]>
Commit | Line | Data |
---|---|---|
752df20e | 1 | /* Common subexpression elimination for GNU compiler. |
f1717362 | 2 | Copyright (C) 1987-2016 Free Software Foundation, Inc. |
752df20e | 3 | |
f12b58b3 | 4 | This file is part of GCC. |
752df20e | 5 | |
f12b58b3 | 6 | GCC is free software; you can redistribute it and/or modify it under |
7 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 8 | Software Foundation; either version 3, or (at your option) any later |
f12b58b3 | 9 | version. |
752df20e | 10 | |
f12b58b3 | 11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
752df20e | 15 | |
16 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
752df20e | 19 | |
752df20e | 20 | #include "config.h" |
405711de | 21 | #include "system.h" |
805e22b2 | 22 | #include "coretypes.h" |
9ef16211 | 23 | #include "backend.h" |
7c29e30e | 24 | #include "target.h" |
752df20e | 25 | #include "rtl.h" |
7c29e30e | 26 | #include "tree.h" |
27 | #include "cfghooks.h" | |
9ef16211 | 28 | #include "df.h" |
7953c610 | 29 | #include "tm_p.h" |
7c29e30e | 30 | #include "insn-config.h" |
42fe97ed | 31 | #include "regs.h" |
7c29e30e | 32 | #include "emit-rtl.h" |
33 | #include "recog.h" | |
94ea8568 | 34 | #include "cfgrtl.h" |
35 | #include "cfganal.h" | |
36 | #include "cfgcleanup.h" | |
d53441c8 | 37 | #include "alias.h" |
ebd9163c | 38 | #include "toplev.h" |
38ccff25 | 39 | #include "params.h" |
d263732c | 40 | #include "rtlhooks-def.h" |
77fce4cd | 41 | #include "tree-pass.h" |
3072d30e | 42 | #include "dbgcnt.h" |
ff17e9ce | 43 | #include "rtl-iter.h" |
752df20e | 44 | |
5fe18e78 | 45 | #ifndef LOAD_EXTEND_OP |
46 | #define LOAD_EXTEND_OP(M) UNKNOWN | |
47 | #endif | |
48 | ||
752df20e | 49 | /* The basic idea of common subexpression elimination is to go |
50 | through the code, keeping a record of expressions that would | |
51 | have the same value at the current scan point, and replacing | |
52 | expressions encountered with the cheapest equivalent expression. | |
53 | ||
54 | It is too complicated to keep track of the different possibilities | |
c863f0f6 | 55 | when control paths merge in this code; so, at each label, we forget all |
56 | that is known and start fresh. This can be described as processing each | |
57 | extended basic block separately. We have a separate pass to perform | |
58 | global CSE. | |
59 | ||
60 | Note CSE can turn a conditional or computed jump into a nop or | |
61 | an unconditional jump. When this occurs we arrange to run the jump | |
62 | optimizer after CSE to delete the unreachable code. | |
752df20e | 63 | |
64 | We use two data structures to record the equivalent expressions: | |
a7f3b1c7 | 65 | a hash table for most expressions, and a vector of "quantity |
66 | numbers" to record equivalent (pseudo) registers. | |
752df20e | 67 | |
68 | The use of the special data structure for registers is desirable | |
69 | because it is faster. It is possible because registers references | |
70 | contain a fairly small number, the register number, taken from | |
71 | a contiguously allocated series, and two register references are | |
72 | identical if they have the same number. General expressions | |
73 | do not have any such thing, so the only way to retrieve the | |
74 | information recorded on an expression other than a register | |
75 | is to keep it in a hash table. | |
76 | ||
77 | Registers and "quantity numbers": | |
cb10db9d | 78 | |
752df20e | 79 | At the start of each basic block, all of the (hardware and pseudo) |
80 | registers used in the function are given distinct quantity | |
81 | numbers to indicate their contents. During scan, when the code | |
82 | copies one register into another, we copy the quantity number. | |
83 | When a register is loaded in any other way, we allocate a new | |
84 | quantity number to describe the value generated by this operation. | |
96d808c2 | 85 | `REG_QTY (N)' records what quantity register N is currently thought |
752df20e | 86 | of as containing. |
87 | ||
1a5bccce | 88 | All real quantity numbers are greater than or equal to zero. |
96d808c2 | 89 | If register N has not been assigned a quantity, `REG_QTY (N)' will |
1a5bccce | 90 | equal -N - 1, which is always negative. |
752df20e | 91 | |
1a5bccce | 92 | Quantity numbers below zero do not exist and none of the `qty_table' |
93 | entries should be referenced with a negative index. | |
752df20e | 94 | |
95 | We also maintain a bidirectional chain of registers for each | |
a7f3b1c7 | 96 | quantity number. The `qty_table` members `first_reg' and `last_reg', |
97 | and `reg_eqv_table' members `next' and `prev' hold these chains. | |
752df20e | 98 | |
99 | The first register in a chain is the one whose lifespan is least local. | |
100 | Among equals, it is the one that was seen first. | |
101 | We replace any equivalent register with that one. | |
102 | ||
103 | If two registers have the same quantity number, it must be true that | |
a7f3b1c7 | 104 | REG expressions with qty_table `mode' must be in the hash table for both |
752df20e | 105 | registers and must be in the same class. |
106 | ||
107 | The converse is not true. Since hard registers may be referenced in | |
108 | any mode, two REG expressions might be equivalent in the hash table | |
109 | but not have the same quantity number if the quantity number of one | |
110 | of the registers is not the same mode as those expressions. | |
cb10db9d | 111 | |
752df20e | 112 | Constants and quantity numbers |
113 | ||
114 | When a quantity has a known constant value, that value is stored | |
a7f3b1c7 | 115 | in the appropriate qty_table `const_rtx'. This is in addition to |
752df20e | 116 | putting the constant in the hash table as is usual for non-regs. |
117 | ||
f9e15121 | 118 | Whether a reg or a constant is preferred is determined by the configuration |
752df20e | 119 | macro CONST_COSTS and will often depend on the constant value. In any |
120 | event, expressions containing constants can be simplified, by fold_rtx. | |
121 | ||
122 | When a quantity has a known nearly constant value (such as an address | |
a7f3b1c7 | 123 | of a stack slot), that value is stored in the appropriate qty_table |
124 | `const_rtx'. | |
752df20e | 125 | |
126 | Integer constants don't have a machine mode. However, cse | |
127 | determines the intended machine mode from the destination | |
128 | of the instruction that moves the constant. The machine mode | |
129 | is recorded in the hash table along with the actual RTL | |
130 | constant expression so that different modes are kept separate. | |
131 | ||
132 | Other expressions: | |
133 | ||
134 | To record known equivalences among expressions in general | |
135 | we use a hash table called `table'. It has a fixed number of buckets | |
136 | that contain chains of `struct table_elt' elements for expressions. | |
137 | These chains connect the elements whose expressions have the same | |
138 | hash codes. | |
139 | ||
140 | Other chains through the same elements connect the elements which | |
141 | currently have equivalent values. | |
142 | ||
143 | Register references in an expression are canonicalized before hashing | |
a7f3b1c7 | 144 | the expression. This is done using `reg_qty' and qty_table `first_reg'. |
752df20e | 145 | The hash code of a register reference is computed using the quantity |
146 | number, not the register number. | |
147 | ||
148 | When the value of an expression changes, it is necessary to remove from the | |
149 | hash table not just that expression but all expressions whose values | |
150 | could be different as a result. | |
151 | ||
152 | 1. If the value changing is in memory, except in special cases | |
153 | ANYTHING referring to memory could be changed. That is because | |
154 | nobody knows where a pointer does not point. | |
155 | The function `invalidate_memory' removes what is necessary. | |
156 | ||
157 | The special cases are when the address is constant or is | |
158 | a constant plus a fixed register such as the frame pointer | |
159 | or a static chain pointer. When such addresses are stored in, | |
160 | we can tell exactly which other such addresses must be invalidated | |
161 | due to overlap. `invalidate' does this. | |
162 | All expressions that refer to non-constant | |
163 | memory addresses are also invalidated. `invalidate_memory' does this. | |
164 | ||
165 | 2. If the value changing is a register, all expressions | |
166 | containing references to that register, and only those, | |
167 | must be removed. | |
168 | ||
169 | Because searching the entire hash table for expressions that contain | |
170 | a register is very slow, we try to figure out when it isn't necessary. | |
171 | Precisely, this is necessary only when expressions have been | |
172 | entered in the hash table using this register, and then the value has | |
173 | changed, and then another expression wants to be added to refer to | |
174 | the register's new value. This sequence of circumstances is rare | |
175 | within any one basic block. | |
176 | ||
96d808c2 | 177 | `REG_TICK' and `REG_IN_TABLE', accessors for members of |
178 | cse_reg_info, are used to detect this case. REG_TICK (i) is | |
179 | incremented whenever a value is stored in register i. | |
180 | REG_IN_TABLE (i) holds -1 if no references to register i have been | |
181 | entered in the table; otherwise, it contains the value REG_TICK (i) | |
182 | had when the references were entered. If we want to enter a | |
183 | reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and | |
184 | remove old references. Until we want to enter a new entry, the | |
185 | mere fact that the two vectors don't match makes the entries be | |
186 | ignored if anyone tries to match them. | |
752df20e | 187 | |
188 | Registers themselves are entered in the hash table as well as in | |
96d808c2 | 189 | the equivalent-register chains. However, `REG_TICK' and |
190 | `REG_IN_TABLE' do not apply to expressions which are simple | |
752df20e | 191 | register references. These expressions are removed from the table |
192 | immediately when they become invalid, and this can be done even if | |
193 | we do not immediately search for all the expressions that refer to | |
194 | the register. | |
195 | ||
196 | A CLOBBER rtx in an instruction invalidates its operand for further | |
197 | reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK | |
198 | invalidates everything that resides in memory. | |
199 | ||
200 | Related expressions: | |
201 | ||
202 | Constant expressions that differ only by an additive integer | |
203 | are called related. When a constant expression is put in | |
204 | the table, the related expression with no constant term | |
205 | is also entered. These are made to point at each other | |
206 | so that it is possible to find out if there exists any | |
207 | register equivalent to an expression related to a given expression. */ | |
cb10db9d | 208 | |
a7f3b1c7 | 209 | /* Length of qty_table vector. We know in advance we will not need |
210 | a quantity number this big. */ | |
752df20e | 211 | |
212 | static int max_qty; | |
213 | ||
214 | /* Next quantity number to be allocated. | |
215 | This is 1 + the largest number needed so far. */ | |
216 | ||
217 | static int next_qty; | |
218 | ||
a7f3b1c7 | 219 | /* Per-qty information tracking. |
752df20e | 220 | |
a7f3b1c7 | 221 | `first_reg' and `last_reg' track the head and tail of the |
222 | chain of registers which currently contain this quantity. | |
752df20e | 223 | |
a7f3b1c7 | 224 | `mode' contains the machine mode of this quantity. |
752df20e | 225 | |
a7f3b1c7 | 226 | `const_rtx' holds the rtx of the constant value of this |
227 | quantity, if known. A summations of the frame/arg pointer | |
228 | and a constant can also be entered here. When this holds | |
229 | a known value, `const_insn' is the insn which stored the | |
230 | constant value. | |
752df20e | 231 | |
a7f3b1c7 | 232 | `comparison_{code,const,qty}' are used to track when a |
233 | comparison between a quantity and some constant or register has | |
234 | been passed. In such a case, we know the results of the comparison | |
235 | in case we see it again. These members record a comparison that | |
236 | is known to be true. `comparison_code' holds the rtx code of such | |
237 | a comparison, else it is set to UNKNOWN and the other two | |
238 | comparison members are undefined. `comparison_const' holds | |
239 | the constant being compared against, or zero if the comparison | |
240 | is not against a constant. `comparison_qty' holds the quantity | |
241 | being compared against when the result is known. If the comparison | |
242 | is not with a register, `comparison_qty' is -1. */ | |
752df20e | 243 | |
a7f3b1c7 | 244 | struct qty_table_elem |
245 | { | |
246 | rtx const_rtx; | |
47f1d198 | 247 | rtx_insn *const_insn; |
a7f3b1c7 | 248 | rtx comparison_const; |
249 | int comparison_qty; | |
02e7a332 | 250 | unsigned int first_reg, last_reg; |
d8b9732d | 251 | /* The sizes of these fields should match the sizes of the |
252 | code and mode fields of struct rtx_def (see rtl.h). */ | |
253 | ENUM_BITFIELD(rtx_code) comparison_code : 16; | |
254 | ENUM_BITFIELD(machine_mode) mode : 8; | |
a7f3b1c7 | 255 | }; |
752df20e | 256 | |
a7f3b1c7 | 257 | /* The table of all qtys, indexed by qty number. */ |
258 | static struct qty_table_elem *qty_table; | |
752df20e | 259 | |
752df20e | 260 | /* For machines that have a CC0, we do not record its value in the hash |
261 | table since its use is guaranteed to be the insn immediately following | |
262 | its definition and any other insn is presumed to invalidate it. | |
263 | ||
c6ddfc69 | 264 | Instead, we store below the current and last value assigned to CC0. |
265 | If it should happen to be a constant, it is stored in preference | |
266 | to the actual assigned value. In case it is a constant, we store | |
267 | the mode in which the constant should be interpreted. */ | |
752df20e | 268 | |
c6ddfc69 | 269 | static rtx this_insn_cc0, prev_insn_cc0; |
3754d046 | 270 | static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode; |
752df20e | 271 | |
272 | /* Insn being scanned. */ | |
273 | ||
47f1d198 | 274 | static rtx_insn *this_insn; |
f529eb25 | 275 | static bool optimize_this_for_speed_p; |
752df20e | 276 | |
2a384a22 | 277 | /* Index by register number, gives the number of the next (or |
278 | previous) register in the chain of registers sharing the same | |
752df20e | 279 | value. |
280 | ||
281 | Or -1 if this register is at the end of the chain. | |
282 | ||
96d808c2 | 283 | If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */ |
a7f3b1c7 | 284 | |
285 | /* Per-register equivalence chain. */ | |
286 | struct reg_eqv_elem | |
287 | { | |
288 | int next, prev; | |
289 | }; | |
752df20e | 290 | |
a7f3b1c7 | 291 | /* The table of all register equivalence chains. */ |
292 | static struct reg_eqv_elem *reg_eqv_table; | |
752df20e | 293 | |
155b05dc | 294 | struct cse_reg_info |
295 | { | |
3bd20490 | 296 | /* The timestamp at which this register is initialized. */ |
297 | unsigned int timestamp; | |
9c4f3716 | 298 | |
299 | /* The quantity number of the register's current contents. */ | |
300 | int reg_qty; | |
301 | ||
302 | /* The number of times the register has been altered in the current | |
303 | basic block. */ | |
304 | int reg_tick; | |
305 | ||
d1264606 | 306 | /* The REG_TICK value at which rtx's containing this register are |
307 | valid in the hash table. If this does not equal the current | |
308 | reg_tick value, such expressions existing in the hash table are | |
309 | invalid. */ | |
310 | int reg_in_table; | |
126fb012 | 311 | |
312 | /* The SUBREG that was set when REG_TICK was last incremented. Set | |
313 | to -1 if the last store was to the whole register, not a subreg. */ | |
70e488ba | 314 | unsigned int subreg_ticked; |
d1264606 | 315 | }; |
752df20e | 316 | |
3bd20490 | 317 | /* A table of cse_reg_info indexed by register numbers. */ |
f9413025 | 318 | static struct cse_reg_info *cse_reg_info_table; |
ac613367 | 319 | |
3bd20490 | 320 | /* The size of the above table. */ |
321 | static unsigned int cse_reg_info_table_size; | |
9c4f3716 | 322 | |
3bd20490 | 323 | /* The index of the first entry that has not been initialized. */ |
324 | static unsigned int cse_reg_info_table_first_uninitialized; | |
752df20e | 325 | |
3bd20490 | 326 | /* The timestamp at the beginning of the current run of |
be22716f | 327 | cse_extended_basic_block. We increment this variable at the beginning of |
328 | the current run of cse_extended_basic_block. The timestamp field of a | |
3bd20490 | 329 | cse_reg_info entry matches the value of this variable if and only |
330 | if the entry has been initialized during the current run of | |
be22716f | 331 | cse_extended_basic_block. */ |
3bd20490 | 332 | static unsigned int cse_reg_info_timestamp; |
752df20e | 333 | |
cb10db9d | 334 | /* A HARD_REG_SET containing all the hard registers for which there is |
752df20e | 335 | currently a REG expression in the hash table. Note the difference |
336 | from the above variables, which indicate if the REG is mentioned in some | |
337 | expression in the table. */ | |
338 | ||
339 | static HARD_REG_SET hard_regs_in_table; | |
340 | ||
283a6b26 | 341 | /* True if CSE has altered the CFG. */ |
342 | static bool cse_cfg_altered; | |
752df20e | 343 | |
283a6b26 | 344 | /* True if CSE has altered conditional jump insns in such a way |
345 | that jump optimization should be redone. */ | |
346 | static bool cse_jumps_altered; | |
752df20e | 347 | |
283a6b26 | 348 | /* True if we put a LABEL_REF into the hash table for an INSN |
349 | without a REG_LABEL_OPERAND, we have to rerun jump after CSE | |
350 | to put in the note. */ | |
351 | static bool recorded_label_ref; | |
26db0da8 | 352 | |
752df20e | 353 | /* canon_hash stores 1 in do_not_record |
354 | if it notices a reference to CC0, PC, or some other volatile | |
355 | subexpression. */ | |
356 | ||
357 | static int do_not_record; | |
358 | ||
359 | /* canon_hash stores 1 in hash_arg_in_memory | |
360 | if it notices a reference to memory within the expression being hashed. */ | |
361 | ||
362 | static int hash_arg_in_memory; | |
363 | ||
752df20e | 364 | /* The hash table contains buckets which are chains of `struct table_elt's, |
365 | each recording one expression's information. | |
366 | That expression is in the `exp' field. | |
367 | ||
7cfb9bcf | 368 | The canon_exp field contains a canonical (from the point of view of |
369 | alias analysis) version of the `exp' field. | |
370 | ||
752df20e | 371 | Those elements with the same hash code are chained in both directions |
372 | through the `next_same_hash' and `prev_same_hash' fields. | |
373 | ||
374 | Each set of expressions with equivalent values | |
375 | are on a two-way chain through the `next_same_value' | |
376 | and `prev_same_value' fields, and all point with | |
377 | the `first_same_value' field at the first element in | |
378 | that chain. The chain is in order of increasing cost. | |
379 | Each element's cost value is in its `cost' field. | |
380 | ||
381 | The `in_memory' field is nonzero for elements that | |
382 | involve any reference to memory. These elements are removed | |
383 | whenever a write is done to an unidentified location in memory. | |
384 | To be safe, we assume that a memory address is unidentified unless | |
385 | the address is either a symbol constant or a constant plus | |
386 | the frame pointer or argument pointer. | |
387 | ||
752df20e | 388 | The `related_value' field is used to connect related expressions |
389 | (that differ by adding an integer). | |
390 | The related expressions are chained in a circular fashion. | |
391 | `related_value' is zero for expressions for which this | |
392 | chain is not useful. | |
393 | ||
394 | The `cost' field stores the cost of this element's expression. | |
d27eb4b1 | 395 | The `regcost' field stores the value returned by approx_reg_cost for |
396 | this element's expression. | |
752df20e | 397 | |
398 | The `is_const' flag is set if the element is a constant (including | |
399 | a fixed address). | |
400 | ||
401 | The `flag' field is used as a temporary during some search routines. | |
402 | ||
403 | The `mode' field is usually the same as GET_MODE (`exp'), but | |
404 | if `exp' is a CONST_INT and has no machine mode then the `mode' | |
405 | field is the mode it was being used as. Each constant is | |
406 | recorded separately for each mode it is used with. */ | |
407 | ||
752df20e | 408 | struct table_elt |
409 | { | |
410 | rtx exp; | |
7cfb9bcf | 411 | rtx canon_exp; |
752df20e | 412 | struct table_elt *next_same_hash; |
413 | struct table_elt *prev_same_hash; | |
414 | struct table_elt *next_same_value; | |
415 | struct table_elt *prev_same_value; | |
416 | struct table_elt *first_same_value; | |
417 | struct table_elt *related_value; | |
418 | int cost; | |
d27eb4b1 | 419 | int regcost; |
d8b9732d | 420 | /* The size of this field should match the size |
421 | of the mode field of struct rtx_def (see rtl.h). */ | |
422 | ENUM_BITFIELD(machine_mode) mode : 8; | |
752df20e | 423 | char in_memory; |
752df20e | 424 | char is_const; |
425 | char flag; | |
426 | }; | |
427 | ||
752df20e | 428 | /* We don't want a lot of buckets, because we rarely have very many |
429 | things stored in the hash table, and a lot of buckets slows | |
430 | down a lot of loops that happen frequently. */ | |
9c4f3716 | 431 | #define HASH_SHIFT 5 |
432 | #define HASH_SIZE (1 << HASH_SHIFT) | |
433 | #define HASH_MASK (HASH_SIZE - 1) | |
752df20e | 434 | |
435 | /* Compute hash code of X in mode M. Special-case case where X is a pseudo | |
436 | register (hard registers may require `do_not_record' to be set). */ | |
437 | ||
438 | #define HASH(X, M) \ | |
8ad4c111 | 439 | ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \ |
9c4f3716 | 440 | ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \ |
441 | : canon_hash (X, M)) & HASH_MASK) | |
752df20e | 442 | |
78d140c9 | 443 | /* Like HASH, but without side-effects. */ |
444 | #define SAFE_HASH(X, M) \ | |
445 | ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \ | |
446 | ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \ | |
447 | : safe_hash (X, M)) & HASH_MASK) | |
448 | ||
d27eb4b1 | 449 | /* Determine whether register number N is considered a fixed register for the |
450 | purpose of approximating register costs. | |
752df20e | 451 | It is desirable to replace other regs with fixed regs, to reduce need for |
452 | non-fixed hard regs. | |
349858d4 | 453 | A reg wins if it is either the frame pointer or designated as fixed. */ |
752df20e | 454 | #define FIXED_REGNO_P(N) \ |
b69007e1 | 455 | ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \ |
4313a67c | 456 | || fixed_regs[N] || global_regs[N]) |
752df20e | 457 | |
458 | /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed | |
2e16c7bd | 459 | hard registers and pointers into the frame are the cheapest with a cost |
460 | of 0. Next come pseudos with a cost of one and other hard registers with | |
461 | a cost of 2. Aside from these special cases, call `rtx_cost'. */ | |
462 | ||
5bbaf5ca | 463 | #define CHEAP_REGNO(N) \ |
9af5ce0c | 464 | (REGNO_PTR_FRAME_P (N) \ |
5bbaf5ca | 465 | || (HARD_REGISTER_NUM_P (N) \ |
c0191571 | 466 | && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS)) |
752df20e | 467 | |
5ae4887d | 468 | #define COST(X, MODE) \ |
469 | (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1)) | |
470 | #define COST_IN(X, MODE, OUTER, OPNO) \ | |
471 | (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO)) | |
752df20e | 472 | |
d1264606 | 473 | /* Get the number of times this register has been updated in this |
474 | basic block. */ | |
475 | ||
3bd20490 | 476 | #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick) |
d1264606 | 477 | |
478 | /* Get the point at which REG was recorded in the table. */ | |
479 | ||
3bd20490 | 480 | #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table) |
d1264606 | 481 | |
126fb012 | 482 | /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a |
483 | SUBREG). */ | |
484 | ||
3bd20490 | 485 | #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked) |
126fb012 | 486 | |
d1264606 | 487 | /* Get the quantity number for REG. */ |
488 | ||
3bd20490 | 489 | #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty) |
d1264606 | 490 | |
752df20e | 491 | /* Determine if the quantity number for register X represents a valid index |
a7f3b1c7 | 492 | into the qty_table. */ |
752df20e | 493 | |
1a5bccce | 494 | #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0) |
752df20e | 495 | |
01c8e4c9 | 496 | /* Compare table_elt X and Y and return true iff X is cheaper than Y. */ |
497 | ||
498 | #define CHEAPER(X, Y) \ | |
499 | (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0) | |
500 | ||
9c4f3716 | 501 | static struct table_elt *table[HASH_SIZE]; |
752df20e | 502 | |
503 | /* Chain of `struct table_elt's made so far for this function | |
504 | but currently removed from the table. */ | |
505 | ||
506 | static struct table_elt *free_element_chain; | |
507 | ||
752df20e | 508 | /* Set to the cost of a constant pool reference if one was found for a |
509 | symbolic constant. If this was found, it means we should try to | |
510 | convert constants into constant pool entries if they don't fit in | |
511 | the insn. */ | |
512 | ||
513 | static int constant_pool_entries_cost; | |
634d45d7 | 514 | static int constant_pool_entries_regcost; |
752df20e | 515 | |
0b09525f | 516 | /* Trace a patch through the CFG. */ |
517 | ||
518 | struct branch_path | |
519 | { | |
520 | /* The basic block for this path entry. */ | |
521 | basic_block bb; | |
522 | }; | |
523 | ||
be22716f | 524 | /* This data describes a block that will be processed by |
525 | cse_extended_basic_block. */ | |
9def8c3e | 526 | |
155b05dc | 527 | struct cse_basic_block_data |
528 | { | |
9def8c3e | 529 | /* Total number of SETs in block. */ |
530 | int nsets; | |
9def8c3e | 531 | /* Size of current branch path, if any. */ |
532 | int path_size; | |
be22716f | 533 | /* Current path, indicating which basic_blocks will be processed. */ |
0b09525f | 534 | struct branch_path *path; |
9def8c3e | 535 | }; |
536 | ||
3072d30e | 537 | |
538 | /* Pointers to the live in/live out bitmaps for the boundaries of the | |
539 | current EBB. */ | |
540 | static bitmap cse_ebb_live_in, cse_ebb_live_out; | |
541 | ||
be22716f | 542 | /* A simple bitmap to track which basic blocks have been visited |
543 | already as part of an already processed extended basic block. */ | |
544 | static sbitmap cse_visited_basic_blocks; | |
545 | ||
8ec3a57b | 546 | static bool fixed_base_plus_p (rtx x); |
5ae4887d | 547 | static int notreg_cost (rtx, machine_mode, enum rtx_code, int); |
069eea26 | 548 | static int preferable (int, int, int, int); |
8ec3a57b | 549 | static void new_basic_block (void); |
3754d046 | 550 | static void make_new_qty (unsigned int, machine_mode); |
8ec3a57b | 551 | static void make_regs_eqv (unsigned int, unsigned int); |
552 | static void delete_reg_equiv (unsigned int); | |
553 | static int mention_regs (rtx); | |
554 | static int insert_regs (rtx, struct table_elt *, int); | |
555 | static void remove_from_table (struct table_elt *, unsigned); | |
d2c970fe | 556 | static void remove_pseudo_from_table (rtx, unsigned); |
3754d046 | 557 | static struct table_elt *lookup (rtx, unsigned, machine_mode); |
558 | static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode); | |
8ec3a57b | 559 | static rtx lookup_as_function (rtx, enum rtx_code); |
01c8e4c9 | 560 | static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned, |
3754d046 | 561 | machine_mode, int, int); |
8ec3a57b | 562 | static struct table_elt *insert (rtx, struct table_elt *, unsigned, |
3754d046 | 563 | machine_mode); |
8ec3a57b | 564 | static void merge_equiv_classes (struct table_elt *, struct table_elt *); |
3754d046 | 565 | static void invalidate (rtx, machine_mode); |
8ec3a57b | 566 | static void remove_invalid_refs (unsigned int); |
567 | static void remove_invalid_subreg_refs (unsigned int, unsigned int, | |
3754d046 | 568 | machine_mode); |
8ec3a57b | 569 | static void rehash_using_reg (rtx); |
570 | static void invalidate_memory (void); | |
571 | static void invalidate_for_call (void); | |
572 | static rtx use_related_value (rtx, struct table_elt *); | |
78d140c9 | 573 | |
3754d046 | 574 | static inline unsigned canon_hash (rtx, machine_mode); |
575 | static inline unsigned safe_hash (rtx, machine_mode); | |
e1ab7874 | 576 | static inline unsigned hash_rtx_string (const char *); |
78d140c9 | 577 | |
47f1d198 | 578 | static rtx canon_reg (rtx, rtx_insn *); |
8ec3a57b | 579 | static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *, |
3754d046 | 580 | machine_mode *, |
581 | machine_mode *); | |
47f1d198 | 582 | static rtx fold_rtx (rtx, rtx_insn *); |
8ec3a57b | 583 | static rtx equiv_constant (rtx); |
47f1d198 | 584 | static void record_jump_equiv (rtx_insn *, bool); |
3754d046 | 585 | static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx, |
8ec3a57b | 586 | int); |
47f1d198 | 587 | static void cse_insn (rtx_insn *); |
be22716f | 588 | static void cse_prescan_path (struct cse_basic_block_data *); |
47f1d198 | 589 | static void invalidate_from_clobbers (rtx_insn *); |
590 | static void invalidate_from_sets_and_clobbers (rtx_insn *); | |
3072d30e | 591 | static rtx cse_process_notes (rtx, rtx, bool *); |
be22716f | 592 | static void cse_extended_basic_block (struct cse_basic_block_data *); |
8ec3a57b | 593 | extern void dump_class (struct table_elt*); |
3bd20490 | 594 | static void get_cse_reg_info_1 (unsigned int regno); |
595 | static struct cse_reg_info * get_cse_reg_info (unsigned int regno); | |
8ec3a57b | 596 | |
597 | static void flush_hash_table (void); | |
47f1d198 | 598 | static bool insn_live_p (rtx_insn *, int *); |
599 | static bool set_live_p (rtx, rtx_insn *, int *); | |
47f1d198 | 600 | static void cse_change_cc_mode_insn (rtx_insn *, rtx); |
601 | static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx); | |
3754d046 | 602 | static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx, |
650d2134 | 603 | bool); |
752df20e | 604 | \f |
d263732c | 605 | |
606 | #undef RTL_HOOKS_GEN_LOWPART | |
607 | #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible | |
608 | ||
609 | static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER; | |
610 | \f | |
ea1760a3 | 611 | /* Nonzero if X has the form (PLUS frame-pointer integer). */ |
805e22b2 | 612 | |
613 | static bool | |
8ec3a57b | 614 | fixed_base_plus_p (rtx x) |
805e22b2 | 615 | { |
616 | switch (GET_CODE (x)) | |
617 | { | |
618 | case REG: | |
619 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx) | |
620 | return true; | |
621 | if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
622 | return true; | |
805e22b2 | 623 | return false; |
624 | ||
625 | case PLUS: | |
971ba038 | 626 | if (!CONST_INT_P (XEXP (x, 1))) |
805e22b2 | 627 | return false; |
628 | return fixed_base_plus_p (XEXP (x, 0)); | |
629 | ||
805e22b2 | 630 | default: |
631 | return false; | |
632 | } | |
633 | } | |
634 | ||
59241190 | 635 | /* Dump the expressions in the equivalence class indicated by CLASSP. |
636 | This function is used only for debugging. */ | |
d2bb3f9d | 637 | DEBUG_FUNCTION void |
8ec3a57b | 638 | dump_class (struct table_elt *classp) |
59241190 | 639 | { |
640 | struct table_elt *elt; | |
641 | ||
642 | fprintf (stderr, "Equivalence chain for "); | |
643 | print_rtl (stderr, classp->exp); | |
644 | fprintf (stderr, ": \n"); | |
cb10db9d | 645 | |
59241190 | 646 | for (elt = classp->first_same_value; elt; elt = elt->next_same_value) |
647 | { | |
648 | print_rtl (stderr, elt->exp); | |
649 | fprintf (stderr, "\n"); | |
650 | } | |
651 | } | |
652 | ||
ff17e9ce | 653 | /* Return an estimate of the cost of the registers used in an rtx. |
654 | This is mostly the number of different REG expressions in the rtx; | |
655 | however for some exceptions like fixed registers we use a cost of | |
656 | 0. If any other hard register reference occurs, return MAX_COST. */ | |
37b8a8d6 | 657 | |
d27eb4b1 | 658 | static int |
ff17e9ce | 659 | approx_reg_cost (const_rtx x) |
d27eb4b1 | 660 | { |
ff17e9ce | 661 | int cost = 0; |
662 | subrtx_iterator::array_type array; | |
663 | FOR_EACH_SUBRTX (iter, array, x, NONCONST) | |
88bc3f54 | 664 | { |
ff17e9ce | 665 | const_rtx x = *iter; |
666 | if (REG_P (x)) | |
88bc3f54 | 667 | { |
ff17e9ce | 668 | unsigned int regno = REGNO (x); |
669 | if (!CHEAP_REGNO (regno)) | |
88bc3f54 | 670 | { |
ff17e9ce | 671 | if (regno < FIRST_PSEUDO_REGISTER) |
672 | { | |
673 | if (targetm.small_register_classes_for_mode_p (GET_MODE (x))) | |
674 | return MAX_COST; | |
675 | cost += 2; | |
676 | } | |
677 | else | |
678 | cost += 1; | |
88bc3f54 | 679 | } |
88bc3f54 | 680 | } |
681 | } | |
88bc3f54 | 682 | return cost; |
d27eb4b1 | 683 | } |
684 | ||
685 | /* Return a negative value if an rtx A, whose costs are given by COST_A | |
686 | and REGCOST_A, is more desirable than an rtx B. | |
687 | Return a positive value if A is less desirable, or 0 if the two are | |
688 | equally good. */ | |
689 | static int | |
069eea26 | 690 | preferable (int cost_a, int regcost_a, int cost_b, int regcost_b) |
d27eb4b1 | 691 | { |
e61c498c | 692 | /* First, get rid of cases involving expressions that are entirely |
589ff9e7 | 693 | unwanted. */ |
694 | if (cost_a != cost_b) | |
695 | { | |
696 | if (cost_a == MAX_COST) | |
697 | return 1; | |
698 | if (cost_b == MAX_COST) | |
699 | return -1; | |
700 | } | |
701 | ||
702 | /* Avoid extending lifetimes of hardregs. */ | |
703 | if (regcost_a != regcost_b) | |
704 | { | |
705 | if (regcost_a == MAX_COST) | |
706 | return 1; | |
707 | if (regcost_b == MAX_COST) | |
708 | return -1; | |
709 | } | |
710 | ||
711 | /* Normal operation costs take precedence. */ | |
d27eb4b1 | 712 | if (cost_a != cost_b) |
713 | return cost_a - cost_b; | |
589ff9e7 | 714 | /* Only if these are identical consider effects on register pressure. */ |
d27eb4b1 | 715 | if (regcost_a != regcost_b) |
716 | return regcost_a - regcost_b; | |
717 | return 0; | |
718 | } | |
719 | ||
de164820 | 720 | /* Internal function, to compute cost when X is not a register; called |
721 | from COST macro to keep it simple. */ | |
722 | ||
723 | static int | |
5ae4887d | 724 | notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno) |
de164820 | 725 | { |
726 | return ((GET_CODE (x) == SUBREG | |
8ad4c111 | 727 | && REG_P (SUBREG_REG (x)) |
5ae4887d | 728 | && GET_MODE_CLASS (mode) == MODE_INT |
de164820 | 729 | && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT |
5ae4887d | 730 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))) |
de164820 | 731 | && subreg_lowpart_p (x) |
5ae4887d | 732 | && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (SUBREG_REG (x)))) |
d27eb4b1 | 733 | ? 0 |
5ae4887d | 734 | : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2); |
de164820 | 735 | } |
736 | ||
cf495191 | 737 | \f |
3bd20490 | 738 | /* Initialize CSE_REG_INFO_TABLE. */ |
9c4f3716 | 739 | |
3bd20490 | 740 | static void |
741 | init_cse_reg_info (unsigned int nregs) | |
742 | { | |
743 | /* Do we need to grow the table? */ | |
744 | if (nregs > cse_reg_info_table_size) | |
d1264606 | 745 | { |
3bd20490 | 746 | unsigned int new_size; |
747 | ||
748 | if (cse_reg_info_table_size < 2048) | |
d1264606 | 749 | { |
3bd20490 | 750 | /* Compute a new size that is a power of 2 and no smaller |
751 | than the large of NREGS and 64. */ | |
752 | new_size = (cse_reg_info_table_size | |
753 | ? cse_reg_info_table_size : 64); | |
754 | ||
755 | while (new_size < nregs) | |
756 | new_size *= 2; | |
d1264606 | 757 | } |
758 | else | |
926f1f1f | 759 | { |
3bd20490 | 760 | /* If we need a big table, allocate just enough to hold |
761 | NREGS registers. */ | |
762 | new_size = nregs; | |
926f1f1f | 763 | } |
9c4f3716 | 764 | |
3bd20490 | 765 | /* Reallocate the table with NEW_SIZE entries. */ |
dd045aee | 766 | free (cse_reg_info_table); |
4c36ffe6 | 767 | cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size); |
3bd20490 | 768 | cse_reg_info_table_size = new_size; |
bee2651c | 769 | cse_reg_info_table_first_uninitialized = 0; |
3bd20490 | 770 | } |
771 | ||
772 | /* Do we have all of the first NREGS entries initialized? */ | |
773 | if (cse_reg_info_table_first_uninitialized < nregs) | |
774 | { | |
775 | unsigned int old_timestamp = cse_reg_info_timestamp - 1; | |
776 | unsigned int i; | |
777 | ||
778 | /* Put the old timestamp on newly allocated entries so that they | |
779 | will all be considered out of date. We do not touch those | |
780 | entries beyond the first NREGS entries to be nice to the | |
781 | virtual memory. */ | |
782 | for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++) | |
783 | cse_reg_info_table[i].timestamp = old_timestamp; | |
d1264606 | 784 | |
3bd20490 | 785 | cse_reg_info_table_first_uninitialized = nregs; |
d1264606 | 786 | } |
3bd20490 | 787 | } |
788 | ||
b5ee2efd | 789 | /* Given REGNO, initialize the cse_reg_info entry for REGNO. */ |
3bd20490 | 790 | |
791 | static void | |
792 | get_cse_reg_info_1 (unsigned int regno) | |
793 | { | |
794 | /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this | |
795 | entry will be considered to have been initialized. */ | |
796 | cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp; | |
797 | ||
798 | /* Initialize the rest of the entry. */ | |
799 | cse_reg_info_table[regno].reg_tick = 1; | |
800 | cse_reg_info_table[regno].reg_in_table = -1; | |
801 | cse_reg_info_table[regno].subreg_ticked = -1; | |
802 | cse_reg_info_table[regno].reg_qty = -regno - 1; | |
803 | } | |
804 | ||
805 | /* Find a cse_reg_info entry for REGNO. */ | |
d1264606 | 806 | |
3bd20490 | 807 | static inline struct cse_reg_info * |
808 | get_cse_reg_info (unsigned int regno) | |
809 | { | |
810 | struct cse_reg_info *p = &cse_reg_info_table[regno]; | |
811 | ||
e02a4f0d | 812 | /* If this entry has not been initialized, go ahead and initialize |
813 | it. */ | |
3bd20490 | 814 | if (p->timestamp != cse_reg_info_timestamp) |
815 | get_cse_reg_info_1 (regno); | |
d1264606 | 816 | |
9c4f3716 | 817 | return p; |
d1264606 | 818 | } |
819 | ||
752df20e | 820 | /* Clear the hash table and initialize each register with its own quantity, |
821 | for a new basic block. */ | |
822 | ||
823 | static void | |
8ec3a57b | 824 | new_basic_block (void) |
752df20e | 825 | { |
19cb6b50 | 826 | int i; |
752df20e | 827 | |
1a5bccce | 828 | next_qty = 0; |
752df20e | 829 | |
b5ee2efd | 830 | /* Invalidate cse_reg_info_table. */ |
3bd20490 | 831 | cse_reg_info_timestamp++; |
752df20e | 832 | |
3bd20490 | 833 | /* Clear out hash table state for this pass. */ |
752df20e | 834 | CLEAR_HARD_REG_SET (hard_regs_in_table); |
835 | ||
836 | /* The per-quantity values used to be initialized here, but it is | |
837 | much faster to initialize each as it is made in `make_new_qty'. */ | |
838 | ||
9c4f3716 | 839 | for (i = 0; i < HASH_SIZE; i++) |
752df20e | 840 | { |
9c4f3716 | 841 | struct table_elt *first; |
842 | ||
843 | first = table[i]; | |
844 | if (first != NULL) | |
752df20e | 845 | { |
9c4f3716 | 846 | struct table_elt *last = first; |
847 | ||
848 | table[i] = NULL; | |
849 | ||
850 | while (last->next_same_hash != NULL) | |
851 | last = last->next_same_hash; | |
852 | ||
853 | /* Now relink this hash entire chain into | |
854 | the free element list. */ | |
855 | ||
856 | last->next_same_hash = free_element_chain; | |
857 | free_element_chain = first; | |
752df20e | 858 | } |
859 | } | |
860 | ||
752df20e | 861 | prev_insn_cc0 = 0; |
752df20e | 862 | } |
863 | ||
a7f3b1c7 | 864 | /* Say that register REG contains a quantity in mode MODE not in any |
865 | register before and initialize that quantity. */ | |
752df20e | 866 | |
867 | static void | |
3754d046 | 868 | make_new_qty (unsigned int reg, machine_mode mode) |
752df20e | 869 | { |
19cb6b50 | 870 | int q; |
