == '1') |
0c1c8ea6 | 8990 | x = gen_unary (outer_op, result_mode, result_mode, x); |
230d793d | 8991 | else |
5f4f0e22 | 8992 | x = gen_binary (outer_op, result_mode, x, GEN_INT (outer_const)); |
230d793d RS |
8993 | } |
8994 | ||
8995 | return x; | |
8996 | } | |
8997 | \f | |
8998 | /* Like recog, but we receive the address of a pointer to a new pattern. | |
8999 | We try to match the rtx that the pointer points to. | |
9000 | If that fails, we may try to modify or replace the pattern, | |
9001 | storing the replacement into the same pointer object. | |
9002 | ||
9003 | Modifications include deletion or addition of CLOBBERs. | |
9004 | ||
9005 | PNOTES is a pointer to a location where any REG_UNUSED notes added for | |
9006 | the CLOBBERs are placed. | |
9007 | ||
a29ca9db RK |
9008 | PADDED_SCRATCHES is set to the number of (clobber (scratch)) patterns |
9009 | we had to add. | |
9010 | ||
230d793d RS |
9011 | The value is the final insn code from the pattern ultimately matched, |
9012 | or -1. */ | |
9013 | ||
9014 | static int | |
a29ca9db | 9015 | recog_for_combine (pnewpat, insn, pnotes, padded_scratches) |
230d793d RS |
9016 | rtx *pnewpat; |
9017 | rtx insn; | |
9018 | rtx *pnotes; | |
a29ca9db | 9019 | int *padded_scratches; |
230d793d RS |
9020 | { |
9021 | register rtx pat = *pnewpat; | |
9022 | int insn_code_number; | |
9023 | int num_clobbers_to_add = 0; | |
9024 | int i; | |
9025 | rtx notes = 0; | |
9026 | ||
a29ca9db RK |
9027 | *padded_scratches = 0; |
9028 | ||
974f4146 RK |
9029 | /* If PAT is a PARALLEL, check to see if it contains the CLOBBER |
9030 | we use to indicate that something didn't match. If we find such a | |
9031 | thing, force rejection. */ | |
d96023cf | 9032 | if (GET_CODE (pat) == PARALLEL) |
974f4146 | 9033 | for (i = XVECLEN (pat, 0) - 1; i >= 0; i--) |
d96023cf RK |
9034 | if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER |
9035 | && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx) | |
974f4146 RK |
9036 | return -1; |
9037 | ||
230d793d RS |
9038 | /* Is the result of combination a valid instruction? */ |
9039 | insn_code_number = recog (pat, insn, &num_clobbers_to_add); | |
9040 | ||
9041 | /* If it isn't, there is the possibility that we previously had an insn | |
9042 | that clobbered some register as a side effect, but the combined | |
9043 | insn doesn't need to do that. So try once more without the clobbers | |
9044 | unless this represents an ASM insn. */ | |
9045 | ||
9046 | if (insn_code_number < 0 && ! check_asm_operands (pat) | |
9047 | && GET_CODE (pat) == PARALLEL) | |
9048 | { | |
9049 | int pos; | |
9050 | ||
9051 | for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++) | |
9052 | if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER) | |
9053 | { | |
9054 | if (i != pos) | |
9055 | SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i)); | |
9056 | pos++; | |
9057 | } | |
9058 | ||
9059 | SUBST_INT (XVECLEN (pat, 0), pos); | |
9060 | ||
9061 | if (pos == 1) | |
9062 | pat = XVECEXP (pat, 0, 0); | |
9063 | ||
9064 | insn_code_number = recog (pat, insn, &num_clobbers_to_add); | |
9065 | } | |
9066 | ||
9067 | /* If we had any clobbers to add, make a new pattern than contains | |
9068 | them. Then check to make sure that all of them are dead. */ | |
9069 | if (num_clobbers_to_add) | |
9070 | { | |
38a448ca RH |
9071 | rtx newpat = gen_rtx_PARALLEL (VOIDmode, |
9072 | gen_rtvec (GET_CODE (pat) == PARALLEL | |
9073 | ? XVECLEN (pat, 0) + num_clobbers_to_add | |
9074 | : num_clobbers_to_add + 1)); | |
230d793d RS |
9075 | |
9076 | if (GET_CODE (pat) == PARALLEL) | |
9077 | for (i = 0; i < XVECLEN (pat, 0); i++) | |
9078 | XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i); | |
9079 | else | |
9080 | XVECEXP (newpat, 0, 0) = pat; | |
9081 | ||
9082 | add_clobbers (newpat, insn_code_number); | |
9083 | ||
9084 | for (i = XVECLEN (newpat, 0) - num_clobbers_to_add; | |
9085 | i < XVECLEN (newpat, 0); i++) | |
9086 | { | |
9087 | if (GET_CODE (XEXP (XVECEXP (newpat, 0, i), 0)) == REG | |
9088 | && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn)) | |
9089 | return -1; | |
a29ca9db RK |
9090 | else if (GET_CODE (XEXP (XVECEXP (newpat, 0, i), 0)) == SCRATCH) |
9091 | (*padded_scratches)++; | |
38a448ca RH |
9092 | notes = gen_rtx_EXPR_LIST (REG_UNUSED, |
9093 | XEXP (XVECEXP (newpat, 0, i), 0), notes); | |
230d793d RS |
9094 | } |
9095 | pat = newpat; | |
9096 | } | |
9097 | ||
9098 | *pnewpat = pat; | |
9099 | *pnotes = notes; | |
9100 | ||
9101 | return insn_code_number; | |
9102 | } | |
9103 | \f | |
9104 | /* Like gen_lowpart but for use by combine. In combine it is not possible | |
9105 | to create any new pseudoregs. However, it is safe to create | |
9106 | invalid memory addresses, because combine will try to recognize | |
9107 | them and all they will do is make the combine attempt fail. | |
9108 | ||
9109 | If for some reason this cannot do its job, an rtx | |
9110 | (clobber (const_int 0)) is returned. | |
9111 | An insn containing that will not be recognized. */ | |
9112 | ||
9113 | #undef gen_lowpart | |
9114 | ||
9115 | static rtx | |
9116 | gen_lowpart_for_combine (mode, x) | |
9117 | enum machine_mode mode; | |
9118 | register rtx x; | |
9119 | { | |
9120 | rtx result; | |
9121 | ||
9122 | if (GET_MODE (x) == mode) | |
9123 | return x; | |
9124 | ||
eae957a8 RK |
9125 | /* We can only support MODE being wider than a word if X is a |
9126 | constant integer or has a mode the same size. */ | |
9127 | ||
9128 | if (GET_MODE_SIZE (mode) > UNITS_PER_WORD | |
9129 | && ! ((GET_MODE (x) == VOIDmode | |
9130 | && (GET_CODE (x) == CONST_INT | |
9131 | || GET_CODE (x) == CONST_DOUBLE)) | |
9132 | || GET_MODE_SIZE (GET_MODE (x)) == GET_MODE_SIZE (mode))) | |
38a448ca | 9133 | return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
230d793d RS |
9134 | |
9135 | /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart | |
9136 | won't know what to do. So we will strip off the SUBREG here and | |
9137 | process normally. */ | |
9138 | if (GET_CODE (x) == SUBREG && GET_CODE (SUBREG_REG (x)) == MEM) | |
9139 | { | |
9140 | x = SUBREG_REG (x); | |
9141 | if (GET_MODE (x) == mode) | |
9142 | return x; | |
9143 | } | |
9144 | ||
9145 | result = gen_lowpart_common (mode, x); | |
64bf47a2 RK |
9146 | if (result != 0 |
9147 | && GET_CODE (result) == SUBREG | |
9148 | && GET_CODE (SUBREG_REG (result)) == REG | |
9149 | && REGNO (SUBREG_REG (result)) >= FIRST_PSEUDO_REGISTER | |
9150 | && (GET_MODE_SIZE (GET_MODE (result)) | |
9151 | != GET_MODE_SIZE (GET_MODE (SUBREG_REG (result))))) | |
b1f21e0a | 9152 | REG_CHANGES_SIZE (REGNO (SUBREG_REG (result))) = 1; |
64bf47a2 | 9153 | |
230d793d RS |
9154 | if (result) |
9155 | return result; | |
9156 | ||
9157 | if (GET_CODE (x) == MEM) | |
9158 | { | |
9159 | register int offset = 0; | |
9160 | rtx new; | |
9161 | ||
9162 | /* Refuse to work on a volatile memory ref or one with a mode-dependent | |
9163 | address. */ | |
9164 | if (MEM_VOLATILE_P (x) || mode_dependent_address_p (XEXP (x, 0))) | |
38a448ca | 9165 | return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
230d793d RS |
9166 | |
9167 | /* If we want to refer to something bigger than the original memref, | |
9168 | generate a perverse subreg instead. That will force a reload | |
9169 | of the original memref X. */ | |
9170 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)) | |
38a448ca | 9171 | return gen_rtx_SUBREG (mode, x, 0); |
230d793d | 9172 | |
f76b9db2 ILT |
9173 | if (WORDS_BIG_ENDIAN) |
9174 | offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD) | |
9175 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)); | |
9176 | if (BYTES_BIG_ENDIAN) | |
9177 | { | |
9178 | /* Adjust the address so that the address-after-the-data is | |
9179 | unchanged. */ | |
9180 | offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode)) | |
9181 | - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x)))); | |
9182 | } | |
38a448ca | 9183 | new = gen_rtx_MEM (mode, plus_constant (XEXP (x, 0), offset)); |
230d793d RS |
9184 | RTX_UNCHANGING_P (new) = RTX_UNCHANGING_P (x); |
9185 | MEM_VOLATILE_P (new) = MEM_VOLATILE_P (x); | |
9186 | MEM_IN_STRUCT_P (new) = MEM_IN_STRUCT_P (x); | |
9187 | return new; | |
9188 | } | |
9189 | ||
9190 | /* If X is a comparison operator, rewrite it in a new mode. This | |
9191 | probably won't match, but may allow further simplifications. */ | |
9192 | else if (GET_RTX_CLASS (GET_CODE (x)) == '<') | |
9193 | return gen_rtx_combine (GET_CODE (x), mode, XEXP (x, 0), XEXP (x, 1)); | |
9194 | ||
9195 | /* If we couldn't simplify X any other way, just enclose it in a | |
9196 | SUBREG. Normally, this SUBREG won't match, but some patterns may | |
a7c99304 | 9197 | include an explicit SUBREG or we may simplify it further in combine. */ |
230d793d | 9198 | else |
dfbe1b2f RK |
9199 | { |
9200 | int word = 0; | |
9201 | ||
9202 | if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) | |
9203 | word = ((GET_MODE_SIZE (GET_MODE (x)) | |
9204 | - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD)) | |
9205 | / UNITS_PER_WORD); | |
38a448ca | 9206 | return gen_rtx_SUBREG (mode, x, word); |
dfbe1b2f | 9207 | } |
230d793d RS |
9208 | } |
9209 | \f | |
9210 | /* Make an rtx expression. This is a subset of gen_rtx and only supports | |
9211 | expressions of 1, 2, or 3 operands, each of which are rtx expressions. | |
9212 | ||
9213 | If the identical expression was previously in the insn (in the undobuf), | |
9214 | it will be returned. Only if it is not found will a new expression | |
9215 | be made. */ | |
9216 | ||
9217 | /*VARARGS2*/ | |
9218 | static rtx | |
4f90e4a0 | 9219 | gen_rtx_combine VPROTO((enum rtx_code code, enum machine_mode mode, ...)) |
230d793d | 9220 | { |
4f90e4a0 | 9221 | #ifndef __STDC__ |
230d793d RS |
9222 | enum rtx_code code; |
9223 | enum machine_mode mode; | |
4f90e4a0 RK |
9224 | #endif |
9225 | va_list p; | |
230d793d RS |
9226 | int n_args; |
9227 | rtx args[3]; | |
b729186a | 9228 | int j; |
230d793d RS |
9229 | char *fmt; |
9230 | rtx rt; | |
241cea85 | 9231 | struct undo *undo; |
230d793d | 9232 | |
4f90e4a0 RK |
9233 | VA_START (p, mode); |
9234 | ||
9235 | #ifndef __STDC__ | |
230d793d RS |
9236 | code = va_arg (p, enum rtx_code); |
9237 | mode = va_arg (p, enum machine_mode); | |
4f90e4a0 RK |
9238 | #endif |
9239 | ||
230d793d RS |
9240 | n_args = GET_RTX_LENGTH (code); |
9241 | fmt = GET_RTX_FORMAT (code); | |
9242 | ||
9243 | if (n_args == 0 || n_args > 3) | |
9244 | abort (); | |
9245 | ||
9246 | /* Get each arg and verify that it is supposed to be an expression. */ | |
9247 | for (j = 0; j < n_args; j++) | |
9248 | { | |
9249 | if (*fmt++ != 'e') | |
9250 | abort (); | |
9251 | ||
9252 | args[j] = va_arg (p, rtx); | |
9253 | } | |
9254 | ||
9255 | /* See if this is in undobuf. Be sure we don't use objects that came | |
9256 | from another insn; this could produce circular rtl structures. */ | |
9257 | ||
241cea85 RK |
9258 | for (undo = undobuf.undos; undo != undobuf.previous_undos; undo = undo->next) |
9259 | if (!undo->is_int | |
9260 | && GET_CODE (undo->old_contents.r) == code | |
9261 | && GET_MODE (undo->old_contents.r) == mode) | |
230d793d RS |
9262 | { |
9263 | for (j = 0; j < n_args; j++) | |
241cea85 | 9264 | if (XEXP (undo->old_contents.r, j) != args[j]) |
230d793d RS |
9265 | break; |
9266 | ||
9267 | if (j == n_args) | |
241cea85 | 9268 | return undo->old_contents.r; |
230d793d RS |
9269 | } |
9270 | ||
9271 | /* Otherwise make a new rtx. We know we have 1, 2, or 3 args. | |
9272 | Use rtx_alloc instead of gen_rtx because it's faster on RISC. */ | |
9273 | rt = rtx_alloc (code); | |
9274 | PUT_MODE (rt, mode); | |
9275 | XEXP (rt, 0) = args[0]; | |
9276 | if (n_args > 1) | |
9277 | { | |
9278 | XEXP (rt, 1) = args[1]; | |
9279 | if (n_args > 2) | |
9280 | XEXP (rt, 2) = args[2]; | |
9281 | } | |
9282 | return rt; | |
9283 | } | |
9284 | ||
9285 | /* These routines make binary and unary operations by first seeing if they | |
9286 | fold; if not, a new expression is allocated. */ | |
9287 | ||
9288 | static rtx | |
9289 | gen_binary (code, mode, op0, op1) | |
9290 | enum rtx_code code; | |
9291 | enum machine_mode mode; | |
9292 | rtx op0, op1; | |
9293 | { | |
9294 | rtx result; | |
1a26b032 RK |
9295 | rtx tem; |
9296 | ||
9297 | if (GET_RTX_CLASS (code) == 'c' | |
9298 | && (GET_CODE (op0) == CONST_INT | |
9299 | || (CONSTANT_P (op0) && GET_CODE (op1) != CONST_INT))) | |
9300 | tem = op0, op0 = op1, op1 = tem; | |
230d793d RS |
9301 | |
9302 | if (GET_RTX_CLASS (code) == '<') | |
9303 | { | |
9304 | enum machine_mode op_mode = GET_MODE (op0); | |
9210df58 RK |
9305 | |
9306 | /* Strip the COMPARE from (REL_OP (compare X Y) 0) to get | |
0f41302f | 9307 | just (REL_OP X Y). */ |
9210df58 RK |
9308 | if (GET_CODE (op0) == COMPARE && op1 == const0_rtx) |
9309 | { | |
9310 | op1 = XEXP (op0, 1); | |
9311 | op0 = XEXP (op0, 0); | |
9312 | op_mode = GET_MODE (op0); | |
9313 | } | |
9314 | ||
230d793d RS |
9315 | if (op_mode == VOIDmode) |
9316 | op_mode = GET_MODE (op1); | |
9317 | result = simplify_relational_operation (code, op_mode, op0, op1); | |
9318 | } | |
9319 | else | |
9320 | result = simplify_binary_operation (code, mode, op0, op1); | |
9321 | ||
9322 | if (result) | |
9323 | return result; | |
9324 | ||
9325 | /* Put complex operands first and constants second. */ | |
9326 | if (GET_RTX_CLASS (code) == 'c' | |
9327 | && ((CONSTANT_P (op0) && GET_CODE (op1) != CONST_INT) | |
9328 | || (GET_RTX_CLASS (GET_CODE (op0)) == 'o' | |
9329 | && GET_RTX_CLASS (GET_CODE (op1)) != 'o') | |
9330 | || (GET_CODE (op0) == SUBREG | |
9331 | && GET_RTX_CLASS (GET_CODE (SUBREG_REG (op0))) == 'o' | |
9332 | && GET_RTX_CLASS (GET_CODE (op1)) != 'o'))) | |
9333 | return gen_rtx_combine (code, mode, op1, op0); | |
9334 | ||
e5e809f4 JL |
9335 | /* If we are turning off bits already known off in OP0, we need not do |
9336 | an AND. */ | |
9337 | else if (code == AND && GET_CODE (op1) == CONST_INT | |
9338 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT | |
9339 | && (nonzero_bits (op0, mode) & ~ INTVAL (op1)) == 0) | |
9340 | return op0; | |
9341 | ||
230d793d RS |
9342 | return gen_rtx_combine (code, mode, op0, op1); |
9343 | } | |
9344 | ||
9345 | static rtx | |
0c1c8ea6 | 9346 | gen_unary (code, mode, op0_mode, op0) |
230d793d | 9347 | enum rtx_code code; |
0c1c8ea6 | 9348 | enum machine_mode mode, op0_mode; |
230d793d RS |
9349 | rtx op0; |
9350 | { | |
0c1c8ea6 | 9351 | rtx result = simplify_unary_operation (code, mode, op0, op0_mode); |
230d793d RS |
9352 | |
9353 | if (result) | |
9354 | return result; | |
9355 | ||
9356 | return gen_rtx_combine (code, mode, op0); | |
9357 | } | |
9358 | \f | |
9359 | /* Simplify a comparison between *POP0 and *POP1 where CODE is the | |
9360 | comparison code that will be tested. | |
9361 | ||
9362 | The result is a possibly different comparison code to use. *POP0 and | |
