]>
Commit | Line | Data |
---|---|---|
dffec8eb | 1 | /* GIMPLE store merging and byte swapping passes. |
99dee823 | 2 | Copyright (C) 2009-2021 Free Software Foundation, Inc. |
f663d9ad KT |
3 | Contributed by ARM Ltd. |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 3, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but | |
13 | WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 | General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
c94c3532 EB |
21 | /* The purpose of the store merging pass is to combine multiple memory stores |
22 | of constant values, values loaded from memory, bitwise operations on those, | |
23 | or bit-field values, to consecutive locations, into fewer wider stores. | |
24 | ||
f663d9ad KT |
25 | For example, if we have a sequence peforming four byte stores to |
26 | consecutive memory locations: | |
27 | [p ] := imm1; | |
28 | [p + 1B] := imm2; | |
29 | [p + 2B] := imm3; | |
30 | [p + 3B] := imm4; | |
31 | we can transform this into a single 4-byte store if the target supports it: | |
c94c3532 | 32 | [p] := imm1:imm2:imm3:imm4 concatenated according to endianness. |
f663d9ad | 33 | |
245f6de1 JJ |
34 | Or: |
35 | [p ] := [q ]; | |
36 | [p + 1B] := [q + 1B]; | |
37 | [p + 2B] := [q + 2B]; | |
38 | [p + 3B] := [q + 3B]; | |
39 | if there is no overlap can be transformed into a single 4-byte | |
40 | load followed by single 4-byte store. | |
41 | ||
42 | Or: | |
43 | [p ] := [q ] ^ imm1; | |
44 | [p + 1B] := [q + 1B] ^ imm2; | |
45 | [p + 2B] := [q + 2B] ^ imm3; | |
46 | [p + 3B] := [q + 3B] ^ imm4; | |
47 | if there is no overlap can be transformed into a single 4-byte | |
48 | load, xored with imm1:imm2:imm3:imm4 and stored using a single 4-byte store. | |
49 | ||
c94c3532 EB |
50 | Or: |
51 | [p:1 ] := imm; | |
52 | [p:31] := val & 0x7FFFFFFF; | |
53 | we can transform this into a single 4-byte store if the target supports it: | |
54 | [p] := imm:(val & 0x7FFFFFFF) concatenated according to endianness. | |
55 | ||
f663d9ad KT |
56 | The algorithm is applied to each basic block in three phases: |
57 | ||
c94c3532 EB |
58 | 1) Scan through the basic block and record assignments to destinations |
59 | that can be expressed as a store to memory of a certain size at a certain | |
60 | bit offset from base expressions we can handle. For bit-fields we also | |
61 | record the surrounding bit region, i.e. bits that could be stored in | |
245f6de1 JJ |
62 | a read-modify-write operation when storing the bit-field. Record store |
63 | chains to different bases in a hash_map (m_stores) and make sure to | |
700d4cb0 | 64 | terminate such chains when appropriate (for example when the stored |
245f6de1 | 65 | values get used subsequently). |
f663d9ad KT |
66 | These stores can be a result of structure element initializers, array stores |
67 | etc. A store_immediate_info object is recorded for every such store. | |
68 | Record as many such assignments to a single base as possible until a | |
69 | statement that interferes with the store sequence is encountered. | |
c94c3532 EB |
70 | Each store has up to 2 operands, which can be a either constant, a memory |
71 | load or an SSA name, from which the value to be stored can be computed. | |
245f6de1 JJ |
72 | At most one of the operands can be a constant. The operands are recorded |
73 | in store_operand_info struct. | |
f663d9ad | 74 | |
c94c3532 | 75 | 2) Analyze the chains of stores recorded in phase 1) (i.e. the vector of |
f663d9ad | 76 | store_immediate_info objects) and coalesce contiguous stores into |
c94c3532 | 77 | merged_store_group objects. For bit-field stores, we don't need to |
245f6de1 JJ |
78 | require the stores to be contiguous, just their surrounding bit regions |
79 | have to be contiguous. If the expression being stored is different | |
80 | between adjacent stores, such as one store storing a constant and | |
81 | following storing a value loaded from memory, or if the loaded memory | |
82 | objects are not adjacent, a new merged_store_group is created as well. | |
f663d9ad KT |
83 | |
84 | For example, given the stores: | |
85 | [p ] := 0; | |
86 | [p + 1B] := 1; | |
87 | [p + 3B] := 0; | |
88 | [p + 4B] := 1; | |
89 | [p + 5B] := 0; | |
90 | [p + 6B] := 0; | |
91 | This phase would produce two merged_store_group objects, one recording the | |
92 | two bytes stored in the memory region [p : p + 1] and another | |
93 | recording the four bytes stored in the memory region [p + 3 : p + 6]. | |
94 | ||
95 | 3) The merged_store_group objects produced in phase 2) are processed | |
96 | to generate the sequence of wider stores that set the contiguous memory | |
97 | regions to the sequence of bytes that correspond to it. This may emit | |
98 | multiple stores per store group to handle contiguous stores that are not | |
99 | of a size that is a power of 2. For example it can try to emit a 40-bit | |
100 | store as a 32-bit store followed by an 8-bit store. | |
c94c3532 EB |
101 | We try to emit as wide stores as we can while respecting STRICT_ALIGNMENT |
102 | or TARGET_SLOW_UNALIGNED_ACCESS settings. | |
f663d9ad KT |
103 | |
104 | Note on endianness and example: | |
105 | Consider 2 contiguous 16-bit stores followed by 2 contiguous 8-bit stores: | |
106 | [p ] := 0x1234; | |
107 | [p + 2B] := 0x5678; | |
108 | [p + 4B] := 0xab; | |
109 | [p + 5B] := 0xcd; | |
110 | ||
111 | The memory layout for little-endian (LE) and big-endian (BE) must be: | |
112 | p |LE|BE| | |
113 | --------- | |
114 | 0 |34|12| | |
115 | 1 |12|34| | |
116 | 2 |78|56| | |
117 | 3 |56|78| | |
118 | 4 |ab|ab| | |
119 | 5 |cd|cd| | |
120 | ||
121 | To merge these into a single 48-bit merged value 'val' in phase 2) | |
122 | on little-endian we insert stores to higher (consecutive) bitpositions | |
123 | into the most significant bits of the merged value. | |
124 | The final merged value would be: 0xcdab56781234 | |
125 | ||
126 | For big-endian we insert stores to higher bitpositions into the least | |
127 | significant bits of the merged value. | |
128 | The final merged value would be: 0x12345678abcd | |
129 | ||
130 | Then, in phase 3), we want to emit this 48-bit value as a 32-bit store | |
131 | followed by a 16-bit store. Again, we must consider endianness when | |
132 | breaking down the 48-bit value 'val' computed above. | |
133 | For little endian we emit: | |
134 | [p] (32-bit) := 0x56781234; // val & 0x0000ffffffff; | |
135 | [p + 4B] (16-bit) := 0xcdab; // (val & 0xffff00000000) >> 32; | |
136 | ||
137 | Whereas for big-endian we emit: | |
138 | [p] (32-bit) := 0x12345678; // (val & 0xffffffff0000) >> 16; | |
139 | [p + 4B] (16-bit) := 0xabcd; // val & 0x00000000ffff; */ | |
140 | ||
141 | #include "config.h" | |
142 | #include "system.h" | |
143 | #include "coretypes.h" | |
144 | #include "backend.h" | |
145 | #include "tree.h" | |
146 | #include "gimple.h" | |
147 | #include "builtins.h" | |
148 | #include "fold-const.h" | |
149 | #include "tree-pass.h" | |
150 | #include "ssa.h" | |
151 | #include "gimple-pretty-print.h" | |
152 | #include "alias.h" | |
153 | #include "fold-const.h" | |
f663d9ad KT |
154 | #include "print-tree.h" |
155 | #include "tree-hash-traits.h" | |
156 | #include "gimple-iterator.h" | |
157 | #include "gimplify.h" | |
c94c3532 | 158 | #include "gimple-fold.h" |
f663d9ad KT |
159 | #include "stor-layout.h" |
160 | #include "timevar.h" | |
629387a6 EB |
161 | #include "cfganal.h" |
162 | #include "cfgcleanup.h" | |
f663d9ad | 163 | #include "tree-cfg.h" |
629387a6 | 164 | #include "except.h" |
f663d9ad KT |
165 | #include "tree-eh.h" |
166 | #include "target.h" | |
aa55dc0c | 167 | #include "gimplify-me.h" |
a62b3dc5 JJ |
168 | #include "rtl.h" |
169 | #include "expr.h" /* For get_bit_range. */ | |
dffec8eb | 170 | #include "optabs-tree.h" |
a95b474a | 171 | #include "dbgcnt.h" |
c22d8787 | 172 | #include "selftest.h" |
f663d9ad KT |
173 | |
174 | /* The maximum size (in bits) of the stores this pass should generate. */ | |
175 | #define MAX_STORE_BITSIZE (BITS_PER_WORD) | |
176 | #define MAX_STORE_BYTES (MAX_STORE_BITSIZE / BITS_PER_UNIT) | |
177 | ||
245f6de1 JJ |
178 | /* Limit to bound the number of aliasing checks for loads with the same |
179 | vuse as the corresponding store. */ | |
180 | #define MAX_STORE_ALIAS_CHECKS 64 | |
181 | ||
f663d9ad KT |
182 | namespace { |
183 | ||
bebadeca | 184 | struct bswap_stat |
dffec8eb JJ |
185 | { |
186 | /* Number of hand-written 16-bit nop / bswaps found. */ | |
187 | int found_16bit; | |
188 | ||
189 | /* Number of hand-written 32-bit nop / bswaps found. */ | |
190 | int found_32bit; | |
191 | ||
192 | /* Number of hand-written 64-bit nop / bswaps found. */ | |
193 | int found_64bit; | |
194 | } nop_stats, bswap_stats; | |
195 | ||
196 | /* A symbolic number structure is used to detect byte permutation and selection | |
197 | patterns of a source. To achieve that, its field N contains an artificial | |
198 | number consisting of BITS_PER_MARKER sized markers tracking where does each | |
199 | byte come from in the source: | |
200 | ||
201 | 0 - target byte has the value 0 | |
202 | FF - target byte has an unknown value (eg. due to sign extension) | |
203 | 1..size - marker value is the byte index in the source (0 for lsb). | |
204 | ||
205 | To detect permutations on memory sources (arrays and structures), a symbolic | |
206 | number is also associated: | |
207 | - a base address BASE_ADDR and an OFFSET giving the address of the source; | |
208 | - a range which gives the difference between the highest and lowest accessed | |
209 | memory location to make such a symbolic number; | |
210 | - the address SRC of the source element of lowest address as a convenience | |
211 | to easily get BASE_ADDR + offset + lowest bytepos; | |
212 | - number of expressions N_OPS bitwise ored together to represent | |
213 | approximate cost of the computation. | |
214 | ||
215 | Note 1: the range is different from size as size reflects the size of the | |
216 | type of the current expression. For instance, for an array char a[], | |
217 | (short) a[0] | (short) a[3] would have a size of 2 but a range of 4 while | |
218 | (short) a[0] | ((short) a[0] << 1) would still have a size of 2 but this | |
219 | time a range of 1. | |
220 | ||
221 | Note 2: for non-memory sources, range holds the same value as size. | |
222 | ||
223 | Note 3: SRC points to the SSA_NAME in case of non-memory source. */ | |
224 | ||
225 | struct symbolic_number { | |
226 | uint64_t n; | |
227 | tree type; | |
228 | tree base_addr; | |
229 | tree offset; | |
4a022c70 | 230 | poly_int64_pod bytepos; |
dffec8eb JJ |
231 | tree src; |
232 | tree alias_set; | |
233 | tree vuse; | |
234 | unsigned HOST_WIDE_INT range; | |
235 | int n_ops; | |
236 | }; | |
237 | ||
238 | #define BITS_PER_MARKER 8 | |
239 | #define MARKER_MASK ((1 << BITS_PER_MARKER) - 1) | |
240 | #define MARKER_BYTE_UNKNOWN MARKER_MASK | |
241 | #define HEAD_MARKER(n, size) \ | |
242 | ((n) & ((uint64_t) MARKER_MASK << (((size) - 1) * BITS_PER_MARKER))) | |
243 | ||
244 | /* The number which the find_bswap_or_nop_1 result should match in | |
245 | order to have a nop. The number is masked according to the size of | |
246 | the symbolic number before using it. */ | |
247 | #define CMPNOP (sizeof (int64_t) < 8 ? 0 : \ | |
248 | (uint64_t)0x08070605 << 32 | 0x04030201) | |
249 | ||
250 | /* The number which the find_bswap_or_nop_1 result should match in | |
251 | order to have a byte swap. The number is masked according to the | |
252 | size of the symbolic number before using it. */ | |
253 | #define CMPXCHG (sizeof (int64_t) < 8 ? 0 : \ | |
254 | (uint64_t)0x01020304 << 32 | 0x05060708) | |
255 | ||
256 | /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic | |
257 | number N. Return false if the requested operation is not permitted | |
258 | on a symbolic number. */ | |
259 | ||
260 | inline bool | |
261 | do_shift_rotate (enum tree_code code, | |
262 | struct symbolic_number *n, | |
263 | int count) | |
264 | { | |
265 | int i, size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; | |
266 | unsigned head_marker; | |
267 | ||
444cda74 JJ |
268 | if (count < 0 |
269 | || count >= TYPE_PRECISION (n->type) | |
270 | || count % BITS_PER_UNIT != 0) | |
dffec8eb JJ |
271 | return false; |
272 | count = (count / BITS_PER_UNIT) * BITS_PER_MARKER; | |
273 | ||
274 | /* Zero out the extra bits of N in order to avoid them being shifted | |
275 | into the significant bits. */ | |
276 | if (size < 64 / BITS_PER_MARKER) | |
277 | n->n &= ((uint64_t) 1 << (size * BITS_PER_MARKER)) - 1; | |
278 | ||
279 | switch (code) | |
280 | { | |
281 | case LSHIFT_EXPR: | |
282 | n->n <<= count; | |
283 | break; | |
284 | case RSHIFT_EXPR: | |
285 | head_marker = HEAD_MARKER (n->n, size); | |
286 | n->n >>= count; | |
287 | /* Arithmetic shift of signed type: result is dependent on the value. */ | |
288 | if (!TYPE_UNSIGNED (n->type) && head_marker) | |
289 | for (i = 0; i < count / BITS_PER_MARKER; i++) | |
290 | n->n |= (uint64_t) MARKER_BYTE_UNKNOWN | |
291 | << ((size - 1 - i) * BITS_PER_MARKER); | |
292 | break; | |
293 | case LROTATE_EXPR: | |
294 | n->n = (n->n << count) | (n->n >> ((size * BITS_PER_MARKER) - count)); | |
295 | break; | |
296 | case RROTATE_EXPR: | |
297 | n->n = (n->n >> count) | (n->n << ((size * BITS_PER_MARKER) - count)); | |
298 | break; | |
299 | default: | |
300 | return false; | |
301 | } | |
302 | /* Zero unused bits for size. */ | |
303 | if (size < 64 / BITS_PER_MARKER) | |
304 | n->n &= ((uint64_t) 1 << (size * BITS_PER_MARKER)) - 1; | |
305 | return true; | |
306 | } | |
307 | ||
308 | /* Perform sanity checking for the symbolic number N and the gimple | |
309 | statement STMT. */ | |
310 | ||
311 | inline bool | |
312 | verify_symbolic_number_p (struct symbolic_number *n, gimple *stmt) | |
313 | { | |
314 | tree lhs_type; | |
315 | ||
316 | lhs_type = gimple_expr_type (stmt); | |
317 | ||
5ea39b24 JJ |
318 | if (TREE_CODE (lhs_type) != INTEGER_TYPE |
319 | && TREE_CODE (lhs_type) != ENUMERAL_TYPE) | |
dffec8eb JJ |
320 | return false; |
321 | ||
322 | if (TYPE_PRECISION (lhs_type) != TYPE_PRECISION (n->type)) | |
323 | return false; | |
324 | ||
325 | return true; | |
326 | } | |
327 | ||
328 | /* Initialize the symbolic number N for the bswap pass from the base element | |
329 | SRC manipulated by the bitwise OR expression. */ | |
330 | ||
331 | bool | |
332 | init_symbolic_number (struct symbolic_number *n, tree src) | |
333 | { | |
334 | int size; | |
335 | ||
336 | if (! INTEGRAL_TYPE_P (TREE_TYPE (src))) | |
337 | return false; | |
338 | ||
339 | n->base_addr = n->offset = n->alias_set = n->vuse = NULL_TREE; | |
340 | n->src = src; | |
341 | ||
342 | /* Set up the symbolic number N by setting each byte to a value between 1 and | |
343 | the byte size of rhs1. The highest order byte is set to n->size and the | |
344 | lowest order byte to 1. */ | |
345 | n->type = TREE_TYPE (src); | |
346 | size = TYPE_PRECISION (n->type); | |
347 | if (size % BITS_PER_UNIT != 0) | |
348 | return false; | |
349 | size /= BITS_PER_UNIT; | |
350 | if (size > 64 / BITS_PER_MARKER) | |
351 | return false; | |
352 | n->range = size; | |
353 | n->n = CMPNOP; | |
354 | n->n_ops = 1; | |
355 | ||
356 | if (size < 64 / BITS_PER_MARKER) | |
357 | n->n &= ((uint64_t) 1 << (size * BITS_PER_MARKER)) - 1; | |
358 | ||
359 | return true; | |
360 | } | |
361 | ||
362 | /* Check if STMT might be a byte swap or a nop from a memory source and returns | |
363 | the answer. If so, REF is that memory source and the base of the memory area | |
364 | accessed and the offset of the access from that base are recorded in N. */ | |
365 | ||
366 | bool | |
367 | find_bswap_or_nop_load (gimple *stmt, tree ref, struct symbolic_number *n) | |
368 | { | |
369 | /* Leaf node is an array or component ref. Memorize its base and | |
370 | offset from base to compare to other such leaf node. */ | |
f37fac2b | 371 | poly_int64 bitsize, bitpos, bytepos; |
dffec8eb JJ |
372 | machine_mode mode; |
373 | int unsignedp, reversep, volatilep; | |
374 | tree offset, base_addr; | |
375 | ||
376 | /* Not prepared to handle PDP endian. */ | |
377 | if (BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN) | |
378 | return false; | |
379 | ||
380 | if (!gimple_assign_load_p (stmt) || gimple_has_volatile_ops (stmt)) | |
381 | return false; | |
382 | ||
383 | base_addr = get_inner_reference (ref, &bitsize, &bitpos, &offset, &mode, | |
384 | &unsignedp, &reversep, &volatilep); | |
385 | ||
4b84d9b8 JJ |
386 | if (TREE_CODE (base_addr) == TARGET_MEM_REF) |
387 | /* Do not rewrite TARGET_MEM_REF. */ | |
388 | return false; | |
389 | else if (TREE_CODE (base_addr) == MEM_REF) | |
dffec8eb | 390 | { |
3fed2ce9 | 391 | poly_offset_int bit_offset = 0; |
dffec8eb JJ |
392 | tree off = TREE_OPERAND (base_addr, 1); |
393 | ||
394 | if (!integer_zerop (off)) | |
395 | { | |
3fed2ce9 RS |
396 | poly_offset_int boff = mem_ref_offset (base_addr); |
397 | boff <<= LOG2_BITS_PER_UNIT; | |
dffec8eb JJ |
398 | bit_offset += boff; |
399 | } | |
400 | ||
401 | base_addr = TREE_OPERAND (base_addr, 0); | |
402 | ||
403 | /* Avoid returning a negative bitpos as this may wreak havoc later. */ | |
3fed2ce9 | 404 | if (maybe_lt (bit_offset, 0)) |
dffec8eb | 405 | { |
3fed2ce9 RS |
406 | tree byte_offset = wide_int_to_tree |
407 | (sizetype, bits_to_bytes_round_down (bit_offset)); | |
408 | bit_offset = num_trailing_bits (bit_offset); | |
dffec8eb | 409 | if (offset) |
3fed2ce9 | 410 | offset = size_binop (PLUS_EXPR, offset, byte_offset); |
dffec8eb | 411 | else |
3fed2ce9 | 412 | offset = byte_offset; |
dffec8eb JJ |
413 | } |
414 | ||
3fed2ce9 | 415 | bitpos += bit_offset.force_shwi (); |
dffec8eb | 416 | } |
4b84d9b8 JJ |
417 | else |
418 | base_addr = build_fold_addr_expr (base_addr); | |
dffec8eb | 419 | |
f37fac2b | 420 | if (!multiple_p (bitpos, BITS_PER_UNIT, &bytepos)) |
dffec8eb | 421 | return false; |
f37fac2b | 422 | if (!multiple_p (bitsize, BITS_PER_UNIT)) |
dffec8eb JJ |
423 | return false; |
424 | if (reversep) | |
425 | return false; | |
426 | ||
427 | if (!init_symbolic_number (n, ref)) | |
428 | return false; | |
429 | n->base_addr = base_addr; | |
430 | n->offset = offset; | |
f37fac2b | 431 | n->bytepos = bytepos; |
dffec8eb JJ |
432 | n->alias_set = reference_alias_ptr_type (ref); |
433 | n->vuse = gimple_vuse (stmt); | |
434 | return true; | |
435 | } | |
436 | ||
437 | /* Compute the symbolic number N representing the result of a bitwise OR on 2 | |
438 | symbolic number N1 and N2 whose source statements are respectively | |
439 | SOURCE_STMT1 and SOURCE_STMT2. */ | |
440 | ||
441 | gimple * | |
442 | perform_symbolic_merge (gimple *source_stmt1, struct symbolic_number *n1, | |
443 | gimple *source_stmt2, struct symbolic_number *n2, | |
444 | struct symbolic_number *n) | |
445 | { | |
446 | int i, size; | |
447 | uint64_t mask; | |
448 | gimple *source_stmt; | |
449 | struct symbolic_number *n_start; | |
450 | ||
451 | tree rhs1 = gimple_assign_rhs1 (source_stmt1); | |
452 | if (TREE_CODE (rhs1) == BIT_FIELD_REF | |
453 | && TREE_CODE (TREE_OPERAND (rhs1, 0)) == SSA_NAME) | |
454 | rhs1 = TREE_OPERAND (rhs1, 0); | |
455 | tree rhs2 = gimple_assign_rhs1 (source_stmt2); | |
456 | if (TREE_CODE (rhs2) == BIT_FIELD_REF | |
457 | && TREE_CODE (TREE_OPERAND (rhs2, 0)) == SSA_NAME) | |
458 | rhs2 = TREE_OPERAND (rhs2, 0); | |
459 | ||
460 | /* Sources are different, cancel bswap if they are not memory location with | |
461 | the same base (array, structure, ...). */ | |
462 | if (rhs1 != rhs2) | |
463 | { | |
464 | uint64_t inc; | |
4a022c70 | 465 | HOST_WIDE_INT start1, start2, start_sub, end_sub, end1, end2, end; |
dffec8eb JJ |
466 | struct symbolic_number *toinc_n_ptr, *n_end; |
467 | basic_block bb1, bb2; | |
468 | ||
469 | if (!n1->base_addr || !n2->base_addr | |
470 | || !operand_equal_p (n1->base_addr, n2->base_addr, 0)) | |
471 | return NULL; | |
472 | ||
473 | if (!n1->offset != !n2->offset | |
474 | || (n1->offset && !operand_equal_p (n1->offset, n2->offset, 0))) | |
475 | return NULL; | |
476 | ||
4a022c70 RS |
477 | start1 = 0; |
478 | if (!(n2->bytepos - n1->bytepos).is_constant (&start2)) | |
479 | return NULL; | |
480 | ||
481 | if (start1 < start2) | |
dffec8eb JJ |
482 | { |
483 | n_start = n1; | |
4a022c70 | 484 | start_sub = start2 - start1; |
dffec8eb JJ |
485 | } |
486 | else | |
487 | { | |
488 | n_start = n2; | |
4a022c70 | 489 | start_sub = start1 - start2; |
dffec8eb JJ |
490 | } |
491 | ||
492 | bb1 = gimple_bb (source_stmt1); | |
493 | bb2 = gimple_bb (source_stmt2); | |
494 | if (dominated_by_p (CDI_DOMINATORS, bb1, bb2)) | |
495 | source_stmt = source_stmt1; | |
496 | else | |
497 | source_stmt = source_stmt2; | |
498 | ||
499 | /* Find the highest address at which a load is performed and | |
500 | compute related info. */ | |
4a022c70 RS |
501 | end1 = start1 + (n1->range - 1); |
502 | end2 = start2 + (n2->range - 1); | |
dffec8eb JJ |
503 | if (end1 < end2) |
504 | { | |
505 | end = end2; | |
506 | end_sub = end2 - end1; | |
507 | } | |
508 | else | |
509 | { | |
510 | end = end1; | |
511 | end_sub = end1 - end2; | |
512 | } | |
513 | n_end = (end2 > end1) ? n2 : n1; | |
514 | ||
515 | /* Find symbolic number whose lsb is the most significant. */ | |
516 | if (BYTES_BIG_ENDIAN) | |
517 | toinc_n_ptr = (n_end == n1) ? n2 : n1; | |
518 | else | |
519 | toinc_n_ptr = (n_start == n1) ? n2 : n1; | |
520 | ||
4a022c70 | 521 | n->range = end - MIN (start1, start2) + 1; |
dffec8eb JJ |
522 | |
523 | /* Check that the range of memory covered can be represented by | |
524 | a symbolic number. */ | |
525 | if (n->range > 64 / BITS_PER_MARKER) | |
526 | return NULL; | |
527 | ||
528 | /* Reinterpret byte marks in symbolic number holding the value of | |
529 | bigger weight according to target endianness. */ | |
530 | inc = BYTES_BIG_ENDIAN ? end_sub : start_sub; | |
531 | size = TYPE_PRECISION (n1->type) / BITS_PER_UNIT; | |
532 | for (i = 0; i < size; i++, inc <<= BITS_PER_MARKER) | |
533 | { | |
534 | unsigned marker | |
535 | = (toinc_n_ptr->n >> (i * BITS_PER_MARKER)) & MARKER_MASK; | |
536 | if (marker && marker != MARKER_BYTE_UNKNOWN) | |
537 | toinc_n_ptr->n += inc; | |
538 | } | |
539 | } | |
540 | else | |
541 | { | |
542 | n->range = n1->range; | |
543 | n_start = n1; | |
544 | source_stmt = source_stmt1; | |
545 | } | |
546 | ||
547 | if (!n1->alias_set | |
548 | || alias_ptr_types_compatible_p (n1->alias_set, n2->alias_set)) | |
549 | n->alias_set = n1->alias_set; | |
550 | else | |
551 | n->alias_set = ptr_type_node; | |
552 | n->vuse = n_start->vuse; | |
553 | n->base_addr = n_start->base_addr; | |
554 | n->offset = n_start->offset; | |
555 | n->src = n_start->src; | |
556 | n->bytepos = n_start->bytepos; | |
557 | n->type = n_start->type; | |
558 | size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; | |
559 | ||
560 | for (i = 0, mask = MARKER_MASK; i < size; i++, mask <<= BITS_PER_MARKER) | |
561 | { | |
562 | uint64_t masked1, masked2; | |
563 | ||
564 | masked1 = n1->n & mask; | |
565 | masked2 = n2->n & mask; | |
566 | if (masked1 && masked2 && masked1 != masked2) | |
567 | return NULL; | |
568 | } | |
569 | n->n = n1->n | n2->n; | |
570 | n->n_ops = n1->n_ops + n2->n_ops; | |
571 | ||
572 | return source_stmt; | |
573 | } | |
574 | ||
575 | /* find_bswap_or_nop_1 invokes itself recursively with N and tries to perform | |
576 | the operation given by the rhs of STMT on the result. If the operation | |
577 | could successfully be executed the function returns a gimple stmt whose | |
578 | rhs's first tree is the expression of the source operand and NULL | |
579 | otherwise. */ | |
580 | ||
581 | gimple * | |
582 | find_bswap_or_nop_1 (gimple *stmt, struct symbolic_number *n, int limit) | |
583 | { | |
584 | enum tree_code code; | |
585 | tree rhs1, rhs2 = NULL; | |
586 | gimple *rhs1_stmt, *rhs2_stmt, *source_stmt1; | |
587 | enum gimple_rhs_class rhs_class; | |
588 | ||
589 | if (!limit || !is_gimple_assign (stmt)) | |
590 | return NULL; | |
591 | ||
592 | rhs1 = gimple_assign_rhs1 (stmt); | |
593 | ||
594 | if (find_bswap_or_nop_load (stmt, rhs1, n)) | |
595 | return stmt; | |
596 | ||
597 | /* Handle BIT_FIELD_REF. */ | |
598 | if (TREE_CODE (rhs1) == BIT_FIELD_REF | |
599 | && TREE_CODE (TREE_OPERAND (rhs1, 0)) == SSA_NAME) | |
600 | { | |
35cf3c55 KZ |
601 | if (!tree_fits_uhwi_p (TREE_OPERAND (rhs1, 1)) |
602 | || !tree_fits_uhwi_p (TREE_OPERAND (rhs1, 2))) | |
603 | return NULL; | |
604 | ||
dffec8eb JJ |
605 | unsigned HOST_WIDE_INT bitsize = tree_to_uhwi (TREE_OPERAND (rhs1, 1)); |
606 | unsigned HOST_WIDE_INT bitpos = tree_to_uhwi (TREE_OPERAND (rhs1, 2)); | |
607 | if (bitpos % BITS_PER_UNIT == 0 | |
608 | && bitsize % BITS_PER_UNIT == 0 | |
609 | && init_symbolic_number (n, TREE_OPERAND (rhs1, 0))) | |
610 | { | |
611 | /* Handle big-endian bit numbering in BIT_FIELD_REF. */ | |
612 | if (BYTES_BIG_ENDIAN) | |
613 | bitpos = TYPE_PRECISION (n->type) - bitpos - bitsize; | |
614 | ||
615 | /* Shift. */ | |
616 | if (!do_shift_rotate (RSHIFT_EXPR, n, bitpos)) | |
617 | return NULL; | |
618 | ||
619 | /* Mask. */ | |
620 | uint64_t mask = 0; | |
621 | uint64_t tmp = (1 << BITS_PER_UNIT) - 1; | |
622 | for (unsigned i = 0; i < bitsize / BITS_PER_UNIT; | |
623 | i++, tmp <<= BITS_PER_UNIT) | |
624 | mask |= (uint64_t) MARKER_MASK << (i * BITS_PER_MARKER); | |
625 | n->n &= mask; | |
626 | ||
627 | /* Convert. */ | |
628 | n->type = TREE_TYPE (rhs1); | |
629 | if (!n->base_addr) | |
630 | n->range = TYPE_PRECISION (n->type) / BITS_PER_UNIT; | |
631 | ||
632 | return verify_symbolic_number_p (n, stmt) ? stmt : NULL; | |
633 | } | |
634 | ||
635 | return NULL; | |
636 | } | |
637 | ||
638 | if (TREE_CODE (rhs1) != SSA_NAME) | |
639 | return NULL; | |
640 | ||
641 | code = gimple_assign_rhs_code (stmt); | |
642 | rhs_class = gimple_assign_rhs_class (stmt); | |
643 | rhs1_stmt = SSA_NAME_DEF_STMT (rhs1); | |
644 | ||
645 | if (rhs_class == GIMPLE_BINARY_RHS) | |
646 | rhs2 = gimple_assign_rhs2 (stmt); | |
647 | ||
648 | /* Handle unary rhs and binary rhs with integer constants as second | |
649 | operand. */ | |
650 | ||
651 | if (rhs_class == GIMPLE_UNARY_RHS | |
652 | || (rhs_class == GIMPLE_BINARY_RHS | |
653 | && TREE_CODE (rhs2) == INTEGER_CST)) | |
654 | { | |
655 | if (code != BIT_AND_EXPR | |
656 | && code != LSHIFT_EXPR | |
657 | && code != RSHIFT_EXPR | |
658 | && code != LROTATE_EXPR | |
659 | && code != RROTATE_EXPR | |
660 | && !CONVERT_EXPR_CODE_P (code)) | |
661 | return NULL; | |
662 | ||
663 | source_stmt1 = find_bswap_or_nop_1 (rhs1_stmt, n, limit - 1); | |
664 | ||
665 | /* If find_bswap_or_nop_1 returned NULL, STMT is a leaf node and | |
666 | we have to initialize the symbolic number. */ | |
667 | if (!source_stmt1) | |
668 | { | |
669 | if (gimple_assign_load_p (stmt) | |
670 | || !init_symbolic_number (n, rhs1)) | |
671 | return NULL; | |
672 | source_stmt1 = stmt; | |
673 | } | |
674 | ||
675 | switch (code) | |
676 | { | |
677 | case BIT_AND_EXPR: | |
678 | { | |
679 | int i, size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; | |
680 | uint64_t val = int_cst_value (rhs2), mask = 0; | |
681 | uint64_t tmp = (1 << BITS_PER_UNIT) - 1; | |
682 | ||
683 | /* Only constants masking full bytes are allowed. */ | |
684 | for (i = 0; i < size; i++, tmp <<= BITS_PER_UNIT) | |
685 | if ((val & tmp) != 0 && (val & tmp) != tmp) | |
686 | return NULL; | |
687 | else if (val & tmp) | |
688 | mask |= (uint64_t) MARKER_MASK << (i * BITS_PER_MARKER); | |
689 | ||
690 | n->n &= mask; | |
691 | } | |
692 | break; | |
693 | case LSHIFT_EXPR: | |
694 | case RSHIFT_EXPR: | |
695 | case LROTATE_EXPR: | |
696 | case RROTATE_EXPR: | |
697 | if (!do_shift_rotate (code, n, (int) TREE_INT_CST_LOW (rhs2))) | |
698 | return NULL; | |
699 | break; | |
700 | CASE_CONVERT: | |
701 | { | |
702 | int i, type_size, old_type_size; | |