871 | struct qty_table_elem *ent; | |
872 | struct reg_eqv_elem *eqv; | |
752df20e | 873 | |
cc636d56 | 874 | gcc_assert (next_qty < max_qty); |
752df20e | 875 | |
d1264606 | 876 | q = REG_QTY (reg) = next_qty++; |
a7f3b1c7 | 877 | ent = &qty_table[q]; |
878 | ent->first_reg = reg; | |
879 | ent->last_reg = reg; | |
880 | ent->mode = mode; | |
47f1d198 | 881 | ent->const_rtx = ent->const_insn = NULL; |
a7f3b1c7 | 882 | ent->comparison_code = UNKNOWN; |
883 | ||
884 | eqv = ®_eqv_table[reg]; | |
885 | eqv->next = eqv->prev = -1; | |
752df20e | 886 | } |
887 | ||
888 | /* Make reg NEW equivalent to reg OLD. | |
889 | OLD is not changing; NEW is. */ | |
890 | ||
891 | static void | |
d328ebdf | 892 | make_regs_eqv (unsigned int new_reg, unsigned int old_reg) |
752df20e | 893 | { |
02e7a332 | 894 | unsigned int lastr, firstr; |
d328ebdf | 895 | int q = REG_QTY (old_reg); |
02e7a332 | 896 | struct qty_table_elem *ent; |
a7f3b1c7 | 897 | |
898 | ent = &qty_table[q]; | |
752df20e | 899 | |
900 | /* Nothing should become eqv until it has a "non-invalid" qty number. */ | |
d328ebdf | 901 | gcc_assert (REGNO_QTY_VALID_P (old_reg)); |
752df20e | 902 | |
d328ebdf | 903 | REG_QTY (new_reg) = q; |
a7f3b1c7 | 904 | firstr = ent->first_reg; |
905 | lastr = ent->last_reg; | |
752df20e | 906 | |
907 | /* Prefer fixed hard registers to anything. Prefer pseudo regs to other | |
908 | hard regs. Among pseudos, if NEW will live longer than any other reg | |
909 | of the same qty, and that is beyond the current basic block, | |
910 | make it the new canonical replacement for this qty. */ | |
911 | if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr)) | |
912 | /* Certain fixed registers might be of the class NO_REGS. This means | |
913 | that not only can they not be allocated by the compiler, but | |
5202ecf2 | 914 | they cannot be used in substitutions or canonicalizations |
752df20e | 915 | either. */ |
d328ebdf | 916 | && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS) |
917 | && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg)) | |
918 | || (new_reg >= FIRST_PSEUDO_REGISTER | |
752df20e | 919 | && (firstr < FIRST_PSEUDO_REGISTER |
d328ebdf | 920 | || (bitmap_bit_p (cse_ebb_live_out, new_reg) |
3072d30e | 921 | && !bitmap_bit_p (cse_ebb_live_out, firstr)) |
d328ebdf | 922 | || (bitmap_bit_p (cse_ebb_live_in, new_reg) |
3072d30e | 923 | && !bitmap_bit_p (cse_ebb_live_in, firstr)))))) |
752df20e | 924 | { |
d328ebdf | 925 | reg_eqv_table[firstr].prev = new_reg; |
926 | reg_eqv_table[new_reg].next = firstr; | |
927 | reg_eqv_table[new_reg].prev = -1; | |
928 | ent->first_reg = new_reg; | |
752df20e | 929 | } |
930 | else | |
931 | { | |
932 | /* If NEW is a hard reg (known to be non-fixed), insert at end. | |
933 | Otherwise, insert before any non-fixed hard regs that are at the | |
934 | end. Registers of class NO_REGS cannot be used as an | |
935 | equivalent for anything. */ | |
a7f3b1c7 | 936 | while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0 |
752df20e | 937 | && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr)) |
d328ebdf | 938 | && new_reg >= FIRST_PSEUDO_REGISTER) |
a7f3b1c7 | 939 | lastr = reg_eqv_table[lastr].prev; |
d328ebdf | 940 | reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next; |
a7f3b1c7 | 941 | if (reg_eqv_table[lastr].next >= 0) |
d328ebdf | 942 | reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg; |
752df20e | 943 | else |
d328ebdf | 944 | qty_table[q].last_reg = new_reg; |
945 | reg_eqv_table[lastr].next = new_reg; | |
946 | reg_eqv_table[new_reg].prev = lastr; | |
752df20e | 947 | } |
948 | } | |
949 | ||
950 | /* Remove REG from its equivalence class. */ | |
951 | ||
952 | static void | |
8ec3a57b | 953 | delete_reg_equiv (unsigned int reg) |
752df20e | 954 | { |
19cb6b50 | 955 | struct qty_table_elem *ent; |
956 | int q = REG_QTY (reg); | |
957 | int p, n; | |
752df20e | 958 | |
7046c09e | 959 | /* If invalid, do nothing. */ |
1a5bccce | 960 | if (! REGNO_QTY_VALID_P (reg)) |
752df20e | 961 | return; |
962 | ||
a7f3b1c7 | 963 | ent = &qty_table[q]; |
964 | ||
965 | p = reg_eqv_table[reg].prev; | |
966 | n = reg_eqv_table[reg].next; | |
7046c09e | 967 | |
752df20e | 968 | if (n != -1) |
a7f3b1c7 | 969 | reg_eqv_table[n].prev = p; |
752df20e | 970 | else |
a7f3b1c7 | 971 | ent->last_reg = p; |
752df20e | 972 | if (p != -1) |
a7f3b1c7 | 973 | reg_eqv_table[p].next = n; |
752df20e | 974 | else |
a7f3b1c7 | 975 | ent->first_reg = n; |
752df20e | 976 | |
1a5bccce | 977 | REG_QTY (reg) = -reg - 1; |
752df20e | 978 | } |
979 | ||
980 | /* Remove any invalid expressions from the hash table | |
981 | that refer to any of the registers contained in expression X. | |
982 | ||
983 | Make sure that newly inserted references to those registers | |
984 | as subexpressions will be considered valid. | |
985 | ||
986 | mention_regs is not called when a register itself | |
987 | is being stored in the table. | |
988 | ||
989 | Return 1 if we have done something that may have changed the hash code | |
990 | of X. */ | |
991 | ||
992 | static int | |
8ec3a57b | 993 | mention_regs (rtx x) |
752df20e | 994 | { |
19cb6b50 | 995 | enum rtx_code code; |
996 | int i, j; | |
997 | const char *fmt; | |
998 | int changed = 0; | |
752df20e | 999 | |
1000 | if (x == 0) | |
c39100fe | 1001 | return 0; |
752df20e | 1002 | |
1003 | code = GET_CODE (x); | |
1004 | if (code == REG) | |
1005 | { | |
02e7a332 | 1006 | unsigned int regno = REGNO (x); |
a2c6f0b7 | 1007 | unsigned int endregno = END_REGNO (x); |
02e7a332 | 1008 | unsigned int i; |
752df20e | 1009 | |
1010 | for (i = regno; i < endregno; i++) | |
1011 | { | |
d1264606 | 1012 | if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i)) |
752df20e | 1013 | remove_invalid_refs (i); |
1014 | ||
d1264606 | 1015 | REG_IN_TABLE (i) = REG_TICK (i); |
126fb012 | 1016 | SUBREG_TICKED (i) = -1; |
752df20e | 1017 | } |
1018 | ||
1019 | return 0; | |
1020 | } | |
1021 | ||
e6860d27 | 1022 | /* If this is a SUBREG, we don't want to discard other SUBREGs of the same |
1023 | pseudo if they don't use overlapping words. We handle only pseudos | |
1024 | here for simplicity. */ | |
8ad4c111 | 1025 | if (code == SUBREG && REG_P (SUBREG_REG (x)) |
e6860d27 | 1026 | && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER) |
1027 | { | |
02e7a332 | 1028 | unsigned int i = REGNO (SUBREG_REG (x)); |
e6860d27 | 1029 | |
d1264606 | 1030 | if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i)) |
e6860d27 | 1031 | { |
126fb012 | 1032 | /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and |
1033 | the last store to this register really stored into this | |
1034 | subreg, then remove the memory of this subreg. | |
1035 | Otherwise, remove any memory of the entire register and | |
1036 | all its subregs from the table. */ | |
1037 | if (REG_TICK (i) - REG_IN_TABLE (i) > 1 | |
70e488ba | 1038 | || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x))) |
e6860d27 | 1039 | remove_invalid_refs (i); |
1040 | else | |
701e46d0 | 1041 | remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x)); |
e6860d27 | 1042 | } |
1043 | ||
d1264606 | 1044 | REG_IN_TABLE (i) = REG_TICK (i); |
70e488ba | 1045 | SUBREG_TICKED (i) = REGNO (SUBREG_REG (x)); |
e6860d27 | 1046 | return 0; |
1047 | } | |
1048 | ||
752df20e | 1049 | /* If X is a comparison or a COMPARE and either operand is a register |
1050 | that does not have a quantity, give it one. This is so that a later | |
1051 | call to record_jump_equiv won't cause X to be assigned a different | |
1052 | hash code and not found in the table after that call. | |
1053 | ||
1054 | It is not necessary to do this here, since rehash_using_reg can | |
1055 | fix up the table later, but doing this here eliminates the need to | |
1056 | call that expensive function in the most common case where the only | |
1057 | use of the register is in the comparison. */ | |
1058 | ||
6720e96c | 1059 | if (code == COMPARE || COMPARISON_P (x)) |
752df20e | 1060 | { |
8ad4c111 | 1061 | if (REG_P (XEXP (x, 0)) |
752df20e | 1062 | && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))) |
4679ade3 | 1063 | if (insert_regs (XEXP (x, 0), NULL, 0)) |
752df20e | 1064 | { |
1065 | rehash_using_reg (XEXP (x, 0)); | |
1066 | changed = 1; | |
1067 | } | |
1068 | ||
8ad4c111 | 1069 | if (REG_P (XEXP (x, 1)) |
752df20e | 1070 | && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1)))) |
4679ade3 | 1071 | if (insert_regs (XEXP (x, 1), NULL, 0)) |
752df20e | 1072 | { |
1073 | rehash_using_reg (XEXP (x, 1)); | |
1074 | changed = 1; | |
1075 | } | |
1076 | } | |
1077 | ||
1078 | fmt = GET_RTX_FORMAT (code); | |
1079 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1080 | if (fmt[i] == 'e') | |
1081 | changed |= mention_regs (XEXP (x, i)); | |
1082 | else if (fmt[i] == 'E') | |
1083 | for (j = 0; j < XVECLEN (x, i); j++) | |
1084 | changed |= mention_regs (XVECEXP (x, i, j)); | |
1085 | ||
1086 | return changed; | |
1087 | } | |
1088 | ||
1089 | /* Update the register quantities for inserting X into the hash table | |
1090 | with a value equivalent to CLASSP. | |
1091 | (If the class does not contain a REG, it is irrelevant.) | |
1092 | If MODIFIED is nonzero, X is a destination; it is being modified. | |
1093 | Note that delete_reg_equiv should be called on a register | |
1094 | before insert_regs is done on that register with MODIFIED != 0. | |
1095 | ||
1096 | Nonzero value means that elements of reg_qty have changed | |
1097 | so X's hash code may be different. */ | |
1098 | ||
1099 | static int | |
8ec3a57b | 1100 | insert_regs (rtx x, struct table_elt *classp, int modified) |
752df20e | 1101 | { |
8ad4c111 | 1102 | if (REG_P (x)) |
752df20e | 1103 | { |
02e7a332 | 1104 | unsigned int regno = REGNO (x); |
1105 | int qty_valid; | |
752df20e | 1106 | |
0aee3bb1 | 1107 | /* If REGNO is in the equivalence table already but is of the |
1108 | wrong mode for that equivalence, don't do anything here. */ | |
1109 | ||
a7f3b1c7 | 1110 | qty_valid = REGNO_QTY_VALID_P (regno); |
1111 | if (qty_valid) | |
1112 | { | |
1113 | struct qty_table_elem *ent = &qty_table[REG_QTY (regno)]; | |
0aee3bb1 | 1114 | |
a7f3b1c7 | 1115 | if (ent->mode != GET_MODE (x)) |
1116 | return 0; | |
1117 | } | |
1118 | ||
1119 | if (modified || ! qty_valid) | |
752df20e | 1120 | { |
1121 | if (classp) | |
1122 | for (classp = classp->first_same_value; | |
1123 | classp != 0; | |
1124 | classp = classp->next_same_value) | |
8ad4c111 | 1125 | if (REG_P (classp->exp) |
752df20e | 1126 | && GET_MODE (classp->exp) == GET_MODE (x)) |
1127 | { | |
412c63b0 | 1128 | unsigned c_regno = REGNO (classp->exp); |
1129 | ||
1130 | gcc_assert (REGNO_QTY_VALID_P (c_regno)); | |
1131 | ||
1132 | /* Suppose that 5 is hard reg and 100 and 101 are | |
1133 | pseudos. Consider | |
1134 | ||
1135 | (set (reg:si 100) (reg:si 5)) | |
1136 | (set (reg:si 5) (reg:si 100)) | |
1137 | (set (reg:di 101) (reg:di 5)) | |
1138 | ||
1139 | We would now set REG_QTY (101) = REG_QTY (5), but the | |
1140 | entry for 5 is in SImode. When we use this later in | |
1141 | copy propagation, we get the register in wrong mode. */ | |
1142 | if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x)) | |
1143 | continue; | |
1144 | ||
1145 | make_regs_eqv (regno, c_regno); | |
752df20e | 1146 | return 1; |
1147 | } | |
1148 | ||
6c1128fe | 1149 | /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger |
1150 | than REG_IN_TABLE to find out if there was only a single preceding | |
1151 | invalidation - for the SUBREG - or another one, which would be | |
1152 | for the full register. However, if we find here that REG_TICK | |
1153 | indicates that the register is invalid, it means that it has | |
1154 | been invalidated in a separate operation. The SUBREG might be used | |
1155 | now (then this is a recursive call), or we might use the full REG | |
1156 | now and a SUBREG of it later. So bump up REG_TICK so that | |
1157 | mention_regs will do the right thing. */ | |
1158 | if (! modified | |
1159 | && REG_IN_TABLE (regno) >= 0 | |
1160 | && REG_TICK (regno) == REG_IN_TABLE (regno) + 1) | |
1161 | REG_TICK (regno)++; | |
a7f3b1c7 | 1162 | make_new_qty (regno, GET_MODE (x)); |
752df20e | 1163 | return 1; |
1164 | } | |
89bbb48f | 1165 | |
1166 | return 0; | |
752df20e | 1167 | } |
50cf1c21 | 1168 | |
1169 | /* If X is a SUBREG, we will likely be inserting the inner register in the | |
1170 | table. If that register doesn't have an assigned quantity number at | |
1171 | this point but does later, the insertion that we will be doing now will | |
1172 | not be accessible because its hash code will have changed. So assign | |
1173 | a quantity number now. */ | |
1174 | ||
8ad4c111 | 1175 | else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x)) |
50cf1c21 | 1176 | && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x)))) |
1177 | { | |
4679ade3 | 1178 | insert_regs (SUBREG_REG (x), NULL, 0); |
e6860d27 | 1179 | mention_regs (x); |
50cf1c21 | 1180 | return 1; |
1181 | } | |
752df20e | 1182 | else |
1183 | return mention_regs (x); | |
1184 | } | |
1185 | \f | |
01c8e4c9 | 1186 | |
1187 | /* Compute upper and lower anchors for CST. Also compute the offset of CST | |
1188 | from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff | |
1189 | CST is equal to an anchor. */ | |
1190 | ||
1191 | static bool | |
1192 | compute_const_anchors (rtx cst, | |
1193 | HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs, | |
1194 | HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs) | |
1195 | { | |
1196 | HOST_WIDE_INT n = INTVAL (cst); | |
1197 | ||
1198 | *lower_base = n & ~(targetm.const_anchor - 1); | |
1199 | if (*lower_base == n) | |
1200 | return false; | |
1201 | ||
1202 | *upper_base = | |
1203 | (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1); | |
1204 | *upper_offs = n - *upper_base; | |
1205 | *lower_offs = n - *lower_base; | |
1206 | return true; | |
1207 | } | |
1208 | ||
1209 | /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */ | |
1210 | ||
1211 | static void | |
1212 | insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs, | |
3754d046 | 1213 | machine_mode mode) |
01c8e4c9 | 1214 | { |
1215 | struct table_elt *elt; | |
1216 | unsigned hash; | |
1217 | rtx anchor_exp; | |
1218 | rtx exp; | |
1219 | ||
1220 | anchor_exp = GEN_INT (anchor); | |
1221 | hash = HASH (anchor_exp, mode); | |
1222 | elt = lookup (anchor_exp, hash, mode); | |
1223 | if (!elt) | |
1224 | elt = insert (anchor_exp, NULL, hash, mode); | |
1225 | ||
29c05e22 | 1226 | exp = plus_constant (mode, reg, offs); |
01c8e4c9 | 1227 | /* REG has just been inserted and the hash codes recomputed. */ |
1228 | mention_regs (exp); | |
1229 | hash = HASH (exp, mode); | |
1230 | ||
1231 | /* Use the cost of the register rather than the whole expression. When | |
1232 | looking up constant anchors we will further offset the corresponding | |
1233 | expression therefore it does not make sense to prefer REGs over | |
1234 | reg-immediate additions. Prefer instead the oldest expression. Also | |
1235 | don't prefer pseudos over hard regs so that we derive constants in | |
1236 | argument registers from other argument registers rather than from the | |
1237 | original pseudo that was used to synthesize the constant. */ | |
5ae4887d | 1238 | insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1); |
01c8e4c9 | 1239 | } |
1240 | ||
1241 | /* The constant CST is equivalent to the register REG. Create | |
1242 | equivalences between the two anchors of CST and the corresponding | |
1243 | register-offset expressions using REG. */ | |
1244 | ||
1245 | static void | |
3754d046 | 1246 | insert_const_anchors (rtx reg, rtx cst, machine_mode mode) |
01c8e4c9 | 1247 | { |
1248 | HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs; | |
1249 | ||
1250 | if (!compute_const_anchors (cst, &lower_base, &lower_offs, | |
1251 | &upper_base, &upper_offs)) | |
1252 | return; | |
1253 | ||
1254 | /* Ignore anchors of value 0. Constants accessible from zero are | |
1255 | simple. */ | |
1256 | if (lower_base != 0) | |
1257 | insert_const_anchor (lower_base, reg, -lower_offs, mode); | |
1258 | ||
1259 | if (upper_base != 0) | |
1260 | insert_const_anchor (upper_base, reg, -upper_offs, mode); | |
1261 | } | |
1262 | ||
1263 | /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of | |
1264 | ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a | |
1265 | valid expression. Return the cheapest and oldest of such expressions. In | |
1266 | *OLD, return how old the resulting expression is compared to the other | |
1267 | equivalent expressions. */ | |
1268 | ||
1269 | static rtx | |
1270 | find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs, | |
1271 | unsigned *old) | |
1272 | { | |
1273 | struct table_elt *elt; | |
1274 | unsigned idx; | |
1275 | struct table_elt *match_elt; | |
1276 | rtx match; | |
1277 | ||
1278 | /* Find the cheapest and *oldest* expression to maximize the chance of | |
1279 | reusing the same pseudo. */ | |
1280 | ||
1281 | match_elt = NULL; | |
1282 | match = NULL_RTX; | |
1283 | for (elt = anchor_elt->first_same_value, idx = 0; | |
1284 | elt; | |
1285 | elt = elt->next_same_value, idx++) | |
1286 | { | |
1287 | if (match_elt && CHEAPER (match_elt, elt)) | |
1288 | return match; | |
1289 | ||
1290 | if (REG_P (elt->exp) | |
1291 | || (GET_CODE (elt->exp) == PLUS | |
1292 | && REG_P (XEXP (elt->exp, 0)) | |
1293 | && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT)) | |
1294 | { | |
1295 | rtx x; | |
1296 | ||
1297 | /* Ignore expressions that are no longer valid. */ | |
1298 | if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false)) | |
1299 | continue; | |
1300 | ||
29c05e22 | 1301 | x = plus_constant (GET_MODE (elt->exp), elt->exp, offs); |
01c8e4c9 | 1302 | if (REG_P (x) |
1303 | || (GET_CODE (x) == PLUS | |
1304 | && IN_RANGE (INTVAL (XEXP (x, 1)), | |
1305 | -targetm.const_anchor, | |
1306 | targetm.const_anchor - 1))) | |
1307 | { | |
1308 | match = x; | |
1309 | match_elt = elt; | |
1310 | *old = idx; | |
1311 | } | |
1312 | } | |
1313 | } | |
1314 | ||
1315 | return match; | |
1316 | } | |
1317 | ||
1318 | /* Try to express the constant SRC_CONST using a register+offset expression | |
1319 | derived from a constant anchor. Return it if successful or NULL_RTX, | |
1320 | otherwise. */ | |
1321 | ||
1322 | static rtx | |
3754d046 | 1323 | try_const_anchors (rtx src_const, machine_mode mode) |
01c8e4c9 | 1324 | { |
1325 | struct table_elt *lower_elt, *upper_elt; | |
1326 | HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs; | |
1327 | rtx lower_anchor_rtx, upper_anchor_rtx; | |
1328 | rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX; | |
1329 | unsigned lower_old, upper_old; | |
1330 | ||
211e63b0 | 1331 | /* CONST_INT is used for CC modes, but we should leave those alone. */ |
1332 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
1333 | return NULL_RTX; | |
1334 | ||
1335 | gcc_assert (SCALAR_INT_MODE_P (mode)); | |
01c8e4c9 | 1336 | if (!compute_const_anchors (src_const, &lower_base, &lower_offs, |
1337 | &upper_base, &upper_offs)) | |
1338 | return NULL_RTX; | |
1339 | ||
1340 | lower_anchor_rtx = GEN_INT (lower_base); | |
1341 | upper_anchor_rtx = GEN_INT (upper_base); | |
1342 | lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode); | |
1343 | upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode); | |
1344 | ||
1345 | if (lower_elt) | |
1346 | lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old); | |
1347 | if (upper_elt) | |
1348 | upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old); | |
1349 | ||
1350 | if (!lower_exp) | |
1351 | return upper_exp; | |
1352 | if (!upper_exp) | |
1353 | return lower_exp; | |
1354 | ||
1355 | /* Return the older expression. */ | |
1356 | return (upper_old > lower_old ? upper_exp : lower_exp); | |
1357 | } | |
1358 | \f | |
752df20e | 1359 | /* Look in or update the hash table. */ |
1360 | ||
752df20e | 1361 | /* Remove table element ELT from use in the table. |
1362 | HASH is its hash code, made using the HASH macro. | |
1363 | It's an argument because often that is known in advance | |
1364 | and we save much time not recomputing it. */ | |
1365 | ||
1366 | static void | |
8ec3a57b | 1367 | remove_from_table (struct table_elt *elt, unsigned int hash) |
752df20e | 1368 | { |
1369 | if (elt == 0) | |
1370 | return; | |
1371 | ||
1372 | /* Mark this element as removed. See cse_insn. */ | |
1373 | elt->first_same_value = 0; | |
1374 | ||
1375 | /* Remove the table element from its equivalence class. */ | |
cb10db9d | 1376 | |
752df20e | 1377 | { |
19cb6b50 | 1378 | struct table_elt *prev = elt->prev_same_value; |
1379 | struct table_elt *next = elt->next_same_value; | |
752df20e | 1380 | |
cb10db9d | 1381 | if (next) |
1382 | next->prev_same_value = prev; | |
752df20e | 1383 | |
1384 | if (prev) | |
1385 | prev->next_same_value = next; | |
1386 | else | |
1387 | { | |
19cb6b50 | 1388 | struct table_elt *newfirst = next; |
752df20e | 1389 | while (next) |
1390 | { | |
1391 | next->first_same_value = newfirst; | |
1392 | next = next->next_same_value; | |
1393 | } | |
1394 | } | |
1395 | } | |
1396 | ||
1397 | /* Remove the table element from its hash bucket. */ | |
1398 | ||
1399 | { | |
19cb6b50 | 1400 | struct table_elt *prev = elt->prev_same_hash; |
1401 | struct table_elt *next = elt->next_same_hash; | |
752df20e | 1402 | |
cb10db9d | 1403 | if (next) |
1404 | next->prev_same_hash = prev; | |
752df20e | 1405 | |
1406 | if (prev) | |
1407 | prev->next_same_hash = next; | |
1408 | else if (table[hash] == elt) | |
1409 | table[hash] = next; | |
1410 | else | |
1411 | { | |
1412 | /* This entry is not in the proper hash bucket. This can happen | |
1413 | when two classes were merged by `merge_equiv_classes'. Search | |
1414 | for the hash bucket that it heads. This happens only very | |
1415 | rarely, so the cost is acceptable. */ | |
9c4f3716 | 1416 | for (hash = 0; hash < HASH_SIZE; hash++) |
752df20e | 1417 | if (table[hash] == elt) |
1418 | table[hash] = next; | |
1419 | } | |
1420 | } | |
1421 | ||
1422 | /* Remove the table element from its related-value circular chain. */ | |
1423 | ||
1424 | if (elt->related_value != 0 && elt->related_value != elt) | |
1425 | { | |
19cb6b50 | 1426 | struct table_elt *p = elt->related_value; |
02e7a332 | 1427 | |
752df20e | 1428 | while (p->related_value != elt) |
1429 | p = p->related_value; | |
1430 | p->related_value = elt->related_value; | |
1431 | if (p->related_value == p) | |
1432 | p->related_value = 0; | |
1433 | } | |
1434 | ||
9c4f3716 | 1435 | /* Now add it to the free element chain. */ |
1436 | elt->next_same_hash = free_element_chain; | |
1437 | free_element_chain = elt; | |
752df20e | 1438 | } |
1439 | ||
d2c970fe | 1440 | /* Same as above, but X is a pseudo-register. */ |
1441 | ||
1442 | static void | |
1443 | remove_pseudo_from_table (rtx x, unsigned int hash) | |
1444 | { | |
1445 | struct table_elt *elt; | |
1446 | ||
1447 | /* Because a pseudo-register can be referenced in more than one | |
1448 | mode, we might have to remove more than one table entry. */ | |
1449 | while ((elt = lookup_for_remove (x, hash, VOIDmode))) | |
1450 | remove_from_table (elt, hash); | |
1451 | } | |
1452 | ||
752df20e | 1453 | /* Look up X in the hash table and return its table element, |
1454 | or 0 if X is not in the table. | |
1455 | ||
1456 | MODE is the machine-mode of X, or if X is an integer constant | |
1457 | with VOIDmode then MODE is the mode with which X will be used. | |
1458 | ||
1459 | Here we are satisfied to find an expression whose tree structure | |
1460 | looks like X. */ | |
1461 | ||
1462 | static struct table_elt * | |
3754d046 | 1463 | lookup (rtx x, unsigned int hash, machine_mode mode) |
752df20e | 1464 | { |
19cb6b50 | 1465 | struct table_elt *p; |
752df20e | 1466 | |
1467 | for (p = table[hash]; p; p = p->next_same_hash) | |
8ad4c111 | 1468 | if (mode == p->mode && ((x == p->exp && REG_P (x)) |
78d140c9 | 1469 | || exp_equiv_p (x, p->exp, !REG_P (x), false))) |
752df20e | 1470 | return p; |
1471 | ||
1472 | return 0; | |
1473 | } | |
1474 | ||
1475 | /* Like `lookup' but don't care whether the table element uses invalid regs. | |
1476 | Also ignore discrepancies in the machine mode of a register. */ | |
1477 | ||
1478 | static struct table_elt * | |
3754d046 | 1479 | lookup_for_remove (rtx x, unsigned int hash, machine_mode mode) |
752df20e | 1480 | { |
19cb6b50 | 1481 | struct table_elt *p; |
752df20e | 1482 | |
8ad4c111 | 1483 | if (REG_P (x)) |
752df20e | 1484 | { |
02e7a332 | 1485 | unsigned int regno = REGNO (x); |
1486 | ||
752df20e | 1487 | /* Don't check the machine mode when comparing registers; |
1488 | invalidating (REG:SI 0) also invalidates (REG:DF 0). */ | |
1489 | for (p = table[hash]; p; p = p->next_same_hash) | |
8ad4c111 | 1490 | if (REG_P (p->exp) |
752df20e | 1491 | && REGNO (p->exp) == regno) |
1492 | return p; | |
1493 | } | |
1494 | else | |
1495 | { | |
1496 | for (p = table[hash]; p; p = p->next_same_hash) | |
78d140c9 | 1497 | if (mode == p->mode |
1498 | && (x == p->exp || exp_equiv_p (x, p->exp, 0, false))) | |
752df20e | 1499 | return p; |
1500 | } | |
1501 | ||
1502 | return 0; | |
1503 | } | |
1504 | ||
1505 | /* Look for an expression equivalent to X and with code CODE. | |
1506 | If one is found, return that expression. */ | |
1507 | ||
1508 | static rtx | |
8ec3a57b | 1509 | lookup_as_function (rtx x, enum rtx_code code) |
752df20e | 1510 | { |
19cb6b50 | 1511 | struct table_elt *p |
78d140c9 | 1512 | = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x)); |
02e7a332 | 1513 | |
752df20e | 1514 | if (p == 0) |
1515 | return 0; | |
1516 | ||
1517 | for (p = p->first_same_value; p; p = p->next_same_value) | |
02e7a332 | 1518 | if (GET_CODE (p->exp) == code |
1519 | /* Make sure this is a valid entry in the table. */ | |
78d140c9 | 1520 | && exp_equiv_p (p->exp, p->exp, 1, false)) |
02e7a332 | 1521 | return p->exp; |
cb10db9d | 1522 | |
752df20e | 1523 | return 0; |
1524 | } | |
1525 | ||
01c8e4c9 | 1526 | /* Insert X in the hash table, assuming HASH is its hash code and |
1527 | CLASSP is an element of the class it should go in (or 0 if a new | |
1528 | class should be made). COST is the code of X and reg_cost is the | |
1529 | cost of registers in X. It is inserted at the proper position to | |
1530 | keep the class in the order cheapest first. | |
752df20e | 1531 | |
1532 | MODE is the machine-mode of X, or if X is an integer constant | |
1533 | with VOIDmode then MODE is the mode with which X will be used. | |
1534 | ||
1535 | For elements of equal cheapness, the most recent one | |
1536 | goes in front, except that the first element in the list | |
1537 | remains first unless a cheaper element is added. The order of | |
1538 | pseudo-registers does not matter, as canon_reg will be called to | |
5202ecf2 | 1539 | find the cheapest when a register is retrieved from the table. |
752df20e | 1540 | |
1541 | The in_memory field in the hash table element is set to 0. | |
1542 | The caller must set it nonzero if appropriate. | |
1543 | ||
1544 | You should call insert_regs (X, CLASSP, MODIFY) before calling here, | |
1545 | and if insert_regs returns a nonzero value | |
1546 | you must then recompute its hash code before calling here. | |
1547 | ||
1548 | If necessary, update table showing constant values of quantities. */ | |
1549 | ||
752df20e | 1550 | static struct table_elt * |
01c8e4c9 | 1551 | insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash, |
3754d046 | 1552 | machine_mode mode, int cost, int reg_cost) |
752df20e | 1553 | { |
19cb6b50 | 1554 | struct table_elt *elt; |
752df20e | 1555 | |
1556 | /* If X is a register and we haven't made a quantity for it, | |
1557 | something is wrong. */ | |
cc636d56 | 1558 | gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x))); |
752df20e | 1559 | |
1560 | /* If X is a hard register, show it is being put in the table. */ | |
8ad4c111 | 1561 | if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER) |
a2c6f0b7 | 1562 | add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x)); |
752df20e | 1563 | |
752df20e | 1564 | /* Put an element for X into the right hash bucket. */ |
1565 | ||
9c4f3716 | 1566 | elt = free_element_chain; |
1567 | if (elt) | |
02e7a332 | 1568 | free_element_chain = elt->next_same_hash; |
9c4f3716 | 1569 | else |
4c36ffe6 | 1570 | elt = XNEW (struct table_elt); |
9c4f3716 | 1571 | |
752df20e | 1572 | elt->exp = x; |
7cfb9bcf | 1573 | elt->canon_exp = NULL_RTX; |
01c8e4c9 | 1574 | elt->cost = cost; |
1575 | elt->regcost = reg_cost; | |
752df20e | 1576 | elt->next_same_value = 0; |
1577 | elt->prev_same_value = 0; | |
1578 | elt->next_same_hash = table[hash]; | |
1579 | elt->prev_same_hash = 0; | |
1580 | elt->related_value = 0; | |
1581 | elt->in_memory = 0; | |
1582 | elt->mode = mode; | |
b04fab2a | 1583 | elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x)); |
752df20e | 1584 | |
1585 | if (table[hash]) | |
1586 | table[hash]->prev_same_hash = elt; | |
1587 | table[hash] = elt; | |
1588 | ||
1589 | /* Put it into the proper value-class. */ | |
1590 | if (classp) | |
1591 | { | |
1592 | classp = classp->first_same_value; | |
1593 | if (CHEAPER (elt, classp)) | |
2358393e | 1594 | /* Insert at the head of the class. */ |
752df20e | 1595 | { |
19cb6b50 | 1596 | struct table_elt *p; |
752df20e | 1597 | elt->next_same_value = classp; |
1598 | classp->prev_same_value = elt; | |
1599 | elt->first_same_value = elt; | |
1600 | ||
1601 | for (p = classp; p; p = p->next_same_value) | |
1602 | p->first_same_value = elt; | |
1603 | } | |
1604 | else | |
1605 | { | |
1606 | /* Insert not at head of the class. */ | |
1607 | /* Put it after the last element cheaper than X. */ | |
19cb6b50 | 1608 | struct table_elt *p, *next; |
02e7a332 | 1609 | |
3c802a1e | 1610 | for (p = classp; |
1611 | (next = p->next_same_value) && CHEAPER (next, elt); | |
1612 | p = next) | |
1613 | ; | |
02e7a332 | 1614 | |
752df20e | 1615 | /* Put it after P and before NEXT. */ |
1616 | elt->next_same_value = next; | |
1617 | if (next) | |
1618 | next->prev_same_value = elt; | |
02e7a332 | 1619 | |
752df20e | 1620 | elt->prev_same_value = p; |
1621 | p->next_same_value = elt; | |
1622 | elt->first_same_value = classp; | |
1623 | } | |
1624 | } | |
1625 | else | |
1626 | elt->first_same_value = elt; | |
1627 | ||
1628 | /* If this is a constant being set equivalent to a register or a register | |
1629 | being set equivalent to a constant, note the constant equivalence. | |
1630 | ||
1631 | If this is a constant, it cannot be equivalent to a different constant, | |
1632 | and a constant is the only thing that can be cheaper than a register. So | |
1633 | we know the register is the head of the class (before the constant was | |
1634 | inserted). | |
1635 | ||
1636 | If this is a register that is not already known equivalent to a | |
1637 | constant, we must check the entire class. | |
1638 | ||
1639 | If this is a register that is already known equivalent to an insn, | |
a7f3b1c7 | 1640 | update the qtys `const_insn' to show that `this_insn' is the latest |
752df20e | 1641 | insn making that quantity equivalent to the constant. */ |
1642 | ||
8ad4c111 | 1643 | if (elt->is_const && classp && REG_P (classp->exp) |
1644 | && !REG_P (x)) | |
752df20e | 1645 | { |
a7f3b1c7 | 1646 | int exp_q = REG_QTY (REGNO (classp->exp)); |
1647 | struct qty_table_elem *exp_ent = &qty_table[exp_q]; | |
1648 | ||
316f48ea | 1649 | exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x); |
a7f3b1c7 | 1650 | exp_ent->const_insn = this_insn; |
752df20e | 1651 | } |
1652 | ||
8ad4c111 | 1653 | else if (REG_P (x) |
a7f3b1c7 | 1654 | && classp |
1655 | && ! qty_table[REG_QTY (REGNO (x))].const_rtx | |
67123c3e | 1656 | && ! elt->is_const) |
752df20e | 1657 | { |
19cb6b50 | 1658 | struct table_elt *p; |
752df20e | 1659 | |
1660 | for (p = classp; p != 0; p = p->next_same_value) | |
1661 | { | |
8ad4c111 | 1662 | if (p->is_const && !REG_P (p->exp)) |
752df20e | 1663 | { |
a7f3b1c7 | 1664 | int x_q = REG_QTY (REGNO (x)); |
1665 | struct qty_table_elem *x_ent = &qty_table[x_q]; | |
1666 | ||
02e7a332 | 1667 | x_ent->const_rtx |
316f48ea | 1668 | = gen_lowpart (GET_MODE (x), p->exp); |
a7f3b1c7 | 1669 | x_ent->const_insn = this_insn; |
752df20e | 1670 | break; |
1671 | } | |
1672 | } | |
1673 | } | |
1674 | ||
8ad4c111 | 1675 | else if (REG_P (x) |
a7f3b1c7 | 1676 | && qty_table[REG_QTY (REGNO (x))].const_rtx |
1677 | && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode) | |
1678 | qty_table[REG_QTY (REGNO (x))].const_insn = this_insn; | |
752df20e | 1679 | |
1680 | /* If this is a constant with symbolic value, | |
1681 | and it has a term with an explicit integer value, | |
1682 | link it up with related expressions. */ | |
1683 | if (GET_CODE (x) == CONST) | |
1684 | { | |
1685 | rtx subexp = get_related_value (x); | |
952bc06d | 1686 | unsigned subhash; |
752df20e | 1687 | struct table_elt *subelt, *subelt_prev; |
1688 | ||
1689 | if (subexp != 0) | |
1690 | { | |
1691 | /* Get the integer-free subexpression in the hash table. */ | |
78d140c9 | 1692 | subhash = SAFE_HASH (subexp, mode); |
752df20e | 1693 | subelt = lookup (subexp, subhash, mode); |
1694 | if (subelt == 0) | |
4679ade3 | 1695 | subelt = insert (subexp, NULL, subhash, mode); |
752df20e | 1696 | /* Initialize SUBELT's circular chain if it has none. */ |
1697 | if (subelt->related_value == 0) | |
1698 | subelt->related_value = subelt; | |
1699 | /* Find the element in the circular chain that precedes SUBELT. */ | |
1700 | subelt_prev = subelt; | |
1701 | while (subelt_prev->related_value != subelt) | |
1702 | subelt_prev = subelt_prev->related_value; | |
1703 | /* Put new ELT into SUBELT's circular chain just before SUBELT. | |
1704 | This way the element that follows SUBELT is the oldest one. */ | |
1705 | elt->related_value = subelt_prev->related_value; | |
1706 | subelt_prev->related_value = elt; | |
1707 | } | |
1708 | } | |
1709 | ||
1710 | return elt; | |
1711 | } | |
01c8e4c9 | 1712 | |
1713 | /* Wrap insert_with_costs by passing the default costs. */ | |
1714 | ||
1715 | static struct table_elt * | |
1716 | insert (rtx x, struct table_elt *classp, unsigned int hash, | |
3754d046 | 1717 | machine_mode mode) |
01c8e4c9 | 1718 | { |
5ae4887d | 1719 | return insert_with_costs (x, classp, hash, mode, |
1720 | COST (x, mode), approx_reg_cost (x)); | |
01c8e4c9 | 1721 | } |
1722 | ||
752df20e | 1723 | \f |
1724 | /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from | |
1725 | CLASS2 into CLASS1. This is done when we have reached an insn which makes | |
1726 | the two classes equivalent. | |
1727 | ||
1728 | CLASS1 will be the surviving class; CLASS2 should not be used after this | |
1729 | call. | |
1730 | ||
1731 | Any invalid entries in CLASS2 will not be copied. */ | |
1732 | ||
1733 | static void | |
8ec3a57b | 1734 | merge_equiv_classes (struct table_elt *class1, struct table_elt *class2) |
752df20e | 1735 | { |
d328ebdf | 1736 | struct table_elt *elt, *next, *new_elt; |
752df20e | 1737 | |
1738 | /* Ensure we start with the head of the classes. */ | |
1739 | class1 = class1->first_same_value; | |
1740 | class2 = class2->first_same_value; | |
1741 | ||
1742 | /* If they were already equal, forget it. */ | |
1743 | if (class1 == class2) | |
1744 | return; | |
1745 | ||
1746 | for (elt = class2; elt; elt = next) | |
1747 | { | |
02e7a332 | 1748 | unsigned int hash; |
752df20e | 1749 | rtx exp = elt->exp; |
3754d046 | 1750 | machine_mode mode = elt->mode; |
752df20e | 1751 | |
1752 | next = elt->next_same_value; | |
1753 | ||
1754 | /* Remove old entry, make a new one in CLASS1's class. | |
1755 | Don't do this for invalid entries as we cannot find their | |
a92771b8 | 1756 | hash code (it also isn't necessary). */ |
78d140c9 | 1757 | if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false)) |
752df20e | 1758 | { |
b57e33a4 | 1759 | bool need_rehash = false; |
1760 | ||
752df20e | 1761 | hash_arg_in_memory = 0; |
752df20e | 1762 | hash = HASH (exp, mode); |
cb10db9d | 1763 | |
8ad4c111 | 1764 | if (REG_P (exp)) |
b57e33a4 | 1765 | { |
1a5bccce | 1766 | need_rehash = REGNO_QTY_VALID_P (REGNO (exp)); |
b57e33a4 | 1767 | delete_reg_equiv (REGNO (exp)); |
1768 | } | |
cb10db9d | 1769 | |
d2c970fe | 1770 | if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER) |
1771 | remove_pseudo_from_table (exp, hash); | |
1772 | else | |
1773 | remove_from_table (elt, hash); | |
752df20e | 1774 | |
b57e33a4 | 1775 | if (insert_regs (exp, class1, 0) || need_rehash) |
1b033cc3 | 1776 | { |
1777 | rehash_using_reg (exp); | |
1778 | hash = HASH (exp, mode); | |
1779 | } | |
d328ebdf | 1780 | new_elt = insert (exp, class1, hash, mode); |
1781 | new_elt->in_memory = hash_arg_in_memory; | |
20d3ff08 | 1782 | if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST) |
1783 | new_elt->cost = MAX_COST; | |
752df20e | 1784 | } |
1785 | } | |
1786 | } | |
1787 | \f | |
53d90e4e | 1788 | /* Flush the entire hash table. */ |
1789 | ||
1790 | static void | |
8ec3a57b | 1791 | flush_hash_table (void) |
53d90e4e | 1792 | { |
1793 | int i; | |
1794 | struct table_elt *p; | |
1795 | ||
9c4f3716 | 1796 | for (i = 0; i < HASH_SIZE; i++) |
53d90e4e | 1797 | for (p = table[i]; p; p = table[i]) |
1798 | { | |
1799 | /* Note that invalidate can remove elements | |
1800 | after P in the current hash chain. */ | |
8ad4c111 | 1801 | if (REG_P (p->exp)) |
4c958a22 | 1802 | invalidate (p->exp, VOIDmode); |
53d90e4e | 1803 | else |
1804 | remove_from_table (p, i); | |
1805 | } | |
1806 | } | |
155b05dc | 1807 | \f |
e4a58c60 | 1808 | /* Check whether an anti dependence exists between X and EXP. MODE and |
1809 | ADDR are as for canon_anti_dependence. */ | |
37b8a8d6 | 1810 | |
e4a58c60 | 1811 | static bool |
3754d046 | 1812 | check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr) |
02b0feeb | 1813 | { |
e4a58c60 | 1814 | subrtx_iterator::array_type array; |
1815 | FOR_EACH_SUBRTX (iter, array, x, NONCONST) | |
1816 | { | |
1817 | const_rtx x = *iter; | |
1818 | if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr)) | |
1819 | return true; | |
1820 | } | |
1821 | return false; | |
02b0feeb | 1822 | } |
1823 | \f | |
155b05dc | 1824 | /* Remove from the hash table, or mark as invalid, all expressions whose |
1825 | values could be altered by storing in X. X is a register, a subreg, or | |
1826 | a memory reference with nonvarying address (because, when a memory | |
1827 | reference with a varying address is stored in, all memory references are | |
1828 | removed by invalidate_memory so specific invalidation is superfluous). | |
1829 | FULL_MODE, if not VOIDmode, indicates that this much should be | |
1830 | invalidated instead of just the amount indicated by the mode of X. This | |
1831 | is only used for bitfield stores into memory. | |
1832 | ||
1833 | A nonvarying address may be just a register or just a symbol reference, | |
1834 | or it may be either of those plus a numeric offset. */ | |
752df20e | 1835 | |
1836 | static void | |
3754d046 | 1837 | invalidate (rtx x, machine_mode full_mode) |
752df20e | 1838 | { |
19cb6b50 | 1839 | int i; |
1840 | struct table_elt *p; | |
56bbdce4 | 1841 | rtx addr; |
752df20e | 1842 | |
155b05dc | 1843 | switch (GET_CODE (x)) |
752df20e | 1844 | { |
155b05dc | 1845 | case REG: |
1846 | { | |
1847 | /* If X is a register, dependencies on its contents are recorded | |
1848 | through the qty number mechanism. Just change the qty number of | |
1849 | the register, mark it as invalid for expressions that refer to it, | |
1850 | and remove it itself. */ | |
02e7a332 | 1851 | unsigned int regno = REGNO (x); |
1852 | unsigned int hash = HASH (x, GET_MODE (x)); | |
752df20e | 1853 | |
155b05dc | 1854 | /* Remove REGNO from any quantity list it might be on and indicate |
1855 | that its value might have changed. If it is a pseudo, remove its | |
1856 | entry from the hash table. | |
752df20e | 1857 | |
155b05dc | 1858 | For a hard register, we do the first two actions above for any |
1859 | additional hard registers corresponding to X. Then, if any of these | |