9363 | *POP1 may be updated. | |
9364 | ||
9365 | It is possible that we might detect that a comparison is either always | |
9366 | true or always false. However, we do not perform general constant | |
5089e22e | 9367 | folding in combine, so this knowledge isn't useful. Such tautologies |
230d793d RS |
9368 | should have been detected earlier. Hence we ignore all such cases. */ |
9369 | ||
9370 | static enum rtx_code | |
9371 | simplify_comparison (code, pop0, pop1) | |
9372 | enum rtx_code code; | |
9373 | rtx *pop0; | |
9374 | rtx *pop1; | |
9375 | { | |
9376 | rtx op0 = *pop0; | |
9377 | rtx op1 = *pop1; | |
9378 | rtx tem, tem1; | |
9379 | int i; | |
9380 | enum machine_mode mode, tmode; | |
9381 | ||
9382 | /* Try a few ways of applying the same transformation to both operands. */ | |
9383 | while (1) | |
9384 | { | |
3a19aabc RK |
9385 | #ifndef WORD_REGISTER_OPERATIONS |
9386 | /* The test below this one won't handle SIGN_EXTENDs on these machines, | |
9387 | so check specially. */ | |
9388 | if (code != GTU && code != GEU && code != LTU && code != LEU | |
9389 | && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT | |
9390 | && GET_CODE (XEXP (op0, 0)) == ASHIFT | |
9391 | && GET_CODE (XEXP (op1, 0)) == ASHIFT | |
9392 | && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG | |
9393 | && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG | |
9394 | && (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))) | |
ad25ba17 | 9395 | == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0)))) |
3a19aabc RK |
9396 | && GET_CODE (XEXP (op0, 1)) == CONST_INT |
9397 | && GET_CODE (XEXP (op1, 1)) == CONST_INT | |
9398 | && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT | |
9399 | && GET_CODE (XEXP (XEXP (op1, 0), 1)) == CONST_INT | |
9400 | && INTVAL (XEXP (op0, 1)) == INTVAL (XEXP (op1, 1)) | |
9401 | && INTVAL (XEXP (op0, 1)) == INTVAL (XEXP (XEXP (op0, 0), 1)) | |
9402 | && INTVAL (XEXP (op0, 1)) == INTVAL (XEXP (XEXP (op1, 0), 1)) | |
9403 | && (INTVAL (XEXP (op0, 1)) | |
9404 | == (GET_MODE_BITSIZE (GET_MODE (op0)) | |
9405 | - (GET_MODE_BITSIZE | |
9406 | (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0)))))))) | |
9407 | { | |
9408 | op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0)); | |
9409 | op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0)); | |
9410 | } | |
9411 | #endif | |
9412 | ||
230d793d RS |
9413 | /* If both operands are the same constant shift, see if we can ignore the |
9414 | shift. We can if the shift is a rotate or if the bits shifted out of | |
951553af | 9415 | this shift are known to be zero for both inputs and if the type of |
230d793d | 9416 | comparison is compatible with the shift. */ |
67232b23 RK |
9417 | if (GET_CODE (op0) == GET_CODE (op1) |
9418 | && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT | |
9419 | && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ)) | |
45620ed4 | 9420 | || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT) |
67232b23 RK |
9421 | && (code != GT && code != LT && code != GE && code != LE)) |
9422 | || (GET_CODE (op0) == ASHIFTRT | |
9423 | && (code != GTU && code != LTU | |
9424 | && code != GEU && code != GEU))) | |
9425 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
9426 | && INTVAL (XEXP (op0, 1)) >= 0 | |
9427 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT | |
9428 | && XEXP (op0, 1) == XEXP (op1, 1)) | |
230d793d RS |
9429 | { |
9430 | enum machine_mode mode = GET_MODE (op0); | |
5f4f0e22 | 9431 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); |
230d793d RS |
9432 | int shift_count = INTVAL (XEXP (op0, 1)); |
9433 | ||
9434 | if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT) | |
9435 | mask &= (mask >> shift_count) << shift_count; | |
45620ed4 | 9436 | else if (GET_CODE (op0) == ASHIFT) |
230d793d RS |
9437 | mask = (mask & (mask << shift_count)) >> shift_count; |
9438 | ||
951553af RK |
9439 | if ((nonzero_bits (XEXP (op0, 0), mode) & ~ mask) == 0 |
9440 | && (nonzero_bits (XEXP (op1, 0), mode) & ~ mask) == 0) | |
230d793d RS |
9441 | op0 = XEXP (op0, 0), op1 = XEXP (op1, 0); |
9442 | else | |
9443 | break; | |
9444 | } | |
9445 | ||
9446 | /* If both operands are AND's of a paradoxical SUBREG by constant, the | |
9447 | SUBREGs are of the same mode, and, in both cases, the AND would | |
9448 | be redundant if the comparison was done in the narrower mode, | |
9449 | do the comparison in the narrower mode (e.g., we are AND'ing with 1 | |
951553af RK |
9450 | and the operand's possibly nonzero bits are 0xffffff01; in that case |
9451 | if we only care about QImode, we don't need the AND). This case | |
9452 | occurs if the output mode of an scc insn is not SImode and | |
7e4dc511 RK |
9453 | STORE_FLAG_VALUE == 1 (e.g., the 386). |
9454 | ||
9455 | Similarly, check for a case where the AND's are ZERO_EXTEND | |
9456 | operations from some narrower mode even though a SUBREG is not | |
9457 | present. */ | |
230d793d RS |
9458 | |
9459 | else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND | |
9460 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
7e4dc511 | 9461 | && GET_CODE (XEXP (op1, 1)) == CONST_INT) |
230d793d | 9462 | { |
7e4dc511 RK |
9463 | rtx inner_op0 = XEXP (op0, 0); |
9464 | rtx inner_op1 = XEXP (op1, 0); | |
9465 | HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1)); | |
9466 | HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1)); | |
9467 | int changed = 0; | |
9468 | ||
9469 | if (GET_CODE (inner_op0) == SUBREG && GET_CODE (inner_op1) == SUBREG | |
9470 | && (GET_MODE_SIZE (GET_MODE (inner_op0)) | |
9471 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner_op0)))) | |
9472 | && (GET_MODE (SUBREG_REG (inner_op0)) | |
9473 | == GET_MODE (SUBREG_REG (inner_op1))) | |
729a2bc6 | 9474 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (inner_op0))) |
7e4dc511 | 9475 | <= HOST_BITS_PER_WIDE_INT) |
01c82bbb | 9476 | && (0 == ((~c0) & nonzero_bits (SUBREG_REG (inner_op0), |
729a2bc6 | 9477 | GET_MODE (SUBREG_REG (inner_op0))))) |
01c82bbb RK |
9478 | && (0 == ((~c1) & nonzero_bits (SUBREG_REG (inner_op1), |
9479 | GET_MODE (SUBREG_REG (inner_op1)))))) | |
7e4dc511 RK |
9480 | { |
9481 | op0 = SUBREG_REG (inner_op0); | |
9482 | op1 = SUBREG_REG (inner_op1); | |
9483 | ||
9484 | /* The resulting comparison is always unsigned since we masked | |
0f41302f | 9485 | off the original sign bit. */ |
7e4dc511 RK |
9486 | code = unsigned_condition (code); |
9487 | ||
9488 | changed = 1; | |
9489 | } | |
230d793d | 9490 | |
7e4dc511 RK |
9491 | else if (c0 == c1) |
9492 | for (tmode = GET_CLASS_NARROWEST_MODE | |
9493 | (GET_MODE_CLASS (GET_MODE (op0))); | |
9494 | tmode != GET_MODE (op0); tmode = GET_MODE_WIDER_MODE (tmode)) | |
9495 | if (c0 == GET_MODE_MASK (tmode)) | |
9496 | { | |
9497 | op0 = gen_lowpart_for_combine (tmode, inner_op0); | |
9498 | op1 = gen_lowpart_for_combine (tmode, inner_op1); | |
66415c8b | 9499 | code = unsigned_condition (code); |
7e4dc511 RK |
9500 | changed = 1; |
9501 | break; | |
9502 | } | |
9503 | ||
9504 | if (! changed) | |
9505 | break; | |
230d793d | 9506 | } |
3a19aabc | 9507 | |
ad25ba17 RK |
9508 | /* If both operands are NOT, we can strip off the outer operation |
9509 | and adjust the comparison code for swapped operands; similarly for | |
9510 | NEG, except that this must be an equality comparison. */ | |
9511 | else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT) | |
9512 | || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG | |
9513 | && (code == EQ || code == NE))) | |
9514 | op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code); | |
3a19aabc | 9515 | |
230d793d RS |
9516 | else |
9517 | break; | |
9518 | } | |
9519 | ||
9520 | /* If the first operand is a constant, swap the operands and adjust the | |
3aceff0d RK |
9521 | comparison code appropriately, but don't do this if the second operand |
9522 | is already a constant integer. */ | |
9523 | if (CONSTANT_P (op0) && GET_CODE (op1) != CONST_INT) | |
230d793d RS |
9524 | { |
9525 | tem = op0, op0 = op1, op1 = tem; | |
9526 | code = swap_condition (code); | |
9527 | } | |
9528 | ||
9529 | /* We now enter a loop during which we will try to simplify the comparison. | |
9530 | For the most part, we only are concerned with comparisons with zero, | |
9531 | but some things may really be comparisons with zero but not start | |
9532 | out looking that way. */ | |
9533 | ||
9534 | while (GET_CODE (op1) == CONST_INT) | |
9535 | { | |
9536 | enum machine_mode mode = GET_MODE (op0); | |
9537 | int mode_width = GET_MODE_BITSIZE (mode); | |
5f4f0e22 | 9538 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); |
230d793d RS |
9539 | int equality_comparison_p; |
9540 | int sign_bit_comparison_p; | |
9541 | int unsigned_comparison_p; | |
5f4f0e22 | 9542 | HOST_WIDE_INT const_op; |
230d793d RS |
9543 | |
9544 | /* We only want to handle integral modes. This catches VOIDmode, | |
9545 | CCmode, and the floating-point modes. An exception is that we | |
9546 | can handle VOIDmode if OP0 is a COMPARE or a comparison | |
9547 | operation. */ | |
9548 | ||
9549 | if (GET_MODE_CLASS (mode) != MODE_INT | |
9550 | && ! (mode == VOIDmode | |
9551 | && (GET_CODE (op0) == COMPARE | |
9552 | || GET_RTX_CLASS (GET_CODE (op0)) == '<'))) | |
9553 | break; | |
9554 | ||
9555 | /* Get the constant we are comparing against and turn off all bits | |
9556 | not on in our mode. */ | |
9557 | const_op = INTVAL (op1); | |
5f4f0e22 | 9558 | if (mode_width <= HOST_BITS_PER_WIDE_INT) |
4803a34a | 9559 | const_op &= mask; |
230d793d RS |
9560 | |
9561 | /* If we are comparing against a constant power of two and the value | |
951553af | 9562 | being compared can only have that single bit nonzero (e.g., it was |
230d793d RS |
9563 | `and'ed with that bit), we can replace this with a comparison |
9564 | with zero. */ | |
9565 | if (const_op | |
9566 | && (code == EQ || code == NE || code == GE || code == GEU | |
9567 | || code == LT || code == LTU) | |
5f4f0e22 | 9568 | && mode_width <= HOST_BITS_PER_WIDE_INT |
230d793d | 9569 | && exact_log2 (const_op) >= 0 |
951553af | 9570 | && nonzero_bits (op0, mode) == const_op) |
230d793d RS |
9571 | { |
9572 | code = (code == EQ || code == GE || code == GEU ? NE : EQ); | |
9573 | op1 = const0_rtx, const_op = 0; | |
9574 | } | |
9575 | ||
d0ab8cd3 RK |
9576 | /* Similarly, if we are comparing a value known to be either -1 or |
9577 | 0 with -1, change it to the opposite comparison against zero. */ | |
9578 | ||
9579 | if (const_op == -1 | |
9580 | && (code == EQ || code == NE || code == GT || code == LE | |
9581 | || code == GEU || code == LTU) | |
9582 | && num_sign_bit_copies (op0, mode) == mode_width) | |
9583 | { | |
9584 | code = (code == EQ || code == LE || code == GEU ? NE : EQ); | |
9585 | op1 = const0_rtx, const_op = 0; | |
9586 | } | |
9587 | ||
230d793d | 9588 | /* Do some canonicalizations based on the comparison code. We prefer |
4803a34a RK |
9589 | comparisons against zero and then prefer equality comparisons. |
9590 | If we can reduce the size of a constant, we will do that too. */ | |
230d793d RS |
9591 | |
9592 | switch (code) | |
9593 | { | |
9594 | case LT: | |
4803a34a RK |
9595 | /* < C is equivalent to <= (C - 1) */ |
9596 | if (const_op > 0) | |
230d793d | 9597 | { |
4803a34a | 9598 | const_op -= 1; |
5f4f0e22 | 9599 | op1 = GEN_INT (const_op); |
230d793d RS |
9600 | code = LE; |
9601 | /* ... fall through to LE case below. */ | |
9602 | } | |
9603 | else | |
9604 | break; | |
9605 | ||
9606 | case LE: | |
4803a34a RK |
9607 | /* <= C is equivalent to < (C + 1); we do this for C < 0 */ |
9608 | if (const_op < 0) | |
9609 | { | |
9610 | const_op += 1; | |
5f4f0e22 | 9611 | op1 = GEN_INT (const_op); |
4803a34a RK |
9612 | code = LT; |
9613 | } | |
230d793d RS |
9614 | |
9615 | /* If we are doing a <= 0 comparison on a value known to have | |
9616 | a zero sign bit, we can replace this with == 0. */ | |
9617 | else if (const_op == 0 | |
5f4f0e22 | 9618 | && mode_width <= HOST_BITS_PER_WIDE_INT |
951553af | 9619 | && (nonzero_bits (op0, mode) |
5f4f0e22 | 9620 | & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0) |
230d793d RS |
9621 | code = EQ; |
9622 | break; | |
9623 | ||
9624 | case GE: | |
0f41302f | 9625 | /* >= C is equivalent to > (C - 1). */ |
4803a34a | 9626 | if (const_op > 0) |
230d793d | 9627 | { |
4803a34a | 9628 | const_op -= 1; |
5f4f0e22 | 9629 | op1 = GEN_INT (const_op); |
230d793d RS |
9630 | code = GT; |
9631 | /* ... fall through to GT below. */ | |
9632 | } | |
9633 | else | |
9634 | break; | |
9635 | ||
9636 | case GT: | |
4803a34a RK |
9637 | /* > C is equivalent to >= (C + 1); we do this for C < 0*/ |
9638 | if (const_op < 0) | |
9639 | { | |
9640 | const_op += 1; | |
5f4f0e22 | 9641 | op1 = GEN_INT (const_op); |
4803a34a RK |
9642 | code = GE; |
9643 | } | |
230d793d RS |
9644 | |
9645 | /* If we are doing a > 0 comparison on a value known to have | |
9646 | a zero sign bit, we can replace this with != 0. */ | |
9647 | else if (const_op == 0 | |
5f4f0e22 | 9648 | && mode_width <= HOST_BITS_PER_WIDE_INT |
951553af | 9649 | && (nonzero_bits (op0, mode) |
5f4f0e22 | 9650 | & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0) |
230d793d RS |
9651 | code = NE; |
9652 | break; | |
9653 | ||
230d793d | 9654 | case LTU: |
4803a34a RK |
9655 | /* < C is equivalent to <= (C - 1). */ |
9656 | if (const_op > 0) | |
9657 | { | |
9658 | const_op -= 1; | |
5f4f0e22 | 9659 | op1 = GEN_INT (const_op); |
4803a34a | 9660 | code = LEU; |
0f41302f | 9661 | /* ... fall through ... */ |
4803a34a | 9662 | } |
d0ab8cd3 RK |
9663 | |
9664 | /* (unsigned) < 0x80000000 is equivalent to >= 0. */ | |
f77aada2 JW |
9665 | else if ((mode_width <= HOST_BITS_PER_WIDE_INT) |
9666 | && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1))) | |
d0ab8cd3 RK |
9667 | { |
9668 | const_op = 0, op1 = const0_rtx; | |
9669 | code = GE; | |
9670 | break; | |
9671 | } | |
4803a34a RK |
9672 | else |
9673 | break; | |
230d793d RS |
9674 | |
9675 | case LEU: | |
9676 | /* unsigned <= 0 is equivalent to == 0 */ | |
9677 | if (const_op == 0) | |
9678 | code = EQ; | |
d0ab8cd3 | 9679 | |
0f41302f | 9680 | /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */ |
f77aada2 JW |
9681 | else if ((mode_width <= HOST_BITS_PER_WIDE_INT) |
9682 | && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1)) | |
d0ab8cd3 RK |
9683 | { |
9684 | const_op = 0, op1 = const0_rtx; | |
9685 | code = GE; | |
9686 | } | |
230d793d RS |
9687 | break; |
9688 | ||
4803a34a RK |
9689 | case GEU: |
9690 | /* >= C is equivalent to < (C - 1). */ | |
9691 | if (const_op > 1) | |
9692 | { | |
9693 | const_op -= 1; | |
5f4f0e22 | 9694 | op1 = GEN_INT (const_op); |
4803a34a | 9695 | code = GTU; |
0f41302f | 9696 | /* ... fall through ... */ |
4803a34a | 9697 | } |
d0ab8cd3 RK |
9698 | |
9699 | /* (unsigned) >= 0x80000000 is equivalent to < 0. */ | |
f77aada2 JW |