703 | tree type; | |
704 | ||
705 | type = gimple_expr_type (stmt); | |
706 | type_size = TYPE_PRECISION (type); | |
707 | if (type_size % BITS_PER_UNIT != 0) | |
708 | return NULL; | |
709 | type_size /= BITS_PER_UNIT; | |
710 | if (type_size > 64 / BITS_PER_MARKER) | |
711 | return NULL; | |
712 | ||
713 | /* Sign extension: result is dependent on the value. */ | |
714 | old_type_size = TYPE_PRECISION (n->type) / BITS_PER_UNIT; | |
715 | if (!TYPE_UNSIGNED (n->type) && type_size > old_type_size | |
716 | && HEAD_MARKER (n->n, old_type_size)) | |
717 | for (i = 0; i < type_size - old_type_size; i++) | |
718 | n->n |= (uint64_t) MARKER_BYTE_UNKNOWN | |
719 | << ((type_size - 1 - i) * BITS_PER_MARKER); | |
720 | ||
721 | if (type_size < 64 / BITS_PER_MARKER) | |
722 | { | |
723 | /* If STMT casts to a smaller type mask out the bits not | |
724 | belonging to the target type. */ | |
725 | n->n &= ((uint64_t) 1 << (type_size * BITS_PER_MARKER)) - 1; | |
726 | } | |
727 | n->type = type; | |
728 | if (!n->base_addr) | |
729 | n->range = type_size; | |
730 | } | |
731 | break; | |
732 | default: | |
733 | return NULL; | |
734 | }; | |
735 | return verify_symbolic_number_p (n, stmt) ? source_stmt1 : NULL; | |
736 | } | |
737 | ||
738 | /* Handle binary rhs. */ | |
739 | ||
740 | if (rhs_class == GIMPLE_BINARY_RHS) | |
741 | { | |
742 | struct symbolic_number n1, n2; | |
743 | gimple *source_stmt, *source_stmt2; | |
744 | ||
745 | if (code != BIT_IOR_EXPR) | |
746 | return NULL; | |
747 | ||
748 | if (TREE_CODE (rhs2) != SSA_NAME) | |
749 | return NULL; | |
750 | ||
751 | rhs2_stmt = SSA_NAME_DEF_STMT (rhs2); | |
752 | ||
753 | switch (code) | |
754 | { | |
755 | case BIT_IOR_EXPR: | |
756 | source_stmt1 = find_bswap_or_nop_1 (rhs1_stmt, &n1, limit - 1); | |
757 | ||
758 | if (!source_stmt1) | |
759 | return NULL; | |
760 | ||
761 | source_stmt2 = find_bswap_or_nop_1 (rhs2_stmt, &n2, limit - 1); | |
762 | ||
763 | if (!source_stmt2) | |
764 | return NULL; | |
765 | ||
766 | if (TYPE_PRECISION (n1.type) != TYPE_PRECISION (n2.type)) | |
767 | return NULL; | |
768 | ||
4b84d9b8 | 769 | if (n1.vuse != n2.vuse) |
dffec8eb JJ |
770 | return NULL; |
771 | ||
772 | source_stmt | |
773 | = perform_symbolic_merge (source_stmt1, &n1, source_stmt2, &n2, n); | |
774 | ||
775 | if (!source_stmt) | |
776 | return NULL; | |
777 | ||
778 | if (!verify_symbolic_number_p (n, stmt)) | |
779 | return NULL; | |
780 | ||
781 | break; | |
782 | default: | |
783 | return NULL; | |
784 | } | |
785 | return source_stmt; | |
786 | } | |
787 | return NULL; | |
788 | } | |
789 | ||
4b84d9b8 JJ |
790 | /* Helper for find_bswap_or_nop and try_coalesce_bswap to compute |
791 | *CMPXCHG, *CMPNOP and adjust *N. */ | |
dffec8eb | 792 | |
4b84d9b8 JJ |
793 | void |
794 | find_bswap_or_nop_finalize (struct symbolic_number *n, uint64_t *cmpxchg, | |
795 | uint64_t *cmpnop) | |
dffec8eb JJ |
796 | { |
797 | unsigned rsize; | |
798 | uint64_t tmpn, mask; | |
dffec8eb | 799 | |
4b84d9b8 JJ |
800 | /* The number which the find_bswap_or_nop_1 result should match in order |
801 | to have a full byte swap. The number is shifted to the right | |
802 | according to the size of the symbolic number before using it. */ | |
803 | *cmpxchg = CMPXCHG; | |
804 | *cmpnop = CMPNOP; | |
dffec8eb JJ |
805 | |
806 | /* Find real size of result (highest non-zero byte). */ | |
807 | if (n->base_addr) | |
808 | for (tmpn = n->n, rsize = 0; tmpn; tmpn >>= BITS_PER_MARKER, rsize++); | |
809 | else | |
810 | rsize = n->range; | |
811 | ||
812 | /* Zero out the bits corresponding to untouched bytes in original gimple | |
813 | expression. */ | |
814 | if (n->range < (int) sizeof (int64_t)) | |
815 | { | |
816 | mask = ((uint64_t) 1 << (n->range * BITS_PER_MARKER)) - 1; | |
4b84d9b8 JJ |
817 | *cmpxchg >>= (64 / BITS_PER_MARKER - n->range) * BITS_PER_MARKER; |
818 | *cmpnop &= mask; | |
dffec8eb JJ |
819 | } |
820 | ||
821 | /* Zero out the bits corresponding to unused bytes in the result of the | |
822 | gimple expression. */ | |
823 | if (rsize < n->range) | |
824 | { | |
825 | if (BYTES_BIG_ENDIAN) | |
826 | { | |
827 | mask = ((uint64_t) 1 << (rsize * BITS_PER_MARKER)) - 1; | |
4b84d9b8 JJ |
828 | *cmpxchg &= mask; |
829 | *cmpnop >>= (n->range - rsize) * BITS_PER_MARKER; | |
dffec8eb JJ |
830 | } |
831 | else | |
832 | { | |
833 | mask = ((uint64_t) 1 << (rsize * BITS_PER_MARKER)) - 1; | |
4b84d9b8 JJ |
834 | *cmpxchg >>= (n->range - rsize) * BITS_PER_MARKER; |
835 | *cmpnop &= mask; | |
dffec8eb JJ |
836 | } |
837 | n->range = rsize; | |
838 | } | |
839 | ||
4b84d9b8 JJ |
840 | n->range *= BITS_PER_UNIT; |
841 | } | |
842 | ||
843 | /* Check if STMT completes a bswap implementation or a read in a given | |
844 | endianness consisting of ORs, SHIFTs and ANDs and sets *BSWAP | |
845 | accordingly. It also sets N to represent the kind of operations | |
846 | performed: size of the resulting expression and whether it works on | |
847 | a memory source, and if so alias-set and vuse. At last, the | |
848 | function returns a stmt whose rhs's first tree is the source | |
849 | expression. */ | |
850 | ||
851 | gimple * | |
852 | find_bswap_or_nop (gimple *stmt, struct symbolic_number *n, bool *bswap) | |
853 | { | |
7f0ce82a KT |
854 | tree type_size = TYPE_SIZE_UNIT (gimple_expr_type (stmt)); |
855 | if (!tree_fits_uhwi_p (type_size)) | |
856 | return NULL; | |
857 | ||
4b84d9b8 JJ |
858 | /* The last parameter determines the depth search limit. It usually |
859 | correlates directly to the number n of bytes to be touched. We | |
0f507a36 | 860 | increase that number by 2 * (log2(n) + 1) here in order to also |
4b84d9b8 JJ |
861 | cover signed -> unsigned conversions of the src operand as can be seen |
862 | in libgcc, and for initial shift/and operation of the src operand. */ | |
7f0ce82a | 863 | int limit = tree_to_uhwi (type_size); |
0f507a36 | 864 | limit += 2 * (1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT) limit)); |
4b84d9b8 JJ |
865 | gimple *ins_stmt = find_bswap_or_nop_1 (stmt, n, limit); |
866 | ||
867 | if (!ins_stmt) | |
cd676dfa JJ |
868 | { |
869 | if (gimple_assign_rhs_code (stmt) != CONSTRUCTOR | |
870 | || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN) | |
871 | return NULL; | |
872 | unsigned HOST_WIDE_INT sz = tree_to_uhwi (type_size) * BITS_PER_UNIT; | |
873 | if (sz != 16 && sz != 32 && sz != 64) | |
874 | return NULL; | |
875 | tree rhs = gimple_assign_rhs1 (stmt); | |
9032d2b2 JJ |
876 | if (CONSTRUCTOR_NELTS (rhs) == 0) |
877 | return NULL; | |
cd676dfa JJ |
878 | tree eltype = TREE_TYPE (TREE_TYPE (rhs)); |
879 | unsigned HOST_WIDE_INT eltsz | |
880 | = int_size_in_bytes (eltype) * BITS_PER_UNIT; | |
881 | if (TYPE_PRECISION (eltype) != eltsz) | |
882 | return NULL; | |
883 | constructor_elt *elt; | |
884 | unsigned int i; | |
885 | tree type = build_nonstandard_integer_type (sz, 1); | |
886 | FOR_EACH_VEC_SAFE_ELT (CONSTRUCTOR_ELTS (rhs), i, elt) | |
887 | { | |
888 | if (TREE_CODE (elt->value) != SSA_NAME | |
889 | || !INTEGRAL_TYPE_P (TREE_TYPE (elt->value))) | |
890 | return NULL; | |
891 | struct symbolic_number n1; | |
892 | gimple *source_stmt | |
893 | = find_bswap_or_nop_1 (SSA_NAME_DEF_STMT (elt->value), &n1, | |
894 | limit - 1); | |
895 | ||
896 | if (!source_stmt) | |
897 | return NULL; | |
898 | ||
899 | n1.type = type; | |
900 | if (!n1.base_addr) | |
901 | n1.range = sz / BITS_PER_UNIT; | |
902 | ||
903 | if (i == 0) | |
904 | { | |
905 | ins_stmt = source_stmt; | |
906 | *n = n1; | |
907 | } | |
908 | else | |
909 | { | |
910 | if (n->vuse != n1.vuse) | |
911 | return NULL; | |
912 | ||
913 | struct symbolic_number n0 = *n; | |
914 | ||
915 | if (!BYTES_BIG_ENDIAN) | |
916 | { | |
917 | if (!do_shift_rotate (LSHIFT_EXPR, &n1, i * eltsz)) | |
918 | return NULL; | |
919 | } | |
920 | else if (!do_shift_rotate (LSHIFT_EXPR, &n0, eltsz)) | |
921 | return NULL; | |
922 | ins_stmt | |
923 | = perform_symbolic_merge (ins_stmt, &n0, source_stmt, &n1, n); | |
924 | ||
925 | if (!ins_stmt) | |
926 | return NULL; | |
927 | } | |
928 | } | |
929 | } | |
4b84d9b8 JJ |
930 | |
931 | uint64_t cmpxchg, cmpnop; | |
932 | find_bswap_or_nop_finalize (n, &cmpxchg, &cmpnop); | |
933 | ||
dffec8eb JJ |
934 | /* A complete byte swap should make the symbolic number to start with |
935 | the largest digit in the highest order byte. Unchanged symbolic | |
936 | number indicates a read with same endianness as target architecture. */ | |
937 | if (n->n == cmpnop) | |
938 | *bswap = false; | |
939 | else if (n->n == cmpxchg) | |
940 | *bswap = true; | |
941 | else | |
942 | return NULL; | |
943 | ||
944 | /* Useless bit manipulation performed by code. */ | |
945 | if (!n->base_addr && n->n == cmpnop && n->n_ops == 1) | |
946 | return NULL; | |
947 | ||
dffec8eb JJ |
948 | return ins_stmt; |
949 | } | |
950 | ||
951 | const pass_data pass_data_optimize_bswap = | |
952 | { | |
953 | GIMPLE_PASS, /* type */ | |
954 | "bswap", /* name */ | |
955 | OPTGROUP_NONE, /* optinfo_flags */ | |
956 | TV_NONE, /* tv_id */ | |
957 | PROP_ssa, /* properties_required */ | |
958 | 0, /* properties_provided */ | |
959 | 0, /* properties_destroyed */ | |
960 | 0, /* todo_flags_start */ | |
961 | 0, /* todo_flags_finish */ | |
962 | }; | |
963 | ||
964 | class pass_optimize_bswap : public gimple_opt_pass | |
965 | { | |
966 | public: | |
967 | pass_optimize_bswap (gcc::context *ctxt) | |
968 | : gimple_opt_pass (pass_data_optimize_bswap, ctxt) | |
969 | {} | |
970 | ||
971 | /* opt_pass methods: */ | |
972 | virtual bool gate (function *) | |
973 | { | |
974 | return flag_expensive_optimizations && optimize && BITS_PER_UNIT == 8; | |
975 | } | |
976 | ||
977 | virtual unsigned int execute (function *); | |
978 | ||
979 | }; // class pass_optimize_bswap | |
980 | ||
d02a8b63 JJ |
981 | /* Helper function for bswap_replace. Build VIEW_CONVERT_EXPR from |
982 | VAL to TYPE. If VAL has different type size, emit a NOP_EXPR cast | |
983 | first. */ | |
984 | ||
985 | static tree | |
986 | bswap_view_convert (gimple_stmt_iterator *gsi, tree type, tree val) | |
987 | { | |
988 | gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (val))); | |
989 | if (TYPE_SIZE (type) != TYPE_SIZE (TREE_TYPE (val))) | |
990 | { | |
991 | HOST_WIDE_INT prec = TREE_INT_CST_LOW (TYPE_SIZE (type)); | |
992 | tree itype = build_nonstandard_integer_type (prec, 1); | |
993 | gimple *g = gimple_build_assign (make_ssa_name (itype), NOP_EXPR, val); | |
994 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
995 | val = gimple_assign_lhs (g); | |
996 | } | |
997 | return build1 (VIEW_CONVERT_EXPR, type, val); | |
998 | } | |
999 | ||
dffec8eb | 1000 | /* Perform the bswap optimization: replace the expression computed in the rhs |
4b84d9b8 JJ |
1001 | of gsi_stmt (GSI) (or if NULL add instead of replace) by an equivalent |
1002 | bswap, load or load + bswap expression. | |
dffec8eb JJ |
1003 | Which of these alternatives replace the rhs is given by N->base_addr (non |
1004 | null if a load is needed) and BSWAP. The type, VUSE and set-alias of the | |
1005 | load to perform are also given in N while the builtin bswap invoke is given | |
4b84d9b8 JJ |
1006 | in FNDEL. Finally, if a load is involved, INS_STMT refers to one of the |
1007 | load statements involved to construct the rhs in gsi_stmt (GSI) and | |
1008 | N->range gives the size of the rhs expression for maintaining some | |
1009 | statistics. | |
dffec8eb | 1010 | |
4b84d9b8 JJ |
1011 | Note that if the replacement involve a load and if gsi_stmt (GSI) is |
1012 | non-NULL, that stmt is moved just after INS_STMT to do the load with the | |
1013 | same VUSE which can lead to gsi_stmt (GSI) changing of basic block. */ | |
dffec8eb | 1014 | |
4b84d9b8 JJ |
1015 | tree |
1016 | bswap_replace (gimple_stmt_iterator gsi, gimple *ins_stmt, tree fndecl, | |
dffec8eb JJ |
1017 | tree bswap_type, tree load_type, struct symbolic_number *n, |
1018 | bool bswap) | |
1019 | { | |
4b84d9b8 | 1020 | tree src, tmp, tgt = NULL_TREE; |
dffec8eb | 1021 | gimple *bswap_stmt; |
cd676dfa | 1022 | tree_code conv_code = NOP_EXPR; |
dffec8eb | 1023 | |
4b84d9b8 | 1024 | gimple *cur_stmt = gsi_stmt (gsi); |
dffec8eb | 1025 | src = n->src; |
4b84d9b8 | 1026 | if (cur_stmt) |
cd676dfa JJ |
1027 | { |
1028 | tgt = gimple_assign_lhs (cur_stmt); | |
1029 | if (gimple_assign_rhs_code (cur_stmt) == CONSTRUCTOR | |
1030 | && tgt | |
1031 | && VECTOR_TYPE_P (TREE_TYPE (tgt))) | |
1032 | conv_code = VIEW_CONVERT_EXPR; | |
1033 | } | |
dffec8eb JJ |
1034 | |
1035 | /* Need to load the value from memory first. */ | |
1036 | if (n->base_addr) | |
1037 | { | |
4b84d9b8 JJ |
1038 | gimple_stmt_iterator gsi_ins = gsi; |
1039 | if (ins_stmt) | |
1040 | gsi_ins = gsi_for_stmt (ins_stmt); | |
dffec8eb JJ |
1041 | tree addr_expr, addr_tmp, val_expr, val_tmp; |
1042 | tree load_offset_ptr, aligned_load_type; | |
4b84d9b8 JJ |
1043 | gimple *load_stmt; |
1044 | unsigned align = get_object_alignment (src); | |
4a022c70 | 1045 | poly_int64 load_offset = 0; |
dffec8eb | 1046 | |
4b84d9b8 JJ |
1047 | if (cur_stmt) |
1048 | { | |
1049 | basic_block ins_bb = gimple_bb (ins_stmt); | |
1050 | basic_block cur_bb = gimple_bb (cur_stmt); | |
1051 | if (!dominated_by_p (CDI_DOMINATORS, cur_bb, ins_bb)) | |
1052 | return NULL_TREE; | |
1053 | ||
1054 | /* Move cur_stmt just before one of the load of the original | |
1055 | to ensure it has the same VUSE. See PR61517 for what could | |
1056 | go wrong. */ | |
1057 | if (gimple_bb (cur_stmt) != gimple_bb (ins_stmt)) | |
1058 | reset_flow_sensitive_info (gimple_assign_lhs (cur_stmt)); | |
1059 | gsi_move_before (&gsi, &gsi_ins); | |
1060 | gsi = gsi_for_stmt (cur_stmt); | |
1061 | } | |
1062 | else | |
1063 | gsi = gsi_ins; | |
dffec8eb JJ |
1064 | |
1065 | /* Compute address to load from and cast according to the size | |
1066 | of the load. */ | |
4b84d9b8 | 1067 | addr_expr = build_fold_addr_expr (src); |
dffec8eb | 1068 | if (is_gimple_mem_ref_addr (addr_expr)) |
4b84d9b8 | 1069 | addr_tmp = unshare_expr (addr_expr); |
dffec8eb JJ |
1070 | else |
1071 | { | |
4b84d9b8 JJ |
1072 | addr_tmp = unshare_expr (n->base_addr); |
1073 | if (!is_gimple_mem_ref_addr (addr_tmp)) | |
1074 | addr_tmp = force_gimple_operand_gsi_1 (&gsi, addr_tmp, | |
1075 | is_gimple_mem_ref_addr, | |
1076 | NULL_TREE, true, | |
1077 | GSI_SAME_STMT); | |
1078 | load_offset = n->bytepos; | |
1079 | if (n->offset) | |
1080 | { | |
1081 | tree off | |
1082 | = force_gimple_operand_gsi (&gsi, unshare_expr (n->offset), | |
1083 | true, NULL_TREE, true, | |
1084 | GSI_SAME_STMT); | |
1085 | gimple *stmt | |
1086 | = gimple_build_assign (make_ssa_name (TREE_TYPE (addr_tmp)), | |
1087 | POINTER_PLUS_EXPR, addr_tmp, off); | |
1088 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1089 | addr_tmp = gimple_assign_lhs (stmt); | |
1090 | } | |
dffec8eb JJ |
1091 | } |
1092 | ||
1093 | /* Perform the load. */ | |
1094 | aligned_load_type = load_type; | |
1095 | if (align < TYPE_ALIGN (load_type)) | |
1096 | aligned_load_type = build_aligned_type (load_type, align); | |
1097 | load_offset_ptr = build_int_cst (n->alias_set, load_offset); | |
1098 | val_expr = fold_build2 (MEM_REF, aligned_load_type, addr_tmp, | |
1099 | load_offset_ptr); | |
1100 | ||
1101 | if (!bswap) | |
1102 | { | |
1103 | if (n->range == 16) | |
1104 | nop_stats.found_16bit++; | |
1105 | else if (n->range == 32) | |
1106 | nop_stats.found_32bit++; | |
1107 | else | |
1108 | { | |
1109 | gcc_assert (n->range == 64); | |
1110 | nop_stats.found_64bit++; | |
1111 | } | |
1112 | ||
1113 | /* Convert the result of load if necessary. */ | |
4b84d9b8 | 1114 | if (tgt && !useless_type_conversion_p (TREE_TYPE (tgt), load_type)) |
dffec8eb JJ |
1115 | { |
1116 | val_tmp = make_temp_ssa_name (aligned_load_type, NULL, | |
1117 | "load_dst"); | |
1118 | load_stmt = gimple_build_assign (val_tmp, val_expr); | |
1119 | gimple_set_vuse (load_stmt, n->vuse); | |
1120 | gsi_insert_before (&gsi, load_stmt, GSI_SAME_STMT); | |
cd676dfa | 1121 | if (conv_code == VIEW_CONVERT_EXPR) |
d02a8b63 | 1122 | val_tmp = bswap_view_convert (&gsi, TREE_TYPE (tgt), val_tmp); |
cd676dfa | 1123 | gimple_assign_set_rhs_with_ops (&gsi, conv_code, val_tmp); |
4b84d9b8 | 1124 | update_stmt (cur_stmt); |
dffec8eb | 1125 | } |
4b84d9b8 | 1126 | else if (cur_stmt) |
dffec8eb JJ |
1127 | { |
1128 | gimple_assign_set_rhs_with_ops (&gsi, MEM_REF, val_expr); | |
1129 | gimple_set_vuse (cur_stmt, n->vuse); | |
4b84d9b8 JJ |
1130 | update_stmt (cur_stmt); |
1131 | } | |
1132 | else | |
1133 | { | |
1134 | tgt = make_ssa_name (load_type); | |
1135 | cur_stmt = gimple_build_assign (tgt, MEM_REF, val_expr); | |
1136 | gimple_set_vuse (cur_stmt, n->vuse); | |
1137 | gsi_insert_before (&gsi, cur_stmt, GSI_SAME_STMT); | |
dffec8eb | 1138 | } |
dffec8eb JJ |
1139 | |
1140 | if (dump_file) | |
1141 | { | |
1142 | fprintf (dump_file, | |
1143 | "%d bit load in target endianness found at: ", | |
1144 | (int) n->range); | |
1145 | print_gimple_stmt (dump_file, cur_stmt, 0); | |
1146 | } | |
4b84d9b8 | 1147 | return tgt; |
dffec8eb JJ |
1148 | } |
1149 | else | |
1150 | { | |
1151 | val_tmp = make_temp_ssa_name (aligned_load_type, NULL, "load_dst"); | |
1152 | load_stmt = gimple_build_assign (val_tmp, val_expr); | |
1153 | gimple_set_vuse (load_stmt, n->vuse); | |
1154 | gsi_insert_before (&gsi, load_stmt, GSI_SAME_STMT); | |
1155 | } | |
1156 | src = val_tmp; | |
1157 | } | |
1158 | else if (!bswap) | |
1159 | { | |
4b84d9b8 JJ |
1160 | gimple *g = NULL; |
1161 | if (tgt && !useless_type_conversion_p (TREE_TYPE (tgt), TREE_TYPE (src))) | |
dffec8eb JJ |
1162 | { |
1163 | if (!is_gimple_val (src)) | |
4b84d9b8 | 1164 | return NULL_TREE; |
cd676dfa | 1165 | if (conv_code == VIEW_CONVERT_EXPR) |
d02a8b63 | 1166 | src = bswap_view_convert (&gsi, TREE_TYPE (tgt), src); |
cd676dfa | 1167 | g = gimple_build_assign (tgt, conv_code, src); |
dffec8eb | 1168 | } |
4b84d9b8 | 1169 | else if (cur_stmt) |
dffec8eb | 1170 | g = gimple_build_assign (tgt, src); |
4b84d9b8 JJ |
1171 | else |
1172 | tgt = src; | |
dffec8eb JJ |
1173 | if (n->range == 16) |
1174 | nop_stats.found_16bit++; | |
1175 | else if (n->range == 32) | |
1176 | nop_stats.found_32bit++; | |
1177 | else | |
1178 | { | |
1179 | gcc_assert (n->range == 64); | |
1180 | nop_stats.found_64bit++; | |
1181 | } | |
1182 | if (dump_file) | |
1183 | { | |
1184 | fprintf (dump_file, | |
1185 | "%d bit reshuffle in target endianness found at: ", | |
1186 | (int) n->range); | |
4b84d9b8 JJ |
1187 | if (cur_stmt) |
1188 | print_gimple_stmt (dump_file, cur_stmt, 0); | |
1189 | else | |
1190 | { | |
4af78ef8 | 1191 | print_generic_expr (dump_file, tgt, TDF_NONE); |
4b84d9b8 JJ |
1192 | fprintf (dump_file, "\n"); |
1193 | } | |
dffec8eb | 1194 | } |
4b84d9b8 JJ |
1195 | if (cur_stmt) |
1196 | gsi_replace (&gsi, g, true); | |
1197 | return tgt; | |
dffec8eb JJ |
1198 | } |
1199 | else if (TREE_CODE (src) == BIT_FIELD_REF) | |
1200 | src = TREE_OPERAND (src, 0); | |
1201 | ||
1202 | if (n->range == 16) | |
1203 | bswap_stats.found_16bit++; | |
1204 | else if (n->range == 32) | |
1205 | bswap_stats.found_32bit++; | |
1206 | else | |
1207 | { | |
1208 | gcc_assert (n->range == 64); | |
1209 | bswap_stats.found_64bit++; | |
1210 | } | |
1211 | ||
1212 | tmp = src; | |
1213 | ||
1214 | /* Convert the src expression if necessary. */ | |
1215 | if (!useless_type_conversion_p (TREE_TYPE (tmp), bswap_type)) | |
1216 | { | |
1217 | gimple *convert_stmt; | |
1218 | ||
1219 | tmp = make_temp_ssa_name (bswap_type, NULL, "bswapsrc"); | |
1220 | convert_stmt = gimple_build_assign (tmp, NOP_EXPR, src); | |
1221 | gsi_insert_before (&gsi, convert_stmt, GSI_SAME_STMT); | |
1222 | } | |
1223 | ||
1224 | /* Canonical form for 16 bit bswap is a rotate expression. Only 16bit values | |
1225 | are considered as rotation of 2N bit values by N bits is generally not | |
1226 | equivalent to a bswap. Consider for instance 0x01020304 r>> 16 which | |
1227 | gives 0x03040102 while a bswap for that value is 0x04030201. */ | |
1228 | if (bswap && n->range == 16) | |
1229 | { | |
1230 | tree count = build_int_cst (NULL, BITS_PER_UNIT); | |
1231 | src = fold_build2 (LROTATE_EXPR, bswap_type, tmp, count); | |
1232 | bswap_stmt = gimple_build_assign (NULL, src); | |
1233 | } | |
1234 | else | |
1235 | bswap_stmt = gimple_build_call (fndecl, 1, tmp); | |
1236 | ||
4b84d9b8 JJ |
1237 | if (tgt == NULL_TREE) |
1238 | tgt = make_ssa_name (bswap_type); | |
dffec8eb JJ |
1239 | tmp = tgt; |
1240 | ||
1241 | /* Convert the result if necessary. */ | |
1242 | if (!useless_type_conversion_p (TREE_TYPE (tgt), bswap_type)) | |
1243 | { | |
1244 | gimple *convert_stmt; | |
1245 | ||
1246 | tmp = make_temp_ssa_name (bswap_type, NULL, "bswapdst"); | |
cd676dfa JJ |
1247 | tree atmp = tmp; |
1248 | if (conv_code == VIEW_CONVERT_EXPR) | |
d02a8b63 | 1249 | atmp = bswap_view_convert (&gsi, TREE_TYPE (tgt), tmp); |
cd676dfa | 1250 | convert_stmt = gimple_build_assign (tgt, conv_code, atmp); |
dffec8eb JJ |
1251 | gsi_insert_after (&gsi, convert_stmt, GSI_SAME_STMT); |
1252 | } | |
1253 | ||
1254 | gimple_set_lhs (bswap_stmt, tmp); | |
1255 | ||
1256 | if (dump_file) | |
1257 | { | |
1258 | fprintf (dump_file, "%d bit bswap implementation found at: ", | |
1259 | (int) n->range); | |
4b84d9b8 JJ |
1260 | if (cur_stmt) |
1261 | print_gimple_stmt (dump_file, cur_stmt, 0); | |
1262 | else | |
1263 | { | |
4af78ef8 | 1264 | print_generic_expr (dump_file, tgt, TDF_NONE); |
4b84d9b8 JJ |
1265 | fprintf (dump_file, "\n"); |
1266 | } | |
dffec8eb JJ |
1267 | } |
1268 | ||
4b84d9b8 JJ |
1269 | if (cur_stmt) |
1270 | { | |
1271 | gsi_insert_after (&gsi, bswap_stmt, GSI_SAME_STMT); | |
1272 | gsi_remove (&gsi, true); | |
1273 | } | |
1274 | else | |
1275 | gsi_insert_before (&gsi, bswap_stmt, GSI_SAME_STMT); | |
1276 | return tgt; | |
dffec8eb JJ |
1277 | } |
1278 | ||
a7553ad6 JJ |
1279 | /* Try to optimize an assignment CUR_STMT with CONSTRUCTOR on the rhs |
1280 | using bswap optimizations. CDI_DOMINATORS need to be | |
1281 | computed on entry. Return true if it has been optimized and | |
1282 | TODO_update_ssa is needed. */ | |
1283 | ||
1284 | static bool | |
1285 | maybe_optimize_vector_constructor (gimple *cur_stmt) | |
1286 | { | |
1287 | tree fndecl = NULL_TREE, bswap_type = NULL_TREE, load_type; | |
1288 | struct symbolic_number n; | |
1289 | bool bswap; | |
1290 | ||
1291 | gcc_assert (is_gimple_assign (cur_stmt) | |
1292 | && gimple_assign_rhs_code (cur_stmt) == CONSTRUCTOR); | |
1293 | ||
1294 | tree rhs = gimple_assign_rhs1 (cur_stmt); | |
1295 | if (!VECTOR_TYPE_P (TREE_TYPE (rhs)) | |
1296 | || !INTEGRAL_TYPE_P (TREE_TYPE (TREE_TYPE (rhs))) | |
1297 | || gimple_assign_lhs (cur_stmt) == NULL_TREE) | |
1298 | return false; | |
1299 | ||
1300 | HOST_WIDE_INT sz = int_size_in_bytes (TREE_TYPE (rhs)) * BITS_PER_UNIT; | |
1301 | switch (sz) | |
1302 | { | |
1303 | case 16: | |
1304 | load_type = bswap_type = uint16_type_node; | |
1305 | break; | |
1306 | case 32: | |
1307 | if (builtin_decl_explicit_p (BUILT_IN_BSWAP32) | |
1308 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing) | |
1309 | { | |
1310 | load_type = uint32_type_node; | |
1311 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
1312 | bswap_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1313 | } | |
1314 | else | |
1315 | return false; | |
1316 | break; | |
1317 | case 64: | |
1318 | if (builtin_decl_explicit_p (BUILT_IN_BSWAP64) | |
1319 | && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing | |
1320 | || (word_mode == SImode | |
1321 | && builtin_decl_explicit_p (BUILT_IN_BSWAP32) | |
1322 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing))) | |
1323 | { | |
1324 | load_type = uint64_type_node; | |
1325 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
1326 | bswap_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1327 | } | |
1328 | else | |
1329 | return false; | |
1330 | break; | |
1331 | default: | |
1332 | return false; | |
1333 | } | |
1334 | ||
1335 | gimple *ins_stmt = find_bswap_or_nop (cur_stmt, &n, &bswap); | |
1336 | if (!ins_stmt || n.range != (unsigned HOST_WIDE_INT) sz) | |
1337 | return false; | |
1338 | ||
1339 | if (bswap && !fndecl && n.range != 16) | |
1340 | return false; | |
1341 | ||
1342 | memset (&nop_stats, 0, sizeof (nop_stats)); | |
1343 | memset (&bswap_stats, 0, sizeof (bswap_stats)); | |
1344 | return bswap_replace (gsi_for_stmt (cur_stmt), ins_stmt, fndecl, | |
1345 | bswap_type, load_type, &n, bswap) != NULL_TREE; | |
1346 | } | |
1347 | ||
dffec8eb JJ |
1348 | /* Find manual byte swap implementations as well as load in a given |
1349 | endianness. Byte swaps are turned into a bswap builtin invokation | |
1350 | while endian loads are converted to bswap builtin invokation or | |
1351 | simple load according to the target endianness. */ | |
1352 | ||
1353 | unsigned int | |
1354 | pass_optimize_bswap::execute (function *fun) | |
1355 | { | |
1356 | basic_block bb; | |
1357 | bool bswap32_p, bswap64_p; | |
1358 | bool changed = false; | |
1359 | tree bswap32_type = NULL_TREE, bswap64_type = NULL_TREE; | |
1360 | ||
1361 | bswap32_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP32) | |
1362 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing); | |
1363 | bswap64_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP64) | |
1364 | && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing | |
1365 | || (bswap32_p && word_mode == SImode))); | |
1366 | ||
1367 | /* Determine the argument type of the builtins. The code later on | |
1368 | assumes that the return and argument type are the same. */ | |
1369 | if (bswap32_p) | |
1370 | { | |
1371 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
1372 | bswap32_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1373 | } | |
1374 | ||
1375 | if (bswap64_p) | |
1376 | { | |
1377 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
1378 | bswap64_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1379 | } | |
1380 | ||
1381 | memset (&nop_stats, 0, sizeof (nop_stats)); | |
1382 | memset (&bswap_stats, 0, sizeof (bswap_stats)); | |
1383 | calculate_dominance_info (CDI_DOMINATORS); | |
1384 | ||
1385 | FOR_EACH_BB_FN (bb, fun) | |
1386 | { | |
1387 | gimple_stmt_iterator gsi; | |
1388 | ||
1389 | /* We do a reverse scan for bswap patterns to make sure we get the | |
1390 | widest match. As bswap pattern matching doesn't handle previously | |
1391 | inserted smaller bswap replacements as sub-patterns, the wider | |
1392 | variant wouldn't be detected. */ | |
1393 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi);) | |
1394 | { | |
1395 | gimple *ins_stmt, *cur_stmt = gsi_stmt (gsi); | |
1396 | tree fndecl = NULL_TREE, bswap_type = NULL_TREE, load_type; | |
1397 | enum tree_code code; | |
1398 | struct symbolic_number n; | |
1399 | bool bswap; | |
1400 | ||
1401 | /* This gsi_prev (&gsi) is not part of the for loop because cur_stmt | |
1402 | might be moved to a different basic block by bswap_replace and gsi | |
1403 | must not points to it if that's the case. Moving the gsi_prev | |
1404 | there make sure that gsi points to the statement previous to | |
1405 | cur_stmt while still making sure that all statements are | |
1406 | considered in this basic block. */ | |
1407 | gsi_prev (&gsi); | |
1408 | ||
1409 | if (!is_gimple_assign (cur_stmt)) | |
1410 | continue; | |
1411 | ||
1412 | code = gimple_assign_rhs_code (cur_stmt); | |
1413 | switch (code) | |
1414 | { | |
1415 | case LROTATE_EXPR: | |
1416 | case RROTATE_EXPR: | |
1417 | if (!tree_fits_uhwi_p (gimple_assign_rhs2 (cur_stmt)) | |
1418 | || tree_to_uhwi (gimple_assign_rhs2 (cur_stmt)) | |
1419 | % BITS_PER_UNIT) | |
1420 | continue; | |
1421 | /* Fall through. */ | |
1422 | case BIT_IOR_EXPR: | |
1423 | break; | |
cd676dfa JJ |
1424 | case CONSTRUCTOR: |
1425 | { | |
1426 | tree rhs = gimple_assign_rhs1 (cur_stmt); | |
