]>
Commit | Line | Data |
---|---|---|
dffec8eb | 1 | /* GIMPLE store merging and byte swapping passes. |
8d9254fc | 2 | Copyright (C) 2009-2020 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 | { | |
854 | /* The last parameter determines the depth search limit. It usually | |
855 | correlates directly to the number n of bytes to be touched. We | |
0f507a36 | 856 | increase that number by 2 * (log2(n) + 1) here in order to also |
4b84d9b8 JJ |
857 | cover signed -> unsigned conversions of the src operand as can be seen |
858 | in libgcc, and for initial shift/and operation of the src operand. */ | |
859 | int limit = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt))); | |
0f507a36 | 860 | limit += 2 * (1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT) limit)); |
4b84d9b8 JJ |
861 | gimple *ins_stmt = find_bswap_or_nop_1 (stmt, n, limit); |
862 | ||
863 | if (!ins_stmt) | |
864 | return NULL; | |
865 | ||
866 | uint64_t cmpxchg, cmpnop; | |
867 | find_bswap_or_nop_finalize (n, &cmpxchg, &cmpnop); | |
868 | ||
dffec8eb JJ |
869 | /* A complete byte swap should make the symbolic number to start with |
870 | the largest digit in the highest order byte. Unchanged symbolic | |
871 | number indicates a read with same endianness as target architecture. */ | |
872 | if (n->n == cmpnop) | |
873 | *bswap = false; | |
874 | else if (n->n == cmpxchg) | |
875 | *bswap = true; | |
876 | else | |
877 | return NULL; | |
878 | ||
879 | /* Useless bit manipulation performed by code. */ | |
880 | if (!n->base_addr && n->n == cmpnop && n->n_ops == 1) | |
881 | return NULL; | |
882 | ||
dffec8eb JJ |
883 | return ins_stmt; |
884 | } | |
885 | ||
886 | const pass_data pass_data_optimize_bswap = | |
887 | { | |
888 | GIMPLE_PASS, /* type */ | |
889 | "bswap", /* name */ | |
890 | OPTGROUP_NONE, /* optinfo_flags */ | |
891 | TV_NONE, /* tv_id */ | |
892 | PROP_ssa, /* properties_required */ | |
893 | 0, /* properties_provided */ | |
894 | 0, /* properties_destroyed */ | |
895 | 0, /* todo_flags_start */ | |
896 | 0, /* todo_flags_finish */ | |
897 | }; | |
898 | ||
899 | class pass_optimize_bswap : public gimple_opt_pass | |
900 | { | |
901 | public: | |
902 | pass_optimize_bswap (gcc::context *ctxt) | |
903 | : gimple_opt_pass (pass_data_optimize_bswap, ctxt) | |
904 | {} | |
905 | ||
906 | /* opt_pass methods: */ | |
907 | virtual bool gate (function *) | |
908 | { | |
909 | return flag_expensive_optimizations && optimize && BITS_PER_UNIT == 8; | |
910 | } | |
911 | ||
912 | virtual unsigned int execute (function *); | |
913 | ||
914 | }; // class pass_optimize_bswap | |
915 | ||
916 | /* Perform the bswap optimization: replace the expression computed in the rhs | |
4b84d9b8 JJ |
917 | of gsi_stmt (GSI) (or if NULL add instead of replace) by an equivalent |
918 | bswap, load or load + bswap expression. | |
dffec8eb JJ |
919 | Which of these alternatives replace the rhs is given by N->base_addr (non |
920 | null if a load is needed) and BSWAP. The type, VUSE and set-alias of the | |
921 | load to perform are also given in N while the builtin bswap invoke is given | |
4b84d9b8 JJ |
922 | in FNDEL. Finally, if a load is involved, INS_STMT refers to one of the |
923 | load statements involved to construct the rhs in gsi_stmt (GSI) and | |
924 | N->range gives the size of the rhs expression for maintaining some | |
925 | statistics. | |
dffec8eb | 926 | |
4b84d9b8 JJ |
927 | Note that if the replacement involve a load and if gsi_stmt (GSI) is |
928 | non-NULL, that stmt is moved just after INS_STMT to do the load with the | |
929 | same VUSE which can lead to gsi_stmt (GSI) changing of basic block. */ | |
dffec8eb | 930 | |
4b84d9b8 JJ |
931 | tree |
932 | bswap_replace (gimple_stmt_iterator gsi, gimple *ins_stmt, tree fndecl, | |
dffec8eb JJ |
933 | tree bswap_type, tree load_type, struct symbolic_number *n, |
934 | bool bswap) | |
935 | { | |
4b84d9b8 | 936 | tree src, tmp, tgt = NULL_TREE; |
dffec8eb JJ |
937 | gimple *bswap_stmt; |
938 | ||
4b84d9b8 | 939 | gimple *cur_stmt = gsi_stmt (gsi); |
dffec8eb | 940 | src = n->src; |
4b84d9b8 JJ |
941 | if (cur_stmt) |
942 | tgt = gimple_assign_lhs (cur_stmt); | |
dffec8eb JJ |
943 | |
944 | /* Need to load the value from memory first. */ | |
945 | if (n->base_addr) | |
946 | { | |
4b84d9b8 JJ |
947 | gimple_stmt_iterator gsi_ins = gsi; |
948 | if (ins_stmt) | |
949 | gsi_ins = gsi_for_stmt (ins_stmt); | |
dffec8eb JJ |
950 | tree addr_expr, addr_tmp, val_expr, val_tmp; |
951 | tree load_offset_ptr, aligned_load_type; | |
4b84d9b8 JJ |
952 | gimple *load_stmt; |
953 | unsigned align = get_object_alignment (src); | |
4a022c70 | 954 | poly_int64 load_offset = 0; |
dffec8eb | 955 | |
4b84d9b8 JJ |
956 | if (cur_stmt) |
957 | { | |
958 | basic_block ins_bb = gimple_bb (ins_stmt); | |
959 | basic_block cur_bb = gimple_bb (cur_stmt); | |
960 | if (!dominated_by_p (CDI_DOMINATORS, cur_bb, ins_bb)) | |
961 | return NULL_TREE; | |
962 | ||
963 | /* Move cur_stmt just before one of the load of the original | |
964 | to ensure it has the same VUSE. See PR61517 for what could | |
965 | go wrong. */ | |
966 | if (gimple_bb (cur_stmt) != gimple_bb (ins_stmt)) | |
967 | reset_flow_sensitive_info (gimple_assign_lhs (cur_stmt)); | |
968 | gsi_move_before (&gsi, &gsi_ins); | |
969 | gsi = gsi_for_stmt (cur_stmt); | |
970 | } | |
971 | else | |
972 | gsi = gsi_ins; | |
dffec8eb JJ |
973 | |
974 | /* Compute address to load from and cast according to the size | |
975 | of the load. */ | |
4b84d9b8 | 976 | addr_expr = build_fold_addr_expr (src); |
dffec8eb | 977 | if (is_gimple_mem_ref_addr (addr_expr)) |
4b84d9b8 | 978 | addr_tmp = unshare_expr (addr_expr); |
dffec8eb JJ |
979 | else |
980 | { | |
4b84d9b8 JJ |
981 | addr_tmp = unshare_expr (n->base_addr); |
982 | if (!is_gimple_mem_ref_addr (addr_tmp)) | |
983 | addr_tmp = force_gimple_operand_gsi_1 (&gsi, addr_tmp, | |
984 | is_gimple_mem_ref_addr, | |
985 | NULL_TREE, true, | |
986 | GSI_SAME_STMT); | |
987 | load_offset = n->bytepos; | |
988 | if (n->offset) | |
989 | { | |
990 | tree off | |
991 | = force_gimple_operand_gsi (&gsi, unshare_expr (n->offset), | |
992 | true, NULL_TREE, true, | |
993 | GSI_SAME_STMT); | |
994 | gimple *stmt | |
995 | = gimple_build_assign (make_ssa_name (TREE_TYPE (addr_tmp)), | |
996 | POINTER_PLUS_EXPR, addr_tmp, off); | |
997 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
998 | addr_tmp = gimple_assign_lhs (stmt); | |
999 | } | |
dffec8eb JJ |
1000 | } |
1001 | ||
1002 | /* Perform the load. */ | |
1003 | aligned_load_type = load_type; | |
1004 | if (align < TYPE_ALIGN (load_type)) | |
1005 | aligned_load_type = build_aligned_type (load_type, align); | |
1006 | load_offset_ptr = build_int_cst (n->alias_set, load_offset); | |
1007 | val_expr = fold_build2 (MEM_REF, aligned_load_type, addr_tmp, | |
1008 | load_offset_ptr); | |
1009 | ||
1010 | if (!bswap) | |
1011 | { | |
1012 | if (n->range == 16) | |
1013 | nop_stats.found_16bit++; | |
1014 | else if (n->range == 32) | |
1015 | nop_stats.found_32bit++; | |
1016 | else | |
1017 | { | |
1018 | gcc_assert (n->range == 64); | |
1019 | nop_stats.found_64bit++; | |
1020 | } | |
1021 | ||
1022 | /* Convert the result of load if necessary. */ | |
4b84d9b8 | 1023 | if (tgt && !useless_type_conversion_p (TREE_TYPE (tgt), load_type)) |
dffec8eb JJ |
1024 | { |
1025 | val_tmp = make_temp_ssa_name (aligned_load_type, NULL, | |
1026 | "load_dst"); | |
1027 | load_stmt = gimple_build_assign (val_tmp, val_expr); | |
1028 | gimple_set_vuse (load_stmt, n->vuse); | |
1029 | gsi_insert_before (&gsi, load_stmt, GSI_SAME_STMT); | |
1030 | gimple_assign_set_rhs_with_ops (&gsi, NOP_EXPR, val_tmp); | |
4b84d9b8 | 1031 | update_stmt (cur_stmt); |
dffec8eb | 1032 | } |
4b84d9b8 | 1033 | else if (cur_stmt) |
dffec8eb JJ |
1034 | { |
1035 | gimple_assign_set_rhs_with_ops (&gsi, MEM_REF, val_expr); | |
1036 | gimple_set_vuse (cur_stmt, n->vuse); | |
4b84d9b8 JJ |
1037 | update_stmt (cur_stmt); |
1038 | } | |
1039 | else | |
1040 | { | |
1041 | tgt = make_ssa_name (load_type); | |
1042 | cur_stmt = gimple_build_assign (tgt, MEM_REF, val_expr); | |
1043 | gimple_set_vuse (cur_stmt, n->vuse); | |
1044 | gsi_insert_before (&gsi, cur_stmt, GSI_SAME_STMT); | |
dffec8eb | 1045 | } |
dffec8eb JJ |
1046 | |
1047 | if (dump_file) | |
1048 | { | |
1049 | fprintf (dump_file, | |
1050 | "%d bit load in target endianness found at: ", | |
1051 | (int) n->range); | |
1052 | print_gimple_stmt (dump_file, cur_stmt, 0); | |
1053 | } | |
4b84d9b8 | 1054 | return tgt; |
dffec8eb JJ |
1055 | } |
1056 | else | |
1057 | { | |
1058 | val_tmp = make_temp_ssa_name (aligned_load_type, NULL, "load_dst"); | |
1059 | load_stmt = gimple_build_assign (val_tmp, val_expr); | |
1060 | gimple_set_vuse (load_stmt, n->vuse); | |
1061 | gsi_insert_before (&gsi, load_stmt, GSI_SAME_STMT); | |
1062 | } | |
1063 | src = val_tmp; | |
1064 | } | |
1065 | else if (!bswap) | |
1066 | { | |
4b84d9b8 JJ |
1067 | gimple *g = NULL; |
1068 | if (tgt && !useless_type_conversion_p (TREE_TYPE (tgt), TREE_TYPE (src))) | |
dffec8eb JJ |
1069 | { |
1070 | if (!is_gimple_val (src)) | |
4b84d9b8 | 1071 | return NULL_TREE; |
dffec8eb JJ |
1072 | g = gimple_build_assign (tgt, NOP_EXPR, src); |
1073 | } | |
4b84d9b8 | 1074 | else if (cur_stmt) |
dffec8eb | 1075 | g = gimple_build_assign (tgt, src); |
4b84d9b8 JJ |
1076 | else |
1077 | tgt = src; | |
dffec8eb JJ |
1078 | if (n->range == 16) |
1079 | nop_stats.found_16bit++; | |
1080 | else if (n->range == 32) | |
1081 | nop_stats.found_32bit++; | |
1082 | else | |
1083 | { | |
1084 | gcc_assert (n->range == 64); | |
1085 | nop_stats.found_64bit++; | |
1086 | } | |
1087 | if (dump_file) | |
1088 | { | |
1089 | fprintf (dump_file, | |
1090 | "%d bit reshuffle in target endianness found at: ", | |
1091 | (int) n->range); | |
4b84d9b8 JJ |
1092 | if (cur_stmt) |
1093 | print_gimple_stmt (dump_file, cur_stmt, 0); | |