9700 | else if ((mode_width <= HOST_BITS_PER_WIDE_INT) |
9701 | && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1))) | |
d0ab8cd3 RK |
9702 | { |
9703 | const_op = 0, op1 = const0_rtx; | |
9704 | code = LT; | |
8b2e69e1 | 9705 | break; |
d0ab8cd3 | 9706 | } |
4803a34a RK |
9707 | else |
9708 | break; | |
9709 | ||
230d793d RS |
9710 | case GTU: |
9711 | /* unsigned > 0 is equivalent to != 0 */ | |
9712 | if (const_op == 0) | |
9713 | code = NE; | |
d0ab8cd3 RK |
9714 | |
9715 | /* (unsigned) > 0x7fffffff is equivalent to < 0. */ | |
f77aada2 JW |
9716 | else if ((mode_width <= HOST_BITS_PER_WIDE_INT) |
9717 | && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1)) | |
d0ab8cd3 RK |
9718 | { |
9719 | const_op = 0, op1 = const0_rtx; | |
9720 | code = LT; | |
9721 | } | |
230d793d | 9722 | break; |
e9a25f70 JL |
9723 | |
9724 | default: | |
9725 | break; | |
230d793d RS |
9726 | } |
9727 | ||
9728 | /* Compute some predicates to simplify code below. */ | |
9729 | ||
9730 | equality_comparison_p = (code == EQ || code == NE); | |
9731 | sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0); | |
9732 | unsigned_comparison_p = (code == LTU || code == LEU || code == GTU | |
9733 | || code == LEU); | |
9734 | ||
6139ff20 RK |
9735 | /* If this is a sign bit comparison and we can do arithmetic in |
9736 | MODE, say that we will only be needing the sign bit of OP0. */ | |
9737 | if (sign_bit_comparison_p | |
9738 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
9739 | op0 = force_to_mode (op0, mode, | |
9740 | ((HOST_WIDE_INT) 1 | |
9741 | << (GET_MODE_BITSIZE (mode) - 1)), | |
e3d616e3 | 9742 | NULL_RTX, 0); |
6139ff20 | 9743 | |
230d793d RS |
9744 | /* Now try cases based on the opcode of OP0. If none of the cases |
9745 | does a "continue", we exit this loop immediately after the | |
9746 | switch. */ | |
9747 | ||
9748 | switch (GET_CODE (op0)) | |
9749 | { | |
9750 | case ZERO_EXTRACT: | |
9751 | /* If we are extracting a single bit from a variable position in | |
9752 | a constant that has only a single bit set and are comparing it | |
9753 | with zero, we can convert this into an equality comparison | |
d7cd794f | 9754 | between the position and the location of the single bit. */ |
230d793d | 9755 | |
230d793d RS |
9756 | if (GET_CODE (XEXP (op0, 0)) == CONST_INT |
9757 | && XEXP (op0, 1) == const1_rtx | |
9758 | && equality_comparison_p && const_op == 0 | |
d7cd794f | 9759 | && (i = exact_log2 (INTVAL (XEXP (op0, 0)))) >= 0) |
230d793d | 9760 | { |
f76b9db2 | 9761 | if (BITS_BIG_ENDIAN) |
d7cd794f | 9762 | #ifdef HAVE_extzv |
f76b9db2 ILT |
9763 | i = (GET_MODE_BITSIZE |
9764 | (insn_operand_mode[(int) CODE_FOR_extzv][1]) - 1 - i); | |
d7cd794f RK |
9765 | #else |
9766 | i = BITS_PER_WORD - 1 - i; | |
230d793d RS |
9767 | #endif |
9768 | ||
9769 | op0 = XEXP (op0, 2); | |
5f4f0e22 | 9770 | op1 = GEN_INT (i); |
230d793d RS |
9771 | const_op = i; |
9772 | ||
9773 | /* Result is nonzero iff shift count is equal to I. */ | |
9774 | code = reverse_condition (code); | |
9775 | continue; | |
9776 | } | |
230d793d | 9777 | |
0f41302f | 9778 | /* ... fall through ... */ |
230d793d RS |
9779 | |
9780 | case SIGN_EXTRACT: | |
9781 | tem = expand_compound_operation (op0); | |
9782 | if (tem != op0) | |
9783 | { | |
9784 | op0 = tem; | |
9785 | continue; | |
9786 | } | |
9787 | break; | |
9788 | ||
9789 | case NOT: | |
9790 | /* If testing for equality, we can take the NOT of the constant. */ | |
9791 | if (equality_comparison_p | |
9792 | && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0) | |
9793 | { | |
9794 | op0 = XEXP (op0, 0); | |
9795 | op1 = tem; | |
9796 | continue; | |
9797 | } | |
9798 | ||
9799 | /* If just looking at the sign bit, reverse the sense of the | |
9800 | comparison. */ | |
9801 | if (sign_bit_comparison_p) | |
9802 | { | |
9803 | op0 = XEXP (op0, 0); | |
9804 | code = (code == GE ? LT : GE); | |
9805 | continue; | |
9806 | } | |
9807 | break; | |
9808 | ||
9809 | case NEG: | |
9810 | /* If testing for equality, we can take the NEG of the constant. */ | |
9811 | if (equality_comparison_p | |
9812 | && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0) | |
9813 | { | |
9814 | op0 = XEXP (op0, 0); | |
9815 | op1 = tem; | |
9816 | continue; | |
9817 | } | |
9818 | ||
9819 | /* The remaining cases only apply to comparisons with zero. */ | |
9820 | if (const_op != 0) | |
9821 | break; | |
9822 | ||
9823 | /* When X is ABS or is known positive, | |
9824 | (neg X) is < 0 if and only if X != 0. */ | |
9825 | ||
9826 | if (sign_bit_comparison_p | |
9827 | && (GET_CODE (XEXP (op0, 0)) == ABS | |
5f4f0e22 | 9828 | || (mode_width <= HOST_BITS_PER_WIDE_INT |
951553af | 9829 | && (nonzero_bits (XEXP (op0, 0), mode) |
5f4f0e22 | 9830 | & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0))) |
230d793d RS |
9831 | { |
9832 | op0 = XEXP (op0, 0); | |
9833 | code = (code == LT ? NE : EQ); | |
9834 | continue; | |
9835 | } | |
9836 | ||
3bed8141 | 9837 | /* If we have NEG of something whose two high-order bits are the |
0f41302f | 9838 | same, we know that "(-a) < 0" is equivalent to "a > 0". */ |
3bed8141 | 9839 | if (num_sign_bit_copies (op0, mode) >= 2) |
230d793d RS |
9840 | { |
9841 | op0 = XEXP (op0, 0); | |
9842 | code = swap_condition (code); | |
9843 | continue; | |
9844 | } | |
9845 | break; | |
9846 | ||
9847 | case ROTATE: | |
9848 | /* If we are testing equality and our count is a constant, we | |
9849 | can perform the inverse operation on our RHS. */ | |
9850 | if (equality_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
9851 | && (tem = simplify_binary_operation (ROTATERT, mode, | |
9852 | op1, XEXP (op0, 1))) != 0) | |
9853 | { | |
9854 | op0 = XEXP (op0, 0); | |
9855 | op1 = tem; | |
9856 | continue; | |
9857 | } | |
9858 | ||
9859 | /* If we are doing a < 0 or >= 0 comparison, it means we are testing | |
9860 | a particular bit. Convert it to an AND of a constant of that | |
9861 | bit. This will be converted into a ZERO_EXTRACT. */ | |
9862 | if (const_op == 0 && sign_bit_comparison_p | |
9863 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
5f4f0e22 | 9864 | && mode_width <= HOST_BITS_PER_WIDE_INT) |
230d793d | 9865 | { |
5f4f0e22 CH |
9866 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), |
9867 | ((HOST_WIDE_INT) 1 | |
9868 | << (mode_width - 1 | |
9869 | - INTVAL (XEXP (op0, 1))))); | |
230d793d RS |
9870 | code = (code == LT ? NE : EQ); |
9871 | continue; | |
9872 | } | |
9873 | ||
0f41302f | 9874 | /* ... fall through ... */ |
230d793d RS |
9875 | |
9876 | case ABS: | |
9877 | /* ABS is ignorable inside an equality comparison with zero. */ | |
9878 | if (const_op == 0 && equality_comparison_p) | |
9879 | { | |
9880 | op0 = XEXP (op0, 0); | |
9881 | continue; | |
9882 | } | |
9883 | break; | |
9884 | ||
9885 | ||
9886 | case SIGN_EXTEND: | |
9887 | /* Can simplify (compare (zero/sign_extend FOO) CONST) | |
9888 | to (compare FOO CONST) if CONST fits in FOO's mode and we | |
9889 | are either testing inequality or have an unsigned comparison | |
9890 | with ZERO_EXTEND or a signed comparison with SIGN_EXTEND. */ | |
9891 | if (! unsigned_comparison_p | |
9892 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0))) | |
5f4f0e22 CH |
9893 | <= HOST_BITS_PER_WIDE_INT) |
9894 | && ((unsigned HOST_WIDE_INT) const_op | |
9895 | < (((HOST_WIDE_INT) 1 | |
9896 | << (GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0))) - 1))))) | |
230d793d RS |
9897 | { |
9898 | op0 = XEXP (op0, 0); | |
9899 | continue; | |
9900 | } | |
9901 | break; | |
9902 | ||
9903 | case SUBREG: | |
a687e897 | 9904 | /* Check for the case where we are comparing A - C1 with C2, |
abc95ed3 | 9905 | both constants are smaller than 1/2 the maximum positive |
a687e897 RK |
9906 | value in MODE, and the comparison is equality or unsigned. |
9907 | In that case, if A is either zero-extended to MODE or has | |
9908 | sufficient sign bits so that the high-order bit in MODE | |
9909 | is a copy of the sign in the inner mode, we can prove that it is | |
9910 | safe to do the operation in the wider mode. This simplifies | |
9911 | many range checks. */ | |
9912 | ||
9913 | if (mode_width <= HOST_BITS_PER_WIDE_INT | |
9914 | && subreg_lowpart_p (op0) | |
9915 | && GET_CODE (SUBREG_REG (op0)) == PLUS | |
9916 | && GET_CODE (XEXP (SUBREG_REG (op0), 1)) == CONST_INT | |
9917 | && INTVAL (XEXP (SUBREG_REG (op0), 1)) < 0 | |
9918 | && (- INTVAL (XEXP (SUBREG_REG (op0), 1)) | |
9919 | < GET_MODE_MASK (mode) / 2) | |
adb7a1cb | 9920 | && (unsigned HOST_WIDE_INT) const_op < GET_MODE_MASK (mode) / 2 |
951553af RK |
9921 | && (0 == (nonzero_bits (XEXP (SUBREG_REG (op0), 0), |
9922 | GET_MODE (SUBREG_REG (op0))) | |
a687e897 RK |
9923 | & ~ GET_MODE_MASK (mode)) |
9924 | || (num_sign_bit_copies (XEXP (SUBREG_REG (op0), 0), | |
9925 | GET_MODE (SUBREG_REG (op0))) | |
9926 | > (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) | |
9927 | - GET_MODE_BITSIZE (mode))))) | |
9928 | { | |
9929 | op0 = SUBREG_REG (op0); | |
9930 | continue; | |
9931 | } | |
9932 | ||
fe0cf571 RK |
9933 | /* If the inner mode is narrower and we are extracting the low part, |
9934 | we can treat the SUBREG as if it were a ZERO_EXTEND. */ | |
9935 | if (subreg_lowpart_p (op0) | |
89f1c7f2 RS |
9936 | && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) < mode_width) |
9937 | /* Fall through */ ; | |
9938 | else | |
230d793d RS |
9939 | break; |
9940 | ||
0f41302f | 9941 | /* ... fall through ... */ |
230d793d RS |
9942 | |
9943 | case ZERO_EXTEND: | |
9944 | if ((unsigned_comparison_p || equality_comparison_p) | |
9945 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0))) | |
5f4f0e22 CH |
9946 | <= HOST_BITS_PER_WIDE_INT) |
9947 | && ((unsigned HOST_WIDE_INT) const_op | |
230d793d RS |
9948 | < GET_MODE_MASK (GET_MODE (XEXP (op0, 0))))) |
9949 | { | |
9950 | op0 = XEXP (op0, 0); | |
9951 | continue; | |
9952 | } | |
9953 | break; | |
9954 | ||
9955 | case PLUS: | |
20fdd649 | 9956 | /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do |
5089e22e | 9957 | this for equality comparisons due to pathological cases involving |
230d793d | 9958 | overflows. */ |
20fdd649 RK |
9959 | if (equality_comparison_p |
9960 | && 0 != (tem = simplify_binary_operation (MINUS, mode, | |
9961 | op1, XEXP (op0, 1)))) | |
230d793d RS |
9962 | { |
9963 | op0 = XEXP (op0, 0); | |
9964 | op1 = tem; | |
9965 | continue; | |
9966 | } | |
9967 | ||
9968 | /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */ | |
9969 | if (const_op == 0 && XEXP (op0, 1) == constm1_rtx | |
9970 | && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p) | |
9971 | { | |
9972 | op0 = XEXP (XEXP (op0, 0), 0); | |
9973 | code = (code == LT ? EQ : NE); | |
9974 | continue; | |
9975 | } | |
9976 | break; | |
9977 | ||
9978 | case MINUS: | |
20fdd649 RK |
9979 | /* (eq (minus A B) C) -> (eq A (plus B C)) or |
9980 | (eq B (minus A C)), whichever simplifies. We can only do | |
9981 | this for equality comparisons due to pathological cases involving | |
9982 | overflows. */ | |
9983 | if (equality_comparison_p | |
9984 | && 0 != (tem = simplify_binary_operation (PLUS, mode, | |
9985 | XEXP (op0, 1), op1))) | |
9986 | { | |
9987 | op0 = XEXP (op0, 0); | |
9988 | op1 = tem; | |
9989 | continue; | |
9990 | } | |
9991 | ||
9992 | if (equality_comparison_p | |
9993 | && 0 != (tem = simplify_binary_operation (MINUS, mode, | |
9994 | XEXP (op0, 0), op1))) | |
9995 | { | |
9996 | op0 = XEXP (op0, 1); | |
9997 | op1 = tem; | |
9998 | continue; | |
9999 | } | |
10000 | ||
230d793d RS |
10001 | /* The sign bit of (minus (ashiftrt X C) X), where C is the number |
10002 | of bits in X minus 1, is one iff X > 0. */ | |
10003 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT | |
10004 | && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT | |
10005 | && INTVAL (XEXP (XEXP (op0, 0), 1)) == mode_width - 1 | |
10006 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1))) | |
10007 | { | |
10008 | op0 = XEXP (op0, 1); | |
10009 | code = (code == GE ? LE : GT); | |
10010 | continue; | |
10011 | } | |
10012 | break; | |
10013 | ||
10014 | case XOR: | |
10015 | /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification | |
10016 | if C is zero or B is a constant. */ | |
10017 | if (equality_comparison_p | |
10018 | && 0 != (tem = simplify_binary_operation (XOR, mode, | |
10019 | XEXP (op0, 1), op1))) | |
10020 | { | |
10021 | op0 = XEXP (op0, 0); | |
10022 | op1 = tem; | |
10023 | continue; | |
10024 | } | |
10025 | break; | |
10026 | ||
10027 | case EQ: case NE: | |
10028 | case LT: case LTU: case LE: case LEU: | |
10029 | case GT: case GTU: case GE: case GEU: | |
10030 | /* We can't do anything if OP0 is a condition code value, rather | |
10031 | than an actual data value. */ | |
10032 | if (const_op != 0 | |
10033 | #ifdef HAVE_cc0 | |
10034 | || XEXP (op0, 0) == cc0_rtx | |
10035 | #endif | |
10036 | || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC) | |
10037 | break; | |
10038 | ||
10039 | /* Get the two operands being compared. */ | |
10040 | if (GET_CODE (XEXP (op0, 0)) == COMPARE) | |
10041 | tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1); | |
10042 | else | |
10043 | tem = XEXP (op0, 0), tem1 = XEXP (op0, 1); | |
10044 | ||
10045 | /* Check for the cases where we simply want the result of the | |
10046 | earlier test or the opposite of that result. */ | |
10047 | if (code == NE | |
10048 | || (code == EQ && reversible_comparison_p (op0)) | |
5f4f0e22 | 10049 | || (GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT |
3f508eca | 10050 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT |
230d793d | 10051 | && (STORE_FLAG_VALUE |
5f4f0e22 CH |
10052 | & (((HOST_WIDE_INT) 1 |
10053 | << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1)))) | |
230d793d RS |
10054 | && (code == LT |
10055 | || (code == GE && reversible_comparison_p (op0))))) | |
10056 | { | |
10057 | code = (code == LT || code == NE | |
10058 | ? GET_CODE (op0) : reverse_condition (GET_CODE (op0))); | |
10059 | op0 = tem, op1 = tem1; | |
10060 | continue; | |
10061 | } | |
10062 | break; | |
10063 | ||
10064 | case IOR: | |
10065 | /* The sign bit of (ior (plus X (const_int -1)) X) is non-zero | |
10066 | iff X <= 0. */ | |
10067 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS | |
10068 | && XEXP (XEXP (op0, 0), 1) == constm1_rtx | |
10069 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1))) | |
10070 | { | |
10071 | op0 = XEXP (op0, 1); | |
10072 | code = (code == GE ? GT : LE); | |
10073 | continue; | |
10074 | } | |
10075 | break; | |
10076 | ||
10077 | case AND: | |
10078 | /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This | |
10079 | will be converted to a ZERO_EXTRACT later. */ | |
10080 | if (const_op == 0 && equality_comparison_p | |
45620ed4 | 10081 | && GET_CODE (XEXP (op0, 0)) == ASHIFT |
230d793d RS |
10082 | && XEXP (XEXP (op0, 0), 0) == const1_rtx) |
10083 | { | |