1427 | if (VECTOR_TYPE_P (TREE_TYPE (rhs)) | |
1428 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_TYPE (rhs)))) | |
1429 | break; | |
1430 | } | |
1431 | continue; | |
dffec8eb JJ |
1432 | default: |
1433 | continue; | |
1434 | } | |
1435 | ||
1436 | ins_stmt = find_bswap_or_nop (cur_stmt, &n, &bswap); | |
1437 | ||
1438 | if (!ins_stmt) | |
1439 | continue; | |
1440 | ||
1441 | switch (n.range) | |
1442 | { | |
1443 | case 16: | |
1444 | /* Already in canonical form, nothing to do. */ | |
1445 | if (code == LROTATE_EXPR || code == RROTATE_EXPR) | |
1446 | continue; | |
1447 | load_type = bswap_type = uint16_type_node; | |
1448 | break; | |
1449 | case 32: | |
1450 | load_type = uint32_type_node; | |
1451 | if (bswap32_p) | |
1452 | { | |
1453 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
1454 | bswap_type = bswap32_type; | |
1455 | } | |
1456 | break; | |
1457 | case 64: | |
1458 | load_type = uint64_type_node; | |
1459 | if (bswap64_p) | |
1460 | { | |
1461 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
1462 | bswap_type = bswap64_type; | |
1463 | } | |
1464 | break; | |
1465 | default: | |
1466 | continue; | |
1467 | } | |
1468 | ||
1469 | if (bswap && !fndecl && n.range != 16) | |
1470 | continue; | |
1471 | ||
4b84d9b8 JJ |
1472 | if (bswap_replace (gsi_for_stmt (cur_stmt), ins_stmt, fndecl, |
1473 | bswap_type, load_type, &n, bswap)) | |
dffec8eb JJ |
1474 | changed = true; |
1475 | } | |
1476 | } | |
1477 | ||
1478 | statistics_counter_event (fun, "16-bit nop implementations found", | |
1479 | nop_stats.found_16bit); | |
1480 | statistics_counter_event (fun, "32-bit nop implementations found", | |
1481 | nop_stats.found_32bit); | |
1482 | statistics_counter_event (fun, "64-bit nop implementations found", | |
1483 | nop_stats.found_64bit); | |
1484 | statistics_counter_event (fun, "16-bit bswap implementations found", | |
1485 | bswap_stats.found_16bit); | |
1486 | statistics_counter_event (fun, "32-bit bswap implementations found", | |
1487 | bswap_stats.found_32bit); | |
1488 | statistics_counter_event (fun, "64-bit bswap implementations found", | |
1489 | bswap_stats.found_64bit); | |
1490 | ||
1491 | return (changed ? TODO_update_ssa : 0); | |
1492 | } | |
1493 | ||
1494 | } // anon namespace | |
1495 | ||
1496 | gimple_opt_pass * | |
1497 | make_pass_optimize_bswap (gcc::context *ctxt) | |
1498 | { | |
1499 | return new pass_optimize_bswap (ctxt); | |
1500 | } | |
1501 | ||
1502 | namespace { | |
1503 | ||
245f6de1 | 1504 | /* Struct recording one operand for the store, which is either a constant, |
c94c3532 EB |
1505 | then VAL represents the constant and all the other fields are zero, or |
1506 | a memory load, then VAL represents the reference, BASE_ADDR is non-NULL | |
1507 | and the other fields also reflect the memory load, or an SSA name, then | |
1508 | VAL represents the SSA name and all the other fields are zero, */ | |
245f6de1 | 1509 | |
6c1dae73 | 1510 | class store_operand_info |
245f6de1 | 1511 | { |
6c1dae73 | 1512 | public: |
245f6de1 JJ |
1513 | tree val; |
1514 | tree base_addr; | |
8a91d545 RS |
1515 | poly_uint64 bitsize; |
1516 | poly_uint64 bitpos; | |
1517 | poly_uint64 bitregion_start; | |
1518 | poly_uint64 bitregion_end; | |
245f6de1 | 1519 | gimple *stmt; |
383ac8dc | 1520 | bool bit_not_p; |
245f6de1 JJ |
1521 | store_operand_info (); |
1522 | }; | |
1523 | ||
1524 | store_operand_info::store_operand_info () | |
1525 | : val (NULL_TREE), base_addr (NULL_TREE), bitsize (0), bitpos (0), | |
383ac8dc | 1526 | bitregion_start (0), bitregion_end (0), stmt (NULL), bit_not_p (false) |
245f6de1 JJ |
1527 | { |
1528 | } | |
1529 | ||
f663d9ad KT |
1530 | /* Struct recording the information about a single store of an immediate |
1531 | to memory. These are created in the first phase and coalesced into | |
1532 | merged_store_group objects in the second phase. */ | |
1533 | ||
6c1dae73 | 1534 | class store_immediate_info |
f663d9ad | 1535 | { |
6c1dae73 | 1536 | public: |
f663d9ad KT |
1537 | unsigned HOST_WIDE_INT bitsize; |
1538 | unsigned HOST_WIDE_INT bitpos; | |
a62b3dc5 JJ |
1539 | unsigned HOST_WIDE_INT bitregion_start; |
1540 | /* This is one past the last bit of the bit region. */ | |
1541 | unsigned HOST_WIDE_INT bitregion_end; | |
f663d9ad KT |
1542 | gimple *stmt; |
1543 | unsigned int order; | |
e362a897 EB |
1544 | /* INTEGER_CST for constant store, STRING_CST for string store, |
1545 | MEM_REF for memory copy, BIT_*_EXPR for logical bitwise operation, | |
1546 | BIT_INSERT_EXPR for bit insertion. | |
4b84d9b8 JJ |
1547 | LROTATE_EXPR if it can be only bswap optimized and |
1548 | ops are not really meaningful. | |
1549 | NOP_EXPR if bswap optimization detected identity, ops | |
1550 | are not meaningful. */ | |
245f6de1 | 1551 | enum tree_code rhs_code; |
4b84d9b8 JJ |
1552 | /* Two fields for bswap optimization purposes. */ |
1553 | struct symbolic_number n; | |
1554 | gimple *ins_stmt; | |
127ef369 | 1555 | /* True if BIT_{AND,IOR,XOR}_EXPR result is inverted before storing. */ |
d60edaba | 1556 | bool bit_not_p; |
127ef369 JJ |
1557 | /* True if ops have been swapped and thus ops[1] represents |
1558 | rhs1 of BIT_{AND,IOR,XOR}_EXPR and ops[0] represents rhs2. */ | |
1559 | bool ops_swapped_p; | |
629387a6 EB |
1560 | /* The index number of the landing pad, or 0 if there is none. */ |
1561 | int lp_nr; | |
245f6de1 JJ |
1562 | /* Operands. For BIT_*_EXPR rhs_code both operands are used, otherwise |
1563 | just the first one. */ | |
1564 | store_operand_info ops[2]; | |
b5926e23 | 1565 | store_immediate_info (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, |
a62b3dc5 | 1566 | unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, |
4b84d9b8 | 1567 | gimple *, unsigned int, enum tree_code, |
629387a6 | 1568 | struct symbolic_number &, gimple *, bool, int, |
245f6de1 JJ |
1569 | const store_operand_info &, |
1570 | const store_operand_info &); | |
f663d9ad KT |
1571 | }; |
1572 | ||
1573 | store_immediate_info::store_immediate_info (unsigned HOST_WIDE_INT bs, | |
b5926e23 | 1574 | unsigned HOST_WIDE_INT bp, |
a62b3dc5 JJ |
1575 | unsigned HOST_WIDE_INT brs, |
1576 | unsigned HOST_WIDE_INT bre, | |
b5926e23 | 1577 | gimple *st, |
245f6de1 JJ |
1578 | unsigned int ord, |
1579 | enum tree_code rhscode, | |
4b84d9b8 JJ |
1580 | struct symbolic_number &nr, |
1581 | gimple *ins_stmtp, | |
d60edaba | 1582 | bool bitnotp, |
629387a6 | 1583 | int nr2, |
245f6de1 JJ |
1584 | const store_operand_info &op0r, |
1585 | const store_operand_info &op1r) | |
a62b3dc5 | 1586 | : bitsize (bs), bitpos (bp), bitregion_start (brs), bitregion_end (bre), |
4b84d9b8 | 1587 | stmt (st), order (ord), rhs_code (rhscode), n (nr), |
629387a6 EB |
1588 | ins_stmt (ins_stmtp), bit_not_p (bitnotp), ops_swapped_p (false), |
1589 | lp_nr (nr2) | |
245f6de1 JJ |
1590 | #if __cplusplus >= 201103L |
1591 | , ops { op0r, op1r } | |
1592 | { | |
1593 | } | |
1594 | #else | |
f663d9ad | 1595 | { |
245f6de1 JJ |
1596 | ops[0] = op0r; |
1597 | ops[1] = op1r; | |
f663d9ad | 1598 | } |
245f6de1 | 1599 | #endif |
f663d9ad KT |
1600 | |
1601 | /* Struct representing a group of stores to contiguous memory locations. | |
1602 | These are produced by the second phase (coalescing) and consumed in the | |
1603 | third phase that outputs the widened stores. */ | |
1604 | ||
6c1dae73 | 1605 | class merged_store_group |
f663d9ad | 1606 | { |
6c1dae73 | 1607 | public: |
f663d9ad KT |
1608 | unsigned HOST_WIDE_INT start; |
1609 | unsigned HOST_WIDE_INT width; | |
a62b3dc5 JJ |
1610 | unsigned HOST_WIDE_INT bitregion_start; |
1611 | unsigned HOST_WIDE_INT bitregion_end; | |
1612 | /* The size of the allocated memory for val and mask. */ | |
f663d9ad | 1613 | unsigned HOST_WIDE_INT buf_size; |
a62b3dc5 | 1614 | unsigned HOST_WIDE_INT align_base; |
8a91d545 | 1615 | poly_uint64 load_align_base[2]; |
f663d9ad KT |
1616 | |
1617 | unsigned int align; | |
245f6de1 | 1618 | unsigned int load_align[2]; |
f663d9ad KT |
1619 | unsigned int first_order; |
1620 | unsigned int last_order; | |
7f5a3982 | 1621 | bool bit_insertion; |
e362a897 | 1622 | bool string_concatenation; |
18e0c3d1 | 1623 | bool only_constants; |
1b3c9813 | 1624 | bool consecutive; |
18e0c3d1 | 1625 | unsigned int first_nonmergeable_order; |
629387a6 | 1626 | int lp_nr; |
f663d9ad | 1627 | |
a62b3dc5 | 1628 | auto_vec<store_immediate_info *> stores; |
f663d9ad KT |
1629 | /* We record the first and last original statements in the sequence because |
1630 | we'll need their vuse/vdef and replacement position. It's easier to keep | |
1631 | track of them separately as 'stores' is reordered by apply_stores. */ | |
1632 | gimple *last_stmt; | |
1633 | gimple *first_stmt; | |
1634 | unsigned char *val; | |
a62b3dc5 | 1635 | unsigned char *mask; |
f663d9ad KT |
1636 | |
1637 | merged_store_group (store_immediate_info *); | |
1638 | ~merged_store_group (); | |
7f5a3982 | 1639 | bool can_be_merged_into (store_immediate_info *); |
f663d9ad KT |
1640 | void merge_into (store_immediate_info *); |
1641 | void merge_overlapping (store_immediate_info *); | |
1642 | bool apply_stores (); | |
a62b3dc5 JJ |
1643 | private: |
1644 | void do_merge (store_immediate_info *); | |
f663d9ad KT |
1645 | }; |
1646 | ||
1647 | /* Debug helper. Dump LEN elements of byte array PTR to FD in hex. */ | |
1648 | ||
1649 | static void | |
1650 | dump_char_array (FILE *fd, unsigned char *ptr, unsigned int len) | |
1651 | { | |
1652 | if (!fd) | |
1653 | return; | |
1654 | ||
1655 | for (unsigned int i = 0; i < len; i++) | |
c94c3532 | 1656 | fprintf (fd, "%02x ", ptr[i]); |
f663d9ad KT |
1657 | fprintf (fd, "\n"); |
1658 | } | |
1659 | ||
f663d9ad KT |
1660 | /* Clear out LEN bits starting from bit START in the byte array |
1661 | PTR. This clears the bits to the *right* from START. | |
1662 | START must be within [0, BITS_PER_UNIT) and counts starting from | |
1663 | the least significant bit. */ | |
1664 | ||
1665 | static void | |
1666 | clear_bit_region_be (unsigned char *ptr, unsigned int start, | |
1667 | unsigned int len) | |
1668 | { | |
1669 | if (len == 0) | |
1670 | return; | |
1671 | /* Clear len bits to the right of start. */ | |
1672 | else if (len <= start + 1) | |
1673 | { | |
1674 | unsigned char mask = (~(~0U << len)); | |
1675 | mask = mask << (start + 1U - len); | |
1676 | ptr[0] &= ~mask; | |
1677 | } | |
1678 | else if (start != BITS_PER_UNIT - 1) | |
1679 | { | |
1680 | clear_bit_region_be (ptr, start, (start % BITS_PER_UNIT) + 1); | |
1681 | clear_bit_region_be (ptr + 1, BITS_PER_UNIT - 1, | |
1682 | len - (start % BITS_PER_UNIT) - 1); | |
1683 | } | |
1684 | else if (start == BITS_PER_UNIT - 1 | |
1685 | && len > BITS_PER_UNIT) | |
1686 | { | |
1687 | unsigned int nbytes = len / BITS_PER_UNIT; | |
a62b3dc5 | 1688 | memset (ptr, 0, nbytes); |
f663d9ad KT |
1689 | if (len % BITS_PER_UNIT != 0) |
1690 | clear_bit_region_be (ptr + nbytes, BITS_PER_UNIT - 1, | |
1691 | len % BITS_PER_UNIT); | |
1692 | } | |
1693 | else | |
1694 | gcc_unreachable (); | |
1695 | } | |
1696 | ||
1697 | /* In the byte array PTR clear the bit region starting at bit | |
1698 | START and is LEN bits wide. | |
1699 | For regions spanning multiple bytes do this recursively until we reach | |
1700 | zero LEN or a region contained within a single byte. */ | |
1701 | ||
1702 | static void | |
1703 | clear_bit_region (unsigned char *ptr, unsigned int start, | |
1704 | unsigned int len) | |
1705 | { | |
1706 | /* Degenerate base case. */ | |
1707 | if (len == 0) | |
1708 | return; | |
1709 | else if (start >= BITS_PER_UNIT) | |
1710 | clear_bit_region (ptr + 1, start - BITS_PER_UNIT, len); | |
1711 | /* Second base case. */ | |
1712 | else if ((start + len) <= BITS_PER_UNIT) | |
1713 | { | |
46a61395 | 1714 | unsigned char mask = (~0U) << (unsigned char) (BITS_PER_UNIT - len); |
f663d9ad KT |
1715 | mask >>= BITS_PER_UNIT - (start + len); |
1716 | ||
1717 | ptr[0] &= ~mask; | |
1718 | ||
1719 | return; | |
1720 | } | |
1721 | /* Clear most significant bits in a byte and proceed with the next byte. */ | |
1722 | else if (start != 0) | |
1723 | { | |
1724 | clear_bit_region (ptr, start, BITS_PER_UNIT - start); | |
1f069ef5 | 1725 | clear_bit_region (ptr + 1, 0, len - (BITS_PER_UNIT - start)); |
f663d9ad KT |
1726 | } |
1727 | /* Whole bytes need to be cleared. */ | |
1728 | else if (start == 0 && len > BITS_PER_UNIT) | |
1729 | { | |
1730 | unsigned int nbytes = len / BITS_PER_UNIT; | |
a848c710 KT |
1731 | /* We could recurse on each byte but we clear whole bytes, so a simple |
1732 | memset will do. */ | |
46a61395 | 1733 | memset (ptr, '\0', nbytes); |
f663d9ad KT |
1734 | /* Clear the remaining sub-byte region if there is one. */ |
1735 | if (len % BITS_PER_UNIT != 0) | |
1736 | clear_bit_region (ptr + nbytes, 0, len % BITS_PER_UNIT); | |
1737 | } | |
1738 | else | |
1739 | gcc_unreachable (); | |
1740 | } | |
1741 | ||
1742 | /* Write BITLEN bits of EXPR to the byte array PTR at | |
1743 | bit position BITPOS. PTR should contain TOTAL_BYTES elements. | |
1744 | Return true if the operation succeeded. */ | |
1745 | ||
1746 | static bool | |
1747 | encode_tree_to_bitpos (tree expr, unsigned char *ptr, int bitlen, int bitpos, | |
46a61395 | 1748 | unsigned int total_bytes) |
f663d9ad KT |
1749 | { |
1750 | unsigned int first_byte = bitpos / BITS_PER_UNIT; | |
ad1de652 JJ |
1751 | bool sub_byte_op_p = ((bitlen % BITS_PER_UNIT) |
1752 | || (bitpos % BITS_PER_UNIT) | |
f4b31647 | 1753 | || !int_mode_for_size (bitlen, 0).exists ()); |
3afd514b JJ |
1754 | bool empty_ctor_p |
1755 | = (TREE_CODE (expr) == CONSTRUCTOR | |
1756 | && CONSTRUCTOR_NELTS (expr) == 0 | |
1757 | && TYPE_SIZE_UNIT (TREE_TYPE (expr)) | |
1758 | && tree_fits_uhwi_p (TYPE_SIZE_UNIT (TREE_TYPE (expr)))); | |
f663d9ad KT |
1759 | |
1760 | if (!sub_byte_op_p) | |
3afd514b JJ |
1761 | { |
1762 | if (first_byte >= total_bytes) | |
1763 | return false; | |
1764 | total_bytes -= first_byte; | |
1765 | if (empty_ctor_p) | |
1766 | { | |
1767 | unsigned HOST_WIDE_INT rhs_bytes | |
1768 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (expr))); | |
1769 | if (rhs_bytes > total_bytes) | |
1770 | return false; | |
1771 | memset (ptr + first_byte, '\0', rhs_bytes); | |
1772 | return true; | |
1773 | } | |
1774 | return native_encode_expr (expr, ptr + first_byte, total_bytes) != 0; | |
1775 | } | |
f663d9ad KT |
1776 | |
1777 | /* LITTLE-ENDIAN | |
1778 | We are writing a non byte-sized quantity or at a position that is not | |
1779 | at a byte boundary. | |
1780 | |--------|--------|--------| ptr + first_byte | |
1781 | ^ ^ | |
1782 | xxx xxxxxxxx xxx< bp> | |
1783 | |______EXPR____| | |
1784 | ||
46a61395 | 1785 | First native_encode_expr EXPR into a temporary buffer and shift each |
f663d9ad KT |
1786 | byte in the buffer by 'bp' (carrying the bits over as necessary). |
1787 | |00000000|00xxxxxx|xxxxxxxx| << bp = |000xxxxx|xxxxxxxx|xxx00000| | |
1788 | <------bitlen---->< bp> | |
1789 | Then we clear the destination bits: | |
1790 | |---00000|00000000|000-----| ptr + first_byte | |
1791 | <-------bitlen--->< bp> | |
1792 | ||
1793 | Finally we ORR the bytes of the shifted EXPR into the cleared region: | |
1794 | |---xxxxx||xxxxxxxx||xxx-----| ptr + first_byte. | |
1795 | ||
1796 | BIG-ENDIAN | |
1797 | We are writing a non byte-sized quantity or at a position that is not | |
1798 | at a byte boundary. | |
1799 | ptr + first_byte |--------|--------|--------| | |
1800 | ^ ^ | |
1801 | <bp >xxx xxxxxxxx xxx | |
1802 | |_____EXPR_____| | |
1803 | ||
46a61395 | 1804 | First native_encode_expr EXPR into a temporary buffer and shift each |
f663d9ad KT |
1805 | byte in the buffer to the right by (carrying the bits over as necessary). |
1806 | We shift by as much as needed to align the most significant bit of EXPR | |
1807 | with bitpos: | |
1808 | |00xxxxxx|xxxxxxxx| >> 3 = |00000xxx|xxxxxxxx|xxxxx000| | |
1809 | <---bitlen----> <bp ><-----bitlen-----> | |
1810 | Then we clear the destination bits: | |
1811 | ptr + first_byte |-----000||00000000||00000---| | |
1812 | <bp ><-------bitlen-----> | |
1813 | ||
1814 | Finally we ORR the bytes of the shifted EXPR into the cleared region: | |
1815 | ptr + first_byte |---xxxxx||xxxxxxxx||xxx-----|. | |
1816 | The awkwardness comes from the fact that bitpos is counted from the | |
1817 | most significant bit of a byte. */ | |
1818 | ||
ef1d3b57 RS |
1819 | /* We must be dealing with fixed-size data at this point, since the |
1820 | total size is also fixed. */ | |
3afd514b JJ |
1821 | unsigned int byte_size; |
1822 | if (empty_ctor_p) | |
1823 | { | |
1824 | unsigned HOST_WIDE_INT rhs_bytes | |
1825 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (expr))); | |
1826 | if (rhs_bytes > total_bytes) | |
1827 | return false; | |
1828 | byte_size = rhs_bytes; | |
1829 | } | |
1830 | else | |
1831 | { | |
1832 | fixed_size_mode mode | |
1833 | = as_a <fixed_size_mode> (TYPE_MODE (TREE_TYPE (expr))); | |
e362a897 EB |
1834 | byte_size |
1835 | = mode == BLKmode | |
1836 | ? tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (expr))) | |
1837 | : GET_MODE_SIZE (mode); | |
3afd514b | 1838 | } |
f663d9ad | 1839 | /* Allocate an extra byte so that we have space to shift into. */ |
3afd514b | 1840 | byte_size++; |
f663d9ad | 1841 | unsigned char *tmpbuf = XALLOCAVEC (unsigned char, byte_size); |
46a61395 | 1842 | memset (tmpbuf, '\0', byte_size); |
f663d9ad | 1843 | /* The store detection code should only have allowed constants that are |
3afd514b JJ |
1844 | accepted by native_encode_expr or empty ctors. */ |
1845 | if (!empty_ctor_p | |
1846 | && native_encode_expr (expr, tmpbuf, byte_size - 1) == 0) | |
f663d9ad KT |
1847 | gcc_unreachable (); |
1848 | ||
1849 | /* The native_encode_expr machinery uses TYPE_MODE to determine how many | |
1850 | bytes to write. This means it can write more than | |
1851 | ROUND_UP (bitlen, BITS_PER_UNIT) / BITS_PER_UNIT bytes (for example | |
1852 | write 8 bytes for a bitlen of 40). Skip the bytes that are not within | |
1853 | bitlen and zero out the bits that are not relevant as well (that may | |
1854 | contain a sign bit due to sign-extension). */ | |
1855 | unsigned int padding | |
1856 | = byte_size - ROUND_UP (bitlen, BITS_PER_UNIT) / BITS_PER_UNIT - 1; | |
ad1de652 JJ |
1857 | /* On big-endian the padding is at the 'front' so just skip the initial |
1858 | bytes. */ | |
1859 | if (BYTES_BIG_ENDIAN) | |
1860 | tmpbuf += padding; | |
1861 | ||
1862 | byte_size -= padding; | |
1863 | ||
1864 | if (bitlen % BITS_PER_UNIT != 0) | |
f663d9ad | 1865 | { |
4b2c06f4 | 1866 | if (BYTES_BIG_ENDIAN) |
ad1de652 JJ |
1867 | clear_bit_region_be (tmpbuf, BITS_PER_UNIT - 1, |
1868 | BITS_PER_UNIT - (bitlen % BITS_PER_UNIT)); | |
1869 | else | |
1870 | clear_bit_region (tmpbuf, bitlen, | |
1871 | byte_size * BITS_PER_UNIT - bitlen); | |
f663d9ad | 1872 | } |
ad1de652 JJ |
1873 | /* Left shifting relies on the last byte being clear if bitlen is |
1874 | a multiple of BITS_PER_UNIT, which might not be clear if | |
1875 | there are padding bytes. */ | |
1876 | else if (!BYTES_BIG_ENDIAN) | |
1877 | tmpbuf[byte_size - 1] = '\0'; | |
f663d9ad KT |
1878 | |
1879 | /* Clear the bit region in PTR where the bits from TMPBUF will be | |
46a61395 | 1880 | inserted into. */ |
f663d9ad KT |
1881 | if (BYTES_BIG_ENDIAN) |
1882 | clear_bit_region_be (ptr + first_byte, | |
1883 | BITS_PER_UNIT - 1 - (bitpos % BITS_PER_UNIT), bitlen); | |
1884 | else | |
1885 | clear_bit_region (ptr + first_byte, bitpos % BITS_PER_UNIT, bitlen); | |
1886 | ||
1887 | int shift_amnt; | |
1888 | int bitlen_mod = bitlen % BITS_PER_UNIT; | |
1889 | int bitpos_mod = bitpos % BITS_PER_UNIT; | |
1890 | ||
1891 | bool skip_byte = false; | |
1892 | if (BYTES_BIG_ENDIAN) | |
1893 | { | |
1894 | /* BITPOS and BITLEN are exactly aligned and no shifting | |
1895 | is necessary. */ | |
1896 | if (bitpos_mod + bitlen_mod == BITS_PER_UNIT | |
1897 | || (bitpos_mod == 0 && bitlen_mod == 0)) | |
1898 | shift_amnt = 0; | |
1899 | /* |. . . . . . . .| | |
1900 | <bp > <blen >. | |
1901 | We always shift right for BYTES_BIG_ENDIAN so shift the beginning | |
1902 | of the value until it aligns with 'bp' in the next byte over. */ | |
1903 | else if (bitpos_mod + bitlen_mod < BITS_PER_UNIT) | |
1904 | { | |
1905 | shift_amnt = bitlen_mod + bitpos_mod; | |
1906 | skip_byte = bitlen_mod != 0; | |
1907 | } | |
1908 | /* |. . . . . . . .| | |
1909 | <----bp---> | |
1910 | <---blen---->. | |
1911 | Shift the value right within the same byte so it aligns with 'bp'. */ | |
1912 | else | |
1913 | shift_amnt = bitlen_mod + bitpos_mod - BITS_PER_UNIT; | |
1914 | } | |
1915 | else | |
1916 | shift_amnt = bitpos % BITS_PER_UNIT; | |
1917 | ||
1918 | /* Create the shifted version of EXPR. */ | |
1919 | if (!BYTES_BIG_ENDIAN) | |
46a61395 | 1920 | { |
8aba425f | 1921 | shift_bytes_in_array_left (tmpbuf, byte_size, shift_amnt); |
46a61395 JJ |
1922 | if (shift_amnt == 0) |
1923 | byte_size--; | |
1924 | } | |
f663d9ad KT |
1925 | else |
1926 | { | |
1927 | gcc_assert (BYTES_BIG_ENDIAN); | |
1928 | shift_bytes_in_array_right (tmpbuf, byte_size, shift_amnt); | |
1929 | /* If shifting right forced us to move into the next byte skip the now | |
1930 | empty byte. */ | |
1931 | if (skip_byte) | |
1932 | { | |
1933 | tmpbuf++; | |
1934 | byte_size--; | |
1935 | } | |
1936 | } | |
1937 | ||
1938 | /* Insert the bits from TMPBUF. */ | |
1939 | for (unsigned int i = 0; i < byte_size; i++) | |
1940 | ptr[first_byte + i] |= tmpbuf[i]; | |
1941 | ||
1942 | return true; | |
1943 | } | |
1944 | ||
1945 | /* Sorting function for store_immediate_info objects. | |
1946 | Sorts them by bitposition. */ | |
1947 | ||
1948 | static int | |
1949 | sort_by_bitpos (const void *x, const void *y) | |
1950 | { | |
1951 | store_immediate_info *const *tmp = (store_immediate_info * const *) x; | |
1952 | store_immediate_info *const *tmp2 = (store_immediate_info * const *) y; | |
1953 | ||
109cca3b | 1954 | if ((*tmp)->bitpos < (*tmp2)->bitpos) |
f663d9ad KT |
1955 | return -1; |
1956 | else if ((*tmp)->bitpos > (*tmp2)->bitpos) | |
1957 | return 1; | |
109cca3b | 1958 | else |
0f0027d1 KT |
1959 | /* If they are the same let's use the order which is guaranteed to |
1960 | be different. */ | |
1961 | return (*tmp)->order - (*tmp2)->order; | |
f663d9ad KT |
1962 | } |
1963 | ||
1964 | /* Sorting function for store_immediate_info objects. | |
1965 | Sorts them by the order field. */ | |
1966 | ||
1967 | static int | |
1968 | sort_by_order (const void *x, const void *y) | |
1969 | { | |
1970 | store_immediate_info *const *tmp = (store_immediate_info * const *) x; | |
1971 | store_immediate_info *const *tmp2 = (store_immediate_info * const *) y; | |
1972 | ||
1973 | if ((*tmp)->order < (*tmp2)->order) | |
1974 | return -1; | |
1975 | else if ((*tmp)->order > (*tmp2)->order) | |
1976 | return 1; | |
1977 | ||
1978 | gcc_unreachable (); | |
1979 | } | |
1980 | ||
1981 | /* Initialize a merged_store_group object from a store_immediate_info | |
1982 | object. */ | |
1983 | ||
1984 | merged_store_group::merged_store_group (store_immediate_info *info) | |
1985 | { | |
1986 | start = info->bitpos; | |
1987 | width = info->bitsize; | |
a62b3dc5 JJ |
1988 | bitregion_start = info->bitregion_start; |
1989 | bitregion_end = info->bitregion_end; | |
f663d9ad KT |
1990 | /* VAL has memory allocated for it in apply_stores once the group |
1991 | width has been finalized. */ | |
1992 | val = NULL; | |
a62b3dc5 | 1993 | mask = NULL; |
e362a897 EB |
1994 | bit_insertion = info->rhs_code == BIT_INSERT_EXPR; |
1995 | string_concatenation = info->rhs_code == STRING_CST; | |
18e0c3d1 | 1996 | only_constants = info->rhs_code == INTEGER_CST; |
1b3c9813 | 1997 | consecutive = true; |
18e0c3d1 | 1998 | first_nonmergeable_order = ~0U; |
629387a6 | 1999 | lp_nr = info->lp_nr; |
a62b3dc5 JJ |
2000 | unsigned HOST_WIDE_INT align_bitpos = 0; |
2001 | get_object_alignment_1 (gimple_assign_lhs (info->stmt), | |
2002 | &align, &align_bitpos); | |
2003 | align_base = start - align_bitpos; | |
245f6de1 JJ |
2004 | for (int i = 0; i < 2; ++i) |
2005 | { | |
2006 | store_operand_info &op = info->ops[i]; | |
2007 | if (op.base_addr == NULL_TREE) | |
2008 | { | |
2009 | load_align[i] = 0; | |
2010 | load_align_base[i] = 0; | |
2011 | } | |
2012 | else | |
2013 | { | |
2014 | get_object_alignment_1 (op.val, &load_align[i], &align_bitpos); | |
2015 | load_align_base[i] = op.bitpos - align_bitpos; | |
2016 | } | |
2017 | } | |
f663d9ad KT |
2018 | stores.create (1); |
2019 | stores.safe_push (info); | |
2020 | last_stmt = info->stmt; | |
2021 | last_order = info->order; | |
2022 | first_stmt = last_stmt; | |
2023 | first_order = last_order; | |
2024 | buf_size = 0; | |
2025 | } | |
2026 | ||
2027 | merged_store_group::~merged_store_group () | |
2028 | { | |
2029 | if (val) | |
2030 | XDELETEVEC (val); | |
2031 | } | |
2032 | ||
7f5a3982 EB |
2033 | /* Return true if the store described by INFO can be merged into the group. */ |
2034 | ||
2035 | bool | |
2036 | merged_store_group::can_be_merged_into (store_immediate_info *info) | |
2037 | { | |
2038 | /* Do not merge bswap patterns. */ | |
2039 | if (info->rhs_code == LROTATE_EXPR) | |
2040 | return false; | |
2041 | ||
629387a6 EB |
2042 | if (info->lp_nr != lp_nr) |
2043 | return false; | |
2044 | ||
7f5a3982 EB |
2045 | /* The canonical case. */ |
2046 | if (info->rhs_code == stores[0]->rhs_code) | |
2047 | return true; | |
2048 | ||
e362a897 | 2049 | /* BIT_INSERT_EXPR is compatible with INTEGER_CST if no STRING_CST. */ |
7f5a3982 | 2050 | if (info->rhs_code == BIT_INSERT_EXPR && stores[0]->rhs_code == INTEGER_CST) |
e362a897 | 2051 | return !string_concatenation; |
7f5a3982 EB |
2052 | |
2053 | if (stores[0]->rhs_code == BIT_INSERT_EXPR && info->rhs_code == INTEGER_CST) | |
e362a897 | 2054 | return !string_concatenation; |
7f5a3982 | 2055 | |
ed01d707 EB |
2056 | /* We can turn MEM_REF into BIT_INSERT_EXPR for bit-field stores, but do it |
2057 | only for small regions since this can generate a lot of instructions. */ | |
7f5a3982 EB |
2058 | if (info->rhs_code == MEM_REF |
2059 | && (stores[0]->rhs_code == INTEGER_CST | |
2060 | || stores[0]->rhs_code == BIT_INSERT_EXPR) | |
2061 | && info->bitregion_start == stores[0]->bitregion_start | |
ed01d707 | 2062 | && info->bitregion_end == stores[0]->bitregion_end |
2815558a | 2063 | && info->bitregion_end - info->bitregion_start <= MAX_FIXED_MODE_SIZE) |
e362a897 | 2064 | return !string_concatenation; |
7f5a3982 EB |
2065 | |
2066 | if (stores[0]->rhs_code == MEM_REF | |
2067 | && (info->rhs_code == INTEGER_CST | |
2068 | || info->rhs_code == BIT_INSERT_EXPR) | |
2069 | && info->bitregion_start == stores[0]->bitregion_start | |
ed01d707 | 2070 | && info->bitregion_end == stores[0]->bitregion_end |
2815558a | 2071 | && info->bitregion_end - info->bitregion_start <= MAX_FIXED_MODE_SIZE) |
e362a897 EB |
2072 | return !string_concatenation; |
2073 | ||
2074 | /* STRING_CST is compatible with INTEGER_CST if no BIT_INSERT_EXPR. */ | |
2075 | if (info->rhs_code == STRING_CST | |
2076 | && stores[0]->rhs_code == INTEGER_CST | |
2077 | && stores[0]->bitsize == CHAR_BIT) | |
2078 | return !bit_insertion; | |
2079 | ||
2080 | if (stores[0]->rhs_code == STRING_CST | |
2081 | && info->rhs_code == INTEGER_CST | |
2082 | && info->bitsize == CHAR_BIT) | |
2083 | return !bit_insertion; | |
7f5a3982 EB |
2084 | |
2085 | return false; | |
2086 | } | |
2087 | ||
a62b3dc5 JJ |
2088 | /* Helper method for merge_into and merge_overlapping to do |
2089 | the common part. */ | |
7f5a3982 | 2090 | |
f663d9ad | 2091 | void |
a62b3dc5 | 2092 | merged_store_group::do_merge (store_immediate_info *info) |
f663d9ad | 2093 | { |
a62b3dc5 JJ |
2094 | bitregion_start = MIN (bitregion_start, info->bitregion_start); |
2095 | bitregion_end = MAX (bitregion_end, info->bitregion_end); | |