1094 | else | |
1095 | { | |
4af78ef8 | 1096 | print_generic_expr (dump_file, tgt, TDF_NONE); |
4b84d9b8 JJ |
1097 | fprintf (dump_file, "\n"); |
1098 | } | |
dffec8eb | 1099 | } |
4b84d9b8 JJ |
1100 | if (cur_stmt) |
1101 | gsi_replace (&gsi, g, true); | |
1102 | return tgt; | |
dffec8eb JJ |
1103 | } |
1104 | else if (TREE_CODE (src) == BIT_FIELD_REF) | |
1105 | src = TREE_OPERAND (src, 0); | |
1106 | ||
1107 | if (n->range == 16) | |
1108 | bswap_stats.found_16bit++; | |
1109 | else if (n->range == 32) | |
1110 | bswap_stats.found_32bit++; | |
1111 | else | |
1112 | { | |
1113 | gcc_assert (n->range == 64); | |
1114 | bswap_stats.found_64bit++; | |
1115 | } | |
1116 | ||
1117 | tmp = src; | |
1118 | ||
1119 | /* Convert the src expression if necessary. */ | |
1120 | if (!useless_type_conversion_p (TREE_TYPE (tmp), bswap_type)) | |
1121 | { | |
1122 | gimple *convert_stmt; | |
1123 | ||
1124 | tmp = make_temp_ssa_name (bswap_type, NULL, "bswapsrc"); | |
1125 | convert_stmt = gimple_build_assign (tmp, NOP_EXPR, src); | |
1126 | gsi_insert_before (&gsi, convert_stmt, GSI_SAME_STMT); | |
1127 | } | |
1128 | ||
1129 | /* Canonical form for 16 bit bswap is a rotate expression. Only 16bit values | |
1130 | are considered as rotation of 2N bit values by N bits is generally not | |
1131 | equivalent to a bswap. Consider for instance 0x01020304 r>> 16 which | |
1132 | gives 0x03040102 while a bswap for that value is 0x04030201. */ | |
1133 | if (bswap && n->range == 16) | |
1134 | { | |
1135 | tree count = build_int_cst (NULL, BITS_PER_UNIT); | |
1136 | src = fold_build2 (LROTATE_EXPR, bswap_type, tmp, count); | |
1137 | bswap_stmt = gimple_build_assign (NULL, src); | |
1138 | } | |
1139 | else | |
1140 | bswap_stmt = gimple_build_call (fndecl, 1, tmp); | |
1141 | ||
4b84d9b8 JJ |
1142 | if (tgt == NULL_TREE) |
1143 | tgt = make_ssa_name (bswap_type); | |
dffec8eb JJ |
1144 | tmp = tgt; |
1145 | ||
1146 | /* Convert the result if necessary. */ | |
1147 | if (!useless_type_conversion_p (TREE_TYPE (tgt), bswap_type)) | |
1148 | { | |
1149 | gimple *convert_stmt; | |
1150 | ||
1151 | tmp = make_temp_ssa_name (bswap_type, NULL, "bswapdst"); | |
1152 | convert_stmt = gimple_build_assign (tgt, NOP_EXPR, tmp); | |
1153 | gsi_insert_after (&gsi, convert_stmt, GSI_SAME_STMT); | |
1154 | } | |
1155 | ||
1156 | gimple_set_lhs (bswap_stmt, tmp); | |
1157 | ||
1158 | if (dump_file) | |
1159 | { | |
1160 | fprintf (dump_file, "%d bit bswap implementation found at: ", | |
1161 | (int) n->range); | |
4b84d9b8 JJ |
1162 | if (cur_stmt) |
1163 | print_gimple_stmt (dump_file, cur_stmt, 0); | |
1164 | else | |
1165 | { | |
4af78ef8 | 1166 | print_generic_expr (dump_file, tgt, TDF_NONE); |
4b84d9b8 JJ |
1167 | fprintf (dump_file, "\n"); |
1168 | } | |
dffec8eb JJ |
1169 | } |
1170 | ||
4b84d9b8 JJ |
1171 | if (cur_stmt) |
1172 | { | |
1173 | gsi_insert_after (&gsi, bswap_stmt, GSI_SAME_STMT); | |
1174 | gsi_remove (&gsi, true); | |
1175 | } | |
1176 | else | |
1177 | gsi_insert_before (&gsi, bswap_stmt, GSI_SAME_STMT); | |
1178 | return tgt; | |
dffec8eb JJ |
1179 | } |
1180 | ||
1181 | /* Find manual byte swap implementations as well as load in a given | |
1182 | endianness. Byte swaps are turned into a bswap builtin invokation | |
1183 | while endian loads are converted to bswap builtin invokation or | |
1184 | simple load according to the target endianness. */ | |
1185 | ||
1186 | unsigned int | |
1187 | pass_optimize_bswap::execute (function *fun) | |
1188 | { | |
1189 | basic_block bb; | |
1190 | bool bswap32_p, bswap64_p; | |
1191 | bool changed = false; | |
1192 | tree bswap32_type = NULL_TREE, bswap64_type = NULL_TREE; | |
1193 | ||
1194 | bswap32_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP32) | |
1195 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing); | |
1196 | bswap64_p = (builtin_decl_explicit_p (BUILT_IN_BSWAP64) | |
1197 | && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing | |
1198 | || (bswap32_p && word_mode == SImode))); | |
1199 | ||
1200 | /* Determine the argument type of the builtins. The code later on | |
1201 | assumes that the return and argument type are the same. */ | |
1202 | if (bswap32_p) | |
1203 | { | |
1204 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
1205 | bswap32_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1206 | } | |
1207 | ||
1208 | if (bswap64_p) | |
1209 | { | |
1210 | tree fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
1211 | bswap64_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
1212 | } | |
1213 | ||
1214 | memset (&nop_stats, 0, sizeof (nop_stats)); | |
1215 | memset (&bswap_stats, 0, sizeof (bswap_stats)); | |
1216 | calculate_dominance_info (CDI_DOMINATORS); | |
1217 | ||
1218 | FOR_EACH_BB_FN (bb, fun) | |
1219 | { | |
1220 | gimple_stmt_iterator gsi; | |
1221 | ||
1222 | /* We do a reverse scan for bswap patterns to make sure we get the | |
1223 | widest match. As bswap pattern matching doesn't handle previously | |
1224 | inserted smaller bswap replacements as sub-patterns, the wider | |
1225 | variant wouldn't be detected. */ | |
1226 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi);) | |
1227 | { | |
1228 | gimple *ins_stmt, *cur_stmt = gsi_stmt (gsi); | |
1229 | tree fndecl = NULL_TREE, bswap_type = NULL_TREE, load_type; | |
1230 | enum tree_code code; | |
1231 | struct symbolic_number n; | |
1232 | bool bswap; | |
1233 | ||
1234 | /* This gsi_prev (&gsi) is not part of the for loop because cur_stmt | |
1235 | might be moved to a different basic block by bswap_replace and gsi | |
1236 | must not points to it if that's the case. Moving the gsi_prev | |
1237 | there make sure that gsi points to the statement previous to | |
1238 | cur_stmt while still making sure that all statements are | |
1239 | considered in this basic block. */ | |
1240 | gsi_prev (&gsi); | |
1241 | ||
1242 | if (!is_gimple_assign (cur_stmt)) | |
1243 | continue; | |
1244 | ||
1245 | code = gimple_assign_rhs_code (cur_stmt); | |
1246 | switch (code) | |
1247 | { | |
1248 | case LROTATE_EXPR: | |
1249 | case RROTATE_EXPR: | |
1250 | if (!tree_fits_uhwi_p (gimple_assign_rhs2 (cur_stmt)) | |
1251 | || tree_to_uhwi (gimple_assign_rhs2 (cur_stmt)) | |
1252 | % BITS_PER_UNIT) | |
1253 | continue; | |
1254 | /* Fall through. */ | |
1255 | case BIT_IOR_EXPR: | |
1256 | break; | |
1257 | default: | |
1258 | continue; | |
1259 | } | |
1260 | ||
1261 | ins_stmt = find_bswap_or_nop (cur_stmt, &n, &bswap); | |
1262 | ||
1263 | if (!ins_stmt) | |
1264 | continue; | |
1265 | ||
1266 | switch (n.range) | |
1267 | { | |
1268 | case 16: | |
1269 | /* Already in canonical form, nothing to do. */ | |
1270 | if (code == LROTATE_EXPR || code == RROTATE_EXPR) | |
1271 | continue; | |
1272 | load_type = bswap_type = uint16_type_node; | |
1273 | break; | |
1274 | case 32: | |
1275 | load_type = uint32_type_node; | |
1276 | if (bswap32_p) | |
1277 | { | |
1278 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
1279 | bswap_type = bswap32_type; | |
1280 | } | |
1281 | break; | |
1282 | case 64: | |
1283 | load_type = uint64_type_node; | |
1284 | if (bswap64_p) | |
1285 | { | |
1286 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
1287 | bswap_type = bswap64_type; | |
1288 | } | |
1289 | break; | |
1290 | default: | |
1291 | continue; | |
1292 | } | |
1293 | ||
1294 | if (bswap && !fndecl && n.range != 16) | |
1295 | continue; | |
1296 | ||
4b84d9b8 JJ |
1297 | if (bswap_replace (gsi_for_stmt (cur_stmt), ins_stmt, fndecl, |
1298 | bswap_type, load_type, &n, bswap)) | |
dffec8eb JJ |
1299 | changed = true; |
1300 | } | |
1301 | } | |
1302 | ||
1303 | statistics_counter_event (fun, "16-bit nop implementations found", | |
1304 | nop_stats.found_16bit); | |
1305 | statistics_counter_event (fun, "32-bit nop implementations found", | |
1306 | nop_stats.found_32bit); | |
1307 | statistics_counter_event (fun, "64-bit nop implementations found", | |
1308 | nop_stats.found_64bit); | |
1309 | statistics_counter_event (fun, "16-bit bswap implementations found", | |
1310 | bswap_stats.found_16bit); | |
1311 | statistics_counter_event (fun, "32-bit bswap implementations found", | |
1312 | bswap_stats.found_32bit); | |
1313 | statistics_counter_event (fun, "64-bit bswap implementations found", | |
1314 | bswap_stats.found_64bit); | |
1315 | ||
1316 | return (changed ? TODO_update_ssa : 0); | |
1317 | } | |
1318 | ||
1319 | } // anon namespace | |
1320 | ||
1321 | gimple_opt_pass * | |
1322 | make_pass_optimize_bswap (gcc::context *ctxt) | |
1323 | { | |
1324 | return new pass_optimize_bswap (ctxt); | |
1325 | } | |
1326 | ||
1327 | namespace { | |
1328 | ||
245f6de1 | 1329 | /* Struct recording one operand for the store, which is either a constant, |
c94c3532 EB |
1330 | then VAL represents the constant and all the other fields are zero, or |
1331 | a memory load, then VAL represents the reference, BASE_ADDR is non-NULL | |
1332 | and the other fields also reflect the memory load, or an SSA name, then | |
1333 | VAL represents the SSA name and all the other fields are zero, */ | |
245f6de1 | 1334 | |
6c1dae73 | 1335 | class store_operand_info |
245f6de1 | 1336 | { |
6c1dae73 | 1337 | public: |
245f6de1 JJ |
1338 | tree val; |
1339 | tree base_addr; | |
8a91d545 RS |
1340 | poly_uint64 bitsize; |
1341 | poly_uint64 bitpos; | |
1342 | poly_uint64 bitregion_start; | |
1343 | poly_uint64 bitregion_end; | |
245f6de1 | 1344 | gimple *stmt; |
383ac8dc | 1345 | bool bit_not_p; |
245f6de1 JJ |
1346 | store_operand_info (); |
1347 | }; | |
1348 | ||
1349 | store_operand_info::store_operand_info () | |
1350 | : val (NULL_TREE), base_addr (NULL_TREE), bitsize (0), bitpos (0), | |
383ac8dc | 1351 | bitregion_start (0), bitregion_end (0), stmt (NULL), bit_not_p (false) |
245f6de1 JJ |
1352 | { |
1353 | } | |
1354 | ||
f663d9ad KT |
1355 | /* Struct recording the information about a single store of an immediate |