10084 | op0 = simplify_and_const_int | |
10085 | (op0, mode, gen_rtx_combine (LSHIFTRT, mode, | |
10086 | XEXP (op0, 1), | |
10087 | XEXP (XEXP (op0, 0), 1)), | |
5f4f0e22 | 10088 | (HOST_WIDE_INT) 1); |
230d793d RS |
10089 | continue; |
10090 | } | |
10091 | ||
10092 | /* If we are comparing (and (lshiftrt X C1) C2) for equality with | |
10093 | zero and X is a comparison and C1 and C2 describe only bits set | |
10094 | in STORE_FLAG_VALUE, we can compare with X. */ | |
10095 | if (const_op == 0 && equality_comparison_p | |
5f4f0e22 | 10096 | && mode_width <= HOST_BITS_PER_WIDE_INT |
230d793d RS |
10097 | && GET_CODE (XEXP (op0, 1)) == CONST_INT |
10098 | && GET_CODE (XEXP (op0, 0)) == LSHIFTRT | |
10099 | && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT | |
10100 | && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0 | |
5f4f0e22 | 10101 | && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT) |
230d793d RS |
10102 | { |
10103 | mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode)) | |
10104 | << INTVAL (XEXP (XEXP (op0, 0), 1))); | |
10105 | if ((~ STORE_FLAG_VALUE & mask) == 0 | |
10106 | && (GET_RTX_CLASS (GET_CODE (XEXP (XEXP (op0, 0), 0))) == '<' | |
10107 | || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0 | |
10108 | && GET_RTX_CLASS (GET_CODE (tem)) == '<'))) | |
10109 | { | |
10110 | op0 = XEXP (XEXP (op0, 0), 0); | |
10111 | continue; | |
10112 | } | |
10113 | } | |
10114 | ||
10115 | /* If we are doing an equality comparison of an AND of a bit equal | |
10116 | to the sign bit, replace this with a LT or GE comparison of | |
10117 | the underlying value. */ | |
10118 | if (equality_comparison_p | |
10119 | && const_op == 0 | |
10120 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
5f4f0e22 | 10121 | && mode_width <= HOST_BITS_PER_WIDE_INT |
230d793d | 10122 | && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode)) |
5f4f0e22 | 10123 | == (HOST_WIDE_INT) 1 << (mode_width - 1))) |
230d793d RS |
10124 | { |
10125 | op0 = XEXP (op0, 0); | |
10126 | code = (code == EQ ? GE : LT); | |
10127 | continue; | |
10128 | } | |
10129 | ||
10130 | /* If this AND operation is really a ZERO_EXTEND from a narrower | |
10131 | mode, the constant fits within that mode, and this is either an | |
10132 | equality or unsigned comparison, try to do this comparison in | |
10133 | the narrower mode. */ | |
10134 | if ((equality_comparison_p || unsigned_comparison_p) | |
10135 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
10136 | && (i = exact_log2 ((INTVAL (XEXP (op0, 1)) | |
10137 | & GET_MODE_MASK (mode)) | |
10138 | + 1)) >= 0 | |
10139 | && const_op >> i == 0 | |
10140 | && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode) | |
10141 | { | |
10142 | op0 = gen_lowpart_for_combine (tmode, XEXP (op0, 0)); | |
10143 | continue; | |
10144 | } | |
e5e809f4 JL |
10145 | |
10146 | /* If this is (and:M1 (subreg:M2 X 0) (const_int C1)) where C1 fits | |
10147 | in both M1 and M2 and the SUBREG is either paradoxical or | |
10148 | represents the low part, permute the SUBREG and the AND and | |
10149 | try again. */ | |
10150 | if (GET_CODE (XEXP (op0, 0)) == SUBREG | |
10151 | && ((mode_width | |
10152 | >= GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (XEXP (op0, 0))))) | |
9ec36da5 JL |
10153 | #ifdef WORD_REGISTER_OPERATIONS |
10154 | || subreg_lowpart_p (XEXP (op0, 0)) | |
10155 | #endif | |
10156 | ) | |
e5e809f4 JL |
10157 | && GET_CODE (XEXP (op0, 1)) == CONST_INT |
10158 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
10159 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (XEXP (op0, 0)))) | |
10160 | <= HOST_BITS_PER_WIDE_INT) | |
10161 | && (INTVAL (XEXP (op0, 1)) & ~ mask) == 0 | |
10162 | && 0 == (~ GET_MODE_MASK (GET_MODE (SUBREG_REG (XEXP (op0, 0)))) | |
9ec36da5 JL |
10163 | & INTVAL (XEXP (op0, 1))) |
10164 | && INTVAL (XEXP (op0, 1)) != mask | |
10165 | && (INTVAL (XEXP (op0, 1)) | |
10166 | != GET_MODE_MASK (GET_MODE (SUBREG_REG (XEXP (op0, 0)))))) | |
e5e809f4 JL |
10167 | |
10168 | { | |
10169 | op0 | |
10170 | = gen_lowpart_for_combine | |
10171 | (mode, | |
10172 | gen_binary (AND, GET_MODE (SUBREG_REG (XEXP (op0, 0))), | |
10173 | SUBREG_REG (XEXP (op0, 0)), XEXP (op0, 1))); | |
10174 | continue; | |
10175 | } | |
10176 | ||
230d793d RS |
10177 | break; |
10178 | ||
10179 | case ASHIFT: | |
45620ed4 | 10180 | /* If we have (compare (ashift FOO N) (const_int C)) and |
230d793d | 10181 | the high order N bits of FOO (N+1 if an inequality comparison) |
951553af | 10182 | are known to be zero, we can do this by comparing FOO with C |
230d793d RS |
10183 | shifted right N bits so long as the low-order N bits of C are |
10184 | zero. */ | |
10185 | if (GET_CODE (XEXP (op0, 1)) == CONST_INT | |
10186 | && INTVAL (XEXP (op0, 1)) >= 0 | |
10187 | && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p) | |
5f4f0e22 CH |
10188 | < HOST_BITS_PER_WIDE_INT) |
10189 | && ((const_op | |
34785d05 | 10190 | & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0) |
5f4f0e22 | 10191 | && mode_width <= HOST_BITS_PER_WIDE_INT |
951553af | 10192 | && (nonzero_bits (XEXP (op0, 0), mode) |
230d793d RS |
10193 | & ~ (mask >> (INTVAL (XEXP (op0, 1)) |
10194 | + ! equality_comparison_p))) == 0) | |
10195 | { | |
10196 | const_op >>= INTVAL (XEXP (op0, 1)); | |
5f4f0e22 | 10197 | op1 = GEN_INT (const_op); |
230d793d RS |
10198 | op0 = XEXP (op0, 0); |
10199 | continue; | |
10200 | } | |
10201 | ||
dfbe1b2f | 10202 | /* If we are doing a sign bit comparison, it means we are testing |
230d793d | 10203 | a particular bit. Convert it to the appropriate AND. */ |
dfbe1b2f | 10204 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT |
5f4f0e22 | 10205 | && mode_width <= HOST_BITS_PER_WIDE_INT) |
230d793d | 10206 | { |
5f4f0e22 CH |
10207 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), |
10208 | ((HOST_WIDE_INT) 1 | |
10209 | << (mode_width - 1 | |
10210 | - INTVAL (XEXP (op0, 1))))); | |
230d793d RS |
10211 | code = (code == LT ? NE : EQ); |
10212 | continue; | |
10213 | } | |
dfbe1b2f RK |
10214 | |
10215 | /* If this an equality comparison with zero and we are shifting | |
10216 | the low bit to the sign bit, we can convert this to an AND of the | |
10217 | low-order bit. */ | |
10218 | if (const_op == 0 && equality_comparison_p | |
10219 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
10220 | && INTVAL (XEXP (op0, 1)) == mode_width - 1) | |
10221 | { | |
5f4f0e22 CH |
10222 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), |
10223 | (HOST_WIDE_INT) 1); | |
dfbe1b2f RK |
10224 | continue; |
10225 | } | |
230d793d RS |
10226 | break; |
10227 | ||
10228 | case ASHIFTRT: | |
d0ab8cd3 RK |
10229 | /* If this is an equality comparison with zero, we can do this |
10230 | as a logical shift, which might be much simpler. */ | |
10231 | if (equality_comparison_p && const_op == 0 | |
10232 | && GET_CODE (XEXP (op0, 1)) == CONST_INT) | |
10233 | { | |
10234 | op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode, | |
10235 | XEXP (op0, 0), | |
10236 | INTVAL (XEXP (op0, 1))); | |
10237 | continue; | |
10238 | } | |
10239 | ||
230d793d RS |
10240 | /* If OP0 is a sign extension and CODE is not an unsigned comparison, |
10241 | do the comparison in a narrower mode. */ | |
10242 | if (! unsigned_comparison_p | |
10243 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
10244 | && GET_CODE (XEXP (op0, 0)) == ASHIFT | |
10245 | && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1) | |
10246 | && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)), | |
22331794 | 10247 | MODE_INT, 1)) != BLKmode |
5f4f0e22 CH |
10248 | && ((unsigned HOST_WIDE_INT) const_op <= GET_MODE_MASK (tmode) |
10249 | || ((unsigned HOST_WIDE_INT) - const_op | |
10250 | <= GET_MODE_MASK (tmode)))) | |
230d793d RS |
10251 | { |
10252 | op0 = gen_lowpart_for_combine (tmode, XEXP (XEXP (op0, 0), 0)); | |
10253 | continue; | |
10254 | } | |
10255 | ||
0f41302f | 10256 | /* ... fall through ... */ |
230d793d RS |
10257 | case LSHIFTRT: |
10258 | /* If we have (compare (xshiftrt FOO N) (const_int C)) and | |
951553af | 10259 | the low order N bits of FOO are known to be zero, we can do this |
230d793d RS |
10260 | by comparing FOO with C shifted left N bits so long as no |
10261 | overflow occurs. */ | |
10262 | if (GET_CODE (XEXP (op0, 1)) == CONST_INT | |
10263 | && INTVAL (XEXP (op0, 1)) >= 0 | |
5f4f0e22 CH |
10264 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT |
10265 | && mode_width <= HOST_BITS_PER_WIDE_INT | |
951553af | 10266 | && (nonzero_bits (XEXP (op0, 0), mode) |
5f4f0e22 | 10267 | & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0 |
230d793d RS |
10268 | && (const_op == 0 |
10269 | || (floor_log2 (const_op) + INTVAL (XEXP (op0, 1)) | |
10270 | < mode_width))) | |
10271 | { | |
10272 | const_op <<= INTVAL (XEXP (op0, 1)); | |
5f4f0e22 | 10273 | op1 = GEN_INT (const_op); |
230d793d RS |
10274 | op0 = XEXP (op0, 0); |
10275 | continue; | |
10276 | } | |
10277 | ||
10278 | /* If we are using this shift to extract just the sign bit, we | |
10279 | can replace this with an LT or GE comparison. */ | |
10280 | if (const_op == 0 | |
10281 | && (equality_comparison_p || sign_bit_comparison_p) | |
10282 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
10283 | && INTVAL (XEXP (op0, 1)) == mode_width - 1) | |
10284 | { | |
10285 | op0 = XEXP (op0, 0); | |
10286 | code = (code == NE || code == GT ? LT : GE); | |
10287 | continue; | |
10288 | } | |
10289 | break; | |
e9a25f70 JL |
10290 | |
10291 | default: | |
10292 | break; | |
230d793d RS |
10293 | } |
10294 | ||
10295 | break; | |
10296 | } | |
10297 | ||
10298 | /* Now make any compound operations involved in this comparison. Then, | |
76d31c63 | 10299 | check for an outmost SUBREG on OP0 that is not doing anything or is |
230d793d RS |
10300 | paradoxical. The latter case can only occur when it is known that the |
10301 | "extra" bits will be zero. Therefore, it is safe to remove the SUBREG. | |
10302 | We can never remove a SUBREG for a non-equality comparison because the | |
10303 | sign bit is in a different place in the underlying object. */ | |
10304 | ||
10305 | op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET); | |
10306 | op1 = make_compound_operation (op1, SET); | |
10307 | ||
10308 | if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0) | |
10309 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
10310 | && (code == NE || code == EQ) | |
10311 | && ((GET_MODE_SIZE (GET_MODE (op0)) | |
10312 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))) | |
10313 | { | |
10314 | op0 = SUBREG_REG (op0); | |
10315 | op1 = gen_lowpart_for_combine (GET_MODE (op0), op1); | |
10316 | } | |
10317 | ||
10318 | else if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0) | |
10319 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
10320 | && (code == NE || code == EQ) | |
ac49a949 RS |
10321 | && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) |
10322 | <= HOST_BITS_PER_WIDE_INT) | |
951553af | 10323 | && (nonzero_bits (SUBREG_REG (op0), GET_MODE (SUBREG_REG (op0))) |
230d793d RS |
10324 | & ~ GET_MODE_MASK (GET_MODE (op0))) == 0 |
10325 | && (tem = gen_lowpart_for_combine (GET_MODE (SUBREG_REG (op0)), | |
10326 | op1), | |
951553af | 10327 | (nonzero_bits (tem, GET_MODE (SUBREG_REG (op0))) |
230d793d RS |
10328 | & ~ GET_MODE_MASK (GET_MODE (op0))) == 0)) |
10329 | op0 = SUBREG_REG (op0), op1 = tem; | |
10330 | ||
10331 | /* We now do the opposite procedure: Some machines don't have compare | |
10332 | insns in all modes. If OP0's mode is an integer mode smaller than a | |
10333 | word and we can't do a compare in that mode, see if there is a larger | |
a687e897 RK |
10334 | mode for which we can do the compare. There are a number of cases in |
10335 | which we can use the wider mode. */ | |
230d793d RS |
10336 | |
10337 | mode = GET_MODE (op0); | |
10338 | if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT | |
10339 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD | |
10340 | && cmp_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) | |
10341 | for (tmode = GET_MODE_WIDER_MODE (mode); | |
5f4f0e22 CH |
10342 | (tmode != VOIDmode |
10343 | && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT); | |
230d793d | 10344 | tmode = GET_MODE_WIDER_MODE (tmode)) |
a687e897 | 10345 | if (cmp_optab->handlers[(int) tmode].insn_code != CODE_FOR_nothing) |
230d793d | 10346 | { |
951553af | 10347 | /* If the only nonzero bits in OP0 and OP1 are those in the |
a687e897 RK |
10348 | narrower mode and this is an equality or unsigned comparison, |
10349 | we can use the wider mode. Similarly for sign-extended | |
7e4dc511 | 10350 | values, in which case it is true for all comparisons. */ |
a687e897 RK |
10351 | if (((code == EQ || code == NE |
10352 | || code == GEU || code == GTU || code == LEU || code == LTU) | |
951553af RK |
10353 | && (nonzero_bits (op0, tmode) & ~ GET_MODE_MASK (mode)) == 0 |
10354 | && (nonzero_bits (op1, tmode) & ~ GET_MODE_MASK (mode)) == 0) | |
7e4dc511 RK |
10355 | || ((num_sign_bit_copies (op0, tmode) |
10356 | > GET_MODE_BITSIZE (tmode) - GET_MODE_BITSIZE (mode)) | |
a687e897 | 10357 | && (num_sign_bit_copies (op1, tmode) |
58744483 | 10358 | > GET_MODE_BITSIZE (tmode) - GET_MODE_BITSIZE (mode)))) |
a687e897 RK |
10359 | { |
10360 | op0 = gen_lowpart_for_combine (tmode, op0); | |
10361 | op1 = gen_lowpart_for_combine (tmode, op1); | |
10362 | break; | |
10363 | } | |
230d793d | 10364 | |
a687e897 RK |
10365 | /* If this is a test for negative, we can make an explicit |
10366 | test of the sign bit. */ | |
10367 | ||
10368 | if (op1 == const0_rtx && (code == LT || code == GE) | |
10369 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
230d793d | 10370 | { |
a687e897 RK |
10371 | op0 = gen_binary (AND, tmode, |
10372 | gen_lowpart_for_combine (tmode, op0), | |
5f4f0e22 CH |
10373 | GEN_INT ((HOST_WIDE_INT) 1 |
10374 | << (GET_MODE_BITSIZE (mode) - 1))); | |
230d793d | 10375 | code = (code == LT) ? NE : EQ; |
a687e897 | 10376 | break; |
230d793d | 10377 | } |
230d793d RS |
10378 | } |
10379 | ||
b7a775b2 RK |
10380 | #ifdef CANONICALIZE_COMPARISON |
10381 | /* If this machine only supports a subset of valid comparisons, see if we | |
10382 | can convert an unsupported one into a supported one. */ | |
10383 | CANONICALIZE_COMPARISON (code, op0, op1); | |
10384 | #endif | |
10385 | ||
230d793d RS |
10386 | *pop0 = op0; |
10387 | *pop1 = op1; | |
10388 | ||
10389 | return code; | |
10390 | } | |
10391 | \f | |
10392 | /* Return 1 if we know that X, a comparison operation, is not operating | |
10393 | on a floating-point value or is EQ or NE, meaning that we can safely | |
10394 | reverse it. */ | |
10395 | ||
10396 | static int | |
10397 | reversible_comparison_p (x) | |
10398 | rtx x; | |
10399 | { | |
10400 | if (TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT | |
7e2a0d8e | 10401 | || flag_fast_math |
230d793d RS |