2096 | ||
2097 | unsigned int this_align; | |
2098 | unsigned HOST_WIDE_INT align_bitpos = 0; | |
2099 | get_object_alignment_1 (gimple_assign_lhs (info->stmt), | |
2100 | &this_align, &align_bitpos); | |
2101 | if (this_align > align) | |
2102 | { | |
2103 | align = this_align; | |
2104 | align_base = info->bitpos - align_bitpos; | |
2105 | } | |
245f6de1 JJ |
2106 | for (int i = 0; i < 2; ++i) |
2107 | { | |
2108 | store_operand_info &op = info->ops[i]; | |
2109 | if (!op.base_addr) | |
2110 | continue; | |
2111 | ||
2112 | get_object_alignment_1 (op.val, &this_align, &align_bitpos); | |
2113 | if (this_align > load_align[i]) | |
2114 | { | |
2115 | load_align[i] = this_align; | |
2116 | load_align_base[i] = op.bitpos - align_bitpos; | |
2117 | } | |
2118 | } | |
f663d9ad | 2119 | |
f663d9ad KT |
2120 | gimple *stmt = info->stmt; |
2121 | stores.safe_push (info); | |
2122 | if (info->order > last_order) | |
2123 | { | |
2124 | last_order = info->order; | |
2125 | last_stmt = stmt; | |
2126 | } | |
2127 | else if (info->order < first_order) | |
2128 | { | |
2129 | first_order = info->order; | |
2130 | first_stmt = stmt; | |
2131 | } | |
e362a897 | 2132 | |
1b3c9813 EB |
2133 | if (info->bitpos != start + width) |
2134 | consecutive = false; | |
2135 | ||
e362a897 EB |
2136 | /* We need to use extraction if there is any bit-field. */ |
2137 | if (info->rhs_code == BIT_INSERT_EXPR) | |
2138 | { | |
2139 | bit_insertion = true; | |
2140 | gcc_assert (!string_concatenation); | |
2141 | } | |
2142 | ||
1b3c9813 | 2143 | /* We want to use concatenation if there is any string. */ |
e362a897 EB |
2144 | if (info->rhs_code == STRING_CST) |
2145 | { | |
2146 | string_concatenation = true; | |
2147 | gcc_assert (!bit_insertion); | |
2148 | } | |
2149 | ||
1b3c9813 EB |
2150 | /* But we cannot use it if we don't have consecutive stores. */ |
2151 | if (!consecutive) | |
2152 | string_concatenation = false; | |
2153 | ||
18e0c3d1 JJ |
2154 | if (info->rhs_code != INTEGER_CST) |
2155 | only_constants = false; | |
f663d9ad KT |
2156 | } |
2157 | ||
a62b3dc5 JJ |
2158 | /* Merge a store recorded by INFO into this merged store. |
2159 | The store is not overlapping with the existing recorded | |
2160 | stores. */ | |
2161 | ||
2162 | void | |
2163 | merged_store_group::merge_into (store_immediate_info *info) | |
2164 | { | |
1b3c9813 EB |
2165 | do_merge (info); |
2166 | ||
a62b3dc5 JJ |
2167 | /* Make sure we're inserting in the position we think we're inserting. */ |
2168 | gcc_assert (info->bitpos >= start + width | |
2169 | && info->bitregion_start <= bitregion_end); | |
2170 | ||
c5679c37 | 2171 | width = info->bitpos + info->bitsize - start; |
a62b3dc5 JJ |
2172 | } |
2173 | ||
f663d9ad KT |
2174 | /* Merge a store described by INFO into this merged store. |
2175 | INFO overlaps in some way with the current store (i.e. it's not contiguous | |
2176 | which is handled by merged_store_group::merge_into). */ | |
2177 | ||
2178 | void | |
2179 | merged_store_group::merge_overlapping (store_immediate_info *info) | |
2180 | { | |
1b3c9813 EB |
2181 | do_merge (info); |
2182 | ||
f663d9ad | 2183 | /* If the store extends the size of the group, extend the width. */ |
a62b3dc5 | 2184 | if (info->bitpos + info->bitsize > start + width) |
c5679c37 | 2185 | width = info->bitpos + info->bitsize - start; |
f663d9ad KT |
2186 | } |
2187 | ||
2188 | /* Go through all the recorded stores in this group in program order and | |
2189 | apply their values to the VAL byte array to create the final merged | |
2190 | value. Return true if the operation succeeded. */ | |
2191 | ||
2192 | bool | |
2193 | merged_store_group::apply_stores () | |
2194 | { | |
e362a897 EB |
2195 | store_immediate_info *info; |
2196 | unsigned int i; | |
2197 | ||
a62b3dc5 JJ |
2198 | /* Make sure we have more than one store in the group, otherwise we cannot |
2199 | merge anything. */ | |
2200 | if (bitregion_start % BITS_PER_UNIT != 0 | |
2201 | || bitregion_end % BITS_PER_UNIT != 0 | |
f663d9ad KT |
2202 | || stores.length () == 1) |
2203 | return false; | |
2204 | ||
e362a897 EB |
2205 | buf_size = (bitregion_end - bitregion_start) / BITS_PER_UNIT; |
2206 | ||
2207 | /* Really do string concatenation for large strings only. */ | |
2208 | if (buf_size <= MOVE_MAX) | |
2209 | string_concatenation = false; | |
2210 | ||
c94c3532 | 2211 | /* Create a power-of-2-sized buffer for native_encode_expr. */ |
e362a897 EB |
2212 | if (!string_concatenation) |
2213 | buf_size = 1 << ceil_log2 (buf_size); | |
2214 | ||
a62b3dc5 JJ |
2215 | val = XNEWVEC (unsigned char, 2 * buf_size); |
2216 | mask = val + buf_size; | |
2217 | memset (val, 0, buf_size); | |
2218 | memset (mask, ~0U, buf_size); | |
f663d9ad | 2219 | |
e362a897 EB |
2220 | stores.qsort (sort_by_order); |
2221 | ||
f663d9ad KT |
2222 | FOR_EACH_VEC_ELT (stores, i, info) |
2223 | { | |
a62b3dc5 | 2224 | unsigned int pos_in_buffer = info->bitpos - bitregion_start; |
c94c3532 | 2225 | tree cst; |
245f6de1 JJ |
2226 | if (info->ops[0].val && info->ops[0].base_addr == NULL_TREE) |
2227 | cst = info->ops[0].val; | |
2228 | else if (info->ops[1].val && info->ops[1].base_addr == NULL_TREE) | |
2229 | cst = info->ops[1].val; | |
c94c3532 EB |
2230 | else |
2231 | cst = NULL_TREE; | |
245f6de1 | 2232 | bool ret = true; |
e362a897 EB |
2233 | if (cst && info->rhs_code != BIT_INSERT_EXPR) |
2234 | ret = encode_tree_to_bitpos (cst, val, info->bitsize, pos_in_buffer, | |
2235 | buf_size); | |
c94c3532 EB |
2236 | unsigned char *m = mask + (pos_in_buffer / BITS_PER_UNIT); |
2237 | if (BYTES_BIG_ENDIAN) | |
2238 | clear_bit_region_be (m, (BITS_PER_UNIT - 1 | |
2239 | - (pos_in_buffer % BITS_PER_UNIT)), | |
2240 | info->bitsize); | |
2241 | else | |
2242 | clear_bit_region (m, pos_in_buffer % BITS_PER_UNIT, info->bitsize); | |
245f6de1 | 2243 | if (cst && dump_file && (dump_flags & TDF_DETAILS)) |
f663d9ad KT |
2244 | { |
2245 | if (ret) | |
2246 | { | |
c94c3532 | 2247 | fputs ("After writing ", dump_file); |
4af78ef8 | 2248 | print_generic_expr (dump_file, cst, TDF_NONE); |
f663d9ad | 2249 | fprintf (dump_file, " of size " HOST_WIDE_INT_PRINT_DEC |
c94c3532 EB |
2250 | " at position %d\n", info->bitsize, pos_in_buffer); |
2251 | fputs (" the merged value contains ", dump_file); | |
f663d9ad | 2252 | dump_char_array (dump_file, val, buf_size); |
c94c3532 EB |
2253 | fputs (" the merged mask contains ", dump_file); |
2254 | dump_char_array (dump_file, mask, buf_size); | |
2255 | if (bit_insertion) | |
2256 | fputs (" bit insertion is required\n", dump_file); | |
e362a897 EB |
2257 | if (string_concatenation) |
2258 | fputs (" string concatenation is required\n", dump_file); | |
f663d9ad KT |
2259 | } |
2260 | else | |
2261 | fprintf (dump_file, "Failed to merge stores\n"); | |
4b84d9b8 | 2262 | } |
f663d9ad KT |
2263 | if (!ret) |
2264 | return false; | |
2265 | } | |
4b84d9b8 | 2266 | stores.qsort (sort_by_bitpos); |
f663d9ad KT |
2267 | return true; |
2268 | } | |
2269 | ||
2270 | /* Structure describing the store chain. */ | |
2271 | ||
6c1dae73 | 2272 | class imm_store_chain_info |
f663d9ad | 2273 | { |
6c1dae73 | 2274 | public: |
50b6d676 AO |
2275 | /* Doubly-linked list that imposes an order on chain processing. |
2276 | PNXP (prev's next pointer) points to the head of a list, or to | |
2277 | the next field in the previous chain in the list. | |
2278 | See pass_store_merging::m_stores_head for more rationale. */ | |
2279 | imm_store_chain_info *next, **pnxp; | |
b5926e23 | 2280 | tree base_addr; |
a62b3dc5 | 2281 | auto_vec<store_immediate_info *> m_store_info; |
f663d9ad KT |
2282 | auto_vec<merged_store_group *> m_merged_store_groups; |
2283 | ||
50b6d676 AO |
2284 | imm_store_chain_info (imm_store_chain_info *&inspt, tree b_a) |
2285 | : next (inspt), pnxp (&inspt), base_addr (b_a) | |
2286 | { | |
2287 | inspt = this; | |
2288 | if (next) | |
2289 | { | |
2290 | gcc_checking_assert (pnxp == next->pnxp); | |
2291 | next->pnxp = &next; | |
2292 | } | |
2293 | } | |
2294 | ~imm_store_chain_info () | |
2295 | { | |
2296 | *pnxp = next; | |
2297 | if (next) | |
2298 | { | |
2299 | gcc_checking_assert (&next == next->pnxp); | |
2300 | next->pnxp = pnxp; | |
2301 | } | |
2302 | } | |
b5926e23 | 2303 | bool terminate_and_process_chain (); |
bd909071 JJ |
2304 | bool try_coalesce_bswap (merged_store_group *, unsigned int, unsigned int, |
2305 | unsigned int); | |
f663d9ad | 2306 | bool coalesce_immediate_stores (); |
b5926e23 RB |
2307 | bool output_merged_store (merged_store_group *); |
2308 | bool output_merged_stores (); | |
f663d9ad KT |
2309 | }; |
2310 | ||
2311 | const pass_data pass_data_tree_store_merging = { | |
2312 | GIMPLE_PASS, /* type */ | |
2313 | "store-merging", /* name */ | |
2314 | OPTGROUP_NONE, /* optinfo_flags */ | |
2315 | TV_GIMPLE_STORE_MERGING, /* tv_id */ | |
2316 | PROP_ssa, /* properties_required */ | |
2317 | 0, /* properties_provided */ | |
2318 | 0, /* properties_destroyed */ | |
2319 | 0, /* todo_flags_start */ | |
2320 | TODO_update_ssa, /* todo_flags_finish */ | |
2321 | }; | |
2322 | ||
2323 | class pass_store_merging : public gimple_opt_pass | |
2324 | { | |
2325 | public: | |
2326 | pass_store_merging (gcc::context *ctxt) | |
faec5f24 | 2327 | : gimple_opt_pass (pass_data_tree_store_merging, ctxt), m_stores_head () |
f663d9ad KT |
2328 | { |
2329 | } | |
2330 | ||
c94c3532 EB |
2331 | /* Pass not supported for PDP-endian, nor for insane hosts or |
2332 | target character sizes where native_{encode,interpret}_expr | |
a62b3dc5 | 2333 | doesn't work properly. */ |
f663d9ad KT |
2334 | virtual bool |
2335 | gate (function *) | |
2336 | { | |
a62b3dc5 | 2337 | return flag_store_merging |
c94c3532 | 2338 | && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN |
a62b3dc5 JJ |
2339 | && CHAR_BIT == 8 |
2340 | && BITS_PER_UNIT == 8; | |
f663d9ad KT |
2341 | } |
2342 | ||
2343 | virtual unsigned int execute (function *); | |
2344 | ||
2345 | private: | |
99b1c316 | 2346 | hash_map<tree_operand_hash, class imm_store_chain_info *> m_stores; |
f663d9ad | 2347 | |
50b6d676 AO |
2348 | /* Form a doubly-linked stack of the elements of m_stores, so that |
2349 | we can iterate over them in a predictable way. Using this order | |
2350 | avoids extraneous differences in the compiler output just because | |
2351 | of tree pointer variations (e.g. different chains end up in | |
2352 | different positions of m_stores, so they are handled in different | |
2353 | orders, so they allocate or release SSA names in different | |
2354 | orders, and when they get reused, subsequent passes end up | |
2355 | getting different SSA names, which may ultimately change | |
2356 | decisions when going out of SSA). */ | |
2357 | imm_store_chain_info *m_stores_head; | |
2358 | ||
629387a6 EB |
2359 | bool process_store (gimple *); |
2360 | bool terminate_and_process_chain (imm_store_chain_info *); | |
383ac8dc | 2361 | bool terminate_all_aliasing_chains (imm_store_chain_info **, gimple *); |
629387a6 | 2362 | bool terminate_and_process_all_chains (); |
f663d9ad KT |
2363 | }; // class pass_store_merging |
2364 | ||
2365 | /* Terminate and process all recorded chains. Return true if any changes | |
2366 | were made. */ | |
2367 | ||
2368 | bool | |
2369 | pass_store_merging::terminate_and_process_all_chains () | |
2370 | { | |
f663d9ad | 2371 | bool ret = false; |
50b6d676 | 2372 | while (m_stores_head) |
629387a6 | 2373 | ret |= terminate_and_process_chain (m_stores_head); |
b119c055 | 2374 | gcc_assert (m_stores.is_empty ()); |
f663d9ad KT |
2375 | return ret; |
2376 | } | |
2377 | ||
383ac8dc JJ |
2378 | /* Terminate all chains that are affected by the statement STMT. |
2379 | CHAIN_INFO is the chain we should ignore from the checks if | |
629387a6 | 2380 | non-NULL. Return true if any changes were made. */ |
f663d9ad KT |
2381 | |
2382 | bool | |
20770eb8 | 2383 | pass_store_merging::terminate_all_aliasing_chains (imm_store_chain_info |
b5926e23 | 2384 | **chain_info, |
f663d9ad KT |
2385 | gimple *stmt) |
2386 | { | |
2387 | bool ret = false; | |
2388 | ||
2389 | /* If the statement doesn't touch memory it can't alias. */ | |
2390 | if (!gimple_vuse (stmt)) | |
2391 | return false; | |
2392 | ||
9e875fd8 | 2393 | tree store_lhs = gimple_store_p (stmt) ? gimple_get_lhs (stmt) : NULL_TREE; |
6b412bf6 RB |
2394 | ao_ref store_lhs_ref; |
2395 | ao_ref_init (&store_lhs_ref, store_lhs); | |
383ac8dc | 2396 | for (imm_store_chain_info *next = m_stores_head, *cur = next; cur; cur = next) |
f663d9ad | 2397 | { |
383ac8dc JJ |
2398 | next = cur->next; |
2399 | ||
2400 | /* We already checked all the stores in chain_info and terminated the | |
2401 | chain if necessary. Skip it here. */ | |
2402 | if (chain_info && *chain_info == cur) | |
2403 | continue; | |
2404 | ||
245f6de1 JJ |
2405 | store_immediate_info *info; |
2406 | unsigned int i; | |
383ac8dc | 2407 | FOR_EACH_VEC_ELT (cur->m_store_info, i, info) |
f663d9ad | 2408 | { |
9e875fd8 | 2409 | tree lhs = gimple_assign_lhs (info->stmt); |
6b412bf6 RB |
2410 | ao_ref lhs_ref; |
2411 | ao_ref_init (&lhs_ref, lhs); | |
2412 | if (ref_maybe_used_by_stmt_p (stmt, &lhs_ref) | |
2413 | || stmt_may_clobber_ref_p_1 (stmt, &lhs_ref) | |
2414 | || (store_lhs && refs_may_alias_p_1 (&store_lhs_ref, | |
2415 | &lhs_ref, false))) | |
f663d9ad | 2416 | { |
245f6de1 | 2417 | if (dump_file && (dump_flags & TDF_DETAILS)) |
f663d9ad | 2418 | { |
245f6de1 JJ |
2419 | fprintf (dump_file, "stmt causes chain termination:\n"); |
2420 | print_gimple_stmt (dump_file, stmt, 0); | |
f663d9ad | 2421 | } |
629387a6 | 2422 | ret |= terminate_and_process_chain (cur); |
245f6de1 | 2423 | break; |
f663d9ad KT |
2424 | } |
2425 | } | |
2426 | } | |
2427 | ||
f663d9ad KT |
2428 | return ret; |
2429 | } | |
2430 | ||
2431 | /* Helper function. Terminate the recorded chain storing to base object | |
2432 | BASE. Return true if the merging and output was successful. The m_stores | |
2433 | entry is removed after the processing in any case. */ | |
2434 | ||
2435 | bool | |
629387a6 | 2436 | pass_store_merging::terminate_and_process_chain (imm_store_chain_info *chain_info) |
f663d9ad | 2437 | { |
b5926e23 RB |
2438 | bool ret = chain_info->terminate_and_process_chain (); |
2439 | m_stores.remove (chain_info->base_addr); | |
2440 | delete chain_info; | |
f663d9ad KT |
2441 | return ret; |
2442 | } | |
2443 | ||
245f6de1 | 2444 | /* Return true if stmts in between FIRST (inclusive) and LAST (exclusive) |
629387a6 EB |
2445 | may clobber REF. FIRST and LAST must have non-NULL vdef. We want to |
2446 | be able to sink load of REF across stores between FIRST and LAST, up | |
2447 | to right before LAST. */ | |
245f6de1 JJ |
2448 | |
2449 | bool | |
2450 | stmts_may_clobber_ref_p (gimple *first, gimple *last, tree ref) | |
2451 | { | |
2452 | ao_ref r; | |
2453 | ao_ref_init (&r, ref); | |
2454 | unsigned int count = 0; | |
2455 | tree vop = gimple_vdef (last); | |
2456 | gimple *stmt; | |
2457 | ||
629387a6 EB |
2458 | /* Return true conservatively if the basic blocks are different. */ |
2459 | if (gimple_bb (first) != gimple_bb (last)) | |
2460 | return true; | |
2461 | ||
245f6de1 JJ |
2462 | do |
2463 | { | |
2464 | stmt = SSA_NAME_DEF_STMT (vop); | |
2465 | if (stmt_may_clobber_ref_p_1 (stmt, &r)) | |
2466 | return true; | |
4b84d9b8 JJ |
2467 | if (gimple_store_p (stmt) |
2468 | && refs_anti_dependent_p (ref, gimple_get_lhs (stmt))) | |
2469 | return true; | |
245f6de1 JJ |
2470 | /* Avoid quadratic compile time by bounding the number of checks |
2471 | we perform. */ | |
2472 | if (++count > MAX_STORE_ALIAS_CHECKS) | |
2473 | return true; | |
2474 | vop = gimple_vuse (stmt); | |
2475 | } | |
2476 | while (stmt != first); | |
629387a6 | 2477 | |
245f6de1 JJ |
2478 | return false; |
2479 | } | |
2480 | ||
2481 | /* Return true if INFO->ops[IDX] is mergeable with the | |
2482 | corresponding loads already in MERGED_STORE group. | |
2483 | BASE_ADDR is the base address of the whole store group. */ | |
2484 | ||
2485 | bool | |
2486 | compatible_load_p (merged_store_group *merged_store, | |
2487 | store_immediate_info *info, | |
2488 | tree base_addr, int idx) | |
2489 | { | |
2490 | store_immediate_info *infof = merged_store->stores[0]; | |
2491 | if (!info->ops[idx].base_addr | |
8a91d545 RS |
2492 | || maybe_ne (info->ops[idx].bitpos - infof->ops[idx].bitpos, |
2493 | info->bitpos - infof->bitpos) | |
245f6de1 JJ |
2494 | || !operand_equal_p (info->ops[idx].base_addr, |
2495 | infof->ops[idx].base_addr, 0)) | |
2496 | return false; | |
2497 | ||
2498 | store_immediate_info *infol = merged_store->stores.last (); | |
2499 | tree load_vuse = gimple_vuse (info->ops[idx].stmt); | |
2500 | /* In this case all vuses should be the same, e.g. | |
2501 | _1 = s.a; _2 = s.b; _3 = _1 | 1; t.a = _3; _4 = _2 | 2; t.b = _4; | |
2502 | or | |
2503 | _1 = s.a; _2 = s.b; t.a = _1; t.b = _2; | |
2504 | and we can emit the coalesced load next to any of those loads. */ | |
2505 | if (gimple_vuse (infof->ops[idx].stmt) == load_vuse | |
2506 | && gimple_vuse (infol->ops[idx].stmt) == load_vuse) | |
2507 | return true; | |
2508 | ||
2509 | /* Otherwise, at least for now require that the load has the same | |
2510 | vuse as the store. See following examples. */ | |
2511 | if (gimple_vuse (info->stmt) != load_vuse) | |
2512 | return false; | |
2513 | ||
2514 | if (gimple_vuse (infof->stmt) != gimple_vuse (infof->ops[idx].stmt) | |
2515 | || (infof != infol | |
2516 | && gimple_vuse (infol->stmt) != gimple_vuse (infol->ops[idx].stmt))) | |
2517 | return false; | |
2518 | ||
2519 | /* If the load is from the same location as the store, already | |
2520 | the construction of the immediate chain info guarantees no intervening | |
2521 | stores, so no further checks are needed. Example: | |
2522 | _1 = s.a; _2 = _1 & -7; s.a = _2; _3 = s.b; _4 = _3 & -7; s.b = _4; */ | |
8a91d545 | 2523 | if (known_eq (info->ops[idx].bitpos, info->bitpos) |
245f6de1 JJ |
2524 | && operand_equal_p (info->ops[idx].base_addr, base_addr, 0)) |
2525 | return true; | |
2526 | ||
2527 | /* Otherwise, we need to punt if any of the loads can be clobbered by any | |
2528 | of the stores in the group, or any other stores in between those. | |
2529 | Previous calls to compatible_load_p ensured that for all the | |
2530 | merged_store->stores IDX loads, no stmts starting with | |
2531 | merged_store->first_stmt and ending right before merged_store->last_stmt | |
2532 | clobbers those loads. */ | |
2533 | gimple *first = merged_store->first_stmt; | |
2534 | gimple *last = merged_store->last_stmt; | |
2535 | unsigned int i; | |
2536 | store_immediate_info *infoc; | |
2537 | /* The stores are sorted by increasing store bitpos, so if info->stmt store | |
2538 | comes before the so far first load, we'll be changing | |
2539 | merged_store->first_stmt. In that case we need to give up if | |
2540 | any of the earlier processed loads clobber with the stmts in the new | |
2541 | range. */ | |
2542 | if (info->order < merged_store->first_order) | |
2543 | { | |
2544 | FOR_EACH_VEC_ELT (merged_store->stores, i, infoc) | |
2545 | if (stmts_may_clobber_ref_p (info->stmt, first, infoc->ops[idx].val)) | |
2546 | return false; | |
2547 | first = info->stmt; | |
2548 | } | |
2549 | /* Similarly, we could change merged_store->last_stmt, so ensure | |
2550 | in that case no stmts in the new range clobber any of the earlier | |
2551 | processed loads. */ | |
2552 | else if (info->order > merged_store->last_order) | |
2553 | { | |
2554 | FOR_EACH_VEC_ELT (merged_store->stores, i, infoc) | |
2555 | if (stmts_may_clobber_ref_p (last, info->stmt, infoc->ops[idx].val)) | |
2556 | return false; | |
2557 | last = info->stmt; | |
2558 | } | |
2559 | /* And finally, we'd be adding a new load to the set, ensure it isn't | |
2560 | clobbered in the new range. */ | |
2561 | if (stmts_may_clobber_ref_p (first, last, info->ops[idx].val)) | |
2562 | return false; | |
2563 | ||
2564 | /* Otherwise, we are looking for: | |
2565 | _1 = s.a; _2 = _1 ^ 15; t.a = _2; _3 = s.b; _4 = _3 ^ 15; t.b = _4; | |
2566 | or | |
2567 | _1 = s.a; t.a = _1; _2 = s.b; t.b = _2; */ | |
2568 | return true; | |
2569 | } | |
2570 | ||
4b84d9b8 JJ |
2571 | /* Add all refs loaded to compute VAL to REFS vector. */ |
2572 | ||
2573 | void | |
2574 | gather_bswap_load_refs (vec<tree> *refs, tree val) | |
2575 | { | |
2576 | if (TREE_CODE (val) != SSA_NAME) | |
2577 | return; | |
2578 | ||
2579 | gimple *stmt = SSA_NAME_DEF_STMT (val); | |
2580 | if (!is_gimple_assign (stmt)) | |
2581 | return; | |
2582 | ||
2583 | if (gimple_assign_load_p (stmt)) | |
2584 | { | |
2585 | refs->safe_push (gimple_assign_rhs1 (stmt)); | |
2586 | return; | |
2587 | } | |
2588 | ||
2589 | switch (gimple_assign_rhs_class (stmt)) | |
2590 | { | |
2591 | case GIMPLE_BINARY_RHS: | |
2592 | gather_bswap_load_refs (refs, gimple_assign_rhs2 (stmt)); | |
2593 | /* FALLTHRU */ | |
2594 | case GIMPLE_UNARY_RHS: | |
2595 | gather_bswap_load_refs (refs, gimple_assign_rhs1 (stmt)); | |
2596 | break; | |
2597 | default: | |
2598 | gcc_unreachable (); | |
2599 | } | |
2600 | } | |
2601 | ||
c5679c37 JJ |
2602 | /* Check if there are any stores in M_STORE_INFO after index I |
2603 | (where M_STORE_INFO must be sorted by sort_by_bitpos) that overlap | |
2604 | a potential group ending with END that have their order | |
4d213bf6 JJ |
2605 | smaller than LAST_ORDER. ALL_INTEGER_CST_P is true if |
2606 | all the stores already merged and the one under consideration | |
2607 | have rhs_code of INTEGER_CST. Return true if there are no such stores. | |
c5679c37 JJ |
2608 | Consider: |
2609 | MEM[(long long int *)p_28] = 0; | |
2610 | MEM[(long long int *)p_28 + 8B] = 0; | |
2611 | MEM[(long long int *)p_28 + 16B] = 0; | |
2612 | MEM[(long long int *)p_28 + 24B] = 0; | |
2613 | _129 = (int) _130; | |
2614 | MEM[(int *)p_28 + 8B] = _129; | |
2615 | MEM[(int *)p_28].a = -1; | |
2616 | We already have | |
2617 | MEM[(long long int *)p_28] = 0; | |
2618 | MEM[(int *)p_28].a = -1; | |
2619 | stmts in the current group and need to consider if it is safe to | |
2620 | add MEM[(long long int *)p_28 + 8B] = 0; store into the same group. | |
2621 | There is an overlap between that store and the MEM[(int *)p_28 + 8B] = _129; | |
2622 | store though, so if we add the MEM[(long long int *)p_28 + 8B] = 0; | |
2623 | into the group and merging of those 3 stores is successful, merged | |
2624 | stmts will be emitted at the latest store from that group, i.e. | |
2625 | LAST_ORDER, which is the MEM[(int *)p_28].a = -1; store. | |
2626 | The MEM[(int *)p_28 + 8B] = _129; store that originally follows | |
2627 | the MEM[(long long int *)p_28 + 8B] = 0; would now be before it, | |
2628 | so we need to refuse merging MEM[(long long int *)p_28 + 8B] = 0; | |
2629 | into the group. That way it will be its own store group and will | |
4d213bf6 | 2630 | not be touched. If ALL_INTEGER_CST_P and there are overlapping |
c5679c37 | 2631 | INTEGER_CST stores, those are mergeable using merge_overlapping, |
bd909071 JJ |
2632 | so don't return false for those. |
2633 | ||
2634 | Similarly, check stores from FIRST_EARLIER (inclusive) to END_EARLIER | |
2635 | (exclusive), whether they don't overlap the bitrange START to END | |
2636 | and have order in between FIRST_ORDER and LAST_ORDER. This is to | |
2637 | prevent merging in cases like: | |
2638 | MEM <char[12]> [&b + 8B] = {}; | |
2639 | MEM[(short *) &b] = 5; | |
2640 | _5 = *x_4(D); | |
2641 | MEM <long long unsigned int> [&b + 2B] = _5; | |
2642 | MEM[(char *)&b + 16B] = 88; | |
2643 | MEM[(int *)&b + 20B] = 1; | |
2644 | The = {} store comes in sort_by_bitpos before the = 88 store, and can't | |
2645 | be merged with it, because the = _5 store overlaps these and is in between | |
2646 | them in sort_by_order ordering. If it was merged, the merged store would | |
2647 | go after the = _5 store and thus change behavior. */ | |
c5679c37 JJ |
2648 | |
2649 | static bool | |
2650 | check_no_overlap (vec<store_immediate_info *> m_store_info, unsigned int i, | |
bd909071 JJ |
2651 | bool all_integer_cst_p, unsigned int first_order, |
2652 | unsigned int last_order, unsigned HOST_WIDE_INT start, | |
2653 | unsigned HOST_WIDE_INT end, unsigned int first_earlier, | |
2654 | unsigned end_earlier) | |
c5679c37 JJ |
2655 | { |
2656 | unsigned int len = m_store_info.length (); | |
bd909071 JJ |
2657 | for (unsigned int j = first_earlier; j < end_earlier; j++) |
2658 | { | |
2659 | store_immediate_info *info = m_store_info[j]; | |
2660 | if (info->order > first_order | |
2661 | && info->order < last_order | |
2662 | && info->bitpos + info->bitsize > start) | |
2663 | return false; | |
2664 | } | |
c5679c37 JJ |
2665 | for (++i; i < len; ++i) |
2666 | { | |
2667 | store_immediate_info *info = m_store_info[i]; | |
2668 | if (info->bitpos >= end) | |
2669 | break; | |
2670 | if (info->order < last_order | |
4d213bf6 | 2671 | && (!all_integer_cst_p || info->rhs_code != INTEGER_CST)) |
c5679c37 JJ |
2672 | return false; |
2673 | } | |
2674 | return true; | |
2675 | } | |
2676 | ||
4b84d9b8 JJ |
2677 | /* Return true if m_store_info[first] and at least one following store |
2678 | form a group which store try_size bitsize value which is byte swapped | |
2679 | from a memory load or some value, or identity from some value. | |
2680 | This uses the bswap pass APIs. */ | |
2681 | ||
2682 | bool | |
2683 | imm_store_chain_info::try_coalesce_bswap (merged_store_group *merged_store, | |
2684 | unsigned int first, | |
bd909071 JJ |
2685 | unsigned int try_size, |
2686 | unsigned int first_earlier) | |
4b84d9b8 JJ |
2687 | { |
2688 | unsigned int len = m_store_info.length (), last = first; | |
2689 | unsigned HOST_WIDE_INT width = m_store_info[first]->bitsize; | |
2690 | if (width >= try_size) | |
2691 | return false; | |
2692 | for (unsigned int i = first + 1; i < len; ++i) | |
2693 | { | |
2694 | if (m_store_info[i]->bitpos != m_store_info[first]->bitpos + width | |
cb76fcd7 | 2695 | || m_store_info[i]->lp_nr != merged_store->lp_nr |
4b84d9b8 JJ |
2696 | || m_store_info[i]->ins_stmt == NULL) |
2697 | return false; | |
2698 | width += m_store_info[i]->bitsize; | |
2699 | if (width >= try_size) | |
2700 | { | |
2701 | last = i; | |
2702 | break; | |
2703 | } | |
2704 | } | |
2705 | if (width != try_size) | |
2706 | return false; | |
2707 | ||
2708 | bool allow_unaligned | |
028d4092 | 2709 | = !STRICT_ALIGNMENT && param_store_merging_allow_unaligned; |
4b84d9b8 JJ |
2710 | /* Punt if the combined store would not be aligned and we need alignment. */ |
2711 | if (!allow_unaligned) | |
2712 | { | |
2713 | unsigned int align = merged_store->align; | |
2714 | unsigned HOST_WIDE_INT align_base = merged_store->align_base; | |
2715 | for (unsigned int i = first + 1; i <= last; ++i) | |
2716 | { | |
2717 | unsigned int this_align; | |
2718 | unsigned HOST_WIDE_INT align_bitpos = 0; | |
2719 | get_object_alignment_1 (gimple_assign_lhs (m_store_info[i]->stmt), | |
2720 | &this_align, &align_bitpos); | |
2721 | if (this_align > align) | |
2722 | { | |
2723 | align = this_align; | |
2724 | align_base = m_store_info[i]->bitpos - align_bitpos; | |
2725 | } | |
2726 | } | |
2727 | unsigned HOST_WIDE_INT align_bitpos | |
2728 | = (m_store_info[first]->bitpos - align_base) & (align - 1); | |
2729 | if (align_bitpos) | |
2730 | align = least_bit_hwi (align_bitpos); | |
2731 | if (align < try_size) | |
2732 | return false; | |
2733 | } | |
2734 | ||
2735 | tree type; | |
2736 | switch (try_size) | |
2737 | { | |
2738 | case 16: type = uint16_type_node; break; | |
2739 | case 32: type = uint32_type_node; break; | |
2740 | case 64: type = uint64_type_node; break; | |
2741 | default: gcc_unreachable (); | |
2742 | } | |
2743 | struct symbolic_number n; | |
2744 | gimple *ins_stmt = NULL; | |
2745 | int vuse_store = -1; | |
2746 | unsigned int first_order = merged_store->first_order; | |