1860 | registers are in the table, we must remove any REG entries that | |
1861 | overlap these registers. */ | |
752df20e | 1862 | |
155b05dc | 1863 | delete_reg_equiv (regno); |
1864 | REG_TICK (regno)++; | |
126fb012 | 1865 | SUBREG_TICKED (regno) = -1; |
f356ea3f | 1866 | |
155b05dc | 1867 | if (regno >= FIRST_PSEUDO_REGISTER) |
d2c970fe | 1868 | remove_pseudo_from_table (x, hash); |
155b05dc | 1869 | else |
1870 | { | |
1871 | HOST_WIDE_INT in_table | |
1872 | = TEST_HARD_REG_BIT (hard_regs_in_table, regno); | |
788bed51 | 1873 | unsigned int endregno = END_REGNO (x); |
02e7a332 | 1874 | unsigned int tregno, tendregno, rn; |
19cb6b50 | 1875 | struct table_elt *p, *next; |
752df20e | 1876 | |
155b05dc | 1877 | CLEAR_HARD_REG_BIT (hard_regs_in_table, regno); |
752df20e | 1878 | |
02e7a332 | 1879 | for (rn = regno + 1; rn < endregno; rn++) |
155b05dc | 1880 | { |
02e7a332 | 1881 | in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn); |
1882 | CLEAR_HARD_REG_BIT (hard_regs_in_table, rn); | |
1883 | delete_reg_equiv (rn); | |
1884 | REG_TICK (rn)++; | |
126fb012 | 1885 | SUBREG_TICKED (rn) = -1; |
155b05dc | 1886 | } |
752df20e | 1887 | |
155b05dc | 1888 | if (in_table) |
9c4f3716 | 1889 | for (hash = 0; hash < HASH_SIZE; hash++) |
155b05dc | 1890 | for (p = table[hash]; p; p = next) |
1891 | { | |
1892 | next = p->next_same_hash; | |
752df20e | 1893 | |
8ad4c111 | 1894 | if (!REG_P (p->exp) |
cb10db9d | 1895 | || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER) |
1896 | continue; | |
1897 | ||
155b05dc | 1898 | tregno = REGNO (p->exp); |
788bed51 | 1899 | tendregno = END_REGNO (p->exp); |
155b05dc | 1900 | if (tendregno > regno && tregno < endregno) |
1901 | remove_from_table (p, hash); | |
1902 | } | |
1903 | } | |
1904 | } | |
752df20e | 1905 | return; |
752df20e | 1906 | |
155b05dc | 1907 | case SUBREG: |
fdb25961 | 1908 | invalidate (SUBREG_REG (x), VOIDmode); |
752df20e | 1909 | return; |
6ede8018 | 1910 | |
155b05dc | 1911 | case PARALLEL: |
cb10db9d | 1912 | for (i = XVECLEN (x, 0) - 1; i >= 0; --i) |
6ede8018 | 1913 | invalidate (XVECEXP (x, 0, i), VOIDmode); |
1914 | return; | |
6ede8018 | 1915 | |
155b05dc | 1916 | case EXPR_LIST: |
1917 | /* This is part of a disjoint return value; extract the location in | |
1918 | question ignoring the offset. */ | |
6ede8018 | 1919 | invalidate (XEXP (x, 0), VOIDmode); |
1920 | return; | |
752df20e | 1921 | |
155b05dc | 1922 | case MEM: |
56bbdce4 | 1923 | addr = canon_rtx (get_addr (XEXP (x, 0))); |
7cfb9bcf | 1924 | /* Calculate the canonical version of X here so that |
1925 | true_dependence doesn't generate new RTL for X on each call. */ | |
1926 | x = canon_rtx (x); | |
1927 | ||
155b05dc | 1928 | /* Remove all hash table elements that refer to overlapping pieces of |
1929 | memory. */ | |
1930 | if (full_mode == VOIDmode) | |
1931 | full_mode = GET_MODE (x); | |
fdb25961 | 1932 | |
9c4f3716 | 1933 | for (i = 0; i < HASH_SIZE; i++) |
752df20e | 1934 | { |
19cb6b50 | 1935 | struct table_elt *next; |
155b05dc | 1936 | |
1937 | for (p = table[i]; p; p = next) | |
1938 | { | |
1939 | next = p->next_same_hash; | |
7cfb9bcf | 1940 | if (p->in_memory) |
1941 | { | |
02b0feeb | 1942 | /* Just canonicalize the expression once; |
1943 | otherwise each time we call invalidate | |
1944 | true_dependence will canonicalize the | |
1945 | expression again. */ | |
1946 | if (!p->canon_exp) | |
1947 | p->canon_exp = canon_rtx (p->exp); | |
e4a58c60 | 1948 | if (check_dependence (p->canon_exp, x, full_mode, addr)) |
7cfb9bcf | 1949 | remove_from_table (p, i); |
7cfb9bcf | 1950 | } |
155b05dc | 1951 | } |
752df20e | 1952 | } |
155b05dc | 1953 | return; |
1954 | ||
1955 | default: | |
cc636d56 | 1956 | gcc_unreachable (); |
752df20e | 1957 | } |
1958 | } | |
7a49a822 | 1959 | |
1960 | /* Invalidate DEST. Used when DEST is not going to be added | |
1961 | into the hash table for some reason, e.g. do_not_record | |
1962 | flagged on it. */ | |
1963 | ||
1964 | static void | |
1965 | invalidate_dest (rtx dest) | |
1966 | { | |
1967 | if (REG_P (dest) | |
1968 | || GET_CODE (dest) == SUBREG | |
1969 | || MEM_P (dest)) | |
1970 | invalidate (dest, VOIDmode); | |
1971 | else if (GET_CODE (dest) == STRICT_LOW_PART | |
1972 | || GET_CODE (dest) == ZERO_EXTRACT) | |
1973 | invalidate (XEXP (dest, 0), GET_MODE (dest)); | |
1974 | } | |
155b05dc | 1975 | \f |
752df20e | 1976 | /* Remove all expressions that refer to register REGNO, |
1977 | since they are already invalid, and we are about to | |
1978 | mark that register valid again and don't want the old | |
1979 | expressions to reappear as valid. */ | |
1980 | ||
1981 | static void | |
8ec3a57b | 1982 | remove_invalid_refs (unsigned int regno) |
752df20e | 1983 | { |
02e7a332 | 1984 | unsigned int i; |
1985 | struct table_elt *p, *next; | |
752df20e | 1986 | |
9c4f3716 | 1987 | for (i = 0; i < HASH_SIZE; i++) |
752df20e | 1988 | for (p = table[i]; p; p = next) |
1989 | { | |
1990 | next = p->next_same_hash; | |
2ec77a7c | 1991 | if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp)) |
752df20e | 1992 | remove_from_table (p, i); |
1993 | } | |
1994 | } | |
e6860d27 | 1995 | |
701e46d0 | 1996 | /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET, |
1997 | and mode MODE. */ | |
e6860d27 | 1998 | static void |
8ec3a57b | 1999 | remove_invalid_subreg_refs (unsigned int regno, unsigned int offset, |
3754d046 | 2000 | machine_mode mode) |
e6860d27 | 2001 | { |
02e7a332 | 2002 | unsigned int i; |
2003 | struct table_elt *p, *next; | |
701e46d0 | 2004 | unsigned int end = offset + (GET_MODE_SIZE (mode) - 1); |
e6860d27 | 2005 | |
9c4f3716 | 2006 | for (i = 0; i < HASH_SIZE; i++) |
e6860d27 | 2007 | for (p = table[i]; p; p = next) |
2008 | { | |
701e46d0 | 2009 | rtx exp = p->exp; |
e6860d27 | 2010 | next = p->next_same_hash; |
cb10db9d | 2011 | |
8ad4c111 | 2012 | if (!REG_P (exp) |
e6860d27 | 2013 | && (GET_CODE (exp) != SUBREG |
8ad4c111 | 2014 | || !REG_P (SUBREG_REG (exp)) |
e6860d27 | 2015 | || REGNO (SUBREG_REG (exp)) != regno |
701e46d0 | 2016 | || (((SUBREG_BYTE (exp) |
2017 | + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset) | |
2018 | && SUBREG_BYTE (exp) <= end)) | |
2ec77a7c | 2019 | && refers_to_regno_p (regno, p->exp)) |
e6860d27 | 2020 | remove_from_table (p, i); |
2021 | } | |
2022 | } | |
752df20e | 2023 | \f |
2024 | /* Recompute the hash codes of any valid entries in the hash table that | |
2025 | reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG. | |
2026 | ||
2027 | This is called when we make a jump equivalence. */ | |
2028 | ||
2029 | static void | |
8ec3a57b | 2030 | rehash_using_reg (rtx x) |
752df20e | 2031 | { |
3c1d7436 | 2032 | unsigned int i; |
752df20e | 2033 | struct table_elt *p, *next; |
952bc06d | 2034 | unsigned hash; |
752df20e | 2035 | |
2036 | if (GET_CODE (x) == SUBREG) | |
2037 | x = SUBREG_REG (x); | |
2038 | ||
2039 | /* If X is not a register or if the register is known not to be in any | |
2040 | valid entries in the table, we have no work to do. */ | |
2041 | ||
8ad4c111 | 2042 | if (!REG_P (x) |
d1264606 | 2043 | || REG_IN_TABLE (REGNO (x)) < 0 |
2044 | || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x))) | |
752df20e | 2045 | return; |
2046 | ||
2047 | /* Scan all hash chains looking for valid entries that mention X. | |
b57e33a4 | 2048 | If we find one and it is in the wrong hash chain, move it. */ |
752df20e | 2049 | |
9c4f3716 | 2050 | for (i = 0; i < HASH_SIZE; i++) |
752df20e | 2051 | for (p = table[i]; p; p = next) |
2052 | { | |
2053 | next = p->next_same_hash; | |
b57e33a4 | 2054 | if (reg_mentioned_p (x, p->exp) |
78d140c9 | 2055 | && exp_equiv_p (p->exp, p->exp, 1, false) |
2056 | && i != (hash = SAFE_HASH (p->exp, p->mode))) | |
752df20e | 2057 | { |
2058 | if (p->next_same_hash) | |
2059 | p->next_same_hash->prev_same_hash = p->prev_same_hash; | |
2060 | ||
2061 | if (p->prev_same_hash) | |
2062 | p->prev_same_hash->next_same_hash = p->next_same_hash; | |
2063 | else | |
2064 | table[i] = p->next_same_hash; | |
2065 | ||
2066 | p->next_same_hash = table[hash]; | |
2067 | p->prev_same_hash = 0; | |
2068 | if (table[hash]) | |
2069 | table[hash]->prev_same_hash = p; | |
2070 | table[hash] = p; | |
2071 | } | |
2072 | } | |
2073 | } | |
2074 | \f | |
752df20e | 2075 | /* Remove from the hash table any expression that is a call-clobbered |
2076 | register. Also update their TICK values. */ | |
2077 | ||
2078 | static void | |
8ec3a57b | 2079 | invalidate_for_call (void) |
752df20e | 2080 | { |
02e7a332 | 2081 | unsigned int regno, endregno; |
2082 | unsigned int i; | |
952bc06d | 2083 | unsigned hash; |
752df20e | 2084 | struct table_elt *p, *next; |
2085 | int in_table = 0; | |
24ec6636 | 2086 | hard_reg_set_iterator hrsi; |
752df20e | 2087 | |
2088 | /* Go through all the hard registers. For each that is clobbered in | |
2089 | a CALL_INSN, remove the register from quantity chains and update | |
2090 | reg_tick if defined. Also see if any of these registers is currently | |
2091 | in the table. */ | |
24ec6636 | 2092 | EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi) |
2093 | { | |
2094 | delete_reg_equiv (regno); | |
2095 | if (REG_TICK (regno) >= 0) | |
2096 | { | |
2097 | REG_TICK (regno)++; | |
2098 | SUBREG_TICKED (regno) = -1; | |
2099 | } | |
2100 | in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0); | |
2101 | } | |
752df20e | 2102 | |
2103 | /* In the case where we have no call-clobbered hard registers in the | |
2104 | table, we are done. Otherwise, scan the table and remove any | |
2105 | entry that overlaps a call-clobbered register. */ | |
2106 | ||
2107 | if (in_table) | |
9c4f3716 | 2108 | for (hash = 0; hash < HASH_SIZE; hash++) |
752df20e | 2109 | for (p = table[hash]; p; p = next) |
2110 | { | |
2111 | next = p->next_same_hash; | |
2112 | ||
8ad4c111 | 2113 | if (!REG_P (p->exp) |
752df20e | 2114 | || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER) |
2115 | continue; | |
2116 | ||
2117 | regno = REGNO (p->exp); | |
788bed51 | 2118 | endregno = END_REGNO (p->exp); |
752df20e | 2119 | |
2120 | for (i = regno; i < endregno; i++) | |
2121 | if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)) | |
2122 | { | |
2123 | remove_from_table (p, hash); | |
2124 | break; | |
2125 | } | |
2126 | } | |
2127 | } | |
2128 | \f | |
2129 | /* Given an expression X of type CONST, | |
2130 | and ELT which is its table entry (or 0 if it | |
2131 | is not in the hash table), | |
2132 | return an alternate expression for X as a register plus integer. | |
2133 | If none can be found, return 0. */ | |
2134 | ||
2135 | static rtx | |
8ec3a57b | 2136 | use_related_value (rtx x, struct table_elt *elt) |
752df20e | 2137 | { |
19cb6b50 | 2138 | struct table_elt *relt = 0; |
2139 | struct table_elt *p, *q; | |
b572011e | 2140 | HOST_WIDE_INT offset; |
752df20e | 2141 | |
2142 | /* First, is there anything related known? | |
2143 | If we have a table element, we can tell from that. | |
2144 | Otherwise, must look it up. */ | |
2145 | ||
2146 | if (elt != 0 && elt->related_value != 0) | |
2147 | relt = elt; | |
2148 | else if (elt == 0 && GET_CODE (x) == CONST) | |
2149 | { | |
2150 | rtx subexp = get_related_value (x); | |
2151 | if (subexp != 0) | |
2152 | relt = lookup (subexp, | |
78d140c9 | 2153 | SAFE_HASH (subexp, GET_MODE (subexp)), |
752df20e | 2154 | GET_MODE (subexp)); |
2155 | } | |
2156 | ||
2157 | if (relt == 0) | |
2158 | return 0; | |
2159 | ||
2160 | /* Search all related table entries for one that has an | |
2161 | equivalent register. */ | |
2162 | ||
2163 | p = relt; | |
2164 | while (1) | |
2165 | { | |
2166 | /* This loop is strange in that it is executed in two different cases. | |
2167 | The first is when X is already in the table. Then it is searching | |
2168 | the RELATED_VALUE list of X's class (RELT). The second case is when | |
2169 | X is not in the table. Then RELT points to a class for the related | |
2170 | value. | |
2171 | ||
2172 | Ensure that, whatever case we are in, that we ignore classes that have | |
2173 | the same value as X. */ | |
2174 | ||
2175 | if (rtx_equal_p (x, p->exp)) | |
2176 | q = 0; | |
2177 | else | |
2178 | for (q = p->first_same_value; q; q = q->next_same_value) | |
8ad4c111 | 2179 | if (REG_P (q->exp)) |
752df20e | 2180 | break; |
2181 | ||
2182 | if (q) | |
2183 | break; | |
2184 | ||
2185 | p = p->related_value; | |
2186 | ||
2187 | /* We went all the way around, so there is nothing to be found. | |
2188 | Alternatively, perhaps RELT was in the table for some other reason | |
2189 | and it has no related values recorded. */ | |
2190 | if (p == relt || p == 0) | |
2191 | break; | |
2192 | } | |
2193 | ||
2194 | if (q == 0) | |
2195 | return 0; | |
2196 | ||
2197 | offset = (get_integer_term (x) - get_integer_term (p->exp)); | |
2198 | /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */ | |
29c05e22 | 2199 | return plus_constant (q->mode, q->exp, offset); |
752df20e | 2200 | } |
2201 | \f | |
e1ab7874 | 2202 | |
d91f2122 | 2203 | /* Hash a string. Just add its bytes up. */ |
2204 | static inline unsigned | |
78d140c9 | 2205 | hash_rtx_string (const char *ps) |
d91f2122 | 2206 | { |
2207 | unsigned hash = 0; | |
d4c5e26d | 2208 | const unsigned char *p = (const unsigned char *) ps; |
2209 | ||
d91f2122 | 2210 | if (p) |
2211 | while (*p) | |
2212 | hash += *p++; | |
2213 | ||
2214 | return hash; | |
2215 | } | |
2216 | ||
48e1416a | 2217 | /* Same as hash_rtx, but call CB on each rtx if it is not NULL. |
e1ab7874 | 2218 | When the callback returns true, we continue with the new rtx. */ |
752df20e | 2219 | |
78d140c9 | 2220 | unsigned |
3754d046 | 2221 | hash_rtx_cb (const_rtx x, machine_mode mode, |
e1ab7874 | 2222 | int *do_not_record_p, int *hash_arg_in_memory_p, |
2223 | bool have_reg_qty, hash_rtx_callback_function cb) | |
752df20e | 2224 | { |
19cb6b50 | 2225 | int i, j; |
2226 | unsigned hash = 0; | |
2227 | enum rtx_code code; | |
2228 | const char *fmt; | |
3754d046 | 2229 | machine_mode newmode; |
e1ab7874 | 2230 | rtx newx; |
752df20e | 2231 | |
78d140c9 | 2232 | /* Used to turn recursion into iteration. We can't rely on GCC's |
2233 | tail-recursion elimination since we need to keep accumulating values | |
2234 | in HASH. */ | |
752df20e | 2235 | repeat: |
2236 | if (x == 0) | |
2237 | return hash; | |
2238 | ||
e1ab7874 | 2239 | /* Invoke the callback first. */ |
48e1416a | 2240 | if (cb != NULL |
e1ab7874 | 2241 | && ((*cb) (x, mode, &newx, &newmode))) |
2242 | { | |
2243 | hash += hash_rtx_cb (newx, newmode, do_not_record_p, | |
2244 | hash_arg_in_memory_p, have_reg_qty, cb); | |
2245 | return hash; | |
2246 | } | |
2247 | ||
752df20e | 2248 | code = GET_CODE (x); |
2249 | switch (code) | |
2250 | { | |
2251 | case REG: | |
2252 | { | |
02e7a332 | 2253 | unsigned int regno = REGNO (x); |
752df20e | 2254 | |
e1ab7874 | 2255 | if (do_not_record_p && !reload_completed) |
752df20e | 2256 | { |
78d140c9 | 2257 | /* On some machines, we can't record any non-fixed hard register, |
2258 | because extending its life will cause reload problems. We | |
2259 | consider ap, fp, sp, gp to be fixed for this purpose. | |
2260 | ||
2261 | We also consider CCmode registers to be fixed for this purpose; | |
2262 | failure to do so leads to failure to simplify 0<100 type of | |
2263 | conditionals. | |
2264 | ||
2265 | On all machines, we can't record any global registers. | |
2266 | Nor should we record any register that is in a small | |
24dd0668 | 2267 | class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */ |
78d140c9 | 2268 | bool record; |
2269 | ||
2270 | if (regno >= FIRST_PSEUDO_REGISTER) | |
2271 | record = true; | |
2272 | else if (x == frame_pointer_rtx | |
2273 | || x == hard_frame_pointer_rtx | |
2274 | || x == arg_pointer_rtx | |
2275 | || x == stack_pointer_rtx | |
2276 | || x == pic_offset_table_rtx) | |
2277 | record = true; | |
2278 | else if (global_regs[regno]) | |
2279 | record = false; | |
2280 | else if (fixed_regs[regno]) | |
2281 | record = true; | |
2282 | else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC) | |
2283 | record = true; | |
ed5527ca | 2284 | else if (targetm.small_register_classes_for_mode_p (GET_MODE (x))) |
78d140c9 | 2285 | record = false; |
24dd0668 | 2286 | else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno))) |
78d140c9 | 2287 | record = false; |
2288 | else | |
2289 | record = true; | |
2290 | ||
2291 | if (!record) | |
2292 | { | |
2293 | *do_not_record_p = 1; | |
2294 | return 0; | |
2295 | } | |
752df20e | 2296 | } |
02e7a332 | 2297 | |
78d140c9 | 2298 | hash += ((unsigned int) REG << 7); |
2299 | hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno); | |
952bc06d | 2300 | return hash; |
752df20e | 2301 | } |
2302 | ||
e6860d27 | 2303 | /* We handle SUBREG of a REG specially because the underlying |
2304 | reg changes its hash value with every value change; we don't | |
2305 | want to have to forget unrelated subregs when one subreg changes. */ | |
2306 | case SUBREG: | |
2307 | { | |
8ad4c111 | 2308 | if (REG_P (SUBREG_REG (x))) |
e6860d27 | 2309 | { |
78d140c9 | 2310 | hash += (((unsigned int) SUBREG << 7) |
701e46d0 | 2311 | + REGNO (SUBREG_REG (x)) |
2312 | + (SUBREG_BYTE (x) / UNITS_PER_WORD)); | |
e6860d27 | 2313 | return hash; |
2314 | } | |
2315 | break; | |
2316 | } | |
2317 | ||
752df20e | 2318 | case CONST_INT: |
78d140c9 | 2319 | hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode |
2320 | + (unsigned int) INTVAL (x)); | |
2321 | return hash; | |
752df20e | 2322 | |
e913b5cd | 2323 | case CONST_WIDE_INT: |
c4050ce7 | 2324 | for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++) |
2325 | hash += CONST_WIDE_INT_ELT (x, i); | |
e913b5cd | 2326 | return hash; |
2327 | ||
752df20e | 2328 | case CONST_DOUBLE: |
2329 | /* This is like the general case, except that it only counts | |
2330 | the integers representing the constant. */ | |
78d140c9 | 2331 | hash += (unsigned int) code + (unsigned int) GET_MODE (x); |
e913b5cd | 2332 | if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode) |
78d140c9 | 2333 | hash += ((unsigned int) CONST_DOUBLE_LOW (x) |
2334 | + (unsigned int) CONST_DOUBLE_HIGH (x)); | |
e913b5cd | 2335 | else |
2336 | hash += real_hash (CONST_DOUBLE_REAL_VALUE (x)); | |
752df20e | 2337 | return hash; |
2338 | ||
e397ad8e | 2339 | case CONST_FIXED: |
2340 | hash += (unsigned int) code + (unsigned int) GET_MODE (x); | |
2341 | hash += fixed_hash (CONST_FIXED_VALUE (x)); | |
2342 | return hash; | |
2343 | ||
886cfd4f | 2344 | case CONST_VECTOR: |
2345 | { | |
2346 | int units; | |
2347 | rtx elt; | |
2348 | ||
2349 | units = CONST_VECTOR_NUNITS (x); | |
2350 | ||
2351 | for (i = 0; i < units; ++i) | |
2352 | { | |
2353 | elt = CONST_VECTOR_ELT (x, i); | |
e1ab7874 | 2354 | hash += hash_rtx_cb (elt, GET_MODE (elt), |
48e1416a | 2355 | do_not_record_p, hash_arg_in_memory_p, |
e1ab7874 | 2356 | have_reg_qty, cb); |
886cfd4f | 2357 | } |
2358 | ||
2359 | return hash; | |
2360 | } | |
2361 | ||
752df20e | 2362 | /* Assume there is only one rtx object for any given label. */ |
2363 | case LABEL_REF: | |
78d140c9 | 2364 | /* We don't hash on the address of the CODE_LABEL to avoid bootstrap |
2365 | differences and differences between each stage's debugging dumps. */ | |
2366 | hash += (((unsigned int) LABEL_REF << 7) | |
b49f2e4b | 2367 | + CODE_LABEL_NUMBER (LABEL_REF_LABEL (x))); |
952bc06d | 2368 | return hash; |
752df20e | 2369 | |
2370 | case SYMBOL_REF: | |
78d140c9 | 2371 | { |
2372 | /* Don't hash on the symbol's address to avoid bootstrap differences. | |
2373 | Different hash values may cause expressions to be recorded in | |
2374 | different orders and thus different registers to be used in the | |
2375 | final assembler. This also avoids differences in the dump files | |
2376 | between various stages. */ | |
2377 | unsigned int h = 0; | |
2378 | const unsigned char *p = (const unsigned char *) XSTR (x, 0); | |
2379 | ||
2380 | while (*p) | |
2381 | h += (h << 7) + *p++; /* ??? revisit */ | |
2382 | ||
2383 | hash += ((unsigned int) SYMBOL_REF << 7) + h; | |
2384 | return hash; | |
2385 | } | |
752df20e | 2386 | |
2387 | case MEM: | |
155b05dc | 2388 | /* We don't record if marked volatile or if BLKmode since we don't |
2389 | know the size of the move. */ | |
e1ab7874 | 2390 | if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)) |
752df20e | 2391 | { |
78d140c9 | 2392 | *do_not_record_p = 1; |
752df20e | 2393 | return 0; |
2394 | } | |
78d140c9 | 2395 | if (hash_arg_in_memory_p && !MEM_READONLY_P (x)) |
2396 | *hash_arg_in_memory_p = 1; | |
805e22b2 | 2397 | |
752df20e | 2398 | /* Now that we have already found this special case, |
2399 | might as well speed it up as much as possible. */ | |
952bc06d | 2400 | hash += (unsigned) MEM; |
752df20e | 2401 | x = XEXP (x, 0); |
2402 | goto repeat; | |
2403 | ||
7002e559 | 2404 | case USE: |
2405 | /* A USE that mentions non-volatile memory needs special | |
2406 | handling since the MEM may be BLKmode which normally | |
2407 | prevents an entry from being made. Pure calls are | |
78d140c9 | 2408 | marked by a USE which mentions BLKmode memory. |
2409 | See calls.c:emit_call_1. */ | |
e16ceb8e | 2410 | if (MEM_P (XEXP (x, 0)) |
7002e559 | 2411 | && ! MEM_VOLATILE_P (XEXP (x, 0))) |
2412 | { | |
d4c5e26d | 2413 | hash += (unsigned) USE; |
7002e559 | 2414 | x = XEXP (x, 0); |
2415 | ||
78d140c9 | 2416 | if (hash_arg_in_memory_p && !MEM_READONLY_P (x)) |
2417 | *hash_arg_in_memory_p = 1; | |
7002e559 | 2418 | |
2419 | /* Now that we have already found this special case, | |
2420 | might as well speed it up as much as possible. */ | |
2421 | hash += (unsigned) MEM; | |
2422 | x = XEXP (x, 0); | |
2423 | goto repeat; | |
2424 | } | |
2425 | break; | |
2426 | ||
752df20e | 2427 | case PRE_DEC: |
2428 | case PRE_INC: | |
2429 | case POST_DEC: | |
2430 | case POST_INC: | |
40988080 | 2431 | case PRE_MODIFY: |
2432 | case POST_MODIFY: | |
752df20e | 2433 | case PC: |
2434 | case CC0: | |
2435 | case CALL: | |
2436 | case UNSPEC_VOLATILE: | |
e1ab7874 | 2437 | if (do_not_record_p) { |
2438 | *do_not_record_p = 1; | |
2439 | return 0; | |
2440 | } | |
2441 | else | |
2442 | return hash; | |
2443 | break; | |
752df20e | 2444 | |
2445 | case ASM_OPERANDS: | |
e1ab7874 | 2446 | if (do_not_record_p && MEM_VOLATILE_P (x)) |
752df20e | 2447 | { |
78d140c9 | 2448 | *do_not_record_p = 1; |
752df20e | 2449 | return 0; |
2450 | } | |
d91f2122 | 2451 | else |
2452 | { | |
2453 | /* We don't want to take the filename and line into account. */ | |
2454 | hash += (unsigned) code + (unsigned) GET_MODE (x) | |
78d140c9 | 2455 | + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x)) |
2456 | + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x)) | |
d91f2122 | 2457 | + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x); |
2458 | ||
2459 | if (ASM_OPERANDS_INPUT_LENGTH (x)) | |
2460 | { | |
2461 | for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++) | |
2462 | { | |
e1ab7874 | 2463 | hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i), |
2464 | GET_MODE (ASM_OPERANDS_INPUT (x, i)), | |
2465 | do_not_record_p, hash_arg_in_memory_p, | |
2466 | have_reg_qty, cb) | |
78d140c9 | 2467 | + hash_rtx_string |
e1ab7874 | 2468 | (ASM_OPERANDS_INPUT_CONSTRAINT (x, i))); |
d91f2122 | 2469 | } |
2470 | ||
78d140c9 | 2471 | hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0)); |
d91f2122 | 2472 | x = ASM_OPERANDS_INPUT (x, 0); |
2473 | mode = GET_MODE (x); | |
2474 | goto repeat; | |
2475 | } | |
2476 | ||
2477 | return hash; | |
2478 | } | |
0dbd1c74 | 2479 | break; |
cb10db9d | 2480 | |
0dbd1c74 | 2481 | default: |
2482 | break; | |
752df20e | 2483 | } |
2484 | ||
2485 | i = GET_RTX_LENGTH (code) - 1; | |
952bc06d | 2486 | hash += (unsigned) code + (unsigned) GET_MODE (x); |
752df20e | 2487 | fmt = GET_RTX_FORMAT (code); |
2488 | for (; i >= 0; i--) | |
2489 | { | |
cc636d56 | 2490 | switch (fmt[i]) |
752df20e | 2491 | { |
cc636d56 | 2492 | case 'e': |
752df20e | 2493 | /* If we are about to do the last recursive call |
2494 | needed at this level, change it into iteration. | |
2495 | This function is called enough to be worth it. */ | |
2496 | if (i == 0) | |
2497 | { | |
78d140c9 | 2498 | x = XEXP (x, i); |
752df20e | 2499 | goto repeat; |
2500 | } | |
48e1416a | 2501 | |
b9c74b4d | 2502 | hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p, |
e1ab7874 | 2503 | hash_arg_in_memory_p, |
2504 | have_reg_qty, cb); | |
cc636d56 | 2505 | break; |
78d140c9 | 2506 | |
cc636d56 | 2507 | case 'E': |
2508 | for (j = 0; j < XVECLEN (x, i); j++) | |
b9c74b4d | 2509 | hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p, |
e1ab7874 | 2510 | hash_arg_in_memory_p, |
2511 | have_reg_qty, cb); | |
cc636d56 | 2512 | break; |
78d140c9 | 2513 | |
cc636d56 | 2514 | case 's': |
2515 | hash += hash_rtx_string (XSTR (x, i)); | |
2516 | break; | |
2517 | ||
2518 | case 'i': | |
2519 | hash += (unsigned int) XINT (x, i); | |
2520 | break; | |
2521 | ||
2522 | case '0': case 't': | |
2523 | /* Unused. */ | |
2524 | break; | |
2525 | ||
2526 | default: | |
2527 | gcc_unreachable (); | |
2528 | } | |
752df20e | 2529 | } |
78d140c9 | 2530 | |
752df20e | 2531 | return hash; |
2532 | } | |
2533 | ||
e1ab7874 | 2534 | /* Hash an rtx. We are careful to make sure the value is never negative. |
2535 | Equivalent registers hash identically. | |
2536 | MODE is used in hashing for CONST_INTs only; | |
2537 | otherwise the mode of X is used. | |
2538 | ||
2539 | Store 1 in DO_NOT_RECORD_P if any subexpression is volatile. | |
2540 | ||
2541 | If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains | |
2e444733 | 2542 | a MEM rtx which does not have the MEM_READONLY_P flag set. |
e1ab7874 | 2543 | |
2544 | Note that cse_insn knows that the hash code of a MEM expression | |
2545 | is just (int) MEM plus the hash code of the address. */ | |
2546 | ||
2547 | unsigned | |
3754d046 | 2548 | hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p, |
e1ab7874 | 2549 | int *hash_arg_in_memory_p, bool have_reg_qty) |
2550 | { | |
2551 | return hash_rtx_cb (x, mode, do_not_record_p, | |
2552 | hash_arg_in_memory_p, have_reg_qty, NULL); | |
2553 | } | |
2554 | ||
78d140c9 | 2555 | /* Hash an rtx X for cse via hash_rtx. |
2556 | Stores 1 in do_not_record if any subexpression is volatile. | |
2557 | Stores 1 in hash_arg_in_memory if X contains a mem rtx which | |
2e444733 | 2558 | does not have the MEM_READONLY_P flag set. */ |
78d140c9 | 2559 | |
2560 | static inline unsigned | |
3754d046 | 2561 | canon_hash (rtx x, machine_mode mode) |
78d140c9 | 2562 | { |
2563 | return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true); | |
2564 | } | |
2565 | ||
2566 | /* Like canon_hash but with no side effects, i.e. do_not_record | |
2567 | and hash_arg_in_memory are not changed. */ | |
752df20e | 2568 | |
78d140c9 | 2569 | static inline unsigned |
3754d046 | 2570 | safe_hash (rtx x, machine_mode mode) |
752df20e | 2571 | { |
78d140c9 | 2572 | int dummy_do_not_record; |
2573 | return hash_rtx (x, mode, &dummy_do_not_record, NULL, true); | |
752df20e | 2574 | } |
2575 | \f | |
2576 | /* Return 1 iff X and Y would canonicalize into the same thing, | |
2577 | without actually constructing the canonicalization of either one. | |
2578 | If VALIDATE is nonzero, | |
2579 | we assume X is an expression being processed from the rtl | |
2580 | and Y was found in the hash table. We check register refs | |
2581 | in Y for being marked as valid. | |
2582 | ||
78d140c9 | 2583 | If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */ |
752df20e | 2584 | |
78d140c9 | 2585 | int |
52d07779 | 2586 | exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse) |
752df20e | 2587 | { |
19cb6b50 | 2588 | int i, j; |
2589 | enum rtx_code code; | |
2590 | const char *fmt; | |
752df20e | 2591 | |
2592 | /* Note: it is incorrect to assume an expression is equivalent to itself | |
2593 | if VALIDATE is nonzero. */ | |
2594 | if (x == y && !validate) | |
2595 | return 1; | |
78d140c9 | 2596 | |
752df20e | 2597 | if (x == 0 || y == 0) |
2598 | return x == y; | |
2599 | ||
2600 | code = GET_CODE (x); | |
2601 | if (code != GET_CODE (y)) | |
78d140c9 | 2602 | return 0; |
752df20e | 2603 | |
2604 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */ | |
2605 | if (GET_MODE (x) != GET_MODE (y)) | |
2606 | return 0; | |
2607 | ||
04ec15fa | 2608 | /* MEMs referring to different address space are not equivalent. */ |
bd1a81f7 | 2609 | if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y)) |
2610 | return 0; | |
2611 | ||
752df20e | 2612 | switch (code) |
2613 | { | |
2614 | case PC: | |
2615 | case CC0: | |
0349edce | 2616 | CASE_CONST_UNIQUE: |
73f5c1e3 | 2617 | return x == y; |
752df20e | 2618 | |
2619 | case LABEL_REF: | |
b49f2e4b | 2620 | return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y); |
752df20e | 2621 | |
d1a40e76 | 2622 | case SYMBOL_REF: |
2623 | return XSTR (x, 0) == XSTR (y, 0); | |
2624 | ||
752df20e | 2625 | case REG: |
78d140c9 | 2626 | if (for_gcse) |
2627 | return REGNO (x) == REGNO (y); | |
2628 | else | |
2629 | { | |
2630 | unsigned int regno = REGNO (y); | |
2631 | unsigned int i; | |
a2c6f0b7 | 2632 | unsigned int endregno = END_REGNO (y); |
752df20e | 2633 | |
78d140c9 | 2634 | /* If the quantities are not the same, the expressions are not |
2635 | equivalent. If there are and we are not to validate, they | |
2636 | are equivalent. Otherwise, ensure all regs are up-to-date. */ | |
752df20e | 2637 | |
78d140c9 | 2638 | if (REG_QTY (REGNO (x)) != REG_QTY (regno)) |
2639 | return 0; | |
2640 | ||
2641 | if (! validate) | |
2642 | return 1; | |
2643 | ||
2644 | for (i = regno; i < endregno; i++) | |
2645 | if (REG_IN_TABLE (i) != REG_TICK (i)) | |
2646 | return 0; | |
752df20e | 2647 | |
752df20e | 2648 | return 1; |
78d140c9 | 2649 | } |
752df20e | 2650 | |
78d140c9 | 2651 | case MEM: |
2652 | if (for_gcse) | |
2653 | { | |
78d140c9 | 2654 | /* A volatile mem should not be considered equivalent to any |
2655 | other. */ | |
2656 | if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y)) | |
2657 | return 0; | |
a79761ff | 2658 | |
2659 | /* Can't merge two expressions in different alias sets, since we | |
2660 | can decide that the expression is transparent in a block when | |
2661 | it isn't, due to it being set with the different alias set. | |
2662 | ||
2663 | Also, can't merge two expressions with different MEM_ATTRS. | |
2664 | They could e.g. be two different entities allocated into the | |
2665 | same space on the stack (see e.g. PR25130). In that case, the | |
2666 | MEM addresses can be the same, even though the two MEMs are | |
2667 | absolutely not equivalent. | |
2668 | ||
2669 | But because really all MEM attributes should be the same for | |
2670 | equivalent MEMs, we just use the invariant that MEMs that have | |
2671 | the same attributes share the same mem_attrs data structure. */ | |
7e304b71 | 2672 | if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y))) |
a79761ff | 2673 | return 0; |
4ee783fc | 2674 | |
2675 | /* If we are handling exceptions, we cannot consider two expressions | |
2676 | with different trapping status as equivalent, because simple_mem | |
2677 | might accept one and reject the other. */ | |
2678 | if (cfun->can_throw_non_call_exceptions | |
2679 | && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))) | |
2680 | return 0; | |
78d140c9 | 2681 | } |
2682 | break; | |
752df20e | 2683 | |
2684 | /* For commutative operations, check both orders. */ | |
2685 | case PLUS: | |
2686 | case MULT: | |
2687 | case AND: | |
2688 | case IOR: | |
2689 | case XOR: | |
2690 | case NE: | |
2691 | case EQ: | |
78d140c9 | 2692 | return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0), |
2693 | validate, for_gcse) | |
752df20e | 2694 | && exp_equiv_p (XEXP (x, 1), XEXP (y, 1), |
78d140c9 | 2695 | validate, for_gcse)) |
752df20e | 2696 | || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1), |
78d140c9 | 2697 | validate, for_gcse) |
752df20e | 2698 | && exp_equiv_p (XEXP (x, 1), XEXP (y, 0), |
78d140c9 | 2699 | validate, for_gcse))); |
cb10db9d | 2700 | |
d91f2122 | 2701 | case ASM_OPERANDS: |
2702 | /* We don't use the generic code below because we want to | |
2703 | disregard filename and line numbers. */ | |
2704 | ||
2705 | /* A volatile asm isn't equivalent to any other. */ | |
2706 | if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y)) | |
2707 | return 0; | |
2708 | ||
2709 | if (GET_MODE (x) != GET_MODE (y) | |
2710 | || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y)) | |
2711 | || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x), | |
2712 | ASM_OPERANDS_OUTPUT_CONSTRAINT (y)) | |
2713 | || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y) | |
2714 | || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y)) | |
2715 | return 0; | |
2716 | ||
2717 | if (ASM_OPERANDS_INPUT_LENGTH (x)) | |
2718 | { | |
2719 | for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) | |
2720 | if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i), | |
2721 | ASM_OPERANDS_INPUT (y, i), | |
78d140c9 | 2722 | validate, for_gcse) |
d91f2122 | 2723 | || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i), |
2724 | ASM_OPERANDS_INPUT_CONSTRAINT (y, i))) | |
2725 | return 0; | |
2726 | } | |
2727 | ||
2728 | return 1; | |
2729 | ||
0dbd1c74 | 2730 | default: |
2731 | break; | |
752df20e | 2732 | } |
2733 | ||
2734 | /* Compare the elements. If any pair of corresponding elements | |
78d140c9 | 2735 | fail to match, return 0 for the whole thing. */ |
752df20e | 2736 | |
2737 | fmt = GET_RTX_FORMAT (code); | |
2738 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2739 | { | |
b572011e | 2740 | switch (fmt[i]) |
752df20e | 2741 | { |
b572011e | 2742 | case 'e': |
78d140c9 | 2743 | if (! exp_equiv_p (XEXP (x, i), XEXP (y, i), |
2744 | validate, for_gcse)) | |
752df20e | 2745 | return 0; |
b572011e | 2746 | break; |
2747 | ||
2748 | case 'E': | |
752df20e | 2749 | if (XVECLEN (x, i) != XVECLEN (y, i)) |
2750 | return 0; | |
2751 | for (j = 0; j < XVECLEN (x, i); j++) | |
2752 | if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j), | |
78d140c9 | 2753 | validate, for_gcse)) |
752df20e | 2754 | return 0; |
b572011e | 2755 | break; |
2756 | ||
2757 | case 's': | |
752df20e | 2758 | if (strcmp (XSTR (x, i), XSTR (y, i))) |
2759 | return 0; | |
b572011e | 2760 | break; |
2761 | ||
2762 | case 'i': | |
752df20e | 2763 | if (XINT (x, i) != XINT (y, i)) |
2764 | return 0; | |
b572011e | 2765 | break; |
2766 | ||
2767 | case 'w': | |
2768 | if (XWINT (x, i) != XWINT (y, i)) | |
2769 | return 0; | |
cb10db9d | 2770 | break; |
b572011e | 2771 | |
2772 | case '0': | |
a4070a91 | 2773 | case 't': |
b572011e | 2774 | break; |
2775 | ||
2776 | default: | |
cc636d56 | 2777 | gcc_unreachable (); |
752df20e | 2778 | } |
cb10db9d | 2779 | } |
b572011e | 2780 | |
752df20e | 2781 | return 1; |
2782 | } | |
2783 | \f | |
1cc37766 | 2784 | /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate |
2785 | the result if necessary. INSN is as for canon_reg. */ | |
2786 | ||
2787 | static void | |
47f1d198 | 2788 | validate_canon_reg (rtx *xloc, rtx_insn *insn) |
1cc37766 | 2789 | { |
3072d30e | 2790 | if (*xloc) |
2791 | { | |
d328ebdf | 2792 | rtx new_rtx = canon_reg (*xloc, insn); |
1cc37766 | 2793 | |
3072d30e | 2794 | /* If replacing pseudo with hard reg or vice versa, ensure the |
2795 | insn remains valid. Likewise if the insn has MATCH_DUPs. */ | |
d328ebdf | 2796 | gcc_assert (insn && new_rtx); |
2797 | validate_change (insn, xloc, new_rtx, 1); | |
3072d30e | 2798 | } |
1cc37766 | 2799 | } |
2800 | ||
752df20e | 2801 | /* Canonicalize an expression: |
2802 | replace each register reference inside it | |
2803 | with the "oldest" equivalent register. | |
2804 | ||
0c0acbaa | 2805 | If INSN is nonzero validate_change is used to ensure that INSN remains valid |
d10cfa8d | 2806 | after we make our substitution. The calls are made with IN_GROUP nonzero |
8d5dd220 | 2807 | so apply_change_group must be called upon the outermost return from this |
2808 | function (unless INSN is zero). The result of apply_change_group can | |
2809 | generally be discarded since the changes we are making are optional. */ | |
752df20e | 2810 | |
2811 | static rtx | |
47f1d198 | 2812 | canon_reg (rtx x, rtx_insn *insn) |
752df20e | 2813 | { |
19cb6b50 | 2814 | int i; |
2815 | enum rtx_code code; | |
2816 | const char *fmt; | |
752df20e | 2817 | |
2818 | if (x == 0) | |
2819 | return x; | |
2820 | ||
2821 | code = GET_CODE (x); | |
2822 | switch (code) | |
2823 | { | |
2824 | case PC: | |
2825 | case CC0: | |
2826 | case CONST: | |
0349edce | 2827 | CASE_CONST_ANY: |
752df20e | 2828 | case SYMBOL_REF: |
2829 | case LABEL_REF: | |
2830 | case ADDR_VEC: | |
2831 | case ADDR_DIFF_VEC: | |
2832 | return x; | |
2833 | ||
2834 | case REG: | |
2835 | { | |
19cb6b50 | 2836 | int first; |
2837 | int q; | |
2838 | struct qty_table_elem *ent; | |
752df20e | 2839 | |
2840 | /* Never replace a hard reg, because hard regs can appear | |
2841 | in more than one machine mode, and we must preserve the mode | |
2842 | of each occurrence. Also, some hard regs appear in | |
2843 | MEMs that are shared and mustn't be altered. Don't try to | |
2844 | replace any reg that maps to a reg of class NO_REGS. */ | |
2845 | if (REGNO (x) < FIRST_PSEUDO_REGISTER | |
2846 | || ! REGNO_QTY_VALID_P (REGNO (x))) | |
2847 | return x; | |
2848 | ||
cb10db9d | 2849 | q = REG_QTY (REGNO (x)); |
a7f3b1c7 | 2850 | ent = &qty_table[q]; |
2851 | first = ent->first_reg; | |
752df20e | 2852 | return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first] |
2853 | : REGNO_REG_CLASS (first) == NO_REGS ? x | |
a7f3b1c7 | 2854 | : gen_rtx_REG (ent->mode, first)); |
752df20e | 2855 | } |
cb10db9d | 2856 | |
0dbd1c74 | 2857 | default: |
2858 | break; | |
752df20e | 2859 | } |
2860 | ||
2861 | fmt = GET_RTX_FORMAT (code); | |
2862 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2863 | { | |
19cb6b50 | 2864 | int j; |
752df20e | 2865 | |
2866 | if (fmt[i] == 'e') | |
1cc37766 | 2867 | validate_canon_reg (&XEXP (x, i), insn); |
752df20e | 2868 | else if (fmt[i] == 'E') |
2869 | for (j = 0; j < XVECLEN (x, i); j++) | |
1cc37766 | 2870 | validate_canon_reg (&XVECEXP (x, i, j), insn); |
752df20e | 2871 | } |
2872 | ||
2873 | return x; | |
2874 | } | |
2875 | \f | |
6a8939cc | 2876 | /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison |
2877 | operation (EQ, NE, GT, etc.), follow it back through the hash table and | |
2878 | what values are being compared. | |
9ce37dcf | 2879 | |
6a8939cc | 2880 | *PARG1 and *PARG2 are updated to contain the rtx representing the values |
2881 | actually being compared. For example, if *PARG1 was (cc0) and *PARG2 | |
2882 | was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were | |
2883 | compared to produce cc0. | |
61b1f5a7 | 2884 | |
6a8939cc | 2885 | The return value is the comparison operator and is either the code of |
2886 | A or the code corresponding to the inverse of the comparison. */ | |
752df20e | 2887 | |
af21a202 | 2888 | static enum rtx_code |
8ec3a57b | 2889 | find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2, |
3754d046 | 2890 | machine_mode *pmode1, machine_mode *pmode2) |
752df20e | 2891 | { |
af21a202 | 2892 | rtx arg1, arg2; |
431205b7 | 2893 | hash_set<rtx> *visited = NULL; |
7d8df2ae | 2894 | /* Set nonzero when we find something of interest. */ |
2895 | rtx x = NULL; | |
9ce37dcf | 2896 | |
af21a202 | 2897 | arg1 = *parg1, arg2 = *parg2; |
752df20e | 2898 | |
af21a202 | 2899 | /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */ |
752df20e | 2900 | |
af21a202 | 2901 | while (arg2 == CONST0_RTX (GET_MODE (arg1))) |
61b1f5a7 | 2902 | { |
af21a202 | 2903 | int reverse_code = 0; |
2904 | struct table_elt *p = 0; | |
308a5ff6 | 2905 | |
7d8df2ae | 2906 | /* Remember state from previous iteration. */ |
2907 | if (x) | |
2908 | { | |
2909 | if (!visited) | |
431205b7 | 2910 | visited = new hash_set<rtx>; |
2911 | visited->add (x); | |
7d8df2ae | 2912 | x = 0; |