1356 | to memory. These are created in the first phase and coalesced into | |
1357 | merged_store_group objects in the second phase. */ | |
1358 | ||
6c1dae73 | 1359 | class store_immediate_info |
f663d9ad | 1360 | { |
6c1dae73 | 1361 | public: |
f663d9ad KT |
1362 | unsigned HOST_WIDE_INT bitsize; |
1363 | unsigned HOST_WIDE_INT bitpos; | |
a62b3dc5 JJ |
1364 | unsigned HOST_WIDE_INT bitregion_start; |
1365 | /* This is one past the last bit of the bit region. */ | |
1366 | unsigned HOST_WIDE_INT bitregion_end; | |
f663d9ad KT |
1367 | gimple *stmt; |
1368 | unsigned int order; | |
c94c3532 EB |
1369 | /* INTEGER_CST for constant stores, MEM_REF for memory copy, |
1370 | BIT_*_EXPR for logical bitwise operation, BIT_INSERT_EXPR | |
1371 | for bit insertion. | |
4b84d9b8 JJ |
1372 | LROTATE_EXPR if it can be only bswap optimized and |
1373 | ops are not really meaningful. | |
1374 | NOP_EXPR if bswap optimization detected identity, ops | |
1375 | are not meaningful. */ | |
245f6de1 | 1376 | enum tree_code rhs_code; |
4b84d9b8 JJ |
1377 | /* Two fields for bswap optimization purposes. */ |
1378 | struct symbolic_number n; | |
1379 | gimple *ins_stmt; | |
127ef369 | 1380 | /* True if BIT_{AND,IOR,XOR}_EXPR result is inverted before storing. */ |
d60edaba | 1381 | bool bit_not_p; |
127ef369 JJ |
1382 | /* True if ops have been swapped and thus ops[1] represents |
1383 | rhs1 of BIT_{AND,IOR,XOR}_EXPR and ops[0] represents rhs2. */ | |
1384 | bool ops_swapped_p; | |
629387a6 EB |
1385 | /* The index number of the landing pad, or 0 if there is none. */ |
1386 | int lp_nr; | |
245f6de1 JJ |
1387 | /* Operands. For BIT_*_EXPR rhs_code both operands are used, otherwise |
1388 | just the first one. */ | |
1389 | store_operand_info ops[2]; | |
b5926e23 | 1390 | store_immediate_info (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, |
a62b3dc5 | 1391 | unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, |
4b84d9b8 | 1392 | gimple *, unsigned int, enum tree_code, |
629387a6 | 1393 | struct symbolic_number &, gimple *, bool, int, |
245f6de1 JJ |
1394 | const store_operand_info &, |
1395 | const store_operand_info &); | |
f663d9ad KT |
1396 | }; |
1397 | ||
1398 | store_immediate_info::store_immediate_info (unsigned HOST_WIDE_INT bs, | |
b5926e23 | 1399 | unsigned HOST_WIDE_INT bp, |
a62b3dc5 JJ |
1400 | unsigned HOST_WIDE_INT brs, |
1401 | unsigned HOST_WIDE_INT bre, | |
b5926e23 | 1402 | gimple *st, |
245f6de1 JJ |
1403 | unsigned int ord, |
1404 | enum tree_code rhscode, | |
4b84d9b8 JJ |
1405 | struct symbolic_number &nr, |
1406 | gimple *ins_stmtp, | |
d60edaba | 1407 | bool bitnotp, |
629387a6 | 1408 | int nr2, |
245f6de1 JJ |
1409 | const store_operand_info &op0r, |
1410 | const store_operand_info &op1r) | |
a62b3dc5 | 1411 | : bitsize (bs), bitpos (bp), bitregion_start (brs), bitregion_end (bre), |
4b84d9b8 | 1412 | stmt (st), order (ord), rhs_code (rhscode), n (nr), |
629387a6 EB |
1413 | ins_stmt (ins_stmtp), bit_not_p (bitnotp), ops_swapped_p (false), |
1414 | lp_nr (nr2) | |
245f6de1 JJ |
1415 | #if __cplusplus >= 201103L |
1416 | , ops { op0r, op1r } | |
1417 | { | |
1418 | } | |
1419 | #else | |
f663d9ad | 1420 | { |
245f6de1 JJ |
1421 | ops[0] = op0r; |
1422 | ops[1] = op1r; | |
f663d9ad | 1423 | } |
245f6de1 | 1424 | #endif |
f663d9ad KT |
1425 | |
1426 | /* Struct representing a group of stores to contiguous memory locations. | |
1427 | These are produced by the second phase (coalescing) and consumed in the | |
1428 | third phase that outputs the widened stores. */ | |
1429 | ||
6c1dae73 | 1430 | class merged_store_group |
f663d9ad | 1431 | { |
6c1dae73 | 1432 | public: |
f663d9ad KT |
1433 | unsigned HOST_WIDE_INT start; |
1434 | unsigned HOST_WIDE_INT width; | |
a62b3dc5 JJ |
1435 | unsigned HOST_WIDE_INT bitregion_start; |
1436 | unsigned HOST_WIDE_INT bitregion_end; | |
1437 | /* The size of the allocated memory for val and mask. */ | |
f663d9ad | 1438 | unsigned HOST_WIDE_INT buf_size; |
a62b3dc5 | 1439 | unsigned HOST_WIDE_INT align_base; |
8a91d545 | 1440 | poly_uint64 load_align_base[2]; |
f663d9ad KT |
1441 | |
1442 | unsigned int align; | |
245f6de1 | 1443 | unsigned int load_align[2]; |
f663d9ad KT |
1444 | unsigned int first_order; |
1445 | unsigned int last_order; | |
7f5a3982 | 1446 | bool bit_insertion; |
18e0c3d1 JJ |
1447 | bool only_constants; |
1448 | unsigned int first_nonmergeable_order; | |
629387a6 | 1449 | int lp_nr; |
f663d9ad | 1450 | |
a62b3dc5 | 1451 | auto_vec<store_immediate_info *> stores; |
f663d9ad KT |
1452 | /* We record the first and last original statements in the sequence because |
1453 | we'll need their vuse/vdef and replacement position. It's easier to keep | |
1454 | track of them separately as 'stores' is reordered by apply_stores. */ | |
1455 | gimple *last_stmt; | |
1456 | gimple *first_stmt; | |
1457 | unsigned char *val; | |
a62b3dc5 | 1458 | unsigned char *mask; |
f663d9ad KT |
1459 | |
1460 | merged_store_group (store_immediate_info *); | |
1461 | ~merged_store_group (); | |
7f5a3982 | 1462 | bool can_be_merged_into (store_immediate_info *); |
f663d9ad KT |
1463 | void merge_into (store_immediate_info *); |
1464 | void merge_overlapping (store_immediate_info *); | |
1465 | bool apply_stores (); | |
a62b3dc5 JJ |
1466 | private: |
1467 | void do_merge (store_immediate_info *); | |
f663d9ad KT |
1468 | }; |
1469 | ||
1470 | /* Debug helper. Dump LEN elements of byte array PTR to FD in hex. */ | |
1471 | ||
1472 | static void | |
1473 | dump_char_array (FILE *fd, unsigned char *ptr, unsigned int len) | |
1474 | { | |
1475 | if (!fd) | |
1476 | return; | |
1477 | ||
1478 | for (unsigned int i = 0; i < len; i++) | |
c94c3532 | 1479 | fprintf (fd, "%02x ", ptr[i]); |
f663d9ad KT |
1480 | fprintf (fd, "\n"); |
1481 | } | |
1482 | ||
f663d9ad KT |
1483 | /* Clear out LEN bits starting from bit START in the byte array |
1484 | PTR. This clears the bits to the *right* from START. | |
1485 | START must be within [0, BITS_PER_UNIT) and counts starting from | |
1486 | the least significant bit. */ | |
1487 | ||
1488 | static void | |
1489 | clear_bit_region_be (unsigned char *ptr, unsigned int start, | |
1490 | unsigned int len) | |
1491 | { | |
1492 | if (len == 0) | |
1493 | return; | |
1494 | /* Clear len bits to the right of start. */ | |
1495 | else if (len <= start + 1) | |
1496 | { | |
1497 | unsigned char mask = (~(~0U << len)); | |
1498 | mask = mask << (start + 1U - len); | |
1499 | ptr[0] &= ~mask; | |
1500 | } | |
1501 | else if (start != BITS_PER_UNIT - 1) | |
1502 | { | |
1503 | clear_bit_region_be (ptr, start, (start % BITS_PER_UNIT) + 1); | |
1504 | clear_bit_region_be (ptr + 1, BITS_PER_UNIT - 1, | |
1505 | len - (start % BITS_PER_UNIT) - 1); | |
1506 | } | |
1507 | else if (start == BITS_PER_UNIT - 1 | |
1508 | && len > BITS_PER_UNIT) | |
1509 | { | |
1510 | unsigned int nbytes = len / BITS_PER_UNIT; | |
a62b3dc5 | 1511 | memset (ptr, 0, nbytes); |
f663d9ad KT |
1512 | if (len % BITS_PER_UNIT != 0) |
1513 | clear_bit_region_be (ptr + nbytes, BITS_PER_UNIT - 1, | |
1514 | len % BITS_PER_UNIT); | |
1515 | } | |
1516 | else | |
1517 | gcc_unreachable (); | |
1518 | } | |
1519 | ||
1520 | /* In the byte array PTR clear the bit region starting at bit | |
1521 | START and is LEN bits wide. | |
1522 | For regions spanning multiple bytes do this recursively until we reach | |
1523 | zero LEN or a region contained within a single byte. */ | |
1524 | ||
1525 | static void | |
1526 | clear_bit_region (unsigned char *ptr, unsigned int start, | |
1527 | unsigned int len) | |
1528 | { | |
1529 | /* Degenerate base case. */ | |
1530 | if (len == 0) | |
1531 | return; | |
1532 | else if (start >= BITS_PER_UNIT) | |
1533 | clear_bit_region (ptr + 1, start - BITS_PER_UNIT, len); | |
1534 | /* Second base case. */ | |
1535 | else if ((start + len) <= BITS_PER_UNIT) | |
1536 | { | |
46a61395 | 1537 | unsigned char mask = (~0U) << (unsigned char) (BITS_PER_UNIT - len); |
f663d9ad KT |
1538 | mask >>= BITS_PER_UNIT - (start + len); |
1539 | ||
1540 | ptr[0] &= ~mask; | |
1541 | ||
1542 | return; | |
1543 | } | |
1544 | /* Clear most significant bits in a byte and proceed with the next byte. */ | |
1545 | else if (start != 0) | |
1546 | { | |
1547 | clear_bit_region (ptr, start, BITS_PER_UNIT - start); | |
1f069ef5 | 1548 | clear_bit_region (ptr + 1, 0, len - (BITS_PER_UNIT - start)); |
f663d9ad KT |
1549 | } |
1550 | /* Whole bytes need to be cleared. */ | |
1551 | else if (start == 0 && len > BITS_PER_UNIT) | |
1552 | { | |
1553 | unsigned int nbytes = len / BITS_PER_UNIT; | |
a848c710 KT |
1554 | /* We could recurse on each byte but we clear whole bytes, so a simple |
1555 | memset will do. */ | |
46a61395 | 1556 | memset (ptr, '\0', nbytes); |
f663d9ad KT |
1557 | /* Clear the remaining sub-byte region if there is one. */ |
1558 | if (len % BITS_PER_UNIT != 0) | |
1559 | clear_bit_region (ptr + nbytes, 0, len % BITS_PER_UNIT); | |
1560 | } | |
1561 | else | |
1562 | gcc_unreachable (); | |
1563 | } | |
1564 | ||
1565 | /* Write BITLEN bits of EXPR to the byte array PTR at | |
1566 | bit position BITPOS. PTR should contain TOTAL_BYTES elements. | |
1567 | Return true if the operation succeeded. */ | |
1568 | ||
1569 | static bool | |
1570 | encode_tree_to_bitpos (tree expr, unsigned char *ptr, int bitlen, int bitpos, | |
46a61395 | 1571 | unsigned int total_bytes) |
f663d9ad KT |
1572 | { |
1573 | unsigned int first_byte = bitpos / BITS_PER_UNIT; | |
ad1de652 JJ |
1574 | bool sub_byte_op_p = ((bitlen % BITS_PER_UNIT) |
1575 | || (bitpos % BITS_PER_UNIT) | |
f4b31647 | 1576 | || !int_mode_for_size (bitlen, 0).exists ()); |
3afd514b JJ |
1577 | bool empty_ctor_p |
1578 | = (TREE_CODE (expr) == CONSTRUCTOR | |
1579 | && CONSTRUCTOR_NELTS (expr) == 0 | |
1580 | && TYPE_SIZE_UNIT (TREE_TYPE (expr)) | |
1581 | && tree_fits_uhwi_p (TYPE_SIZE_UNIT (TREE_TYPE (expr)))); | |
f663d9ad KT |
1582 | |
1583 | if (!sub_byte_op_p) | |