2747 | unsigned int last_order = merged_store->last_order; | |
2748 | gimple *first_stmt = merged_store->first_stmt; | |
2749 | gimple *last_stmt = merged_store->last_stmt; | |
c5679c37 | 2750 | unsigned HOST_WIDE_INT end = merged_store->start + merged_store->width; |
4b84d9b8 JJ |
2751 | store_immediate_info *infof = m_store_info[first]; |
2752 | ||
2753 | for (unsigned int i = first; i <= last; ++i) | |
2754 | { | |
2755 | store_immediate_info *info = m_store_info[i]; | |
2756 | struct symbolic_number this_n = info->n; | |
2757 | this_n.type = type; | |
2758 | if (!this_n.base_addr) | |
2759 | this_n.range = try_size / BITS_PER_UNIT; | |
30fa8e9c JJ |
2760 | else |
2761 | /* Update vuse in case it has changed by output_merged_stores. */ | |
2762 | this_n.vuse = gimple_vuse (info->ins_stmt); | |
4b84d9b8 JJ |
2763 | unsigned int bitpos = info->bitpos - infof->bitpos; |
2764 | if (!do_shift_rotate (LSHIFT_EXPR, &this_n, | |
2765 | BYTES_BIG_ENDIAN | |
2766 | ? try_size - info->bitsize - bitpos | |
2767 | : bitpos)) | |
2768 | return false; | |
aa11164a | 2769 | if (this_n.base_addr && vuse_store) |
4b84d9b8 JJ |
2770 | { |
2771 | unsigned int j; | |
2772 | for (j = first; j <= last; ++j) | |
2773 | if (this_n.vuse == gimple_vuse (m_store_info[j]->stmt)) | |
2774 | break; | |
2775 | if (j > last) | |
2776 | { | |
2777 | if (vuse_store == 1) | |
2778 | return false; | |
2779 | vuse_store = 0; | |
2780 | } | |
2781 | } | |
2782 | if (i == first) | |
2783 | { | |
2784 | n = this_n; | |
2785 | ins_stmt = info->ins_stmt; | |
2786 | } | |
2787 | else | |
2788 | { | |
c5679c37 | 2789 | if (n.base_addr && n.vuse != this_n.vuse) |
4b84d9b8 | 2790 | { |
c5679c37 JJ |
2791 | if (vuse_store == 0) |
2792 | return false; | |
2793 | vuse_store = 1; | |
4b84d9b8 | 2794 | } |
c5679c37 JJ |
2795 | if (info->order > last_order) |
2796 | { | |
2797 | last_order = info->order; | |
2798 | last_stmt = info->stmt; | |
2799 | } | |
2800 | else if (info->order < first_order) | |
2801 | { | |
2802 | first_order = info->order; | |
2803 | first_stmt = info->stmt; | |
2804 | } | |
2805 | end = MAX (end, info->bitpos + info->bitsize); | |
4b84d9b8 JJ |
2806 | |
2807 | ins_stmt = perform_symbolic_merge (ins_stmt, &n, info->ins_stmt, | |
2808 | &this_n, &n); | |
2809 | if (ins_stmt == NULL) | |
2810 | return false; | |
2811 | } | |
2812 | } | |
2813 | ||
2814 | uint64_t cmpxchg, cmpnop; | |
2815 | find_bswap_or_nop_finalize (&n, &cmpxchg, &cmpnop); | |
2816 | ||
2817 | /* A complete byte swap should make the symbolic number to start with | |
2818 | the largest digit in the highest order byte. Unchanged symbolic | |
2819 | number indicates a read with same endianness as target architecture. */ | |
2820 | if (n.n != cmpnop && n.n != cmpxchg) | |
2821 | return false; | |
2822 | ||
2823 | if (n.base_addr == NULL_TREE && !is_gimple_val (n.src)) | |
2824 | return false; | |
2825 | ||
bd909071 JJ |
2826 | if (!check_no_overlap (m_store_info, last, false, first_order, last_order, |
2827 | merged_store->start, end, first_earlier, first)) | |
c5679c37 JJ |
2828 | return false; |
2829 | ||
4b84d9b8 JJ |
2830 | /* Don't handle memory copy this way if normal non-bswap processing |
2831 | would handle it too. */ | |
2832 | if (n.n == cmpnop && (unsigned) n.n_ops == last - first + 1) | |
2833 | { | |
2834 | unsigned int i; | |
2835 | for (i = first; i <= last; ++i) | |
2836 | if (m_store_info[i]->rhs_code != MEM_REF) | |
2837 | break; | |
2838 | if (i == last + 1) | |
2839 | return false; | |
2840 | } | |
2841 | ||
2842 | if (n.n == cmpxchg) | |
2843 | switch (try_size) | |
2844 | { | |
2845 | case 16: | |
2846 | /* Will emit LROTATE_EXPR. */ | |
2847 | break; | |
2848 | case 32: | |
2849 | if (builtin_decl_explicit_p (BUILT_IN_BSWAP32) | |
2850 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing) | |
2851 | break; | |
2852 | return false; | |
2853 | case 64: | |
2854 | if (builtin_decl_explicit_p (BUILT_IN_BSWAP64) | |
2855 | && optab_handler (bswap_optab, DImode) != CODE_FOR_nothing) | |
2856 | break; | |
2857 | return false; | |
2858 | default: | |
2859 | gcc_unreachable (); | |
2860 | } | |
2861 | ||
2862 | if (!allow_unaligned && n.base_addr) | |
2863 | { | |
2864 | unsigned int align = get_object_alignment (n.src); | |
2865 | if (align < try_size) | |
2866 | return false; | |
2867 | } | |
2868 | ||
2869 | /* If each load has vuse of the corresponding store, need to verify | |
2870 | the loads can be sunk right before the last store. */ | |
2871 | if (vuse_store == 1) | |
2872 | { | |
2873 | auto_vec<tree, 64> refs; | |
2874 | for (unsigned int i = first; i <= last; ++i) | |
2875 | gather_bswap_load_refs (&refs, | |
2876 | gimple_assign_rhs1 (m_store_info[i]->stmt)); | |
2877 | ||
2878 | unsigned int i; | |
2879 | tree ref; | |
2880 | FOR_EACH_VEC_ELT (refs, i, ref) | |
2881 | if (stmts_may_clobber_ref_p (first_stmt, last_stmt, ref)) | |
2882 | return false; | |
2883 | n.vuse = NULL_TREE; | |
2884 | } | |
2885 | ||
2886 | infof->n = n; | |
2887 | infof->ins_stmt = ins_stmt; | |
2888 | for (unsigned int i = first; i <= last; ++i) | |
2889 | { | |
2890 | m_store_info[i]->rhs_code = n.n == cmpxchg ? LROTATE_EXPR : NOP_EXPR; | |
2891 | m_store_info[i]->ops[0].base_addr = NULL_TREE; | |
2892 | m_store_info[i]->ops[1].base_addr = NULL_TREE; | |
2893 | if (i != first) | |
2894 | merged_store->merge_into (m_store_info[i]); | |
2895 | } | |
2896 | ||
2897 | return true; | |
2898 | } | |
2899 | ||
f663d9ad KT |
2900 | /* Go through the candidate stores recorded in m_store_info and merge them |
2901 | into merged_store_group objects recorded into m_merged_store_groups | |
2902 | representing the widened stores. Return true if coalescing was successful | |
2903 | and the number of widened stores is fewer than the original number | |
2904 | of stores. */ | |
2905 | ||
2906 | bool | |
2907 | imm_store_chain_info::coalesce_immediate_stores () | |
2908 | { | |
2909 | /* Anything less can't be processed. */ | |
2910 | if (m_store_info.length () < 2) | |
2911 | return false; | |
2912 | ||
2913 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
c94c3532 | 2914 | fprintf (dump_file, "Attempting to coalesce %u stores in chain\n", |
f663d9ad KT |
2915 | m_store_info.length ()); |
2916 | ||
2917 | store_immediate_info *info; | |
4b84d9b8 | 2918 | unsigned int i, ignore = 0; |
bd909071 JJ |
2919 | unsigned int first_earlier = 0; |
2920 | unsigned int end_earlier = 0; | |
f663d9ad KT |
2921 | |
2922 | /* Order the stores by the bitposition they write to. */ | |
2923 | m_store_info.qsort (sort_by_bitpos); | |
2924 | ||
2925 | info = m_store_info[0]; | |
2926 | merged_store_group *merged_store = new merged_store_group (info); | |
c94c3532 EB |
2927 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2928 | fputs ("New store group\n", dump_file); | |
f663d9ad KT |
2929 | |
2930 | FOR_EACH_VEC_ELT (m_store_info, i, info) | |
2931 | { | |
3afd514b JJ |
2932 | unsigned HOST_WIDE_INT new_bitregion_start, new_bitregion_end; |
2933 | ||
4b84d9b8 | 2934 | if (i <= ignore) |
c94c3532 | 2935 | goto done; |
f663d9ad | 2936 | |
bd909071 JJ |
2937 | while (first_earlier < end_earlier |
2938 | && (m_store_info[first_earlier]->bitpos | |
2939 | + m_store_info[first_earlier]->bitsize | |
2940 | <= merged_store->start)) | |
2941 | first_earlier++; | |
2942 | ||
4b84d9b8 JJ |
2943 | /* First try to handle group of stores like: |
2944 | p[0] = data >> 24; | |
2945 | p[1] = data >> 16; | |
2946 | p[2] = data >> 8; | |
2947 | p[3] = data; | |
2948 | using the bswap framework. */ | |
2949 | if (info->bitpos == merged_store->start + merged_store->width | |
2950 | && merged_store->stores.length () == 1 | |
2951 | && merged_store->stores[0]->ins_stmt != NULL | |
cb76fcd7 | 2952 | && info->lp_nr == merged_store->lp_nr |
4b84d9b8 JJ |
2953 | && info->ins_stmt != NULL) |
2954 | { | |
2955 | unsigned int try_size; | |
2956 | for (try_size = 64; try_size >= 16; try_size >>= 1) | |
bd909071 JJ |
2957 | if (try_coalesce_bswap (merged_store, i - 1, try_size, |
2958 | first_earlier)) | |
4b84d9b8 JJ |
2959 | break; |
2960 | ||
2961 | if (try_size >= 16) | |
2962 | { | |
2963 | ignore = i + merged_store->stores.length () - 1; | |
2964 | m_merged_store_groups.safe_push (merged_store); | |
2965 | if (ignore < m_store_info.length ()) | |
bd909071 JJ |
2966 | { |
2967 | merged_store = new merged_store_group (m_store_info[ignore]); | |
2968 | end_earlier = ignore; | |
2969 | } | |
4b84d9b8 JJ |
2970 | else |
2971 | merged_store = NULL; | |
c94c3532 | 2972 | goto done; |
4b84d9b8 JJ |
2973 | } |
2974 | } | |
2975 | ||
3afd514b JJ |
2976 | new_bitregion_start |
2977 | = MIN (merged_store->bitregion_start, info->bitregion_start); | |
2978 | new_bitregion_end | |
2979 | = MAX (merged_store->bitregion_end, info->bitregion_end); | |
2980 | ||
2981 | if (info->order >= merged_store->first_nonmergeable_order | |
2982 | || (((new_bitregion_end - new_bitregion_start + 1) / BITS_PER_UNIT) | |
028d4092 | 2983 | > (unsigned) param_store_merging_max_size)) |
18e0c3d1 JJ |
2984 | ; |
2985 | ||
f663d9ad KT |
2986 | /* |---store 1---| |
2987 | |---store 2---| | |
4b84d9b8 | 2988 | Overlapping stores. */ |
18e0c3d1 | 2989 | else if (IN_RANGE (info->bitpos, merged_store->start, |
4d213bf6 JJ |
2990 | merged_store->start + merged_store->width - 1) |
2991 | /* |---store 1---||---store 2---| | |
2992 | Handle also the consecutive INTEGER_CST stores case here, | |
2993 | as we have here the code to deal with overlaps. */ | |
2994 | || (info->bitregion_start <= merged_store->bitregion_end | |
2995 | && info->rhs_code == INTEGER_CST | |
2996 | && merged_store->only_constants | |
2997 | && merged_store->can_be_merged_into (info))) | |
f663d9ad | 2998 | { |
245f6de1 | 2999 | /* Only allow overlapping stores of constants. */ |
629387a6 EB |
3000 | if (info->rhs_code == INTEGER_CST |
3001 | && merged_store->only_constants | |
3002 | && info->lp_nr == merged_store->lp_nr) | |
245f6de1 | 3003 | { |
bd909071 JJ |
3004 | unsigned int first_order |
3005 | = MIN (merged_store->first_order, info->order); | |
6cd4c66e JJ |
3006 | unsigned int last_order |
3007 | = MAX (merged_store->last_order, info->order); | |
3008 | unsigned HOST_WIDE_INT end | |
3009 | = MAX (merged_store->start + merged_store->width, | |
3010 | info->bitpos + info->bitsize); | |
bd909071 JJ |
3011 | if (check_no_overlap (m_store_info, i, true, first_order, |
3012 | last_order, merged_store->start, end, | |
3013 | first_earlier, end_earlier)) | |
6cd4c66e JJ |
3014 | { |
3015 | /* check_no_overlap call above made sure there are no | |
3016 | overlapping stores with non-INTEGER_CST rhs_code | |
3017 | in between the first and last of the stores we've | |
3018 | just merged. If there are any INTEGER_CST rhs_code | |
3019 | stores in between, we need to merge_overlapping them | |
3020 | even if in the sort_by_bitpos order there are other | |
3021 | overlapping stores in between. Keep those stores as is. | |
3022 | Example: | |
3023 | MEM[(int *)p_28] = 0; | |
3024 | MEM[(char *)p_28 + 3B] = 1; | |
3025 | MEM[(char *)p_28 + 1B] = 2; | |
3026 | MEM[(char *)p_28 + 2B] = MEM[(char *)p_28 + 6B]; | |
3027 | We can't merge the zero store with the store of two and | |
3028 | not merge anything else, because the store of one is | |
3029 | in the original order in between those two, but in | |
3030 | store_by_bitpos order it comes after the last store that | |
3031 | we can't merge with them. We can merge the first 3 stores | |
3032 | and keep the last store as is though. */ | |
18e0c3d1 JJ |
3033 | unsigned int len = m_store_info.length (); |
3034 | unsigned int try_order = last_order; | |
3035 | unsigned int first_nonmergeable_order; | |
3036 | unsigned int k; | |
3037 | bool last_iter = false; | |
3038 | int attempts = 0; | |
3039 | do | |
6cd4c66e | 3040 | { |
18e0c3d1 | 3041 | unsigned int max_order = 0; |
bd909071 | 3042 | unsigned int min_order = first_order; |
18e0c3d1 JJ |
3043 | unsigned first_nonmergeable_int_order = ~0U; |
3044 | unsigned HOST_WIDE_INT this_end = end; | |
3045 | k = i; | |
3046 | first_nonmergeable_order = ~0U; | |
3047 | for (unsigned int j = i + 1; j < len; ++j) | |
6cd4c66e | 3048 | { |
18e0c3d1 JJ |
3049 | store_immediate_info *info2 = m_store_info[j]; |
3050 | if (info2->bitpos >= this_end) | |
3051 | break; | |
3052 | if (info2->order < try_order) | |
6cd4c66e | 3053 | { |
4119cd69 JJ |
3054 | if (info2->rhs_code != INTEGER_CST |
3055 | || info2->lp_nr != merged_store->lp_nr) | |
18e0c3d1 JJ |
3056 | { |
3057 | /* Normally check_no_overlap makes sure this | |
3058 | doesn't happen, but if end grows below, | |
3059 | then we need to process more stores than | |
3060 | check_no_overlap verified. Example: | |
3061 | MEM[(int *)p_5] = 0; | |
3062 | MEM[(short *)p_5 + 3B] = 1; | |
3063 | MEM[(char *)p_5 + 4B] = _9; | |
3064 | MEM[(char *)p_5 + 2B] = 2; */ | |
3065 | k = 0; | |
3066 | break; | |
3067 | } | |
3068 | k = j; | |
bd909071 | 3069 | min_order = MIN (min_order, info2->order); |
18e0c3d1 JJ |
3070 | this_end = MAX (this_end, |
3071 | info2->bitpos + info2->bitsize); | |
6cd4c66e | 3072 | } |
18e0c3d1 | 3073 | else if (info2->rhs_code == INTEGER_CST |
4119cd69 | 3074 | && info2->lp_nr == merged_store->lp_nr |
18e0c3d1 JJ |
3075 | && !last_iter) |
3076 | { | |
3077 | max_order = MAX (max_order, info2->order + 1); | |
3078 | first_nonmergeable_int_order | |
3079 | = MIN (first_nonmergeable_int_order, | |
3080 | info2->order); | |
3081 | } | |
3082 | else | |
3083 | first_nonmergeable_order | |
3084 | = MIN (first_nonmergeable_order, info2->order); | |
6cd4c66e | 3085 | } |
bd909071 JJ |
3086 | if (k > i |
3087 | && !check_no_overlap (m_store_info, len - 1, true, | |
3088 | min_order, try_order, | |
3089 | merged_store->start, this_end, | |
3090 | first_earlier, end_earlier)) | |
3091 | k = 0; | |
18e0c3d1 JJ |
3092 | if (k == 0) |
3093 | { | |
3094 | if (last_order == try_order) | |
3095 | break; | |
3096 | /* If this failed, but only because we grew | |
3097 | try_order, retry with the last working one, | |
3098 | so that we merge at least something. */ | |
3099 | try_order = last_order; | |
3100 | last_iter = true; | |
3101 | continue; | |
3102 | } | |
3103 | last_order = try_order; | |
3104 | /* Retry with a larger try_order to see if we could | |
3105 | merge some further INTEGER_CST stores. */ | |
3106 | if (max_order | |
3107 | && (first_nonmergeable_int_order | |
3108 | < first_nonmergeable_order)) | |
3109 | { | |
3110 | try_order = MIN (max_order, | |
3111 | first_nonmergeable_order); | |
3112 | try_order | |
3113 | = MIN (try_order, | |
3114 | merged_store->first_nonmergeable_order); | |
3115 | if (try_order > last_order && ++attempts < 16) | |
3116 | continue; | |
3117 | } | |
3118 | first_nonmergeable_order | |
3119 | = MIN (first_nonmergeable_order, | |
3120 | first_nonmergeable_int_order); | |
3121 | end = this_end; | |
3122 | break; | |
6cd4c66e | 3123 | } |
18e0c3d1 | 3124 | while (1); |
6cd4c66e JJ |
3125 | |
3126 | if (k != 0) | |
3127 | { | |
3128 | merged_store->merge_overlapping (info); | |
3129 | ||
18e0c3d1 JJ |
3130 | merged_store->first_nonmergeable_order |
3131 | = MIN (merged_store->first_nonmergeable_order, | |
3132 | first_nonmergeable_order); | |
3133 | ||
6cd4c66e JJ |
3134 | for (unsigned int j = i + 1; j <= k; j++) |
3135 | { | |
3136 | store_immediate_info *info2 = m_store_info[j]; | |
3137 | gcc_assert (info2->bitpos < end); | |
3138 | if (info2->order < last_order) | |
3139 | { | |
3140 | gcc_assert (info2->rhs_code == INTEGER_CST); | |
18e0c3d1 JJ |
3141 | if (info != info2) |
3142 | merged_store->merge_overlapping (info2); | |
6cd4c66e JJ |
3143 | } |
3144 | /* Other stores are kept and not merged in any | |
3145 | way. */ | |
3146 | } | |
3147 | ignore = k; | |
3148 | goto done; | |
3149 | } | |
3150 | } | |
245f6de1 | 3151 | } |
f663d9ad | 3152 | } |
245f6de1 JJ |
3153 | /* |---store 1---||---store 2---| |
3154 | This store is consecutive to the previous one. | |
3155 | Merge it into the current store group. There can be gaps in between | |
3156 | the stores, but there can't be gaps in between bitregions. */ | |
c94c3532 | 3157 | else if (info->bitregion_start <= merged_store->bitregion_end |
7f5a3982 | 3158 | && merged_store->can_be_merged_into (info)) |
f663d9ad | 3159 | { |
245f6de1 JJ |
3160 | store_immediate_info *infof = merged_store->stores[0]; |
3161 | ||
3162 | /* All the rhs_code ops that take 2 operands are commutative, | |
3163 | swap the operands if it could make the operands compatible. */ | |
3164 | if (infof->ops[0].base_addr | |
3165 | && infof->ops[1].base_addr | |
3166 | && info->ops[0].base_addr | |
3167 | && info->ops[1].base_addr | |
8a91d545 RS |
3168 | && known_eq (info->ops[1].bitpos - infof->ops[0].bitpos, |
3169 | info->bitpos - infof->bitpos) | |
245f6de1 JJ |
3170 | && operand_equal_p (info->ops[1].base_addr, |
3171 | infof->ops[0].base_addr, 0)) | |
127ef369 JJ |
3172 | { |
3173 | std::swap (info->ops[0], info->ops[1]); | |
3174 | info->ops_swapped_p = true; | |
3175 | } | |
4d213bf6 | 3176 | if (check_no_overlap (m_store_info, i, false, |
bd909071 | 3177 | MIN (merged_store->first_order, info->order), |
a7fe6482 | 3178 | MAX (merged_store->last_order, info->order), |
bd909071 | 3179 | merged_store->start, |
a7fe6482 | 3180 | MAX (merged_store->start + merged_store->width, |
bd909071 JJ |
3181 | info->bitpos + info->bitsize), |
3182 | first_earlier, end_earlier)) | |
245f6de1 | 3183 | { |
7f5a3982 EB |
3184 | /* Turn MEM_REF into BIT_INSERT_EXPR for bit-field stores. */ |
3185 | if (info->rhs_code == MEM_REF && infof->rhs_code != MEM_REF) | |
3186 | { | |
3187 | info->rhs_code = BIT_INSERT_EXPR; | |
3188 | info->ops[0].val = gimple_assign_rhs1 (info->stmt); | |
3189 | info->ops[0].base_addr = NULL_TREE; | |
3190 | } | |
3191 | else if (infof->rhs_code == MEM_REF && info->rhs_code != MEM_REF) | |
3192 | { | |
3193 | store_immediate_info *infoj; | |
3194 | unsigned int j; | |
3195 | FOR_EACH_VEC_ELT (merged_store->stores, j, infoj) | |
3196 | { | |
3197 | infoj->rhs_code = BIT_INSERT_EXPR; | |
3198 | infoj->ops[0].val = gimple_assign_rhs1 (infoj->stmt); | |
3199 | infoj->ops[0].base_addr = NULL_TREE; | |
3200 | } | |
e362a897 | 3201 | merged_store->bit_insertion = true; |
7f5a3982 EB |
3202 | } |
3203 | if ((infof->ops[0].base_addr | |
3204 | ? compatible_load_p (merged_store, info, base_addr, 0) | |
3205 | : !info->ops[0].base_addr) | |
3206 | && (infof->ops[1].base_addr | |
3207 | ? compatible_load_p (merged_store, info, base_addr, 1) | |
3208 | : !info->ops[1].base_addr)) | |
3209 | { | |
3210 | merged_store->merge_into (info); | |
3211 | goto done; | |
3212 | } | |
245f6de1 JJ |
3213 | } |
3214 | } | |
f663d9ad | 3215 | |
245f6de1 JJ |
3216 | /* |---store 1---| <gap> |---store 2---|. |
3217 | Gap between stores or the rhs not compatible. Start a new group. */ | |
f663d9ad | 3218 | |
245f6de1 JJ |
3219 | /* Try to apply all the stores recorded for the group to determine |
3220 | the bitpattern they write and discard it if that fails. | |
3221 | This will also reject single-store groups. */ | |
c94c3532 | 3222 | if (merged_store->apply_stores ()) |
245f6de1 | 3223 | m_merged_store_groups.safe_push (merged_store); |
c94c3532 EB |
3224 | else |
3225 | delete merged_store; | |
f663d9ad | 3226 | |
245f6de1 | 3227 | merged_store = new merged_store_group (info); |
bd909071 | 3228 | end_earlier = i; |
c94c3532 EB |
3229 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3230 | fputs ("New store group\n", dump_file); | |
3231 | ||
3232 | done: | |
3233 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3234 | { | |
3235 | fprintf (dump_file, "Store %u:\nbitsize:" HOST_WIDE_INT_PRINT_DEC | |
3236 | " bitpos:" HOST_WIDE_INT_PRINT_DEC " val:", | |
3237 | i, info->bitsize, info->bitpos); | |
3238 | print_generic_expr (dump_file, gimple_assign_rhs1 (info->stmt)); | |
3239 | fputc ('\n', dump_file); | |
3240 | } | |
f663d9ad KT |
3241 | } |
3242 | ||
a62b3dc5 | 3243 | /* Record or discard the last store group. */ |
4b84d9b8 JJ |
3244 | if (merged_store) |
3245 | { | |
c94c3532 | 3246 | if (merged_store->apply_stores ()) |
4b84d9b8 | 3247 | m_merged_store_groups.safe_push (merged_store); |
c94c3532 EB |
3248 | else |
3249 | delete merged_store; | |
4b84d9b8 | 3250 | } |
f663d9ad KT |
3251 | |
3252 | gcc_assert (m_merged_store_groups.length () <= m_store_info.length ()); | |
c94c3532 | 3253 | |
f663d9ad KT |
3254 | bool success |
3255 | = !m_merged_store_groups.is_empty () | |
3256 | && m_merged_store_groups.length () < m_store_info.length (); | |
3257 | ||
3258 | if (success && dump_file) | |
c94c3532 | 3259 | fprintf (dump_file, "Coalescing successful!\nMerged into %u stores\n", |
a62b3dc5 | 3260 | m_merged_store_groups.length ()); |
f663d9ad KT |
3261 | |
3262 | return success; | |
3263 | } | |
3264 | ||
245f6de1 JJ |
3265 | /* Return the type to use for the merged stores or loads described by STMTS. |
3266 | This is needed to get the alias sets right. If IS_LOAD, look for rhs, | |
3267 | otherwise lhs. Additionally set *CLIQUEP and *BASEP to MR_DEPENDENCE_* | |
3268 | of the MEM_REFs if any. */ | |
f663d9ad KT |
3269 | |
3270 | static tree | |
245f6de1 JJ |
3271 | get_alias_type_for_stmts (vec<gimple *> &stmts, bool is_load, |
3272 | unsigned short *cliquep, unsigned short *basep) | |
f663d9ad KT |
3273 | { |
3274 | gimple *stmt; | |
3275 | unsigned int i; | |
245f6de1 JJ |
3276 | tree type = NULL_TREE; |
3277 | tree ret = NULL_TREE; | |
3278 | *cliquep = 0; | |
3279 | *basep = 0; | |
f663d9ad KT |
3280 | |
3281 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
3282 | { | |
245f6de1 JJ |
3283 | tree ref = is_load ? gimple_assign_rhs1 (stmt) |
3284 | : gimple_assign_lhs (stmt); | |
3285 | tree type1 = reference_alias_ptr_type (ref); | |
3286 | tree base = get_base_address (ref); | |
f663d9ad | 3287 | |
245f6de1 JJ |
3288 | if (i == 0) |
3289 | { | |
3290 | if (TREE_CODE (base) == MEM_REF) | |
3291 | { | |
3292 | *cliquep = MR_DEPENDENCE_CLIQUE (base); | |
3293 | *basep = MR_DEPENDENCE_BASE (base); | |
3294 | } | |
3295 | ret = type = type1; | |
3296 | continue; | |
3297 | } | |
f663d9ad | 3298 | if (!alias_ptr_types_compatible_p (type, type1)) |
245f6de1 JJ |
3299 | ret = ptr_type_node; |
3300 | if (TREE_CODE (base) != MEM_REF | |
3301 | || *cliquep != MR_DEPENDENCE_CLIQUE (base) | |
3302 | || *basep != MR_DEPENDENCE_BASE (base)) | |
3303 | { | |
3304 | *cliquep = 0; | |
3305 | *basep = 0; | |
3306 | } | |
f663d9ad | 3307 | } |
245f6de1 | 3308 | return ret; |
f663d9ad KT |
3309 | } |
3310 | ||
3311 | /* Return the location_t information we can find among the statements | |
3312 | in STMTS. */ | |
3313 | ||
3314 | static location_t | |
245f6de1 | 3315 | get_location_for_stmts (vec<gimple *> &stmts) |
f663d9ad KT |
3316 | { |
3317 | gimple *stmt; | |
3318 | unsigned int i; | |
3319 | ||
3320 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
3321 | if (gimple_has_location (stmt)) | |
3322 | return gimple_location (stmt); | |
3323 | ||
3324 | return UNKNOWN_LOCATION; | |
3325 | } | |
3326 | ||
3327 | /* Used to decribe a store resulting from splitting a wide store in smaller | |
3328 | regularly-sized stores in split_group. */ | |
3329 | ||
6c1dae73 | 3330 | class split_store |
f663d9ad | 3331 | { |
6c1dae73 | 3332 | public: |
f663d9ad KT |
3333 | unsigned HOST_WIDE_INT bytepos; |
3334 | unsigned HOST_WIDE_INT size; | |
3335 | unsigned HOST_WIDE_INT align; | |
245f6de1 | 3336 | auto_vec<store_immediate_info *> orig_stores; |
a62b3dc5 JJ |
3337 | /* True if there is a single orig stmt covering the whole split store. */ |
3338 | bool orig; | |
f663d9ad KT |
3339 | split_store (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, |
3340 | unsigned HOST_WIDE_INT); | |
3341 | }; | |
3342 | ||
3343 | /* Simple constructor. */ | |
3344 | ||
3345 | split_store::split_store (unsigned HOST_WIDE_INT bp, | |
3346 | unsigned HOST_WIDE_INT sz, | |
3347 | unsigned HOST_WIDE_INT al) | |
a62b3dc5 | 3348 | : bytepos (bp), size (sz), align (al), orig (false) |
f663d9ad | 3349 | { |
245f6de1 | 3350 | orig_stores.create (0); |
f663d9ad KT |
3351 | } |
3352 | ||
245f6de1 JJ |
3353 | /* Record all stores in GROUP that write to the region starting at BITPOS and |
3354 | is of size BITSIZE. Record infos for such statements in STORES if | |
3355 | non-NULL. The stores in GROUP must be sorted by bitposition. Return INFO | |
5384a802 JJ |
3356 | if there is exactly one original store in the range (in that case ignore |
3357 | clobber stmts, unless there are only clobber stmts). */ | |
f663d9ad | 3358 | |
a62b3dc5 | 3359 | static store_immediate_info * |
99b1c316 | 3360 | find_constituent_stores (class merged_store_group *group, |
245f6de1 JJ |
3361 | vec<store_immediate_info *> *stores, |
3362 | unsigned int *first, | |
3363 | unsigned HOST_WIDE_INT bitpos, | |
3364 | unsigned HOST_WIDE_INT bitsize) | |
f663d9ad | 3365 | { |
a62b3dc5 | 3366 | store_immediate_info *info, *ret = NULL; |
f663d9ad | 3367 | unsigned int i; |
a62b3dc5 JJ |
3368 | bool second = false; |
3369 | bool update_first = true; | |
f663d9ad | 3370 | unsigned HOST_WIDE_INT end = bitpos + bitsize; |
a62b3dc5 | 3371 | for (i = *first; group->stores.iterate (i, &info); ++i) |
f663d9ad KT |
3372 | { |
3373 | unsigned HOST_WIDE_INT stmt_start = info->bitpos; | |
3374 | unsigned HOST_WIDE_INT stmt_end = stmt_start + info->bitsize; | |
a62b3dc5 JJ |
3375 | if (stmt_end <= bitpos) |
3376 | { | |
3377 | /* BITPOS passed to this function never decreases from within the | |
3378 | same split_group call, so optimize and don't scan info records | |
3379 | which are known to end before or at BITPOS next time. | |
3380 | Only do it if all stores before this one also pass this. */ | |
3381 | if (update_first) | |
3382 | *first = i + 1; | |
3383 | continue; | |
3384 | } | |
3385 | else | |
3386 | update_first = false; | |
3387 | ||
f663d9ad | 3388 | /* The stores in GROUP are ordered by bitposition so if we're past |
a62b3dc5 JJ |
3389 | the region for this group return early. */ |
3390 | if (stmt_start >= end) | |
3391 | return ret; | |
3392 | ||
5384a802 JJ |
3393 | if (gimple_clobber_p (info->stmt)) |
3394 | { | |
3395 | if (stores) | |
3396 | stores->safe_push (info); | |
3397 | if (ret == NULL) | |
3398 | ret = info; | |
3399 | continue; | |
3400 | } | |
245f6de1 | 3401 | if (stores) |
a62b3dc5 | 3402 | { |
245f6de1 | 3403 | stores->safe_push (info); |
5384a802 | 3404 | if (ret && !gimple_clobber_p (ret->stmt)) |
a62b3dc5 JJ |
3405 | { |
3406 | ret = NULL; | |
3407 | second = true; | |
3408 | } | |
3409 | } | |
5384a802 | 3410 | else if (ret && !gimple_clobber_p (ret->stmt)) |
a62b3dc5 JJ |
3411 | return NULL; |
3412 | if (!second) | |
3413 | ret = info; | |
f663d9ad | 3414 | } |
a62b3dc5 | 3415 | return ret; |
f663d9ad KT |
3416 | } |
3417 | ||
d7a9512e JJ |
3418 | /* Return how many SSA_NAMEs used to compute value to store in the INFO |
3419 | store have multiple uses. If any SSA_NAME has multiple uses, also | |
3420 | count statements needed to compute it. */ | |
3421 | ||
3422 | static unsigned | |
3423 | count_multiple_uses (store_immediate_info *info) | |
3424 | { | |
3425 | gimple *stmt = info->stmt; | |
3426 | unsigned ret = 0; | |
3427 | switch (info->rhs_code) | |
3428 | { | |
3429 | case INTEGER_CST: | |
e362a897 | 3430 | case STRING_CST: |
d7a9512e JJ |
3431 | return 0; |
3432 | case BIT_AND_EXPR: | |
3433 | case BIT_IOR_EXPR: | |
3434 | case BIT_XOR_EXPR: | |
d60edaba JJ |
3435 | if (info->bit_not_p) |
3436 | { | |
3437 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
3438 | ret = 1; /* Fall through below to return | |
3439 | the BIT_NOT_EXPR stmt and then | |
3440 | BIT_{AND,IOR,XOR}_EXPR and anything it | |
3441 | uses. */ | |
3442 | else | |
3443 | /* stmt is after this the BIT_NOT_EXPR. */ | |
3444 | stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3445 | } | |