10402 | || GET_CODE (x) == NE || GET_CODE (x) == EQ) |
10403 | return 1; | |
10404 | ||
10405 | switch (GET_MODE_CLASS (GET_MODE (XEXP (x, 0)))) | |
10406 | { | |
10407 | case MODE_INT: | |
3ad2180a RK |
10408 | case MODE_PARTIAL_INT: |
10409 | case MODE_COMPLEX_INT: | |
230d793d RS |
10410 | return 1; |
10411 | ||
10412 | case MODE_CC: | |
9210df58 RK |
10413 | /* If the mode of the condition codes tells us that this is safe, |
10414 | we need look no further. */ | |
10415 | if (REVERSIBLE_CC_MODE (GET_MODE (XEXP (x, 0)))) | |
10416 | return 1; | |
10417 | ||
10418 | /* Otherwise try and find where the condition codes were last set and | |
10419 | use that. */ | |
230d793d RS |
10420 | x = get_last_value (XEXP (x, 0)); |
10421 | return (x && GET_CODE (x) == COMPARE | |
3ad2180a | 10422 | && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 0)))); |
e9a25f70 JL |
10423 | |
10424 | default: | |
10425 | return 0; | |
230d793d | 10426 | } |
230d793d RS |
10427 | } |
10428 | \f | |
10429 | /* Utility function for following routine. Called when X is part of a value | |
10430 | being stored into reg_last_set_value. Sets reg_last_set_table_tick | |
10431 | for each register mentioned. Similar to mention_regs in cse.c */ | |
10432 | ||
10433 | static void | |
10434 | update_table_tick (x) | |
10435 | rtx x; | |
10436 | { | |
10437 | register enum rtx_code code = GET_CODE (x); | |
10438 | register char *fmt = GET_RTX_FORMAT (code); | |
10439 | register int i; | |
10440 | ||
10441 | if (code == REG) | |
10442 | { | |
10443 | int regno = REGNO (x); | |
10444 | int endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
10445 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
10446 | ||
10447 | for (i = regno; i < endregno; i++) | |
10448 | reg_last_set_table_tick[i] = label_tick; | |
10449 | ||
10450 | return; | |
10451 | } | |
10452 | ||
10453 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
10454 | /* Note that we can't have an "E" in values stored; see | |
10455 | get_last_value_validate. */ | |
10456 | if (fmt[i] == 'e') | |
10457 | update_table_tick (XEXP (x, i)); | |
10458 | } | |
10459 | ||
10460 | /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we | |
10461 | are saying that the register is clobbered and we no longer know its | |
7988fd36 RK |
10462 | value. If INSN is zero, don't update reg_last_set; this is only permitted |
10463 | with VALUE also zero and is used to invalidate the register. */ | |
230d793d RS |
10464 | |
10465 | static void | |
10466 | record_value_for_reg (reg, insn, value) | |
10467 | rtx reg; | |
10468 | rtx insn; | |
10469 | rtx value; | |
10470 | { | |
10471 | int regno = REGNO (reg); | |
10472 | int endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
10473 | ? HARD_REGNO_NREGS (regno, GET_MODE (reg)) : 1); | |
10474 | int i; | |
10475 | ||
10476 | /* If VALUE contains REG and we have a previous value for REG, substitute | |
10477 | the previous value. */ | |
10478 | if (value && insn && reg_overlap_mentioned_p (reg, value)) | |
10479 | { | |
10480 | rtx tem; | |
10481 | ||
10482 | /* Set things up so get_last_value is allowed to see anything set up to | |
10483 | our insn. */ | |
10484 | subst_low_cuid = INSN_CUID (insn); | |
10485 | tem = get_last_value (reg); | |
10486 | ||
10487 | if (tem) | |
10488 | value = replace_rtx (copy_rtx (value), reg, tem); | |
10489 | } | |
10490 | ||
10491 | /* For each register modified, show we don't know its value, that | |
ef026f91 RS |
10492 | we don't know about its bitwise content, that its value has been |
10493 | updated, and that we don't know the location of the death of the | |
10494 | register. */ | |
230d793d RS |
10495 | for (i = regno; i < endregno; i ++) |
10496 | { | |
10497 | if (insn) | |
10498 | reg_last_set[i] = insn; | |
10499 | reg_last_set_value[i] = 0; | |
ef026f91 RS |
10500 | reg_last_set_mode[i] = 0; |
10501 | reg_last_set_nonzero_bits[i] = 0; | |
10502 | reg_last_set_sign_bit_copies[i] = 0; | |
230d793d RS |
10503 | reg_last_death[i] = 0; |
10504 | } | |
10505 | ||
10506 | /* Mark registers that are being referenced in this value. */ | |
10507 | if (value) | |
10508 | update_table_tick (value); | |
10509 | ||
10510 | /* Now update the status of each register being set. | |
10511 | If someone is using this register in this block, set this register | |
10512 | to invalid since we will get confused between the two lives in this | |
10513 | basic block. This makes using this register always invalid. In cse, we | |
10514 | scan the table to invalidate all entries using this register, but this | |
10515 | is too much work for us. */ | |
10516 | ||
10517 | for (i = regno; i < endregno; i++) | |
10518 | { | |
10519 | reg_last_set_label[i] = label_tick; | |
10520 | if (value && reg_last_set_table_tick[i] == label_tick) | |
10521 | reg_last_set_invalid[i] = 1; | |
10522 | else | |
10523 | reg_last_set_invalid[i] = 0; | |
10524 | } | |
10525 | ||
10526 | /* The value being assigned might refer to X (like in "x++;"). In that | |
10527 | case, we must replace it with (clobber (const_int 0)) to prevent | |
10528 | infinite loops. */ | |
9a893315 | 10529 | if (value && ! get_last_value_validate (&value, insn, |
230d793d RS |
10530 | reg_last_set_label[regno], 0)) |
10531 | { | |
10532 | value = copy_rtx (value); | |
9a893315 JW |
10533 | if (! get_last_value_validate (&value, insn, |
10534 | reg_last_set_label[regno], 1)) | |
230d793d RS |
10535 | value = 0; |
10536 | } | |
10537 | ||
55310dad RK |
10538 | /* For the main register being modified, update the value, the mode, the |
10539 | nonzero bits, and the number of sign bit copies. */ | |
10540 | ||
230d793d RS |
10541 | reg_last_set_value[regno] = value; |
10542 | ||
55310dad RK |
10543 | if (value) |
10544 | { | |
2afabb48 | 10545 | subst_low_cuid = INSN_CUID (insn); |
55310dad RK |
10546 | reg_last_set_mode[regno] = GET_MODE (reg); |
10547 | reg_last_set_nonzero_bits[regno] = nonzero_bits (value, GET_MODE (reg)); | |
10548 | reg_last_set_sign_bit_copies[regno] | |
10549 | = num_sign_bit_copies (value, GET_MODE (reg)); | |
10550 | } | |
230d793d RS |
10551 | } |
10552 | ||
10553 | /* Used for communication between the following two routines. */ | |
10554 | static rtx record_dead_insn; | |
10555 | ||
10556 | /* Called via note_stores from record_dead_and_set_regs to handle one | |
10557 | SET or CLOBBER in an insn. */ | |
10558 | ||
10559 | static void | |
10560 | record_dead_and_set_regs_1 (dest, setter) | |
10561 | rtx dest, setter; | |
10562 | { | |
ca89d290 RK |
10563 | if (GET_CODE (dest) == SUBREG) |
10564 | dest = SUBREG_REG (dest); | |
10565 | ||
230d793d RS |
10566 | if (GET_CODE (dest) == REG) |
10567 | { | |
10568 | /* If we are setting the whole register, we know its value. Otherwise | |
10569 | show that we don't know the value. We can handle SUBREG in | |
10570 | some cases. */ | |
10571 | if (GET_CODE (setter) == SET && dest == SET_DEST (setter)) | |
10572 | record_value_for_reg (dest, record_dead_insn, SET_SRC (setter)); | |
10573 | else if (GET_CODE (setter) == SET | |
10574 | && GET_CODE (SET_DEST (setter)) == SUBREG | |
10575 | && SUBREG_REG (SET_DEST (setter)) == dest | |
90bf8081 | 10576 | && GET_MODE_BITSIZE (GET_MODE (dest)) <= BITS_PER_WORD |
230d793d | 10577 | && subreg_lowpart_p (SET_DEST (setter))) |
d0ab8cd3 RK |
10578 | record_value_for_reg (dest, record_dead_insn, |
10579 | gen_lowpart_for_combine (GET_MODE (dest), | |
10580 | SET_SRC (setter))); | |
230d793d | 10581 | else |
5f4f0e22 | 10582 | record_value_for_reg (dest, record_dead_insn, NULL_RTX); |
230d793d RS |
10583 | } |
10584 | else if (GET_CODE (dest) == MEM | |
10585 | /* Ignore pushes, they clobber nothing. */ | |
10586 | && ! push_operand (dest, GET_MODE (dest))) | |
10587 | mem_last_set = INSN_CUID (record_dead_insn); | |
10588 | } | |
10589 | ||
10590 | /* Update the records of when each REG was most recently set or killed | |
10591 | for the things done by INSN. This is the last thing done in processing | |
10592 | INSN in the combiner loop. | |
10593 | ||
ef026f91 RS |
10594 | We update reg_last_set, reg_last_set_value, reg_last_set_mode, |
10595 | reg_last_set_nonzero_bits, reg_last_set_sign_bit_copies, reg_last_death, | |
10596 | and also the similar information mem_last_set (which insn most recently | |
10597 | modified memory) and last_call_cuid (which insn was the most recent | |
10598 | subroutine call). */ | |
230d793d RS |
10599 | |
10600 | static void | |
10601 | record_dead_and_set_regs (insn) | |
10602 | rtx insn; | |
10603 | { | |
10604 | register rtx link; | |
55310dad RK |
10605 | int i; |
10606 | ||
230d793d RS |
10607 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) |
10608 | { | |
dbc131f3 RK |
10609 | if (REG_NOTE_KIND (link) == REG_DEAD |
10610 | && GET_CODE (XEXP (link, 0)) == REG) | |
10611 | { | |
10612 | int regno = REGNO (XEXP (link, 0)); | |
10613 | int endregno | |
10614 | = regno + (regno < FIRST_PSEUDO_REGISTER | |
10615 | ? HARD_REGNO_NREGS (regno, GET_MODE (XEXP (link, 0))) | |
10616 | : 1); | |
dbc131f3 RK |
10617 | |
10618 | for (i = regno; i < endregno; i++) | |
10619 | reg_last_death[i] = insn; | |
10620 | } | |
230d793d | 10621 | else if (REG_NOTE_KIND (link) == REG_INC) |
5f4f0e22 | 10622 | record_value_for_reg (XEXP (link, 0), insn, NULL_RTX); |
230d793d RS |
10623 | } |
10624 | ||
10625 | if (GET_CODE (insn) == CALL_INSN) | |
55310dad RK |
10626 | { |
10627 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
10628 | if (call_used_regs[i]) | |
10629 | { | |
10630 | reg_last_set_value[i] = 0; | |
ef026f91 RS |
10631 | reg_last_set_mode[i] = 0; |
10632 | reg_last_set_nonzero_bits[i] = 0; | |
10633 | reg_last_set_sign_bit_copies[i] = 0; | |
55310dad RK |
10634 | reg_last_death[i] = 0; |
10635 | } | |
10636 | ||
10637 | last_call_cuid = mem_last_set = INSN_CUID (insn); | |
10638 | } | |
230d793d RS |
10639 | |
10640 | record_dead_insn = insn; | |
10641 | note_stores (PATTERN (insn), record_dead_and_set_regs_1); | |
10642 | } | |
10643 | \f | |
10644 | /* Utility routine for the following function. Verify that all the registers | |
10645 | mentioned in *LOC are valid when *LOC was part of a value set when | |
10646 | label_tick == TICK. Return 0 if some are not. | |
10647 | ||
10648 | If REPLACE is non-zero, replace the invalid reference with | |
10649 | (clobber (const_int 0)) and return 1. This replacement is useful because | |
10650 | we often can get useful information about the form of a value (e.g., if | |
10651 | it was produced by a shift that always produces -1 or 0) even though | |
10652 | we don't know exactly what registers it was produced from. */ | |
10653 | ||
10654 | static int | |
9a893315 | 10655 | get_last_value_validate (loc, insn, tick, replace) |
230d793d | 10656 | rtx *loc; |
9a893315 | 10657 | rtx insn; |
230d793d RS |
10658 | int tick; |
10659 | int replace; | |
10660 | { | |
10661 | rtx x = *loc; | |
10662 | char *fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
10663 | int len = GET_RTX_LENGTH (GET_CODE (x)); | |
10664 | int i; | |
10665 | ||
10666 | if (GET_CODE (x) == REG) | |
10667 | { | |
10668 | int regno = REGNO (x); | |
10669 | int endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
10670 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
10671 | int j; | |
10672 | ||
10673 | for (j = regno; j < endregno; j++) | |
10674 | if (reg_last_set_invalid[j] | |
10675 | /* If this is a pseudo-register that was only set once, it is | |
10676 | always valid. */ | |
b1f21e0a | 10677 | || (! (regno >= FIRST_PSEUDO_REGISTER && REG_N_SETS (regno) == 1) |
230d793d RS |
10678 | && reg_last_set_label[j] > tick)) |
10679 | { | |
10680 | if (replace) | |
38a448ca | 10681 | *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
230d793d RS |
10682 | return replace; |
10683 | } | |
10684 | ||
10685 | return 1; | |
10686 | } | |
9a893315 JW |
10687 | /* If this is a memory reference, make sure that there were |
10688 | no stores after it that might have clobbered the value. We don't | |
10689 | have alias info, so we assume any store invalidates it. */ | |
10690 | else if (GET_CODE (x) == MEM && ! RTX_UNCHANGING_P (x) | |
10691 | && INSN_CUID (insn) <= mem_last_set) | |
10692 | { | |
10693 | if (replace) | |
38a448ca | 10694 | *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
9a893315 JW |
10695 | return replace; |
10696 | } | |
230d793d RS |
10697 | |
10698 | for (i = 0; i < len; i++) | |
10699 | if ((fmt[i] == 'e' | |
9a893315 | 10700 | && get_last_value_validate (&XEXP (x, i), insn, tick, replace) == 0) |
230d793d RS |
10701 | /* Don't bother with these. They shouldn't occur anyway. */ |
10702 | || fmt[i] == 'E') | |
10703 | return 0; | |
10704 | ||
10705 | /* If we haven't found a reason for it to be invalid, it is valid. */ | |
10706 | return 1; | |
10707 | } | |
10708 | ||
10709 | /* Get the last value assigned to X, if known. Some registers | |
10710 | in the value may be replaced with (clobber (const_int 0)) if their value | |
10711 | is known longer known reliably. */ | |
10712 | ||
10713 | static rtx | |
10714 | get_last_value (x) | |
10715 | rtx x; | |
10716 | { | |
10717 | int regno; | |
10718 | rtx value; | |
10719 | ||
10720 | /* If this is a non-paradoxical SUBREG, get the value of its operand and | |
10721 | then convert it to the desired mode. If this is a paradoxical SUBREG, | |
0f41302f | 10722 | we cannot predict what values the "extra" bits might have. */ |
230d793d RS |
10723 | if (GET_CODE (x) == SUBREG |
10724 | && subreg_lowpart_p (x) | |
10725 | && (GET_MODE_SIZE (GET_MODE (x)) | |
10726 | <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))) | |
10727 | && (value = get_last_value (SUBREG_REG (x))) != 0) | |
10728 | return gen_lowpart_for_combine (GET_MODE (x), value); | |
10729 | ||
10730 | if (GET_CODE (x) != REG) | |
10731 | return 0; | |
10732 | ||
10733 | regno = REGNO (x); | |
10734 | value = reg_last_set_value[regno]; | |
10735 | ||
0f41302f MS |
10736 | /* If we don't have a value or if it isn't for this basic block, |
10737 | return 0. */ | |
230d793d RS |
10738 | |
10739 | if (value == 0 | |
b1f21e0a | 10740 | || (REG_N_SETS (regno) != 1 |
55310dad | 10741 | && reg_last_set_label[regno] != label_tick)) |
230d793d RS |
10742 | return 0; |
10743 | ||
4255220d | 10744 | /* If the value was set in a later insn than the ones we are processing, |
4090a6b3 RK |
10745 | we can't use it even if the register was only set once, but make a quick |
10746 | check to see if the previous insn set it to something. This is commonly | |
0d9641d1 JW |
10747 | the case when the same pseudo is used by repeated insns. |
10748 | ||
10749 | This does not work if there exists an instruction which is temporarily | |
10750 | not on the insn chain. */ | |
d0ab8cd3 | 10751 | |
bcd49eb7 | 10752 | if (INSN_CUID (reg_last_set[regno]) >= subst_low_cuid) |
d0ab8cd3 RK |
10753 | { |
10754 | rtx insn, set; | |
10755 | ||
bcd49eb7 JW |