3afd514b JJ |
1584 | { |
1585 | if (first_byte >= total_bytes) | |
1586 | return false; | |
1587 | total_bytes -= first_byte; | |
1588 | if (empty_ctor_p) | |
1589 | { | |
1590 | unsigned HOST_WIDE_INT rhs_bytes | |
1591 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (expr))); | |
1592 | if (rhs_bytes > total_bytes) | |
1593 | return false; | |
1594 | memset (ptr + first_byte, '\0', rhs_bytes); | |
1595 | return true; | |
1596 | } | |
1597 | return native_encode_expr (expr, ptr + first_byte, total_bytes) != 0; | |
1598 | } | |
f663d9ad KT |
1599 | |
1600 | /* LITTLE-ENDIAN | |
1601 | We are writing a non byte-sized quantity or at a position that is not | |
1602 | at a byte boundary. | |
1603 | |--------|--------|--------| ptr + first_byte | |
1604 | ^ ^ | |
1605 | xxx xxxxxxxx xxx< bp> | |
1606 | |______EXPR____| | |
1607 | ||
46a61395 | 1608 | First native_encode_expr EXPR into a temporary buffer and shift each |
f663d9ad KT |
1609 | byte in the buffer by 'bp' (carrying the bits over as necessary). |
1610 | |00000000|00xxxxxx|xxxxxxxx| << bp = |000xxxxx|xxxxxxxx|xxx00000| | |
1611 | <------bitlen---->< bp> | |
1612 | Then we clear the destination bits: | |
1613 | |---00000|00000000|000-----| ptr + first_byte | |
1614 | <-------bitlen--->< bp> | |
1615 | ||
1616 | Finally we ORR the bytes of the shifted EXPR into the cleared region: | |
1617 | |---xxxxx||xxxxxxxx||xxx-----| ptr + first_byte. | |
1618 | ||
1619 | BIG-ENDIAN | |
1620 | We are writing a non byte-sized quantity or at a position that is not | |
1621 | at a byte boundary. | |
1622 | ptr + first_byte |--------|--------|--------| | |
1623 | ^ ^ | |
1624 | <bp >xxx xxxxxxxx xxx | |
1625 | |_____EXPR_____| | |
1626 | ||
46a61395 | 1627 | First native_encode_expr EXPR into a temporary buffer and shift each |
f663d9ad KT |
1628 | byte in the buffer to the right by (carrying the bits over as necessary). |
1629 | We shift by as much as needed to align the most significant bit of EXPR | |
1630 | with bitpos: | |
1631 | |00xxxxxx|xxxxxxxx| >> 3 = |00000xxx|xxxxxxxx|xxxxx000| | |
1632 | <---bitlen----> <bp ><-----bitlen-----> | |
1633 | Then we clear the destination bits: | |
1634 | ptr + first_byte |-----000||00000000||00000---| | |
1635 | <bp ><-------bitlen-----> | |
1636 | ||
1637 | Finally we ORR the bytes of the shifted EXPR into the cleared region: | |
1638 | ptr + first_byte |---xxxxx||xxxxxxxx||xxx-----|. | |
1639 | The awkwardness comes from the fact that bitpos is counted from the | |
1640 | most significant bit of a byte. */ | |
1641 | ||
ef1d3b57 RS |
1642 | /* We must be dealing with fixed-size data at this point, since the |
1643 | total size is also fixed. */ | |
3afd514b JJ |
1644 | unsigned int byte_size; |
1645 | if (empty_ctor_p) | |
1646 | { | |
1647 | unsigned HOST_WIDE_INT rhs_bytes | |
1648 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (expr))); | |
1649 | if (rhs_bytes > total_bytes) | |
1650 | return false; | |
1651 | byte_size = rhs_bytes; | |
1652 | } | |
1653 | else | |
1654 | { | |
1655 | fixed_size_mode mode | |
1656 | = as_a <fixed_size_mode> (TYPE_MODE (TREE_TYPE (expr))); | |
1657 | byte_size = GET_MODE_SIZE (mode); | |
1658 | } | |
f663d9ad | 1659 | /* Allocate an extra byte so that we have space to shift into. */ |
3afd514b | 1660 | byte_size++; |
f663d9ad | 1661 | unsigned char *tmpbuf = XALLOCAVEC (unsigned char, byte_size); |
46a61395 | 1662 | memset (tmpbuf, '\0', byte_size); |
f663d9ad | 1663 | /* The store detection code should only have allowed constants that are |
3afd514b JJ |
1664 | accepted by native_encode_expr or empty ctors. */ |
1665 | if (!empty_ctor_p | |
1666 | && native_encode_expr (expr, tmpbuf, byte_size - 1) == 0) | |
f663d9ad KT |
1667 | gcc_unreachable (); |
1668 | ||
1669 | /* The native_encode_expr machinery uses TYPE_MODE to determine how many | |
1670 | bytes to write. This means it can write more than | |
1671 | ROUND_UP (bitlen, BITS_PER_UNIT) / BITS_PER_UNIT bytes (for example | |
1672 | write 8 bytes for a bitlen of 40). Skip the bytes that are not within | |
1673 | bitlen and zero out the bits that are not relevant as well (that may | |
1674 | contain a sign bit due to sign-extension). */ | |
1675 | unsigned int padding | |
1676 | = byte_size - ROUND_UP (bitlen, BITS_PER_UNIT) / BITS_PER_UNIT - 1; | |
ad1de652 JJ |
1677 | /* On big-endian the padding is at the 'front' so just skip the initial |
1678 | bytes. */ | |
1679 | if (BYTES_BIG_ENDIAN) | |
1680 | tmpbuf += padding; | |
1681 | ||
1682 | byte_size -= padding; | |
1683 | ||
1684 | if (bitlen % BITS_PER_UNIT != 0) | |
f663d9ad | 1685 | { |
4b2c06f4 | 1686 | if (BYTES_BIG_ENDIAN) |
ad1de652 JJ |
1687 | clear_bit_region_be (tmpbuf, BITS_PER_UNIT - 1, |
1688 | BITS_PER_UNIT - (bitlen % BITS_PER_UNIT)); | |
1689 | else | |
1690 | clear_bit_region (tmpbuf, bitlen, | |
1691 | byte_size * BITS_PER_UNIT - bitlen); | |
f663d9ad | 1692 | } |
ad1de652 JJ |
1693 | /* Left shifting relies on the last byte being clear if bitlen is |
1694 | a multiple of BITS_PER_UNIT, which might not be clear if | |
1695 | there are padding bytes. */ | |
1696 | else if (!BYTES_BIG_ENDIAN) | |
1697 | tmpbuf[byte_size - 1] = '\0'; | |
f663d9ad KT |
1698 | |
1699 | /* Clear the bit region in PTR where the bits from TMPBUF will be | |
46a61395 | 1700 | inserted into. */ |
f663d9ad KT |
1701 | if (BYTES_BIG_ENDIAN) |
1702 | clear_bit_region_be (ptr + first_byte, | |
1703 | BITS_PER_UNIT - 1 - (bitpos % BITS_PER_UNIT), bitlen); | |
1704 | else | |
1705 | clear_bit_region (ptr + first_byte, bitpos % BITS_PER_UNIT, bitlen); | |
1706 | ||
1707 | int shift_amnt; | |
1708 | int bitlen_mod = bitlen % BITS_PER_UNIT; | |
1709 | int bitpos_mod = bitpos % BITS_PER_UNIT; | |
1710 | ||
1711 | bool skip_byte = false; | |
1712 | if (BYTES_BIG_ENDIAN) | |
1713 | { | |
1714 | /* BITPOS and BITLEN are exactly aligned and no shifting | |
1715 | is necessary. */ | |
1716 | if (bitpos_mod + bitlen_mod == BITS_PER_UNIT | |
1717 | || (bitpos_mod == 0 && bitlen_mod == 0)) | |
1718 | shift_amnt = 0; | |
1719 | /* |. . . . . . . .| | |
1720 | <bp > <blen >. | |
1721 | We always shift right for BYTES_BIG_ENDIAN so shift the beginning | |
1722 | of the value until it aligns with 'bp' in the next byte over. */ | |
1723 | else if (bitpos_mod + bitlen_mod < BITS_PER_UNIT) | |
1724 | { | |
1725 | shift_amnt = bitlen_mod + bitpos_mod; | |
1726 | skip_byte = bitlen_mod != 0; | |
1727 | } | |
1728 | /* |. . . . . . . .| | |
1729 | <----bp---> | |
1730 | <---blen---->. | |
1731 | Shift the value right within the same byte so it aligns with 'bp'. */ | |
1732 | else | |
1733 | shift_amnt = bitlen_mod + bitpos_mod - BITS_PER_UNIT; | |
1734 | } | |
1735 | else | |
1736 | shift_amnt = bitpos % BITS_PER_UNIT; | |
1737 | ||
1738 | /* Create the shifted version of EXPR. */ | |
1739 | if (!BYTES_BIG_ENDIAN) | |
46a61395 | 1740 | { |
8aba425f | 1741 | shift_bytes_in_array_left (tmpbuf, byte_size, shift_amnt); |
46a61395 JJ |
1742 | if (shift_amnt == 0) |
1743 | byte_size--; | |
1744 | } | |
f663d9ad KT |
1745 | else |
1746 | { | |
1747 | gcc_assert (BYTES_BIG_ENDIAN); | |
1748 | shift_bytes_in_array_right (tmpbuf, byte_size, shift_amnt); | |
1749 | /* If shifting right forced us to move into the next byte skip the now | |
1750 | empty byte. */ | |
1751 | if (skip_byte) | |
1752 | { | |
1753 | tmpbuf++; | |
1754 | byte_size--; | |
1755 | } | |
1756 | } | |
1757 | ||
1758 | /* Insert the bits from TMPBUF. */ | |
1759 | for (unsigned int i = 0; i < byte_size; i++) | |
1760 | ptr[first_byte + i] |= tmpbuf[i]; | |
1761 | ||
1762 | return true; | |
1763 | } | |
1764 | ||
1765 | /* Sorting function for store_immediate_info objects. | |
1766 | Sorts them by bitposition. */ | |
1767 | ||
1768 | static int | |
1769 | sort_by_bitpos (const void *x, const void *y) | |
1770 | { | |
1771 | store_immediate_info *const *tmp = (store_immediate_info * const *) x; | |
1772 | store_immediate_info *const *tmp2 = (store_immediate_info * const *) y; | |
1773 | ||
109cca3b | 1774 | if ((*tmp)->bitpos < (*tmp2)->bitpos) |
f663d9ad KT |
1775 | return -1; |
1776 | else if ((*tmp)->bitpos > (*tmp2)->bitpos) | |
1777 | return 1; | |
109cca3b | 1778 | else |
0f0027d1 KT |
1779 | /* If they are the same let's use the order which is guaranteed to |
1780 | be different. */ | |
1781 | return (*tmp)->order - (*tmp2)->order; | |
f663d9ad KT |
1782 | } |
1783 | ||
1784 | /* Sorting function for store_immediate_info objects. | |
1785 | Sorts them by the order field. */ | |
1786 | ||
1787 | static int | |
1788 | sort_by_order (const void *x, const void *y) | |
1789 | { | |
1790 | store_immediate_info *const *tmp = (store_immediate_info * const *) x; | |
1791 | store_immediate_info *const *tmp2 = (store_immediate_info * const *) y; | |
1792 | ||
1793 | if ((*tmp)->order < (*tmp2)->order) | |
1794 | return -1; | |
1795 | else if ((*tmp)->order > (*tmp2)->order) | |
1796 | return 1; | |
1797 | ||
1798 | gcc_unreachable (); | |
1799 | } | |
1800 | ||
1801 | /* Initialize a merged_store_group object from a store_immediate_info | |
1802 | object. */ | |
1803 | ||
1804 | merged_store_group::merged_store_group (store_immediate_info *info) | |
1805 | { | |
1806 | start = info->bitpos; | |
1807 | width = info->bitsize; | |
a62b3dc5 JJ |
1808 | bitregion_start = info->bitregion_start; |
1809 | bitregion_end = info->bitregion_end; | |
f663d9ad KT |
1810 | /* VAL has memory allocated for it in apply_stores once the group |
1811 | width has been finalized. */ | |
1812 | val = NULL; | |
a62b3dc5 | 1813 | mask = NULL; |
c94c3532 | 1814 | bit_insertion = false; |
18e0c3d1 JJ |
1815 | only_constants = info->rhs_code == INTEGER_CST; |
1816 | first_nonmergeable_order = ~0U; | |
629387a6 | 1817 | lp_nr = info->lp_nr; |
a62b3dc5 JJ |
1818 | unsigned HOST_WIDE_INT align_bitpos = 0; |
1819 | get_object_alignment_1 (gimple_assign_lhs (info->stmt), | |
1820 | &align, &align_bitpos); | |