d7a9512e JJ |
3446 | if (!has_single_use (gimple_assign_rhs1 (stmt))) |
3447 | { | |
3448 | ret += 1 + info->ops[0].bit_not_p; | |
3449 | if (info->ops[1].base_addr) | |
3450 | ret += 1 + info->ops[1].bit_not_p; | |
3451 | return ret + 1; | |
3452 | } | |
3453 | stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3454 | /* stmt is now the BIT_*_EXPR. */ | |
3455 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
127ef369 JJ |
3456 | ret += 1 + info->ops[info->ops_swapped_p].bit_not_p; |
3457 | else if (info->ops[info->ops_swapped_p].bit_not_p) | |
d7a9512e JJ |
3458 | { |
3459 | gimple *stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3460 | if (!has_single_use (gimple_assign_rhs1 (stmt2))) | |
3461 | ++ret; | |
3462 | } | |
3463 | if (info->ops[1].base_addr == NULL_TREE) | |
127ef369 JJ |
3464 | { |
3465 | gcc_checking_assert (!info->ops_swapped_p); | |
3466 | return ret; | |
3467 | } | |
d7a9512e | 3468 | if (!has_single_use (gimple_assign_rhs2 (stmt))) |
127ef369 JJ |
3469 | ret += 1 + info->ops[1 - info->ops_swapped_p].bit_not_p; |
3470 | else if (info->ops[1 - info->ops_swapped_p].bit_not_p) | |
d7a9512e JJ |
3471 | { |
3472 | gimple *stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs2 (stmt)); | |
3473 | if (!has_single_use (gimple_assign_rhs1 (stmt2))) | |
3474 | ++ret; | |
3475 | } | |
3476 | return ret; | |
3477 | case MEM_REF: | |
3478 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
3479 | return 1 + info->ops[0].bit_not_p; | |
3480 | else if (info->ops[0].bit_not_p) | |
3481 | { | |
3482 | stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3483 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
3484 | return 1; | |
3485 | } | |
3486 | return 0; | |
c94c3532 EB |
3487 | case BIT_INSERT_EXPR: |
3488 | return has_single_use (gimple_assign_rhs1 (stmt)) ? 0 : 1; | |
d7a9512e JJ |
3489 | default: |
3490 | gcc_unreachable (); | |
3491 | } | |
3492 | } | |
3493 | ||
f663d9ad | 3494 | /* Split a merged store described by GROUP by populating the SPLIT_STORES |
a62b3dc5 JJ |
3495 | vector (if non-NULL) with split_store structs describing the byte offset |
3496 | (from the base), the bit size and alignment of each store as well as the | |
3497 | original statements involved in each such split group. | |
f663d9ad KT |
3498 | This is to separate the splitting strategy from the statement |
3499 | building/emission/linking done in output_merged_store. | |
a62b3dc5 | 3500 | Return number of new stores. |
245f6de1 JJ |
3501 | If ALLOW_UNALIGNED_STORE is false, then all stores must be aligned. |
3502 | If ALLOW_UNALIGNED_LOAD is false, then all loads must be aligned. | |
3afd514b JJ |
3503 | BZERO_FIRST may be true only when the first store covers the whole group |
3504 | and clears it; if BZERO_FIRST is true, keep that first store in the set | |
3505 | unmodified and emit further stores for the overrides only. | |
a62b3dc5 JJ |
3506 | If SPLIT_STORES is NULL, it is just a dry run to count number of |
3507 | new stores. */ | |
f663d9ad | 3508 | |
a62b3dc5 | 3509 | static unsigned int |
245f6de1 | 3510 | split_group (merged_store_group *group, bool allow_unaligned_store, |
3afd514b | 3511 | bool allow_unaligned_load, bool bzero_first, |
99b1c316 | 3512 | vec<split_store *> *split_stores, |
d7a9512e JJ |
3513 | unsigned *total_orig, |
3514 | unsigned *total_new) | |
f663d9ad | 3515 | { |
a62b3dc5 JJ |
3516 | unsigned HOST_WIDE_INT pos = group->bitregion_start; |
3517 | unsigned HOST_WIDE_INT size = group->bitregion_end - pos; | |
f663d9ad | 3518 | unsigned HOST_WIDE_INT bytepos = pos / BITS_PER_UNIT; |
a62b3dc5 JJ |
3519 | unsigned HOST_WIDE_INT group_align = group->align; |
3520 | unsigned HOST_WIDE_INT align_base = group->align_base; | |
245f6de1 | 3521 | unsigned HOST_WIDE_INT group_load_align = group_align; |
d7a9512e | 3522 | bool any_orig = false; |
f663d9ad | 3523 | |
f663d9ad KT |
3524 | gcc_assert ((size % BITS_PER_UNIT == 0) && (pos % BITS_PER_UNIT == 0)); |
3525 | ||
e362a897 EB |
3526 | /* For bswap framework using sets of stores, all the checking has been done |
3527 | earlier in try_coalesce_bswap and the result always needs to be emitted | |
3528 | as a single store. Likewise for string concatenation, */ | |
4b84d9b8 | 3529 | if (group->stores[0]->rhs_code == LROTATE_EXPR |
e362a897 EB |
3530 | || group->stores[0]->rhs_code == NOP_EXPR |
3531 | || group->string_concatenation) | |
4b84d9b8 | 3532 | { |
3afd514b | 3533 | gcc_assert (!bzero_first); |
4b84d9b8 JJ |
3534 | if (total_orig) |
3535 | { | |
3536 | /* Avoid the old/new stmt count heuristics. It should be | |
3537 | always beneficial. */ | |
3538 | total_new[0] = 1; | |
3539 | total_orig[0] = 2; | |
3540 | } | |
3541 | ||
3542 | if (split_stores) | |
3543 | { | |
3544 | unsigned HOST_WIDE_INT align_bitpos | |
3545 | = (group->start - align_base) & (group_align - 1); | |
3546 | unsigned HOST_WIDE_INT align = group_align; | |
3547 | if (align_bitpos) | |
3548 | align = least_bit_hwi (align_bitpos); | |
3549 | bytepos = group->start / BITS_PER_UNIT; | |
99b1c316 | 3550 | split_store *store |
4b84d9b8 JJ |
3551 | = new split_store (bytepos, group->width, align); |
3552 | unsigned int first = 0; | |
3553 | find_constituent_stores (group, &store->orig_stores, | |
3554 | &first, group->start, group->width); | |
3555 | split_stores->safe_push (store); | |
3556 | } | |
3557 | ||
3558 | return 1; | |
3559 | } | |
3560 | ||
a62b3dc5 | 3561 | unsigned int ret = 0, first = 0; |
f663d9ad | 3562 | unsigned HOST_WIDE_INT try_pos = bytepos; |
f663d9ad | 3563 | |
d7a9512e JJ |
3564 | if (total_orig) |
3565 | { | |
3566 | unsigned int i; | |
3567 | store_immediate_info *info = group->stores[0]; | |
3568 | ||
3569 | total_new[0] = 0; | |
3570 | total_orig[0] = 1; /* The orig store. */ | |
3571 | info = group->stores[0]; | |
3572 | if (info->ops[0].base_addr) | |
a6fbd154 | 3573 | total_orig[0]++; |
d7a9512e | 3574 | if (info->ops[1].base_addr) |
a6fbd154 | 3575 | total_orig[0]++; |
d7a9512e JJ |
3576 | switch (info->rhs_code) |
3577 | { | |
3578 | case BIT_AND_EXPR: | |
3579 | case BIT_IOR_EXPR: | |
3580 | case BIT_XOR_EXPR: | |
3581 | total_orig[0]++; /* The orig BIT_*_EXPR stmt. */ | |
3582 | break; | |
3583 | default: | |
3584 | break; | |
3585 | } | |
3586 | total_orig[0] *= group->stores.length (); | |
3587 | ||
3588 | FOR_EACH_VEC_ELT (group->stores, i, info) | |
a6fbd154 JJ |
3589 | { |
3590 | total_new[0] += count_multiple_uses (info); | |
3591 | total_orig[0] += (info->bit_not_p | |
3592 | + info->ops[0].bit_not_p | |
3593 | + info->ops[1].bit_not_p); | |
3594 | } | |
d7a9512e JJ |
3595 | } |
3596 | ||
245f6de1 JJ |
3597 | if (!allow_unaligned_load) |
3598 | for (int i = 0; i < 2; ++i) | |
3599 | if (group->load_align[i]) | |
3600 | group_load_align = MIN (group_load_align, group->load_align[i]); | |
3601 | ||
3afd514b JJ |
3602 | if (bzero_first) |
3603 | { | |
5384a802 JJ |
3604 | store_immediate_info *gstore; |
3605 | FOR_EACH_VEC_ELT (group->stores, first, gstore) | |
3606 | if (!gimple_clobber_p (gstore->stmt)) | |
3607 | break; | |
3608 | ++first; | |
3afd514b JJ |
3609 | ret = 1; |
3610 | if (split_stores) | |
3611 | { | |
99b1c316 | 3612 | split_store *store |
5384a802 JJ |
3613 | = new split_store (bytepos, gstore->bitsize, align_base); |
3614 | store->orig_stores.safe_push (gstore); | |
3afd514b JJ |
3615 | store->orig = true; |
3616 | any_orig = true; | |
3617 | split_stores->safe_push (store); | |
3618 | } | |
3619 | } | |
3620 | ||
f663d9ad KT |
3621 | while (size > 0) |
3622 | { | |
245f6de1 | 3623 | if ((allow_unaligned_store || group_align <= BITS_PER_UNIT) |
3afd514b JJ |
3624 | && (group->mask[try_pos - bytepos] == (unsigned char) ~0U |
3625 | || (bzero_first && group->val[try_pos - bytepos] == 0))) | |
a62b3dc5 JJ |
3626 | { |
3627 | /* Skip padding bytes. */ | |
3628 | ++try_pos; | |
3629 | size -= BITS_PER_UNIT; | |
3630 | continue; | |
3631 | } | |
3632 | ||
f663d9ad | 3633 | unsigned HOST_WIDE_INT try_bitpos = try_pos * BITS_PER_UNIT; |
a62b3dc5 JJ |
3634 | unsigned int try_size = MAX_STORE_BITSIZE, nonmasked; |
3635 | unsigned HOST_WIDE_INT align_bitpos | |
3636 | = (try_bitpos - align_base) & (group_align - 1); | |
3637 | unsigned HOST_WIDE_INT align = group_align; | |
5384a802 | 3638 | bool found_orig = false; |
a62b3dc5 JJ |
3639 | if (align_bitpos) |
3640 | align = least_bit_hwi (align_bitpos); | |
245f6de1 | 3641 | if (!allow_unaligned_store) |
a62b3dc5 | 3642 | try_size = MIN (try_size, align); |
245f6de1 JJ |
3643 | if (!allow_unaligned_load) |
3644 | { | |
3645 | /* If we can't do or don't want to do unaligned stores | |
3646 | as well as loads, we need to take the loads into account | |
3647 | as well. */ | |
3648 | unsigned HOST_WIDE_INT load_align = group_load_align; | |
3649 | align_bitpos = (try_bitpos - align_base) & (load_align - 1); | |
3650 | if (align_bitpos) | |
3651 | load_align = least_bit_hwi (align_bitpos); | |
3652 | for (int i = 0; i < 2; ++i) | |
3653 | if (group->load_align[i]) | |
3654 | { | |
8a91d545 RS |
3655 | align_bitpos |
3656 | = known_alignment (try_bitpos | |
3657 | - group->stores[0]->bitpos | |
3658 | + group->stores[0]->ops[i].bitpos | |
3659 | - group->load_align_base[i]); | |
3660 | if (align_bitpos & (group_load_align - 1)) | |
245f6de1 JJ |
3661 | { |
3662 | unsigned HOST_WIDE_INT a = least_bit_hwi (align_bitpos); | |
3663 | load_align = MIN (load_align, a); | |
3664 | } | |
3665 | } | |
3666 | try_size = MIN (try_size, load_align); | |
3667 | } | |
a62b3dc5 | 3668 | store_immediate_info *info |
245f6de1 | 3669 | = find_constituent_stores (group, NULL, &first, try_bitpos, try_size); |
5384a802 | 3670 | if (info && !gimple_clobber_p (info->stmt)) |
a62b3dc5 JJ |
3671 | { |
3672 | /* If there is just one original statement for the range, see if | |
3673 | we can just reuse the original store which could be even larger | |
3674 | than try_size. */ | |
3675 | unsigned HOST_WIDE_INT stmt_end | |
3676 | = ROUND_UP (info->bitpos + info->bitsize, BITS_PER_UNIT); | |
245f6de1 JJ |
3677 | info = find_constituent_stores (group, NULL, &first, try_bitpos, |
3678 | stmt_end - try_bitpos); | |
a62b3dc5 JJ |
3679 | if (info && info->bitpos >= try_bitpos) |
3680 | { | |
5384a802 JJ |
3681 | store_immediate_info *info2 = NULL; |
3682 | unsigned int first_copy = first; | |
3683 | if (info->bitpos > try_bitpos | |
3684 | && stmt_end - try_bitpos <= try_size) | |
3685 | { | |
3686 | info2 = find_constituent_stores (group, NULL, &first_copy, | |
3687 | try_bitpos, | |
3688 | info->bitpos - try_bitpos); | |
3689 | gcc_assert (info2 == NULL || gimple_clobber_p (info2->stmt)); | |
3690 | } | |
3691 | if (info2 == NULL && stmt_end - try_bitpos < try_size) | |
3692 | { | |
3693 | info2 = find_constituent_stores (group, NULL, &first_copy, | |
3694 | stmt_end, | |
3695 | (try_bitpos + try_size) | |
3696 | - stmt_end); | |
3697 | gcc_assert (info2 == NULL || gimple_clobber_p (info2->stmt)); | |
3698 | } | |
3699 | if (info2 == NULL) | |
3700 | { | |
3701 | try_size = stmt_end - try_bitpos; | |
3702 | found_orig = true; | |
3703 | goto found; | |
3704 | } | |
a62b3dc5 JJ |
3705 | } |
3706 | } | |
f663d9ad | 3707 | |
a62b3dc5 JJ |
3708 | /* Approximate store bitsize for the case when there are no padding |
3709 | bits. */ | |
3710 | while (try_size > size) | |
3711 | try_size /= 2; | |
3712 | /* Now look for whole padding bytes at the end of that bitsize. */ | |
3713 | for (nonmasked = try_size / BITS_PER_UNIT; nonmasked > 0; --nonmasked) | |
3714 | if (group->mask[try_pos - bytepos + nonmasked - 1] | |
3afd514b JJ |
3715 | != (unsigned char) ~0U |
3716 | && (!bzero_first | |
3717 | || group->val[try_pos - bytepos + nonmasked - 1] != 0)) | |
a62b3dc5 | 3718 | break; |
5384a802 | 3719 | if (nonmasked == 0 || (info && gimple_clobber_p (info->stmt))) |
a62b3dc5 JJ |
3720 | { |
3721 | /* If entire try_size range is padding, skip it. */ | |
3722 | try_pos += try_size / BITS_PER_UNIT; | |
3723 | size -= try_size; | |
3724 | continue; | |
3725 | } | |
3726 | /* Otherwise try to decrease try_size if second half, last 3 quarters | |
3727 | etc. are padding. */ | |
3728 | nonmasked *= BITS_PER_UNIT; | |
3729 | while (nonmasked <= try_size / 2) | |
3730 | try_size /= 2; | |
245f6de1 | 3731 | if (!allow_unaligned_store && group_align > BITS_PER_UNIT) |
a62b3dc5 JJ |
3732 | { |
3733 | /* Now look for whole padding bytes at the start of that bitsize. */ | |
3734 | unsigned int try_bytesize = try_size / BITS_PER_UNIT, masked; | |
3735 | for (masked = 0; masked < try_bytesize; ++masked) | |
3afd514b JJ |
3736 | if (group->mask[try_pos - bytepos + masked] != (unsigned char) ~0U |
3737 | && (!bzero_first | |
3738 | || group->val[try_pos - bytepos + masked] != 0)) | |
a62b3dc5 JJ |
3739 | break; |
3740 | masked *= BITS_PER_UNIT; | |
3741 | gcc_assert (masked < try_size); | |
3742 | if (masked >= try_size / 2) | |
3743 | { | |
3744 | while (masked >= try_size / 2) | |
3745 | { | |
3746 | try_size /= 2; | |
3747 | try_pos += try_size / BITS_PER_UNIT; | |
3748 | size -= try_size; | |
3749 | masked -= try_size; | |
3750 | } | |
3751 | /* Need to recompute the alignment, so just retry at the new | |
3752 | position. */ | |
3753 | continue; | |
3754 | } | |
3755 | } | |
3756 | ||
3757 | found: | |
3758 | ++ret; | |
f663d9ad | 3759 | |
a62b3dc5 JJ |
3760 | if (split_stores) |
3761 | { | |
99b1c316 | 3762 | split_store *store |
a62b3dc5 | 3763 | = new split_store (try_pos, try_size, align); |
245f6de1 JJ |
3764 | info = find_constituent_stores (group, &store->orig_stores, |
3765 | &first, try_bitpos, try_size); | |
a62b3dc5 | 3766 | if (info |
5384a802 | 3767 | && !gimple_clobber_p (info->stmt) |
a62b3dc5 | 3768 | && info->bitpos >= try_bitpos |
5384a802 JJ |
3769 | && info->bitpos + info->bitsize <= try_bitpos + try_size |
3770 | && (store->orig_stores.length () == 1 | |
3771 | || found_orig | |
3772 | || (info->bitpos == try_bitpos | |
3773 | && (info->bitpos + info->bitsize | |
3774 | == try_bitpos + try_size)))) | |
d7a9512e JJ |
3775 | { |
3776 | store->orig = true; | |
3777 | any_orig = true; | |
3778 | } | |
a62b3dc5 JJ |
3779 | split_stores->safe_push (store); |
3780 | } | |
3781 | ||
3782 | try_pos += try_size / BITS_PER_UNIT; | |
f663d9ad | 3783 | size -= try_size; |
f663d9ad | 3784 | } |
a62b3dc5 | 3785 | |
d7a9512e JJ |
3786 | if (total_orig) |
3787 | { | |
a6fbd154 | 3788 | unsigned int i; |
99b1c316 | 3789 | split_store *store; |
d7a9512e JJ |
3790 | /* If we are reusing some original stores and any of the |
3791 | original SSA_NAMEs had multiple uses, we need to subtract | |
3792 | those now before we add the new ones. */ | |
3793 | if (total_new[0] && any_orig) | |
3794 | { | |
d7a9512e JJ |
3795 | FOR_EACH_VEC_ELT (*split_stores, i, store) |
3796 | if (store->orig) | |
3797 | total_new[0] -= count_multiple_uses (store->orig_stores[0]); | |
3798 | } | |
3799 | total_new[0] += ret; /* The new store. */ | |
3800 | store_immediate_info *info = group->stores[0]; | |
3801 | if (info->ops[0].base_addr) | |
a6fbd154 | 3802 | total_new[0] += ret; |
d7a9512e | 3803 | if (info->ops[1].base_addr) |
a6fbd154 | 3804 | total_new[0] += ret; |
d7a9512e JJ |
3805 | switch (info->rhs_code) |
3806 | { | |
3807 | case BIT_AND_EXPR: | |
3808 | case BIT_IOR_EXPR: | |
3809 | case BIT_XOR_EXPR: | |
3810 | total_new[0] += ret; /* The new BIT_*_EXPR stmt. */ | |
3811 | break; | |
3812 | default: | |
3813 | break; | |
3814 | } | |
a6fbd154 JJ |
3815 | FOR_EACH_VEC_ELT (*split_stores, i, store) |
3816 | { | |
3817 | unsigned int j; | |
3818 | bool bit_not_p[3] = { false, false, false }; | |
3819 | /* If all orig_stores have certain bit_not_p set, then | |
3820 | we'd use a BIT_NOT_EXPR stmt and need to account for it. | |
3821 | If some orig_stores have certain bit_not_p set, then | |
3822 | we'd use a BIT_XOR_EXPR with a mask and need to account for | |
3823 | it. */ | |
3824 | FOR_EACH_VEC_ELT (store->orig_stores, j, info) | |
3825 | { | |
3826 | if (info->ops[0].bit_not_p) | |
3827 | bit_not_p[0] = true; | |
3828 | if (info->ops[1].bit_not_p) | |
3829 | bit_not_p[1] = true; | |
3830 | if (info->bit_not_p) | |
3831 | bit_not_p[2] = true; | |
3832 | } | |
3833 | total_new[0] += bit_not_p[0] + bit_not_p[1] + bit_not_p[2]; | |
3834 | } | |
3835 | ||
d7a9512e JJ |
3836 | } |
3837 | ||
a62b3dc5 | 3838 | return ret; |
f663d9ad KT |
3839 | } |
3840 | ||
a6fbd154 JJ |
3841 | /* Return the operation through which the operand IDX (if < 2) or |
3842 | result (IDX == 2) should be inverted. If NOP_EXPR, no inversion | |
3843 | is done, if BIT_NOT_EXPR, all bits are inverted, if BIT_XOR_EXPR, | |
3844 | the bits should be xored with mask. */ | |
3845 | ||
3846 | static enum tree_code | |
3847 | invert_op (split_store *split_store, int idx, tree int_type, tree &mask) | |
3848 | { | |
3849 | unsigned int i; | |
3850 | store_immediate_info *info; | |
3851 | unsigned int cnt = 0; | |
e215422f | 3852 | bool any_paddings = false; |
a6fbd154 JJ |
3853 | FOR_EACH_VEC_ELT (split_store->orig_stores, i, info) |
3854 | { | |
3855 | bool bit_not_p = idx < 2 ? info->ops[idx].bit_not_p : info->bit_not_p; | |
3856 | if (bit_not_p) | |
e215422f JJ |
3857 | { |
3858 | ++cnt; | |
3859 | tree lhs = gimple_assign_lhs (info->stmt); | |
3860 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)) | |
3861 | && TYPE_PRECISION (TREE_TYPE (lhs)) < info->bitsize) | |
3862 | any_paddings = true; | |
3863 | } | |
a6fbd154 JJ |
3864 | } |
3865 | mask = NULL_TREE; | |
3866 | if (cnt == 0) | |
3867 | return NOP_EXPR; | |
e215422f | 3868 | if (cnt == split_store->orig_stores.length () && !any_paddings) |
a6fbd154 JJ |
3869 | return BIT_NOT_EXPR; |
3870 | ||
3871 | unsigned HOST_WIDE_INT try_bitpos = split_store->bytepos * BITS_PER_UNIT; | |
3872 | unsigned buf_size = split_store->size / BITS_PER_UNIT; | |
3873 | unsigned char *buf | |
3874 | = XALLOCAVEC (unsigned char, buf_size); | |
3875 | memset (buf, ~0U, buf_size); | |
3876 | FOR_EACH_VEC_ELT (split_store->orig_stores, i, info) | |
3877 | { | |
3878 | bool bit_not_p = idx < 2 ? info->ops[idx].bit_not_p : info->bit_not_p; | |
3879 | if (!bit_not_p) | |
3880 | continue; | |
3881 | /* Clear regions with bit_not_p and invert afterwards, rather than | |
3882 | clear regions with !bit_not_p, so that gaps in between stores aren't | |
3883 | set in the mask. */ | |
3884 | unsigned HOST_WIDE_INT bitsize = info->bitsize; | |
e215422f | 3885 | unsigned HOST_WIDE_INT prec = bitsize; |
a6fbd154 | 3886 | unsigned int pos_in_buffer = 0; |
e215422f JJ |
3887 | if (any_paddings) |
3888 | { | |
3889 | tree lhs = gimple_assign_lhs (info->stmt); | |
3890 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)) | |
3891 | && TYPE_PRECISION (TREE_TYPE (lhs)) < bitsize) | |
3892 | prec = TYPE_PRECISION (TREE_TYPE (lhs)); | |
3893 | } | |
a6fbd154 JJ |
3894 | if (info->bitpos < try_bitpos) |
3895 | { | |
3896 | gcc_assert (info->bitpos + bitsize > try_bitpos); | |
e215422f JJ |
3897 | if (!BYTES_BIG_ENDIAN) |
3898 | { | |
3899 | if (prec <= try_bitpos - info->bitpos) | |
3900 | continue; | |
3901 | prec -= try_bitpos - info->bitpos; | |
3902 | } | |
3903 | bitsize -= try_bitpos - info->bitpos; | |
3904 | if (BYTES_BIG_ENDIAN && prec > bitsize) | |
3905 | prec = bitsize; | |
a6fbd154 JJ |
3906 | } |
3907 | else | |
3908 | pos_in_buffer = info->bitpos - try_bitpos; | |
e215422f JJ |
3909 | if (prec < bitsize) |
3910 | { | |
3911 | /* If this is a bool inversion, invert just the least significant | |
3912 | prec bits rather than all bits of it. */ | |
3913 | if (BYTES_BIG_ENDIAN) | |
3914 | { | |
3915 | pos_in_buffer += bitsize - prec; | |
3916 | if (pos_in_buffer >= split_store->size) | |
3917 | continue; | |
3918 | } | |
3919 | bitsize = prec; | |
3920 | } | |
a6fbd154 JJ |
3921 | if (pos_in_buffer + bitsize > split_store->size) |
3922 | bitsize = split_store->size - pos_in_buffer; | |
3923 | unsigned char *p = buf + (pos_in_buffer / BITS_PER_UNIT); | |
3924 | if (BYTES_BIG_ENDIAN) | |
3925 | clear_bit_region_be (p, (BITS_PER_UNIT - 1 | |
3926 | - (pos_in_buffer % BITS_PER_UNIT)), bitsize); | |
3927 | else | |
3928 | clear_bit_region (p, pos_in_buffer % BITS_PER_UNIT, bitsize); | |
3929 | } | |
3930 | for (unsigned int i = 0; i < buf_size; ++i) | |
3931 | buf[i] = ~buf[i]; | |
3932 | mask = native_interpret_expr (int_type, buf, buf_size); | |
3933 | return BIT_XOR_EXPR; | |
3934 | } | |
3935 | ||
f663d9ad KT |
3936 | /* Given a merged store group GROUP output the widened version of it. |
3937 | The store chain is against the base object BASE. | |
3938 | Try store sizes of at most MAX_STORE_BITSIZE bits wide and don't output | |
3939 | unaligned stores for STRICT_ALIGNMENT targets or if it's too expensive. | |
3940 | Make sure that the number of statements output is less than the number of | |
3941 | original statements. If a better sequence is possible emit it and | |
3942 | return true. */ | |
3943 | ||
3944 | bool | |
b5926e23 | 3945 | imm_store_chain_info::output_merged_store (merged_store_group *group) |
f663d9ad | 3946 | { |
e362a897 | 3947 | const unsigned HOST_WIDE_INT start_byte_pos |
a62b3dc5 | 3948 | = group->bitregion_start / BITS_PER_UNIT; |
f663d9ad KT |
3949 | unsigned int orig_num_stmts = group->stores.length (); |
3950 | if (orig_num_stmts < 2) | |
3951 | return false; | |
3952 | ||
245f6de1 | 3953 | bool allow_unaligned_store |
028d4092 | 3954 | = !STRICT_ALIGNMENT && param_store_merging_allow_unaligned; |
245f6de1 | 3955 | bool allow_unaligned_load = allow_unaligned_store; |
3afd514b | 3956 | bool bzero_first = false; |
5384a802 JJ |
3957 | store_immediate_info *store; |
3958 | unsigned int num_clobber_stmts = 0; | |
3959 | if (group->stores[0]->rhs_code == INTEGER_CST) | |
3960 | { | |
e362a897 | 3961 | unsigned int i; |
5384a802 JJ |
3962 | FOR_EACH_VEC_ELT (group->stores, i, store) |
3963 | if (gimple_clobber_p (store->stmt)) | |
3964 | num_clobber_stmts++; | |
3965 | else if (TREE_CODE (gimple_assign_rhs1 (store->stmt)) == CONSTRUCTOR | |
3966 | && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (store->stmt)) == 0 | |
3967 | && group->start == store->bitpos | |
3968 | && group->width == store->bitsize | |
3969 | && (group->start % BITS_PER_UNIT) == 0 | |
3970 | && (group->width % BITS_PER_UNIT) == 0) | |
3971 | { | |
3972 | bzero_first = true; | |
3973 | break; | |
3974 | } | |
3975 | else | |
3976 | break; | |
3977 | FOR_EACH_VEC_ELT_FROM (group->stores, i, store, i) | |
3978 | if (gimple_clobber_p (store->stmt)) | |
3979 | num_clobber_stmts++; | |
3980 | if (num_clobber_stmts == orig_num_stmts) | |
3981 | return false; | |
3982 | orig_num_stmts -= num_clobber_stmts; | |
3983 | } | |
3afd514b | 3984 | if (allow_unaligned_store || bzero_first) |
a62b3dc5 JJ |
3985 | { |
3986 | /* If unaligned stores are allowed, see how many stores we'd emit | |
3987 | for unaligned and how many stores we'd emit for aligned stores. | |
3afd514b JJ |
3988 | Only use unaligned stores if it allows fewer stores than aligned. |
3989 | Similarly, if there is a whole region clear first, prefer expanding | |
3990 | it together compared to expanding clear first followed by merged | |
3991 | further stores. */ | |
21f65995 | 3992 | unsigned cnt[4] = { ~0U, ~0U, ~0U, ~0U }; |
3afd514b JJ |
3993 | int pass_min = 0; |
3994 | for (int pass = 0; pass < 4; ++pass) | |
3995 | { | |
3996 | if (!allow_unaligned_store && (pass & 1) != 0) | |
3997 | continue; | |
3998 | if (!bzero_first && (pass & 2) != 0) | |
3999 | continue; | |
4000 | cnt[pass] = split_group (group, (pass & 1) != 0, | |
4001 | allow_unaligned_load, (pass & 2) != 0, | |
4002 | NULL, NULL, NULL); | |
4003 | if (cnt[pass] < cnt[pass_min]) | |
4004 | pass_min = pass; | |
4005 | } | |
4006 | if ((pass_min & 1) == 0) | |
245f6de1 | 4007 | allow_unaligned_store = false; |
3afd514b JJ |
4008 | if ((pass_min & 2) == 0) |
4009 | bzero_first = false; | |
a62b3dc5 | 4010 | } |
e362a897 EB |
4011 | |
4012 | auto_vec<class split_store *, 32> split_stores; | |
4013 | split_store *split_store; | |
4014 | unsigned total_orig, total_new, i; | |
3afd514b | 4015 | split_group (group, allow_unaligned_store, allow_unaligned_load, bzero_first, |
d7a9512e | 4016 | &split_stores, &total_orig, &total_new); |
a62b3dc5 | 4017 | |
5384a802 JJ |
4018 | /* Determine if there is a clobber covering the whole group at the start, |
4019 | followed by proposed split stores that cover the whole group. In that | |
4020 | case, prefer the transformation even if | |
4021 | split_stores.length () == orig_num_stmts. */ | |
4022 | bool clobber_first = false; | |
4023 | if (num_clobber_stmts | |
4024 | && gimple_clobber_p (group->stores[0]->stmt) | |
4025 | && group->start == group->stores[0]->bitpos | |
4026 | && group->width == group->stores[0]->bitsize | |
4027 | && (group->start % BITS_PER_UNIT) == 0 | |
4028 | && (group->width % BITS_PER_UNIT) == 0) | |
4029 | { | |
4030 | clobber_first = true; | |
4031 | unsigned HOST_WIDE_INT pos = group->start / BITS_PER_UNIT; | |
4032 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
4033 | if (split_store->bytepos != pos) | |
4034 | { | |
4035 | clobber_first = false; | |
4036 | break; | |
4037 | } | |
4038 | else | |
4039 | pos += split_store->size / BITS_PER_UNIT; | |
4040 | if (pos != (group->start + group->width) / BITS_PER_UNIT) | |
4041 | clobber_first = false; | |
4042 | } | |
4043 | ||
4044 | if (split_stores.length () >= orig_num_stmts + clobber_first) | |
a62b3dc5 | 4045 | { |
5384a802 | 4046 | |
a62b3dc5 JJ |
4047 | /* We didn't manage to reduce the number of statements. Bail out. */ |
4048 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
d7a9512e JJ |
4049 | fprintf (dump_file, "Exceeded original number of stmts (%u)." |
4050 | " Not profitable to emit new sequence.\n", | |
4051 | orig_num_stmts); | |
dd172744 RB |
4052 | FOR_EACH_VEC_ELT (split_stores, i, split_store) |
4053 | delete split_store; | |
a62b3dc5 JJ |
4054 | return false; |
4055 | } | |
d7a9512e JJ |
4056 | if (total_orig <= total_new) |
4057 | { | |
4058 | /* If number of estimated new statements is above estimated original | |
4059 | statements, bail out too. */ | |
4060 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4061 | fprintf (dump_file, "Estimated number of original stmts (%u)" | |
4062 | " not larger than estimated number of new" | |
4063 | " stmts (%u).\n", | |
4064 | total_orig, total_new); | |
dd172744 RB |
4065 | FOR_EACH_VEC_ELT (split_stores, i, split_store) |
4066 | delete split_store; | |
4b84d9b8 | 4067 | return false; |
d7a9512e | 4068 | } |
5384a802 JJ |
4069 | if (group->stores[0]->rhs_code == INTEGER_CST) |
4070 | { | |
4071 | bool all_orig = true; | |
4072 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
4073 | if (!split_store->orig) | |
4074 | { | |
4075 | all_orig = false; | |
4076 | break; | |
4077 | } | |
4078 | if (all_orig) | |
4079 | { | |
4080 | unsigned int cnt = split_stores.length (); | |
4081 | store_immediate_info *store; | |
4082 | FOR_EACH_VEC_ELT (group->stores, i, store) | |
4083 | if (gimple_clobber_p (store->stmt)) | |
4084 | ++cnt; | |
4085 | /* Punt if we wouldn't make any real changes, i.e. keep all | |
4086 | orig stmts + all clobbers. */ | |
4087 | if (cnt == group->stores.length ()) | |
4088 | { | |
4089 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4090 | fprintf (dump_file, "Exceeded original number of stmts (%u)." | |
4091 | " Not profitable to emit new sequence.\n", | |
4092 | orig_num_stmts); | |
4093 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
4094 | delete split_store; | |
4095 | return false; | |
4096 | } | |
4097 | } | |
4098 | } | |
f663d9ad KT |
4099 | |
4100 | gimple_stmt_iterator last_gsi = gsi_for_stmt (group->last_stmt); | |
4101 | gimple_seq seq = NULL; | |
f663d9ad KT |
4102 | tree last_vdef, new_vuse; |
4103 | last_vdef = gimple_vdef (group->last_stmt); | |
4104 | new_vuse = gimple_vuse (group->last_stmt); | |
4b84d9b8 JJ |
4105 | tree bswap_res = NULL_TREE; |
4106 | ||