2913 | } | |
2914 | ||
af21a202 | 2915 | /* If arg1 is a COMPARE, extract the comparison arguments from it. |
2916 | On machines with CC0, this is the only case that can occur, since | |
2917 | fold_rtx will return the COMPARE or item being compared with zero | |
2918 | when given CC0. */ | |
308a5ff6 | 2919 | |
af21a202 | 2920 | if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx) |
2921 | x = arg1; | |
308a5ff6 | 2922 | |
af21a202 | 2923 | /* If ARG1 is a comparison operator and CODE is testing for |
2924 | STORE_FLAG_VALUE, get the inner arguments. */ | |
61b1f5a7 | 2925 | |
6720e96c | 2926 | else if (COMPARISON_P (arg1)) |
752df20e | 2927 | { |
aa870c1b | 2928 | #ifdef FLOAT_STORE_FLAG_VALUE |
2929 | REAL_VALUE_TYPE fsfv; | |
2930 | #endif | |
2931 | ||
af21a202 | 2932 | if (code == NE |
2933 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT | |
2934 | && code == LT && STORE_FLAG_VALUE == -1) | |
2935 | #ifdef FLOAT_STORE_FLAG_VALUE | |
95204692 | 2936 | || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1)) |
aa870c1b | 2937 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
2938 | REAL_VALUE_NEGATIVE (fsfv))) | |
752df20e | 2939 | #endif |
61b1f5a7 | 2940 | ) |
af21a202 | 2941 | x = arg1; |
2942 | else if (code == EQ | |
2943 | || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT | |
2944 | && code == GE && STORE_FLAG_VALUE == -1) | |
2945 | #ifdef FLOAT_STORE_FLAG_VALUE | |
95204692 | 2946 | || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1)) |
aa870c1b | 2947 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
2948 | REAL_VALUE_NEGATIVE (fsfv))) | |
af21a202 | 2949 | #endif |
2950 | ) | |
2951 | x = arg1, reverse_code = 1; | |
752df20e | 2952 | } |
2953 | ||
af21a202 | 2954 | /* ??? We could also check for |
752df20e | 2955 | |
af21a202 | 2956 | (ne (and (eq (...) (const_int 1))) (const_int 0)) |
752df20e | 2957 | |
af21a202 | 2958 | and related forms, but let's wait until we see them occurring. */ |
752df20e | 2959 | |
af21a202 | 2960 | if (x == 0) |
2961 | /* Look up ARG1 in the hash table and see if it has an equivalence | |
2962 | that lets us see what is being compared. */ | |
78d140c9 | 2963 | p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1)); |
cb10db9d | 2964 | if (p) |
e9a91a9e | 2965 | { |
2966 | p = p->first_same_value; | |
2967 | ||
2968 | /* If what we compare is already known to be constant, that is as | |
2969 | good as it gets. | |
2970 | We need to break the loop in this case, because otherwise we | |
2971 | can have an infinite loop when looking at a reg that is known | |
2972 | to be a constant which is the same as a comparison of a reg | |
2973 | against zero which appears later in the insn stream, which in | |
2974 | turn is constant and the same as the comparison of the first reg | |
2975 | against zero... */ | |
2976 | if (p->is_const) | |
2977 | break; | |
2978 | } | |
752df20e | 2979 | |
af21a202 | 2980 | for (; p; p = p->next_same_value) |
752df20e | 2981 | { |
3754d046 | 2982 | machine_mode inner_mode = GET_MODE (p->exp); |
aa870c1b | 2983 | #ifdef FLOAT_STORE_FLAG_VALUE |
2984 | REAL_VALUE_TYPE fsfv; | |
2985 | #endif | |
752df20e | 2986 | |
af21a202 | 2987 | /* If the entry isn't valid, skip it. */ |
78d140c9 | 2988 | if (! exp_equiv_p (p->exp, p->exp, 1, false)) |
af21a202 | 2989 | continue; |
51356f86 | 2990 | |
7d8df2ae | 2991 | /* If it's a comparison we've used before, skip it. */ |
431205b7 | 2992 | if (visited && visited->contains (p->exp)) |
7a49726a | 2993 | continue; |
2994 | ||
6a8939cc | 2995 | if (GET_CODE (p->exp) == COMPARE |
2996 | /* Another possibility is that this machine has a compare insn | |
2997 | that includes the comparison code. In that case, ARG1 would | |
2998 | be equivalent to a comparison operation that would set ARG1 to | |
2999 | either STORE_FLAG_VALUE or zero. If this is an NE operation, | |
3000 | ORIG_CODE is the actual comparison being done; if it is an EQ, | |
3001 | we must reverse ORIG_CODE. On machine with a negative value | |
3002 | for STORE_FLAG_VALUE, also look at LT and GE operations. */ | |
3003 | || ((code == NE | |
3004 | || (code == LT | |
f92430e0 | 3005 | && val_signbit_known_set_p (inner_mode, |
3006 | STORE_FLAG_VALUE)) | |
af21a202 | 3007 | #ifdef FLOAT_STORE_FLAG_VALUE |
6a8939cc | 3008 | || (code == LT |
cee7491d | 3009 | && SCALAR_FLOAT_MODE_P (inner_mode) |
aa870c1b | 3010 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
3011 | REAL_VALUE_NEGATIVE (fsfv))) | |
af21a202 | 3012 | #endif |
6a8939cc | 3013 | ) |
6720e96c | 3014 | && COMPARISON_P (p->exp))) |
752df20e | 3015 | { |
af21a202 | 3016 | x = p->exp; |
3017 | break; | |
3018 | } | |
3019 | else if ((code == EQ | |
3020 | || (code == GE | |
f92430e0 | 3021 | && val_signbit_known_set_p (inner_mode, |
3022 | STORE_FLAG_VALUE)) | |
af21a202 | 3023 | #ifdef FLOAT_STORE_FLAG_VALUE |
3024 | || (code == GE | |
cee7491d | 3025 | && SCALAR_FLOAT_MODE_P (inner_mode) |
aa870c1b | 3026 | && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)), |
3027 | REAL_VALUE_NEGATIVE (fsfv))) | |
af21a202 | 3028 | #endif |
3029 | ) | |
6720e96c | 3030 | && COMPARISON_P (p->exp)) |
af21a202 | 3031 | { |
3032 | reverse_code = 1; | |
3033 | x = p->exp; | |
3034 | break; | |
752df20e | 3035 | } |
3036 | ||
805e22b2 | 3037 | /* If this non-trapping address, e.g. fp + constant, the |
3038 | equivalent is a better operand since it may let us predict | |
3039 | the value of the comparison. */ | |
3040 | else if (!rtx_addr_can_trap_p (p->exp)) | |
af21a202 | 3041 | { |
3042 | arg1 = p->exp; | |
3043 | continue; | |
3044 | } | |
752df20e | 3045 | } |
752df20e | 3046 | |
af21a202 | 3047 | /* If we didn't find a useful equivalence for ARG1, we are done. |
3048 | Otherwise, set up for the next iteration. */ | |
3049 | if (x == 0) | |
3050 | break; | |
752df20e | 3051 | |
47ae02b7 | 3052 | /* If we need to reverse the comparison, make sure that is |
6d1304b6 | 3053 | possible -- we can't necessarily infer the value of GE from LT |
3054 | with floating-point operands. */ | |
af21a202 | 3055 | if (reverse_code) |
7da6ea0c | 3056 | { |
3057 | enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX); | |
3058 | if (reversed == UNKNOWN) | |
3059 | break; | |
d4c5e26d | 3060 | else |
3061 | code = reversed; | |
7da6ea0c | 3062 | } |
6720e96c | 3063 | else if (COMPARISON_P (x)) |
7da6ea0c | 3064 | code = GET_CODE (x); |
3065 | arg1 = XEXP (x, 0), arg2 = XEXP (x, 1); | |
752df20e | 3066 | } |
3067 | ||
af21a202 | 3068 | /* Return our results. Return the modes from before fold_rtx |
3069 | because fold_rtx might produce const_int, and then it's too late. */ | |
3070 | *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2); | |
3071 | *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0); | |
3072 | ||
7d8df2ae | 3073 | if (visited) |
431205b7 | 3074 | delete visited; |
af21a202 | 3075 | return code; |
752df20e | 3076 | } |
3077 | \f | |
42a3a38b | 3078 | /* If X is a nontrivial arithmetic operation on an argument for which |
3079 | a constant value can be determined, return the result of operating | |
3080 | on that value, as a constant. Otherwise, return X, possibly with | |
3081 | one or more operands changed to a forward-propagated constant. | |
18b14db6 | 3082 | |
42a3a38b | 3083 | If X is a register whose contents are known, we do NOT return |
3084 | those contents here; equiv_constant is called to perform that task. | |
3085 | For SUBREGs and MEMs, we do that both here and in equiv_constant. | |
752df20e | 3086 | |
3087 | INSN is the insn that we may be modifying. If it is 0, make a copy | |
3088 | of X before modifying it. */ | |
3089 | ||
3090 | static rtx | |
47f1d198 | 3091 | fold_rtx (rtx x, rtx_insn *insn) |
752df20e | 3092 | { |
19cb6b50 | 3093 | enum rtx_code code; |
3754d046 | 3094 | machine_mode mode; |
19cb6b50 | 3095 | const char *fmt; |
3096 | int i; | |
d328ebdf | 3097 | rtx new_rtx = 0; |
42a3a38b | 3098 | int changed = 0; |
752df20e | 3099 | |
42a3a38b | 3100 | /* Operands of X. */ |
a561ec10 | 3101 | /* Workaround -Wmaybe-uninitialized false positive during |
3102 | profiledbootstrap by initializing them. */ | |
3103 | rtx folded_arg0 = NULL_RTX; | |
3104 | rtx folded_arg1 = NULL_RTX; | |
752df20e | 3105 | |
3106 | /* Constant equivalents of first three operands of X; | |
3107 | 0 when no such equivalent is known. */ | |
3108 | rtx const_arg0; | |
3109 | rtx const_arg1; | |
3110 | rtx const_arg2; | |
3111 | ||
3112 | /* The mode of the first operand of X. We need this for sign and zero | |
3113 | extends. */ | |
3754d046 | 3114 | machine_mode mode_arg0; |
752df20e | 3115 | |
3116 | if (x == 0) | |
3117 | return x; | |
3118 | ||
42a3a38b | 3119 | /* Try to perform some initial simplifications on X. */ |
752df20e | 3120 | code = GET_CODE (x); |
3121 | switch (code) | |
3122 | { | |
42a3a38b | 3123 | case MEM: |
3124 | case SUBREG: | |
b803a3c1 | 3125 | /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning |
3126 | than it would in other contexts. Basically its mode does not | |
3127 | signify the size of the object read. That information is carried | |
3128 | by size operand. If we happen to have a MEM of the appropriate | |
3129 | mode in our tables with a constant value we could simplify the | |
3130 | extraction incorrectly if we allowed substitution of that value | |
3131 | for the MEM. */ | |
3132 | case ZERO_EXTRACT: | |
3133 | case SIGN_EXTRACT: | |
d328ebdf | 3134 | if ((new_rtx = equiv_constant (x)) != NULL_RTX) |
3135 | return new_rtx; | |
42a3a38b | 3136 | return x; |
3137 | ||
752df20e | 3138 | case CONST: |
0349edce | 3139 | CASE_CONST_ANY: |
752df20e | 3140 | case SYMBOL_REF: |
3141 | case LABEL_REF: | |
3142 | case REG: | |
97108156 | 3143 | case PC: |
752df20e | 3144 | /* No use simplifying an EXPR_LIST |
3145 | since they are used only for lists of args | |
3146 | in a function call's REG_EQUAL note. */ | |
3147 | case EXPR_LIST: | |
3148 | return x; | |
3149 | ||
752df20e | 3150 | case CC0: |
3151 | return prev_insn_cc0; | |
752df20e | 3152 | |
c97a7837 | 3153 | case ASM_OPERANDS: |
d239a9ad | 3154 | if (insn) |
3155 | { | |
3156 | for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) | |
3157 | validate_change (insn, &ASM_OPERANDS_INPUT (x, i), | |
3158 | fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0); | |
3159 | } | |
42a3a38b | 3160 | return x; |
3161 | ||
42a3a38b | 3162 | case CALL: |
93516111 | 3163 | if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0))) |
42a3a38b | 3164 | return x; |
c97a7837 | 3165 | break; |
cb10db9d | 3166 | |
42a3a38b | 3167 | /* Anything else goes through the loop below. */ |
0dbd1c74 | 3168 | default: |
3169 | break; | |
752df20e | 3170 | } |
3171 | ||
42a3a38b | 3172 | mode = GET_MODE (x); |
752df20e | 3173 | const_arg0 = 0; |
3174 | const_arg1 = 0; | |
3175 | const_arg2 = 0; | |
3176 | mode_arg0 = VOIDmode; | |
3177 | ||
3178 | /* Try folding our operands. | |
3179 | Then see which ones have constant values known. */ | |
3180 | ||
3181 | fmt = GET_RTX_FORMAT (code); | |
3182 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3183 | if (fmt[i] == 'e') | |
3184 | { | |
42a3a38b | 3185 | rtx folded_arg = XEXP (x, i), const_arg; |
3754d046 | 3186 | machine_mode mode_arg = GET_MODE (folded_arg); |
06320855 | 3187 | |
3188 | switch (GET_CODE (folded_arg)) | |
3189 | { | |
3190 | case MEM: | |
3191 | case REG: | |
3192 | case SUBREG: | |
3193 | const_arg = equiv_constant (folded_arg); | |
3194 | break; | |
3195 | ||
3196 | case CONST: | |
0349edce | 3197 | CASE_CONST_ANY: |
06320855 | 3198 | case SYMBOL_REF: |
3199 | case LABEL_REF: | |
06320855 | 3200 | const_arg = folded_arg; |
3201 | break; | |
3202 | ||
06320855 | 3203 | case CC0: |
77cb85b2 | 3204 | /* The cc0-user and cc0-setter may be in different blocks if |
3205 | the cc0-setter potentially traps. In that case PREV_INSN_CC0 | |
3206 | will have been cleared as we exited the block with the | |
3207 | setter. | |
3208 | ||
3209 | While we could potentially track cc0 in this case, it just | |
3210 | doesn't seem to be worth it given that cc0 targets are not | |
3211 | terribly common or important these days and trapping math | |
3212 | is rarely used. The combination of those two conditions | |
3213 | necessary to trip this situation is exceedingly rare in the | |
3214 | real world. */ | |
3215 | if (!prev_insn_cc0) | |
3216 | { | |
3217 | const_arg = NULL_RTX; | |
3218 | } | |
3219 | else | |
3220 | { | |
3221 | folded_arg = prev_insn_cc0; | |
3222 | mode_arg = prev_insn_cc0_mode; | |
3223 | const_arg = equiv_constant (folded_arg); | |
3224 | } | |
06320855 | 3225 | break; |
06320855 | 3226 | |
3227 | default: | |
3228 | folded_arg = fold_rtx (folded_arg, insn); | |
3229 | const_arg = equiv_constant (folded_arg); | |
3230 | break; | |
3231 | } | |
752df20e | 3232 | |
3233 | /* For the first three operands, see if the operand | |
3234 | is constant or equivalent to a constant. */ | |
3235 | switch (i) | |
3236 | { | |
3237 | case 0: | |
3238 | folded_arg0 = folded_arg; | |
3239 | const_arg0 = const_arg; | |
3240 | mode_arg0 = mode_arg; | |
3241 | break; | |
3242 | case 1: | |
3243 | folded_arg1 = folded_arg; | |
3244 | const_arg1 = const_arg; | |
3245 | break; | |
3246 | case 2: | |
3247 | const_arg2 = const_arg; | |
3248 | break; | |
3249 | } | |
3250 | ||
42a3a38b | 3251 | /* Pick the least expensive of the argument and an equivalent constant |
3252 | argument. */ | |
3253 | if (const_arg != 0 | |
3254 | && const_arg != folded_arg | |
5ae4887d | 3255 | && (COST_IN (const_arg, mode_arg, code, i) |
3256 | <= COST_IN (folded_arg, mode_arg, code, i)) | |
f35e401c | 3257 | |
8f1e01cb | 3258 | /* It's not safe to substitute the operand of a conversion |
3259 | operator with a constant, as the conversion's identity | |
fe24f256 | 3260 | depends upon the mode of its operand. This optimization |
8f1e01cb | 3261 | is handled by the call to simplify_unary_operation. */ |
42a3a38b | 3262 | && (GET_RTX_CLASS (code) != RTX_UNARY |
3263 | || GET_MODE (const_arg) == mode_arg0 | |
3264 | || (code != ZERO_EXTEND | |
3265 | && code != SIGN_EXTEND | |
3266 | && code != TRUNCATE | |
3267 | && code != FLOAT_TRUNCATE | |
3268 | && code != FLOAT_EXTEND | |
3269 | && code != FLOAT | |
3270 | && code != FIX | |
3271 | && code != UNSIGNED_FLOAT | |
3272 | && code != UNSIGNED_FIX))) | |
3273 | folded_arg = const_arg; | |
3274 | ||
3275 | if (folded_arg == XEXP (x, i)) | |
3276 | continue; | |
752df20e | 3277 | |
42a3a38b | 3278 | if (insn == NULL_RTX && !changed) |
3279 | x = copy_rtx (x); | |
3280 | changed = 1; | |
4f34fbd6 | 3281 | validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1); |
c08e043f | 3282 | } |
752df20e | 3283 | |
42a3a38b | 3284 | if (changed) |
752df20e | 3285 | { |
42a3a38b | 3286 | /* Canonicalize X if necessary, and keep const_argN and folded_argN |
3287 | consistent with the order in X. */ | |
3288 | if (canonicalize_change_group (insn, x)) | |
752df20e | 3289 | { |
c586a5e3 | 3290 | std::swap (const_arg0, const_arg1); |
3291 | std::swap (folded_arg0, folded_arg1); | |
752df20e | 3292 | } |
42a3a38b | 3293 | |
3294 | apply_change_group (); | |
752df20e | 3295 | } |
3296 | ||
3297 | /* If X is an arithmetic operation, see if we can simplify it. */ | |
3298 | ||
3299 | switch (GET_RTX_CLASS (code)) | |
3300 | { | |
6720e96c | 3301 | case RTX_UNARY: |
528b0df8 | 3302 | { |
528b0df8 | 3303 | /* We can't simplify extension ops unless we know the |
3304 | original mode. */ | |
3305 | if ((code == ZERO_EXTEND || code == SIGN_EXTEND) | |
3306 | && mode_arg0 == VOIDmode) | |
3307 | break; | |
3308 | ||
d328ebdf | 3309 | new_rtx = simplify_unary_operation (code, mode, |
2bde5b8e | 3310 | const_arg0 ? const_arg0 : folded_arg0, |
3311 | mode_arg0); | |
528b0df8 | 3312 | } |
752df20e | 3313 | break; |
cb10db9d | 3314 | |
6720e96c | 3315 | case RTX_COMPARE: |
3316 | case RTX_COMM_COMPARE: | |
752df20e | 3317 | /* See what items are actually being compared and set FOLDED_ARG[01] |
3318 | to those values and CODE to the actual comparison code. If any are | |
3319 | constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't | |
3320 | do anything if both operands are already known to be constant. */ | |
3321 | ||
5b2d8298 | 3322 | /* ??? Vector mode comparisons are not supported yet. */ |
3323 | if (VECTOR_MODE_P (mode)) | |
3324 | break; | |
3325 | ||
752df20e | 3326 | if (const_arg0 == 0 || const_arg1 == 0) |
3327 | { | |
3328 | struct table_elt *p0, *p1; | |
ecb6ee6d | 3329 | rtx true_rtx, false_rtx; |
3754d046 | 3330 | machine_mode mode_arg1; |
50cf1c21 | 3331 | |
95204692 | 3332 | if (SCALAR_FLOAT_MODE_P (mode)) |
50cf1c21 | 3333 | { |
ecb6ee6d | 3334 | #ifdef FLOAT_STORE_FLAG_VALUE |
d5f9611d | 3335 | true_rtx = (const_double_from_real_value |
d4c5e26d | 3336 | (FLOAT_STORE_FLAG_VALUE (mode), mode)); |
ecb6ee6d | 3337 | #else |
3338 | true_rtx = NULL_RTX; | |
3339 | #endif | |
9c811526 | 3340 | false_rtx = CONST0_RTX (mode); |
50cf1c21 | 3341 | } |
ecb6ee6d | 3342 | else |
3343 | { | |
3344 | true_rtx = const_true_rtx; | |
3345 | false_rtx = const0_rtx; | |
3346 | } | |
752df20e | 3347 | |
5c4c31e3 | 3348 | code = find_comparison_args (code, &folded_arg0, &folded_arg1, |
3349 | &mode_arg0, &mode_arg1); | |
752df20e | 3350 | |
5c4c31e3 | 3351 | /* If the mode is VOIDmode or a MODE_CC mode, we don't know |
3352 | what kinds of things are being compared, so we can't do | |
3353 | anything with this comparison. */ | |
752df20e | 3354 | |
3355 | if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC) | |
3356 | break; | |
3357 | ||
58a1adea | 3358 | const_arg0 = equiv_constant (folded_arg0); |
3359 | const_arg1 = equiv_constant (folded_arg1); | |
3360 | ||
a92771b8 | 3361 | /* If we do not now have two constants being compared, see |
3362 | if we can nevertheless deduce some things about the | |
3363 | comparison. */ | |
752df20e | 3364 | if (const_arg0 == 0 || const_arg1 == 0) |
3365 | { | |
9d3874a6 | 3366 | if (const_arg1 != NULL) |
3367 | { | |
3368 | rtx cheapest_simplification; | |
3369 | int cheapest_cost; | |
3370 | rtx simp_result; | |
3371 | struct table_elt *p; | |
3372 | ||
3373 | /* See if we can find an equivalent of folded_arg0 | |
3374 | that gets us a cheaper expression, possibly a | |
3375 | constant through simplifications. */ | |
3376 | p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0), | |
3377 | mode_arg0); | |
48e1416a | 3378 | |
9d3874a6 | 3379 | if (p != NULL) |
3380 | { | |
3381 | cheapest_simplification = x; | |
5ae4887d | 3382 | cheapest_cost = COST (x, mode); |
9d3874a6 | 3383 | |
3384 | for (p = p->first_same_value; p != NULL; p = p->next_same_value) | |
3385 | { | |
3386 | int cost; | |
3387 | ||
3388 | /* If the entry isn't valid, skip it. */ | |
3389 | if (! exp_equiv_p (p->exp, p->exp, 1, false)) | |
3390 | continue; | |
3391 | ||
3392 | /* Try to simplify using this equivalence. */ | |
3393 | simp_result | |
3394 | = simplify_relational_operation (code, mode, | |
3395 | mode_arg0, | |
3396 | p->exp, | |
3397 | const_arg1); | |
3398 | ||
3399 | if (simp_result == NULL) | |
3400 | continue; | |
3401 | ||
5ae4887d | 3402 | cost = COST (simp_result, mode); |
9d3874a6 | 3403 | if (cost < cheapest_cost) |
3404 | { | |
3405 | cheapest_cost = cost; | |
3406 | cheapest_simplification = simp_result; | |
3407 | } | |
3408 | } | |
3409 | ||
3410 | /* If we have a cheaper expression now, use that | |
3411 | and try folding it further, from the top. */ | |
3412 | if (cheapest_simplification != x) | |
045ed337 | 3413 | return fold_rtx (copy_rtx (cheapest_simplification), |
3414 | insn); | |
9d3874a6 | 3415 | } |
3416 | } | |
3417 | ||
03a563f6 | 3418 | /* See if the two operands are the same. */ |
3419 | ||
3bac3cce | 3420 | if ((REG_P (folded_arg0) |
3421 | && REG_P (folded_arg1) | |
3422 | && (REG_QTY (REGNO (folded_arg0)) | |
3423 | == REG_QTY (REGNO (folded_arg1)))) | |
03a563f6 | 3424 | || ((p0 = lookup (folded_arg0, |
78d140c9 | 3425 | SAFE_HASH (folded_arg0, mode_arg0), |
3426 | mode_arg0)) | |
03a563f6 | 3427 | && (p1 = lookup (folded_arg1, |
78d140c9 | 3428 | SAFE_HASH (folded_arg1, mode_arg0), |
3429 | mode_arg0)) | |
03a563f6 | 3430 | && p0->first_same_value == p1->first_same_value)) |
3bac3cce | 3431 | folded_arg1 = folded_arg0; |
752df20e | 3432 | |
3433 | /* If FOLDED_ARG0 is a register, see if the comparison we are | |
3434 | doing now is either the same as we did before or the reverse | |
3435 | (we only check the reverse if not floating-point). */ | |
8ad4c111 | 3436 | else if (REG_P (folded_arg0)) |
752df20e | 3437 | { |
d1264606 | 3438 | int qty = REG_QTY (REGNO (folded_arg0)); |
752df20e | 3439 | |
a7f3b1c7 | 3440 | if (REGNO_QTY_VALID_P (REGNO (folded_arg0))) |
3441 | { | |
3442 | struct qty_table_elem *ent = &qty_table[qty]; | |
3443 | ||
3444 | if ((comparison_dominates_p (ent->comparison_code, code) | |
a4110d9a | 3445 | || (! FLOAT_MODE_P (mode_arg0) |
3446 | && comparison_dominates_p (ent->comparison_code, | |
3447 | reverse_condition (code)))) | |
a7f3b1c7 | 3448 | && (rtx_equal_p (ent->comparison_const, folded_arg1) |
3449 | || (const_arg1 | |
3450 | && rtx_equal_p (ent->comparison_const, | |
3451 | const_arg1)) | |
8ad4c111 | 3452 | || (REG_P (folded_arg1) |
a7f3b1c7 | 3453 | && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty)))) |
ecb6ee6d | 3454 | { |
3455 | if (comparison_dominates_p (ent->comparison_code, code)) | |
3456 | { | |
3457 | if (true_rtx) | |
3458 | return true_rtx; | |
3459 | else | |
3460 | break; | |
3461 | } | |
3462 | else | |
3463 | return false_rtx; | |
3464 | } | |
a7f3b1c7 | 3465 | } |
752df20e | 3466 | } |
3467 | } | |
3468 | } | |
3469 | ||
3470 | /* If we are comparing against zero, see if the first operand is | |
3471 | equivalent to an IOR with a constant. If so, we may be able to | |
3472 | determine the result of this comparison. */ | |
3bac3cce | 3473 | if (const_arg1 == const0_rtx && !const_arg0) |
752df20e | 3474 | { |
3475 | rtx y = lookup_as_function (folded_arg0, IOR); | |
3476 | rtx inner_const; | |
3477 | ||
3478 | if (y != 0 | |
3479 | && (inner_const = equiv_constant (XEXP (y, 1))) != 0 | |
971ba038 | 3480 | && CONST_INT_P (inner_const) |
752df20e | 3481 | && INTVAL (inner_const) != 0) |
3bac3cce | 3482 | folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const); |
752df20e | 3483 | } |
3484 | ||
ac503e50 | 3485 | { |
b9b50b55 | 3486 | rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0); |
3487 | rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1); | |
3488 | new_rtx = simplify_relational_operation (code, mode, mode_arg0, | |
3489 | op0, op1); | |
ac503e50 | 3490 | } |
752df20e | 3491 | break; |
3492 | ||
6720e96c | 3493 | case RTX_BIN_ARITH: |
3494 | case RTX_COMM_ARITH: | |
752df20e | 3495 | switch (code) |
3496 | { | |
3497 | case PLUS: | |
3498 | /* If the second operand is a LABEL_REF, see if the first is a MINUS | |
3499 | with that LABEL_REF as its second operand. If so, the result is | |
3500 | the first operand of that MINUS. This handles switches with an | |
3501 | ADDR_DIFF_VEC table. */ | |
3502 | if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF) | |
3503 | { | |
e6d1f05b | 3504 | rtx y |
3505 | = GET_CODE (folded_arg0) == MINUS ? folded_arg0 | |
b74befc5 | 3506 | : lookup_as_function (folded_arg0, MINUS); |
752df20e | 3507 | |
3508 | if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF | |
b49f2e4b | 3509 | && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg1)) |
752df20e | 3510 | return XEXP (y, 0); |
528b0df8 | 3511 | |
3512 | /* Now try for a CONST of a MINUS like the above. */ | |
e6d1f05b | 3513 | if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0 |
3514 | : lookup_as_function (folded_arg0, CONST))) != 0 | |
528b0df8 | 3515 | && GET_CODE (XEXP (y, 0)) == MINUS |
3516 | && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF | |
b49f2e4b | 3517 | && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg1)) |
528b0df8 | 3518 | return XEXP (XEXP (y, 0), 0); |
752df20e | 3519 | } |
f7cf73ed | 3520 | |
e6d1f05b | 3521 | /* Likewise if the operands are in the other order. */ |
3522 | if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF) | |
3523 | { | |
3524 | rtx y | |
3525 | = GET_CODE (folded_arg1) == MINUS ? folded_arg1 | |
b74befc5 | 3526 | : lookup_as_function (folded_arg1, MINUS); |
e6d1f05b | 3527 | |
3528 | if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF | |
b49f2e4b | 3529 | && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg0)) |
e6d1f05b | 3530 | return XEXP (y, 0); |
3531 | ||
3532 | /* Now try for a CONST of a MINUS like the above. */ | |
3533 | if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1 | |
3534 | : lookup_as_function (folded_arg1, CONST))) != 0 | |
3535 | && GET_CODE (XEXP (y, 0)) == MINUS | |
3536 | && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF | |
b49f2e4b | 3537 | && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg0)) |
e6d1f05b | 3538 | return XEXP (XEXP (y, 0), 0); |
3539 | } | |
3540 | ||
f7cf73ed | 3541 | /* If second operand is a register equivalent to a negative |
3542 | CONST_INT, see if we can find a register equivalent to the | |
3543 | positive constant. Make a MINUS if so. Don't do this for | |
337bf63c | 3544 | a non-negative constant since we might then alternate between |
3fb1e43b | 3545 | choosing positive and negative constants. Having the positive |
337bf63c | 3546 | constant previously-used is the more common case. Be sure |
3547 | the resulting constant is non-negative; if const_arg1 were | |
3548 | the smallest negative number this would overflow: depending | |
3549 | on the mode, this would either just be the same value (and | |
3550 | hence not save anything) or be incorrect. */ | |
971ba038 | 3551 | if (const_arg1 != 0 && CONST_INT_P (const_arg1) |
337bf63c | 3552 | && INTVAL (const_arg1) < 0 |
aaa2446c | 3553 | /* This used to test |
3554 | ||
b74befc5 | 3555 | -INTVAL (const_arg1) >= 0 |
aaa2446c | 3556 | |
3557 | But The Sun V5.0 compilers mis-compiled that test. So | |
3558 | instead we test for the problematic value in a more direct | |
3559 | manner and hope the Sun compilers get it correct. */ | |
76d98649 | 3560 | && INTVAL (const_arg1) != |
edc19fd0 | 3561 | (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1)) |
8ad4c111 | 3562 | && REG_P (folded_arg1)) |
f7cf73ed | 3563 | { |
b74befc5 | 3564 | rtx new_const = GEN_INT (-INTVAL (const_arg1)); |
f7cf73ed | 3565 | struct table_elt *p |
78d140c9 | 3566 | = lookup (new_const, SAFE_HASH (new_const, mode), mode); |
f7cf73ed | 3567 | |
3568 | if (p) | |
3569 | for (p = p->first_same_value; p; p = p->next_same_value) | |
8ad4c111 | 3570 | if (REG_P (p->exp)) |
af21a202 | 3571 | return simplify_gen_binary (MINUS, mode, folded_arg0, |
47f1d198 | 3572 | canon_reg (p->exp, NULL)); |
f7cf73ed | 3573 | } |
5c4c31e3 | 3574 | goto from_plus; |
3575 | ||
3576 | case MINUS: | |
3577 | /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2). | |
3578 | If so, produce (PLUS Z C2-C). */ | |
971ba038 | 3579 | if (const_arg1 != 0 && CONST_INT_P (const_arg1)) |
5c4c31e3 | 3580 | { |
3581 | rtx y = lookup_as_function (XEXP (x, 0), PLUS); | |
971ba038 | 3582 | if (y && CONST_INT_P (XEXP (y, 1))) |
29c05e22 | 3583 | return fold_rtx (plus_constant (mode, copy_rtx (y), |
a66a39f2 | 3584 | -INTVAL (const_arg1)), |
47f1d198 | 3585 | NULL); |
5c4c31e3 | 3586 | } |
752df20e | 3587 | |
b74befc5 | 3588 | /* Fall through. */ |
752df20e | 3589 | |
5c4c31e3 | 3590 | from_plus: |
752df20e | 3591 | case SMIN: case SMAX: case UMIN: case UMAX: |
3592 | case IOR: case AND: case XOR: | |
7a4fa2a1 | 3593 | case MULT: |
752df20e | 3594 | case ASHIFT: case LSHIFTRT: case ASHIFTRT: |
3595 | /* If we have (<op> <reg> <const_int>) for an associative OP and REG | |
3596 | is known to be of similar form, we may be able to replace the | |
3597 | operation with a combined operation. This may eliminate the | |
3598 | intermediate operation if every use is simplified in this way. | |
3599 | Note that the similar optimization done by combine.c only works | |
3600 | if the intermediate operation's result has only one reference. */ | |
3601 | ||
8ad4c111 | 3602 | if (REG_P (folded_arg0) |
971ba038 | 3603 | && const_arg1 && CONST_INT_P (const_arg1)) |
752df20e | 3604 | { |
3605 | int is_shift | |
3606 | = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT); | |
8f353ea8 | 3607 | rtx y, inner_const, new_const; |
6026d749 | 3608 | rtx canon_const_arg1 = const_arg1; |
752df20e | 3609 | enum rtx_code associate_code; |
752df20e | 3610 | |
0518a465 | 3611 | if (is_shift |
ded805e6 | 3612 | && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode) |
0518a465 | 3613 | || INTVAL (const_arg1) < 0)) |
3614 | { | |
3615 | if (SHIFT_COUNT_TRUNCATED) | |
6026d749 | 3616 | canon_const_arg1 = GEN_INT (INTVAL (const_arg1) |
3617 | & (GET_MODE_BITSIZE (mode) | |
3618 | - 1)); | |
0518a465 | 3619 | else |
3620 | break; | |
3621 | } | |
3622 | ||
8f353ea8 | 3623 | y = lookup_as_function (folded_arg0, code); |
0518a465 | 3624 | if (y == 0) |
3625 | break; | |
0518a465 | 3626 | |
3627 | /* If we have compiled a statement like | |
3628 | "if (x == (x & mask1))", and now are looking at | |
3629 | "x & mask2", we will have a case where the first operand | |
3630 | of Y is the same as our first operand. Unless we detect | |
3631 | this case, an infinite loop will result. */ | |
3632 | if (XEXP (y, 0) == folded_arg0) | |
752df20e | 3633 | break; |
3634 | ||
8f353ea8 | 3635 | inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0)); |
971ba038 | 3636 | if (!inner_const || !CONST_INT_P (inner_const)) |
8f353ea8 | 3637 | break; |
3638 | ||
752df20e | 3639 | /* Don't associate these operations if they are a PLUS with the |
3640 | same constant and it is a power of two. These might be doable | |
3641 | with a pre- or post-increment. Similarly for two subtracts of | |
3642 | identical powers of two with post decrement. */ | |
3643 | ||
9cae6d48 | 3644 | if (code == PLUS && const_arg1 == inner_const |
e4e498cf | 3645 | && ((HAVE_PRE_INCREMENT |
3646 | && exact_log2 (INTVAL (const_arg1)) >= 0) | |
3647 | || (HAVE_POST_INCREMENT | |
3648 | && exact_log2 (INTVAL (const_arg1)) >= 0) | |
3649 | || (HAVE_PRE_DECREMENT | |
3650 | && exact_log2 (- INTVAL (const_arg1)) >= 0) | |
3651 | || (HAVE_POST_DECREMENT | |
3652 | && exact_log2 (- INTVAL (const_arg1)) >= 0))) | |
752df20e | 3653 | break; |
3654 | ||
e7323ddd | 3655 | /* ??? Vector mode shifts by scalar |
3656 | shift operand are not supported yet. */ | |
3657 | if (is_shift && VECTOR_MODE_P (mode)) | |
3658 | break; | |
3659 | ||
0518a465 | 3660 | if (is_shift |
ded805e6 | 3661 | && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode) |
0518a465 | 3662 | || INTVAL (inner_const) < 0)) |
3663 | { | |
3664 | if (SHIFT_COUNT_TRUNCATED) | |
3665 | inner_const = GEN_INT (INTVAL (inner_const) | |
3666 | & (GET_MODE_BITSIZE (mode) - 1)); | |
3667 | else | |
3668 | break; | |
3669 | } | |
3670 | ||
752df20e | 3671 | /* Compute the code used to compose the constants. For example, |
7a4fa2a1 | 3672 | A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */ |
752df20e | 3673 | |
7a4fa2a1 | 3674 | associate_code = (is_shift || code == MINUS ? PLUS : code); |
752df20e | 3675 | |
3676 | new_const = simplify_binary_operation (associate_code, mode, | |
6026d749 | 3677 | canon_const_arg1, |
3678 | inner_const); | |
752df20e | 3679 | |
3680 | if (new_const == 0) | |
3681 | break; | |
3682 | ||
3683 | /* If we are associating shift operations, don't let this | |
94ad8c53 | 3684 | produce a shift of the size of the object or larger. |
3685 | This could occur when we follow a sign-extend by a right | |
3686 | shift on a machine that does a sign-extend as a pair | |
3687 | of shifts. */ | |
752df20e | 3688 | |
0518a465 | 3689 | if (is_shift |
971ba038 | 3690 | && CONST_INT_P (new_const) |
ded805e6 | 3691 | && INTVAL (new_const) >= GET_MODE_PRECISION (mode)) |
94ad8c53 | 3692 | { |
3693 | /* As an exception, we can turn an ASHIFTRT of this | |
3694 | form into a shift of the number of bits - 1. */ | |
3695 | if (code == ASHIFTRT) | |
3696 | new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1); | |
0518a465 | 3697 | else if (!side_effects_p (XEXP (y, 0))) |
3698 | return CONST0_RTX (mode); | |
94ad8c53 | 3699 | else |
3700 | break; | |
3701 | } | |
752df20e | 3702 | |
3703 | y = copy_rtx (XEXP (y, 0)); | |
3704 | ||
3705 | /* If Y contains our first operand (the most common way this | |
3706 | can happen is if Y is a MEM), we would do into an infinite | |
3707 | loop if we tried to fold it. So don't in that case. */ | |
3708 | ||
3709 | if (! reg_mentioned_p (folded_arg0, y)) | |
3710 | y = fold_rtx (y, insn); | |
3711 | ||
af21a202 | 3712 | return simplify_gen_binary (code, mode, y, new_const); |
752df20e | 3713 | } |
0dbd1c74 | 3714 | break; |
3715 | ||
7a4fa2a1 | 3716 | case DIV: case UDIV: |
3717 | /* ??? The associative optimization performed immediately above is | |
3718 | also possible for DIV and UDIV using associate_code of MULT. | |
3719 | However, we would need extra code to verify that the | |
3720 | multiplication does not overflow, that is, there is no overflow | |
3721 | in the calculation of new_const. */ | |
3722 | break; | |
3723 | ||
0dbd1c74 | 3724 | default: |
3725 | break; | |
752df20e | 3726 | } |
3727 | ||
d328ebdf | 3728 | new_rtx = simplify_binary_operation (code, mode, |
752df20e | 3729 | const_arg0 ? const_arg0 : folded_arg0, |
3730 | const_arg1 ? const_arg1 : folded_arg1); | |
3731 | break; | |
3732 | ||
6720e96c | 3733 | case RTX_OBJ: |
752df20e | 3734 | /* (lo_sum (high X) X) is simply X. */ |
3735 | if (code == LO_SUM && const_arg0 != 0 | |
3736 | && GET_CODE (const_arg0) == HIGH | |
3737 | && rtx_equal_p (XEXP (const_arg0, 0), const_arg1)) | |
3738 | return const_arg1; | |
3739 | break; | |
3740 | ||
6720e96c | 3741 | case RTX_TERNARY: |
3742 | case RTX_BITFIELD_OPS: | |
d328ebdf | 3743 | new_rtx = simplify_ternary_operation (code, mode, mode_arg0, |
752df20e | 3744 | const_arg0 ? const_arg0 : folded_arg0, |
3745 | const_arg1 ? const_arg1 : folded_arg1, | |
3746 | const_arg2 ? const_arg2 : XEXP (x, 2)); | |
3747 | break; | |
dd5ff96d | 3748 | |
6720e96c | 3749 | default: |
3750 | break; | |
752df20e | 3751 | } |
3752 | ||
d328ebdf | 3753 | return new_rtx ? new_rtx : x; |
752df20e | 3754 | } |
3755 | \f | |
3756 | /* Return a constant value currently equivalent to X. | |
3757 | Return 0 if we don't know one. */ | |
3758 | ||
3759 | static rtx | |
8ec3a57b | 3760 | equiv_constant (rtx x) |
752df20e | 3761 | { |
8ad4c111 | 3762 | if (REG_P (x) |
a7f3b1c7 | 3763 | && REGNO_QTY_VALID_P (REGNO (x))) |
3764 | { | |
3765 | int x_q = REG_QTY (REGNO (x)); | |
3766 | struct qty_table_elem *x_ent = &qty_table[x_q]; | |
3767 | ||
3768 | if (x_ent->const_rtx) | |
316f48ea | 3769 | x = gen_lowpart (GET_MODE (x), x_ent->const_rtx); |
a7f3b1c7 | 3770 | } |
752df20e | 3771 | |
f2f6be45 | 3772 | if (x == 0 || CONSTANT_P (x)) |
752df20e | 3773 | return x; |
3774 | ||
42a3a38b | 3775 | if (GET_CODE (x) == SUBREG) |
3776 | { | |
3754d046 | 3777 | machine_mode mode = GET_MODE (x); |
3778 | machine_mode imode = GET_MODE (SUBREG_REG (x)); | |
d328ebdf | 3779 | rtx new_rtx; |
42a3a38b | 3780 | |
3781 | /* See if we previously assigned a constant value to this SUBREG. */ | |
d328ebdf | 3782 | if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0 |
e913b5cd | 3783 | || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0 |
d328ebdf | 3784 | || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0 |
3785 | || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0) | |
3786 | return new_rtx; | |
42a3a38b | 3787 | |
5216d9e8 | 3788 | /* If we didn't and if doing so makes sense, see if we previously |
3789 | assigned a constant value to the enclosing word mode SUBREG. */ | |
3790 | if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode) | |
3791 | && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode)) | |
3792 | { | |
3793 | int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode); | |
3794 | if (byte >= 0 && (byte % UNITS_PER_WORD) == 0) | |
3795 | { | |
3796 | rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte); | |
3797 | new_rtx = lookup_as_function (y, CONST_INT); | |
3798 | if (new_rtx) | |
3799 | return gen_lowpart (mode, new_rtx); | |
3800 | } | |
3801 | } | |
3802 | ||
3a966565 | 3803 | /* Otherwise see if we already have a constant for the inner REG, |
3804 | and if that is enough to calculate an equivalent constant for | |
3805 | the subreg. Note that the upper bits of paradoxical subregs | |
3806 | are undefined, so they cannot be said to equal anything. */ | |
42a3a38b | 3807 | if (REG_P (SUBREG_REG (x)) |
3a966565 | 3808 | && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode) |
d328ebdf | 3809 | && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0) |
5216d9e8 | 3810 | return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x)); |
42a3a38b | 3811 | |
3812 | return 0; | |
3813 | } | |
3814 | ||
3815 | /* If X is a MEM, see if it is a constant-pool reference, or look it up in | |
3816 | the hash table in case its value was seen before. */ | |
e516eaa9 | 3817 | |
e16ceb8e | 3818 | if (MEM_P (x)) |
e516eaa9 | 3819 | { |
3820 | struct table_elt *elt; | |
3821 | ||
42a3a38b | 3822 | x = avoid_constant_pool_reference (x); |
e516eaa9 | 3823 | if (CONSTANT_P (x)) |
3824 | return x; | |
3825 | ||
78d140c9 | 3826 | elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x)); |
e516eaa9 | 3827 | if (elt == 0) |
3828 | return 0; | |
3829 | ||
3830 | for (elt = elt->first_same_value; elt; elt = elt->next_same_value) | |
3831 | if (elt->is_const && CONSTANT_P (elt->exp)) | |
3832 | return elt->exp; | |
3833 | } | |
3834 | ||
752df20e | 3835 | return 0; |
3836 | } | |
3837 | \f | |
bbe0b6d7 | 3838 | /* Given INSN, a jump insn, TAKEN indicates if we are following the |
3839 | "taken" branch. | |
752df20e | 3840 | |
3841 | In certain cases, this can cause us to add an equivalence. For example, | |
cb10db9d | 3842 | if we are following the taken case of |
8ec3a57b | 3843 | if (i == 2) |
752df20e | 3844 | we can add the fact that `i' and '2' are now equivalent. |
3845 | ||
3846 | In any case, we can record that this comparison was passed. If the same | |
3847 | comparison is seen later, we will know its value. */ | |
3848 | ||
3849 | static void | |
47f1d198 | 3850 | record_jump_equiv (rtx_insn *insn, bool taken) |
752df20e | 3851 | { |
3852 | int cond_known_true; | |
3853 | rtx op0, op1; | |
b2816317 | 3854 | rtx set; |
3754d046 | 3855 | machine_mode mode, mode0, mode1; |
752df20e | 3856 | int reversed_nonequality = 0; |
3857 | enum rtx_code code; | |
3858 | ||
3859 | /* Ensure this is the right kind of insn. */ | |
bbe0b6d7 | 3860 | gcc_assert (any_condjump_p (insn)); |
3861 | ||
b2816317 | 3862 | set = pc_set (insn); |
752df20e | 3863 | |
3864 | /* See if this jump condition is known true or false. */ | |
3865 | if (taken) | |
b2816317 | 3866 | cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx); |
752df20e | 3867 | else |
b2816317 | 3868 | cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx); |
752df20e | 3869 | |
3870 | /* Get the type of comparison being done and the operands being compared. | |
3871 | If we had to reverse a non-equality condition, record that fact so we | |
3872 | know that it isn't valid for floating-point. */ | |
b2816317 | 3873 | code = GET_CODE (XEXP (SET_SRC (set), 0)); |
3874 | op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn); | |
3875 | op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn); | |
752df20e | 3876 | |
40c562ee | 3877 | /* On a cc0 target the cc0-setter and cc0-user may end up in different |
3878 | blocks. When that happens the tracking of the cc0-setter via | |
3879 | PREV_INSN_CC0 is spoiled. That means that fold_rtx may return | |
3880 | NULL_RTX. In those cases, there's nothing to record. */ | |
3881 | if (op0 == NULL_RTX || op1 == NULL_RTX) | |
3882 | return; | |
3883 | ||
5c4c31e3 | 3884 | code = find_comparison_args (code, &op0, &op1, &mode0, &mode1); |
752df20e | 3885 | if (! cond_known_true) |
3886 | { | |