1821 | align_base = start - align_bitpos; | |
245f6de1 JJ |
1822 | for (int i = 0; i < 2; ++i) |
1823 | { | |
1824 | store_operand_info &op = info->ops[i]; | |
1825 | if (op.base_addr == NULL_TREE) | |
1826 | { | |
1827 | load_align[i] = 0; | |
1828 | load_align_base[i] = 0; | |
1829 | } | |
1830 | else | |
1831 | { | |
1832 | get_object_alignment_1 (op.val, &load_align[i], &align_bitpos); | |
1833 | load_align_base[i] = op.bitpos - align_bitpos; | |
1834 | } | |
1835 | } | |
f663d9ad KT |
1836 | stores.create (1); |
1837 | stores.safe_push (info); | |
1838 | last_stmt = info->stmt; | |
1839 | last_order = info->order; | |
1840 | first_stmt = last_stmt; | |
1841 | first_order = last_order; | |
1842 | buf_size = 0; | |
1843 | } | |
1844 | ||
1845 | merged_store_group::~merged_store_group () | |
1846 | { | |
1847 | if (val) | |
1848 | XDELETEVEC (val); | |
1849 | } | |
1850 | ||
7f5a3982 EB |
1851 | /* Return true if the store described by INFO can be merged into the group. */ |
1852 | ||
1853 | bool | |
1854 | merged_store_group::can_be_merged_into (store_immediate_info *info) | |
1855 | { | |
1856 | /* Do not merge bswap patterns. */ | |
1857 | if (info->rhs_code == LROTATE_EXPR) | |
1858 | return false; | |
1859 | ||
629387a6 EB |
1860 | if (info->lp_nr != lp_nr) |
1861 | return false; | |
1862 | ||
7f5a3982 EB |
1863 | /* The canonical case. */ |
1864 | if (info->rhs_code == stores[0]->rhs_code) | |
1865 | return true; | |
1866 | ||
1867 | /* BIT_INSERT_EXPR is compatible with INTEGER_CST. */ | |
1868 | if (info->rhs_code == BIT_INSERT_EXPR && stores[0]->rhs_code == INTEGER_CST) | |
1869 | return true; | |
1870 | ||
1871 | if (stores[0]->rhs_code == BIT_INSERT_EXPR && info->rhs_code == INTEGER_CST) | |
1872 | return true; | |
1873 | ||
ed01d707 EB |
1874 | /* We can turn MEM_REF into BIT_INSERT_EXPR for bit-field stores, but do it |
1875 | only for small regions since this can generate a lot of instructions. */ | |
7f5a3982 EB |
1876 | if (info->rhs_code == MEM_REF |
1877 | && (stores[0]->rhs_code == INTEGER_CST | |
1878 | || stores[0]->rhs_code == BIT_INSERT_EXPR) | |
1879 | && info->bitregion_start == stores[0]->bitregion_start | |
ed01d707 | 1880 | && info->bitregion_end == stores[0]->bitregion_end |
2815558a | 1881 | && info->bitregion_end - info->bitregion_start <= MAX_FIXED_MODE_SIZE) |
7f5a3982 EB |
1882 | return true; |
1883 | ||
1884 | if (stores[0]->rhs_code == MEM_REF | |
1885 | && (info->rhs_code == INTEGER_CST | |
1886 | || info->rhs_code == BIT_INSERT_EXPR) | |
1887 | && info->bitregion_start == stores[0]->bitregion_start | |
ed01d707 | 1888 | && info->bitregion_end == stores[0]->bitregion_end |
2815558a | 1889 | && info->bitregion_end - info->bitregion_start <= MAX_FIXED_MODE_SIZE) |
7f5a3982 EB |
1890 | return true; |
1891 | ||
1892 | return false; | |
1893 | } | |
1894 | ||
a62b3dc5 JJ |
1895 | /* Helper method for merge_into and merge_overlapping to do |
1896 | the common part. */ | |
7f5a3982 | 1897 | |
f663d9ad | 1898 | void |
a62b3dc5 | 1899 | merged_store_group::do_merge (store_immediate_info *info) |
f663d9ad | 1900 | { |
a62b3dc5 JJ |
1901 | bitregion_start = MIN (bitregion_start, info->bitregion_start); |
1902 | bitregion_end = MAX (bitregion_end, info->bitregion_end); | |
1903 | ||
1904 | unsigned int this_align; | |
1905 | unsigned HOST_WIDE_INT align_bitpos = 0; | |
1906 | get_object_alignment_1 (gimple_assign_lhs (info->stmt), | |
1907 | &this_align, &align_bitpos); | |
1908 | if (this_align > align) | |
1909 | { | |
1910 | align = this_align; | |
1911 | align_base = info->bitpos - align_bitpos; | |
1912 | } | |
245f6de1 JJ |
1913 | for (int i = 0; i < 2; ++i) |
1914 | { | |
1915 | store_operand_info &op = info->ops[i]; | |
1916 | if (!op.base_addr) | |
1917 | continue; | |
1918 | ||
1919 | get_object_alignment_1 (op.val, &this_align, &align_bitpos); | |
1920 | if (this_align > load_align[i]) | |
1921 | { | |
1922 | load_align[i] = this_align; | |
1923 | load_align_base[i] = op.bitpos - align_bitpos; | |
1924 | } | |
1925 | } | |
f663d9ad | 1926 | |
f663d9ad KT |
1927 | gimple *stmt = info->stmt; |
1928 | stores.safe_push (info); | |
1929 | if (info->order > last_order) | |
1930 | { | |
1931 | last_order = info->order; | |
1932 | last_stmt = stmt; | |
1933 | } | |
1934 | else if (info->order < first_order) | |
1935 | { | |
1936 | first_order = info->order; | |
1937 | first_stmt = stmt; | |
1938 | } | |
18e0c3d1 JJ |
1939 | if (info->rhs_code != INTEGER_CST) |
1940 | only_constants = false; | |
f663d9ad KT |
1941 | } |
1942 | ||
a62b3dc5 JJ |
1943 | /* Merge a store recorded by INFO into this merged store. |
1944 | The store is not overlapping with the existing recorded | |
1945 | stores. */ | |
1946 | ||
1947 | void | |
1948 | merged_store_group::merge_into (store_immediate_info *info) | |
1949 | { | |
a62b3dc5 JJ |
1950 | /* Make sure we're inserting in the position we think we're inserting. */ |
1951 | gcc_assert (info->bitpos >= start + width | |
1952 | && info->bitregion_start <= bitregion_end); | |
1953 | ||
c5679c37 | 1954 | width = info->bitpos + info->bitsize - start; |
a62b3dc5 JJ |
1955 | do_merge (info); |
1956 | } | |
1957 | ||
f663d9ad KT |
1958 | /* Merge a store described by INFO into this merged store. |
1959 | INFO overlaps in some way with the current store (i.e. it's not contiguous | |
1960 | which is handled by merged_store_group::merge_into). */ | |
1961 | ||
1962 | void | |
1963 | merged_store_group::merge_overlapping (store_immediate_info *info) | |
1964 | { | |
f663d9ad | 1965 | /* If the store extends the size of the group, extend the width. */ |
a62b3dc5 | 1966 | if (info->bitpos + info->bitsize > start + width) |
c5679c37 | 1967 | width = info->bitpos + info->bitsize - start; |
f663d9ad | 1968 | |
a62b3dc5 | 1969 | do_merge (info); |
f663d9ad KT |
1970 | } |
1971 | ||
1972 | /* Go through all the recorded stores in this group in program order and | |
1973 | apply their values to the VAL byte array to create the final merged | |
1974 | value. Return true if the operation succeeded. */ | |
1975 | ||
1976 | bool | |
1977 | merged_store_group::apply_stores () | |
1978 | { | |
a62b3dc5 JJ |
1979 | /* Make sure we have more than one store in the group, otherwise we cannot |
1980 | merge anything. */ | |
1981 | if (bitregion_start % BITS_PER_UNIT != 0 | |
1982 | || bitregion_end % BITS_PER_UNIT != 0 | |
f663d9ad KT |
1983 | || stores.length () == 1) |
1984 | return false; | |
1985 | ||
1986 | stores.qsort (sort_by_order); | |
a62b3dc5 | 1987 | store_immediate_info *info; |
f663d9ad | 1988 | unsigned int i; |
c94c3532 EB |
1989 | /* Create a power-of-2-sized buffer for native_encode_expr. */ |
1990 | buf_size = 1 << ceil_log2 ((bitregion_end - bitregion_start) / BITS_PER_UNIT); | |
a62b3dc5 JJ |
1991 | val = XNEWVEC (unsigned char, 2 * buf_size); |
1992 | mask = val + buf_size; | |
1993 | memset (val, 0, buf_size); | |
1994 | memset (mask, ~0U, buf_size); | |
f663d9ad KT |
1995 | |
1996 | FOR_EACH_VEC_ELT (stores, i, info) | |
1997 | { | |
a62b3dc5 | 1998 | unsigned int pos_in_buffer = info->bitpos - bitregion_start; |
c94c3532 | 1999 | tree cst; |
245f6de1 JJ |
2000 | if (info->ops[0].val && info->ops[0].base_addr == NULL_TREE) |
2001 | cst = info->ops[0].val; | |
2002 | else if (info->ops[1].val && info->ops[1].base_addr == NULL_TREE) | |
2003 | cst = info->ops[1].val; | |
c94c3532 EB |
2004 | else |
2005 | cst = NULL_TREE; | |
245f6de1 JJ |
2006 | bool ret = true; |
2007 | if (cst) | |
c94c3532 EB |
2008 | { |
2009 | if (info->rhs_code == BIT_INSERT_EXPR) | |
2010 | bit_insertion = true; | |
2011 | else | |
2012 | ret = encode_tree_to_bitpos (cst, val, info->bitsize, | |
2013 | pos_in_buffer, buf_size); | |
2014 | } | |
2015 | unsigned char *m = mask + (pos_in_buffer / BITS_PER_UNIT); | |
2016 | if (BYTES_BIG_ENDIAN) | |
2017 | clear_bit_region_be (m, (BITS_PER_UNIT - 1 | |
2018 | - (pos_in_buffer % BITS_PER_UNIT)), | |
2019 | info->bitsize); | |
2020 | else | |
2021 | clear_bit_region (m, pos_in_buffer % BITS_PER_UNIT, info->bitsize); | |
245f6de1 | 2022 | if (cst && dump_file && (dump_flags & TDF_DETAILS)) |
f663d9ad KT |
2023 | { |
2024 | if (ret) | |
2025 | { | |
c94c3532 | 2026 | fputs ("After writing ", dump_file); |
4af78ef8 | 2027 | print_generic_expr (dump_file, cst, TDF_NONE); |
f663d9ad | 2028 | fprintf (dump_file, " of size " HOST_WIDE_INT_PRINT_DEC |
c94c3532 EB |
2029 | " at position %d\n", info->bitsize, pos_in_buffer); |
2030 | fputs (" the merged value contains ", dump_file); | |
f663d9ad | 2031 | dump_char_array (dump_file, val, buf_size); |
c94c3532 EB |
2032 | fputs (" the merged mask contains ", dump_file); |
2033 | dump_char_array (dump_file, mask, buf_size); | |
2034 | if (bit_insertion) | |
2035 | fputs (" bit insertion is required\n", dump_file); | |
f663d9ad KT |
2036 | } |
2037 | else | |
2038 | fprintf (dump_file, "Failed to merge stores\n"); | |
4b84d9b8 | 2039 | } |
f663d9ad KT |
2040 | if (!ret) |
2041 | return false; | |
2042 | } | |
4b84d9b8 | 2043 | stores.qsort (sort_by_bitpos); |
f663d9ad KT |
2044 | return true; |
2045 | } | |
2046 | ||
2047 | /* Structure describing the store chain. */ | |
2048 | ||
6c1dae73 | 2049 | class imm_store_chain_info |
f663d9ad | 2050 | { |
6c1dae73 | 2051 | public: |
50b6d676 AO |
2052 | /* Doubly-linked list that imposes an order on chain processing. |
2053 | PNXP (prev's next pointer) points to the head of a list, or to | |
2054 | the next field in the previous chain in the list. | |
2055 | See pass_store_merging::m_stores_head for more rationale. */ | |
2056 | imm_store_chain_info *next, **pnxp; | |
b5926e23 | 2057 | tree base_addr; |
a62b3dc5 | 2058 | auto_vec<store_immediate_info *> m_store_info; |
f663d9ad KT |
2059 | auto_vec<merged_store_group *> m_merged_store_groups; |
2060 | ||
50b6d676 AO |
2061 | imm_store_chain_info (imm_store_chain_info *&inspt, tree b_a) |
2062 | : next (inspt), pnxp (&inspt), base_addr (b_a) | |