5384a802 JJ |
4107 | /* Clobbers are not removed. */ |
4108 | if (gimple_clobber_p (group->last_stmt)) | |
4109 | { | |
4110 | new_vuse = make_ssa_name (gimple_vop (cfun), group->last_stmt); | |
4111 | gimple_set_vdef (group->last_stmt, new_vuse); | |
4112 | } | |
4113 | ||
4b84d9b8 JJ |
4114 | if (group->stores[0]->rhs_code == LROTATE_EXPR |
4115 | || group->stores[0]->rhs_code == NOP_EXPR) | |
4116 | { | |
4117 | tree fndecl = NULL_TREE, bswap_type = NULL_TREE, load_type; | |
4118 | gimple *ins_stmt = group->stores[0]->ins_stmt; | |
4119 | struct symbolic_number *n = &group->stores[0]->n; | |
4120 | bool bswap = group->stores[0]->rhs_code == LROTATE_EXPR; | |
4121 | ||
4122 | switch (n->range) | |
4123 | { | |
4124 | case 16: | |
4125 | load_type = bswap_type = uint16_type_node; | |
4126 | break; | |
4127 | case 32: | |
4128 | load_type = uint32_type_node; | |
4129 | if (bswap) | |
4130 | { | |
4131 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
4132 | bswap_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
4133 | } | |
4134 | break; | |
4135 | case 64: | |
4136 | load_type = uint64_type_node; | |
4137 | if (bswap) | |
4138 | { | |
4139 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
4140 | bswap_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
4141 | } | |
4142 | break; | |
4143 | default: | |
4144 | gcc_unreachable (); | |
4145 | } | |
4146 | ||
4147 | /* If the loads have each vuse of the corresponding store, | |
4148 | we've checked the aliasing already in try_coalesce_bswap and | |
4149 | we want to sink the need load into seq. So need to use new_vuse | |
4150 | on the load. */ | |
30fa8e9c | 4151 | if (n->base_addr) |
4b84d9b8 | 4152 | { |
30fa8e9c JJ |
4153 | if (n->vuse == NULL) |
4154 | { | |
4155 | n->vuse = new_vuse; | |
4156 | ins_stmt = NULL; | |
4157 | } | |
4158 | else | |
4159 | /* Update vuse in case it has changed by output_merged_stores. */ | |
4160 | n->vuse = gimple_vuse (ins_stmt); | |
4b84d9b8 JJ |
4161 | } |
4162 | bswap_res = bswap_replace (gsi_start (seq), ins_stmt, fndecl, | |
4163 | bswap_type, load_type, n, bswap); | |
4164 | gcc_assert (bswap_res); | |
4165 | } | |
f663d9ad KT |
4166 | |
4167 | gimple *stmt = NULL; | |
245f6de1 | 4168 | auto_vec<gimple *, 32> orig_stmts; |
4b84d9b8 JJ |
4169 | gimple_seq this_seq; |
4170 | tree addr = force_gimple_operand_1 (unshare_expr (base_addr), &this_seq, | |
aa55dc0c | 4171 | is_gimple_mem_ref_addr, NULL_TREE); |
4b84d9b8 | 4172 | gimple_seq_add_seq_without_update (&seq, this_seq); |
245f6de1 JJ |
4173 | |
4174 | tree load_addr[2] = { NULL_TREE, NULL_TREE }; | |
4175 | gimple_seq load_seq[2] = { NULL, NULL }; | |
4176 | gimple_stmt_iterator load_gsi[2] = { gsi_none (), gsi_none () }; | |
4177 | for (int j = 0; j < 2; ++j) | |
4178 | { | |
4179 | store_operand_info &op = group->stores[0]->ops[j]; | |
4180 | if (op.base_addr == NULL_TREE) | |
4181 | continue; | |
4182 | ||
4183 | store_immediate_info *infol = group->stores.last (); | |
4184 | if (gimple_vuse (op.stmt) == gimple_vuse (infol->ops[j].stmt)) | |
4185 | { | |
97031af7 JJ |
4186 | /* We can't pick the location randomly; while we've verified |
4187 | all the loads have the same vuse, they can be still in different | |
4188 | basic blocks and we need to pick the one from the last bb: | |
4189 | int x = q[0]; | |
4190 | if (x == N) return; | |
4191 | int y = q[1]; | |
4192 | p[0] = x; | |
4193 | p[1] = y; | |
4194 | otherwise if we put the wider load at the q[0] load, we might | |
4195 | segfault if q[1] is not mapped. */ | |
4196 | basic_block bb = gimple_bb (op.stmt); | |
4197 | gimple *ostmt = op.stmt; | |
4198 | store_immediate_info *info; | |
4199 | FOR_EACH_VEC_ELT (group->stores, i, info) | |
4200 | { | |
4201 | gimple *tstmt = info->ops[j].stmt; | |
4202 | basic_block tbb = gimple_bb (tstmt); | |
4203 | if (dominated_by_p (CDI_DOMINATORS, tbb, bb)) | |
4204 | { | |
4205 | ostmt = tstmt; | |
4206 | bb = tbb; | |
4207 | } | |
4208 | } | |
4209 | load_gsi[j] = gsi_for_stmt (ostmt); | |
245f6de1 JJ |
4210 | load_addr[j] |
4211 | = force_gimple_operand_1 (unshare_expr (op.base_addr), | |
4212 | &load_seq[j], is_gimple_mem_ref_addr, | |
4213 | NULL_TREE); | |
4214 | } | |
4215 | else if (operand_equal_p (base_addr, op.base_addr, 0)) | |
4216 | load_addr[j] = addr; | |
4217 | else | |
3e2927a1 | 4218 | { |
3e2927a1 JJ |
4219 | load_addr[j] |
4220 | = force_gimple_operand_1 (unshare_expr (op.base_addr), | |
4221 | &this_seq, is_gimple_mem_ref_addr, | |
4222 | NULL_TREE); | |
4223 | gimple_seq_add_seq_without_update (&seq, this_seq); | |
4224 | } | |
245f6de1 JJ |
4225 | } |
4226 | ||
f663d9ad KT |
4227 | FOR_EACH_VEC_ELT (split_stores, i, split_store) |
4228 | { | |
e362a897 EB |
4229 | const unsigned HOST_WIDE_INT try_size = split_store->size; |
4230 | const unsigned HOST_WIDE_INT try_pos = split_store->bytepos; | |
4231 | const unsigned HOST_WIDE_INT try_bitpos = try_pos * BITS_PER_UNIT; | |
4232 | const unsigned HOST_WIDE_INT try_align = split_store->align; | |
4233 | const unsigned HOST_WIDE_INT try_offset = try_pos - start_byte_pos; | |
a62b3dc5 JJ |
4234 | tree dest, src; |
4235 | location_t loc; | |
e362a897 | 4236 | |
a62b3dc5 JJ |
4237 | if (split_store->orig) |
4238 | { | |
5384a802 JJ |
4239 | /* If there is just a single non-clobber constituent store |
4240 | which covers the whole area, just reuse the lhs and rhs. */ | |
4241 | gimple *orig_stmt = NULL; | |
4242 | store_immediate_info *store; | |
4243 | unsigned int j; | |
4244 | FOR_EACH_VEC_ELT (split_store->orig_stores, j, store) | |
4245 | if (!gimple_clobber_p (store->stmt)) | |
4246 | { | |
4247 | orig_stmt = store->stmt; | |
4248 | break; | |
4249 | } | |
245f6de1 JJ |
4250 | dest = gimple_assign_lhs (orig_stmt); |
4251 | src = gimple_assign_rhs1 (orig_stmt); | |
4252 | loc = gimple_location (orig_stmt); | |
a62b3dc5 JJ |
4253 | } |
4254 | else | |
4255 | { | |
245f6de1 JJ |
4256 | store_immediate_info *info; |
4257 | unsigned short clique, base; | |
4258 | unsigned int k; | |
4259 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4260 | orig_stmts.safe_push (info->stmt); | |
a62b3dc5 | 4261 | tree offset_type |
245f6de1 | 4262 | = get_alias_type_for_stmts (orig_stmts, false, &clique, &base); |
e362a897 | 4263 | tree dest_type; |
245f6de1 JJ |
4264 | loc = get_location_for_stmts (orig_stmts); |
4265 | orig_stmts.truncate (0); | |
a62b3dc5 | 4266 | |
e362a897 EB |
4267 | if (group->string_concatenation) |
4268 | dest_type | |
4269 | = build_array_type_nelts (char_type_node, | |
4270 | try_size / BITS_PER_UNIT); | |
4271 | else | |
4272 | { | |
4273 | dest_type = build_nonstandard_integer_type (try_size, UNSIGNED); | |
4274 | dest_type = build_aligned_type (dest_type, try_align); | |
4275 | } | |
4276 | dest = fold_build2 (MEM_REF, dest_type, addr, | |
a62b3dc5 | 4277 | build_int_cst (offset_type, try_pos)); |
245f6de1 JJ |
4278 | if (TREE_CODE (dest) == MEM_REF) |
4279 | { | |
4280 | MR_DEPENDENCE_CLIQUE (dest) = clique; | |
4281 | MR_DEPENDENCE_BASE (dest) = base; | |
4282 | } | |
4283 | ||
c94c3532 | 4284 | tree mask; |
e362a897 | 4285 | if (bswap_res || group->string_concatenation) |
c94c3532 EB |
4286 | mask = integer_zero_node; |
4287 | else | |
e362a897 EB |
4288 | mask = native_interpret_expr (dest_type, |
4289 | group->mask + try_offset, | |
4b84d9b8 | 4290 | group->buf_size); |
245f6de1 JJ |
4291 | |
4292 | tree ops[2]; | |
4293 | for (int j = 0; | |
4294 | j < 1 + (split_store->orig_stores[0]->ops[1].val != NULL_TREE); | |
4295 | ++j) | |
4296 | { | |
4297 | store_operand_info &op = split_store->orig_stores[0]->ops[j]; | |
4b84d9b8 JJ |
4298 | if (bswap_res) |
4299 | ops[j] = bswap_res; | |
e362a897 EB |
4300 | else if (group->string_concatenation) |
4301 | { | |
4302 | ops[j] = build_string (try_size / BITS_PER_UNIT, | |
4303 | (const char *) group->val + try_offset); | |
4304 | TREE_TYPE (ops[j]) = dest_type; | |
4305 | } | |
4b84d9b8 | 4306 | else if (op.base_addr) |
245f6de1 JJ |
4307 | { |
4308 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4309 | orig_stmts.safe_push (info->ops[j].stmt); | |
4310 | ||
4311 | offset_type = get_alias_type_for_stmts (orig_stmts, true, | |
4312 | &clique, &base); | |
4313 | location_t load_loc = get_location_for_stmts (orig_stmts); | |
4314 | orig_stmts.truncate (0); | |
4315 | ||
4316 | unsigned HOST_WIDE_INT load_align = group->load_align[j]; | |
4317 | unsigned HOST_WIDE_INT align_bitpos | |
c94c3532 | 4318 | = known_alignment (try_bitpos |
8a91d545 RS |
4319 | - split_store->orig_stores[0]->bitpos |
4320 | + op.bitpos); | |
4321 | if (align_bitpos & (load_align - 1)) | |
245f6de1 JJ |
4322 | load_align = least_bit_hwi (align_bitpos); |
4323 | ||
4324 | tree load_int_type | |
4325 | = build_nonstandard_integer_type (try_size, UNSIGNED); | |
4326 | load_int_type | |
4327 | = build_aligned_type (load_int_type, load_align); | |
4328 | ||
8a91d545 | 4329 | poly_uint64 load_pos |
c94c3532 | 4330 | = exact_div (try_bitpos |
8a91d545 RS |
4331 | - split_store->orig_stores[0]->bitpos |
4332 | + op.bitpos, | |
4333 | BITS_PER_UNIT); | |
245f6de1 JJ |
4334 | ops[j] = fold_build2 (MEM_REF, load_int_type, load_addr[j], |
4335 | build_int_cst (offset_type, load_pos)); | |
4336 | if (TREE_CODE (ops[j]) == MEM_REF) | |
4337 | { | |
4338 | MR_DEPENDENCE_CLIQUE (ops[j]) = clique; | |
4339 | MR_DEPENDENCE_BASE (ops[j]) = base; | |
4340 | } | |
4341 | if (!integer_zerop (mask)) | |
4342 | /* The load might load some bits (that will be masked off | |
4343 | later on) uninitialized, avoid -W*uninitialized | |
4344 | warnings in that case. */ | |
4345 | TREE_NO_WARNING (ops[j]) = 1; | |
4346 | ||
e362a897 | 4347 | stmt = gimple_build_assign (make_ssa_name (dest_type), ops[j]); |
245f6de1 JJ |
4348 | gimple_set_location (stmt, load_loc); |
4349 | if (gsi_bb (load_gsi[j])) | |
4350 | { | |
4351 | gimple_set_vuse (stmt, gimple_vuse (op.stmt)); | |
4352 | gimple_seq_add_stmt_without_update (&load_seq[j], stmt); | |
4353 | } | |
4354 | else | |
4355 | { | |
4356 | gimple_set_vuse (stmt, new_vuse); | |
4357 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4358 | } | |
4359 | ops[j] = gimple_assign_lhs (stmt); | |
a6fbd154 JJ |
4360 | tree xor_mask; |
4361 | enum tree_code inv_op | |
e362a897 | 4362 | = invert_op (split_store, j, dest_type, xor_mask); |
a6fbd154 | 4363 | if (inv_op != NOP_EXPR) |
383ac8dc | 4364 | { |
e362a897 | 4365 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
a6fbd154 | 4366 | inv_op, ops[j], xor_mask); |
383ac8dc JJ |
4367 | gimple_set_location (stmt, load_loc); |
4368 | ops[j] = gimple_assign_lhs (stmt); | |
4369 | ||
4370 | if (gsi_bb (load_gsi[j])) | |
4371 | gimple_seq_add_stmt_without_update (&load_seq[j], | |
4372 | stmt); | |
4373 | else | |
4374 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4375 | } | |
245f6de1 JJ |
4376 | } |
4377 | else | |
e362a897 EB |
4378 | ops[j] = native_interpret_expr (dest_type, |
4379 | group->val + try_offset, | |
245f6de1 JJ |
4380 | group->buf_size); |
4381 | } | |
4382 | ||
4383 | switch (split_store->orig_stores[0]->rhs_code) | |
4384 | { | |
4385 | case BIT_AND_EXPR: | |
4386 | case BIT_IOR_EXPR: | |
4387 | case BIT_XOR_EXPR: | |
4388 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4389 | { | |
4390 | tree rhs1 = gimple_assign_rhs1 (info->stmt); | |
4391 | orig_stmts.safe_push (SSA_NAME_DEF_STMT (rhs1)); | |
4392 | } | |
4393 | location_t bit_loc; | |
4394 | bit_loc = get_location_for_stmts (orig_stmts); | |
4395 | orig_stmts.truncate (0); | |
4396 | ||
4397 | stmt | |
e362a897 | 4398 | = gimple_build_assign (make_ssa_name (dest_type), |
245f6de1 JJ |
4399 | split_store->orig_stores[0]->rhs_code, |
4400 | ops[0], ops[1]); | |
4401 | gimple_set_location (stmt, bit_loc); | |
4402 | /* If there is just one load and there is a separate | |
4403 | load_seq[0], emit the bitwise op right after it. */ | |
4404 | if (load_addr[1] == NULL_TREE && gsi_bb (load_gsi[0])) | |
4405 | gimple_seq_add_stmt_without_update (&load_seq[0], stmt); | |
4406 | /* Otherwise, if at least one load is in seq, we need to | |
4407 | emit the bitwise op right before the store. If there | |
4408 | are two loads and are emitted somewhere else, it would | |
4409 | be better to emit the bitwise op as early as possible; | |
4410 | we don't track where that would be possible right now | |
4411 | though. */ | |
4412 | else | |
4413 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4414 | src = gimple_assign_lhs (stmt); | |
a6fbd154 JJ |
4415 | tree xor_mask; |
4416 | enum tree_code inv_op; | |
e362a897 | 4417 | inv_op = invert_op (split_store, 2, dest_type, xor_mask); |
a6fbd154 | 4418 | if (inv_op != NOP_EXPR) |
d60edaba | 4419 | { |
e362a897 | 4420 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
a6fbd154 | 4421 | inv_op, src, xor_mask); |
d60edaba JJ |
4422 | gimple_set_location (stmt, bit_loc); |
4423 | if (load_addr[1] == NULL_TREE && gsi_bb (load_gsi[0])) | |
4424 | gimple_seq_add_stmt_without_update (&load_seq[0], stmt); | |
4425 | else | |
4426 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4427 | src = gimple_assign_lhs (stmt); | |
4428 | } | |
245f6de1 | 4429 | break; |
4b84d9b8 JJ |
4430 | case LROTATE_EXPR: |
4431 | case NOP_EXPR: | |
4432 | src = ops[0]; | |
4433 | if (!is_gimple_val (src)) | |
4434 | { | |
4435 | stmt = gimple_build_assign (make_ssa_name (TREE_TYPE (src)), | |
4436 | src); | |
4437 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4438 | src = gimple_assign_lhs (stmt); | |
4439 | } | |
e362a897 | 4440 | if (!useless_type_conversion_p (dest_type, TREE_TYPE (src))) |
4b84d9b8 | 4441 | { |
e362a897 | 4442 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
4b84d9b8 JJ |
4443 | NOP_EXPR, src); |
4444 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4445 | src = gimple_assign_lhs (stmt); | |
4446 | } | |
e362a897 | 4447 | inv_op = invert_op (split_store, 2, dest_type, xor_mask); |
be52ac73 JJ |
4448 | if (inv_op != NOP_EXPR) |
4449 | { | |
e362a897 | 4450 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
be52ac73 JJ |
4451 | inv_op, src, xor_mask); |
4452 | gimple_set_location (stmt, loc); | |
4453 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4454 | src = gimple_assign_lhs (stmt); | |
4455 | } | |
4b84d9b8 | 4456 | break; |
245f6de1 JJ |
4457 | default: |
4458 | src = ops[0]; | |
4459 | break; | |
4460 | } | |
4461 | ||
c94c3532 EB |
4462 | /* If bit insertion is required, we use the source as an accumulator |
4463 | into which the successive bit-field values are manually inserted. | |
4464 | FIXME: perhaps use BIT_INSERT_EXPR instead in some cases? */ | |
4465 | if (group->bit_insertion) | |
4466 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4467 | if (info->rhs_code == BIT_INSERT_EXPR | |
4468 | && info->bitpos < try_bitpos + try_size | |
4469 | && info->bitpos + info->bitsize > try_bitpos) | |
4470 | { | |
4471 | /* Mask, truncate, convert to final type, shift and ior into | |
4472 | the accumulator. Note that every step can be a no-op. */ | |
4473 | const HOST_WIDE_INT start_gap = info->bitpos - try_bitpos; | |
4474 | const HOST_WIDE_INT end_gap | |
4475 | = (try_bitpos + try_size) - (info->bitpos + info->bitsize); | |
4476 | tree tem = info->ops[0].val; | |
ed01d707 EB |
4477 | if (!INTEGRAL_TYPE_P (TREE_TYPE (tem))) |
4478 | { | |
4479 | const unsigned HOST_WIDE_INT size | |
4480 | = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (tem))); | |
4481 | tree integer_type | |
4482 | = build_nonstandard_integer_type (size, UNSIGNED); | |
4483 | tem = gimple_build (&seq, loc, VIEW_CONVERT_EXPR, | |
4484 | integer_type, tem); | |
4485 | } | |
c14add82 EB |
4486 | if (TYPE_PRECISION (TREE_TYPE (tem)) <= info->bitsize) |
4487 | { | |
4488 | tree bitfield_type | |
4489 | = build_nonstandard_integer_type (info->bitsize, | |
4490 | UNSIGNED); | |
4491 | tem = gimple_convert (&seq, loc, bitfield_type, tem); | |
4492 | } | |
4493 | else if ((BYTES_BIG_ENDIAN ? start_gap : end_gap) > 0) | |
c94c3532 EB |
4494 | { |
4495 | const unsigned HOST_WIDE_INT imask | |
4496 | = (HOST_WIDE_INT_1U << info->bitsize) - 1; | |
4497 | tem = gimple_build (&seq, loc, | |
4498 | BIT_AND_EXPR, TREE_TYPE (tem), tem, | |
4499 | build_int_cst (TREE_TYPE (tem), | |
4500 | imask)); | |
4501 | } | |
4502 | const HOST_WIDE_INT shift | |
4503 | = (BYTES_BIG_ENDIAN ? end_gap : start_gap); | |
4504 | if (shift < 0) | |
4505 | tem = gimple_build (&seq, loc, | |
4506 | RSHIFT_EXPR, TREE_TYPE (tem), tem, | |
4507 | build_int_cst (NULL_TREE, -shift)); | |
e362a897 | 4508 | tem = gimple_convert (&seq, loc, dest_type, tem); |
c94c3532 EB |
4509 | if (shift > 0) |
4510 | tem = gimple_build (&seq, loc, | |
e362a897 | 4511 | LSHIFT_EXPR, dest_type, tem, |
c94c3532 EB |
4512 | build_int_cst (NULL_TREE, shift)); |
4513 | src = gimple_build (&seq, loc, | |
e362a897 | 4514 | BIT_IOR_EXPR, dest_type, tem, src); |
c94c3532 EB |
4515 | } |
4516 | ||
a62b3dc5 JJ |
4517 | if (!integer_zerop (mask)) |
4518 | { | |
e362a897 | 4519 | tree tem = make_ssa_name (dest_type); |
a62b3dc5 JJ |
4520 | tree load_src = unshare_expr (dest); |
4521 | /* The load might load some or all bits uninitialized, | |
4522 | avoid -W*uninitialized warnings in that case. | |
4523 | As optimization, it would be nice if all the bits are | |
4524 | provably uninitialized (no stores at all yet or previous | |
4525 | store a CLOBBER) we'd optimize away the load and replace | |
4526 | it e.g. with 0. */ | |
4527 | TREE_NO_WARNING (load_src) = 1; | |
4528 | stmt = gimple_build_assign (tem, load_src); | |
4529 | gimple_set_location (stmt, loc); | |
4530 | gimple_set_vuse (stmt, new_vuse); | |
4531 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4532 | ||
4533 | /* FIXME: If there is a single chunk of zero bits in mask, | |
4534 | perhaps use BIT_INSERT_EXPR instead? */ | |
e362a897 | 4535 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
a62b3dc5 JJ |
4536 | BIT_AND_EXPR, tem, mask); |
4537 | gimple_set_location (stmt, loc); | |
4538 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4539 | tem = gimple_assign_lhs (stmt); | |
4540 | ||
245f6de1 | 4541 | if (TREE_CODE (src) == INTEGER_CST) |
e362a897 | 4542 | src = wide_int_to_tree (dest_type, |
245f6de1 JJ |
4543 | wi::bit_and_not (wi::to_wide (src), |
4544 | wi::to_wide (mask))); | |
4545 | else | |
4546 | { | |
4547 | tree nmask | |
e362a897 | 4548 | = wide_int_to_tree (dest_type, |
245f6de1 | 4549 | wi::bit_not (wi::to_wide (mask))); |
e362a897 | 4550 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
245f6de1 JJ |
4551 | BIT_AND_EXPR, src, nmask); |
4552 | gimple_set_location (stmt, loc); | |
4553 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4554 | src = gimple_assign_lhs (stmt); | |
4555 | } | |
e362a897 | 4556 | stmt = gimple_build_assign (make_ssa_name (dest_type), |
a62b3dc5 JJ |
4557 | BIT_IOR_EXPR, tem, src); |
4558 | gimple_set_location (stmt, loc); | |
4559 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4560 | src = gimple_assign_lhs (stmt); | |
4561 | } | |
4562 | } | |
f663d9ad KT |
4563 | |
4564 | stmt = gimple_build_assign (dest, src); | |
4565 | gimple_set_location (stmt, loc); | |
4566 | gimple_set_vuse (stmt, new_vuse); | |
4567 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4568 | ||
629387a6 EB |
4569 | if (group->lp_nr && stmt_could_throw_p (cfun, stmt)) |
4570 | add_stmt_to_eh_lp (stmt, group->lp_nr); | |
4571 | ||
f663d9ad KT |
4572 | tree new_vdef; |
4573 | if (i < split_stores.length () - 1) | |
a62b3dc5 | 4574 | new_vdef = make_ssa_name (gimple_vop (cfun), stmt); |
f663d9ad KT |
4575 | else |
4576 | new_vdef = last_vdef; | |
4577 | ||
4578 | gimple_set_vdef (stmt, new_vdef); | |
4579 | SSA_NAME_DEF_STMT (new_vdef) = stmt; | |
4580 | new_vuse = new_vdef; | |
4581 | } | |
4582 | ||
4583 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
4584 | delete split_store; | |
4585 | ||
f663d9ad KT |
4586 | gcc_assert (seq); |
4587 | if (dump_file) | |
4588 | { | |
4589 | fprintf (dump_file, | |
c94c3532 | 4590 | "New sequence of %u stores to replace old one of %u stores\n", |
a62b3dc5 | 4591 | split_stores.length (), orig_num_stmts); |
f663d9ad KT |
4592 | if (dump_flags & TDF_DETAILS) |
4593 | print_gimple_seq (dump_file, seq, 0, TDF_VOPS | TDF_MEMSYMS); | |
4594 | } | |
629387a6 | 4595 | |
5384a802 JJ |
4596 | if (gimple_clobber_p (group->last_stmt)) |
4597 | update_stmt (group->last_stmt); | |
4598 | ||
629387a6 EB |
4599 | if (group->lp_nr > 0) |
4600 | { | |
4601 | /* We're going to insert a sequence of (potentially) throwing stores | |
4602 | into an active EH region. This means that we're going to create | |
4603 | new basic blocks with EH edges pointing to the post landing pad | |
4604 | and, therefore, to have to update its PHI nodes, if any. For the | |
4605 | virtual PHI node, we're going to use the VDEFs created above, but | |
4606 | for the other nodes, we need to record the original reaching defs. */ | |
4607 | eh_landing_pad lp = get_eh_landing_pad_from_number (group->lp_nr); | |
4608 | basic_block lp_bb = label_to_block (cfun, lp->post_landing_pad); | |
4609 | basic_block last_bb = gimple_bb (group->last_stmt); | |
4610 | edge last_edge = find_edge (last_bb, lp_bb); | |
4611 | auto_vec<tree, 16> last_defs; | |
4612 | gphi_iterator gpi; | |
4613 | for (gpi = gsi_start_phis (lp_bb); !gsi_end_p (gpi); gsi_next (&gpi)) | |
4614 | { | |
4615 | gphi *phi = gpi.phi (); | |
4616 | tree last_def; | |
4617 | if (virtual_operand_p (gimple_phi_result (phi))) | |
4618 | last_def = NULL_TREE; | |
4619 | else | |
4620 | last_def = gimple_phi_arg_def (phi, last_edge->dest_idx); | |
4621 | last_defs.safe_push (last_def); | |
4622 | } | |
4623 | ||
4624 | /* Do the insertion. Then, if new basic blocks have been created in the | |
4625 | process, rewind the chain of VDEFs create above to walk the new basic | |
4626 | blocks and update the corresponding arguments of the PHI nodes. */ | |
4627 | update_modified_stmts (seq); | |
4628 | if (gimple_find_sub_bbs (seq, &last_gsi)) | |
4629 | while (last_vdef != gimple_vuse (group->last_stmt)) | |
4630 | { | |
4631 | gimple *stmt = SSA_NAME_DEF_STMT (last_vdef); | |
4632 | if (stmt_could_throw_p (cfun, stmt)) | |
4633 | { | |
4634 | edge new_edge = find_edge (gimple_bb (stmt), lp_bb); | |
4635 | unsigned int i; | |
4636 | for (gpi = gsi_start_phis (lp_bb), i = 0; | |
4637 | !gsi_end_p (gpi); | |
4638 | gsi_next (&gpi), i++) | |
4639 | { | |
4640 | gphi *phi = gpi.phi (); | |
4641 | tree new_def; | |
4642 | if (virtual_operand_p (gimple_phi_result (phi))) | |
4643 | new_def = last_vdef; | |
4644 | else | |
4645 | new_def = last_defs[i]; | |
4646 | add_phi_arg (phi, new_def, new_edge, UNKNOWN_LOCATION); | |
4647 | } | |
4648 | } | |
4649 | last_vdef = gimple_vuse (stmt); | |
4650 | } | |
4651 | } | |
4652 | else | |
4653 | gsi_insert_seq_after (&last_gsi, seq, GSI_SAME_STMT); | |
4654 | ||
245f6de1 JJ |
4655 | for (int j = 0; j < 2; ++j) |
4656 | if (load_seq[j]) | |
4657 | gsi_insert_seq_after (&load_gsi[j], load_seq[j], GSI_SAME_STMT); | |
f663d9ad KT |
4658 | |
4659 | return true; | |
4660 | } | |
4661 | ||
4662 | /* Process the merged_store_group objects created in the coalescing phase. | |
4663 | The stores are all against the base object BASE. | |
4664 | Try to output the widened stores and delete the original statements if | |
4665 | successful. Return true iff any changes were made. */ | |
4666 | ||
4667 | bool | |
b5926e23 | 4668 | imm_store_chain_info::output_merged_stores () |
f663d9ad KT |
4669 | { |
4670 | unsigned int i; | |
4671 | merged_store_group *merged_store; | |
4672 | bool ret = false; | |
4673 | FOR_EACH_VEC_ELT (m_merged_store_groups, i, merged_store) | |
4674 | { | |
a95b474a ML |
4675 | if (dbg_cnt (store_merging) |
4676 | && output_merged_store (merged_store)) | |
f663d9ad KT |
4677 | { |
4678 | unsigned int j; | |
4679 | store_immediate_info *store; | |
4680 | FOR_EACH_VEC_ELT (merged_store->stores, j, store) | |
4681 | { | |
4682 | gimple *stmt = store->stmt; | |
4683 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
5384a802 JJ |
4684 | /* Don't remove clobbers, they are still useful even if |
4685 | everything is overwritten afterwards. */ | |
4686 | if (gimple_clobber_p (stmt)) | |
4687 | continue; | |
f663d9ad | 4688 | gsi_remove (&gsi, true); |
629387a6 EB |
4689 | if (store->lp_nr) |
4690 | remove_stmt_from_eh_lp (stmt); | |
f663d9ad KT |
4691 | if (stmt != merged_store->last_stmt) |
4692 | { | |
4693 | unlink_stmt_vdef (stmt); | |
4694 | release_defs (stmt); | |
4695 | } | |
4696 | } | |
4697 | ret = true; | |
4698 | } | |
4699 | } | |
4700 | if (ret && dump_file) | |
4701 | fprintf (dump_file, "Merging successful!\n"); | |
4702 | ||
4703 | return ret; | |
4704 | } | |
4705 | ||
4706 | /* Coalesce the store_immediate_info objects recorded against the base object | |
4707 | BASE in the first phase and output them. | |
4708 | Delete the allocated structures. | |
4709 | Return true if any changes were made. */ | |
4710 | ||
4711 | bool | |
b5926e23 | 4712 | imm_store_chain_info::terminate_and_process_chain () |
f663d9ad KT |
4713 | { |
4714 | /* Process store chain. */ | |
4715 | bool ret = false; | |
4716 | if (m_store_info.length () > 1) | |
4717 | { | |
4718 | ret = coalesce_immediate_stores (); | |
4719 | if (ret) | |
b5926e23 | 4720 | ret = output_merged_stores (); |
f663d9ad KT |
4721 | } |
4722 | ||
4723 | /* Delete all the entries we allocated ourselves. */ | |
4724 | store_immediate_info *info; | |
4725 | unsigned int i; | |
4726 | FOR_EACH_VEC_ELT (m_store_info, i, info) | |
4727 | delete info; | |
4728 | ||
4729 | merged_store_group *merged_info; | |
4730 | FOR_EACH_VEC_ELT (m_merged_store_groups, i, merged_info) | |
4731 | delete merged_info; | |
4732 | ||
4733 | return ret; | |
4734 | } | |
4735 | ||
4736 | /* Return true iff LHS is a destination potentially interesting for | |
4737 | store merging. In practice these are the codes that get_inner_reference | |
4738 | can process. */ | |
4739 | ||
4740 | static bool | |
4741 | lhs_valid_for_store_merging_p (tree lhs) | |
4742 | { | |
629387a6 | 4743 | if (DECL_P (lhs)) |
f663d9ad KT |
4744 | return true; |
4745 | ||
629387a6 EB |
4746 | switch (TREE_CODE (lhs)) |
4747 | { | |
4748 | case ARRAY_REF: | |
4749 | case ARRAY_RANGE_REF: | |
4750 | case BIT_FIELD_REF: | |
4751 | case COMPONENT_REF: | |
4752 | case MEM_REF: | |
e362a897 | 4753 | case VIEW_CONVERT_EXPR: |
629387a6 EB |
4754 | return true; |
4755 | default: | |
4756 | return false; | |
4757 | } | |
4758 | ||
4759 | gcc_unreachable (); | |
f663d9ad KT |
4760 | } |
4761 | ||
4762 | /* Return true if the tree RHS is a constant we want to consider | |
4763 | during store merging. In practice accept all codes that | |
4764 | native_encode_expr accepts. */ | |
4765 | ||
4766 | static bool | |
4767 | rhs_valid_for_store_merging_p (tree rhs) | |
4768 | { | |
cf098191 | 4769 | unsigned HOST_WIDE_INT size; |
3afd514b | 4770 | if (TREE_CODE (rhs) == CONSTRUCTOR |
3afd514b JJ |
4771 | && CONSTRUCTOR_NELTS (rhs) == 0 |
4772 | && TYPE_SIZE_UNIT (TREE_TYPE (rhs)) | |
4773 | && tree_fits_uhwi_p (TYPE_SIZE_UNIT (TREE_TYPE (rhs)))) | |
4774 | return true; | |
cf098191 RS |
4775 | return (GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (rhs))).is_constant (&size) |
4776 | && native_encode_expr (rhs, NULL, size) != 0); | |
f663d9ad KT |
4777 | } |
4778 | ||
629387a6 EB |
4779 | /* Adjust *PBITPOS, *PBITREGION_START and *PBITREGION_END by BYTE_OFF bytes |
4780 | and return true on success or false on failure. */ | |
4781 | ||
4782 | static bool | |
4783 | adjust_bit_pos (poly_offset_int byte_off, | |
4784 | poly_int64 *pbitpos, | |
4785 | poly_uint64 *pbitregion_start, | |
4786 | poly_uint64 *pbitregion_end) | |
4787 | { | |
4788 | poly_offset_int bit_off = byte_off << LOG2_BITS_PER_UNIT; | |
4789 | bit_off += *pbitpos; | |
4790 | ||
4791 | if (known_ge (bit_off, 0) && bit_off.to_shwi (pbitpos)) | |
4792 | { | |