7da6ea0c | 3887 | code = reversed_comparison_code_parts (code, op0, op1, insn); |
a4110d9a | 3888 | |
3889 | /* Don't remember if we can't find the inverse. */ | |
3890 | if (code == UNKNOWN) | |
3891 | return; | |
752df20e | 3892 | } |
3893 | ||
3894 | /* The mode is the mode of the non-constant. */ | |
5c4c31e3 | 3895 | mode = mode0; |
3896 | if (mode1 != VOIDmode) | |
3897 | mode = mode1; | |
752df20e | 3898 | |
3899 | record_jump_cond (code, mode, op0, op1, reversed_nonequality); | |
3900 | } | |
3901 | ||
cfa1a80d | 3902 | /* Yet another form of subreg creation. In this case, we want something in |
3903 | MODE, and we should assume OP has MODE iff it is naturally modeless. */ | |
3904 | ||
3905 | static rtx | |
3754d046 | 3906 | record_jump_cond_subreg (machine_mode mode, rtx op) |
cfa1a80d | 3907 | { |
3754d046 | 3908 | machine_mode op_mode = GET_MODE (op); |
cfa1a80d | 3909 | if (op_mode == mode || op_mode == VOIDmode) |
3910 | return op; | |
3911 | return lowpart_subreg (mode, op, op_mode); | |
3912 | } | |
3913 | ||
752df20e | 3914 | /* We know that comparison CODE applied to OP0 and OP1 in MODE is true. |
3915 | REVERSED_NONEQUALITY is nonzero if CODE had to be swapped. | |
3916 | Make any useful entries we can with that information. Called from | |
3917 | above function and called recursively. */ | |
3918 | ||
3919 | static void | |
3754d046 | 3920 | record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0, |
8ec3a57b | 3921 | rtx op1, int reversed_nonequality) |
752df20e | 3922 | { |
952bc06d | 3923 | unsigned op0_hash, op1_hash; |
0af17926 | 3924 | int op0_in_memory, op1_in_memory; |
752df20e | 3925 | struct table_elt *op0_elt, *op1_elt; |
3926 | ||
3927 | /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG, | |
3928 | we know that they are also equal in the smaller mode (this is also | |
3929 | true for all smaller modes whether or not there is a SUBREG, but | |
f5d1f9f9 | 3930 | is not worth testing for with no SUBREG). */ |
752df20e | 3931 | |
3c5cc27f | 3932 | /* Note that GET_MODE (op0) may not equal MODE. */ |
b537bfdb | 3933 | if (code == EQ && paradoxical_subreg_p (op0)) |
752df20e | 3934 | { |
3754d046 | 3935 | machine_mode inner_mode = GET_MODE (SUBREG_REG (op0)); |
cfa1a80d | 3936 | rtx tem = record_jump_cond_subreg (inner_mode, op1); |
3937 | if (tem) | |
3938 | record_jump_cond (code, mode, SUBREG_REG (op0), tem, | |
3939 | reversed_nonequality); | |
752df20e | 3940 | } |
3941 | ||
b537bfdb | 3942 | if (code == EQ && paradoxical_subreg_p (op1)) |
752df20e | 3943 | { |
3754d046 | 3944 | machine_mode inner_mode = GET_MODE (SUBREG_REG (op1)); |
cfa1a80d | 3945 | rtx tem = record_jump_cond_subreg (inner_mode, op0); |
3946 | if (tem) | |
3947 | record_jump_cond (code, mode, SUBREG_REG (op1), tem, | |
3948 | reversed_nonequality); | |
752df20e | 3949 | } |
3950 | ||
cb10db9d | 3951 | /* Similarly, if this is an NE comparison, and either is a SUBREG |
752df20e | 3952 | making a smaller mode, we know the whole thing is also NE. */ |
3953 | ||
3c5cc27f | 3954 | /* Note that GET_MODE (op0) may not equal MODE; |
3955 | if we test MODE instead, we can get an infinite recursion | |
3956 | alternating between two modes each wider than MODE. */ | |
3957 | ||
752df20e | 3958 | if (code == NE && GET_CODE (op0) == SUBREG |
3959 | && subreg_lowpart_p (op0) | |
3c5cc27f | 3960 | && (GET_MODE_SIZE (GET_MODE (op0)) |
3961 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))) | |
752df20e | 3962 | { |
3754d046 | 3963 | machine_mode inner_mode = GET_MODE (SUBREG_REG (op0)); |
cfa1a80d | 3964 | rtx tem = record_jump_cond_subreg (inner_mode, op1); |
3965 | if (tem) | |
3966 | record_jump_cond (code, mode, SUBREG_REG (op0), tem, | |
3967 | reversed_nonequality); | |
752df20e | 3968 | } |
3969 | ||
3970 | if (code == NE && GET_CODE (op1) == SUBREG | |
3971 | && subreg_lowpart_p (op1) | |
3c5cc27f | 3972 | && (GET_MODE_SIZE (GET_MODE (op1)) |
3973 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1))))) | |
752df20e | 3974 | { |
3754d046 | 3975 | machine_mode inner_mode = GET_MODE (SUBREG_REG (op1)); |
cfa1a80d | 3976 | rtx tem = record_jump_cond_subreg (inner_mode, op0); |
3977 | if (tem) | |
3978 | record_jump_cond (code, mode, SUBREG_REG (op1), tem, | |
3979 | reversed_nonequality); | |
752df20e | 3980 | } |
3981 | ||
3982 | /* Hash both operands. */ | |
3983 | ||
3984 | do_not_record = 0; | |
3985 | hash_arg_in_memory = 0; | |
952bc06d | 3986 | op0_hash = HASH (op0, mode); |
752df20e | 3987 | op0_in_memory = hash_arg_in_memory; |
752df20e | 3988 | |
3989 | if (do_not_record) | |
3990 | return; | |
3991 | ||
3992 | do_not_record = 0; | |
3993 | hash_arg_in_memory = 0; | |
952bc06d | 3994 | op1_hash = HASH (op1, mode); |
752df20e | 3995 | op1_in_memory = hash_arg_in_memory; |
cb10db9d | 3996 | |
752df20e | 3997 | if (do_not_record) |
3998 | return; | |
3999 | ||
4000 | /* Look up both operands. */ | |
952bc06d | 4001 | op0_elt = lookup (op0, op0_hash, mode); |
4002 | op1_elt = lookup (op1, op1_hash, mode); | |
752df20e | 4003 | |
9f8339f2 | 4004 | /* If both operands are already equivalent or if they are not in the |
4005 | table but are identical, do nothing. */ | |
4006 | if ((op0_elt != 0 && op1_elt != 0 | |
4007 | && op0_elt->first_same_value == op1_elt->first_same_value) | |
4008 | || op0 == op1 || rtx_equal_p (op0, op1)) | |
4009 | return; | |
4010 | ||
752df20e | 4011 | /* If we aren't setting two things equal all we can do is save this |
5b620701 | 4012 | comparison. Similarly if this is floating-point. In the latter |
4013 | case, OP1 might be zero and both -0.0 and 0.0 are equal to it. | |
4014 | If we record the equality, we might inadvertently delete code | |
4015 | whose intent was to change -0 to +0. */ | |
4016 | ||
c1712420 | 4017 | if (code != EQ || FLOAT_MODE_P (GET_MODE (op0))) |
752df20e | 4018 | { |
a7f3b1c7 | 4019 | struct qty_table_elem *ent; |
4020 | int qty; | |
4021 | ||
752df20e | 4022 | /* If we reversed a floating-point comparison, if OP0 is not a |
4023 | register, or if OP1 is neither a register or constant, we can't | |
4024 | do anything. */ | |
4025 | ||
8ad4c111 | 4026 | if (!REG_P (op1)) |
752df20e | 4027 | op1 = equiv_constant (op1); |
4028 | ||
c1712420 | 4029 | if ((reversed_nonequality && FLOAT_MODE_P (mode)) |
8ad4c111 | 4030 | || !REG_P (op0) || op1 == 0) |
752df20e | 4031 | return; |
4032 | ||
4033 | /* Put OP0 in the hash table if it isn't already. This gives it a | |
4034 | new quantity number. */ | |
4035 | if (op0_elt == 0) | |
4036 | { | |
4679ade3 | 4037 | if (insert_regs (op0, NULL, 0)) |
752df20e | 4038 | { |
4039 | rehash_using_reg (op0); | |
952bc06d | 4040 | op0_hash = HASH (op0, mode); |
a45f1da6 | 4041 | |
4042 | /* If OP0 is contained in OP1, this changes its hash code | |
4043 | as well. Faster to rehash than to check, except | |
4044 | for the simple case of a constant. */ | |
4045 | if (! CONSTANT_P (op1)) | |
952bc06d | 4046 | op1_hash = HASH (op1,mode); |
752df20e | 4047 | } |
4048 | ||
4679ade3 | 4049 | op0_elt = insert (op0, NULL, op0_hash, mode); |
752df20e | 4050 | op0_elt->in_memory = op0_in_memory; |
752df20e | 4051 | } |
4052 | ||
a7f3b1c7 | 4053 | qty = REG_QTY (REGNO (op0)); |
4054 | ent = &qty_table[qty]; | |
4055 | ||
4056 | ent->comparison_code = code; | |
8ad4c111 | 4057 | if (REG_P (op1)) |
752df20e | 4058 | { |
95f65c26 | 4059 | /* Look it up again--in case op0 and op1 are the same. */ |
952bc06d | 4060 | op1_elt = lookup (op1, op1_hash, mode); |
95f65c26 | 4061 | |
752df20e | 4062 | /* Put OP1 in the hash table so it gets a new quantity number. */ |
4063 | if (op1_elt == 0) | |
4064 | { | |
4679ade3 | 4065 | if (insert_regs (op1, NULL, 0)) |
752df20e | 4066 | { |
4067 | rehash_using_reg (op1); | |
952bc06d | 4068 | op1_hash = HASH (op1, mode); |
752df20e | 4069 | } |
4070 | ||
4679ade3 | 4071 | op1_elt = insert (op1, NULL, op1_hash, mode); |
752df20e | 4072 | op1_elt->in_memory = op1_in_memory; |
752df20e | 4073 | } |
4074 | ||
a7f3b1c7 | 4075 | ent->comparison_const = NULL_RTX; |
4076 | ent->comparison_qty = REG_QTY (REGNO (op1)); | |
752df20e | 4077 | } |
4078 | else | |
4079 | { | |
a7f3b1c7 | 4080 | ent->comparison_const = op1; |
4081 | ent->comparison_qty = -1; | |
752df20e | 4082 | } |
4083 | ||
4084 | return; | |
4085 | } | |
4086 | ||
56e155ea | 4087 | /* If either side is still missing an equivalence, make it now, |
4088 | then merge the equivalences. */ | |
752df20e | 4089 | |
752df20e | 4090 | if (op0_elt == 0) |
4091 | { | |
4679ade3 | 4092 | if (insert_regs (op0, NULL, 0)) |
752df20e | 4093 | { |
4094 | rehash_using_reg (op0); | |
952bc06d | 4095 | op0_hash = HASH (op0, mode); |
752df20e | 4096 | } |
4097 | ||
4679ade3 | 4098 | op0_elt = insert (op0, NULL, op0_hash, mode); |
752df20e | 4099 | op0_elt->in_memory = op0_in_memory; |
752df20e | 4100 | } |
4101 | ||
4102 | if (op1_elt == 0) | |
4103 | { | |
4679ade3 | 4104 | if (insert_regs (op1, NULL, 0)) |
752df20e | 4105 | { |
4106 | rehash_using_reg (op1); | |
952bc06d | 4107 | op1_hash = HASH (op1, mode); |
752df20e | 4108 | } |
4109 | ||
4679ade3 | 4110 | op1_elt = insert (op1, NULL, op1_hash, mode); |
752df20e | 4111 | op1_elt->in_memory = op1_in_memory; |
752df20e | 4112 | } |
56e155ea | 4113 | |
4114 | merge_equiv_classes (op0_elt, op1_elt); | |
752df20e | 4115 | } |
4116 | \f | |
4117 | /* CSE processing for one instruction. | |
2aca5650 | 4118 | |
4119 | Most "true" common subexpressions are mostly optimized away in GIMPLE, | |
4120 | but the few that "leak through" are cleaned up by cse_insn, and complex | |
4121 | addressing modes are often formed here. | |
4122 | ||
4123 | The main function is cse_insn, and between here and that function | |
4124 | a couple of helper functions is defined to keep the size of cse_insn | |
4125 | within reasonable proportions. | |
4126 | ||
4127 | Data is shared between the main and helper functions via STRUCT SET, | |
4128 | that contains all data related for every set in the instruction that | |
4129 | is being processed. | |
4130 | ||
4131 | Note that cse_main processes all sets in the instruction. Most | |
4132 | passes in GCC only process simple SET insns or single_set insns, but | |
4133 | CSE processes insns with multiple sets as well. */ | |
752df20e | 4134 | |
4135 | /* Data on one SET contained in the instruction. */ | |
4136 | ||
4137 | struct set | |
4138 | { | |
4139 | /* The SET rtx itself. */ | |
4140 | rtx rtl; | |
4141 | /* The SET_SRC of the rtx (the original value, if it is changing). */ | |
4142 | rtx src; | |
4143 | /* The hash-table element for the SET_SRC of the SET. */ | |
4144 | struct table_elt *src_elt; | |
952bc06d | 4145 | /* Hash value for the SET_SRC. */ |
4146 | unsigned src_hash; | |
4147 | /* Hash value for the SET_DEST. */ | |
4148 | unsigned dest_hash; | |
752df20e | 4149 | /* The SET_DEST, with SUBREG, etc., stripped. */ |
4150 | rtx inner_dest; | |
cb10db9d | 4151 | /* Nonzero if the SET_SRC is in memory. */ |
752df20e | 4152 | char src_in_memory; |
752df20e | 4153 | /* Nonzero if the SET_SRC contains something |
4154 | whose value cannot be predicted and understood. */ | |
4155 | char src_volatile; | |
d8b9732d | 4156 | /* Original machine mode, in case it becomes a CONST_INT. |
4157 | The size of this field should match the size of the mode | |
4158 | field of struct rtx_def (see rtl.h). */ | |
4159 | ENUM_BITFIELD(machine_mode) mode : 8; | |
752df20e | 4160 | /* A constant equivalent for SET_SRC, if any. */ |
4161 | rtx src_const; | |
952bc06d | 4162 | /* Hash value of constant equivalent for SET_SRC. */ |
4163 | unsigned src_const_hash; | |
752df20e | 4164 | /* Table entry for constant equivalent for SET_SRC, if any. */ |
4165 | struct table_elt *src_const_elt; | |
977ffed2 | 4166 | /* Table entry for the destination address. */ |
4167 | struct table_elt *dest_addr_elt; | |
752df20e | 4168 | }; |
2aca5650 | 4169 | \f |
4170 | /* Special handling for (set REG0 REG1) where REG0 is the | |
4171 | "cheapest", cheaper than REG1. After cse, REG1 will probably not | |
4172 | be used in the sequel, so (if easily done) change this insn to | |
4173 | (set REG1 REG0) and replace REG1 with REG0 in the previous insn | |
4174 | that computed their value. Then REG1 will become a dead store | |
4175 | and won't cloud the situation for later optimizations. | |
4176 | ||
4177 | Do not make this change if REG1 is a hard register, because it will | |
4178 | then be used in the sequel and we may be changing a two-operand insn | |
4179 | into a three-operand insn. | |
4180 | ||
4181 | This is the last transformation that cse_insn will try to do. */ | |
752df20e | 4182 | |
4183 | static void | |
47f1d198 | 4184 | try_back_substitute_reg (rtx set, rtx_insn *insn) |
752df20e | 4185 | { |
2aca5650 | 4186 | rtx dest = SET_DEST (set); |
4187 | rtx src = SET_SRC (set); | |
752df20e | 4188 | |
2aca5650 | 4189 | if (REG_P (dest) |
4190 | && REG_P (src) && ! HARD_REGISTER_P (src) | |
4191 | && REGNO_QTY_VALID_P (REGNO (src))) | |
4192 | { | |
4193 | int src_q = REG_QTY (REGNO (src)); | |
4194 | struct qty_table_elem *src_ent = &qty_table[src_q]; | |
752df20e | 4195 | |
2aca5650 | 4196 | if (src_ent->first_reg == REGNO (dest)) |
4197 | { | |
4198 | /* Scan for the previous nonnote insn, but stop at a basic | |
4199 | block boundary. */ | |
47f1d198 | 4200 | rtx_insn *prev = insn; |
4201 | rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn)); | |
2aca5650 | 4202 | do |
4203 | { | |
4204 | prev = PREV_INSN (prev); | |
4205 | } | |
4206 | while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev))); | |
752df20e | 4207 | |
2aca5650 | 4208 | /* Do not swap the registers around if the previous instruction |
4209 | attaches a REG_EQUIV note to REG1. | |
752df20e | 4210 | |
2aca5650 | 4211 | ??? It's not entirely clear whether we can transfer a REG_EQUIV |
4212 | from the pseudo that originally shadowed an incoming argument | |
4213 | to another register. Some uses of REG_EQUIV might rely on it | |
4214 | being attached to REG1 rather than REG2. | |
752df20e | 4215 | |
2aca5650 | 4216 | This section previously turned the REG_EQUIV into a REG_EQUAL |
4217 | note. We cannot do that because REG_EQUIV may provide an | |
4218 | uninitialized stack slot when REG_PARM_STACK_SPACE is used. */ | |
4219 | if (NONJUMP_INSN_P (prev) | |
4220 | && GET_CODE (PATTERN (prev)) == SET | |
4221 | && SET_DEST (PATTERN (prev)) == src | |
4222 | && ! find_reg_note (prev, REG_EQUIV, NULL_RTX)) | |
4223 | { | |
4224 | rtx note; | |
4225 | ||
4226 | validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1); | |
4227 | validate_change (insn, &SET_DEST (set), src, 1); | |
4228 | validate_change (insn, &SET_SRC (set), dest, 1); | |
4229 | apply_change_group (); | |
4230 | ||
4231 | /* If INSN has a REG_EQUAL note, and this note mentions | |
4232 | REG0, then we must delete it, because the value in | |
4233 | REG0 has changed. If the note's value is REG1, we must | |
4234 | also delete it because that is now this insn's dest. */ | |
4235 | note = find_reg_note (insn, REG_EQUAL, NULL_RTX); | |
4236 | if (note != 0 | |
4237 | && (reg_mentioned_p (dest, XEXP (note, 0)) | |
4238 | || rtx_equal_p (src, XEXP (note, 0)))) | |
4239 | remove_note (insn, note); | |
4240 | } | |
ddaf7ad3 | 4241 | } |
b84155cd | 4242 | } |
2aca5650 | 4243 | } |
4244 | \f | |
4245 | /* Record all the SETs in this instruction into SETS_PTR, | |
4246 | and return the number of recorded sets. */ | |
4247 | static int | |
47f1d198 | 4248 | find_sets_in_insn (rtx_insn *insn, struct set **psets) |
2aca5650 | 4249 | { |
4250 | struct set *sets = *psets; | |
4251 | int n_sets = 0; | |
4252 | rtx x = PATTERN (insn); | |
b84155cd | 4253 | |
752df20e | 4254 | if (GET_CODE (x) == SET) |
4255 | { | |
752df20e | 4256 | /* Ignore SETs that are unconditional jumps. |
4257 | They never need cse processing, so this does not hurt. | |
4258 | The reason is not efficiency but rather | |
4259 | so that we can test at the end for instructions | |
4260 | that have been simplified to unconditional jumps | |
4261 | and not be misled by unchanged instructions | |
4262 | that were unconditional jumps to begin with. */ | |
4263 | if (SET_DEST (x) == pc_rtx | |
4264 | && GET_CODE (SET_SRC (x)) == LABEL_REF) | |
4265 | ; | |
752df20e | 4266 | /* Don't count call-insns, (set (reg 0) (call ...)), as a set. |
4267 | The hard function value register is used only once, to copy to | |
2aca5650 | 4268 | someplace else, so it isn't worth cse'ing. */ |
752df20e | 4269 | else if (GET_CODE (SET_SRC (x)) == CALL) |
2aca5650 | 4270 | ; |
752df20e | 4271 | else |
2aca5650 | 4272 | sets[n_sets++].rtl = x; |
752df20e | 4273 | } |
4274 | else if (GET_CODE (x) == PARALLEL) | |
4275 | { | |
2aca5650 | 4276 | int i, lim = XVECLEN (x, 0); |
cb10db9d | 4277 | |
20d3ff08 | 4278 | /* Go over the expressions of the PARALLEL in forward order, to |
2aca5650 | 4279 | put them in the same order in the SETS array. */ |
752df20e | 4280 | for (i = 0; i < lim; i++) |
4281 | { | |
19cb6b50 | 4282 | rtx y = XVECEXP (x, 0, i); |
752df20e | 4283 | if (GET_CODE (y) == SET) |
4284 | { | |
8d5dd220 | 4285 | /* As above, we ignore unconditional jumps and call-insns and |
4286 | ignore the result of apply_change_group. */ | |
2aca5650 | 4287 | if (SET_DEST (y) == pc_rtx |
4288 | && GET_CODE (SET_SRC (y)) == LABEL_REF) | |
4289 | ; | |
4290 | else if (GET_CODE (SET_SRC (y)) == CALL) | |
752df20e | 4291 | ; |
4292 | else | |
4293 | sets[n_sets++].rtl = y; | |
4294 | } | |
752df20e | 4295 | } |
4296 | } | |
2aca5650 | 4297 | |
4298 | return n_sets; | |
4299 | } | |
4300 | \f | |
c0ac34cf | 4301 | /* Subroutine of canonicalize_insn. X is an ASM_OPERANDS in INSN. */ |
4302 | ||
4303 | static void | |
4304 | canon_asm_operands (rtx x, rtx_insn *insn) | |
4305 | { | |
4306 | for (int i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) | |
4307 | { | |
4308 | rtx input = ASM_OPERANDS_INPUT (x, i); | |
4309 | if (!(REG_P (input) && HARD_REGISTER_P (input))) | |
4310 | { | |
4311 | input = canon_reg (input, insn); | |
4312 | validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1); | |
4313 | } | |
4314 | } | |
4315 | } | |
4316 | ||
2aca5650 | 4317 | /* Where possible, substitute every register reference in the N_SETS |
47ae02b7 | 4318 | number of SETS in INSN with the canonical register. |
2aca5650 | 4319 | |
4320 | Register canonicalization propagatest the earliest register (i.e. | |
4321 | one that is set before INSN) with the same value. This is a very | |
4322 | useful, simple form of CSE, to clean up warts from expanding GIMPLE | |
4323 | to RTL. For instance, a CONST for an address is usually expanded | |
4324 | multiple times to loads into different registers, thus creating many | |
4325 | subexpressions of the form: | |
4326 | ||
4327 | (set (reg1) (some_const)) | |
4328 | (set (mem (... reg1 ...) (thing))) | |
4329 | (set (reg2) (some_const)) | |
4330 | (set (mem (... reg2 ...) (thing))) | |
4331 | ||
4332 | After canonicalizing, the code takes the following form: | |
4333 | ||
4334 | (set (reg1) (some_const)) | |
4335 | (set (mem (... reg1 ...) (thing))) | |
4336 | (set (reg2) (some_const)) | |
4337 | (set (mem (... reg1 ...) (thing))) | |
4338 | ||
4339 | The set to reg2 is now trivially dead, and the memory reference (or | |
4340 | address, or whatever) may be a candidate for further CSEing. | |
4341 | ||
4342 | In this function, the result of apply_change_group can be ignored; | |
4343 | see canon_reg. */ | |
4344 | ||
4345 | static void | |
47f1d198 | 4346 | canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets) |
2aca5650 | 4347 | { |
4348 | struct set *sets = *psets; | |
4349 | rtx tem; | |
4350 | rtx x = PATTERN (insn); | |
4351 | int i; | |
4352 | ||
4353 | if (CALL_P (insn)) | |
4354 | { | |
4355 | for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1)) | |
c8010b80 | 4356 | if (GET_CODE (XEXP (tem, 0)) != SET) |
4357 | XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn); | |
2aca5650 | 4358 | } |
4359 | ||
4360 | if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL) | |
4361 | { | |
4362 | canon_reg (SET_SRC (x), insn); | |
4363 | apply_change_group (); | |
4364 | fold_rtx (SET_SRC (x), insn); | |
4365 | } | |
752df20e | 4366 | else if (GET_CODE (x) == CLOBBER) |
4367 | { | |
2aca5650 | 4368 | /* If we clobber memory, canon the address. |
4369 | This does nothing when a register is clobbered | |
4370 | because we have already invalidated the reg. */ | |
e16ceb8e | 4371 | if (MEM_P (XEXP (x, 0))) |
3072d30e | 4372 | canon_reg (XEXP (x, 0), insn); |
752df20e | 4373 | } |
752df20e | 4374 | else if (GET_CODE (x) == USE |
8ad4c111 | 4375 | && ! (REG_P (XEXP (x, 0)) |
752df20e | 4376 | && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER)) |
2aca5650 | 4377 | /* Canonicalize a USE of a pseudo register or memory location. */ |
e126ac03 | 4378 | canon_reg (x, insn); |
4379 | else if (GET_CODE (x) == ASM_OPERANDS) | |
c0ac34cf | 4380 | canon_asm_operands (x, insn); |
752df20e | 4381 | else if (GET_CODE (x) == CALL) |
4382 | { | |
4383 | canon_reg (x, insn); | |
8b82837b | 4384 | apply_change_group (); |
752df20e | 4385 | fold_rtx (x, insn); |
4386 | } | |
9845d120 | 4387 | else if (DEBUG_INSN_P (insn)) |
4388 | canon_reg (PATTERN (insn), insn); | |
2aca5650 | 4389 | else if (GET_CODE (x) == PARALLEL) |
4390 | { | |
4391 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
4392 | { | |
4393 | rtx y = XVECEXP (x, 0, i); | |
4394 | if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL) | |
4395 | { | |
4396 | canon_reg (SET_SRC (y), insn); | |
4397 | apply_change_group (); | |
4398 | fold_rtx (SET_SRC (y), insn); | |
4399 | } | |
4400 | else if (GET_CODE (y) == CLOBBER) | |
4401 | { | |
4402 | if (MEM_P (XEXP (y, 0))) | |
4403 | canon_reg (XEXP (y, 0), insn); | |
4404 | } | |
4405 | else if (GET_CODE (y) == USE | |
4406 | && ! (REG_P (XEXP (y, 0)) | |
4407 | && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER)) | |
4408 | canon_reg (y, insn); | |
c0ac34cf | 4409 | else if (GET_CODE (y) == ASM_OPERANDS) |
4410 | canon_asm_operands (y, insn); | |
2aca5650 | 4411 | else if (GET_CODE (y) == CALL) |
4412 | { | |
4413 | canon_reg (y, insn); | |
4414 | apply_change_group (); | |
4415 | fold_rtx (y, insn); | |
4416 | } | |
4417 | } | |
4418 | } | |
752df20e | 4419 | |
384770d0 | 4420 | if (n_sets == 1 && REG_NOTES (insn) != 0 |
2aca5650 | 4421 | && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0) |
24d87432 | 4422 | { |
2aca5650 | 4423 | /* We potentially will process this insn many times. Therefore, |
4424 | drop the REG_EQUAL note if it is equal to the SET_SRC of the | |
4425 | unique set in INSN. | |
4426 | ||
4427 | Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART, | |
4428 | because cse_insn handles those specially. */ | |
4429 | if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART | |
4430 | && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))) | |
4431 | remove_note (insn, tem); | |
4432 | else | |
4433 | { | |
4434 | canon_reg (XEXP (tem, 0), insn); | |
4435 | apply_change_group (); | |
4436 | XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn); | |
4437 | df_notes_rescan (insn); | |
4438 | } | |
24d87432 | 4439 | } |
752df20e | 4440 | |
4441 | /* Canonicalize sources and addresses of destinations. | |
4442 | We do this in a separate pass to avoid problems when a MATCH_DUP is | |
4443 | present in the insn pattern. In that case, we want to ensure that | |
4444 | we don't break the duplicate nature of the pattern. So we will replace | |
4445 | both operands at the same time. Otherwise, we would fail to find an | |
4446 | equivalent substitution in the loop calling validate_change below. | |
752df20e | 4447 | |
4448 | We used to suppress canonicalization of DEST if it appears in SRC, | |
8b82837b | 4449 | but we don't do this any more. */ |
752df20e | 4450 | |
4451 | for (i = 0; i < n_sets; i++) | |
4452 | { | |
4453 | rtx dest = SET_DEST (sets[i].rtl); | |
4454 | rtx src = SET_SRC (sets[i].rtl); | |
d328ebdf | 4455 | rtx new_rtx = canon_reg (src, insn); |
752df20e | 4456 | |
d328ebdf | 4457 | validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1); |
752df20e | 4458 | |
476d094d | 4459 | if (GET_CODE (dest) == ZERO_EXTRACT) |
752df20e | 4460 | { |
4461 | validate_change (insn, &XEXP (dest, 1), | |
8b82837b | 4462 | canon_reg (XEXP (dest, 1), insn), 1); |
752df20e | 4463 | validate_change (insn, &XEXP (dest, 2), |
8b82837b | 4464 | canon_reg (XEXP (dest, 2), insn), 1); |
752df20e | 4465 | } |
4466 | ||
476d094d | 4467 | while (GET_CODE (dest) == SUBREG |
752df20e | 4468 | || GET_CODE (dest) == ZERO_EXTRACT |
476d094d | 4469 | || GET_CODE (dest) == STRICT_LOW_PART) |
752df20e | 4470 | dest = XEXP (dest, 0); |
4471 | ||
e16ceb8e | 4472 | if (MEM_P (dest)) |
752df20e | 4473 | canon_reg (dest, insn); |
4474 | } | |
4475 | ||
8b82837b | 4476 | /* Now that we have done all the replacements, we can apply the change |
4477 | group and see if they all work. Note that this will cause some | |
4478 | canonicalizations that would have worked individually not to be applied | |
4479 | because some other canonicalization didn't work, but this should not | |
cb10db9d | 4480 | occur often. |
8d5dd220 | 4481 | |
4482 | The result of apply_change_group can be ignored; see canon_reg. */ | |
8b82837b | 4483 | |
4484 | apply_change_group (); | |
2aca5650 | 4485 | } |
4486 | \f | |
4487 | /* Main function of CSE. | |
4488 | First simplify sources and addresses of all assignments | |
4489 | in the instruction, using previously-computed equivalents values. | |
4490 | Then install the new sources and destinations in the table | |
4491 | of available values. */ | |
4492 | ||
4493 | static void | |
47f1d198 | 4494 | cse_insn (rtx_insn *insn) |
2aca5650 | 4495 | { |
4496 | rtx x = PATTERN (insn); | |
4497 | int i; | |
4498 | rtx tem; | |
4499 | int n_sets = 0; | |
4500 | ||
4501 | rtx src_eqv = 0; | |
4502 | struct table_elt *src_eqv_elt = 0; | |
4503 | int src_eqv_volatile = 0; | |
4504 | int src_eqv_in_memory = 0; | |
4505 | unsigned src_eqv_hash = 0; | |
4506 | ||
4507 | struct set *sets = (struct set *) 0; | |
4508 | ||
4509 | if (GET_CODE (x) == SET) | |
4510 | sets = XALLOCA (struct set); | |
4511 | else if (GET_CODE (x) == PARALLEL) | |
4512 | sets = XALLOCAVEC (struct set, XVECLEN (x, 0)); | |
4513 | ||
4514 | this_insn = insn; | |
2aca5650 | 4515 | /* Records what this insn does to set CC0. */ |
4516 | this_insn_cc0 = 0; | |
4517 | this_insn_cc0_mode = VOIDmode; | |
2aca5650 | 4518 | |
4519 | /* Find all regs explicitly clobbered in this insn, | |
4520 | to ensure they are not replaced with any other regs | |
4521 | elsewhere in this insn. */ | |
4522 | invalidate_from_sets_and_clobbers (insn); | |
4523 | ||
4524 | /* Record all the SETs in this instruction. */ | |
4525 | n_sets = find_sets_in_insn (insn, &sets); | |
4526 | ||
4527 | /* Substitute the canonical register where possible. */ | |
4528 | canonicalize_insn (insn, &sets, n_sets); | |
4529 | ||
4530 | /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV, | |
f2c7e335 | 4531 | if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The |
4532 | latter condition is necessary because SRC_EQV is handled specially for | |
4533 | this case, and if it isn't set, then there will be no equivalence | |
4534 | for the destination. */ | |
2aca5650 | 4535 | if (n_sets == 1 && REG_NOTES (insn) != 0 |
6c1fc504 | 4536 | && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0) |
f2c7e335 | 4537 | { |
f2c7e335 | 4538 | |
6c1fc504 | 4539 | if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT |
4540 | && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)) | |
4541 | || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART)) | |
4542 | src_eqv = copy_rtx (XEXP (tem, 0)); | |
f2c7e335 | 4543 | /* If DEST is of the form ZERO_EXTACT, as in: |
4544 | (set (zero_extract:SI (reg:SI 119) | |
4545 | (const_int 16 [0x10]) | |
4546 | (const_int 16 [0x10])) | |
4547 | (const_int 51154 [0xc7d2])) | |
4548 | REG_EQUAL note will specify the value of register (reg:SI 119) at this | |
4549 | point. Note that this is different from SRC_EQV. We can however | |
4550 | calculate SRC_EQV with the position and width of ZERO_EXTRACT. */ | |
4551 | else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT | |
a76e121f | 4552 | && CONST_INT_P (XEXP (tem, 0)) |
f2c7e335 | 4553 | && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1)) |
4554 | && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2))) | |
4555 | { | |
4556 | rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0); | |
4557 | rtx width = XEXP (SET_DEST (sets[0].rtl), 1); | |
4558 | rtx pos = XEXP (SET_DEST (sets[0].rtl), 2); | |
a76e121f | 4559 | HOST_WIDE_INT val = INTVAL (XEXP (tem, 0)); |
f2c7e335 | 4560 | HOST_WIDE_INT mask; |
4561 | unsigned int shift; | |
4562 | if (BITS_BIG_ENDIAN) | |
4563 | shift = GET_MODE_PRECISION (GET_MODE (dest_reg)) | |
4564 | - INTVAL (pos) - INTVAL (width); | |
4565 | else | |
4566 | shift = INTVAL (pos); | |
4567 | if (INTVAL (width) == HOST_BITS_PER_WIDE_INT) | |
7097b942 | 4568 | mask = HOST_WIDE_INT_M1; |
f2c7e335 | 4569 | else |
edc19fd0 | 4570 | mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1; |
f2c7e335 | 4571 | val = (val >> shift) & mask; |
4572 | src_eqv = GEN_INT (val); | |
4573 | } | |
4574 | } | |
8b82837b | 4575 | |
752df20e | 4576 | /* Set sets[i].src_elt to the class each source belongs to. |
4577 | Detect assignments from or to volatile things | |
4578 | and set set[i] to zero so they will be ignored | |
4579 | in the rest of this function. | |
4580 | ||
4581 | Nothing in this loop changes the hash table or the register chains. */ | |
4582 | ||
4583 | for (i = 0; i < n_sets; i++) | |
4584 | { | |
a49d9163 | 4585 | bool repeat = false; |
663f4248 | 4586 | bool mem_noop_insn = false; |
19cb6b50 | 4587 | rtx src, dest; |
4588 | rtx src_folded; | |
4589 | struct table_elt *elt = 0, *p; | |
3754d046 | 4590 | machine_mode mode; |
752df20e | 4591 | rtx src_eqv_here; |
4592 | rtx src_const = 0; | |
4593 | rtx src_related = 0; | |
01c8e4c9 | 4594 | bool src_related_is_const_anchor = false; |
752df20e | 4595 | struct table_elt *src_const_elt = 0; |
fb561825 | 4596 | int src_cost = MAX_COST; |
4597 | int src_eqv_cost = MAX_COST; | |
4598 | int src_folded_cost = MAX_COST; | |
4599 | int src_related_cost = MAX_COST; | |
4600 | int src_elt_cost = MAX_COST; | |
4601 | int src_regcost = MAX_COST; | |
4602 | int src_eqv_regcost = MAX_COST; | |
4603 | int src_folded_regcost = MAX_COST; | |
4604 | int src_related_regcost = MAX_COST; | |
4605 | int src_elt_regcost = MAX_COST; | |
d10cfa8d | 4606 | /* Set nonzero if we need to call force_const_mem on with the |
752df20e | 4607 | contents of src_folded before using it. */ |
4608 | int src_folded_force_flag = 0; | |
4609 | ||
4610 | dest = SET_DEST (sets[i].rtl); | |
4611 | src = SET_SRC (sets[i].rtl); | |
4612 | ||
4613 | /* If SRC is a constant that has no machine mode, | |
4614 | hash it with the destination's machine mode. | |
4615 | This way we can keep different modes separate. */ | |
4616 | ||
4617 | mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src); | |
4618 | sets[i].mode = mode; | |
4619 | ||
4620 | if (src_eqv) | |
4621 | { | |
3754d046 | 4622 | machine_mode eqvmode = mode; |
752df20e | 4623 | if (GET_CODE (dest) == STRICT_LOW_PART) |
4624 | eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0))); | |
4625 | do_not_record = 0; | |
4626 | hash_arg_in_memory = 0; | |
952bc06d | 4627 | src_eqv_hash = HASH (src_eqv, eqvmode); |
752df20e | 4628 | |
4629 | /* Find the equivalence class for the equivalent expression. */ | |
4630 | ||
4631 | if (!do_not_record) | |
952bc06d | 4632 | src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode); |
752df20e | 4633 | |
4634 | src_eqv_volatile = do_not_record; | |
4635 | src_eqv_in_memory = hash_arg_in_memory; | |
752df20e | 4636 | } |
4637 | ||
4638 | /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the | |
4639 | value of the INNER register, not the destination. So it is not | |
fd6efe25 | 4640 | a valid substitution for the source. But save it for later. */ |
752df20e | 4641 | if (GET_CODE (dest) == STRICT_LOW_PART) |
4642 | src_eqv_here = 0; | |
4643 | else | |
4644 | src_eqv_here = src_eqv; | |
4645 | ||
4646 | /* Simplify and foldable subexpressions in SRC. Then get the fully- | |
4647 | simplified result, which may not necessarily be valid. */ | |
afabf5e0 | 4648 | src_folded = fold_rtx (src, NULL); |
752df20e | 4649 | |
c93674f2 | 4650 | #if 0 |
4651 | /* ??? This caused bad code to be generated for the m68k port with -O2. | |
4652 | Suppose src is (CONST_INT -1), and that after truncation src_folded | |
4653 | is (CONST_INT 3). Suppose src_folded is then used for src_const. | |
4654 | At the end we will add src and src_const to the same equivalence | |
4655 | class. We now have 3 and -1 on the same equivalence class. This | |
4656 | causes later instructions to be mis-optimized. */ | |
752df20e | 4657 | /* If storing a constant in a bitfield, pre-truncate the constant |
4658 | so we will be able to record it later. */ | |
476d094d | 4659 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT) |
752df20e | 4660 | { |
4661 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
4662 | ||
971ba038 | 4663 | if (CONST_INT_P (src) |
4664 | && CONST_INT_P (width) | |
b572011e | 4665 | && INTVAL (width) < HOST_BITS_PER_WIDE_INT |
4666 | && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width)))) | |
4667 | src_folded | |
edc19fd0 | 4668 | = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1 |
b572011e | 4669 | << INTVAL (width)) - 1)); |
752df20e | 4670 | } |
c93674f2 | 4671 | #endif |
752df20e | 4672 | |
4673 | /* Compute SRC's hash code, and also notice if it | |
4674 | should not be recorded at all. In that case, | |
4675 | prevent any further processing of this assignment. */ | |
4676 | do_not_record = 0; | |
4677 | hash_arg_in_memory = 0; | |
752df20e | 4678 | |
4679 | sets[i].src = src; | |
952bc06d | 4680 | sets[i].src_hash = HASH (src, mode); |
752df20e | 4681 | sets[i].src_volatile = do_not_record; |
4682 | sets[i].src_in_memory = hash_arg_in_memory; | |
752df20e | 4683 | |
6ea5a450 | 4684 | /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is |
75f84104 | 4685 | a pseudo, do not record SRC. Using SRC as a replacement for |
4686 | anything else will be incorrect in that situation. Note that | |
4687 | this usually occurs only for stack slots, in which case all the | |
4688 | RTL would be referring to SRC, so we don't lose any optimization | |
4689 | opportunities by not having SRC in the hash table. */ | |
6ea5a450 | 4690 | |
e16ceb8e | 4691 | if (MEM_P (src) |
75f84104 | 4692 | && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0 |
8ad4c111 | 4693 | && REG_P (dest) |
75f84104 | 4694 | && REGNO (dest) >= FIRST_PSEUDO_REGISTER) |
6ea5a450 | 4695 | sets[i].src_volatile = 1; |
4696 | ||
f3f18244 | 4697 | else if (GET_CODE (src) == ASM_OPERANDS |
4698 | && GET_CODE (x) == PARALLEL) | |
20d3ff08 | 4699 | { |
4700 | /* Do not record result of a non-volatile inline asm with | |
4701 | more than one result. */ | |
4702 | if (n_sets > 1) | |
4703 | sets[i].src_volatile = 1; | |
4704 | ||
4705 | int j, lim = XVECLEN (x, 0); | |
4706 | for (j = 0; j < lim; j++) | |
4707 | { | |
4708 | rtx y = XVECEXP (x, 0, j); | |
4709 | /* And do not record result of a non-volatile inline asm | |
4710 | with "memory" clobber. */ | |
4711 | if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0))) | |
4712 | { | |
4713 | sets[i].src_volatile = 1; | |
4714 | break; | |
4715 | } | |
4716 | } | |
4717 | } | |
f3f18244 | 4718 | |
c538053c | 4719 | #if 0 |
4720 | /* It is no longer clear why we used to do this, but it doesn't | |
4721 | appear to still be needed. So let's try without it since this | |
4722 | code hurts cse'ing widened ops. */ | |
5f3447b0 | 4723 | /* If source is a paradoxical subreg (such as QI treated as an SI), |
752df20e | 4724 | treat it as volatile. It may do the work of an SI in one context |
4725 | where the extra bits are not being used, but cannot replace an SI | |
4726 | in general. */ | |
b537bfdb | 4727 | if (paradoxical_subreg_p (src)) |
752df20e | 4728 | sets[i].src_volatile = 1; |
c538053c | 4729 | #endif |
752df20e | 4730 | |
4731 | /* Locate all possible equivalent forms for SRC. Try to replace | |
4732 | SRC in the insn with each cheaper equivalent. | |
4733 | ||
4734 | We have the following types of equivalents: SRC itself, a folded | |
4735 | version, a value given in a REG_EQUAL note, or a value related | |
4736 | to a constant. | |
4737 | ||
4738 | Each of these equivalents may be part of an additional class | |
4739 | of equivalents (if more than one is in the table, they must be in | |
4740 | the same class; we check for this). | |
4741 | ||
4742 | If the source is volatile, we don't do any table lookups. | |
4743 | ||
4744 | We note any constant equivalent for possible later use in a | |
4745 | REG_NOTE. */ | |
4746 | ||
4747 | if (!sets[i].src_volatile) | |
952bc06d | 4748 | elt = lookup (src, sets[i].src_hash, mode); |
752df20e | 4749 | |
4750 | sets[i].src_elt = elt; | |
4751 | ||
4752 | if (elt && src_eqv_here && src_eqv_elt) | |
cb10db9d | 4753 | { |
4754 | if (elt->first_same_value != src_eqv_elt->first_same_value) | |
752df20e | 4755 | { |
4756 | /* The REG_EQUAL is indicating that two formerly distinct | |
4757 | classes are now equivalent. So merge them. */ | |
4758 | merge_equiv_classes (elt, src_eqv_elt); | |
952bc06d | 4759 | src_eqv_hash = HASH (src_eqv, elt->mode); |
4760 | src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode); | |
752df20e | 4761 | } |
4762 | ||
cb10db9d | 4763 | src_eqv_here = 0; |
4764 | } | |
752df20e | 4765 | |
4766 | else if (src_eqv_elt) | |
cb10db9d | 4767 | elt = src_eqv_elt; |
752df20e | 4768 | |
4769 | /* Try to find a constant somewhere and record it in `src_const'. | |
4770 | Record its table element, if any, in `src_const_elt'. Look in | |
4771 | any known equivalences first. (If the constant is not in the | |
952bc06d | 4772 | table, also set `sets[i].src_const_hash'). */ |
752df20e | 4773 | if (elt) |
cb10db9d | 4774 | for (p = elt->first_same_value; p; p = p->next_same_value) |
752df20e | 4775 | if (p->is_const) |
4776 | { | |
4777 | src_const = p->exp; | |
4778 | src_const_elt = elt; | |
4779 | break; | |
4780 | } | |
4781 | ||
4782 | if (src_const == 0 | |
4783 | && (CONSTANT_P (src_folded) | |
cb10db9d | 4784 | /* Consider (minus (label_ref L1) (label_ref L2)) as |
752df20e | 4785 | "constant" here so we will record it. This allows us |
4786 | to fold switch statements when an ADDR_DIFF_VEC is used. */ | |
4787 | || (GET_CODE (src_folded) == MINUS | |
4788 | && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF | |
4789 | && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF))) | |
4790 | src_const = src_folded, src_const_elt = elt; | |
4791 | else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here)) | |
4792 | src_const = src_eqv_here, src_const_elt = src_eqv_elt; | |
4793 | ||
4794 | /* If we don't know if the constant is in the table, get its | |
4795 | hash code and look it up. */ | |
4796 | if (src_const && src_const_elt == 0) | |
4797 | { | |
952bc06d | 4798 | sets[i].src_const_hash = HASH (src_const, mode); |
4799 | src_const_elt = lookup (src_const, sets[i].src_const_hash, mode); | |
752df20e | 4800 | } |
4801 | ||
4802 | sets[i].src_const = src_const; | |
4803 | sets[i].src_const_elt = src_const_elt; | |
4804 | ||
4805 | /* If the constant and our source are both in the table, mark them as | |
4806 | equivalent. Otherwise, if a constant is in the table but the source | |
4807 | isn't, set ELT to it. */ | |
4808 | if (src_const_elt && elt | |
4809 | && src_const_elt->first_same_value != elt->first_same_value) | |
4810 | merge_equiv_classes (elt, src_const_elt); | |
4811 | else if (src_const_elt && elt == 0) | |
4812 | elt = src_const_elt; | |
4813 | ||
4814 | /* See if there is a register linearly related to a constant | |
4815 | equivalent of SRC. */ | |
4816 | if (src_const | |
4817 | && (GET_CODE (src_const) == CONST | |