10756 | /* We can not do anything useful in this case, because there is |
10757 | an instruction which is not on the insn chain. */ | |
10758 | if (subst_prev_insn) | |
10759 | return 0; | |
10760 | ||
4255220d JW |
10761 | /* Skip over USE insns. They are not useful here, and they may have |
10762 | been made by combine, in which case they do not have a INSN_CUID | |
d6c80562 | 10763 | value. We can't use prev_real_insn, because that would incorrectly |
e340018d JW |
10764 | take us backwards across labels. Skip over BARRIERs also, since |
10765 | they could have been made by combine. If we see one, we must be | |
10766 | optimizing dead code, so it doesn't matter what we do. */ | |
d6c80562 JW |
10767 | for (insn = prev_nonnote_insn (subst_insn); |
10768 | insn && ((GET_CODE (insn) == INSN | |
10769 | && GET_CODE (PATTERN (insn)) == USE) | |
e340018d | 10770 | || GET_CODE (insn) == BARRIER |
4255220d | 10771 | || INSN_CUID (insn) >= subst_low_cuid); |
d6c80562 | 10772 | insn = prev_nonnote_insn (insn)) |
3adde2a5 | 10773 | ; |
d0ab8cd3 RK |
10774 | |
10775 | if (insn | |
10776 | && (set = single_set (insn)) != 0 | |
10777 | && rtx_equal_p (SET_DEST (set), x)) | |
10778 | { | |
10779 | value = SET_SRC (set); | |
10780 | ||
10781 | /* Make sure that VALUE doesn't reference X. Replace any | |
ddd5a7c1 | 10782 | explicit references with a CLOBBER. If there are any remaining |
d0ab8cd3 RK |
10783 | references (rare), don't use the value. */ |
10784 | ||
10785 | if (reg_mentioned_p (x, value)) | |
10786 | value = replace_rtx (copy_rtx (value), x, | |
38a448ca | 10787 | gen_rtx_CLOBBER (GET_MODE (x), const0_rtx)); |
d0ab8cd3 RK |
10788 | |
10789 | if (reg_overlap_mentioned_p (x, value)) | |
10790 | return 0; | |
10791 | } | |
10792 | else | |
10793 | return 0; | |
10794 | } | |
10795 | ||
10796 | /* If the value has all its registers valid, return it. */ | |
9a893315 JW |
10797 | if (get_last_value_validate (&value, reg_last_set[regno], |
10798 | reg_last_set_label[regno], 0)) | |
230d793d RS |
10799 | return value; |
10800 | ||
10801 | /* Otherwise, make a copy and replace any invalid register with | |
10802 | (clobber (const_int 0)). If that fails for some reason, return 0. */ | |
10803 | ||
10804 | value = copy_rtx (value); | |
9a893315 JW |
10805 | if (get_last_value_validate (&value, reg_last_set[regno], |
10806 | reg_last_set_label[regno], 1)) | |
230d793d RS |
10807 | return value; |
10808 | ||
10809 | return 0; | |
10810 | } | |
10811 | \f | |
10812 | /* Return nonzero if expression X refers to a REG or to memory | |
10813 | that is set in an instruction more recent than FROM_CUID. */ | |
10814 | ||
10815 | static int | |
10816 | use_crosses_set_p (x, from_cuid) | |
10817 | register rtx x; | |
10818 | int from_cuid; | |
10819 | { | |
10820 | register char *fmt; | |
10821 | register int i; | |
10822 | register enum rtx_code code = GET_CODE (x); | |
10823 | ||
10824 | if (code == REG) | |
10825 | { | |
10826 | register int regno = REGNO (x); | |
e28f5732 RK |
10827 | int endreg = regno + (regno < FIRST_PSEUDO_REGISTER |
10828 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
10829 | ||
230d793d RS |
10830 | #ifdef PUSH_ROUNDING |
10831 | /* Don't allow uses of the stack pointer to be moved, | |
10832 | because we don't know whether the move crosses a push insn. */ | |
10833 | if (regno == STACK_POINTER_REGNUM) | |
10834 | return 1; | |
10835 | #endif | |
e28f5732 RK |
10836 | for (;regno < endreg; regno++) |
10837 | if (reg_last_set[regno] | |
10838 | && INSN_CUID (reg_last_set[regno]) > from_cuid) | |
10839 | return 1; | |
10840 | return 0; | |
230d793d RS |
10841 | } |
10842 | ||
10843 | if (code == MEM && mem_last_set > from_cuid) | |
10844 | return 1; | |
10845 | ||
10846 | fmt = GET_RTX_FORMAT (code); | |
10847 | ||
10848 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
10849 | { | |
10850 | if (fmt[i] == 'E') | |
10851 | { | |
10852 | register int j; | |
10853 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
10854 | if (use_crosses_set_p (XVECEXP (x, i, j), from_cuid)) | |
10855 | return 1; | |
10856 | } | |
10857 | else if (fmt[i] == 'e' | |
10858 | && use_crosses_set_p (XEXP (x, i), from_cuid)) | |
10859 | return 1; | |
10860 | } | |
10861 | return 0; | |
10862 | } | |
10863 | \f | |
10864 | /* Define three variables used for communication between the following | |
10865 | routines. */ | |
10866 | ||
10867 | static int reg_dead_regno, reg_dead_endregno; | |
10868 | static int reg_dead_flag; | |
10869 | ||
10870 | /* Function called via note_stores from reg_dead_at_p. | |
10871 | ||
ddd5a7c1 | 10872 | If DEST is within [reg_dead_regno, reg_dead_endregno), set |
230d793d RS |
10873 | reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */ |
10874 | ||
10875 | static void | |
10876 | reg_dead_at_p_1 (dest, x) | |
10877 | rtx dest; | |
10878 | rtx x; | |
10879 | { | |
10880 | int regno, endregno; | |
10881 | ||
10882 | if (GET_CODE (dest) != REG) | |
10883 | return; | |
10884 | ||
10885 | regno = REGNO (dest); | |
10886 | endregno = regno + (regno < FIRST_PSEUDO_REGISTER | |
10887 | ? HARD_REGNO_NREGS (regno, GET_MODE (dest)) : 1); | |
10888 | ||
10889 | if (reg_dead_endregno > regno && reg_dead_regno < endregno) | |
10890 | reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1; | |
10891 | } | |
10892 | ||
10893 | /* Return non-zero if REG is known to be dead at INSN. | |
10894 | ||
10895 | We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER | |
10896 | referencing REG, it is dead. If we hit a SET referencing REG, it is | |
10897 | live. Otherwise, see if it is live or dead at the start of the basic | |
6e25d159 RK |
10898 | block we are in. Hard regs marked as being live in NEWPAT_USED_REGS |
10899 | must be assumed to be always live. */ | |
230d793d RS |
10900 | |
10901 | static int | |
10902 | reg_dead_at_p (reg, insn) | |
10903 | rtx reg; | |
10904 | rtx insn; | |
10905 | { | |
10906 | int block, i; | |
10907 | ||
10908 | /* Set variables for reg_dead_at_p_1. */ | |
10909 | reg_dead_regno = REGNO (reg); | |
10910 | reg_dead_endregno = reg_dead_regno + (reg_dead_regno < FIRST_PSEUDO_REGISTER | |
10911 | ? HARD_REGNO_NREGS (reg_dead_regno, | |
10912 | GET_MODE (reg)) | |
10913 | : 1); | |
10914 | ||
10915 | reg_dead_flag = 0; | |
10916 | ||
6e25d159 RK |
10917 | /* Check that reg isn't mentioned in NEWPAT_USED_REGS. */ |
10918 | if (reg_dead_regno < FIRST_PSEUDO_REGISTER) | |
10919 | { | |
10920 | for (i = reg_dead_regno; i < reg_dead_endregno; i++) | |
10921 | if (TEST_HARD_REG_BIT (newpat_used_regs, i)) | |
10922 | return 0; | |
10923 | } | |
10924 | ||
230d793d RS |
10925 | /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, label, or |
10926 | beginning of function. */ | |
60715d0b | 10927 | for (; insn && GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != BARRIER; |
230d793d RS |
10928 | insn = prev_nonnote_insn (insn)) |
10929 | { | |
10930 | note_stores (PATTERN (insn), reg_dead_at_p_1); | |
10931 | if (reg_dead_flag) | |
10932 | return reg_dead_flag == 1 ? 1 : 0; | |
10933 | ||
10934 | if (find_regno_note (insn, REG_DEAD, reg_dead_regno)) | |
10935 | return 1; | |
10936 | } | |
10937 | ||
10938 | /* Get the basic block number that we were in. */ | |
10939 | if (insn == 0) | |
10940 | block = 0; | |
10941 | else | |
10942 | { | |
10943 | for (block = 0; block < n_basic_blocks; block++) | |
10944 | if (insn == basic_block_head[block]) | |
10945 | break; | |
10946 | ||
10947 | if (block == n_basic_blocks) | |
10948 | return 0; | |
10949 | } | |
10950 | ||
10951 | for (i = reg_dead_regno; i < reg_dead_endregno; i++) | |
8e08106d | 10952 | if (REGNO_REG_SET_P (basic_block_live_at_start[block], i)) |
230d793d RS |
10953 | return 0; |
10954 | ||
10955 | return 1; | |
10956 | } | |
6e25d159 RK |
10957 | \f |
10958 | /* Note hard registers in X that are used. This code is similar to | |
10959 | that in flow.c, but much simpler since we don't care about pseudos. */ | |
10960 | ||
10961 | static void | |
10962 | mark_used_regs_combine (x) | |
10963 | rtx x; | |
10964 | { | |
10965 | register RTX_CODE code = GET_CODE (x); | |
10966 | register int regno; | |
10967 | int i; | |
10968 | ||
10969 | switch (code) | |
10970 | { | |
10971 | case LABEL_REF: | |
10972 | case SYMBOL_REF: | |
10973 | case CONST_INT: | |
10974 | case CONST: | |
10975 | case CONST_DOUBLE: | |
10976 | case PC: | |
10977 | case ADDR_VEC: | |
10978 | case ADDR_DIFF_VEC: | |
10979 | case ASM_INPUT: | |
10980 | #ifdef HAVE_cc0 | |
10981 | /* CC0 must die in the insn after it is set, so we don't need to take | |
10982 | special note of it here. */ | |
10983 | case CC0: | |
10984 | #endif | |
10985 | return; | |
10986 | ||
10987 | case CLOBBER: | |
10988 | /* If we are clobbering a MEM, mark any hard registers inside the | |
10989 | address as used. */ | |
10990 | if (GET_CODE (XEXP (x, 0)) == MEM) | |
10991 | mark_used_regs_combine (XEXP (XEXP (x, 0), 0)); | |
10992 | return; | |
10993 | ||
10994 | case REG: | |
10995 | regno = REGNO (x); | |
10996 | /* A hard reg in a wide mode may really be multiple registers. | |
10997 | If so, mark all of them just like the first. */ | |
10998 | if (regno < FIRST_PSEUDO_REGISTER) | |
10999 | { | |
11000 | /* None of this applies to the stack, frame or arg pointers */ | |
11001 | if (regno == STACK_POINTER_REGNUM | |
11002 | #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM | |
11003 | || regno == HARD_FRAME_POINTER_REGNUM | |
11004 | #endif | |
11005 | #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM | |
11006 | || (regno == ARG_POINTER_REGNUM && fixed_regs[regno]) | |
11007 | #endif | |
11008 | || regno == FRAME_POINTER_REGNUM) | |
11009 | return; | |
11010 | ||
11011 | i = HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
11012 | while (i-- > 0) | |
11013 | SET_HARD_REG_BIT (newpat_used_regs, regno + i); | |
11014 | } | |
11015 | return; | |
11016 | ||
11017 | case SET: | |
11018 | { | |
11019 | /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in | |
11020 | the address. */ | |
11021 | register rtx testreg = SET_DEST (x); | |
11022 | ||
e048778f RK |
11023 | while (GET_CODE (testreg) == SUBREG |
11024 | || GET_CODE (testreg) == ZERO_EXTRACT | |
11025 | || GET_CODE (testreg) == SIGN_EXTRACT | |
11026 | || GET_CODE (testreg) == STRICT_LOW_PART) | |
6e25d159 RK |
11027 | testreg = XEXP (testreg, 0); |
11028 | ||
11029 | if (GET_CODE (testreg) == MEM) | |
11030 | mark_used_regs_combine (XEXP (testreg, 0)); | |
11031 | ||
11032 | mark_used_regs_combine (SET_SRC (x)); | |
6e25d159 | 11033 | } |
e9a25f70 JL |
11034 | return; |
11035 | ||
11036 | default: | |
11037 | break; | |
6e25d159 RK |
11038 | } |
11039 | ||
11040 | /* Recursively scan the operands of this expression. */ | |
11041 | ||
11042 | { | |
11043 | register char *fmt = GET_RTX_FORMAT (code); | |
11044 | ||
11045 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
11046 | { | |
11047 | if (fmt[i] == 'e') | |
11048 | mark_used_regs_combine (XEXP (x, i)); | |
11049 | else if (fmt[i] == 'E') | |
11050 | { | |
11051 | register int j; | |
11052 | ||
11053 | for (j = 0; j < XVECLEN (x, i); j++) | |
11054 | mark_used_regs_combine (XVECEXP (x, i, j)); | |
11055 | } | |
11056 | } | |
11057 | } | |
11058 | } | |
11059 | ||
230d793d RS |
11060 | \f |
11061 | /* Remove register number REGNO from the dead registers list of INSN. | |
11062 | ||
11063 | Return the note used to record the death, if there was one. */ | |
11064 | ||
11065 | rtx | |
11066 | remove_death (regno, insn) | |
11067 | int regno; | |
11068 | rtx insn; | |
11069 | { | |
11070 | register rtx note = find_regno_note (insn, REG_DEAD, regno); | |
11071 | ||
11072 | if (note) | |
1a26b032 | 11073 | { |
b1f21e0a | 11074 | REG_N_DEATHS (regno)--; |
1a26b032 RK |
11075 | remove_note (insn, note); |
11076 | } | |
230d793d RS |
11077 | |
11078 | return note; | |
11079 | } | |
11080 | ||
11081 | /* For each register (hardware or pseudo) used within expression X, if its | |
11082 | death is in an instruction with cuid between FROM_CUID (inclusive) and | |
11083 | TO_INSN (exclusive), put a REG_DEAD note for that register in the | |
11084 | list headed by PNOTES. | |
11085 | ||
6eb12cef RK |
11086 | That said, don't move registers killed by maybe_kill_insn. |
11087 | ||
230d793d RS |
11088 | This is done when X is being merged by combination into TO_INSN. These |
11089 | notes will then be distributed as needed. */ | |
11090 | ||
11091 | static void | |
6eb12cef | 11092 | move_deaths (x, maybe_kill_insn, from_cuid, to_insn, pnotes) |
230d793d | 11093 | rtx x; |
6eb12cef | 11094 | rtx maybe_kill_insn; |
230d793d RS |
11095 | int from_cuid; |
11096 | rtx to_insn; | |
11097 | rtx *pnotes; | |
11098 | { | |
11099 | register char *fmt; | |
11100 | register int len, i; | |
11101 | register enum rtx_code code = GET_CODE (x); | |
11102 | ||
11103 | if (code == REG) | |
11104 | { | |
11105 | register int regno = REGNO (x); | |
11106 | register rtx where_dead = reg_last_death[regno]; | |
e340018d JW |
11107 | register rtx before_dead, after_dead; |
11108 | ||
6eb12cef RK |
11109 | /* Don't move the register if it gets killed in between from and to */ |
11110 | if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn) | |
11111 | && !reg_referenced_p (x, maybe_kill_insn)) | |
11112 | return; | |
11113 | ||
e340018d JW |
11114 | /* WHERE_DEAD could be a USE insn made by combine, so first we |
11115 | make sure that we have insns with valid INSN_CUID values. */ | |
11116 | before_dead = where_dead; | |
11117 | while (before_dead && INSN_UID (before_dead) > max_uid_cuid) | |
11118 | before_dead = PREV_INSN (before_dead); | |
11119 | after_dead = where_dead; | |
11120 | while (after_dead && INSN_UID (after_dead) > max_uid_cuid) | |
11121 | after_dead = NEXT_INSN (after_dead); | |
11122 | ||
11123 | if (before_dead && after_dead | |
11124 | && INSN_CUID (before_dead) >= from_cuid | |
11125 | && (INSN_CUID (after_dead) < INSN_CUID (to_insn) | |
11126 | || (where_dead != after_dead | |
11127 | && INSN_CUID (after_dead) == INSN_CUID (to_insn)))) | |
230d793d | 11128 | { |
dbc131f3 | 11129 | rtx note = remove_death (regno, where_dead); |
230d793d RS |
11130 | |
11131 | /* It is possible for the call above to return 0. This can occur | |
11132 | when reg_last_death points to I2 or I1 that we combined with. | |
dbc131f3 RK |
11133 | In that case make a new note. |
11134 | ||
11135 | We must also check for the case where X is a hard register | |
11136 | and NOTE is a death note for a range of hard registers | |
11137 | including X. In that case, we must put REG_DEAD notes for | |
11138 | the remaining registers in place of NOTE. */ | |
11139 | ||
11140 | if (note != 0 && regno < FIRST_PSEUDO_REGISTER | |
11141 | && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0))) | |
24e46fc4 | 11142 | > GET_MODE_SIZE (GET_MODE (x)))) |
dbc131f3 RK |
11143 | { |
11144 | int deadregno = REGNO (XEXP (note, 0)); | |