2063 | { | |
2064 | inspt = this; | |
2065 | if (next) | |
2066 | { | |
2067 | gcc_checking_assert (pnxp == next->pnxp); | |
2068 | next->pnxp = &next; | |
2069 | } | |
2070 | } | |
2071 | ~imm_store_chain_info () | |
2072 | { | |
2073 | *pnxp = next; | |
2074 | if (next) | |
2075 | { | |
2076 | gcc_checking_assert (&next == next->pnxp); | |
2077 | next->pnxp = pnxp; | |
2078 | } | |
2079 | } | |
b5926e23 | 2080 | bool terminate_and_process_chain (); |
4b84d9b8 | 2081 | bool try_coalesce_bswap (merged_store_group *, unsigned int, unsigned int); |
f663d9ad | 2082 | bool coalesce_immediate_stores (); |
b5926e23 RB |
2083 | bool output_merged_store (merged_store_group *); |
2084 | bool output_merged_stores (); | |
f663d9ad KT |
2085 | }; |
2086 | ||
2087 | const pass_data pass_data_tree_store_merging = { | |
2088 | GIMPLE_PASS, /* type */ | |
2089 | "store-merging", /* name */ | |
2090 | OPTGROUP_NONE, /* optinfo_flags */ | |
2091 | TV_GIMPLE_STORE_MERGING, /* tv_id */ | |
2092 | PROP_ssa, /* properties_required */ | |
2093 | 0, /* properties_provided */ | |
2094 | 0, /* properties_destroyed */ | |
2095 | 0, /* todo_flags_start */ | |
2096 | TODO_update_ssa, /* todo_flags_finish */ | |
2097 | }; | |
2098 | ||
2099 | class pass_store_merging : public gimple_opt_pass | |
2100 | { | |
2101 | public: | |
2102 | pass_store_merging (gcc::context *ctxt) | |
faec5f24 | 2103 | : gimple_opt_pass (pass_data_tree_store_merging, ctxt), m_stores_head () |
f663d9ad KT |
2104 | { |
2105 | } | |
2106 | ||
c94c3532 EB |
2107 | /* Pass not supported for PDP-endian, nor for insane hosts or |
2108 | target character sizes where native_{encode,interpret}_expr | |
a62b3dc5 | 2109 | doesn't work properly. */ |
f663d9ad KT |
2110 | virtual bool |
2111 | gate (function *) | |
2112 | { | |
a62b3dc5 | 2113 | return flag_store_merging |
c94c3532 | 2114 | && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN |
a62b3dc5 JJ |
2115 | && CHAR_BIT == 8 |
2116 | && BITS_PER_UNIT == 8; | |
f663d9ad KT |
2117 | } |
2118 | ||
2119 | virtual unsigned int execute (function *); | |
2120 | ||
2121 | private: | |
99b1c316 | 2122 | hash_map<tree_operand_hash, class imm_store_chain_info *> m_stores; |
f663d9ad | 2123 | |
50b6d676 AO |
2124 | /* Form a doubly-linked stack of the elements of m_stores, so that |
2125 | we can iterate over them in a predictable way. Using this order | |
2126 | avoids extraneous differences in the compiler output just because | |
2127 | of tree pointer variations (e.g. different chains end up in | |
2128 | different positions of m_stores, so they are handled in different | |
2129 | orders, so they allocate or release SSA names in different | |
2130 | orders, and when they get reused, subsequent passes end up | |
2131 | getting different SSA names, which may ultimately change | |
2132 | decisions when going out of SSA). */ | |
2133 | imm_store_chain_info *m_stores_head; | |
2134 | ||
629387a6 EB |
2135 | bool process_store (gimple *); |
2136 | bool terminate_and_process_chain (imm_store_chain_info *); | |
383ac8dc | 2137 | bool terminate_all_aliasing_chains (imm_store_chain_info **, gimple *); |
629387a6 | 2138 | bool terminate_and_process_all_chains (); |
f663d9ad KT |
2139 | }; // class pass_store_merging |
2140 | ||
2141 | /* Terminate and process all recorded chains. Return true if any changes | |
2142 | were made. */ | |
2143 | ||
2144 | bool | |
2145 | pass_store_merging::terminate_and_process_all_chains () | |
2146 | { | |
f663d9ad | 2147 | bool ret = false; |
50b6d676 | 2148 | while (m_stores_head) |
629387a6 | 2149 | ret |= terminate_and_process_chain (m_stores_head); |
b119c055 | 2150 | gcc_assert (m_stores.is_empty ()); |
f663d9ad KT |
2151 | return ret; |
2152 | } | |
2153 | ||
383ac8dc JJ |
2154 | /* Terminate all chains that are affected by the statement STMT. |
2155 | CHAIN_INFO is the chain we should ignore from the checks if | |
629387a6 | 2156 | non-NULL. Return true if any changes were made. */ |
f663d9ad KT |
2157 | |
2158 | bool | |
20770eb8 | 2159 | pass_store_merging::terminate_all_aliasing_chains (imm_store_chain_info |
b5926e23 | 2160 | **chain_info, |
f663d9ad KT |
2161 | gimple *stmt) |
2162 | { | |
2163 | bool ret = false; | |
2164 | ||
2165 | /* If the statement doesn't touch memory it can't alias. */ | |
2166 | if (!gimple_vuse (stmt)) | |
2167 | return false; | |
2168 | ||
9e875fd8 | 2169 | tree store_lhs = gimple_store_p (stmt) ? gimple_get_lhs (stmt) : NULL_TREE; |
6b412bf6 RB |
2170 | ao_ref store_lhs_ref; |
2171 | ao_ref_init (&store_lhs_ref, store_lhs); | |
383ac8dc | 2172 | for (imm_store_chain_info *next = m_stores_head, *cur = next; cur; cur = next) |
f663d9ad | 2173 | { |
383ac8dc JJ |
2174 | next = cur->next; |
2175 | ||
2176 | /* We already checked all the stores in chain_info and terminated the | |
2177 | chain if necessary. Skip it here. */ | |
2178 | if (chain_info && *chain_info == cur) | |
2179 | continue; | |
2180 | ||
245f6de1 JJ |
2181 | store_immediate_info *info; |
2182 | unsigned int i; | |
383ac8dc | 2183 | FOR_EACH_VEC_ELT (cur->m_store_info, i, info) |
f663d9ad | 2184 | { |
9e875fd8 | 2185 | tree lhs = gimple_assign_lhs (info->stmt); |
6b412bf6 RB |
2186 | ao_ref lhs_ref; |
2187 | ao_ref_init (&lhs_ref, lhs); | |
2188 | if (ref_maybe_used_by_stmt_p (stmt, &lhs_ref) | |
2189 | || stmt_may_clobber_ref_p_1 (stmt, &lhs_ref) | |
2190 | || (store_lhs && refs_may_alias_p_1 (&store_lhs_ref, | |
2191 | &lhs_ref, false))) | |
f663d9ad | 2192 | { |
245f6de1 | 2193 | if (dump_file && (dump_flags & TDF_DETAILS)) |
f663d9ad | 2194 | { |
245f6de1 JJ |
2195 | fprintf (dump_file, "stmt causes chain termination:\n"); |
2196 | print_gimple_stmt (dump_file, stmt, 0); | |
f663d9ad | 2197 | } |
629387a6 | 2198 | ret |= terminate_and_process_chain (cur); |
245f6de1 | 2199 | break; |
f663d9ad KT |
2200 | } |
2201 | } | |
2202 | } | |
2203 | ||
f663d9ad KT |
2204 | return ret; |
2205 | } | |
2206 | ||
2207 | /* Helper function. Terminate the recorded chain storing to base object | |
2208 | BASE. Return true if the merging and output was successful. The m_stores | |
2209 | entry is removed after the processing in any case. */ | |
2210 | ||
2211 | bool | |
629387a6 | 2212 | pass_store_merging::terminate_and_process_chain (imm_store_chain_info *chain_info) |
f663d9ad | 2213 | { |
b5926e23 RB |
2214 | bool ret = chain_info->terminate_and_process_chain (); |
2215 | m_stores.remove (chain_info->base_addr); | |
2216 | delete chain_info; | |
f663d9ad KT |
2217 | return ret; |
2218 | } | |
2219 | ||
245f6de1 | 2220 | /* Return true if stmts in between FIRST (inclusive) and LAST (exclusive) |
629387a6 EB |
2221 | may clobber REF. FIRST and LAST must have non-NULL vdef. We want to |
2222 | be able to sink load of REF across stores between FIRST and LAST, up | |
2223 | to right before LAST. */ | |
245f6de1 JJ |
2224 | |
2225 | bool | |
2226 | stmts_may_clobber_ref_p (gimple *first, gimple *last, tree ref) | |
2227 | { | |
2228 | ao_ref r; | |
2229 | ao_ref_init (&r, ref); | |
2230 | unsigned int count = 0; | |
2231 | tree vop = gimple_vdef (last); | |
2232 | gimple *stmt; | |
2233 | ||
629387a6 EB |
2234 | /* Return true conservatively if the basic blocks are different. */ |
2235 | if (gimple_bb (first) != gimple_bb (last)) | |
2236 | return true; | |
2237 | ||
245f6de1 JJ |
2238 | do |
2239 | { | |
2240 | stmt = SSA_NAME_DEF_STMT (vop); | |
2241 | if (stmt_may_clobber_ref_p_1 (stmt, &r)) | |
2242 | return true; | |
4b84d9b8 JJ |
2243 | if (gimple_store_p (stmt) |
2244 | && refs_anti_dependent_p (ref, gimple_get_lhs (stmt))) | |
2245 | return true; | |
245f6de1 JJ |
2246 | /* Avoid quadratic compile time by bounding the number of checks |
2247 | we perform. */ | |
2248 | if (++count > MAX_STORE_ALIAS_CHECKS) | |
2249 | return true; | |
2250 | vop = gimple_vuse (stmt); | |
2251 | } | |
2252 | while (stmt != first); | |
629387a6 | 2253 | |
245f6de1 JJ |
2254 | return false; |
2255 | } | |
2256 | ||
2257 | /* Return true if INFO->ops[IDX] is mergeable with the | |
2258 | corresponding loads already in MERGED_STORE group. | |
2259 | BASE_ADDR is the base address of the whole store group. */ | |
2260 | ||
2261 | bool | |
2262 | compatible_load_p (merged_store_group *merged_store, | |
2263 | store_immediate_info *info, | |
2264 | tree base_addr, int idx) | |
2265 | { | |
2266 | store_immediate_info *infof = merged_store->stores[0]; | |
2267 | if (!info->ops[idx].base_addr | |
8a91d545 RS |
2268 | || maybe_ne (info->ops[idx].bitpos - infof->ops[idx].bitpos, |
2269 | info->bitpos - infof->bitpos) | |
245f6de1 JJ |
2270 | || !operand_equal_p (info->ops[idx].base_addr, |
2271 | infof->ops[idx].base_addr, 0)) | |
2272 | return false; | |
2273 | ||
2274 | store_immediate_info *infol = merged_store->stores.last (); | |
2275 | tree load_vuse = gimple_vuse (info->ops[idx].stmt); | |
2276 | /* In this case all vuses should be the same, e.g. | |
2277 | _1 = s.a; _2 = s.b; _3 = _1 | 1; t.a = _3; _4 = _2 | 2; t.b = _4; | |
2278 | or | |
2279 | _1 = s.a; _2 = s.b; t.a = _1; t.b = _2; | |
2280 | and we can emit the coalesced load next to any of those loads. */ | |
2281 | if (gimple_vuse (infof->ops[idx].stmt) == load_vuse | |
2282 | && gimple_vuse (infol->ops[idx].stmt) == load_vuse) | |
2283 | return true; | |
2284 | ||
2285 | /* Otherwise, at least for now require that the load has the same | |
2286 | vuse as the store. See following examples. */ | |
2287 | if (gimple_vuse (info->stmt) != load_vuse) | |
2288 | return false; | |
2289 | ||
2290 | if (gimple_vuse (infof->stmt) != gimple_vuse (infof->ops[idx].stmt) | |
2291 | || (infof != infol | |
2292 | && gimple_vuse (infol->stmt) != gimple_vuse (infol->ops[idx].stmt))) | |
2293 | return false; | |
2294 | ||
2295 | /* If the load is from the same location as the store, already | |
2296 | the construction of the immediate chain info guarantees no intervening | |
2297 | stores, so no further checks are needed. Example: | |