4818 | || (src_const_elt && src_const_elt->related_value != 0))) | |
cb10db9d | 4819 | { |
4820 | src_related = use_related_value (src_const, src_const_elt); | |
4821 | if (src_related) | |
4822 | { | |
752df20e | 4823 | struct table_elt *src_related_elt |
cb10db9d | 4824 | = lookup (src_related, HASH (src_related, mode), mode); |
752df20e | 4825 | if (src_related_elt && elt) |
cb10db9d | 4826 | { |
752df20e | 4827 | if (elt->first_same_value |
4828 | != src_related_elt->first_same_value) | |
cb10db9d | 4829 | /* This can occur when we previously saw a CONST |
752df20e | 4830 | involving a SYMBOL_REF and then see the SYMBOL_REF |
4831 | twice. Merge the involved classes. */ | |
4832 | merge_equiv_classes (elt, src_related_elt); | |
4833 | ||
cb10db9d | 4834 | src_related = 0; |
752df20e | 4835 | src_related_elt = 0; |
cb10db9d | 4836 | } |
4837 | else if (src_related_elt && elt == 0) | |
4838 | elt = src_related_elt; | |
752df20e | 4839 | } |
cb10db9d | 4840 | } |
752df20e | 4841 | |
4023cea7 | 4842 | /* See if we have a CONST_INT that is already in a register in a |
4843 | wider mode. */ | |
4844 | ||
971ba038 | 4845 | if (src_const && src_related == 0 && CONST_INT_P (src_const) |
4023cea7 | 4846 | && GET_MODE_CLASS (mode) == MODE_INT |
ded805e6 | 4847 | && GET_MODE_PRECISION (mode) < BITS_PER_WORD) |
4023cea7 | 4848 | { |
3754d046 | 4849 | machine_mode wider_mode; |
4023cea7 | 4850 | |
4851 | for (wider_mode = GET_MODE_WIDER_MODE (mode); | |
1595fd95 | 4852 | wider_mode != VOIDmode |
ded805e6 | 4853 | && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD |
4023cea7 | 4854 | && src_related == 0; |
4855 | wider_mode = GET_MODE_WIDER_MODE (wider_mode)) | |
4856 | { | |
4857 | struct table_elt *const_elt | |
4858 | = lookup (src_const, HASH (src_const, wider_mode), wider_mode); | |
4859 | ||
4860 | if (const_elt == 0) | |
4861 | continue; | |
4862 | ||
4863 | for (const_elt = const_elt->first_same_value; | |
4864 | const_elt; const_elt = const_elt->next_same_value) | |
8ad4c111 | 4865 | if (REG_P (const_elt->exp)) |
4023cea7 | 4866 | { |
8b172e0e | 4867 | src_related = gen_lowpart (mode, const_elt->exp); |
4023cea7 | 4868 | break; |
4869 | } | |
4870 | } | |
4871 | } | |
4872 | ||
f9e15121 | 4873 | /* Another possibility is that we have an AND with a constant in |
4874 | a mode narrower than a word. If so, it might have been generated | |
4875 | as part of an "if" which would narrow the AND. If we already | |
4876 | have done the AND in a wider mode, we can use a SUBREG of that | |
4877 | value. */ | |
4878 | ||
4879 | if (flag_expensive_optimizations && ! src_related | |
971ba038 | 4880 | && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1)) |
f9e15121 | 4881 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD) |
4882 | { | |
3754d046 | 4883 | machine_mode tmode; |
941522d6 | 4884 | rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1)); |
f9e15121 | 4885 | |
4886 | for (tmode = GET_MODE_WIDER_MODE (mode); | |
4887 | GET_MODE_SIZE (tmode) <= UNITS_PER_WORD; | |
4888 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
4889 | { | |
316f48ea | 4890 | rtx inner = gen_lowpart (tmode, XEXP (src, 0)); |
f9e15121 | 4891 | struct table_elt *larger_elt; |
4892 | ||
4893 | if (inner) | |
4894 | { | |
4895 | PUT_MODE (new_and, tmode); | |
4896 | XEXP (new_and, 0) = inner; | |
4897 | larger_elt = lookup (new_and, HASH (new_and, tmode), tmode); | |
4898 | if (larger_elt == 0) | |
4899 | continue; | |
4900 | ||
4901 | for (larger_elt = larger_elt->first_same_value; | |
4902 | larger_elt; larger_elt = larger_elt->next_same_value) | |
8ad4c111 | 4903 | if (REG_P (larger_elt->exp)) |
f9e15121 | 4904 | { |
4905 | src_related | |
316f48ea | 4906 | = gen_lowpart (mode, larger_elt->exp); |
f9e15121 | 4907 | break; |
4908 | } | |
4909 | ||
4910 | if (src_related) | |
4911 | break; | |
4912 | } | |
4913 | } | |
4914 | } | |
c13941f4 | 4915 | |
c13941f4 | 4916 | /* See if a MEM has already been loaded with a widening operation; |
4917 | if it has, we can use a subreg of that. Many CISC machines | |
4918 | also have such operations, but this is only likely to be | |
5aedf60c | 4919 | beneficial on these machines. */ |
cb10db9d | 4920 | |
b74befc5 | 4921 | if (flag_expensive_optimizations && src_related == 0 |
c13941f4 | 4922 | && (GET_MODE_SIZE (mode) < UNITS_PER_WORD) |
4923 | && GET_MODE_CLASS (mode) == MODE_INT | |
e16ceb8e | 4924 | && MEM_P (src) && ! do_not_record |
21f1e711 | 4925 | && LOAD_EXTEND_OP (mode) != UNKNOWN) |
c13941f4 | 4926 | { |
89333dfe | 4927 | struct rtx_def memory_extend_buf; |
4928 | rtx memory_extend_rtx = &memory_extend_buf; | |
3754d046 | 4929 | machine_mode tmode; |
cb10db9d | 4930 | |
c13941f4 | 4931 | /* Set what we are trying to extend and the operation it might |
4932 | have been extended with. */ | |
9af5ce0c | 4933 | memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx)); |
c13941f4 | 4934 | PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode)); |
4935 | XEXP (memory_extend_rtx, 0) = src; | |
cb10db9d | 4936 | |
c13941f4 | 4937 | for (tmode = GET_MODE_WIDER_MODE (mode); |
4938 | GET_MODE_SIZE (tmode) <= UNITS_PER_WORD; | |
4939 | tmode = GET_MODE_WIDER_MODE (tmode)) | |
4940 | { | |
4941 | struct table_elt *larger_elt; | |
cb10db9d | 4942 | |
c13941f4 | 4943 | PUT_MODE (memory_extend_rtx, tmode); |
cb10db9d | 4944 | larger_elt = lookup (memory_extend_rtx, |
c13941f4 | 4945 | HASH (memory_extend_rtx, tmode), tmode); |
4946 | if (larger_elt == 0) | |
4947 | continue; | |
cb10db9d | 4948 | |
c13941f4 | 4949 | for (larger_elt = larger_elt->first_same_value; |
4950 | larger_elt; larger_elt = larger_elt->next_same_value) | |
8ad4c111 | 4951 | if (REG_P (larger_elt->exp)) |
c13941f4 | 4952 | { |
8b172e0e | 4953 | src_related = gen_lowpart (mode, larger_elt->exp); |
c13941f4 | 4954 | break; |
4955 | } | |
cb10db9d | 4956 | |
c13941f4 | 4957 | if (src_related) |
4958 | break; | |
4959 | } | |
4960 | } | |
cb10db9d | 4961 | |
01c8e4c9 | 4962 | /* Try to express the constant using a register+offset expression |
4963 | derived from a constant anchor. */ | |
4964 | ||
4965 | if (targetm.const_anchor | |
4966 | && !src_related | |
4967 | && src_const | |
4968 | && GET_CODE (src_const) == CONST_INT) | |
4969 | { | |
4970 | src_related = try_const_anchors (src_const, mode); | |
4971 | src_related_is_const_anchor = src_related != NULL_RTX; | |
4972 | } | |
4973 | ||
4974 | ||
752df20e | 4975 | if (src == src_folded) |
cb10db9d | 4976 | src_folded = 0; |
752df20e | 4977 | |
d10cfa8d | 4978 | /* At this point, ELT, if nonzero, points to a class of expressions |
752df20e | 4979 | equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED, |
d10cfa8d | 4980 | and SRC_RELATED, if nonzero, each contain additional equivalent |
752df20e | 4981 | expressions. Prune these latter expressions by deleting expressions |
4982 | already in the equivalence class. | |
4983 | ||
4984 | Check for an equivalent identical to the destination. If found, | |
4985 | this is the preferred equivalent since it will likely lead to | |
4986 | elimination of the insn. Indicate this by placing it in | |
4987 | `src_related'. */ | |
4988 | ||
cb10db9d | 4989 | if (elt) |
4990 | elt = elt->first_same_value; | |
752df20e | 4991 | for (p = elt; p; p = p->next_same_value) |
cb10db9d | 4992 | { |
752df20e | 4993 | enum rtx_code code = GET_CODE (p->exp); |
4994 | ||
4995 | /* If the expression is not valid, ignore it. Then we do not | |
4996 | have to check for validity below. In most cases, we can use | |
4997 | `rtx_equal_p', since canonicalization has already been done. */ | |
78d140c9 | 4998 | if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false)) |
752df20e | 4999 | continue; |
5000 | ||
47ac60a3 | 5001 | /* Also skip paradoxical subregs, unless that's what we're |
5002 | looking for. */ | |
b537bfdb | 5003 | if (paradoxical_subreg_p (p->exp) |
47ac60a3 | 5004 | && ! (src != 0 |
5005 | && GET_CODE (src) == SUBREG | |
5006 | && GET_MODE (src) == GET_MODE (p->exp) | |
5007 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
5008 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp)))))) | |
5009 | continue; | |
5010 | ||
cb10db9d | 5011 | if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp)) |
752df20e | 5012 | src = 0; |
cb10db9d | 5013 | else if (src_folded && GET_CODE (src_folded) == code |
752df20e | 5014 | && rtx_equal_p (src_folded, p->exp)) |
5015 | src_folded = 0; | |
cb10db9d | 5016 | else if (src_eqv_here && GET_CODE (src_eqv_here) == code |
752df20e | 5017 | && rtx_equal_p (src_eqv_here, p->exp)) |
5018 | src_eqv_here = 0; | |
cb10db9d | 5019 | else if (src_related && GET_CODE (src_related) == code |
752df20e | 5020 | && rtx_equal_p (src_related, p->exp)) |
5021 | src_related = 0; | |
5022 | ||
5023 | /* This is the same as the destination of the insns, we want | |
5024 | to prefer it. Copy it to src_related. The code below will | |
5025 | then give it a negative cost. */ | |
5026 | if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest)) | |
5027 | src_related = dest; | |
cb10db9d | 5028 | } |
752df20e | 5029 | |
5030 | /* Find the cheapest valid equivalent, trying all the available | |
5031 | possibilities. Prefer items not in the hash table to ones | |
5032 | that are when they are equal cost. Note that we can never | |
5033 | worsen an insn as the current contents will also succeed. | |
e2ef73d2 | 5034 | If we find an equivalent identical to the destination, use it as best, |
a92771b8 | 5035 | since this insn will probably be eliminated in that case. */ |
752df20e | 5036 | if (src) |
5037 | { | |
5038 | if (rtx_equal_p (src, dest)) | |
589ff9e7 | 5039 | src_cost = src_regcost = -1; |
752df20e | 5040 | else |
d27eb4b1 | 5041 | { |
5ae4887d | 5042 | src_cost = COST (src, mode); |
d27eb4b1 | 5043 | src_regcost = approx_reg_cost (src); |
5044 | } | |
752df20e | 5045 | } |
5046 | ||
5047 | if (src_eqv_here) | |
5048 | { | |
5049 | if (rtx_equal_p (src_eqv_here, dest)) | |
589ff9e7 | 5050 | src_eqv_cost = src_eqv_regcost = -1; |
752df20e | 5051 | else |
d27eb4b1 | 5052 | { |
5ae4887d | 5053 | src_eqv_cost = COST (src_eqv_here, mode); |
d27eb4b1 | 5054 | src_eqv_regcost = approx_reg_cost (src_eqv_here); |
5055 | } | |
752df20e | 5056 | } |
5057 | ||
5058 | if (src_folded) | |
5059 | { | |
5060 | if (rtx_equal_p (src_folded, dest)) | |
589ff9e7 | 5061 | src_folded_cost = src_folded_regcost = -1; |
752df20e | 5062 | else |
d27eb4b1 | 5063 | { |
5ae4887d | 5064 | src_folded_cost = COST (src_folded, mode); |
d27eb4b1 | 5065 | src_folded_regcost = approx_reg_cost (src_folded); |
5066 | } | |
752df20e | 5067 | } |
5068 | ||
5069 | if (src_related) | |
5070 | { | |
5071 | if (rtx_equal_p (src_related, dest)) | |
589ff9e7 | 5072 | src_related_cost = src_related_regcost = -1; |
752df20e | 5073 | else |
d27eb4b1 | 5074 | { |
5ae4887d | 5075 | src_related_cost = COST (src_related, mode); |
d27eb4b1 | 5076 | src_related_regcost = approx_reg_cost (src_related); |
01c8e4c9 | 5077 | |
5078 | /* If a const-anchor is used to synthesize a constant that | |
5079 | normally requires multiple instructions then slightly prefer | |
5080 | it over the original sequence. These instructions are likely | |
5081 | to become redundant now. We can't compare against the cost | |
5082 | of src_eqv_here because, on MIPS for example, multi-insn | |
5083 | constants have zero cost; they are assumed to be hoisted from | |
5084 | loops. */ | |
5085 | if (src_related_is_const_anchor | |
5086 | && src_related_cost == src_cost | |
5087 | && src_eqv_here) | |
5088 | src_related_cost--; | |
d27eb4b1 | 5089 | } |
752df20e | 5090 | } |
5091 | ||
5092 | /* If this was an indirect jump insn, a known label will really be | |
5093 | cheaper even though it looks more expensive. */ | |
5094 | if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF) | |
fb561825 | 5095 | src_folded = src_const, src_folded_cost = src_folded_regcost = -1; |
cb10db9d | 5096 | |
752df20e | 5097 | /* Terminate loop when replacement made. This must terminate since |
5098 | the current contents will be tested and will always be valid. */ | |
5099 | while (1) | |
cb10db9d | 5100 | { |
5101 | rtx trial; | |
752df20e | 5102 | |
cb10db9d | 5103 | /* Skip invalid entries. */ |
8ad4c111 | 5104 | while (elt && !REG_P (elt->exp) |
78d140c9 | 5105 | && ! exp_equiv_p (elt->exp, elt->exp, 1, false)) |
cb10db9d | 5106 | elt = elt->next_same_value; |
47ac60a3 | 5107 | |
5108 | /* A paradoxical subreg would be bad here: it'll be the right | |
5109 | size, but later may be adjusted so that the upper bits aren't | |
5110 | what we want. So reject it. */ | |
5111 | if (elt != 0 | |
b537bfdb | 5112 | && paradoxical_subreg_p (elt->exp) |
47ac60a3 | 5113 | /* It is okay, though, if the rtx we're trying to match |
5114 | will ignore any of the bits we can't predict. */ | |
5115 | && ! (src != 0 | |
5116 | && GET_CODE (src) == SUBREG | |
5117 | && GET_MODE (src) == GET_MODE (elt->exp) | |
5118 | && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))) | |
5119 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp)))))) | |
5120 | { | |
5121 | elt = elt->next_same_value; | |
5122 | continue; | |
5123 | } | |
cb10db9d | 5124 | |
d4c5e26d | 5125 | if (elt) |
d27eb4b1 | 5126 | { |
5127 | src_elt_cost = elt->cost; | |
5128 | src_elt_regcost = elt->regcost; | |
5129 | } | |
752df20e | 5130 | |
d4c5e26d | 5131 | /* Find cheapest and skip it for the next time. For items |
752df20e | 5132 | of equal cost, use this order: |
5133 | src_folded, src, src_eqv, src_related and hash table entry. */ | |
fb561825 | 5134 | if (src_folded |
069eea26 | 5135 | && preferable (src_folded_cost, src_folded_regcost, |
5136 | src_cost, src_regcost) <= 0 | |
5137 | && preferable (src_folded_cost, src_folded_regcost, | |
5138 | src_eqv_cost, src_eqv_regcost) <= 0 | |
5139 | && preferable (src_folded_cost, src_folded_regcost, | |
5140 | src_related_cost, src_related_regcost) <= 0 | |
5141 | && preferable (src_folded_cost, src_folded_regcost, | |
5142 | src_elt_cost, src_elt_regcost) <= 0) | |
752df20e | 5143 | { |
589ff9e7 | 5144 | trial = src_folded, src_folded_cost = MAX_COST; |
752df20e | 5145 | if (src_folded_force_flag) |
d4a75790 | 5146 | { |
5147 | rtx forced = force_const_mem (mode, trial); | |
5148 | if (forced) | |
5149 | trial = forced; | |
5150 | } | |
752df20e | 5151 | } |
fb561825 | 5152 | else if (src |
069eea26 | 5153 | && preferable (src_cost, src_regcost, |
5154 | src_eqv_cost, src_eqv_regcost) <= 0 | |
5155 | && preferable (src_cost, src_regcost, | |
5156 | src_related_cost, src_related_regcost) <= 0 | |
5157 | && preferable (src_cost, src_regcost, | |
5158 | src_elt_cost, src_elt_regcost) <= 0) | |
589ff9e7 | 5159 | trial = src, src_cost = MAX_COST; |
fb561825 | 5160 | else if (src_eqv_here |
069eea26 | 5161 | && preferable (src_eqv_cost, src_eqv_regcost, |
5162 | src_related_cost, src_related_regcost) <= 0 | |
5163 | && preferable (src_eqv_cost, src_eqv_regcost, | |
5164 | src_elt_cost, src_elt_regcost) <= 0) | |
0806b508 | 5165 | trial = src_eqv_here, src_eqv_cost = MAX_COST; |
fb561825 | 5166 | else if (src_related |
069eea26 | 5167 | && preferable (src_related_cost, src_related_regcost, |
5168 | src_elt_cost, src_elt_regcost) <= 0) | |
0806b508 | 5169 | trial = src_related, src_related_cost = MAX_COST; |
cb10db9d | 5170 | else |
752df20e | 5171 | { |
0806b508 | 5172 | trial = elt->exp; |
752df20e | 5173 | elt = elt->next_same_value; |
589ff9e7 | 5174 | src_elt_cost = MAX_COST; |
752df20e | 5175 | } |
5176 | ||
5fe61d21 | 5177 | /* Avoid creation of overlapping memory moves. */ |
663f4248 | 5178 | if (MEM_P (trial) && MEM_P (dest) && !rtx_equal_p (trial, dest)) |
5fe61d21 | 5179 | { |
5180 | rtx src, dest; | |
5181 | ||
5182 | /* BLKmode moves are not handled by cse anyway. */ | |
5183 | if (GET_MODE (trial) == BLKmode) | |
5184 | break; | |
5185 | ||
5186 | src = canon_rtx (trial); | |
5187 | dest = canon_rtx (SET_DEST (sets[i].rtl)); | |
5188 | ||
5189 | if (!MEM_P (src) || !MEM_P (dest) | |
a84256aa | 5190 | || !nonoverlapping_memrefs_p (src, dest, false)) |
5fe61d21 | 5191 | break; |
5192 | } | |
5193 | ||
a49d9163 | 5194 | /* Try to optimize |
5195 | (set (reg:M N) (const_int A)) | |
5196 | (set (reg:M2 O) (const_int B)) | |
5197 | (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D)) | |
5198 | (reg:M2 O)). */ | |
5199 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT | |
5200 | && CONST_INT_P (trial) | |
5201 | && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1)) | |
5202 | && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2)) | |
5203 | && REG_P (XEXP (SET_DEST (sets[i].rtl), 0)) | |
ded805e6 | 5204 | && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl))) |
a49d9163 | 5205 | >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))) |
5206 | && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)) | |
5207 | + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2)) | |
5208 | <= HOST_BITS_PER_WIDE_INT)) | |
5209 | { | |
5210 | rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0); | |
5211 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
5212 | rtx pos = XEXP (SET_DEST (sets[i].rtl), 2); | |
5213 | unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg)); | |
5214 | struct table_elt *dest_elt | |
5215 | = lookup (dest_reg, dest_hash, GET_MODE (dest_reg)); | |
5216 | rtx dest_cst = NULL; | |
5217 | ||
5218 | if (dest_elt) | |
5219 | for (p = dest_elt->first_same_value; p; p = p->next_same_value) | |
5220 | if (p->is_const && CONST_INT_P (p->exp)) | |
5221 | { | |
5222 | dest_cst = p->exp; | |
5223 | break; | |
5224 | } | |
5225 | if (dest_cst) | |
5226 | { | |
5227 | HOST_WIDE_INT val = INTVAL (dest_cst); | |
5228 | HOST_WIDE_INT mask; | |
5229 | unsigned int shift; | |
5230 | if (BITS_BIG_ENDIAN) | |
ded805e6 | 5231 | shift = GET_MODE_PRECISION (GET_MODE (dest_reg)) |
a49d9163 | 5232 | - INTVAL (pos) - INTVAL (width); |
5233 | else | |
5234 | shift = INTVAL (pos); | |
5235 | if (INTVAL (width) == HOST_BITS_PER_WIDE_INT) | |
7097b942 | 5236 | mask = HOST_WIDE_INT_M1; |
a49d9163 | 5237 | else |
edc19fd0 | 5238 | mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1; |
a49d9163 | 5239 | val &= ~(mask << shift); |
5240 | val |= (INTVAL (trial) & mask) << shift; | |
5241 | val = trunc_int_for_mode (val, GET_MODE (dest_reg)); | |
5242 | validate_unshare_change (insn, &SET_DEST (sets[i].rtl), | |
5243 | dest_reg, 1); | |
5244 | validate_unshare_change (insn, &SET_SRC (sets[i].rtl), | |
5245 | GEN_INT (val), 1); | |
5246 | if (apply_change_group ()) | |
5247 | { | |
5248 | rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX); | |
5249 | if (note) | |
5250 | { | |
5251 | remove_note (insn, note); | |
5252 | df_notes_rescan (insn); | |
5253 | } | |
5254 | src_eqv = NULL_RTX; | |
5255 | src_eqv_elt = NULL; | |
5256 | src_eqv_volatile = 0; | |
5257 | src_eqv_in_memory = 0; | |
5258 | src_eqv_hash = 0; | |
5259 | repeat = true; | |
5260 | break; | |
5261 | } | |
5262 | } | |
5263 | } | |
5264 | ||
752df20e | 5265 | /* We don't normally have an insn matching (set (pc) (pc)), so |
5266 | check for this separately here. We will delete such an | |
5267 | insn below. | |
5268 | ||
0f48207f | 5269 | For other cases such as a table jump or conditional jump |
5270 | where we know the ultimate target, go ahead and replace the | |
5271 | operand. While that may not make a valid insn, we will | |
5272 | reemit the jump below (and also insert any necessary | |
5273 | barriers). */ | |
752df20e | 5274 | if (n_sets == 1 && dest == pc_rtx |
5275 | && (trial == pc_rtx | |
5276 | || (GET_CODE (trial) == LABEL_REF | |
5277 | && ! condjump_p (insn)))) | |
5278 | { | |
806351c6 | 5279 | /* Don't substitute non-local labels, this confuses CFG. */ |
5280 | if (GET_CODE (trial) == LABEL_REF | |
5281 | && LABEL_REF_NONLOCAL_P (trial)) | |
5282 | continue; | |
5283 | ||
0f48207f | 5284 | SET_SRC (sets[i].rtl) = trial; |
283a6b26 | 5285 | cse_jumps_altered = true; |
752df20e | 5286 | break; |
5287 | } | |
cb10db9d | 5288 | |
663f4248 | 5289 | /* Similarly, lots of targets don't allow no-op |
5290 | (set (mem x) (mem x)) moves. */ | |
5291 | else if (n_sets == 1 | |
5292 | && MEM_P (trial) | |
5293 | && MEM_P (dest) | |
5294 | && rtx_equal_p (trial, dest) | |
5295 | && !side_effects_p (dest) | |
5296 | && (cfun->can_delete_dead_exceptions | |
5297 | || insn_nothrow_p (insn))) | |
5298 | { | |
5299 | SET_SRC (sets[i].rtl) = trial; | |
5300 | mem_noop_insn = true; | |
5301 | break; | |
5302 | } | |
5303 | ||
0ab04fbf | 5304 | /* Reject certain invalid forms of CONST that we create. */ |
5305 | else if (CONSTANT_P (trial) | |
5306 | && GET_CODE (trial) == CONST | |
5307 | /* Reject cases that will cause decode_rtx_const to | |
5308 | die. On the alpha when simplifying a switch, we | |
5309 | get (const (truncate (minus (label_ref) | |
5310 | (label_ref)))). */ | |
5311 | && (GET_CODE (XEXP (trial, 0)) == TRUNCATE | |
5312 | /* Likewise on IA-64, except without the | |
5313 | truncate. */ | |
5314 | || (GET_CODE (XEXP (trial, 0)) == MINUS | |
5315 | && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF | |
5316 | && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF))) | |
5317 | /* Do nothing for this case. */ | |
5318 | ; | |
5319 | ||
752df20e | 5320 | /* Look for a substitution that makes a valid insn. */ |
20d3ff08 | 5321 | else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl), |
5322 | trial, 0)) | |
e2ef73d2 | 5323 | { |
d328ebdf | 5324 | rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn); |
e30d7fb3 | 5325 | |
8d5dd220 | 5326 | /* The result of apply_change_group can be ignored; see |
5327 | canon_reg. */ | |
5328 | ||
d328ebdf | 5329 | validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1); |
e443ebaf | 5330 | apply_change_group (); |
be22716f | 5331 | |
e2ef73d2 | 5332 | break; |
5333 | } | |
752df20e | 5334 | |
cb10db9d | 5335 | /* If we previously found constant pool entries for |
752df20e | 5336 | constants and this is a constant, try making a |
5337 | pool entry. Put it in src_folded unless we already have done | |
5338 | this since that is where it likely came from. */ | |
5339 | ||
5340 | else if (constant_pool_entries_cost | |
5341 | && CONSTANT_P (trial) | |
88f6e1a4 | 5342 | && (src_folded == 0 |
e16ceb8e | 5343 | || (!MEM_P (src_folded) |
88f6e1a4 | 5344 | && ! src_folded_force_flag)) |
ea0cb7ae | 5345 | && GET_MODE_CLASS (mode) != MODE_CC |
5346 | && mode != VOIDmode) | |
752df20e | 5347 | { |
5348 | src_folded_force_flag = 1; | |
5349 | src_folded = trial; | |
5350 | src_folded_cost = constant_pool_entries_cost; | |
634d45d7 | 5351 | src_folded_regcost = constant_pool_entries_regcost; |
752df20e | 5352 | } |
cb10db9d | 5353 | } |
752df20e | 5354 | |
a49d9163 | 5355 | /* If we changed the insn too much, handle this set from scratch. */ |
5356 | if (repeat) | |
5357 | { | |
5358 | i--; | |
5359 | continue; | |
5360 | } | |
5361 | ||
752df20e | 5362 | src = SET_SRC (sets[i].rtl); |
5363 | ||
5364 | /* In general, it is good to have a SET with SET_SRC == SET_DEST. | |
5365 | However, there is an important exception: If both are registers | |
5366 | that are not the head of their equivalence class, replace SET_SRC | |
5367 | with the head of the class. If we do not do this, we will have | |
5368 | both registers live over a portion of the basic block. This way, | |
5369 | their lifetimes will likely abut instead of overlapping. */ | |
8ad4c111 | 5370 | if (REG_P (dest) |
a7f3b1c7 | 5371 | && REGNO_QTY_VALID_P (REGNO (dest))) |
752df20e | 5372 | { |
a7f3b1c7 | 5373 | int dest_q = REG_QTY (REGNO (dest)); |
5374 | struct qty_table_elem *dest_ent = &qty_table[dest_q]; | |
5375 | ||
5376 | if (dest_ent->mode == GET_MODE (dest) | |
5377 | && dest_ent->first_reg != REGNO (dest) | |
8ad4c111 | 5378 | && REG_P (src) && REGNO (src) == REGNO (dest) |
a7f3b1c7 | 5379 | /* Don't do this if the original insn had a hard reg as |
5380 | SET_SRC or SET_DEST. */ | |
8ad4c111 | 5381 | && (!REG_P (sets[i].src) |
a7f3b1c7 | 5382 | || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER) |
8ad4c111 | 5383 | && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER)) |
a7f3b1c7 | 5384 | /* We can't call canon_reg here because it won't do anything if |
5385 | SRC is a hard register. */ | |
05b1716f | 5386 | { |
a7f3b1c7 | 5387 | int src_q = REG_QTY (REGNO (src)); |
5388 | struct qty_table_elem *src_ent = &qty_table[src_q]; | |
5389 | int first = src_ent->first_reg; | |
5390 | rtx new_src | |
5391 | = (first >= FIRST_PSEUDO_REGISTER | |
5392 | ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first)); | |
5393 | ||
5394 | /* We must use validate-change even for this, because this | |
5395 | might be a special no-op instruction, suitable only to | |
5396 | tag notes onto. */ | |
5397 | if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0)) | |
5398 | { | |
5399 | src = new_src; | |
5400 | /* If we had a constant that is cheaper than what we are now | |
5401 | setting SRC to, use that constant. We ignored it when we | |
5402 | thought we could make this into a no-op. */ | |
5ae4887d | 5403 | if (src_const && COST (src_const, mode) < COST (src, mode) |
cb10db9d | 5404 | && validate_change (insn, &SET_SRC (sets[i].rtl), |
5405 | src_const, 0)) | |
a7f3b1c7 | 5406 | src = src_const; |
5407 | } | |
05b1716f | 5408 | } |
752df20e | 5409 | } |
5410 | ||
5411 | /* If we made a change, recompute SRC values. */ | |
5412 | if (src != sets[i].src) | |
cb10db9d | 5413 | { |
cb10db9d | 5414 | do_not_record = 0; |
5415 | hash_arg_in_memory = 0; | |
752df20e | 5416 | sets[i].src = src; |
cb10db9d | 5417 | sets[i].src_hash = HASH (src, mode); |
5418 | sets[i].src_volatile = do_not_record; | |
5419 | sets[i].src_in_memory = hash_arg_in_memory; | |
5420 | sets[i].src_elt = lookup (src, sets[i].src_hash, mode); | |
5421 | } | |
752df20e | 5422 | |
5423 | /* If this is a single SET, we are setting a register, and we have an | |
a24ec999 | 5424 | equivalent constant, we want to add a REG_EQUAL note if the constant |
5425 | is different from the source. We don't want to do it for a constant | |
5426 | pseudo since verifying that this pseudo hasn't been eliminated is a | |
5427 | pain; moreover such a note won't help anything. | |
f5d1f9f9 | 5428 | |
5429 | Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF))) | |
5430 | which can be created for a reference to a compile time computable | |
5431 | entry in a jump table. */ | |
a24ec999 | 5432 | if (n_sets == 1 |
5433 | && REG_P (dest) | |
5434 | && src_const | |
8ad4c111 | 5435 | && !REG_P (src_const) |
a24ec999 | 5436 | && !(GET_CODE (src_const) == SUBREG |
5437 | && REG_P (SUBREG_REG (src_const))) | |
5438 | && !(GET_CODE (src_const) == CONST | |
5439 | && GET_CODE (XEXP (src_const, 0)) == MINUS | |
5440 | && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF | |
5441 | && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF) | |
5442 | && !rtx_equal_p (src, src_const)) | |
752df20e | 5443 | { |
a24ec999 | 5444 | /* Make sure that the rtx is not shared. */ |
5445 | src_const = copy_rtx (src_const); | |
c69ad724 | 5446 | |
a24ec999 | 5447 | /* Record the actual constant value in a REG_EQUAL note, |
5448 | making a new one if one does not already exist. */ | |
5449 | set_unique_reg_note (insn, REG_EQUAL, src_const); | |
5450 | df_notes_rescan (insn); | |
752df20e | 5451 | } |
5452 | ||
5453 | /* Now deal with the destination. */ | |
5454 | do_not_record = 0; | |
752df20e | 5455 | |
476d094d | 5456 | /* Look within any ZERO_EXTRACT to the MEM or REG within it. */ |
5457 | while (GET_CODE (dest) == SUBREG | |
752df20e | 5458 | || GET_CODE (dest) == ZERO_EXTRACT |
752df20e | 5459 | || GET_CODE (dest) == STRICT_LOW_PART) |
8f4cc641 | 5460 | dest = XEXP (dest, 0); |
752df20e | 5461 | |
5462 | sets[i].inner_dest = dest; | |
5463 | ||
e16ceb8e | 5464 | if (MEM_P (dest)) |
752df20e | 5465 | { |
ea0cb7ae | 5466 | #ifdef PUSH_ROUNDING |
5467 | /* Stack pushes invalidate the stack pointer. */ | |
5468 | rtx addr = XEXP (dest, 0); | |
6720e96c | 5469 | if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC |
ea0cb7ae | 5470 | && XEXP (addr, 0) == stack_pointer_rtx) |
4c958a22 | 5471 | invalidate (stack_pointer_rtx, VOIDmode); |
ea0cb7ae | 5472 | #endif |
752df20e | 5473 | dest = fold_rtx (dest, insn); |
752df20e | 5474 | } |
5475 | ||
5476 | /* Compute the hash code of the destination now, | |
5477 | before the effects of this instruction are recorded, | |
5478 | since the register values used in the address computation | |
5479 | are those before this instruction. */ | |
952bc06d | 5480 | sets[i].dest_hash = HASH (dest, mode); |
752df20e | 5481 | |
5482 | /* Don't enter a bit-field in the hash table | |
5483 | because the value in it after the store | |
5484 | may not equal what was stored, due to truncation. */ | |
5485 | ||
476d094d | 5486 | if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT) |
752df20e | 5487 | { |
5488 | rtx width = XEXP (SET_DEST (sets[i].rtl), 1); | |
5489 | ||
971ba038 | 5490 | if (src_const != 0 && CONST_INT_P (src_const) |
5491 | && CONST_INT_P (width) | |
b572011e | 5492 | && INTVAL (width) < HOST_BITS_PER_WIDE_INT |
5493 | && ! (INTVAL (src_const) | |
561f0ec8 | 5494 | & (HOST_WIDE_INT_M1U << INTVAL (width)))) |
752df20e | 5495 | /* Exception: if the value is constant, |
5496 | and it won't be truncated, record it. */ | |
5497 | ; | |
5498 | else | |
5499 | { | |
5500 | /* This is chosen so that the destination will be invalidated | |
5501 | but no new value will be recorded. | |
5502 | We must invalidate because sometimes constant | |
5503 | values can be recorded for bitfields. */ | |
5504 | sets[i].src_elt = 0; | |
5505 | sets[i].src_volatile = 1; | |
5506 | src_eqv = 0; | |
5507 | src_eqv_elt = 0; | |
5508 | } | |
5509 | } | |
5510 | ||
5511 | /* If only one set in a JUMP_INSN and it is now a no-op, we can delete | |
5512 | the insn. */ | |
5513 | else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx) | |
5514 | { | |
25999090 | 5515 | /* One less use of the label this insn used to jump to. */ |
bfa8ea12 | 5516 | cse_cfg_altered |= delete_insn_and_edges (insn); |
283a6b26 | 5517 | cse_jumps_altered = true; |
752df20e | 5518 | /* No more processing for this set. */ |
5519 | sets[i].rtl = 0; | |
5520 | } | |
5521 | ||
663f4248 | 5522 | /* Similarly for no-op MEM moves. */ |
5523 | else if (mem_noop_insn) | |
5524 | { | |
5525 | if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) | |
5526 | cse_cfg_altered = true; | |
bfa8ea12 | 5527 | cse_cfg_altered |= delete_insn_and_edges (insn); |
663f4248 | 5528 | /* No more processing for this set. */ |
5529 | sets[i].rtl = 0; | |
5530 | } | |
5531 | ||
752df20e | 5532 | /* If this SET is now setting PC to a label, we know it used to |
0f48207f | 5533 | be a conditional or computed branch. */ |
9d95b2b0 | 5534 | else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF |
5535 | && !LABEL_REF_NONLOCAL_P (src)) | |
752df20e | 5536 | { |
0f48207f | 5537 | /* We reemit the jump in as many cases as possible just in |
5538 | case the form of an unconditional jump is significantly | |
5539 | different than a computed jump or conditional jump. | |
5540 | ||
5541 | If this insn has multiple sets, then reemitting the | |
5542 | jump is nontrivial. So instead we just force rerecognition | |
5543 | and hope for the best. */ | |
5544 | if (n_sets == 1) | |
752df20e | 5545 | { |
9ed997be | 5546 | rtx_jump_insn *new_rtx; |
32a6f3ca | 5547 | rtx note; |
743ce3f8 | 5548 | |
1d5ad681 | 5549 | rtx_insn *seq = targetm.gen_jump (XEXP (src, 0)); |
5550 | new_rtx = emit_jump_insn_before (seq, insn); | |
d328ebdf | 5551 | JUMP_LABEL (new_rtx) = XEXP (src, 0); |
752df20e | 5552 | LABEL_NUSES (XEXP (src, 0))++; |
9074c68b | 5553 | |
5554 | /* Make sure to copy over REG_NON_LOCAL_GOTO. */ | |
5555 | note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0); | |
5556 | if (note) | |
5557 | { | |
5558 | XEXP (note, 1) = NULL_RTX; | |
d328ebdf | 5559 | REG_NOTES (new_rtx) = note; |
9074c68b | 5560 | } |
5561 | ||
bfa8ea12 | 5562 | cse_cfg_altered |= delete_insn_and_edges (insn); |
32a6f3ca | 5563 | insn = new_rtx; |
752df20e | 5564 | } |
d578a436 | 5565 | else |
d578a436 | 5566 | INSN_CODE (insn) = -1; |
752df20e | 5567 | |
283a6b26 | 5568 | /* Do not bother deleting any unreachable code, let jump do it. */ |
5569 | cse_jumps_altered = true; | |
752df20e | 5570 | sets[i].rtl = 0; |
5571 | } | |
5572 | ||
8cdd0f84 | 5573 | /* If destination is volatile, invalidate it and then do no further |
5574 | processing for this assignment. */ | |
752df20e | 5575 | |
5576 | else if (do_not_record) | |
8cdd0f84 | 5577 | { |
7a49a822 | 5578 | invalidate_dest (dest); |
8cdd0f84 | 5579 | sets[i].rtl = 0; |
5580 | } | |
752df20e | 5581 | |
5582 | if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl)) | |
7a49a822 | 5583 | { |
5584 | do_not_record = 0; | |
5585 | sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode); | |
5586 | if (do_not_record) | |
5587 | { | |
5588 | invalidate_dest (SET_DEST (sets[i].rtl)); | |
5589 | sets[i].rtl = 0; | |
5590 | } | |
5591 | } | |
752df20e | 5592 | |
752df20e | 5593 | /* If setting CC0, record what it was set to, or a constant, if it |
5594 | is equivalent to a constant. If it is being set to a floating-point | |
5595 | value, make a COMPARE with the appropriate constant of 0. If we | |
5596 | don't do this, later code can interpret this as a test against | |
5597 | const0_rtx, which can cause problems if we try to put it into an | |
5598 | insn as a floating-point operand. */ | |
5599 | if (dest == cc0_rtx) | |
5600 | { | |
5601 | this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src; | |
5602 | this_insn_cc0_mode = mode; | |
c1712420 | 5603 | if (FLOAT_MODE_P (mode)) |
941522d6 | 5604 | this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0, |
5605 | CONST0_RTX (mode)); | |
752df20e | 5606 | } |
752df20e | 5607 | } |
5608 | ||
5609 | /* Now enter all non-volatile source expressions in the hash table | |
5610 | if they are not already present. | |
5611 | Record their equivalence classes in src_elt. | |
5612 | This way we can insert the corresponding destinations into | |
5613 | the same classes even if the actual sources are no longer in them | |
5614 | (having been invalidated). */ | |
5615 | ||
5616 | if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile | |
5617 | && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl))) | |
5618 | { | |
19cb6b50 | 5619 | struct table_elt *elt; |
5620 | struct table_elt *classp = sets[0].src_elt; | |
752df20e | 5621 | rtx dest = SET_DEST (sets[0].rtl); |
3754d046 | 5622 | machine_mode eqvmode = GET_MODE (dest); |
752df20e | 5623 | |
5624 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
5625 | { | |
5626 | eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0))); | |
5627 | classp = 0; | |
5628 | } | |
5629 | if (insert_regs (src_eqv, classp, 0)) | |
1b033cc3 | 5630 | { |
5631 | rehash_using_reg (src_eqv); | |
5632 | src_eqv_hash = HASH (src_eqv, eqvmode); | |
5633 | } | |
952bc06d | 5634 | elt = insert (src_eqv, classp, src_eqv_hash, eqvmode); |
752df20e | 5635 | elt->in_memory = src_eqv_in_memory; |
752df20e | 5636 | src_eqv_elt = elt; |
c697ea36 | 5637 | |
5638 | /* Check to see if src_eqv_elt is the same as a set source which | |
5639 | does not yet have an elt, and if so set the elt of the set source | |
5640 | to src_eqv_elt. */ | |
5641 | for (i = 0; i < n_sets; i++) | |
cf541778 | 5642 | if (sets[i].rtl && sets[i].src_elt == 0 |
5643 | && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv)) | |
c697ea36 | 5644 | sets[i].src_elt = src_eqv_elt; |
752df20e | 5645 | } |
5646 | ||
5647 | for (i = 0; i < n_sets; i++) | |
5648 | if (sets[i].rtl && ! sets[i].src_volatile | |
5649 | && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl))) | |
5650 | { | |
5651 | if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART) | |
5652 | { | |
5653 | /* REG_EQUAL in setting a STRICT_LOW_PART | |
5654 | gives an equivalent for the entire destination register, | |
5655 | not just for the subreg being stored in now. | |
5656 | This is a more interesting equivalence, so we arrange later | |
5657 | to treat the entire reg as the destination. */ | |
5658 | sets[i].src_elt = src_eqv_elt; | |
952bc06d | 5659 | sets[i].src_hash = src_eqv_hash; |
752df20e | 5660 | } |
5661 | else | |
5662 | { | |
5663 | /* Insert source and constant equivalent into hash table, if not | |
5664 | already present. */ | |
19cb6b50 | 5665 | struct table_elt *classp = src_eqv_elt; |
5666 | rtx src = sets[i].src; | |
5667 | rtx dest = SET_DEST (sets[i].rtl); | |
3754d046 | 5668 | machine_mode mode |
752df20e | 5669 | = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src); |
5670 | ||
3c512ee7 | 5671 | /* It's possible that we have a source value known to be |
5672 | constant but don't have a REG_EQUAL note on the insn. | |
5673 | Lack of a note will mean src_eqv_elt will be NULL. This | |
5674 | can happen where we've generated a SUBREG to access a | |
5675 | CONST_INT that is already in a register in a wider mode. | |
5676 | Ensure that the source expression is put in the proper | |
5677 | constant class. */ | |
5678 | if (!classp) | |
5679 | classp = sets[i].src_const_elt; | |
5680 | ||
cf541778 | 5681 | if (sets[i].src_elt == 0) |
752df20e | 5682 | { |
1e5b92fa | 5683 | struct table_elt *elt; |
cf541778 | 5684 | |
1e5b92fa | 5685 | /* Note that these insert_regs calls cannot remove |
5686 | any of the src_elt's, because they would have failed to | |
5687 | match if not still valid. */ | |
5688 | if (insert_regs (src, classp, 0)) | |
5689 | { | |
5690 | rehash_using_reg (src); | |
5691 | sets[i].src_hash = HASH (src, mode); | |
1b033cc3 | 5692 | } |
1e5b92fa | 5693 | elt = insert (src, classp, sets[i].src_hash, mode); |
5694 | elt->in_memory = sets[i].src_in_memory; | |
20d3ff08 | 5695 | /* If inline asm has any clobbers, ensure we only reuse |
5696 | existing inline asms and never try to put the ASM_OPERANDS | |
5697 | into an insn that isn't inline asm. */ | |
5698 | if (GET_CODE (src) == ASM_OPERANDS | |
5699 | && GET_CODE (x) == PARALLEL) | |
5700 | elt->cost = MAX_COST; | |
1e5b92fa | 5701 | sets[i].src_elt = classp = elt; |
752df20e | 5702 | } |
752df20e | 5703 | if (sets[i].src_const && sets[i].src_const_elt == 0 |
5704 | && src != sets[i].src_const | |
5705 | && ! rtx_equal_p (sets[i].src_const, src)) | |
5706 | sets[i].src_elt = insert (sets[i].src_const, classp, | |
952bc06d | 5707 | sets[i].src_const_hash, mode); |
752df20e | 5708 | } |
5709 | } | |
5710 | else if (sets[i].src_elt == 0) | |
5711 | /* If we did not insert the source into the hash table (e.g., it was | |
5712 | volatile), note the equivalence class for the REG_EQUAL value, if any, | |
5713 | so that the destination goes into that class. */ | |
5714 | sets[i].src_elt = src_eqv_elt; | |
5715 | ||
977ffed2 | 5716 | /* Record destination addresses in the hash table. This allows us to |
5717 | check if they are invalidated by other sets. */ | |
5718 | for (i = 0; i < n_sets; i++) | |
5719 | { | |
5720 | if (sets[i].rtl) | |
5721 | { | |
5722 | rtx x = sets[i].inner_dest; | |
5723 | struct table_elt *elt; | |
3754d046 | 5724 | machine_mode mode; |
977ffed2 | 5725 | unsigned hash; |
5726 | ||
5727 | if (MEM_P (x)) | |
5728 | { | |
5729 | x = XEXP (x, 0); | |
5730 | mode = GET_MODE (x); | |
5731 | hash = HASH (x, mode); | |
5732 | elt = lookup (x, hash, mode); | |
5733 | if (!elt) | |
5734 | { | |
5735 | if (insert_regs (x, NULL, 0)) | |
5736 | { | |
06320855 | 5737 | rtx dest = SET_DEST (sets[i].rtl); |
5738 | ||
977ffed2 | 5739 | rehash_using_reg (x); |
5740 | hash = HASH (x, mode); | |
06320855 | 5741 | sets[i].dest_hash = HASH (dest, GET_MODE (dest)); |
977ffed2 | 5742 | } |
5743 | elt = insert (x, NULL, hash, mode); | |
5744 | } | |
5745 | ||
5746 | sets[i].dest_addr_elt = elt; | |
5747 | } | |
5748 | else | |
5749 | sets[i].dest_addr_elt = NULL; | |
5750 | } | |
5751 | } | |
5752 | ||
2aca5650 | 5753 | invalidate_from_clobbers (insn); |
8b82837b | 5754 | |
cb10db9d | 5755 | /* Some registers are invalidated by subroutine calls. Memory is |