11145 | int deadend | |
11146 | = (deadregno + HARD_REGNO_NREGS (deadregno, | |
11147 | GET_MODE (XEXP (note, 0)))); | |
11148 | int ourend = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
11149 | int i; | |
11150 | ||
11151 | for (i = deadregno; i < deadend; i++) | |
11152 | if (i < regno || i >= ourend) | |
11153 | REG_NOTES (where_dead) | |
38a448ca RH |
11154 | = gen_rtx_EXPR_LIST (REG_DEAD, |
11155 | gen_rtx_REG (reg_raw_mode[i], i), | |
11156 | REG_NOTES (where_dead)); | |
dbc131f3 | 11157 | } |
24e46fc4 JW |
11158 | /* If we didn't find any note, or if we found a REG_DEAD note that |
11159 | covers only part of the given reg, and we have a multi-reg hard | |
fabd69e8 RK |
11160 | register, then to be safe we must check for REG_DEAD notes |
11161 | for each register other than the first. They could have | |
11162 | their own REG_DEAD notes lying around. */ | |
24e46fc4 JW |
11163 | else if ((note == 0 |
11164 | || (note != 0 | |
11165 | && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0))) | |
11166 | < GET_MODE_SIZE (GET_MODE (x))))) | |
11167 | && regno < FIRST_PSEUDO_REGISTER | |
fabd69e8 RK |
11168 | && HARD_REGNO_NREGS (regno, GET_MODE (x)) > 1) |
11169 | { | |
11170 | int ourend = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
24e46fc4 | 11171 | int i, offset; |
fabd69e8 RK |
11172 | rtx oldnotes = 0; |
11173 | ||
24e46fc4 JW |
11174 | if (note) |
11175 | offset = HARD_REGNO_NREGS (regno, GET_MODE (XEXP (note, 0))); | |
11176 | else | |
11177 | offset = 1; | |
11178 | ||
11179 | for (i = regno + offset; i < ourend; i++) | |
38a448ca | 11180 | move_deaths (gen_rtx_REG (reg_raw_mode[i], i), |
6eb12cef | 11181 | maybe_kill_insn, from_cuid, to_insn, &oldnotes); |
fabd69e8 | 11182 | } |
230d793d | 11183 | |
dbc131f3 | 11184 | if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x)) |
230d793d RS |
11185 | { |
11186 | XEXP (note, 1) = *pnotes; | |
11187 | *pnotes = note; | |
11188 | } | |
11189 | else | |
38a448ca | 11190 | *pnotes = gen_rtx_EXPR_LIST (REG_DEAD, x, *pnotes); |
1a26b032 | 11191 | |
b1f21e0a | 11192 | REG_N_DEATHS (regno)++; |
230d793d RS |
11193 | } |
11194 | ||
11195 | return; | |
11196 | } | |
11197 | ||
11198 | else if (GET_CODE (x) == SET) | |
11199 | { | |
11200 | rtx dest = SET_DEST (x); | |
11201 | ||
6eb12cef | 11202 | move_deaths (SET_SRC (x), maybe_kill_insn, from_cuid, to_insn, pnotes); |
230d793d | 11203 | |
a7c99304 RK |
11204 | /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG |
11205 | that accesses one word of a multi-word item, some | |
11206 | piece of everything register in the expression is used by | |
11207 | this insn, so remove any old death. */ | |
11208 | ||
11209 | if (GET_CODE (dest) == ZERO_EXTRACT | |
11210 | || GET_CODE (dest) == STRICT_LOW_PART | |
11211 | || (GET_CODE (dest) == SUBREG | |
11212 | && (((GET_MODE_SIZE (GET_MODE (dest)) | |
11213 | + UNITS_PER_WORD - 1) / UNITS_PER_WORD) | |
11214 | == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) | |
11215 | + UNITS_PER_WORD - 1) / UNITS_PER_WORD)))) | |
230d793d | 11216 | { |
6eb12cef | 11217 | move_deaths (dest, maybe_kill_insn, from_cuid, to_insn, pnotes); |
a7c99304 | 11218 | return; |
230d793d RS |
11219 | } |
11220 | ||
a7c99304 RK |
11221 | /* If this is some other SUBREG, we know it replaces the entire |
11222 | value, so use that as the destination. */ | |
11223 | if (GET_CODE (dest) == SUBREG) | |
11224 | dest = SUBREG_REG (dest); | |
11225 | ||
11226 | /* If this is a MEM, adjust deaths of anything used in the address. | |
11227 | For a REG (the only other possibility), the entire value is | |
11228 | being replaced so the old value is not used in this insn. */ | |
230d793d RS |
11229 | |
11230 | if (GET_CODE (dest) == MEM) | |
6eb12cef RK |
11231 | move_deaths (XEXP (dest, 0), maybe_kill_insn, from_cuid, |
11232 | to_insn, pnotes); | |
230d793d RS |
11233 | return; |
11234 | } | |
11235 | ||
11236 | else if (GET_CODE (x) == CLOBBER) | |
11237 | return; | |
11238 | ||
11239 | len = GET_RTX_LENGTH (code); | |
11240 | fmt = GET_RTX_FORMAT (code); | |
11241 | ||
11242 | for (i = 0; i < len; i++) | |
11243 | { | |
11244 | if (fmt[i] == 'E') | |
11245 | { | |
11246 | register int j; | |
11247 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
6eb12cef RK |
11248 | move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_cuid, |
11249 | to_insn, pnotes); | |
230d793d RS |
11250 | } |
11251 | else if (fmt[i] == 'e') | |
6eb12cef | 11252 | move_deaths (XEXP (x, i), maybe_kill_insn, from_cuid, to_insn, pnotes); |
230d793d RS |
11253 | } |
11254 | } | |
11255 | \f | |
a7c99304 RK |
11256 | /* Return 1 if X is the target of a bit-field assignment in BODY, the |
11257 | pattern of an insn. X must be a REG. */ | |
230d793d RS |
11258 | |
11259 | static int | |
a7c99304 RK |
11260 | reg_bitfield_target_p (x, body) |
11261 | rtx x; | |
230d793d RS |
11262 | rtx body; |
11263 | { | |
11264 | int i; | |
11265 | ||
11266 | if (GET_CODE (body) == SET) | |
a7c99304 RK |
11267 | { |
11268 | rtx dest = SET_DEST (body); | |
11269 | rtx target; | |
11270 | int regno, tregno, endregno, endtregno; | |
11271 | ||
11272 | if (GET_CODE (dest) == ZERO_EXTRACT) | |
11273 | target = XEXP (dest, 0); | |
11274 | else if (GET_CODE (dest) == STRICT_LOW_PART) | |
11275 | target = SUBREG_REG (XEXP (dest, 0)); | |
11276 | else | |
11277 | return 0; | |
11278 | ||
11279 | if (GET_CODE (target) == SUBREG) | |
11280 | target = SUBREG_REG (target); | |
11281 | ||
11282 | if (GET_CODE (target) != REG) | |
11283 | return 0; | |
11284 | ||
11285 | tregno = REGNO (target), regno = REGNO (x); | |
11286 | if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER) | |
11287 | return target == x; | |
11288 | ||
11289 | endtregno = tregno + HARD_REGNO_NREGS (tregno, GET_MODE (target)); | |
11290 | endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
11291 | ||
11292 | return endregno > tregno && regno < endtregno; | |
11293 | } | |
230d793d RS |
11294 | |
11295 | else if (GET_CODE (body) == PARALLEL) | |
11296 | for (i = XVECLEN (body, 0) - 1; i >= 0; i--) | |
a7c99304 | 11297 | if (reg_bitfield_target_p (x, XVECEXP (body, 0, i))) |
230d793d RS |
11298 | return 1; |
11299 | ||
11300 | return 0; | |
11301 | } | |
11302 | \f | |
11303 | /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them | |
11304 | as appropriate. I3 and I2 are the insns resulting from the combination | |
11305 | insns including FROM (I2 may be zero). | |
11306 | ||
11307 | ELIM_I2 and ELIM_I1 are either zero or registers that we know will | |
11308 | not need REG_DEAD notes because they are being substituted for. This | |
11309 | saves searching in the most common cases. | |
11310 | ||
11311 | Each note in the list is either ignored or placed on some insns, depending | |
11312 | on the type of note. */ | |
11313 | ||
11314 | static void | |
11315 | distribute_notes (notes, from_insn, i3, i2, elim_i2, elim_i1) | |
11316 | rtx notes; | |
11317 | rtx from_insn; | |
11318 | rtx i3, i2; | |
11319 | rtx elim_i2, elim_i1; | |
11320 | { | |
11321 | rtx note, next_note; | |
11322 | rtx tem; | |
11323 | ||
11324 | for (note = notes; note; note = next_note) | |
11325 | { | |
11326 | rtx place = 0, place2 = 0; | |
11327 | ||
11328 | /* If this NOTE references a pseudo register, ensure it references | |
11329 | the latest copy of that register. */ | |
11330 | if (XEXP (note, 0) && GET_CODE (XEXP (note, 0)) == REG | |
11331 | && REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER) | |
11332 | XEXP (note, 0) = regno_reg_rtx[REGNO (XEXP (note, 0))]; | |
11333 | ||
11334 | next_note = XEXP (note, 1); | |
11335 | switch (REG_NOTE_KIND (note)) | |
11336 | { | |
c9903b44 DE |
11337 | case REG_BR_PROB: |
11338 | case REG_EXEC_COUNT: | |
11339 | /* Doesn't matter much where we put this, as long as it's somewhere. | |
11340 | It is preferable to keep these notes on branches, which is most | |
11341 | likely to be i3. */ | |
11342 | place = i3; | |
11343 | break; | |
11344 | ||
230d793d | 11345 | case REG_UNUSED: |
07d0cbdd | 11346 | /* Any clobbers for i3 may still exist, and so we must process |
176c9e6b JW |
11347 | REG_UNUSED notes from that insn. |
11348 | ||
11349 | Any clobbers from i2 or i1 can only exist if they were added by | |
11350 | recog_for_combine. In that case, recog_for_combine created the | |
11351 | necessary REG_UNUSED notes. Trying to keep any original | |
11352 | REG_UNUSED notes from these insns can cause incorrect output | |
11353 | if it is for the same register as the original i3 dest. | |
11354 | In that case, we will notice that the register is set in i3, | |
11355 | and then add a REG_UNUSED note for the destination of i3, which | |
07d0cbdd JW |
11356 | is wrong. However, it is possible to have REG_UNUSED notes from |
11357 | i2 or i1 for register which were both used and clobbered, so | |
11358 | we keep notes from i2 or i1 if they will turn into REG_DEAD | |
11359 | notes. */ | |
176c9e6b | 11360 | |
230d793d RS |
11361 | /* If this register is set or clobbered in I3, put the note there |
11362 | unless there is one already. */ | |
07d0cbdd | 11363 | if (reg_set_p (XEXP (note, 0), PATTERN (i3))) |
230d793d | 11364 | { |
07d0cbdd JW |
11365 | if (from_insn != i3) |
11366 | break; | |
11367 | ||
230d793d RS |
11368 | if (! (GET_CODE (XEXP (note, 0)) == REG |
11369 | ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0))) | |
11370 | : find_reg_note (i3, REG_UNUSED, XEXP (note, 0)))) | |
11371 | place = i3; | |
11372 | } | |
11373 | /* Otherwise, if this register is used by I3, then this register | |
11374 | now dies here, so we must put a REG_DEAD note here unless there | |
11375 | is one already. */ | |
11376 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)) | |
11377 | && ! (GET_CODE (XEXP (note, 0)) == REG | |
11378 | ? find_regno_note (i3, REG_DEAD, REGNO (XEXP (note, 0))) | |
11379 | : find_reg_note (i3, REG_DEAD, XEXP (note, 0)))) | |
11380 | { | |
11381 | PUT_REG_NOTE_KIND (note, REG_DEAD); | |
11382 | place = i3; | |
11383 | } | |
11384 | break; | |
11385 | ||
11386 | case REG_EQUAL: | |
11387 | case REG_EQUIV: | |
11388 | case REG_NONNEG: | |
9ae8ffe7 | 11389 | case REG_NOALIAS: |
230d793d RS |
11390 | /* These notes say something about results of an insn. We can |
11391 | only support them if they used to be on I3 in which case they | |
a687e897 RK |
11392 | remain on I3. Otherwise they are ignored. |
11393 | ||
11394 | If the note refers to an expression that is not a constant, we | |
11395 | must also ignore the note since we cannot tell whether the | |
11396 | equivalence is still true. It might be possible to do | |
11397 | slightly better than this (we only have a problem if I2DEST | |
11398 | or I1DEST is present in the expression), but it doesn't | |
11399 | seem worth the trouble. */ | |
11400 | ||
11401 | if (from_insn == i3 | |
11402 | && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0)))) | |
230d793d RS |
11403 | place = i3; |
11404 | break; | |
11405 | ||
11406 | case REG_INC: | |
11407 | case REG_NO_CONFLICT: | |
11408 | case REG_LABEL: | |
11409 | /* These notes say something about how a register is used. They must | |
11410 | be present on any use of the register in I2 or I3. */ | |
11411 | if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))) | |
11412 | place = i3; | |
11413 | ||
11414 | if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2))) | |
11415 | { | |
11416 | if (place) | |
11417 | place2 = i2; | |
11418 | else | |
11419 | place = i2; | |
11420 | } | |
11421 | break; | |
11422 | ||
11423 | case REG_WAS_0: | |
11424 | /* It is too much trouble to try to see if this note is still | |
11425 | correct in all situations. It is better to simply delete it. */ | |
11426 | break; | |
11427 | ||
11428 | case REG_RETVAL: | |
11429 | /* If the insn previously containing this note still exists, | |
11430 | put it back where it was. Otherwise move it to the previous | |
11431 | insn. Adjust the corresponding REG_LIBCALL note. */ | |
11432 | if (GET_CODE (from_insn) != NOTE) | |
11433 | place = from_insn; | |
11434 | else | |
11435 | { | |
5f4f0e22 | 11436 | tem = find_reg_note (XEXP (note, 0), REG_LIBCALL, NULL_RTX); |
230d793d RS |
11437 | place = prev_real_insn (from_insn); |
11438 | if (tem && place) | |
11439 | XEXP (tem, 0) = place; | |
11440 | } | |
11441 | break; | |
11442 | ||
11443 | case REG_LIBCALL: | |
11444 | /* This is handled similarly to REG_RETVAL. */ | |
11445 | if (GET_CODE (from_insn) != NOTE) | |
11446 | place = from_insn; | |
11447 | else | |
11448 | { | |
5f4f0e22 | 11449 | tem = find_reg_note (XEXP (note, 0), REG_RETVAL, NULL_RTX); |
230d793d RS |
11450 | place = next_real_insn (from_insn); |
11451 | if (tem && place) | |
11452 | XEXP (tem, 0) = place; | |
11453 | } | |
11454 | break; | |
11455 | ||
11456 | case REG_DEAD: | |
11457 | /* If the register is used as an input in I3, it dies there. | |
11458 | Similarly for I2, if it is non-zero and adjacent to I3. | |
11459 | ||
11460 | If the register is not used as an input in either I3 or I2 | |
11461 | and it is not one of the registers we were supposed to eliminate, | |
11462 | there are two possibilities. We might have a non-adjacent I2 | |
11463 | or we might have somehow eliminated an additional register | |
11464 | from a computation. For example, we might have had A & B where | |
11465 | we discover that B will always be zero. In this case we will | |
11466 | eliminate the reference to A. | |
11467 | ||
11468 | In both cases, we must search to see if we can find a previous | |
11469 | use of A and put the death note there. */ | |
11470 | ||
6e2d1486 RK |
11471 | if (from_insn |
11472 | && GET_CODE (from_insn) == CALL_INSN | |
11473 | && find_reg_fusage (from_insn, USE, XEXP (note, 0))) | |
11474 | place = from_insn; | |
11475 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))) | |
230d793d RS |
11476 | place = i3; |
11477 | else if (i2 != 0 && next_nonnote_insn (i2) == i3 | |
11478 | && reg_referenced_p (XEXP (note, 0), PATTERN (i2))) | |
11479 | place = i2; | |
11480 | ||
11481 | if (XEXP (note, 0) == elim_i2 || XEXP (note, 0) == elim_i1) | |
11482 | break; | |
11483 | ||
510dd77e RK |
11484 | /* If the register is used in both I2 and I3 and it dies in I3, |
11485 | we might have added another reference to it. If reg_n_refs | |
11486 | was 2, bump it to 3. This has to be correct since the | |
11487 | register must have been set somewhere. The reason this is | |
11488 | done is because local-alloc.c treats 2 references as a | |
11489 | special case. */ | |
11490 | ||
11491 | if (place == i3 && i2 != 0 && GET_CODE (XEXP (note, 0)) == REG | |
b1f21e0a | 11492 | && REG_N_REFS (REGNO (XEXP (note, 0)))== 2 |
510dd77e | 11493 | && reg_referenced_p (XEXP (note, 0), PATTERN (i2))) |
b1f21e0a | 11494 | REG_N_REFS (REGNO (XEXP (note, 0))) = 3; |
510dd77e | 11495 | |
230d793d | 11496 | if (place == 0) |
38d8473f RK |
11497 | { |