2298 | _1 = s.a; _2 = _1 & -7; s.a = _2; _3 = s.b; _4 = _3 & -7; s.b = _4; */ | |
8a91d545 | 2299 | if (known_eq (info->ops[idx].bitpos, info->bitpos) |
245f6de1 JJ |
2300 | && operand_equal_p (info->ops[idx].base_addr, base_addr, 0)) |
2301 | return true; | |
2302 | ||
2303 | /* Otherwise, we need to punt if any of the loads can be clobbered by any | |
2304 | of the stores in the group, or any other stores in between those. | |
2305 | Previous calls to compatible_load_p ensured that for all the | |
2306 | merged_store->stores IDX loads, no stmts starting with | |
2307 | merged_store->first_stmt and ending right before merged_store->last_stmt | |
2308 | clobbers those loads. */ | |
2309 | gimple *first = merged_store->first_stmt; | |
2310 | gimple *last = merged_store->last_stmt; | |
2311 | unsigned int i; | |
2312 | store_immediate_info *infoc; | |
2313 | /* The stores are sorted by increasing store bitpos, so if info->stmt store | |
2314 | comes before the so far first load, we'll be changing | |
2315 | merged_store->first_stmt. In that case we need to give up if | |
2316 | any of the earlier processed loads clobber with the stmts in the new | |
2317 | range. */ | |
2318 | if (info->order < merged_store->first_order) | |
2319 | { | |
2320 | FOR_EACH_VEC_ELT (merged_store->stores, i, infoc) | |
2321 | if (stmts_may_clobber_ref_p (info->stmt, first, infoc->ops[idx].val)) | |
2322 | return false; | |
2323 | first = info->stmt; | |
2324 | } | |
2325 | /* Similarly, we could change merged_store->last_stmt, so ensure | |
2326 | in that case no stmts in the new range clobber any of the earlier | |
2327 | processed loads. */ | |
2328 | else if (info->order > merged_store->last_order) | |
2329 | { | |
2330 | FOR_EACH_VEC_ELT (merged_store->stores, i, infoc) | |
2331 | if (stmts_may_clobber_ref_p (last, info->stmt, infoc->ops[idx].val)) | |
2332 | return false; | |
2333 | last = info->stmt; | |
2334 | } | |
2335 | /* And finally, we'd be adding a new load to the set, ensure it isn't | |
2336 | clobbered in the new range. */ | |
2337 | if (stmts_may_clobber_ref_p (first, last, info->ops[idx].val)) | |
2338 | return false; | |
2339 | ||
2340 | /* Otherwise, we are looking for: | |
2341 | _1 = s.a; _2 = _1 ^ 15; t.a = _2; _3 = s.b; _4 = _3 ^ 15; t.b = _4; | |
2342 | or | |
2343 | _1 = s.a; t.a = _1; _2 = s.b; t.b = _2; */ | |
2344 | return true; | |
2345 | } | |
2346 | ||
4b84d9b8 JJ |
2347 | /* Add all refs loaded to compute VAL to REFS vector. */ |
2348 | ||
2349 | void | |
2350 | gather_bswap_load_refs (vec<tree> *refs, tree val) | |
2351 | { | |
2352 | if (TREE_CODE (val) != SSA_NAME) | |
2353 | return; | |
2354 | ||
2355 | gimple *stmt = SSA_NAME_DEF_STMT (val); | |
2356 | if (!is_gimple_assign (stmt)) | |
2357 | return; | |
2358 | ||
2359 | if (gimple_assign_load_p (stmt)) | |
2360 | { | |
2361 | refs->safe_push (gimple_assign_rhs1 (stmt)); | |
2362 | return; | |
2363 | } | |
2364 | ||
2365 | switch (gimple_assign_rhs_class (stmt)) | |
2366 | { | |
2367 | case GIMPLE_BINARY_RHS: | |
2368 | gather_bswap_load_refs (refs, gimple_assign_rhs2 (stmt)); | |
2369 | /* FALLTHRU */ | |
2370 | case GIMPLE_UNARY_RHS: | |
2371 | gather_bswap_load_refs (refs, gimple_assign_rhs1 (stmt)); | |
2372 | break; | |
2373 | default: | |
2374 | gcc_unreachable (); | |
2375 | } | |
2376 | } | |
2377 | ||
c5679c37 JJ |
2378 | /* Check if there are any stores in M_STORE_INFO after index I |
2379 | (where M_STORE_INFO must be sorted by sort_by_bitpos) that overlap | |
2380 | a potential group ending with END that have their order | |
4d213bf6 JJ |
2381 | smaller than LAST_ORDER. ALL_INTEGER_CST_P is true if |
2382 | all the stores already merged and the one under consideration | |
2383 | have rhs_code of INTEGER_CST. Return true if there are no such stores. | |
c5679c37 JJ |
2384 | Consider: |
2385 | MEM[(long long int *)p_28] = 0; | |
2386 | MEM[(long long int *)p_28 + 8B] = 0; | |
2387 | MEM[(long long int *)p_28 + 16B] = 0; | |
2388 | MEM[(long long int *)p_28 + 24B] = 0; | |
2389 | _129 = (int) _130; | |
2390 | MEM[(int *)p_28 + 8B] = _129; | |
2391 | MEM[(int *)p_28].a = -1; | |
2392 | We already have | |
2393 | MEM[(long long int *)p_28] = 0; | |
2394 | MEM[(int *)p_28].a = -1; | |
2395 | stmts in the current group and need to consider if it is safe to | |
2396 | add MEM[(long long int *)p_28 + 8B] = 0; store into the same group. | |
2397 | There is an overlap between that store and the MEM[(int *)p_28 + 8B] = _129; | |
2398 | store though, so if we add the MEM[(long long int *)p_28 + 8B] = 0; | |
2399 | into the group and merging of those 3 stores is successful, merged | |
2400 | stmts will be emitted at the latest store from that group, i.e. | |
2401 | LAST_ORDER, which is the MEM[(int *)p_28].a = -1; store. | |
2402 | The MEM[(int *)p_28 + 8B] = _129; store that originally follows | |
2403 | the MEM[(long long int *)p_28 + 8B] = 0; would now be before it, | |
2404 | so we need to refuse merging MEM[(long long int *)p_28 + 8B] = 0; | |
2405 | into the group. That way it will be its own store group and will | |
4d213bf6 | 2406 | not be touched. If ALL_INTEGER_CST_P and there are overlapping |
c5679c37 JJ |
2407 | INTEGER_CST stores, those are mergeable using merge_overlapping, |
2408 | so don't return false for those. */ | |
2409 | ||
2410 | static bool | |
2411 | check_no_overlap (vec<store_immediate_info *> m_store_info, unsigned int i, | |
4d213bf6 | 2412 | bool all_integer_cst_p, unsigned int last_order, |
c5679c37 JJ |
2413 | unsigned HOST_WIDE_INT end) |
2414 | { | |
2415 | unsigned int len = m_store_info.length (); | |
2416 | for (++i; i < len; ++i) | |
2417 | { | |
2418 | store_immediate_info *info = m_store_info[i]; | |
2419 | if (info->bitpos >= end) | |
2420 | break; | |
2421 | if (info->order < last_order | |
4d213bf6 | 2422 | && (!all_integer_cst_p || info->rhs_code != INTEGER_CST)) |
c5679c37 JJ |
2423 | return false; |
2424 | } | |
2425 | return true; | |
2426 | } | |
2427 | ||
4b84d9b8 JJ |
2428 | /* Return true if m_store_info[first] and at least one following store |
2429 | form a group which store try_size bitsize value which is byte swapped | |
2430 | from a memory load or some value, or identity from some value. | |
2431 | This uses the bswap pass APIs. */ | |
2432 | ||
2433 | bool | |
2434 | imm_store_chain_info::try_coalesce_bswap (merged_store_group *merged_store, | |
2435 | unsigned int first, | |
2436 | unsigned int try_size) | |
2437 | { | |
2438 | unsigned int len = m_store_info.length (), last = first; | |
2439 | unsigned HOST_WIDE_INT width = m_store_info[first]->bitsize; | |
2440 | if (width >= try_size) | |
2441 | return false; | |
2442 | for (unsigned int i = first + 1; i < len; ++i) | |
2443 | { | |
2444 | if (m_store_info[i]->bitpos != m_store_info[first]->bitpos + width | |
cb76fcd7 | 2445 | || m_store_info[i]->lp_nr != merged_store->lp_nr |
4b84d9b8 JJ |
2446 | || m_store_info[i]->ins_stmt == NULL) |
2447 | return false; | |
2448 | width += m_store_info[i]->bitsize; | |
2449 | if (width >= try_size) | |
2450 | { | |
2451 | last = i; | |
2452 | break; | |
2453 | } | |
2454 | } | |
2455 | if (width != try_size) | |
2456 | return false; | |
2457 | ||
2458 | bool allow_unaligned | |
028d4092 | 2459 | = !STRICT_ALIGNMENT && param_store_merging_allow_unaligned; |
4b84d9b8 JJ |
2460 | /* Punt if the combined store would not be aligned and we need alignment. */ |
2461 | if (!allow_unaligned) | |
2462 | { | |
2463 | unsigned int align = merged_store->align; | |
2464 | unsigned HOST_WIDE_INT align_base = merged_store->align_base; | |
2465 | for (unsigned int i = first + 1; i <= last; ++i) | |
2466 | { | |
2467 | unsigned int this_align; | |
2468 | unsigned HOST_WIDE_INT align_bitpos = 0; | |
2469 | get_object_alignment_1 (gimple_assign_lhs (m_store_info[i]->stmt), | |
2470 | &this_align, &align_bitpos); | |
2471 | if (this_align > align) | |
2472 | { | |
2473 | align = this_align; | |
2474 | align_base = m_store_info[i]->bitpos - align_bitpos; | |
2475 | } | |
2476 | } | |
2477 | unsigned HOST_WIDE_INT align_bitpos | |
2478 | = (m_store_info[first]->bitpos - align_base) & (align - 1); | |
2479 | if (align_bitpos) | |
2480 | align = least_bit_hwi (align_bitpos); | |
2481 | if (align < try_size) | |
2482 | return false; | |
2483 | } | |
2484 | ||
2485 | tree type; | |
2486 | switch (try_size) | |
2487 | { | |
2488 | case 16: type = uint16_type_node; break; | |
2489 | case 32: type = uint32_type_node; break; | |
2490 | case 64: type = uint64_type_node; break; | |
2491 | default: gcc_unreachable (); | |
2492 | } | |
2493 | struct symbolic_number n; | |
2494 | gimple *ins_stmt = NULL; | |
2495 | int vuse_store = -1; | |
2496 | unsigned int first_order = merged_store->first_order; | |
2497 | unsigned int last_order = merged_store->last_order; | |
2498 | gimple *first_stmt = merged_store->first_stmt; | |
2499 | gimple *last_stmt = merged_store->last_stmt; | |
c5679c37 | 2500 | unsigned HOST_WIDE_INT end = merged_store->start + merged_store->width; |
4b84d9b8 JJ |
2501 | store_immediate_info *infof = m_store_info[first]; |
2502 | ||
2503 | for (unsigned int i = first; i <= last; ++i) | |
2504 | { | |
2505 | store_immediate_info *info = m_store_info[i]; | |
2506 | struct symbolic_number this_n = info->n; | |
2507 | this_n.type = type; | |
2508 | if (!this_n.base_addr) | |
2509 | this_n.range = try_size / BITS_PER_UNIT; | |
30fa8e9c JJ |
2510 | else |
2511 | /* Update vuse in case it has changed by output_merged_stores. */ | |
2512 | this_n.vuse = gimple_vuse (info->ins_stmt); | |
4b84d9b8 JJ |
2513 | unsigned int bitpos = info->bitpos - infof->bitpos; |
2514 | if (!do_shift_rotate (LSHIFT_EXPR, &this_n, | |
2515 | BYTES_BIG_ENDIAN | |
2516 | ? try_size - info->bitsize - bitpos | |
2517 | : bitpos)) | |
2518 | return false; | |
aa11164a | 2519 | if (this_n.base_addr && vuse_store) |
4b84d9b8 JJ |
2520 | { |
2521 | unsigned int j; | |