8b82837b | 5756 | invalidated by non-constant calls. */ |
5757 | ||
6d7dc5b9 | 5758 | if (CALL_P (insn)) |
752df20e | 5759 | { |
9c2a0c05 | 5760 | if (!(RTL_CONST_OR_PURE_CALL_P (insn))) |
ea0cb7ae | 5761 | invalidate_memory (); |
33698dfe | 5762 | else |
5763 | /* For const/pure calls, invalidate any argument slots, because | |
5764 | those are owned by the callee. */ | |
5765 | for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1)) | |
5766 | if (GET_CODE (XEXP (tem, 0)) == USE | |
5767 | && MEM_P (XEXP (XEXP (tem, 0), 0))) | |
5768 | invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode); | |
752df20e | 5769 | invalidate_for_call (); |
5770 | } | |
5771 | ||
5772 | /* Now invalidate everything set by this instruction. | |
5773 | If a SUBREG or other funny destination is being set, | |
5774 | sets[i].rtl is still nonzero, so here we invalidate the reg | |
5775 | a part of which is being set. */ | |
5776 | ||
5777 | for (i = 0; i < n_sets; i++) | |
5778 | if (sets[i].rtl) | |
5779 | { | |
fdb25961 | 5780 | /* We can't use the inner dest, because the mode associated with |
5781 | a ZERO_EXTRACT is significant. */ | |
19cb6b50 | 5782 | rtx dest = SET_DEST (sets[i].rtl); |
752df20e | 5783 | |
5784 | /* Needed for registers to remove the register from its | |
5785 | previous quantity's chain. | |
5786 | Needed for memory if this is a nonvarying address, unless | |
5787 | we have just done an invalidate_memory that covers even those. */ | |
8ad4c111 | 5788 | if (REG_P (dest) || GET_CODE (dest) == SUBREG) |
fdb25961 | 5789 | invalidate (dest, VOIDmode); |
e16ceb8e | 5790 | else if (MEM_P (dest)) |
2046d6d5 | 5791 | invalidate (dest, VOIDmode); |
319134e7 | 5792 | else if (GET_CODE (dest) == STRICT_LOW_PART |
5793 | || GET_CODE (dest) == ZERO_EXTRACT) | |
fdb25961 | 5794 | invalidate (XEXP (dest, 0), GET_MODE (dest)); |
752df20e | 5795 | } |
5796 | ||
be22716f | 5797 | /* Don't cse over a call to setjmp; on some machines (eg VAX) |
5798 | the regs restored by the longjmp come from a later time | |
5799 | than the setjmp. */ | |
5800 | if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL)) | |
5801 | { | |
5802 | flush_hash_table (); | |
5803 | goto done; | |
5804 | } | |
5805 | ||
752df20e | 5806 | /* Make sure registers mentioned in destinations |
5807 | are safe for use in an expression to be inserted. | |
5808 | This removes from the hash table | |
5809 | any invalid entry that refers to one of these registers. | |
5810 | ||
5811 | We don't care about the return value from mention_regs because | |
5812 | we are going to hash the SET_DEST values unconditionally. */ | |
5813 | ||
5814 | for (i = 0; i < n_sets; i++) | |
e6860d27 | 5815 | { |
5816 | if (sets[i].rtl) | |
5817 | { | |
5818 | rtx x = SET_DEST (sets[i].rtl); | |
5819 | ||
8ad4c111 | 5820 | if (!REG_P (x)) |
e6860d27 | 5821 | mention_regs (x); |
5822 | else | |
5823 | { | |
5824 | /* We used to rely on all references to a register becoming | |
5825 | inaccessible when a register changes to a new quantity, | |
5826 | since that changes the hash code. However, that is not | |
9c4f3716 | 5827 | safe, since after HASH_SIZE new quantities we get a |
e6860d27 | 5828 | hash 'collision' of a register with its own invalid |
5829 | entries. And since SUBREGs have been changed not to | |
5830 | change their hash code with the hash code of the register, | |
5831 | it wouldn't work any longer at all. So we have to check | |
5832 | for any invalid references lying around now. | |
5833 | This code is similar to the REG case in mention_regs, | |
5834 | but it knows that reg_tick has been incremented, and | |
5835 | it leaves reg_in_table as -1 . */ | |
02e7a332 | 5836 | unsigned int regno = REGNO (x); |
a2c6f0b7 | 5837 | unsigned int endregno = END_REGNO (x); |
02e7a332 | 5838 | unsigned int i; |
e6860d27 | 5839 | |
5840 | for (i = regno; i < endregno; i++) | |
5841 | { | |
d1264606 | 5842 | if (REG_IN_TABLE (i) >= 0) |
e6860d27 | 5843 | { |
5844 | remove_invalid_refs (i); | |
d1264606 | 5845 | REG_IN_TABLE (i) = -1; |
e6860d27 | 5846 | } |
5847 | } | |
5848 | } | |
5849 | } | |
5850 | } | |
752df20e | 5851 | |
5852 | /* We may have just removed some of the src_elt's from the hash table. | |
977ffed2 | 5853 | So replace each one with the current head of the same class. |
5854 | Also check if destination addresses have been removed. */ | |
752df20e | 5855 | |
5856 | for (i = 0; i < n_sets; i++) | |
5857 | if (sets[i].rtl) | |
5858 | { | |
977ffed2 | 5859 | if (sets[i].dest_addr_elt |
5860 | && sets[i].dest_addr_elt->first_same_value == 0) | |
5861 | { | |
d249588e | 5862 | /* The elt was removed, which means this destination is not |
977ffed2 | 5863 | valid after this instruction. */ |
5864 | sets[i].rtl = NULL_RTX; | |
5865 | } | |
5866 | else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0) | |
752df20e | 5867 | /* If elt was removed, find current head of same class, |
5868 | or 0 if nothing remains of that class. */ | |
5869 | { | |
19cb6b50 | 5870 | struct table_elt *elt = sets[i].src_elt; |
752df20e | 5871 | |
5872 | while (elt && elt->prev_same_value) | |
5873 | elt = elt->prev_same_value; | |
5874 | ||
5875 | while (elt && elt->first_same_value == 0) | |
5876 | elt = elt->next_same_value; | |
5877 | sets[i].src_elt = elt ? elt->first_same_value : 0; | |
5878 | } | |
5879 | } | |
5880 | ||
5881 | /* Now insert the destinations into their equivalence classes. */ | |
5882 | ||
5883 | for (i = 0; i < n_sets; i++) | |
5884 | if (sets[i].rtl) | |
5885 | { | |
19cb6b50 | 5886 | rtx dest = SET_DEST (sets[i].rtl); |
19cb6b50 | 5887 | struct table_elt *elt; |
752df20e | 5888 | |
5889 | /* Don't record value if we are not supposed to risk allocating | |
5890 | floating-point values in registers that might be wider than | |
5891 | memory. */ | |
5892 | if ((flag_float_store | |
e16ceb8e | 5893 | && MEM_P (dest) |
c1712420 | 5894 | && FLOAT_MODE_P (GET_MODE (dest))) |
6510de05 | 5895 | /* Don't record BLKmode values, because we don't know the |
5896 | size of it, and can't be sure that other BLKmode values | |
5897 | have the same or smaller size. */ | |
5898 | || GET_MODE (dest) == BLKmode | |
752df20e | 5899 | /* If we didn't put a REG_EQUAL value or a source into the hash |
5900 | table, there is no point is recording DEST. */ | |
619142e5 | 5901 | || sets[i].src_elt == 0 |
5902 | /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND | |
5903 | or SIGN_EXTEND, don't record DEST since it can cause | |
5904 | some tracking to be wrong. | |
5905 | ||
5906 | ??? Think about this more later. */ | |
b537bfdb | 5907 | || (paradoxical_subreg_p (dest) |
619142e5 | 5908 | && (GET_CODE (sets[i].src) == SIGN_EXTEND |
5909 | || GET_CODE (sets[i].src) == ZERO_EXTEND))) | |
752df20e | 5910 | continue; |
5911 | ||
5912 | /* STRICT_LOW_PART isn't part of the value BEING set, | |
5913 | and neither is the SUBREG inside it. | |
5914 | Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */ | |
5915 | if (GET_CODE (dest) == STRICT_LOW_PART) | |
5916 | dest = SUBREG_REG (XEXP (dest, 0)); | |
5917 | ||
8ad4c111 | 5918 | if (REG_P (dest) || GET_CODE (dest) == SUBREG) |
752df20e | 5919 | /* Registers must also be inserted into chains for quantities. */ |
5920 | if (insert_regs (dest, sets[i].src_elt, 1)) | |
1b033cc3 | 5921 | { |
5922 | /* If `insert_regs' changes something, the hash code must be | |
5923 | recalculated. */ | |
5924 | rehash_using_reg (dest); | |
5925 | sets[i].dest_hash = HASH (dest, GET_MODE (dest)); | |
5926 | } | |
752df20e | 5927 | |
e8825bb0 | 5928 | elt = insert (dest, sets[i].src_elt, |
5929 | sets[i].dest_hash, GET_MODE (dest)); | |
a97275a9 | 5930 | |
01c8e4c9 | 5931 | /* If this is a constant, insert the constant anchors with the |
5932 | equivalent register-offset expressions using register DEST. */ | |
5933 | if (targetm.const_anchor | |
5934 | && REG_P (dest) | |
5935 | && SCALAR_INT_MODE_P (GET_MODE (dest)) | |
5936 | && GET_CODE (sets[i].src_elt->exp) == CONST_INT) | |
5937 | insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest)); | |
5938 | ||
e16ceb8e | 5939 | elt->in_memory = (MEM_P (sets[i].inner_dest) |
b04fab2a | 5940 | && !MEM_READONLY_P (sets[i].inner_dest)); |
26830081 | 5941 | |
e516eaa9 | 5942 | /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no |
5943 | narrower than M2, and both M1 and M2 are the same number of words, | |
5944 | we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so | |
5945 | make that equivalence as well. | |
752df20e | 5946 | |
316f48ea | 5947 | However, BAR may have equivalences for which gen_lowpart |
5948 | will produce a simpler value than gen_lowpart applied to | |
752df20e | 5949 | BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all |
cb10db9d | 5950 | BAR's equivalences. If we don't get a simplified form, make |
752df20e | 5951 | the SUBREG. It will not be used in an equivalence, but will |
5952 | cause two similar assignments to be detected. | |
5953 | ||
5954 | Note the loop below will find SUBREG_REG (DEST) since we have | |
5955 | already entered SRC and DEST of the SET in the table. */ | |
5956 | ||
5957 | if (GET_CODE (dest) == SUBREG | |
e82e6abc | 5958 | && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1) |
5959 | / UNITS_PER_WORD) | |
cb10db9d | 5960 | == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD) |
752df20e | 5961 | && (GET_MODE_SIZE (GET_MODE (dest)) |
5962 | >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))) | |
5963 | && sets[i].src_elt != 0) | |
5964 | { | |
3754d046 | 5965 | machine_mode new_mode = GET_MODE (SUBREG_REG (dest)); |
752df20e | 5966 | struct table_elt *elt, *classp = 0; |
5967 | ||
5968 | for (elt = sets[i].src_elt->first_same_value; elt; | |
5969 | elt = elt->next_same_value) | |
5970 | { | |
5971 | rtx new_src = 0; | |
952bc06d | 5972 | unsigned src_hash; |
752df20e | 5973 | struct table_elt *src_elt; |
cdc84acd | 5974 | int byte = 0; |
752df20e | 5975 | |
5976 | /* Ignore invalid entries. */ | |
8ad4c111 | 5977 | if (!REG_P (elt->exp) |
78d140c9 | 5978 | && ! exp_equiv_p (elt->exp, elt->exp, 1, false)) |
752df20e | 5979 | continue; |
5980 | ||
38b13a9b | 5981 | /* We may have already been playing subreg games. If the |
5982 | mode is already correct for the destination, use it. */ | |
5983 | if (GET_MODE (elt->exp) == new_mode) | |
5984 | new_src = elt->exp; | |
5985 | else | |
5986 | { | |
5987 | /* Calculate big endian correction for the SUBREG_BYTE. | |
5988 | We have already checked that M1 (GET_MODE (dest)) | |
5989 | is not narrower than M2 (new_mode). */ | |
5990 | if (BYTES_BIG_ENDIAN) | |
5991 | byte = (GET_MODE_SIZE (GET_MODE (dest)) | |
5992 | - GET_MODE_SIZE (new_mode)); | |
5993 | ||
5994 | new_src = simplify_gen_subreg (new_mode, elt->exp, | |
5995 | GET_MODE (dest), byte); | |
5996 | } | |
5997 | ||
cdc84acd | 5998 | /* The call to simplify_gen_subreg fails if the value |
5999 | is VOIDmode, yet we can't do any simplification, e.g. | |
6000 | for EXPR_LISTs denoting function call results. | |
6001 | It is invalid to construct a SUBREG with a VOIDmode | |
6002 | SUBREG_REG, hence a zero new_src means we can't do | |
6003 | this substitution. */ | |
6004 | if (! new_src) | |
6005 | continue; | |
752df20e | 6006 | |
6007 | src_hash = HASH (new_src, new_mode); | |
6008 | src_elt = lookup (new_src, src_hash, new_mode); | |
6009 | ||
6010 | /* Put the new source in the hash table is if isn't | |
6011 | already. */ | |
6012 | if (src_elt == 0) | |
6013 | { | |
6014 | if (insert_regs (new_src, classp, 0)) | |
1b033cc3 | 6015 | { |
6016 | rehash_using_reg (new_src); | |
6017 | src_hash = HASH (new_src, new_mode); | |
6018 | } | |
752df20e | 6019 | src_elt = insert (new_src, classp, src_hash, new_mode); |
6020 | src_elt->in_memory = elt->in_memory; | |
20d3ff08 | 6021 | if (GET_CODE (new_src) == ASM_OPERANDS |
6022 | && elt->cost == MAX_COST) | |
6023 | src_elt->cost = MAX_COST; | |
752df20e | 6024 | } |
6025 | else if (classp && classp != src_elt->first_same_value) | |
cb10db9d | 6026 | /* Show that two things that we've seen before are |
752df20e | 6027 | actually the same. */ |
6028 | merge_equiv_classes (src_elt, classp); | |
6029 | ||
6030 | classp = src_elt->first_same_value; | |
7720c877 | 6031 | /* Ignore invalid entries. */ |
6032 | while (classp | |
8ad4c111 | 6033 | && !REG_P (classp->exp) |
78d140c9 | 6034 | && ! exp_equiv_p (classp->exp, classp->exp, 1, false)) |
7720c877 | 6035 | classp = classp->next_same_value; |
752df20e | 6036 | } |
6037 | } | |
6038 | } | |
6039 | ||
01a22203 | 6040 | /* Special handling for (set REG0 REG1) where REG0 is the |
6041 | "cheapest", cheaper than REG1. After cse, REG1 will probably not | |
6042 | be used in the sequel, so (if easily done) change this insn to | |
6043 | (set REG1 REG0) and replace REG1 with REG0 in the previous insn | |
6044 | that computed their value. Then REG1 will become a dead store | |
6045 | and won't cloud the situation for later optimizations. | |
752df20e | 6046 | |
6047 | Do not make this change if REG1 is a hard register, because it will | |
6048 | then be used in the sequel and we may be changing a two-operand insn | |
6049 | into a three-operand insn. | |
6050 | ||
1e5b92fa | 6051 | Also do not do this if we are operating on a copy of INSN. */ |
752df20e | 6052 | |
2aca5650 | 6053 | if (n_sets == 1 && sets[0].rtl) |
6054 | try_back_substitute_reg (sets[0].rtl, insn); | |
752df20e | 6055 | |
be22716f | 6056 | done:; |
752df20e | 6057 | } |
6058 | \f | |
59241190 | 6059 | /* Remove from the hash table all expressions that reference memory. */ |
155b05dc | 6060 | |
752df20e | 6061 | static void |
8ec3a57b | 6062 | invalidate_memory (void) |
752df20e | 6063 | { |
19cb6b50 | 6064 | int i; |
6065 | struct table_elt *p, *next; | |
752df20e | 6066 | |
9c4f3716 | 6067 | for (i = 0; i < HASH_SIZE; i++) |
ea0cb7ae | 6068 | for (p = table[i]; p; p = next) |
6069 | { | |
6070 | next = p->next_same_hash; | |
6071 | if (p->in_memory) | |
6072 | remove_from_table (p, i); | |
6073 | } | |
6074 | } | |
6075 | ||
2aca5650 | 6076 | /* Perform invalidation on the basis of everything about INSN, |
752df20e | 6077 | except for invalidating the actual places that are SET in it. |
6078 | This includes the places CLOBBERed, and anything that might | |
2aca5650 | 6079 | alias with something that is SET or CLOBBERed. */ |
752df20e | 6080 | |
6081 | static void | |
47f1d198 | 6082 | invalidate_from_clobbers (rtx_insn *insn) |
752df20e | 6083 | { |
2aca5650 | 6084 | rtx x = PATTERN (insn); |
6085 | ||
752df20e | 6086 | if (GET_CODE (x) == CLOBBER) |
6087 | { | |
6088 | rtx ref = XEXP (x, 0); | |
ea0cb7ae | 6089 | if (ref) |
6090 | { | |
8ad4c111 | 6091 | if (REG_P (ref) || GET_CODE (ref) == SUBREG |
e16ceb8e | 6092 | || MEM_P (ref)) |
ea0cb7ae | 6093 | invalidate (ref, VOIDmode); |
6094 | else if (GET_CODE (ref) == STRICT_LOW_PART | |
6095 | || GET_CODE (ref) == ZERO_EXTRACT) | |
6096 | invalidate (XEXP (ref, 0), GET_MODE (ref)); | |
6097 | } | |
752df20e | 6098 | } |
6099 | else if (GET_CODE (x) == PARALLEL) | |
6100 | { | |
19cb6b50 | 6101 | int i; |
752df20e | 6102 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) |
6103 | { | |
19cb6b50 | 6104 | rtx y = XVECEXP (x, 0, i); |
752df20e | 6105 | if (GET_CODE (y) == CLOBBER) |
6106 | { | |
6107 | rtx ref = XEXP (y, 0); | |
8ad4c111 | 6108 | if (REG_P (ref) || GET_CODE (ref) == SUBREG |
e16ceb8e | 6109 | || MEM_P (ref)) |
ea0cb7ae | 6110 | invalidate (ref, VOIDmode); |
6111 | else if (GET_CODE (ref) == STRICT_LOW_PART | |
6112 | || GET_CODE (ref) == ZERO_EXTRACT) | |
6113 | invalidate (XEXP (ref, 0), GET_MODE (ref)); | |
752df20e | 6114 | } |
6115 | } | |
6116 | } | |
6117 | } | |
6118 | \f | |
2aca5650 | 6119 | /* Perform invalidation on the basis of everything about INSN. |
6120 | This includes the places CLOBBERed, and anything that might | |
6121 | alias with something that is SET or CLOBBERed. */ | |
6122 | ||
6123 | static void | |
47f1d198 | 6124 | invalidate_from_sets_and_clobbers (rtx_insn *insn) |
2aca5650 | 6125 | { |
6126 | rtx tem; | |
6127 | rtx x = PATTERN (insn); | |
6128 | ||
6129 | if (CALL_P (insn)) | |
6130 | { | |
6131 | for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1)) | |
6132 | if (GET_CODE (XEXP (tem, 0)) == CLOBBER) | |
6133 | invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode); | |
6134 | } | |
6135 | ||
6136 | /* Ensure we invalidate the destination register of a CALL insn. | |
6137 | This is necessary for machines where this register is a fixed_reg, | |
6138 | because no other code would invalidate it. */ | |
6139 | if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL) | |
6140 | invalidate (SET_DEST (x), VOIDmode); | |
6141 | ||
6142 | else if (GET_CODE (x) == PARALLEL) | |
6143 | { | |
6144 | int i; | |
6145 | ||
6146 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
6147 | { | |
6148 | rtx y = XVECEXP (x, 0, i); | |
6149 | if (GET_CODE (y) == CLOBBER) | |
6150 | { | |
6151 | rtx clobbered = XEXP (y, 0); | |
6152 | ||
6153 | if (REG_P (clobbered) | |
6154 | || GET_CODE (clobbered) == SUBREG) | |
6155 | invalidate (clobbered, VOIDmode); | |
6156 | else if (GET_CODE (clobbered) == STRICT_LOW_PART | |
6157 | || GET_CODE (clobbered) == ZERO_EXTRACT) | |
6158 | invalidate (XEXP (clobbered, 0), GET_MODE (clobbered)); | |
6159 | } | |
6160 | else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL) | |
6161 | invalidate (SET_DEST (y), VOIDmode); | |
6162 | } | |
6163 | } | |
6164 | } | |
6165 | \f | |
752df20e | 6166 | /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes |
6167 | and replace any registers in them with either an equivalent constant | |
6168 | or the canonical form of the register. If we are inside an address, | |
6169 | only do this if the address remains valid. | |
6170 | ||
6171 | OBJECT is 0 except when within a MEM in which case it is the MEM. | |
6172 | ||
6173 | Return the replacement for X. */ | |
6174 | ||
6175 | static rtx | |
3072d30e | 6176 | cse_process_notes_1 (rtx x, rtx object, bool *changed) |
752df20e | 6177 | { |
6178 | enum rtx_code code = GET_CODE (x); | |
d2ca078f | 6179 | const char *fmt = GET_RTX_FORMAT (code); |
752df20e | 6180 | int i; |
6181 | ||
6182 | switch (code) | |
6183 | { | |
752df20e | 6184 | case CONST: |
6185 | case SYMBOL_REF: | |
6186 | case LABEL_REF: | |
0349edce | 6187 | CASE_CONST_ANY: |
752df20e | 6188 | case PC: |
6189 | case CC0: | |
6190 | case LO_SUM: | |
6191 | return x; | |
6192 | ||
6193 | case MEM: | |
a344307e | 6194 | validate_change (x, &XEXP (x, 0), |
3072d30e | 6195 | cse_process_notes (XEXP (x, 0), x, changed), 0); |
752df20e | 6196 | return x; |
6197 | ||
6198 | case EXPR_LIST: | |
752df20e | 6199 | if (REG_NOTE_KIND (x) == REG_EQUAL) |
3072d30e | 6200 | XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed); |
9eb946de | 6201 | /* Fall through. */ |
6202 | ||
6203 | case INSN_LIST: | |
6204 | case INT_LIST: | |
752df20e | 6205 | if (XEXP (x, 1)) |
3072d30e | 6206 | XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed); |
752df20e | 6207 | return x; |
6208 | ||
21c77c5d | 6209 | case SIGN_EXTEND: |
6210 | case ZERO_EXTEND: | |
5afa7a07 | 6211 | case SUBREG: |
21c77c5d | 6212 | { |
d328ebdf | 6213 | rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed); |
21c77c5d | 6214 | /* We don't substitute VOIDmode constants into these rtx, |
6215 | since they would impede folding. */ | |
d328ebdf | 6216 | if (GET_MODE (new_rtx) != VOIDmode) |
6217 | validate_change (object, &XEXP (x, 0), new_rtx, 0); | |
21c77c5d | 6218 | return x; |
6219 | } | |
6220 | ||
d733203b | 6221 | case UNSIGNED_FLOAT: |
6222 | { | |
6223 | rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed); | |
6224 | /* We don't substitute negative VOIDmode constants into these rtx, | |
6225 | since they would impede folding. */ | |
6226 | if (GET_MODE (new_rtx) != VOIDmode | |
6227 | || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0) | |
6228 | || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0)) | |
6229 | validate_change (object, &XEXP (x, 0), new_rtx, 0); | |
6230 | return x; | |
6231 | } | |
6232 | ||
752df20e | 6233 | case REG: |
d1264606 | 6234 | i = REG_QTY (REGNO (x)); |
752df20e | 6235 | |
6236 | /* Return a constant or a constant register. */ | |
a7f3b1c7 | 6237 | if (REGNO_QTY_VALID_P (REGNO (x))) |
752df20e | 6238 | { |
a7f3b1c7 | 6239 | struct qty_table_elem *ent = &qty_table[i]; |
6240 | ||
6241 | if (ent->const_rtx != NULL_RTX | |
6242 | && (CONSTANT_P (ent->const_rtx) | |
8ad4c111 | 6243 | || REG_P (ent->const_rtx))) |
a7f3b1c7 | 6244 | { |
d328ebdf | 6245 | rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx); |
6246 | if (new_rtx) | |
6247 | return copy_rtx (new_rtx); | |
a7f3b1c7 | 6248 | } |
752df20e | 6249 | } |
6250 | ||
6251 | /* Otherwise, canonicalize this register. */ | |
47f1d198 | 6252 | return canon_reg (x, NULL); |
cb10db9d | 6253 | |
0dbd1c74 | 6254 | default: |
6255 | break; | |
752df20e | 6256 | } |
6257 | ||
6258 | for (i = 0; i < GET_RTX_LENGTH (code); i++) | |
6259 | if (fmt[i] == 'e') | |
6260 | validate_change (object, &XEXP (x, i), | |
3072d30e | 6261 | cse_process_notes (XEXP (x, i), object, changed), 0); |
752df20e | 6262 | |
6263 | return x; | |
6264 | } | |
3072d30e | 6265 | |
6266 | static rtx | |
6267 | cse_process_notes (rtx x, rtx object, bool *changed) | |
6268 | { | |
d328ebdf | 6269 | rtx new_rtx = cse_process_notes_1 (x, object, changed); |
6270 | if (new_rtx != x) | |
3072d30e | 6271 | *changed = true; |
d328ebdf | 6272 | return new_rtx; |
3072d30e | 6273 | } |
6274 | ||
752df20e | 6275 | \f |
be22716f | 6276 | /* Find a path in the CFG, starting with FIRST_BB to perform CSE on. |
752df20e | 6277 | |
be22716f | 6278 | DATA is a pointer to a struct cse_basic_block_data, that is used to |
6279 | describe the path. | |
6280 | It is filled with a queue of basic blocks, starting with FIRST_BB | |
6281 | and following a trace through the CFG. | |
48e1416a | 6282 | |
be22716f | 6283 | If all paths starting at FIRST_BB have been followed, or no new path |
6284 | starting at FIRST_BB can be constructed, this function returns FALSE. | |
6285 | Otherwise, DATA->path is filled and the function returns TRUE indicating | |
6286 | that a path to follow was found. | |
752df20e | 6287 | |
7920eed5 | 6288 | If FOLLOW_JUMPS is false, the maximum path length is 1 and the only |
be22716f | 6289 | block in the path will be FIRST_BB. */ |
752df20e | 6290 | |
be22716f | 6291 | static bool |
6292 | cse_find_path (basic_block first_bb, struct cse_basic_block_data *data, | |
6293 | int follow_jumps) | |
752df20e | 6294 | { |
be22716f | 6295 | basic_block bb; |
6296 | edge e; | |
6297 | int path_size; | |
48e1416a | 6298 | |
08b7917c | 6299 | bitmap_set_bit (cse_visited_basic_blocks, first_bb->index); |
752df20e | 6300 | |
be22716f | 6301 | /* See if there is a previous path. */ |
6302 | path_size = data->path_size; | |
6303 | ||
6304 | /* There is a previous path. Make sure it started with FIRST_BB. */ | |
6305 | if (path_size) | |
6306 | gcc_assert (data->path[0].bb == first_bb); | |
6307 | ||
6308 | /* There was only one basic block in the last path. Clear the path and | |
6309 | return, so that paths starting at another basic block can be tried. */ | |
6310 | if (path_size == 1) | |
6311 | { | |
6312 | path_size = 0; | |
6313 | goto done; | |
6314 | } | |
6315 | ||
6316 | /* If the path was empty from the beginning, construct a new path. */ | |
6317 | if (path_size == 0) | |
6318 | data->path[path_size++].bb = first_bb; | |
6319 | else | |
752df20e | 6320 | { |
be22716f | 6321 | /* Otherwise, path_size must be equal to or greater than 2, because |
6322 | a previous path exists that is at least two basic blocks long. | |
6323 | ||
6324 | Update the previous branch path, if any. If the last branch was | |
6325 | previously along the branch edge, take the fallthrough edge now. */ | |
6326 | while (path_size >= 2) | |
752df20e | 6327 | { |
be22716f | 6328 | basic_block last_bb_in_path, previous_bb_in_path; |
6329 | edge e; | |
6330 | ||
6331 | --path_size; | |
6332 | last_bb_in_path = data->path[path_size].bb; | |
6333 | previous_bb_in_path = data->path[path_size - 1].bb; | |
6334 | ||
6335 | /* If we previously followed a path along the branch edge, try | |
6336 | the fallthru edge now. */ | |
6337 | if (EDGE_COUNT (previous_bb_in_path->succs) == 2 | |
6338 | && any_condjump_p (BB_END (previous_bb_in_path)) | |
6339 | && (e = find_edge (previous_bb_in_path, last_bb_in_path)) | |
6340 | && e == BRANCH_EDGE (previous_bb_in_path)) | |
6341 | { | |
6342 | bb = FALLTHRU_EDGE (previous_bb_in_path)->dest; | |
34154e27 | 6343 | if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun) |
3752d411 | 6344 | && single_pred_p (bb) |
6345 | /* We used to assert here that we would only see blocks | |
6346 | that we have not visited yet. But we may end up | |
6347 | visiting basic blocks twice if the CFG has changed | |
6348 | in this run of cse_main, because when the CFG changes | |
6349 | the topological sort of the CFG also changes. A basic | |
6350 | blocks that previously had more than two predecessors | |
6351 | may now have a single predecessor, and become part of | |
6352 | a path that starts at another basic block. | |
6353 | ||
6354 | We still want to visit each basic block only once, so | |
6355 | halt the path here if we have already visited BB. */ | |
08b7917c | 6356 | && !bitmap_bit_p (cse_visited_basic_blocks, bb->index)) |
be22716f | 6357 | { |
08b7917c | 6358 | bitmap_set_bit (cse_visited_basic_blocks, bb->index); |
be22716f | 6359 | data->path[path_size++].bb = bb; |
6360 | break; | |
6361 | } | |
6362 | } | |
6363 | ||
6364 | data->path[path_size].bb = NULL; | |
6365 | } | |
6366 | ||
6367 | /* If only one block remains in the path, bail. */ | |
6368 | if (path_size == 1) | |
6369 | { | |
6370 | path_size = 0; | |
6371 | goto done; | |
752df20e | 6372 | } |
752df20e | 6373 | } |
6374 | ||
be22716f | 6375 | /* Extend the path if possible. */ |
6376 | if (follow_jumps) | |
752df20e | 6377 | { |
be22716f | 6378 | bb = data->path[path_size - 1].bb; |
6379 | while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH)) | |
6380 | { | |
6381 | if (single_succ_p (bb)) | |
6382 | e = single_succ_edge (bb); | |
6383 | else if (EDGE_COUNT (bb->succs) == 2 | |
6384 | && any_condjump_p (BB_END (bb))) | |
6385 | { | |
6386 | /* First try to follow the branch. If that doesn't lead | |
6387 | to a useful path, follow the fallthru edge. */ | |
6388 | e = BRANCH_EDGE (bb); | |
6389 | if (!single_pred_p (e->dest)) | |
6390 | e = FALLTHRU_EDGE (bb); | |
6391 | } | |
6392 | else | |
6393 | e = NULL; | |
752df20e | 6394 | |
d1ff492e | 6395 | if (e |
4c43a998 | 6396 | && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label) |
34154e27 | 6397 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
3752d411 | 6398 | && single_pred_p (e->dest) |
6399 | /* Avoid visiting basic blocks twice. The large comment | |
6400 | above explains why this can happen. */ | |
08b7917c | 6401 | && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index)) |
be22716f | 6402 | { |
6403 | basic_block bb2 = e->dest; | |
08b7917c | 6404 | bitmap_set_bit (cse_visited_basic_blocks, bb2->index); |
be22716f | 6405 | data->path[path_size++].bb = bb2; |
6406 | bb = bb2; | |
6407 | } | |
6408 | else | |
6409 | bb = NULL; | |
6410 | } | |
6411 | } | |
6412 | ||
6413 | done: | |
6414 | data->path_size = path_size; | |
6415 | return path_size != 0; | |
6416 | } | |
6417 | \f | |
6418 | /* Dump the path in DATA to file F. NSETS is the number of sets | |
6419 | in the path. */ | |
6420 | ||
6421 | static void | |
6422 | cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f) | |
6423 | { | |
6424 | int path_entry; | |
6425 | ||
6426 | fprintf (f, ";; Following path with %d sets: ", nsets); | |
6427 | for (path_entry = 0; path_entry < data->path_size; path_entry++) | |
6428 | fprintf (f, "%d ", (data->path[path_entry].bb)->index); | |
6429 | fputc ('\n', dump_file); | |
6430 | fflush (f); | |
6431 | } | |
6432 | ||
99013338 | 6433 | \f |
6434 | /* Return true if BB has exception handling successor edges. */ | |
6435 | ||
6436 | static bool | |
6437 | have_eh_succ_edges (basic_block bb) | |
6438 | { | |
6439 | edge e; | |
6440 | edge_iterator ei; | |
6441 | ||
6442 | FOR_EACH_EDGE (e, ei, bb->succs) | |
6443 | if (e->flags & EDGE_EH) | |
6444 | return true; | |
6445 | ||
6446 | return false; | |
6447 | } | |
6448 | ||
be22716f | 6449 | \f |
6450 | /* Scan to the end of the path described by DATA. Return an estimate of | |
3072d30e | 6451 | the total number of SETs of all insns in the path. */ |
be22716f | 6452 | |
6453 | static void | |
6454 | cse_prescan_path (struct cse_basic_block_data *data) | |
6455 | { | |
6456 | int nsets = 0; | |
be22716f | 6457 | int path_size = data->path_size; |
6458 | int path_entry; | |
6459 | ||
6460 | /* Scan to end of each basic block in the path. */ | |
48e1416a | 6461 | for (path_entry = 0; path_entry < path_size; path_entry++) |
be22716f | 6462 | { |
6463 | basic_block bb; | |
47f1d198 | 6464 | rtx_insn *insn; |
dfcbcd81 | 6465 | |
be22716f | 6466 | bb = data->path[path_entry].bb; |
752df20e | 6467 | |
be22716f | 6468 | FOR_BB_INSNS (bb, insn) |
752df20e | 6469 | { |
be22716f | 6470 | if (!INSN_P (insn)) |
6471 | continue; | |
cb10db9d | 6472 | |
be22716f | 6473 | /* A PARALLEL can have lots of SETs in it, |
6474 | especially if it is really an ASM_OPERANDS. */ | |
6475 | if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
6476 | nsets += XVECLEN (PATTERN (insn), 0); | |
6477 | else | |
6478 | nsets += 1; | |
752df20e | 6479 | } |
be22716f | 6480 | } |
6481 | ||
be22716f | 6482 | data->nsets = nsets; |
6483 | } | |
6484 | \f | |
9206d997 | 6485 | /* Return true if the pattern of INSN uses a LABEL_REF for which |
6486 | there isn't a REG_LABEL_OPERAND note. */ | |
6487 | ||
6488 | static bool | |
6489 | check_for_label_ref (rtx_insn *insn) | |
6490 | { | |
6491 | /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND | |
6492 | note for it, we must rerun jump since it needs to place the note. If | |
6493 | this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain, | |
6494 | don't do this since no REG_LABEL_OPERAND will be added. */ | |
6495 | subrtx_iterator::array_type array; | |
6496 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL) | |
6497 | { | |
6498 | const_rtx x = *iter; | |
6499 | if (GET_CODE (x) == LABEL_REF | |
6500 | && !LABEL_REF_NONLOCAL_P (x) | |
6501 | && (!JUMP_P (insn) | |
b49f2e4b | 6502 | || !label_is_jump_target_p (LABEL_REF_LABEL (x), insn)) |
6503 | && LABEL_P (LABEL_REF_LABEL (x)) | |
6504 | && INSN_UID (LABEL_REF_LABEL (x)) != 0 | |
6505 | && !find_reg_note (insn, REG_LABEL_OPERAND, LABEL_REF_LABEL (x))) | |
9206d997 | 6506 | return true; |
6507 | } | |
6508 | return false; | |
6509 | } | |
6510 | ||
be22716f | 6511 | /* Process a single extended basic block described by EBB_DATA. */ |
752df20e | 6512 | |
be22716f | 6513 | static void |
6514 | cse_extended_basic_block (struct cse_basic_block_data *ebb_data) | |
6515 | { | |
6516 | int path_size = ebb_data->path_size; | |
6517 | int path_entry; | |
6518 | int num_insns = 0; | |
6519 | ||
6520 | /* Allocate the space needed by qty_table. */ | |
6521 | qty_table = XNEWVEC (struct qty_table_elem, max_qty); | |
6522 | ||
6523 | new_basic_block (); | |
deb2741b | 6524 | cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb); |
6525 | cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb); | |
be22716f | 6526 | for (path_entry = 0; path_entry < path_size; path_entry++) |
6527 | { | |
6528 | basic_block bb; | |
47f1d198 | 6529 | rtx_insn *insn; |
be22716f | 6530 | |
6531 | bb = ebb_data->path[path_entry].bb; | |
b357aba8 | 6532 | |
6533 | /* Invalidate recorded information for eh regs if there is an EH | |
6534 | edge pointing to that bb. */ | |
6535 | if (bb_has_eh_pred (bb)) | |
6536 | { | |
f1c570a6 | 6537 | df_ref def; |
b357aba8 | 6538 | |
f1c570a6 | 6539 | FOR_EACH_ARTIFICIAL_DEF (def, bb->index) |
6540 | if (DF_REF_FLAGS (def) & DF_REF_AT_TOP) | |
6541 | invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def))); | |
b357aba8 | 6542 | } |
6543 | ||
396a4a1d | 6544 | optimize_this_for_speed_p = optimize_bb_for_speed_p (bb); |
201f6961 | 6545 | FOR_BB_INSNS (bb, insn) |
752df20e | 6546 | { |
be22716f | 6547 | /* If we have processed 1,000 insns, flush the hash table to |
6548 | avoid extreme quadratic behavior. We must not include NOTEs | |
6549 | in the count since there may be more of them when generating | |
6550 | debugging information. If we clear the table at different | |
6551 | times, code generated with -g -O might be different than code | |
6552 | generated with -O but not -g. | |
6553 | ||
6554 | FIXME: This is a real kludge and needs to be done some other | |
6555 | way. */ | |
9845d120 | 6556 | if (NONDEBUG_INSN_P (insn) |
be22716f | 6557 | && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS)) |
6558 | { | |
6559 | flush_hash_table (); | |
6560 | num_insns = 0; | |
6561 | } | |
752df20e | 6562 | |
be22716f | 6563 | if (INSN_P (insn)) |
752df20e | 6564 | { |
be22716f | 6565 | /* Process notes first so we have all notes in canonical forms |
6566 | when looking for duplicate operations. */ | |
6567 | if (REG_NOTES (insn)) | |
3072d30e | 6568 | { |
6569 | bool changed = false; | |
6570 | REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), | |
6571 | NULL_RTX, &changed); | |
6572 | if (changed) | |
6573 | df_notes_rescan (insn); | |
6574 | } | |
be22716f | 6575 | |
1e5b92fa | 6576 | cse_insn (insn); |
be22716f | 6577 | |
be22716f | 6578 | /* If we haven't already found an insn where we added a LABEL_REF, |
6579 | check this one. */ | |
283a6b26 | 6580 | if (INSN_P (insn) && !recorded_label_ref |
9206d997 | 6581 | && check_for_label_ref (insn)) |
283a6b26 | 6582 | recorded_label_ref = true; |
c6ddfc69 | 6583 | |
ff900b8e | 6584 | if (HAVE_cc0 && NONDEBUG_INSN_P (insn)) |
c6ddfc69 | 6585 | { |
5542b661 | 6586 | /* If the previous insn sets CC0 and this insn no |
6587 | longer references CC0, delete the previous insn. | |
6588 | Here we use fact that nothing expects CC0 to be | |
6589 | valid over an insn, which is true until the final | |
6590 | pass. */ | |
47f1d198 | 6591 | rtx_insn *prev_insn; |
6592 | rtx tem; | |
5542b661 | 6593 | |
6594 | prev_insn = prev_nonnote_nondebug_insn (insn); | |
6595 | if (prev_insn && NONJUMP_INSN_P (prev_insn) | |
6596 | && (tem = single_set (prev_insn)) != NULL_RTX | |
6597 | && SET_DEST (tem) == cc0_rtx | |
6598 | && ! reg_mentioned_p (cc0_rtx, PATTERN (insn))) | |
6599 | delete_insn (prev_insn); | |
6600 | ||
6601 | /* If this insn is not the last insn in the basic | |
6602 | block, it will be PREV_INSN(insn) in the next | |
6603 | iteration. If we recorded any CC0-related | |
6604 | information for this insn, remember it. */ | |
6605 | if (insn != BB_END (bb)) | |
6606 | { | |
6607 | prev_insn_cc0 = this_insn_cc0; | |
6608 | prev_insn_cc0_mode = this_insn_cc0_mode; | |
6609 | } | |
c6ddfc69 | 6610 | } |
be22716f | 6611 | } |
6612 | } | |
752df20e | 6613 | |
99013338 | 6614 | /* With non-call exceptions, we are not always able to update |
6615 | the CFG properly inside cse_insn. So clean up possibly | |
6616 | redundant EH edges here. */ | |
cbeb677e | 6617 | if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb)) |
283a6b26 | 6618 | cse_cfg_altered |= purge_dead_edges (bb); |
99013338 | 6619 | |
be22716f | 6620 | /* If we changed a conditional jump, we may have terminated |
6621 | the path we are following. Check that by verifying that | |
6622 | the edge we would take still exists. If the edge does | |
6623 | not exist anymore, purge the remainder of the path. | |
6624 | Note that this will cause us to return to the caller. */ | |
6625 | if (path_entry < path_size - 1) | |
6626 | { | |
6627 | basic_block next_bb = ebb_data->path[path_entry + 1].bb; | |
6628 | if (!find_edge (bb, next_bb)) | |
5b58e627 | 6629 | { |
6630 | do | |
6631 | { | |
6632 | path_size--; | |
6633 | ||
6634 | /* If we truncate the path, we must also reset the | |
6635 | visited bit on the remaining blocks in the path, | |
6636 | or we will never visit them at all. */ | |
08b7917c | 6637 | bitmap_clear_bit (cse_visited_basic_blocks, |
5b58e627 | 6638 | ebb_data->path[path_size].bb->index); |
6639 | ebb_data->path[path_size].bb = NULL; | |
6640 | } | |
6641 | while (path_size - 1 != path_entry); | |
6642 | ebb_data->path_size = path_size; | |
6643 | } | |
752df20e | 6644 | } |
752df20e | 6645 | |
be22716f | 6646 | /* If this is a conditional jump insn, record any known |
6647 | equivalences due to the condition being tested. */ | |
6648 | insn = BB_END (bb); | |
6649 | if (path_entry < path_size - 1 | |
6650 | && JUMP_P (insn) | |
6651 | && single_set (insn) | |
6652 | && any_condjump_p (insn)) | |
6653 | { | |
6654 | basic_block next_bb = ebb_data->path[path_entry + 1].bb; | |
6655 | bool taken = (next_bb == BRANCH_EDGE (bb)->dest); | |
6656 | record_jump_equiv (insn, taken); | |
6657 | } | |
c6ddfc69 | 6658 | |
c6ddfc69 | 6659 | /* Clear the CC0-tracking related insns, they can't provide |
6660 | useful information across basic block boundaries. */ | |
6661 | prev_insn_cc0 = 0; | |
be22716f | 6662 | } |
752df20e | 6663 | |
be22716f | 6664 | gcc_assert (next_qty <= max_qty); |
752df20e | 6665 | |
be22716f | 6666 | free (qty_table); |
752df20e | 6667 | } |
3072d30e | 6668 | |
752df20e | 6669 | \f |
752df20e | 6670 | /* Perform cse on the instructions of a function. |
6671 | F is the first instruction. | |
6672 | NREGS is one plus the highest pseudo-reg number used in the instruction. | |
6673 | ||
283a6b26 | 6674 | Return 2 if jump optimizations should be redone due to simplifications |
6675 | in conditional jump instructions. | |
6676 | Return 1 if the CFG should be cleaned up because it has been modified. | |
6677 | Return 0 otherwise. */ | |
752df20e | 6678 | |
d2bb3f9d | 6679 | static int |
47f1d198 | 6680 | cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs) |
752df20e | 6681 | { |
be22716f | 6682 | struct cse_basic_block_data ebb_data; |
6683 | basic_block bb; | |
fe672ac0 | 6684 | int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
be22716f | 6685 | int i, n_blocks; |
752df20e | 6686 | |
c28211ae | 6687 | /* CSE doesn't use dominane info but can invalidate it in different ways. |
6688 | For simplicity free dominance info here. */ | |
6689 | free_dominance_info (CDI_DOMINATORS); | |
6690 | ||
3072d30e | 6691 | df_set_flags (DF_LR_RUN_DCE); |
264adf90 | 6692 | df_note_add_problem (); |
3072d30e | 6693 | df_analyze (); |
6694 | df_set_flags (DF_DEFER_INSN_RESCAN); | |
6695 | ||
6696 | reg_scan (get_insns (), max_reg_num ()); | |
3bd20490 | 6697 | init_cse_reg_info (nregs); |
6698 | ||
be22716f | 6699 | ebb_data.path = XNEWVEC (struct branch_path, |
6700 | PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH)); | |
38ccff25 | 6701 | |
283a6b26 | 6702 | cse_cfg_altered = false; |
6703 | cse_jumps_altered = false; | |
6704 | recorded_label_ref = false; | |
752df20e | 6705 | constant_pool_entries_cost = 0; |
634d45d7 | 6706 | constant_pool_entries_regcost = 0; |
be22716f | 6707 | ebb_data.path_size = 0; |
6708 | ebb_data.nsets = 0; | |
d263732c | 6709 | rtl_hooks = cse_rtl_hooks; |
752df20e | 6710 | |
6711 | init_recog (); | |
ea0cb7ae | 6712 | init_alias_analysis (); |
752df20e | 6713 | |
4c36ffe6 | 6714 | reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs); |
752df20e | 6715 | |
be22716f | 6716 | /* Set up the table of already visited basic blocks. */ |
fe672ac0 | 6717 | cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
53c5d9d4 | 6718 | bitmap_clear (cse_visited_basic_blocks); |
752df20e | 6719 | |
99013338 | 6720 | /* Loop over basic blocks in reverse completion order (RPO), |
be22716f | 6721 | excluding the ENTRY and EXIT blocks. */ |
5b58e627 | 6722 | n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false); |
be22716f | 6723 | i = 0; |
6724 | while (i < n_blocks) | |
752df20e | 6725 | { |
99013338 | 6726 | /* Find the first block in the RPO queue that we have not yet |