4793 | if (maybe_ne (*pbitregion_end, 0U)) | |
4794 | { | |
4795 | bit_off = byte_off << LOG2_BITS_PER_UNIT; | |
4796 | bit_off += *pbitregion_start; | |
4797 | if (bit_off.to_uhwi (pbitregion_start)) | |
4798 | { | |
4799 | bit_off = byte_off << LOG2_BITS_PER_UNIT; | |
4800 | bit_off += *pbitregion_end; | |
4801 | if (!bit_off.to_uhwi (pbitregion_end)) | |
4802 | *pbitregion_end = 0; | |
4803 | } | |
4804 | else | |
4805 | *pbitregion_end = 0; | |
4806 | } | |
4807 | return true; | |
4808 | } | |
4809 | else | |
4810 | return false; | |
4811 | } | |
4812 | ||
245f6de1 JJ |
4813 | /* If MEM is a memory reference usable for store merging (either as |
4814 | store destination or for loads), return the non-NULL base_addr | |
4815 | and set *PBITSIZE, *PBITPOS, *PBITREGION_START and *PBITREGION_END. | |
4816 | Otherwise return NULL, *PBITPOS should be still valid even for that | |
4817 | case. */ | |
4818 | ||
4819 | static tree | |
8a91d545 RS |
4820 | mem_valid_for_store_merging (tree mem, poly_uint64 *pbitsize, |
4821 | poly_uint64 *pbitpos, | |
4822 | poly_uint64 *pbitregion_start, | |
4823 | poly_uint64 *pbitregion_end) | |
245f6de1 | 4824 | { |
8a91d545 RS |
4825 | poly_int64 bitsize, bitpos; |
4826 | poly_uint64 bitregion_start = 0, bitregion_end = 0; | |
245f6de1 JJ |
4827 | machine_mode mode; |
4828 | int unsignedp = 0, reversep = 0, volatilep = 0; | |
4829 | tree offset; | |
4830 | tree base_addr = get_inner_reference (mem, &bitsize, &bitpos, &offset, &mode, | |
4831 | &unsignedp, &reversep, &volatilep); | |
4832 | *pbitsize = bitsize; | |
8a91d545 | 4833 | if (known_eq (bitsize, 0)) |
245f6de1 JJ |
4834 | return NULL_TREE; |
4835 | ||
4836 | if (TREE_CODE (mem) == COMPONENT_REF | |
4837 | && DECL_BIT_FIELD_TYPE (TREE_OPERAND (mem, 1))) | |
4838 | { | |
4839 | get_bit_range (&bitregion_start, &bitregion_end, mem, &bitpos, &offset); | |
8a91d545 RS |
4840 | if (maybe_ne (bitregion_end, 0U)) |
4841 | bitregion_end += 1; | |
245f6de1 JJ |
4842 | } |
4843 | ||
4844 | if (reversep) | |
4845 | return NULL_TREE; | |
4846 | ||
4847 | /* We do not want to rewrite TARGET_MEM_REFs. */ | |
4848 | if (TREE_CODE (base_addr) == TARGET_MEM_REF) | |
4849 | return NULL_TREE; | |
4850 | /* In some cases get_inner_reference may return a | |
4851 | MEM_REF [ptr + byteoffset]. For the purposes of this pass | |
4852 | canonicalize the base_addr to MEM_REF [ptr] and take | |
4853 | byteoffset into account in the bitpos. This occurs in | |
4854 | PR 23684 and this way we can catch more chains. */ | |
4855 | else if (TREE_CODE (base_addr) == MEM_REF) | |
4856 | { | |
629387a6 EB |
4857 | if (!adjust_bit_pos (mem_ref_offset (base_addr), &bitpos, |
4858 | &bitregion_start, &bitregion_end)) | |
245f6de1 JJ |
4859 | return NULL_TREE; |
4860 | base_addr = TREE_OPERAND (base_addr, 0); | |
4861 | } | |
4862 | /* get_inner_reference returns the base object, get at its | |
4863 | address now. */ | |
4864 | else | |
4865 | { | |
8a91d545 | 4866 | if (maybe_lt (bitpos, 0)) |
245f6de1 JJ |
4867 | return NULL_TREE; |
4868 | base_addr = build_fold_addr_expr (base_addr); | |
4869 | } | |
4870 | ||
629387a6 | 4871 | if (offset) |
245f6de1 JJ |
4872 | { |
4873 | /* If the access is variable offset then a base decl has to be | |
4874 | address-taken to be able to emit pointer-based stores to it. | |
4875 | ??? We might be able to get away with re-using the original | |
4876 | base up to the first variable part and then wrapping that inside | |
4877 | a BIT_FIELD_REF. */ | |
4878 | tree base = get_base_address (base_addr); | |
629387a6 | 4879 | if (!base || (DECL_P (base) && !TREE_ADDRESSABLE (base))) |
245f6de1 JJ |
4880 | return NULL_TREE; |
4881 | ||
629387a6 EB |
4882 | /* Similarly to above for the base, remove constant from the offset. */ |
4883 | if (TREE_CODE (offset) == PLUS_EXPR | |
4884 | && TREE_CODE (TREE_OPERAND (offset, 1)) == INTEGER_CST | |
4885 | && adjust_bit_pos (wi::to_poly_offset (TREE_OPERAND (offset, 1)), | |
4886 | &bitpos, &bitregion_start, &bitregion_end)) | |
4887 | offset = TREE_OPERAND (offset, 0); | |
4888 | ||
245f6de1 JJ |
4889 | base_addr = build2 (POINTER_PLUS_EXPR, TREE_TYPE (base_addr), |
4890 | base_addr, offset); | |
4891 | } | |
4892 | ||
629387a6 EB |
4893 | if (known_eq (bitregion_end, 0U)) |
4894 | { | |
4895 | bitregion_start = round_down_to_byte_boundary (bitpos); | |
4896 | bitregion_end = round_up_to_byte_boundary (bitpos + bitsize); | |
4897 | } | |
4898 | ||
245f6de1 JJ |
4899 | *pbitsize = bitsize; |
4900 | *pbitpos = bitpos; | |
4901 | *pbitregion_start = bitregion_start; | |
4902 | *pbitregion_end = bitregion_end; | |
4903 | return base_addr; | |
4904 | } | |
4905 | ||
4906 | /* Return true if STMT is a load that can be used for store merging. | |
4907 | In that case fill in *OP. BITSIZE, BITPOS, BITREGION_START and | |
4908 | BITREGION_END are properties of the corresponding store. */ | |
4909 | ||
4910 | static bool | |
4911 | handled_load (gimple *stmt, store_operand_info *op, | |
8a91d545 RS |
4912 | poly_uint64 bitsize, poly_uint64 bitpos, |
4913 | poly_uint64 bitregion_start, poly_uint64 bitregion_end) | |
245f6de1 | 4914 | { |
383ac8dc | 4915 | if (!is_gimple_assign (stmt)) |
245f6de1 | 4916 | return false; |
383ac8dc JJ |
4917 | if (gimple_assign_rhs_code (stmt) == BIT_NOT_EXPR) |
4918 | { | |
4919 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
4920 | if (TREE_CODE (rhs1) == SSA_NAME | |
383ac8dc JJ |
4921 | && handled_load (SSA_NAME_DEF_STMT (rhs1), op, bitsize, bitpos, |
4922 | bitregion_start, bitregion_end)) | |
4923 | { | |
d60edaba JJ |
4924 | /* Don't allow _1 = load; _2 = ~1; _3 = ~_2; which should have |
4925 | been optimized earlier, but if allowed here, would confuse the | |
4926 | multiple uses counting. */ | |
4927 | if (op->bit_not_p) | |
4928 | return false; | |
383ac8dc JJ |
4929 | op->bit_not_p = !op->bit_not_p; |
4930 | return true; | |
4931 | } | |
4932 | return false; | |
4933 | } | |
4934 | if (gimple_vuse (stmt) | |
4935 | && gimple_assign_load_p (stmt) | |
36bbc05d | 4936 | && !stmt_can_throw_internal (cfun, stmt) |
245f6de1 JJ |
4937 | && !gimple_has_volatile_ops (stmt)) |
4938 | { | |
4939 | tree mem = gimple_assign_rhs1 (stmt); | |
4940 | op->base_addr | |
4941 | = mem_valid_for_store_merging (mem, &op->bitsize, &op->bitpos, | |
4942 | &op->bitregion_start, | |
4943 | &op->bitregion_end); | |
4944 | if (op->base_addr != NULL_TREE | |
8a91d545 RS |
4945 | && known_eq (op->bitsize, bitsize) |
4946 | && multiple_p (op->bitpos - bitpos, BITS_PER_UNIT) | |
4947 | && known_ge (op->bitpos - op->bitregion_start, | |
4948 | bitpos - bitregion_start) | |
4949 | && known_ge (op->bitregion_end - op->bitpos, | |
4950 | bitregion_end - bitpos)) | |
245f6de1 JJ |
4951 | { |
4952 | op->stmt = stmt; | |
4953 | op->val = mem; | |
383ac8dc | 4954 | op->bit_not_p = false; |
245f6de1 JJ |
4955 | return true; |
4956 | } | |
4957 | } | |
4958 | return false; | |
4959 | } | |
4960 | ||
629387a6 EB |
4961 | /* Return the index number of the landing pad for STMT, if any. */ |
4962 | ||
4963 | static int | |
4964 | lp_nr_for_store (gimple *stmt) | |
4965 | { | |
4966 | if (!cfun->can_throw_non_call_exceptions || !cfun->eh) | |
4967 | return 0; | |
4968 | ||
4969 | if (!stmt_could_throw_p (cfun, stmt)) | |
4970 | return 0; | |
4971 | ||
4972 | return lookup_stmt_eh_lp (stmt); | |
4973 | } | |
4974 | ||
245f6de1 | 4975 | /* Record the store STMT for store merging optimization if it can be |
629387a6 | 4976 | optimized. Return true if any changes were made. */ |
245f6de1 | 4977 | |
629387a6 | 4978 | bool |
245f6de1 JJ |
4979 | pass_store_merging::process_store (gimple *stmt) |
4980 | { | |
4981 | tree lhs = gimple_assign_lhs (stmt); | |
4982 | tree rhs = gimple_assign_rhs1 (stmt); | |
2c832ffe SSF |
4983 | poly_uint64 bitsize, bitpos = 0; |
4984 | poly_uint64 bitregion_start = 0, bitregion_end = 0; | |
245f6de1 JJ |
4985 | tree base_addr |
4986 | = mem_valid_for_store_merging (lhs, &bitsize, &bitpos, | |
4987 | &bitregion_start, &bitregion_end); | |
8a91d545 | 4988 | if (known_eq (bitsize, 0U)) |
629387a6 | 4989 | return false; |
245f6de1 JJ |
4990 | |
4991 | bool invalid = (base_addr == NULL_TREE | |
8a91d545 RS |
4992 | || (maybe_gt (bitsize, |
4993 | (unsigned int) MAX_BITSIZE_MODE_ANY_INT) | |
3afd514b JJ |
4994 | && TREE_CODE (rhs) != INTEGER_CST |
4995 | && (TREE_CODE (rhs) != CONSTRUCTOR | |
4996 | || CONSTRUCTOR_NELTS (rhs) != 0))); | |
245f6de1 | 4997 | enum tree_code rhs_code = ERROR_MARK; |
d60edaba | 4998 | bool bit_not_p = false; |
4b84d9b8 JJ |
4999 | struct symbolic_number n; |
5000 | gimple *ins_stmt = NULL; | |
245f6de1 JJ |
5001 | store_operand_info ops[2]; |
5002 | if (invalid) | |
5003 | ; | |
e362a897 EB |
5004 | else if (TREE_CODE (rhs) == STRING_CST) |
5005 | { | |
5006 | rhs_code = STRING_CST; | |
5007 | ops[0].val = rhs; | |
5008 | } | |
245f6de1 JJ |
5009 | else if (rhs_valid_for_store_merging_p (rhs)) |
5010 | { | |
5011 | rhs_code = INTEGER_CST; | |
5012 | ops[0].val = rhs; | |
5013 | } | |
e362a897 | 5014 | else if (TREE_CODE (rhs) == SSA_NAME) |
245f6de1 JJ |
5015 | { |
5016 | gimple *def_stmt = SSA_NAME_DEF_STMT (rhs), *def_stmt1, *def_stmt2; | |
5017 | if (!is_gimple_assign (def_stmt)) | |
5018 | invalid = true; | |
5019 | else if (handled_load (def_stmt, &ops[0], bitsize, bitpos, | |
5020 | bitregion_start, bitregion_end)) | |
5021 | rhs_code = MEM_REF; | |
d60edaba JJ |
5022 | else if (gimple_assign_rhs_code (def_stmt) == BIT_NOT_EXPR) |
5023 | { | |
5024 | tree rhs1 = gimple_assign_rhs1 (def_stmt); | |
5025 | if (TREE_CODE (rhs1) == SSA_NAME | |
5026 | && is_gimple_assign (SSA_NAME_DEF_STMT (rhs1))) | |
5027 | { | |
5028 | bit_not_p = true; | |
5029 | def_stmt = SSA_NAME_DEF_STMT (rhs1); | |
5030 | } | |
5031 | } | |
c94c3532 | 5032 | |
d60edaba | 5033 | if (rhs_code == ERROR_MARK && !invalid) |
245f6de1 JJ |
5034 | switch ((rhs_code = gimple_assign_rhs_code (def_stmt))) |
5035 | { | |
5036 | case BIT_AND_EXPR: | |
5037 | case BIT_IOR_EXPR: | |
5038 | case BIT_XOR_EXPR: | |
5039 | tree rhs1, rhs2; | |
5040 | rhs1 = gimple_assign_rhs1 (def_stmt); | |
5041 | rhs2 = gimple_assign_rhs2 (def_stmt); | |
5042 | invalid = true; | |
d7a9512e | 5043 | if (TREE_CODE (rhs1) != SSA_NAME) |
245f6de1 JJ |
5044 | break; |
5045 | def_stmt1 = SSA_NAME_DEF_STMT (rhs1); | |
5046 | if (!is_gimple_assign (def_stmt1) | |
5047 | || !handled_load (def_stmt1, &ops[0], bitsize, bitpos, | |
5048 | bitregion_start, bitregion_end)) | |
5049 | break; | |
5050 | if (rhs_valid_for_store_merging_p (rhs2)) | |
5051 | ops[1].val = rhs2; | |
d7a9512e | 5052 | else if (TREE_CODE (rhs2) != SSA_NAME) |
245f6de1 JJ |
5053 | break; |
5054 | else | |
5055 | { | |
5056 | def_stmt2 = SSA_NAME_DEF_STMT (rhs2); | |
5057 | if (!is_gimple_assign (def_stmt2)) | |
5058 | break; | |
5059 | else if (!handled_load (def_stmt2, &ops[1], bitsize, bitpos, | |
5060 | bitregion_start, bitregion_end)) | |
5061 | break; | |
5062 | } | |
5063 | invalid = false; | |
5064 | break; | |
5065 | default: | |
5066 | invalid = true; | |
5067 | break; | |
5068 | } | |
c94c3532 | 5069 | |
8a91d545 RS |
5070 | unsigned HOST_WIDE_INT const_bitsize; |
5071 | if (bitsize.is_constant (&const_bitsize) | |
c94c3532 | 5072 | && (const_bitsize % BITS_PER_UNIT) == 0 |
8a91d545 | 5073 | && const_bitsize <= 64 |
c94c3532 | 5074 | && multiple_p (bitpos, BITS_PER_UNIT)) |
4b84d9b8 JJ |
5075 | { |
5076 | ins_stmt = find_bswap_or_nop_1 (def_stmt, &n, 12); | |
5077 | if (ins_stmt) | |
5078 | { | |
5079 | uint64_t nn = n.n; | |
5080 | for (unsigned HOST_WIDE_INT i = 0; | |
8a91d545 RS |
5081 | i < const_bitsize; |
5082 | i += BITS_PER_UNIT, nn >>= BITS_PER_MARKER) | |
4b84d9b8 JJ |
5083 | if ((nn & MARKER_MASK) == 0 |
5084 | || (nn & MARKER_MASK) == MARKER_BYTE_UNKNOWN) | |
5085 | { | |
5086 | ins_stmt = NULL; | |
5087 | break; | |
5088 | } | |
5089 | if (ins_stmt) | |
5090 | { | |
5091 | if (invalid) | |
5092 | { | |
5093 | rhs_code = LROTATE_EXPR; | |
5094 | ops[0].base_addr = NULL_TREE; | |
5095 | ops[1].base_addr = NULL_TREE; | |
5096 | } | |
5097 | invalid = false; | |
5098 | } | |
5099 | } | |
5100 | } | |
c94c3532 EB |
5101 | |
5102 | if (invalid | |
5103 | && bitsize.is_constant (&const_bitsize) | |
5104 | && ((const_bitsize % BITS_PER_UNIT) != 0 | |
5105 | || !multiple_p (bitpos, BITS_PER_UNIT)) | |
ed01d707 | 5106 | && const_bitsize <= MAX_FIXED_MODE_SIZE) |
c94c3532 | 5107 | { |
c14add82 | 5108 | /* Bypass a conversion to the bit-field type. */ |
31a5d8c5 EB |
5109 | if (!bit_not_p |
5110 | && is_gimple_assign (def_stmt) | |
5111 | && CONVERT_EXPR_CODE_P (rhs_code)) | |
c94c3532 EB |
5112 | { |
5113 | tree rhs1 = gimple_assign_rhs1 (def_stmt); | |
5114 | if (TREE_CODE (rhs1) == SSA_NAME | |
c14add82 | 5115 | && INTEGRAL_TYPE_P (TREE_TYPE (rhs1))) |
c94c3532 EB |
5116 | rhs = rhs1; |
5117 | } | |
5118 | rhs_code = BIT_INSERT_EXPR; | |
31a5d8c5 | 5119 | bit_not_p = false; |
c94c3532 EB |
5120 | ops[0].val = rhs; |
5121 | ops[0].base_addr = NULL_TREE; | |
5122 | ops[1].base_addr = NULL_TREE; | |
5123 | invalid = false; | |
5124 | } | |
245f6de1 | 5125 | } |
e362a897 EB |
5126 | else |
5127 | invalid = true; | |
245f6de1 | 5128 | |
8a91d545 RS |
5129 | unsigned HOST_WIDE_INT const_bitsize, const_bitpos; |
5130 | unsigned HOST_WIDE_INT const_bitregion_start, const_bitregion_end; | |
5131 | if (invalid | |
5132 | || !bitsize.is_constant (&const_bitsize) | |
5133 | || !bitpos.is_constant (&const_bitpos) | |
5134 | || !bitregion_start.is_constant (&const_bitregion_start) | |
5135 | || !bitregion_end.is_constant (&const_bitregion_end)) | |
629387a6 | 5136 | return terminate_all_aliasing_chains (NULL, stmt); |
245f6de1 | 5137 | |
4b84d9b8 JJ |
5138 | if (!ins_stmt) |
5139 | memset (&n, 0, sizeof (n)); | |
5140 | ||
99b1c316 | 5141 | class imm_store_chain_info **chain_info = NULL; |
629387a6 | 5142 | bool ret = false; |
383ac8dc JJ |
5143 | if (base_addr) |
5144 | chain_info = m_stores.get (base_addr); | |
5145 | ||
245f6de1 JJ |
5146 | store_immediate_info *info; |
5147 | if (chain_info) | |
5148 | { | |
5149 | unsigned int ord = (*chain_info)->m_store_info.length (); | |
8a91d545 RS |
5150 | info = new store_immediate_info (const_bitsize, const_bitpos, |
5151 | const_bitregion_start, | |
5152 | const_bitregion_end, | |
5153 | stmt, ord, rhs_code, n, ins_stmt, | |
629387a6 EB |
5154 | bit_not_p, lp_nr_for_store (stmt), |
5155 | ops[0], ops[1]); | |
245f6de1 JJ |
5156 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5157 | { | |
5158 | fprintf (dump_file, "Recording immediate store from stmt:\n"); | |
5159 | print_gimple_stmt (dump_file, stmt, 0); | |
5160 | } | |
5161 | (*chain_info)->m_store_info.safe_push (info); | |
629387a6 | 5162 | ret |= terminate_all_aliasing_chains (chain_info, stmt); |
245f6de1 JJ |
5163 | /* If we reach the limit of stores to merge in a chain terminate and |
5164 | process the chain now. */ | |
5165 | if ((*chain_info)->m_store_info.length () | |
028d4092 | 5166 | == (unsigned int) param_max_stores_to_merge) |
245f6de1 JJ |
5167 | { |
5168 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5169 | fprintf (dump_file, | |
5170 | "Reached maximum number of statements to merge:\n"); | |
629387a6 | 5171 | ret |= terminate_and_process_chain (*chain_info); |
245f6de1 | 5172 | } |
629387a6 | 5173 | return ret; |
245f6de1 JJ |
5174 | } |
5175 | ||
5176 | /* Store aliases any existing chain? */ | |
629387a6 | 5177 | ret |= terminate_all_aliasing_chains (NULL, stmt); |
245f6de1 | 5178 | /* Start a new chain. */ |
99b1c316 | 5179 | class imm_store_chain_info *new_chain |
245f6de1 | 5180 | = new imm_store_chain_info (m_stores_head, base_addr); |
8a91d545 RS |
5181 | info = new store_immediate_info (const_bitsize, const_bitpos, |
5182 | const_bitregion_start, | |
5183 | const_bitregion_end, | |
5184 | stmt, 0, rhs_code, n, ins_stmt, | |
629387a6 EB |
5185 | bit_not_p, lp_nr_for_store (stmt), |
5186 | ops[0], ops[1]); | |
245f6de1 JJ |
5187 | new_chain->m_store_info.safe_push (info); |
5188 | m_stores.put (base_addr, new_chain); | |
5189 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5190 | { | |
5191 | fprintf (dump_file, "Starting new chain with statement:\n"); | |
5192 | print_gimple_stmt (dump_file, stmt, 0); | |
5193 | fprintf (dump_file, "The base object is:\n"); | |
5194 | print_generic_expr (dump_file, base_addr); | |
5195 | fprintf (dump_file, "\n"); | |
5196 | } | |
629387a6 EB |
5197 | return ret; |
5198 | } | |
5199 | ||
5200 | /* Return true if STMT is a store valid for store merging. */ | |
5201 | ||
5202 | static bool | |
5203 | store_valid_for_store_merging_p (gimple *stmt) | |
5204 | { | |
5205 | return gimple_assign_single_p (stmt) | |
5206 | && gimple_vdef (stmt) | |
5207 | && lhs_valid_for_store_merging_p (gimple_assign_lhs (stmt)) | |
5384a802 | 5208 | && (!gimple_has_volatile_ops (stmt) || gimple_clobber_p (stmt)); |
629387a6 EB |
5209 | } |
5210 | ||
5211 | enum basic_block_status { BB_INVALID, BB_VALID, BB_EXTENDED_VALID }; | |
5212 | ||
5213 | /* Return the status of basic block BB wrt store merging. */ | |
5214 | ||
5215 | static enum basic_block_status | |
5216 | get_status_for_store_merging (basic_block bb) | |
5217 | { | |
5218 | unsigned int num_statements = 0; | |
a7553ad6 | 5219 | unsigned int num_constructors = 0; |
629387a6 EB |
5220 | gimple_stmt_iterator gsi; |
5221 | edge e; | |
5222 | ||
5223 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
5224 | { | |
5225 | gimple *stmt = gsi_stmt (gsi); | |
5226 | ||
5227 | if (is_gimple_debug (stmt)) | |
5228 | continue; | |
5229 | ||
5230 | if (store_valid_for_store_merging_p (stmt) && ++num_statements >= 2) | |
5231 | break; | |
a7553ad6 JJ |
5232 | |
5233 | if (is_gimple_assign (stmt) | |
5234 | && gimple_assign_rhs_code (stmt) == CONSTRUCTOR) | |
5235 | { | |
5236 | tree rhs = gimple_assign_rhs1 (stmt); | |
5237 | if (VECTOR_TYPE_P (TREE_TYPE (rhs)) | |
5238 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_TYPE (rhs))) | |
5239 | && gimple_assign_lhs (stmt) != NULL_TREE) | |
5240 | { | |
5241 | HOST_WIDE_INT sz | |
5242 | = int_size_in_bytes (TREE_TYPE (rhs)) * BITS_PER_UNIT; | |
5243 | if (sz == 16 || sz == 32 || sz == 64) | |
5244 | { | |
5245 | num_constructors = 1; | |
5246 | break; | |
5247 | } | |
5248 | } | |
5249 | } | |
629387a6 EB |
5250 | } |
5251 | ||
a7553ad6 | 5252 | if (num_statements == 0 && num_constructors == 0) |
629387a6 EB |
5253 | return BB_INVALID; |
5254 | ||
5255 | if (cfun->can_throw_non_call_exceptions && cfun->eh | |
5256 | && store_valid_for_store_merging_p (gimple_seq_last_stmt (bb_seq (bb))) | |
5257 | && (e = find_fallthru_edge (bb->succs)) | |
5258 | && e->dest == bb->next_bb) | |
5259 | return BB_EXTENDED_VALID; | |
5260 | ||
a7553ad6 | 5261 | return (num_statements >= 2 || num_constructors) ? BB_VALID : BB_INVALID; |
245f6de1 JJ |
5262 | } |
5263 | ||
f663d9ad | 5264 | /* Entry point for the pass. Go over each basic block recording chains of |
245f6de1 JJ |
5265 | immediate stores. Upon encountering a terminating statement (as defined |
5266 | by stmt_terminates_chain_p) process the recorded stores and emit the widened | |
5267 | variants. */ | |
f663d9ad KT |
5268 | |
5269 | unsigned int | |
5270 | pass_store_merging::execute (function *fun) | |
5271 | { | |
5272 | basic_block bb; | |
5273 | hash_set<gimple *> orig_stmts; | |
629387a6 EB |
5274 | bool changed = false, open_chains = false; |
5275 | ||
5276 | /* If the function can throw and catch non-call exceptions, we'll be trying | |
5277 | to merge stores across different basic blocks so we need to first unsplit | |
5278 | the EH edges in order to streamline the CFG of the function. */ | |
5279 | if (cfun->can_throw_non_call_exceptions && cfun->eh) | |
5280 | unsplit_eh_edges (); | |
f663d9ad | 5281 | |
4b84d9b8 JJ |
5282 | calculate_dominance_info (CDI_DOMINATORS); |
5283 | ||
f663d9ad KT |
5284 | FOR_EACH_BB_FN (bb, fun) |
5285 | { | |
629387a6 | 5286 | const basic_block_status bb_status = get_status_for_store_merging (bb); |
f663d9ad | 5287 | gimple_stmt_iterator gsi; |
f663d9ad | 5288 | |
629387a6 EB |
5289 | if (open_chains && (bb_status == BB_INVALID || !single_pred_p (bb))) |
5290 | { | |
5291 | changed |= terminate_and_process_all_chains (); | |
5292 | open_chains = false; | |
f663d9ad KT |
5293 | } |
5294 | ||
629387a6 | 5295 | if (bb_status == BB_INVALID) |
f663d9ad KT |
5296 | continue; |
5297 | ||
5298 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5299 | fprintf (dump_file, "Processing basic block <%d>:\n", bb->index); | |
5300 | ||
a7553ad6 | 5301 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); ) |
f663d9ad KT |
5302 | { |
5303 | gimple *stmt = gsi_stmt (gsi); | |
a7553ad6 | 5304 | gsi_next (&gsi); |
f663d9ad | 5305 | |
50b6d676 AO |
5306 | if (is_gimple_debug (stmt)) |
5307 | continue; | |
5308 | ||
5384a802 | 5309 | if (gimple_has_volatile_ops (stmt) && !gimple_clobber_p (stmt)) |
f663d9ad KT |
5310 | { |
5311 | /* Terminate all chains. */ | |
5312 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5313 | fprintf (dump_file, "Volatile access terminates " | |
5314 | "all chains\n"); | |
629387a6 EB |
5315 | changed |= terminate_and_process_all_chains (); |
5316 | open_chains = false; | |
f663d9ad KT |
5317 | continue; |
5318 | } | |
5319 | ||
a7553ad6 JJ |
5320 | if (is_gimple_assign (stmt) |
5321 | && gimple_assign_rhs_code (stmt) == CONSTRUCTOR | |
5322 | && maybe_optimize_vector_constructor (stmt)) | |
5323 | continue; | |
5324 | ||
629387a6 EB |
5325 | if (store_valid_for_store_merging_p (stmt)) |
5326 | changed |= process_store (stmt); | |
245f6de1 | 5327 | else |
629387a6 EB |
5328 | changed |= terminate_all_aliasing_chains (NULL, stmt); |
5329 | } | |
5330 | ||
5331 | if (bb_status == BB_EXTENDED_VALID) | |
5332 | open_chains = true; | |
5333 | else | |
5334 | { | |
5335 | changed |= terminate_and_process_all_chains (); | |
5336 | open_chains = false; | |
f663d9ad | 5337 | } |
f663d9ad | 5338 | } |
629387a6 EB |
5339 | |
5340 | if (open_chains) | |
5341 | changed |= terminate_and_process_all_chains (); | |
5342 | ||
5343 | /* If the function can throw and catch non-call exceptions and something | |
5344 | changed during the pass, then the CFG has (very likely) changed too. */ | |
5345 | if (cfun->can_throw_non_call_exceptions && cfun->eh && changed) | |
5346 | { | |
5347 | free_dominance_info (CDI_DOMINATORS); | |
5348 | return TODO_cleanup_cfg; | |
5349 | } | |
5350 | ||
f663d9ad KT |
5351 | return 0; |
5352 | } | |
5353 | ||
5354 | } // anon namespace | |
5355 | ||
5356 | /* Construct and return a store merging pass object. */ | |
5357 | ||
5358 | gimple_opt_pass * | |
5359 | make_pass_store_merging (gcc::context *ctxt) | |
5360 | { | |
5361 | return new pass_store_merging (ctxt); | |
5362 | } | |
c22d8787 KT |
5363 | |
5364 | #if CHECKING_P | |
5365 | ||
5366 | namespace selftest { | |
5367 | ||
5368 | /* Selftests for store merging helpers. */ | |
5369 | ||
5370 | /* Assert that all elements of the byte arrays X and Y, both of length N | |
5371 | are equal. */ | |
5372 | ||
5373 | static void | |
5374 | verify_array_eq (unsigned char *x, unsigned char *y, unsigned int n) | |
5375 | { | |
5376 | for (unsigned int i = 0; i < n; i++) | |
5377 | { | |
5378 | if (x[i] != y[i]) | |
5379 | { | |
5380 | fprintf (stderr, "Arrays do not match. X:\n"); | |
5381 | dump_char_array (stderr, x, n); | |
5382 | fprintf (stderr, "Y:\n"); | |
5383 | dump_char_array (stderr, y, n); | |
5384 | } | |
5385 | ASSERT_EQ (x[i], y[i]); | |
5386 | } | |
5387 | } | |
5388 | ||
8aba425f | 5389 | /* Test shift_bytes_in_array_left and that it carries bits across between |
c22d8787 KT |
5390 | bytes correctly. */ |
5391 | ||
5392 | static void | |
8aba425f | 5393 | verify_shift_bytes_in_array_left (void) |
c22d8787 KT |
5394 | { |
5395 | /* byte 1 | byte 0 | |
5396 | 00011111 | 11100000. */ | |
5397 | unsigned char orig[2] = { 0xe0, 0x1f }; | |
5398 | unsigned char in[2]; | |
5399 | memcpy (in, orig, sizeof orig); | |
5400 | ||
5401 | unsigned char expected[2] = { 0x80, 0x7f }; | |
8aba425f | 5402 | shift_bytes_in_array_left (in, sizeof (in), 2); |
c22d8787 KT |
5403 | verify_array_eq (in, expected, sizeof (in)); |
5404 | ||
5405 | memcpy (in, orig, sizeof orig); | |
5406 | memcpy (expected, orig, sizeof orig); | |
5407 | /* Check that shifting by zero doesn't change anything. */ | |
8aba425f | 5408 | shift_bytes_in_array_left (in, sizeof (in), 0); |
c22d8787 KT |
5409 | verify_array_eq (in, expected, sizeof (in)); |
5410 | ||
5411 | } | |
5412 | ||
5413 | /* Test shift_bytes_in_array_right and that it carries bits across between | |
5414 | bytes correctly. */ | |
5415 | ||
5416 | static void | |
5417 | verify_shift_bytes_in_array_right (void) | |
5418 | { | |
5419 | /* byte 1 | byte 0 | |
5420 | 00011111 | 11100000. */ | |
5421 | unsigned char orig[2] = { 0x1f, 0xe0}; | |
5422 | unsigned char in[2]; | |
5423 | memcpy (in, orig, sizeof orig); | |
5424 | unsigned char expected[2] = { 0x07, 0xf8}; | |
5425 | shift_bytes_in_array_right (in, sizeof (in), 2); | |
5426 | verify_array_eq (in, expected, sizeof (in)); | |
5427 | ||
5428 | memcpy (in, orig, sizeof orig); | |
5429 | memcpy (expected, orig, sizeof orig); | |
5430 | /* Check that shifting by zero doesn't change anything. */ | |
5431 | shift_bytes_in_array_right (in, sizeof (in), 0); | |
5432 | verify_array_eq (in, expected, sizeof (in)); | |
5433 | } | |
5434 | ||
5435 | /* Test clear_bit_region that it clears exactly the bits asked and | |
5436 | nothing more. */ | |
5437 | ||
5438 | static void | |
5439 | verify_clear_bit_region (void) | |
5440 | { | |
5441 | /* Start with all bits set and test clearing various patterns in them. */ | |
5442 | unsigned char orig[3] = { 0xff, 0xff, 0xff}; | |
5443 | unsigned char in[3]; | |
5444 | unsigned char expected[3]; | |
5445 | memcpy (in, orig, sizeof in); | |
5446 | ||
5447 | /* Check zeroing out all the bits. */ | |
5448 | clear_bit_region (in, 0, 3 * BITS_PER_UNIT); | |
5449 | expected[0] = expected[1] = expected[2] = 0; | |
5450 | verify_array_eq (in, expected, sizeof in); | |
5451 | ||
5452 | memcpy (in, orig, sizeof in); | |
5453 | /* Leave the first and last bits intact. */ | |
5454 | clear_bit_region (in, 1, 3 * BITS_PER_UNIT - 2); | |
5455 | expected[0] = 0x1; | |
5456 | expected[1] = 0; | |
5457 | expected[2] = 0x80; | |
5458 | verify_array_eq (in, expected, sizeof in); | |
5459 | } | |
5460 | ||
5384a802 | 5461 | /* Test clear_bit_region_be that it clears exactly the bits asked and |
c22d8787 KT |
5462 | nothing more. */ |
5463 | ||
5464 | static void | |
5465 | verify_clear_bit_region_be (void) | |
5466 | { | |
5467 | /* Start with all bits set and test clearing various patterns in them. */ | |
5468 | unsigned char orig[3] = { 0xff, 0xff, 0xff}; | |
5469 | unsigned char in[3]; | |
5470 | unsigned char expected[3]; | |
5471 | memcpy (in, orig, sizeof in); | |
5472 | ||
5473 | /* Check zeroing out all the bits. */ | |
5474 | clear_bit_region_be (in, BITS_PER_UNIT - 1, 3 * BITS_PER_UNIT); | |
5475 | expected[0] = expected[1] = expected[2] = 0; | |
5476 | verify_array_eq (in, expected, sizeof in); | |
5477 | ||
5478 | memcpy (in, orig, sizeof in); | |
5479 | /* Leave the first and last bits intact. */ | |
5480 | clear_bit_region_be (in, BITS_PER_UNIT - 2, 3 * BITS_PER_UNIT - 2); | |
5481 | expected[0] = 0x80; | |
5482 | expected[1] = 0; | |
5483 | expected[2] = 0x1; | |
5484 | verify_array_eq (in, expected, sizeof in); | |
5485 | } | |
5486 | ||
5487 | ||
5488 | /* Run all of the selftests within this file. */ | |
5489 | ||
5490 | void | |
5491 | store_merging_c_tests (void) | |
5492 | { | |
8aba425f | 5493 | verify_shift_bytes_in_array_left (); |
c22d8787 KT |
5494 | verify_shift_bytes_in_array_right (); |
5495 | verify_clear_bit_region (); | |
5496 | verify_clear_bit_region_be (); | |
5497 | } | |
5498 | ||
5499 | } // namespace selftest | |
5500 | #endif /* CHECKING_P. */ |