11498 | for (tem = prev_nonnote_insn (i3); | |
11499 | place == 0 && tem | |
11500 | && (GET_CODE (tem) == INSN || GET_CODE (tem) == CALL_INSN); | |
11501 | tem = prev_nonnote_insn (tem)) | |
11502 | { | |
11503 | /* If the register is being set at TEM, see if that is all | |
11504 | TEM is doing. If so, delete TEM. Otherwise, make this | |
11505 | into a REG_UNUSED note instead. */ | |
11506 | if (reg_set_p (XEXP (note, 0), PATTERN (tem))) | |
11507 | { | |
11508 | rtx set = single_set (tem); | |
e5e809f4 JL |
11509 | rtx inner_dest = 0; |
11510 | ||
11511 | if (set != 0) | |
11512 | for (inner_dest = SET_DEST (set); | |
11513 | GET_CODE (inner_dest) == STRICT_LOW_PART | |
11514 | || GET_CODE (inner_dest) == SUBREG | |
11515 | || GET_CODE (inner_dest) == ZERO_EXTRACT; | |
11516 | inner_dest = XEXP (inner_dest, 0)) | |
11517 | ; | |
38d8473f RK |
11518 | |
11519 | /* Verify that it was the set, and not a clobber that | |
11520 | modified the register. */ | |
11521 | ||
11522 | if (set != 0 && ! side_effects_p (SET_SRC (set)) | |
e5e809f4 | 11523 | && rtx_equal_p (XEXP (note, 0), inner_dest)) |
38d8473f RK |
11524 | { |
11525 | /* Move the notes and links of TEM elsewhere. | |
11526 | This might delete other dead insns recursively. | |
11527 | First set the pattern to something that won't use | |
11528 | any register. */ | |
11529 | ||
11530 | PATTERN (tem) = pc_rtx; | |
11531 | ||
11532 | distribute_notes (REG_NOTES (tem), tem, tem, | |
11533 | NULL_RTX, NULL_RTX, NULL_RTX); | |
11534 | distribute_links (LOG_LINKS (tem)); | |
11535 | ||
11536 | PUT_CODE (tem, NOTE); | |
11537 | NOTE_LINE_NUMBER (tem) = NOTE_INSN_DELETED; | |
11538 | NOTE_SOURCE_FILE (tem) = 0; | |
11539 | } | |
e5e809f4 JL |
11540 | /* If the register is both set and used here, put the |
11541 | REG_DEAD note here, but place a REG_UNUSED note | |
11542 | here too unless there already is one. */ | |
11543 | else if (reg_referenced_p (XEXP (note, 0), | |
11544 | PATTERN (tem))) | |
11545 | { | |
11546 | place = tem; | |
11547 | ||
11548 | if (! find_regno_note (tem, REG_UNUSED, | |
11549 | REGNO (XEXP (note, 0)))) | |
11550 | REG_NOTES (tem) | |
9e6a5703 JC |
11551 | = gen_rtx_EXPR_LIST (REG_UNUSED, |
11552 | XEXP (note, 0), | |
11553 | REG_NOTES (tem)); | |
e5e809f4 | 11554 | } |
38d8473f RK |
11555 | else |
11556 | { | |
11557 | PUT_REG_NOTE_KIND (note, REG_UNUSED); | |
11558 | ||
11559 | /* If there isn't already a REG_UNUSED note, put one | |
11560 | here. */ | |
11561 | if (! find_regno_note (tem, REG_UNUSED, | |
11562 | REGNO (XEXP (note, 0)))) | |
11563 | place = tem; | |
11564 | break; | |
230d793d RS |
11565 | } |
11566 | } | |
13018fad RE |
11567 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem)) |
11568 | || (GET_CODE (tem) == CALL_INSN | |
11569 | && find_reg_fusage (tem, USE, XEXP (note, 0)))) | |
230d793d RS |
11570 | { |
11571 | place = tem; | |
932d1119 RK |
11572 | |
11573 | /* If we are doing a 3->2 combination, and we have a | |
11574 | register which formerly died in i3 and was not used | |
11575 | by i2, which now no longer dies in i3 and is used in | |
11576 | i2 but does not die in i2, and place is between i2 | |
11577 | and i3, then we may need to move a link from place to | |
11578 | i2. */ | |
a8908849 RK |
11579 | if (i2 && INSN_UID (place) <= max_uid_cuid |
11580 | && INSN_CUID (place) > INSN_CUID (i2) | |
932d1119 RK |
11581 | && from_insn && INSN_CUID (from_insn) > INSN_CUID (i2) |
11582 | && reg_referenced_p (XEXP (note, 0), PATTERN (i2))) | |
11583 | { | |
11584 | rtx links = LOG_LINKS (place); | |
11585 | LOG_LINKS (place) = 0; | |
11586 | distribute_links (links); | |
11587 | } | |
230d793d RS |
11588 | break; |
11589 | } | |
38d8473f RK |
11590 | } |
11591 | ||
11592 | /* If we haven't found an insn for the death note and it | |
11593 | is still a REG_DEAD note, but we have hit a CODE_LABEL, | |
11594 | insert a USE insn for the register at that label and | |
11595 | put the death node there. This prevents problems with | |
11596 | call-state tracking in caller-save.c. */ | |
11597 | if (REG_NOTE_KIND (note) == REG_DEAD && place == 0 && tem != 0) | |
e2cce0cf RK |
11598 | { |
11599 | place | |
38a448ca | 11600 | = emit_insn_after (gen_rtx_USE (VOIDmode, XEXP (note, 0)), |
e2cce0cf RK |
11601 | tem); |
11602 | ||
11603 | /* If this insn was emitted between blocks, then update | |
11604 | basic_block_head of the current block to include it. */ | |
11605 | if (basic_block_end[this_basic_block - 1] == tem) | |
11606 | basic_block_head[this_basic_block] = place; | |
11607 | } | |
38d8473f | 11608 | } |
230d793d RS |
11609 | |
11610 | /* If the register is set or already dead at PLACE, we needn't do | |
e5e809f4 JL |
11611 | anything with this note if it is still a REG_DEAD note. |
11612 | We can here if it is set at all, not if is it totally replace, | |
11613 | which is what `dead_or_set_p' checks, so also check for it being | |
11614 | set partially. */ | |
11615 | ||
230d793d | 11616 | |
230d793d RS |
11617 | if (place && REG_NOTE_KIND (note) == REG_DEAD) |
11618 | { | |
11619 | int regno = REGNO (XEXP (note, 0)); | |
11620 | ||
11621 | if (dead_or_set_p (place, XEXP (note, 0)) | |
11622 | || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place))) | |
11623 | { | |
11624 | /* Unless the register previously died in PLACE, clear | |
11625 | reg_last_death. [I no longer understand why this is | |
11626 | being done.] */ | |
11627 | if (reg_last_death[regno] != place) | |
11628 | reg_last_death[regno] = 0; | |
11629 | place = 0; | |
11630 | } | |
11631 | else | |
11632 | reg_last_death[regno] = place; | |
11633 | ||
11634 | /* If this is a death note for a hard reg that is occupying | |
11635 | multiple registers, ensure that we are still using all | |
11636 | parts of the object. If we find a piece of the object | |
11637 | that is unused, we must add a USE for that piece before | |
11638 | PLACE and put the appropriate REG_DEAD note on it. | |
11639 | ||
11640 | An alternative would be to put a REG_UNUSED for the pieces | |
11641 | on the insn that set the register, but that can't be done if | |
11642 | it is not in the same block. It is simpler, though less | |
11643 | efficient, to add the USE insns. */ | |
11644 | ||
11645 | if (place && regno < FIRST_PSEUDO_REGISTER | |
11646 | && HARD_REGNO_NREGS (regno, GET_MODE (XEXP (note, 0))) > 1) | |
11647 | { | |
11648 | int endregno | |
11649 | = regno + HARD_REGNO_NREGS (regno, | |
11650 | GET_MODE (XEXP (note, 0))); | |
11651 | int all_used = 1; | |
11652 | int i; | |
11653 | ||
11654 | for (i = regno; i < endregno; i++) | |
9fd5bb62 JW |
11655 | if (! refers_to_regno_p (i, i + 1, PATTERN (place), 0) |
11656 | && ! find_regno_fusage (place, USE, i)) | |
230d793d | 11657 | { |
38a448ca | 11658 | rtx piece = gen_rtx_REG (reg_raw_mode[i], i); |
28f6d3af RK |
11659 | rtx p; |
11660 | ||
11661 | /* See if we already placed a USE note for this | |
11662 | register in front of PLACE. */ | |
11663 | for (p = place; | |
11664 | GET_CODE (PREV_INSN (p)) == INSN | |
11665 | && GET_CODE (PATTERN (PREV_INSN (p))) == USE; | |
11666 | p = PREV_INSN (p)) | |
11667 | if (rtx_equal_p (piece, | |
11668 | XEXP (PATTERN (PREV_INSN (p)), 0))) | |
11669 | { | |
11670 | p = 0; | |
11671 | break; | |
11672 | } | |
11673 | ||
11674 | if (p) | |
11675 | { | |
11676 | rtx use_insn | |
38a448ca RH |
11677 | = emit_insn_before (gen_rtx_USE (VOIDmode, |
11678 | piece), | |
28f6d3af RK |
11679 | p); |
11680 | REG_NOTES (use_insn) | |
38a448ca RH |
11681 | = gen_rtx_EXPR_LIST (REG_DEAD, piece, |
11682 | REG_NOTES (use_insn)); | |
28f6d3af | 11683 | } |
230d793d | 11684 | |
5089e22e | 11685 | all_used = 0; |
230d793d RS |
11686 | } |
11687 | ||
a394b17b JW |
11688 | /* Check for the case where the register dying partially |
11689 | overlaps the register set by this insn. */ | |
11690 | if (all_used) | |
11691 | for (i = regno; i < endregno; i++) | |
11692 | if (dead_or_set_regno_p (place, i)) | |
11693 | { | |
11694 | all_used = 0; | |
11695 | break; | |
11696 | } | |
11697 | ||
230d793d RS |
11698 | if (! all_used) |
11699 | { | |
11700 | /* Put only REG_DEAD notes for pieces that are | |
11701 | still used and that are not already dead or set. */ | |
11702 | ||
11703 | for (i = regno; i < endregno; i++) | |
11704 | { | |
38a448ca | 11705 | rtx piece = gen_rtx_REG (reg_raw_mode[i], i); |
230d793d | 11706 | |
17cbf358 JW |
11707 | if ((reg_referenced_p (piece, PATTERN (place)) |
11708 | || (GET_CODE (place) == CALL_INSN | |
11709 | && find_reg_fusage (place, USE, piece))) | |
230d793d RS |
11710 | && ! dead_or_set_p (place, piece) |
11711 | && ! reg_bitfield_target_p (piece, | |
11712 | PATTERN (place))) | |
38a448ca RH |
11713 | REG_NOTES (place) |
11714 | = gen_rtx_EXPR_LIST (REG_DEAD, | |
11715 | piece, REG_NOTES (place)); | |
230d793d RS |
11716 | } |
11717 | ||
11718 | place = 0; | |
11719 | } | |
11720 | } | |
11721 | } | |
11722 | break; | |
11723 | ||
11724 | default: | |
11725 | /* Any other notes should not be present at this point in the | |
11726 | compilation. */ | |
11727 | abort (); | |
11728 | } | |
11729 | ||
11730 | if (place) | |
11731 | { | |
11732 | XEXP (note, 1) = REG_NOTES (place); | |
11733 | REG_NOTES (place) = note; | |
11734 | } | |
1a26b032 RK |
11735 | else if ((REG_NOTE_KIND (note) == REG_DEAD |
11736 | || REG_NOTE_KIND (note) == REG_UNUSED) | |
11737 | && GET_CODE (XEXP (note, 0)) == REG) | |
b1f21e0a | 11738 | REG_N_DEATHS (REGNO (XEXP (note, 0)))--; |
230d793d RS |
11739 | |
11740 | if (place2) | |
1a26b032 RK |
11741 | { |
11742 | if ((REG_NOTE_KIND (note) == REG_DEAD | |
11743 | || REG_NOTE_KIND (note) == REG_UNUSED) | |
11744 | && GET_CODE (XEXP (note, 0)) == REG) | |
b1f21e0a | 11745 | REG_N_DEATHS (REGNO (XEXP (note, 0)))++; |
1a26b032 | 11746 | |
38a448ca RH |
11747 | REG_NOTES (place2) = gen_rtx_fmt_ee (GET_CODE (note), |
11748 | REG_NOTE_KIND (note), | |
11749 | XEXP (note, 0), | |
11750 | REG_NOTES (place2)); | |
1a26b032 | 11751 | } |
230d793d RS |
11752 | } |
11753 | } | |
11754 | \f | |
11755 | /* Similarly to above, distribute the LOG_LINKS that used to be present on | |
5089e22e RS |
11756 | I3, I2, and I1 to new locations. This is also called in one case to |
11757 | add a link pointing at I3 when I3's destination is changed. */ | |
230d793d RS |
11758 | |
11759 | static void | |
11760 | distribute_links (links) | |
11761 | rtx links; | |
11762 | { | |
11763 | rtx link, next_link; | |
11764 | ||
11765 | for (link = links; link; link = next_link) | |
11766 | { | |
11767 | rtx place = 0; | |
11768 | rtx insn; | |
11769 | rtx set, reg; | |
11770 | ||
11771 | next_link = XEXP (link, 1); | |
11772 | ||
11773 | /* If the insn that this link points to is a NOTE or isn't a single | |
11774 | set, ignore it. In the latter case, it isn't clear what we | |
11775 | can do other than ignore the link, since we can't tell which | |
11776 | register it was for. Such links wouldn't be used by combine | |
11777 | anyway. | |
11778 | ||
11779 | It is not possible for the destination of the target of the link to | |
11780 | have been changed by combine. The only potential of this is if we | |
11781 | replace I3, I2, and I1 by I3 and I2. But in that case the | |
11782 | destination of I2 also remains unchanged. */ | |
11783 | ||
11784 | if (GET_CODE (XEXP (link, 0)) == NOTE | |
11785 | || (set = single_set (XEXP (link, 0))) == 0) | |
11786 | continue; | |
11787 | ||
11788 | reg = SET_DEST (set); | |
11789 | while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT | |
11790 | || GET_CODE (reg) == SIGN_EXTRACT | |
11791 | || GET_CODE (reg) == STRICT_LOW_PART) | |
11792 | reg = XEXP (reg, 0); | |
11793 | ||
11794 | /* A LOG_LINK is defined as being placed on the first insn that uses | |
11795 | a register and points to the insn that sets the register. Start | |
11796 | searching at the next insn after the target of the link and stop | |
11797 | when we reach a set of the register or the end of the basic block. | |
11798 | ||
11799 | Note that this correctly handles the link that used to point from | |
5089e22e | 11800 | I3 to I2. Also note that not much searching is typically done here |
230d793d RS |
11801 | since most links don't point very far away. */ |
11802 | ||
11803 | for (insn = NEXT_INSN (XEXP (link, 0)); | |
0d4d42c3 RK |
11804 | (insn && (this_basic_block == n_basic_blocks - 1 |
11805 | || basic_block_head[this_basic_block + 1] != insn)); | |
230d793d RS |
11806 | insn = NEXT_INSN (insn)) |
11807 | if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
11808 | && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
11809 | { | |
11810 | if (reg_referenced_p (reg, PATTERN (insn))) | |
11811 | place = insn; | |
11812 | break; | |
11813 | } | |
6e2d1486 RK |
11814 | else if (GET_CODE (insn) == CALL_INSN |
11815 | && find_reg_fusage (insn, USE, reg)) | |
11816 | { | |
11817 | place = insn; | |
11818 | break; | |
11819 | } | |
230d793d RS |
11820 | |
11821 | /* If we found a place to put the link, place it there unless there | |
11822 | is already a link to the same insn as LINK at that point. */ | |
11823 | ||
11824 | if (place) | |
11825 | { | |
11826 | rtx link2; | |
11827 | ||
11828 | for (link2 = LOG_LINKS (place); link2; link2 = XEXP (link2, 1)) | |
11829 | if (XEXP (link2, 0) == XEXP (link, 0)) | |
11830 | break; | |
11831 | ||
11832 | if (link2 == 0) | |
11833 | { | |
11834 | XEXP (link, 1) = LOG_LINKS (place); | |
11835 | LOG_LINKS (place) = link; | |
abe6e52f RK |
11836 | |
11837 | /* Set added_links_insn to the earliest insn we added a | |
11838 | link to. */ | |
11839 | if (added_links_insn == 0 | |
11840 | || INSN_CUID (added_links_insn) > INSN_CUID (place)) | |
11841 | added_links_insn = place; | |
230d793d RS |
11842 | } |
11843 | } | |
11844 | } | |
11845 | } | |
11846 | \f | |
1427d6d2 RK |
11847 | /* Compute INSN_CUID for INSN, which is an insn made by combine. */ |
11848 | ||
11849 | static int | |
11850 | insn_cuid (insn) | |
11851 | rtx insn; | |
11852 | { | |
11853 | while (insn != 0 && INSN_UID (insn) > max_uid_cuid | |
11854 | && GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == USE) | |
11855 | insn = NEXT_INSN (insn); | |
11856 | ||
11857 | if (INSN_UID (insn) > max_uid_cuid) | |
11858 | abort (); | |
11859 | ||
11860 | return INSN_CUID (insn); | |
11861 | } | |
11862 | \f | |
230d793d RS |
11863 | void |
11864 | dump_combine_stats (file) | |
11865 | FILE *file; | |
11866 | { | |
11867 | fprintf | |
11868 | (file, | |
11869 | ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n", | |
11870 | combine_attempts, combine_merges, combine_extras, combine_successes); | |
11871 | } | |
11872 | ||
11873 | void | |
11874 | dump_combine_total_stats (file) | |
11875 | FILE *file; | |
11876 | { | |
11877 | fprintf | |
11878 | (file, | |
11879 | "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n", | |
11880 | total_attempts, total_merges, total_extras, total_successes); | |
11881 | } |