be22716f | 6727 | processed before. */ |
6728 | do | |
0dbd1c74 | 6729 | { |
f5a6b05f | 6730 | bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]); |
0dbd1c74 | 6731 | } |
08b7917c | 6732 | while (bitmap_bit_p (cse_visited_basic_blocks, bb->index) |
be22716f | 6733 | && i < n_blocks); |
752df20e | 6734 | |
be22716f | 6735 | /* Find all paths starting with BB, and process them. */ |
6736 | while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps)) | |
752df20e | 6737 | { |
be22716f | 6738 | /* Pre-scan the path. */ |
6739 | cse_prescan_path (&ebb_data); | |
752df20e | 6740 | |
be22716f | 6741 | /* If this basic block has no sets, skip it. */ |
6742 | if (ebb_data.nsets == 0) | |
6743 | continue; | |
752df20e | 6744 | |
7920eed5 | 6745 | /* Get a reasonable estimate for the maximum number of qty's |
be22716f | 6746 | needed for this path. For this, we take the number of sets |
6747 | and multiply that by MAX_RECOG_OPERANDS. */ | |
6748 | max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS; | |
752df20e | 6749 | |
be22716f | 6750 | /* Dump the path we're about to process. */ |
6751 | if (dump_file) | |
6752 | cse_dump_path (&ebb_data, ebb_data.nsets, dump_file); | |
541a035f | 6753 | |
be22716f | 6754 | cse_extended_basic_block (&ebb_data); |
752df20e | 6755 | } |
752df20e | 6756 | } |
6757 | ||
be22716f | 6758 | /* Clean up. */ |
6759 | end_alias_analysis (); | |
be22716f | 6760 | free (reg_eqv_table); |
6761 | free (ebb_data.path); | |
6762 | sbitmap_free (cse_visited_basic_blocks); | |
5b58e627 | 6763 | free (rc_order); |
be22716f | 6764 | rtl_hooks = general_rtl_hooks; |
ef866782 | 6765 | |
283a6b26 | 6766 | if (cse_jumps_altered || recorded_label_ref) |
6767 | return 2; | |
6768 | else if (cse_cfg_altered) | |
6769 | return 1; | |
6770 | else | |
6771 | return 0; | |
752df20e | 6772 | } |
6773 | \f | |
6774 | /* Count the number of times registers are used (not set) in X. | |
6775 | COUNTS is an array in which we accumulate the count, INCR is how much | |
e6bf10d8 | 6776 | we count each register usage. |
6777 | ||
6778 | Don't count a usage of DEST, which is the SET_DEST of a SET which | |
6779 | contains X in its SET_SRC. This is because such a SET does not | |
6780 | modify the liveness of DEST. | |
46313beb | 6781 | DEST is set to pc_rtx for a trapping insn, or for an insn with side effects. |
6782 | We must then count uses of a SET_DEST regardless, because the insn can't be | |
6783 | deleted here. */ | |
752df20e | 6784 | |
6785 | static void | |
e6bf10d8 | 6786 | count_reg_usage (rtx x, int *counts, rtx dest, int incr) |
752df20e | 6787 | { |
b84155cd | 6788 | enum rtx_code code; |
ce32fe65 | 6789 | rtx note; |
d2ca078f | 6790 | const char *fmt; |
752df20e | 6791 | int i, j; |
6792 | ||
b84155cd | 6793 | if (x == 0) |
6794 | return; | |
6795 | ||
6796 | switch (code = GET_CODE (x)) | |
752df20e | 6797 | { |
6798 | case REG: | |
e6bf10d8 | 6799 | if (x != dest) |
6800 | counts[REGNO (x)] += incr; | |
752df20e | 6801 | return; |
6802 | ||
6803 | case PC: | |
6804 | case CC0: | |
6805 | case CONST: | |
0349edce | 6806 | CASE_CONST_ANY: |
752df20e | 6807 | case SYMBOL_REF: |
6808 | case LABEL_REF: | |
a51d039e | 6809 | return; |
6810 | ||
cb10db9d | 6811 | case CLOBBER: |
a51d039e | 6812 | /* If we are clobbering a MEM, mark any registers inside the address |
6813 | as being used. */ | |
e16ceb8e | 6814 | if (MEM_P (XEXP (x, 0))) |
e6bf10d8 | 6815 | count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr); |
752df20e | 6816 | return; |
6817 | ||
6818 | case SET: | |
6819 | /* Unless we are setting a REG, count everything in SET_DEST. */ | |
8ad4c111 | 6820 | if (!REG_P (SET_DEST (x))) |
e6bf10d8 | 6821 | count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr); |
6822 | count_reg_usage (SET_SRC (x), counts, | |
6823 | dest ? dest : SET_DEST (x), | |
6824 | incr); | |
752df20e | 6825 | return; |
6826 | ||
9845d120 | 6827 | case DEBUG_INSN: |
6828 | return; | |
6829 | ||
b84155cd | 6830 | case CALL_INSN: |
752df20e | 6831 | case INSN: |
6832 | case JUMP_INSN: | |
bc0dfc8d | 6833 | /* We expect dest to be NULL_RTX here. If the insn may throw, |
46313beb | 6834 | or if it cannot be deleted due to side-effects, mark this fact |
6835 | by setting DEST to pc_rtx. */ | |
bc0dfc8d | 6836 | if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x)) |
6837 | || side_effects_p (PATTERN (x))) | |
e6bf10d8 | 6838 | dest = pc_rtx; |
6839 | if (code == CALL_INSN) | |
6840 | count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr); | |
6841 | count_reg_usage (PATTERN (x), counts, dest, incr); | |
752df20e | 6842 | |
6843 | /* Things used in a REG_EQUAL note aren't dead since loop may try to | |
6844 | use them. */ | |
6845 | ||
ce32fe65 | 6846 | note = find_reg_equal_equiv_note (x); |
6847 | if (note) | |
86178c33 | 6848 | { |
6849 | rtx eqv = XEXP (note, 0); | |
6850 | ||
6851 | if (GET_CODE (eqv) == EXPR_LIST) | |
6852 | /* This REG_EQUAL note describes the result of a function call. | |
6853 | Process all the arguments. */ | |
6854 | do | |
6855 | { | |
e6bf10d8 | 6856 | count_reg_usage (XEXP (eqv, 0), counts, dest, incr); |
86178c33 | 6857 | eqv = XEXP (eqv, 1); |
6858 | } | |
6859 | while (eqv && GET_CODE (eqv) == EXPR_LIST); | |
6860 | else | |
e6bf10d8 | 6861 | count_reg_usage (eqv, counts, dest, incr); |
86178c33 | 6862 | } |
752df20e | 6863 | return; |
6864 | ||
d5f9786f | 6865 | case EXPR_LIST: |
6866 | if (REG_NOTE_KIND (x) == REG_EQUAL | |
6867 | || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE) | |
6868 | /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)), | |
6869 | involving registers in the address. */ | |
6870 | || GET_CODE (XEXP (x, 0)) == CLOBBER) | |
e6bf10d8 | 6871 | count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr); |
d5f9786f | 6872 | |
e6bf10d8 | 6873 | count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr); |
d5f9786f | 6874 | return; |
6875 | ||
16d4da86 | 6876 | case ASM_OPERANDS: |
16d4da86 | 6877 | /* Iterate over just the inputs, not the constraints as well. */ |
6878 | for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--) | |
e6bf10d8 | 6879 | count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr); |
16d4da86 | 6880 | return; |
6881 | ||
752df20e | 6882 | case INSN_LIST: |
b3578ae7 | 6883 | case INT_LIST: |
cc636d56 | 6884 | gcc_unreachable (); |
cb10db9d | 6885 | |
0dbd1c74 | 6886 | default: |
6887 | break; | |
752df20e | 6888 | } |
6889 | ||
6890 | fmt = GET_RTX_FORMAT (code); | |
6891 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
6892 | { | |
6893 | if (fmt[i] == 'e') | |
e6bf10d8 | 6894 | count_reg_usage (XEXP (x, i), counts, dest, incr); |
752df20e | 6895 | else if (fmt[i] == 'E') |
6896 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
e6bf10d8 | 6897 | count_reg_usage (XVECEXP (x, i, j), counts, dest, incr); |
752df20e | 6898 | } |
6899 | } | |
6900 | \f | |
a52dfddb | 6901 | /* Return true if X is a dead register. */ |
9845d120 | 6902 | |
a52dfddb | 6903 | static inline int |
a51848dc | 6904 | is_dead_reg (const_rtx x, int *counts) |
9845d120 | 6905 | { |
9845d120 | 6906 | return (REG_P (x) |
6907 | && REGNO (x) >= FIRST_PSEUDO_REGISTER | |
6908 | && counts[REGNO (x)] == 0); | |
6909 | } | |
6910 | ||
6d866f03 | 6911 | /* Return true if set is live. */ |
6912 | static bool | |
47f1d198 | 6913 | set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */ |
8ec3a57b | 6914 | int *counts) |
6d866f03 | 6915 | { |
9ed997be | 6916 | rtx_insn *tem; |
6d866f03 | 6917 | |
6918 | if (set_noop_p (set)) | |
6919 | ; | |
6920 | ||
6d866f03 | 6921 | else if (GET_CODE (SET_DEST (set)) == CC0 |
6922 | && !side_effects_p (SET_SRC (set)) | |
5542b661 | 6923 | && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX |
6d866f03 | 6924 | || !INSN_P (tem) |
6925 | || !reg_referenced_p (cc0_rtx, PATTERN (tem)))) | |
6926 | return false; | |
a52dfddb | 6927 | else if (!is_dead_reg (SET_DEST (set), counts) |
e8825bb0 | 6928 | || side_effects_p (SET_SRC (set))) |
6d866f03 | 6929 | return true; |
6930 | return false; | |
6931 | } | |
6932 | ||
6933 | /* Return true if insn is live. */ | |
6934 | ||
6935 | static bool | |
47f1d198 | 6936 | insn_live_p (rtx_insn *insn, int *counts) |
6d866f03 | 6937 | { |
6938 | int i; | |
bc0dfc8d | 6939 | if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
8ca56a3b | 6940 | return true; |
6941 | else if (GET_CODE (PATTERN (insn)) == SET) | |
6fc669ae | 6942 | return set_live_p (PATTERN (insn), insn, counts); |
6d866f03 | 6943 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) |
6fc669ae | 6944 | { |
6945 | for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) | |
6946 | { | |
6947 | rtx elt = XVECEXP (PATTERN (insn), 0, i); | |
6d866f03 | 6948 | |
6fc669ae | 6949 | if (GET_CODE (elt) == SET) |
6950 | { | |
6951 | if (set_live_p (elt, insn, counts)) | |
6952 | return true; | |
6953 | } | |
6954 | else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE) | |
6955 | return true; | |
6956 | } | |
6957 | return false; | |
6958 | } | |
9845d120 | 6959 | else if (DEBUG_INSN_P (insn)) |
6960 | { | |
47f1d198 | 6961 | rtx_insn *next; |
9845d120 | 6962 | |
6963 | for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next)) | |
6964 | if (NOTE_P (next)) | |
6965 | continue; | |
6966 | else if (!DEBUG_INSN_P (next)) | |
6967 | return true; | |
6968 | else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next)) | |
6969 | return false; | |
6970 | ||
9845d120 | 6971 | return true; |
6972 | } | |
6d866f03 | 6973 | else |
6974 | return true; | |
6975 | } | |
6976 | ||
a52dfddb | 6977 | /* Count the number of stores into pseudo. Callback for note_stores. */ |
6978 | ||
6979 | static void | |
6980 | count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data) | |
6981 | { | |
6982 | int *counts = (int *) data; | |
6983 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) | |
6984 | counts[REGNO (x)]++; | |
6985 | } | |
6986 | ||
a51848dc | 6987 | /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead |
6988 | pseudo doesn't have a replacement. COUNTS[X] is zero if register X | |
6989 | is dead and REPLACEMENTS[X] is null if it has no replacemenet. | |
6990 | Set *SEEN_REPL to true if we see a dead register that does have | |
6991 | a replacement. */ | |
a52dfddb | 6992 | |
a51848dc | 6993 | static bool |
6994 | is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements, | |
6995 | bool *seen_repl) | |
a52dfddb | 6996 | { |
a51848dc | 6997 | subrtx_iterator::array_type array; |
6998 | FOR_EACH_SUBRTX (iter, array, pat, NONCONST) | |
a52dfddb | 6999 | { |
a51848dc | 7000 | const_rtx x = *iter; |
7001 | if (is_dead_reg (x, counts)) | |
7002 | { | |
7003 | if (replacements && replacements[REGNO (x)] != NULL_RTX) | |
7004 | *seen_repl = true; | |
7005 | else | |
7006 | return true; | |
7007 | } | |
a52dfddb | 7008 | } |
a51848dc | 7009 | return false; |
a52dfddb | 7010 | } |
7011 | ||
7012 | /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR. | |
7013 | Callback for simplify_replace_fn_rtx. */ | |
7014 | ||
7015 | static rtx | |
7016 | replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data) | |
7017 | { | |
7018 | rtx *replacements = (rtx *) data; | |
7019 | ||
7020 | if (REG_P (x) | |
7021 | && REGNO (x) >= FIRST_PSEUDO_REGISTER | |
7022 | && replacements[REGNO (x)] != NULL_RTX) | |
7023 | { | |
7024 | if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)])) | |
7025 | return replacements[REGNO (x)]; | |
7026 | return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)], | |
7027 | GET_MODE (replacements[REGNO (x)])); | |
7028 | } | |
7029 | return NULL_RTX; | |
7030 | } | |
7031 | ||
752df20e | 7032 | /* Scan all the insns and delete any that are dead; i.e., they store a register |
7033 | that is never used or they copy a register to itself. | |
7034 | ||
33752533 | 7035 | This is used to remove insns made obviously dead by cse, loop or other |
7036 | optimizations. It improves the heuristics in loop since it won't try to | |
7037 | move dead invariants out of loops or make givs for dead quantities. The | |
7038 | remaining passes of the compilation are also sped up. */ | |
752df20e | 7039 | |
fb20d6fa | 7040 | int |
f2f648a3 | 7041 | delete_trivially_dead_insns (rtx_insn *insns, int nreg) |
752df20e | 7042 | { |
b9cf3f63 | 7043 | int *counts; |
f2f648a3 | 7044 | rtx_insn *insn, *prev; |
a52dfddb | 7045 | rtx *replacements = NULL; |
2aaf7099 | 7046 | int ndead = 0; |
752df20e | 7047 | |
fb20d6fa | 7048 | timevar_push (TV_DELETE_TRIVIALLY_DEAD); |
752df20e | 7049 | /* First count the number of times each register is used. */ |
a52dfddb | 7050 | if (MAY_HAVE_DEBUG_INSNS) |
7051 | { | |
7052 | counts = XCNEWVEC (int, nreg * 3); | |
7053 | for (insn = insns; insn; insn = NEXT_INSN (insn)) | |
7054 | if (DEBUG_INSN_P (insn)) | |
7055 | count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg, | |
7056 | NULL_RTX, 1); | |
7057 | else if (INSN_P (insn)) | |
7058 | { | |
7059 | count_reg_usage (insn, counts, NULL_RTX, 1); | |
7060 | note_stores (PATTERN (insn), count_stores, counts + nreg * 2); | |
7061 | } | |
7062 | /* If there can be debug insns, COUNTS are 3 consecutive arrays. | |
7063 | First one counts how many times each pseudo is used outside | |
7064 | of debug insns, second counts how many times each pseudo is | |
7065 | used in debug insns and third counts how many times a pseudo | |
7066 | is stored. */ | |
7067 | } | |
7068 | else | |
7069 | { | |
7070 | counts = XCNEWVEC (int, nreg); | |
7071 | for (insn = insns; insn; insn = NEXT_INSN (insn)) | |
7072 | if (INSN_P (insn)) | |
7073 | count_reg_usage (insn, counts, NULL_RTX, 1); | |
7074 | /* If no debug insns can be present, COUNTS is just an array | |
7075 | which counts how many times each pseudo is used. */ | |
7076 | } | |
0d1f9fde | 7077 | /* Pseudo PIC register should be considered as used due to possible |
7078 | new usages generated. */ | |
7079 | if (!reload_completed | |
7080 | && pic_offset_table_rtx | |
7081 | && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER) | |
7082 | counts[REGNO (pic_offset_table_rtx)]++; | |
2aaf7099 | 7083 | /* Go from the last insn to the first and delete insns that only set unused |
7084 | registers or copy a register to itself. As we delete an insn, remove | |
7085 | usage counts for registers it uses. | |
af21a202 | 7086 | |
2aaf7099 | 7087 | The first jump optimization pass may leave a real insn as the last |
7088 | insn in the function. We must not skip that insn or we may end | |
a52dfddb | 7089 | up deleting code that is not really dead. |
7090 | ||
7091 | If some otherwise unused register is only used in DEBUG_INSNs, | |
7092 | try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before | |
7093 | the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR | |
7094 | has been created for the unused register, replace it with | |
7095 | the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */ | |
4ac6fa85 | 7096 | for (insn = get_last_insn (); insn; insn = prev) |
2aaf7099 | 7097 | { |
7098 | int live_insn = 0; | |
752df20e | 7099 | |
4ac6fa85 | 7100 | prev = PREV_INSN (insn); |
7101 | if (!INSN_P (insn)) | |
7102 | continue; | |
752df20e | 7103 | |
1e5b92fa | 7104 | live_insn = insn_live_p (insn, counts); |
752df20e | 7105 | |
2aaf7099 | 7106 | /* If this is a dead insn, delete it and show registers in it aren't |
7107 | being used. */ | |
752df20e | 7108 | |
3072d30e | 7109 | if (! live_insn && dbg_cnt (delete_trivial_dead)) |
2aaf7099 | 7110 | { |
a52dfddb | 7111 | if (DEBUG_INSN_P (insn)) |
7112 | count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg, | |
7113 | NULL_RTX, -1); | |
7114 | else | |
7115 | { | |
7116 | rtx set; | |
7117 | if (MAY_HAVE_DEBUG_INSNS | |
7118 | && (set = single_set (insn)) != NULL_RTX | |
7119 | && is_dead_reg (SET_DEST (set), counts) | |
7120 | /* Used at least once in some DEBUG_INSN. */ | |
7121 | && counts[REGNO (SET_DEST (set)) + nreg] > 0 | |
7122 | /* And set exactly once. */ | |
7123 | && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1 | |
7124 | && !side_effects_p (SET_SRC (set)) | |
7125 | && asm_noperands (PATTERN (insn)) < 0) | |
7126 | { | |
e149ca56 | 7127 | rtx dval, bind_var_loc; |
7128 | rtx_insn *bind; | |
a52dfddb | 7129 | |
7130 | /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */ | |
7131 | dval = make_debug_expr_from_rtl (SET_DEST (set)); | |
7132 | ||
7133 | /* Emit a debug bind insn before the insn in which | |
7134 | reg dies. */ | |
e149ca56 | 7135 | bind_var_loc = |
7136 | gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)), | |
7137 | DEBUG_EXPR_TREE_DECL (dval), | |
7138 | SET_SRC (set), | |
7139 | VAR_INIT_STATUS_INITIALIZED); | |
7140 | count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1); | |
7141 | ||
7142 | bind = emit_debug_insn_before (bind_var_loc, insn); | |
a52dfddb | 7143 | df_insn_rescan (bind); |
7144 | ||
7145 | if (replacements == NULL) | |
7146 | replacements = XCNEWVEC (rtx, nreg); | |
7147 | replacements[REGNO (SET_DEST (set))] = dval; | |
7148 | } | |
7149 | ||
7150 | count_reg_usage (insn, counts, NULL_RTX, -1); | |
7151 | ndead++; | |
7152 | } | |
bfa8ea12 | 7153 | cse_cfg_altered |= delete_insn_and_edges (insn); |
2aaf7099 | 7154 | } |
d4c5e26d | 7155 | } |
b9cf3f63 | 7156 | |
a52dfddb | 7157 | if (MAY_HAVE_DEBUG_INSNS) |
7158 | { | |
a52dfddb | 7159 | for (insn = get_last_insn (); insn; insn = PREV_INSN (insn)) |
7160 | if (DEBUG_INSN_P (insn)) | |
7161 | { | |
7162 | /* If this debug insn references a dead register that wasn't replaced | |
7163 | with an DEBUG_EXPR, reset the DEBUG_INSN. */ | |
a51848dc | 7164 | bool seen_repl = false; |
7165 | if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn), | |
7166 | counts, replacements, &seen_repl)) | |
a52dfddb | 7167 | { |
7168 | INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC (); | |
7169 | df_insn_rescan (insn); | |
7170 | } | |
a51848dc | 7171 | else if (seen_repl) |
a52dfddb | 7172 | { |
7173 | INSN_VAR_LOCATION_LOC (insn) | |
7174 | = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn), | |
7175 | NULL_RTX, replace_dead_reg, | |
7176 | replacements); | |
7177 | df_insn_rescan (insn); | |
7178 | } | |
7179 | } | |
dd045aee | 7180 | free (replacements); |
a52dfddb | 7181 | } |
7182 | ||
450d042a | 7183 | if (dump_file && ndead) |
2aaf7099 | 7184 | fprintf (dump_file, "Deleted %i trivially dead insns\n", |
7185 | ndead); | |
b9cf3f63 | 7186 | /* Clean up. */ |
7187 | free (counts); | |
fb20d6fa | 7188 | timevar_pop (TV_DELETE_TRIVIALLY_DEAD); |
7189 | return ndead; | |
752df20e | 7190 | } |
124ac4e4 | 7191 | |
2342ac7b | 7192 | /* If LOC contains references to NEWREG in a different mode, change them |
7193 | to use NEWREG instead. */ | |
124ac4e4 | 7194 | |
2342ac7b | 7195 | static void |
7196 | cse_change_cc_mode (subrtx_ptr_iterator::array_type &array, | |
41805aed | 7197 | rtx *loc, rtx_insn *insn, rtx newreg) |
124ac4e4 | 7198 | { |
2342ac7b | 7199 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) |
124ac4e4 | 7200 | { |
2342ac7b | 7201 | rtx *loc = *iter; |
7202 | rtx x = *loc; | |
7203 | if (x | |
7204 | && REG_P (x) | |
7205 | && REGNO (x) == REGNO (newreg) | |
7206 | && GET_MODE (x) != GET_MODE (newreg)) | |
7207 | { | |
7208 | validate_change (insn, loc, newreg, 1); | |
7209 | iter.skip_subrtxes (); | |
7210 | } | |
124ac4e4 | 7211 | } |
124ac4e4 | 7212 | } |
7213 | ||
b866694e | 7214 | /* Change the mode of any reference to the register REGNO (NEWREG) to |
7215 | GET_MODE (NEWREG) in INSN. */ | |
7216 | ||
7217 | static void | |
47f1d198 | 7218 | cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg) |
b866694e | 7219 | { |
b866694e | 7220 | int success; |
7221 | ||
7222 | if (!INSN_P (insn)) | |
7223 | return; | |
7224 | ||
2342ac7b | 7225 | subrtx_ptr_iterator::array_type array; |
7226 | cse_change_cc_mode (array, &PATTERN (insn), insn, newreg); | |
7227 | cse_change_cc_mode (array, ®_NOTES (insn), insn, newreg); | |
48e1416a | 7228 | |
b866694e | 7229 | /* If the following assertion was triggered, there is most probably |
7230 | something wrong with the cc_modes_compatible back end function. | |
7231 | CC modes only can be considered compatible if the insn - with the mode | |
7232 | replaced by any of the compatible modes - can still be recognized. */ | |
7233 | success = apply_change_group (); | |
7234 | gcc_assert (success); | |
7235 | } | |
7236 | ||
124ac4e4 | 7237 | /* Change the mode of any reference to the register REGNO (NEWREG) to |
7238 | GET_MODE (NEWREG), starting at START. Stop before END. Stop at | |
4362e8e0 | 7239 | any instruction which modifies NEWREG. */ |
124ac4e4 | 7240 | |
7241 | static void | |
47f1d198 | 7242 | cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg) |
124ac4e4 | 7243 | { |
47f1d198 | 7244 | rtx_insn *insn; |
124ac4e4 | 7245 | |
7246 | for (insn = start; insn != end; insn = NEXT_INSN (insn)) | |
7247 | { | |
7248 | if (! INSN_P (insn)) | |
7249 | continue; | |
7250 | ||
4362e8e0 | 7251 | if (reg_set_p (newreg, insn)) |
124ac4e4 | 7252 | return; |
7253 | ||
b866694e | 7254 | cse_change_cc_mode_insn (insn, newreg); |
124ac4e4 | 7255 | } |
7256 | } | |
7257 | ||
7258 | /* BB is a basic block which finishes with CC_REG as a condition code | |
7259 | register which is set to CC_SRC. Look through the successors of BB | |
7260 | to find blocks which have a single predecessor (i.e., this one), | |
7261 | and look through those blocks for an assignment to CC_REG which is | |
7262 | equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are | |
7263 | permitted to change the mode of CC_SRC to a compatible mode. This | |
7264 | returns VOIDmode if no equivalent assignments were found. | |
7265 | Otherwise it returns the mode which CC_SRC should wind up with. | |
650d2134 | 7266 | ORIG_BB should be the same as BB in the outermost cse_cc_succs call, |
7267 | but is passed unmodified down to recursive calls in order to prevent | |
7268 | endless recursion. | |
124ac4e4 | 7269 | |
7270 | The main complexity in this function is handling the mode issues. | |
7271 | We may have more than one duplicate which we can eliminate, and we | |
7272 | try to find a mode which will work for multiple duplicates. */ | |
7273 | ||
3754d046 | 7274 | static machine_mode |
650d2134 | 7275 | cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src, |
7276 | bool can_change_mode) | |
124ac4e4 | 7277 | { |
7278 | bool found_equiv; | |
3754d046 | 7279 | machine_mode mode; |
124ac4e4 | 7280 | unsigned int insn_count; |
7281 | edge e; | |
47f1d198 | 7282 | rtx_insn *insns[2]; |
3754d046 | 7283 | machine_mode modes[2]; |
47f1d198 | 7284 | rtx_insn *last_insns[2]; |
124ac4e4 | 7285 | unsigned int i; |
7286 | rtx newreg; | |
cd665a06 | 7287 | edge_iterator ei; |
124ac4e4 | 7288 | |
7289 | /* We expect to have two successors. Look at both before picking | |
7290 | the final mode for the comparison. If we have more successors | |
7291 | (i.e., some sort of table jump, although that seems unlikely), | |
7292 | then we require all beyond the first two to use the same | |
7293 | mode. */ | |
7294 | ||
7295 | found_equiv = false; | |
7296 | mode = GET_MODE (cc_src); | |
7297 | insn_count = 0; | |
cd665a06 | 7298 | FOR_EACH_EDGE (e, ei, bb->succs) |
124ac4e4 | 7299 | { |
47f1d198 | 7300 | rtx_insn *insn; |
7301 | rtx_insn *end; | |
124ac4e4 | 7302 | |
7303 | if (e->flags & EDGE_COMPLEX) | |
7304 | continue; | |
7305 | ||
cd665a06 | 7306 | if (EDGE_COUNT (e->dest->preds) != 1 |
34154e27 | 7307 | || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
650d2134 | 7308 | /* Avoid endless recursion on unreachable blocks. */ |
7309 | || e->dest == orig_bb) | |
124ac4e4 | 7310 | continue; |
7311 | ||
7312 | end = NEXT_INSN (BB_END (e->dest)); | |
7313 | for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn)) | |
7314 | { | |
7315 | rtx set; | |
7316 | ||
7317 | if (! INSN_P (insn)) | |
7318 | continue; | |
7319 | ||
7320 | /* If CC_SRC is modified, we have to stop looking for | |
7321 | something which uses it. */ | |
7322 | if (modified_in_p (cc_src, insn)) | |
7323 | break; | |
7324 | ||
7325 | /* Check whether INSN sets CC_REG to CC_SRC. */ | |
7326 | set = single_set (insn); | |
7327 | if (set | |
8ad4c111 | 7328 | && REG_P (SET_DEST (set)) |
124ac4e4 | 7329 | && REGNO (SET_DEST (set)) == REGNO (cc_reg)) |
7330 | { | |
7331 | bool found; | |
3754d046 | 7332 | machine_mode set_mode; |
7333 | machine_mode comp_mode; | |
124ac4e4 | 7334 | |
7335 | found = false; | |
7336 | set_mode = GET_MODE (SET_SRC (set)); | |
7337 | comp_mode = set_mode; | |
7338 | if (rtx_equal_p (cc_src, SET_SRC (set))) | |
7339 | found = true; | |
7340 | else if (GET_CODE (cc_src) == COMPARE | |
7341 | && GET_CODE (SET_SRC (set)) == COMPARE | |
960670fc | 7342 | && mode != set_mode |
124ac4e4 | 7343 | && rtx_equal_p (XEXP (cc_src, 0), |
7344 | XEXP (SET_SRC (set), 0)) | |
7345 | && rtx_equal_p (XEXP (cc_src, 1), | |
7346 | XEXP (SET_SRC (set), 1))) | |
48e1416a | 7347 | |
124ac4e4 | 7348 | { |
883b2e73 | 7349 | comp_mode = targetm.cc_modes_compatible (mode, set_mode); |
124ac4e4 | 7350 | if (comp_mode != VOIDmode |
7351 | && (can_change_mode || comp_mode == mode)) | |
7352 | found = true; | |
7353 | } | |
7354 | ||
7355 | if (found) | |
7356 | { | |
7357 | found_equiv = true; | |
960670fc | 7358 | if (insn_count < ARRAY_SIZE (insns)) |
124ac4e4 | 7359 | { |
7360 | insns[insn_count] = insn; | |
7361 | modes[insn_count] = set_mode; | |
7362 | last_insns[insn_count] = end; | |
7363 | ++insn_count; | |
7364 | ||
960670fc | 7365 | if (mode != comp_mode) |
7366 | { | |
cc636d56 | 7367 | gcc_assert (can_change_mode); |
960670fc | 7368 | mode = comp_mode; |
b866694e | 7369 | |
7370 | /* The modified insn will be re-recognized later. */ | |
960670fc | 7371 | PUT_MODE (cc_src, mode); |
7372 | } | |
124ac4e4 | 7373 | } |
7374 | else | |
7375 | { | |
7376 | if (set_mode != mode) | |
960670fc | 7377 | { |
7378 | /* We found a matching expression in the | |
7379 | wrong mode, but we don't have room to | |
7380 | store it in the array. Punt. This case | |
7381 | should be rare. */ | |
7382 | break; | |
7383 | } | |
124ac4e4 | 7384 | /* INSN sets CC_REG to a value equal to CC_SRC |
7385 | with the right mode. We can simply delete | |
7386 | it. */ | |
7387 | delete_insn (insn); | |
7388 | } | |
7389 | ||
7390 | /* We found an instruction to delete. Keep looking, | |
7391 | in the hopes of finding a three-way jump. */ | |
7392 | continue; | |
7393 | } | |
7394 | ||
7395 | /* We found an instruction which sets the condition | |
7396 | code, so don't look any farther. */ | |
7397 | break; | |
7398 | } | |
7399 | ||
7400 | /* If INSN sets CC_REG in some other way, don't look any | |
7401 | farther. */ | |
7402 | if (reg_set_p (cc_reg, insn)) | |
7403 | break; | |
7404 | } | |
7405 | ||
7406 | /* If we fell off the bottom of the block, we can keep looking | |
7407 | through successors. We pass CAN_CHANGE_MODE as false because | |
7408 | we aren't prepared to handle compatibility between the | |
7409 | further blocks and this block. */ | |
7410 | if (insn == end) | |
7411 | { | |
3754d046 | 7412 | machine_mode submode; |
960670fc | 7413 | |
650d2134 | 7414 | submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false); |
960670fc | 7415 | if (submode != VOIDmode) |
7416 | { | |
cc636d56 | 7417 | gcc_assert (submode == mode); |
960670fc | 7418 | found_equiv = true; |
7419 | can_change_mode = false; | |
7420 | } | |
124ac4e4 | 7421 | } |
7422 | } | |
7423 | ||
7424 | if (! found_equiv) | |
7425 | return VOIDmode; | |
7426 | ||
7427 | /* Now INSN_COUNT is the number of instructions we found which set | |
7428 | CC_REG to a value equivalent to CC_SRC. The instructions are in | |
7429 | INSNS. The modes used by those instructions are in MODES. */ | |
7430 | ||
7431 | newreg = NULL_RTX; | |
7432 | for (i = 0; i < insn_count; ++i) | |
7433 | { | |
7434 | if (modes[i] != mode) | |
7435 | { | |
7436 | /* We need to change the mode of CC_REG in INSNS[i] and | |
7437 | subsequent instructions. */ | |
7438 | if (! newreg) | |
7439 | { | |
7440 | if (GET_MODE (cc_reg) == mode) | |
7441 | newreg = cc_reg; | |
7442 | else | |
7443 | newreg = gen_rtx_REG (mode, REGNO (cc_reg)); | |
7444 | } | |
7445 | cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i], | |
7446 | newreg); | |
7447 | } | |
7448 | ||
bfa8ea12 | 7449 | cse_cfg_altered |= delete_insn_and_edges (insns[i]); |
124ac4e4 | 7450 | } |
7451 | ||
7452 | return mode; | |
7453 | } | |
7454 | ||
7455 | /* If we have a fixed condition code register (or two), walk through | |
7456 | the instructions and try to eliminate duplicate assignments. */ | |
7457 | ||
66c2c707 | 7458 | static void |
124ac4e4 | 7459 | cse_condition_code_reg (void) |
7460 | { | |
7461 | unsigned int cc_regno_1; | |
7462 | unsigned int cc_regno_2; | |
7463 | rtx cc_reg_1; | |
7464 | rtx cc_reg_2; | |
7465 | basic_block bb; | |
7466 | ||
883b2e73 | 7467 | if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2)) |
124ac4e4 | 7468 | return; |
7469 | ||
7470 | cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1); | |
7471 | if (cc_regno_2 != INVALID_REGNUM) | |
7472 | cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2); | |
7473 | else | |
7474 | cc_reg_2 = NULL_RTX; | |
7475 | ||
fc00614f | 7476 | FOR_EACH_BB_FN (bb, cfun) |
124ac4e4 | 7477 | { |
47f1d198 | 7478 | rtx_insn *last_insn; |
124ac4e4 | 7479 | rtx cc_reg; |
47f1d198 | 7480 | rtx_insn *insn; |
7481 | rtx_insn *cc_src_insn; | |
124ac4e4 | 7482 | rtx cc_src; |
3754d046 | 7483 | machine_mode mode; |
7484 | machine_mode orig_mode; | |
124ac4e4 | 7485 | |
7486 | /* Look for blocks which end with a conditional jump based on a | |
7487 | condition code register. Then look for the instruction which | |
7488 | sets the condition code register. Then look through the | |
7489 | successor blocks for instructions which set the condition | |
7490 | code register to the same value. There are other possible | |
7491 | uses of the condition code register, but these are by far the | |
7492 | most common and the ones which we are most likely to be able | |
7493 | to optimize. */ | |
7494 | ||
7495 | last_insn = BB_END (bb); | |
6d7dc5b9 | 7496 | if (!JUMP_P (last_insn)) |
124ac4e4 | 7497 | continue; |
7498 | ||
7499 | if (reg_referenced_p (cc_reg_1, PATTERN (last_insn))) | |
7500 | cc_reg = cc_reg_1; | |
7501 | else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn))) | |
7502 | cc_reg = cc_reg_2; | |
7503 | else | |
7504 | continue; | |
7505 | ||
47f1d198 | 7506 | cc_src_insn = NULL; |
124ac4e4 | 7507 | cc_src = NULL_RTX; |
7508 | for (insn = PREV_INSN (last_insn); | |
7509 | insn && insn != PREV_INSN (BB_HEAD (bb)); | |
7510 | insn = PREV_INSN (insn)) | |
7511 | { | |
7512 | rtx set; | |
7513 | ||
7514 | if (! INSN_P (insn)) | |
7515 | continue; | |
7516 | set = single_set (insn); | |
7517 | if (set | |
8ad4c111 | 7518 | && REG_P (SET_DEST (set)) |
124ac4e4 | 7519 | && REGNO (SET_DEST (set)) == REGNO (cc_reg)) |
7520 | { | |
7521 | cc_src_insn = insn; | |
7522 | cc_src = SET_SRC (set); | |
7523 | break; | |
7524 | } | |
7525 | else if (reg_set_p (cc_reg, insn)) | |
7526 | break; | |
7527 | } | |
7528 | ||
7529 | if (! cc_src_insn) | |
7530 | continue; | |
7531 | ||
7532 | if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn))) | |
7533 | continue; | |
7534 | ||
7535 | /* Now CC_REG is a condition code register used for a | |
7536 | conditional jump at the end of the block, and CC_SRC, in | |
7537 | CC_SRC_INSN, is the value to which that condition code | |
7538 | register is set, and CC_SRC is still meaningful at the end of | |
7539 | the basic block. */ | |
7540 | ||
960670fc | 7541 | orig_mode = GET_MODE (cc_src); |
650d2134 | 7542 | mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true); |
960670fc | 7543 | if (mode != VOIDmode) |
124ac4e4 | 7544 | { |
cc636d56 | 7545 | gcc_assert (mode == GET_MODE (cc_src)); |
960670fc | 7546 | if (mode != orig_mode) |
4362e8e0 | 7547 | { |
7548 | rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg)); | |
7549 | ||
b866694e | 7550 | cse_change_cc_mode_insn (cc_src_insn, newreg); |
4362e8e0 | 7551 | |
7552 | /* Do the same in the following insns that use the | |
7553 | current value of CC_REG within BB. */ | |
7554 | cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn), | |
7555 | NEXT_INSN (last_insn), | |
7556 | newreg); | |
7557 | } | |
124ac4e4 | 7558 | } |
7559 | } | |
7560 | } | |
77fce4cd | 7561 | \f |
7562 | ||
7563 | /* Perform common subexpression elimination. Nonzero value from | |
7564 | `cse_main' means that jumps were simplified and some code may now | |
7565 | be unreachable, so do jump optimization again. */ | |
2a1990e9 | 7566 | static unsigned int |
77fce4cd | 7567 | rest_of_handle_cse (void) |
7568 | { | |
7569 | int tem; | |
3072d30e | 7570 | |
77fce4cd | 7571 | if (dump_file) |
562d71e8 | 7572 | dump_flow_info (dump_file, dump_flags); |
77fce4cd | 7573 | |
3f5be5f4 | 7574 | tem = cse_main (get_insns (), max_reg_num ()); |
77fce4cd | 7575 | |
7576 | /* If we are not running more CSE passes, then we are no longer | |
7577 | expecting CSE to be run. But always rerun it in a cheap mode. */ | |
7578 | cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse; | |
7579 | ||
283a6b26 | 7580 | if (tem == 2) |
7581 | { | |
7582 | timevar_push (TV_JUMP); | |
7583 | rebuild_jump_labels (get_insns ()); | |
03a400fb | 7584 | cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED); |
283a6b26 | 7585 | timevar_pop (TV_JUMP); |
7586 | } | |
7587 | else if (tem == 1 || optimize > 1) | |
03a400fb | 7588 | cse_cfg_altered |= cleanup_cfg (0); |
be22716f | 7589 | |
2a1990e9 | 7590 | return 0; |
77fce4cd | 7591 | } |
7592 | ||
cbe8bda8 | 7593 | namespace { |
7594 | ||
7595 | const pass_data pass_data_cse = | |
77fce4cd | 7596 | { |
cbe8bda8 | 7597 | RTL_PASS, /* type */ |
7598 | "cse1", /* name */ | |
7599 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 7600 | TV_CSE, /* tv_id */ |
7601 | 0, /* properties_required */ | |
7602 | 0, /* properties_provided */ | |
7603 | 0, /* properties_destroyed */ | |
7604 | 0, /* todo_flags_start */ | |
8b88439e | 7605 | TODO_df_finish, /* todo_flags_finish */ |
77fce4cd | 7606 | }; |
7607 | ||
cbe8bda8 | 7608 | class pass_cse : public rtl_opt_pass |
7609 | { | |
7610 | public: | |
9af5ce0c | 7611 | pass_cse (gcc::context *ctxt) |
7612 | : rtl_opt_pass (pass_data_cse, ctxt) | |
cbe8bda8 | 7613 | {} |
7614 | ||
7615 | /* opt_pass methods: */ | |
31315c24 | 7616 | virtual bool gate (function *) { return optimize > 0; } |
65b0537f | 7617 | virtual unsigned int execute (function *) { return rest_of_handle_cse (); } |
cbe8bda8 | 7618 | |
7619 | }; // class pass_cse | |
7620 | ||
7621 | } // anon namespace | |
7622 | ||
7623 | rtl_opt_pass * | |
7624 | make_pass_cse (gcc::context *ctxt) | |
7625 | { | |
7626 | return new pass_cse (ctxt); | |
7627 | } | |
7628 | ||
77fce4cd | 7629 | |
77fce4cd | 7630 | /* Run second CSE pass after loop optimizations. */ |
2a1990e9 | 7631 | static unsigned int |
77fce4cd | 7632 | rest_of_handle_cse2 (void) |
7633 | { | |
7634 | int tem; | |
7635 | ||
7636 | if (dump_file) | |
562d71e8 | 7637 | dump_flow_info (dump_file, dump_flags); |
77fce4cd | 7638 | |
3f5be5f4 | 7639 | tem = cse_main (get_insns (), max_reg_num ()); |
77fce4cd | 7640 | |
7641 | /* Run a pass to eliminate duplicated assignments to condition code | |
7642 | registers. We have to run this after bypass_jumps, because it | |
7643 | makes it harder for that pass to determine whether a jump can be | |
7644 | bypassed safely. */ | |
7645 | cse_condition_code_reg (); | |
7646 | ||
77fce4cd | 7647 | delete_trivially_dead_insns (get_insns (), max_reg_num ()); |
7648 | ||
283a6b26 | 7649 | if (tem == 2) |
77fce4cd | 7650 | { |
7651 | timevar_push (TV_JUMP); | |
7652 | rebuild_jump_labels (get_insns ()); | |
03a400fb | 7653 | cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED); |
77fce4cd | 7654 | timevar_pop (TV_JUMP); |
7655 | } | |
283a6b26 | 7656 | else if (tem == 1) |
03a400fb | 7657 | cse_cfg_altered |= cleanup_cfg (0); |
283a6b26 | 7658 | |
77fce4cd | 7659 | cse_not_expected = 1; |
2a1990e9 | 7660 | return 0; |
77fce4cd | 7661 | } |
7662 | ||
7663 | ||
cbe8bda8 | 7664 | namespace { |
7665 | ||
7666 | const pass_data pass_data_cse2 = | |
77fce4cd | 7667 | { |
cbe8bda8 | 7668 | RTL_PASS, /* type */ |
7669 | "cse2", /* name */ | |
7670 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 7671 | TV_CSE2, /* tv_id */ |
7672 | 0, /* properties_required */ | |
7673 | 0, /* properties_provided */ | |
7674 | 0, /* properties_destroyed */ | |
7675 | 0, /* todo_flags_start */ | |
8b88439e | 7676 | TODO_df_finish, /* todo_flags_finish */ |
77fce4cd | 7677 | }; |
d743aba2 | 7678 | |
cbe8bda8 | 7679 | class pass_cse2 : public rtl_opt_pass |
7680 | { | |
7681 | public: | |
9af5ce0c | 7682 | pass_cse2 (gcc::context *ctxt) |
7683 | : rtl_opt_pass (pass_data_cse2, ctxt) | |
cbe8bda8 | 7684 | {} |
7685 | ||
7686 | /* opt_pass methods: */ | |
31315c24 | 7687 | virtual bool gate (function *) |
7688 | { | |
7689 | return optimize > 0 && flag_rerun_cse_after_loop; | |
7690 | } | |
7691 | ||
65b0537f | 7692 | virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); } |
cbe8bda8 | 7693 | |
7694 | }; // class pass_cse2 | |
7695 | ||
7696 | } // anon namespace | |
7697 | ||
7698 | rtl_opt_pass * | |
7699 | make_pass_cse2 (gcc::context *ctxt) | |
7700 | { | |
7701 | return new pass_cse2 (ctxt); | |
7702 | } | |
7703 | ||
d743aba2 | 7704 | /* Run second CSE pass after loop optimizations. */ |
7705 | static unsigned int | |
7706 | rest_of_handle_cse_after_global_opts (void) | |
7707 | { | |
7708 | int save_cfj; | |
7709 | int tem; | |
7710 | ||
7711 | /* We only want to do local CSE, so don't follow jumps. */ | |
7712 | save_cfj = flag_cse_follow_jumps; | |
7713 | flag_cse_follow_jumps = 0; | |
7714 | ||
7715 | rebuild_jump_labels (get_insns ()); | |
7716 | tem = cse_main (get_insns (), max_reg_num ()); | |
bfa8ea12 | 7717 | cse_cfg_altered |= purge_all_dead_edges (); |
d743aba2 | 7718 | delete_trivially_dead_insns (get_insns (), max_reg_num ()); |
7719 | ||
7720 | cse_not_expected = !flag_rerun_cse_after_loop; | |
7721 | ||
7722 | /* If cse altered any jumps, rerun jump opts to clean things up. */ | |
7723 | if (tem == 2) | |
7724 | { | |
7725 | timevar_push (TV_JUMP); | |
7726 | rebuild_jump_labels (get_insns ()); | |
03a400fb | 7727 | cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED); |
d743aba2 | 7728 | timevar_pop (TV_JUMP); |
7729 | } | |
7730 | else if (tem == 1) | |
03a400fb | 7731 | cse_cfg_altered |= cleanup_cfg (0); |
d743aba2 | 7732 | |
7733 | flag_cse_follow_jumps = save_cfj; | |
7734 | return 0; | |
7735 | } | |
7736 | ||
cbe8bda8 | 7737 | namespace { |
7738 | ||
7739 | const pass_data pass_data_cse_after_global_opts = | |
d743aba2 | 7740 | { |
cbe8bda8 | 7741 | RTL_PASS, /* type */ |
7742 | "cse_local", /* name */ | |
7743 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 7744 | TV_CSE, /* tv_id */ |
7745 | 0, /* properties_required */ | |
7746 | 0, /* properties_provided */ | |
7747 | 0, /* properties_destroyed */ | |
7748 | 0, /* todo_flags_start */ | |
8b88439e | 7749 | TODO_df_finish, /* todo_flags_finish */ |
d743aba2 | 7750 | }; |
cbe8bda8 | 7751 | |
7752 | class pass_cse_after_global_opts : public rtl_opt_pass | |
7753 | { | |
7754 | public: | |
9af5ce0c | 7755 | pass_cse_after_global_opts (gcc::context *ctxt) |
7756 | : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt) | |
cbe8bda8 | 7757 | {} |
7758 | ||
7759 | /* opt_pass methods: */ | |
31315c24 | 7760 | virtual bool gate (function *) |
7761 | { | |
7762 | return optimize > 0 && flag_rerun_cse_after_global_opts; | |
7763 | } | |
7764 | ||
65b0537f | 7765 | virtual unsigned int execute (function *) |
7766 | { | |
7767 | return rest_of_handle_cse_after_global_opts (); | |
7768 | } | |
cbe8bda8 | 7769 | |
7770 | }; // class pass_cse_after_global_opts | |
7771 | ||
7772 | } // anon namespace | |
7773 | ||
7774 | rtl_opt_pass * | |
7775 | make_pass_cse_after_global_opts (gcc::context *ctxt) | |
7776 | { | |
7777 | return new pass_cse_after_global_opts (ctxt); | |
7778 | } |