2522 | for (j = first; j <= last; ++j) | |
2523 | if (this_n.vuse == gimple_vuse (m_store_info[j]->stmt)) | |
2524 | break; | |
2525 | if (j > last) | |
2526 | { | |
2527 | if (vuse_store == 1) | |
2528 | return false; | |
2529 | vuse_store = 0; | |
2530 | } | |
2531 | } | |
2532 | if (i == first) | |
2533 | { | |
2534 | n = this_n; | |
2535 | ins_stmt = info->ins_stmt; | |
2536 | } | |
2537 | else | |
2538 | { | |
c5679c37 | 2539 | if (n.base_addr && n.vuse != this_n.vuse) |
4b84d9b8 | 2540 | { |
c5679c37 JJ |
2541 | if (vuse_store == 0) |
2542 | return false; | |
2543 | vuse_store = 1; | |
4b84d9b8 | 2544 | } |
c5679c37 JJ |
2545 | if (info->order > last_order) |
2546 | { | |
2547 | last_order = info->order; | |
2548 | last_stmt = info->stmt; | |
2549 | } | |
2550 | else if (info->order < first_order) | |
2551 | { | |
2552 | first_order = info->order; | |
2553 | first_stmt = info->stmt; | |
2554 | } | |
2555 | end = MAX (end, info->bitpos + info->bitsize); | |
4b84d9b8 JJ |
2556 | |
2557 | ins_stmt = perform_symbolic_merge (ins_stmt, &n, info->ins_stmt, | |
2558 | &this_n, &n); | |
2559 | if (ins_stmt == NULL) | |
2560 | return false; | |
2561 | } | |
2562 | } | |
2563 | ||
2564 | uint64_t cmpxchg, cmpnop; | |
2565 | find_bswap_or_nop_finalize (&n, &cmpxchg, &cmpnop); | |
2566 | ||
2567 | /* A complete byte swap should make the symbolic number to start with | |
2568 | the largest digit in the highest order byte. Unchanged symbolic | |
2569 | number indicates a read with same endianness as target architecture. */ | |
2570 | if (n.n != cmpnop && n.n != cmpxchg) | |
2571 | return false; | |
2572 | ||
2573 | if (n.base_addr == NULL_TREE && !is_gimple_val (n.src)) | |
2574 | return false; | |
2575 | ||
4d213bf6 | 2576 | if (!check_no_overlap (m_store_info, last, false, last_order, end)) |
c5679c37 JJ |
2577 | return false; |
2578 | ||
4b84d9b8 JJ |
2579 | /* Don't handle memory copy this way if normal non-bswap processing |
2580 | would handle it too. */ | |
2581 | if (n.n == cmpnop && (unsigned) n.n_ops == last - first + 1) | |
2582 | { | |
2583 | unsigned int i; | |
2584 | for (i = first; i <= last; ++i) | |
2585 | if (m_store_info[i]->rhs_code != MEM_REF) | |
2586 | break; | |
2587 | if (i == last + 1) | |
2588 | return false; | |
2589 | } | |
2590 | ||
2591 | if (n.n == cmpxchg) | |
2592 | switch (try_size) | |
2593 | { | |
2594 | case 16: | |
2595 | /* Will emit LROTATE_EXPR. */ | |
2596 | break; | |
2597 | case 32: | |
2598 | if (builtin_decl_explicit_p (BUILT_IN_BSWAP32) | |
2599 | && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing) | |
2600 | break; | |
2601 | return false; | |
2602 | case 64: | |
2603 | if (builtin_decl_explicit_p (BUILT_IN_BSWAP64) | |
2604 | && optab_handler (bswap_optab, DImode) != CODE_FOR_nothing) | |
2605 | break; | |
2606 | return false; | |
2607 | default: | |
2608 | gcc_unreachable (); | |
2609 | } | |
2610 | ||
2611 | if (!allow_unaligned && n.base_addr) | |
2612 | { | |
2613 | unsigned int align = get_object_alignment (n.src); | |
2614 | if (align < try_size) | |
2615 | return false; | |
2616 | } | |
2617 | ||
2618 | /* If each load has vuse of the corresponding store, need to verify | |
2619 | the loads can be sunk right before the last store. */ | |
2620 | if (vuse_store == 1) | |
2621 | { | |
2622 | auto_vec<tree, 64> refs; | |
2623 | for (unsigned int i = first; i <= last; ++i) | |
2624 | gather_bswap_load_refs (&refs, | |
2625 | gimple_assign_rhs1 (m_store_info[i]->stmt)); | |
2626 | ||
2627 | unsigned int i; | |
2628 | tree ref; | |
2629 | FOR_EACH_VEC_ELT (refs, i, ref) | |
2630 | if (stmts_may_clobber_ref_p (first_stmt, last_stmt, ref)) | |
2631 | return false; | |
2632 | n.vuse = NULL_TREE; | |
2633 | } | |
2634 | ||
2635 | infof->n = n; | |
2636 | infof->ins_stmt = ins_stmt; | |
2637 | for (unsigned int i = first; i <= last; ++i) | |
2638 | { | |
2639 | m_store_info[i]->rhs_code = n.n == cmpxchg ? LROTATE_EXPR : NOP_EXPR; | |
2640 | m_store_info[i]->ops[0].base_addr = NULL_TREE; | |
2641 | m_store_info[i]->ops[1].base_addr = NULL_TREE; | |
2642 | if (i != first) | |
2643 | merged_store->merge_into (m_store_info[i]); | |
2644 | } | |
2645 | ||
2646 | return true; | |
2647 | } | |
2648 | ||
f663d9ad KT |
2649 | /* Go through the candidate stores recorded in m_store_info and merge them |
2650 | into merged_store_group objects recorded into m_merged_store_groups | |
2651 | representing the widened stores. Return true if coalescing was successful | |
2652 | and the number of widened stores is fewer than the original number | |
2653 | of stores. */ | |
2654 | ||
2655 | bool | |
2656 | imm_store_chain_info::coalesce_immediate_stores () | |
2657 | { | |
2658 | /* Anything less can't be processed. */ | |
2659 | if (m_store_info.length () < 2) | |
2660 | return false; | |
2661 | ||
2662 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
c94c3532 | 2663 | fprintf (dump_file, "Attempting to coalesce %u stores in chain\n", |
f663d9ad KT |
2664 | m_store_info.length ()); |
2665 | ||
2666 | store_immediate_info *info; | |
4b84d9b8 | 2667 | unsigned int i, ignore = 0; |
f663d9ad KT |
2668 | |
2669 | /* Order the stores by the bitposition they write to. */ | |
2670 | m_store_info.qsort (sort_by_bitpos); | |
2671 | ||
2672 | info = m_store_info[0]; | |
2673 | merged_store_group *merged_store = new merged_store_group (info); | |
c94c3532 EB |
2674 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2675 | fputs ("New store group\n", dump_file); | |
f663d9ad KT |
2676 | |
2677 | FOR_EACH_VEC_ELT (m_store_info, i, info) | |
2678 | { | |
3afd514b JJ |
2679 | unsigned HOST_WIDE_INT new_bitregion_start, new_bitregion_end; |
2680 | ||
4b84d9b8 | 2681 | if (i <= ignore) |
c94c3532 | 2682 | goto done; |
f663d9ad | 2683 | |
4b84d9b8 JJ |
2684 | /* First try to handle group of stores like: |
2685 | p[0] = data >> 24; | |
2686 | p[1] = data >> 16; | |
2687 | p[2] = data >> 8; | |
2688 | p[3] = data; | |
2689 | using the bswap framework. */ | |
2690 | if (info->bitpos == merged_store->start + merged_store->width | |
2691 | && merged_store->stores.length () == 1 | |
2692 | && merged_store->stores[0]->ins_stmt != NULL | |
cb76fcd7 | 2693 | && info->lp_nr == merged_store->lp_nr |
4b84d9b8 JJ |
2694 | && info->ins_stmt != NULL) |
2695 | { | |
2696 | unsigned int try_size; | |
2697 | for (try_size = 64; try_size >= 16; try_size >>= 1) | |
2698 | if (try_coalesce_bswap (merged_store, i - 1, try_size)) | |
2699 | break; | |
2700 | ||
2701 | if (try_size >= 16) | |
2702 | { | |
2703 | ignore = i + merged_store->stores.length () - 1; | |
2704 | m_merged_store_groups.safe_push (merged_store); | |
2705 | if (ignore < m_store_info.length ()) | |
2706 | merged_store = new merged_store_group (m_store_info[ignore]); | |
2707 | else | |
2708 | merged_store = NULL; | |
c94c3532 | 2709 | goto done; |
4b84d9b8 JJ |
2710 | } |
2711 | } | |
2712 | ||
3afd514b JJ |
2713 | new_bitregion_start |
2714 | = MIN (merged_store->bitregion_start, info->bitregion_start); | |
2715 | new_bitregion_end | |
2716 | = MAX (merged_store->bitregion_end, info->bitregion_end); | |
2717 | ||
2718 | if (info->order >= merged_store->first_nonmergeable_order | |
2719 | || (((new_bitregion_end - new_bitregion_start + 1) / BITS_PER_UNIT) | |
028d4092 | 2720 | > (unsigned) param_store_merging_max_size)) |
18e0c3d1 JJ |
2721 | ; |
2722 | ||
f663d9ad KT |
2723 | /* |---store 1---| |
2724 | |---store 2---| | |
4b84d9b8 | 2725 | Overlapping stores. */ |
18e0c3d1 | 2726 | else if (IN_RANGE (info->bitpos, merged_store->start, |
4d213bf6 JJ |
2727 | merged_store->start + merged_store->width - 1) |
2728 | /* |---store 1---||---store 2---| | |
2729 | Handle also the consecutive INTEGER_CST stores case here, | |
2730 | as we have here the code to deal with overlaps. */ | |
2731 | || (info->bitregion_start <= merged_store->bitregion_end | |
2732 | && info->rhs_code == INTEGER_CST | |
2733 | && merged_store->only_constants | |
2734 | && merged_store->can_be_merged_into (info))) | |
f663d9ad | 2735 | { |
245f6de1 | 2736 | /* Only allow overlapping stores of constants. */ |
629387a6 EB |
2737 | if (info->rhs_code == INTEGER_CST |
2738 | && merged_store->only_constants | |
2739 | && info->lp_nr == merged_store->lp_nr) | |
245f6de1 | 2740 | { |
6cd4c66e JJ |
2741 | unsigned int last_order |
2742 | = MAX (merged_store->last_order, info->order); | |
2743 | unsigned HOST_WIDE_INT end | |
2744 | = MAX (merged_store->start + merged_store->width, | |
2745 | info->bitpos + info->bitsize); | |
4d213bf6 | 2746 | if (check_no_overlap (m_store_info, i, true, last_order, end)) |
6cd4c66e JJ |
2747 | { |
2748 | /* check_no_overlap call above made sure there are no | |
2749 | overlapping stores with non-INTEGER_CST rhs_code | |
2750 | in between the first and last of the stores we've | |
2751 | just merged. If there are any INTEGER_CST rhs_code | |
2752 | stores in between, we need to merge_overlapping them | |
2753 | even if in the sort_by_bitpos order there are other | |
2754 | overlapping stores in between. Keep those stores as is. | |
2755 | Example: | |
2756 | MEM[(int *)p_28] = 0; | |
2757 | MEM[(char *)p_28 + 3B] = 1; | |
2758 | MEM[(char *)p_28 + 1B] = 2; | |
2759 | MEM[(char *)p_28 + 2B] = MEM[(char *)p_28 + 6B]; | |
2760 | We can't merge the zero store with the store of two and | |
2761 | not merge anything else, because the store of one is | |
2762 | in the original order in between those two, but in | |
2763 | store_by_bitpos order it comes after the last store that | |
2764 | we can't merge with them. We can merge the first 3 stores | |
2765 | and keep the last store as is though. */ | |
18e0c3d1 JJ |
2766 | unsigned int len = m_store_info.length (); |
2767 | unsigned int try_order = last_order; | |
2768 | unsigned int first_nonmergeable_order; | |
2769 | unsigned int k; | |
2770 | bool last_iter = false; | |
2771 | int attempts = 0; | |
2772 | do | |
6cd4c66e | 2773 | { |
18e0c3d1 JJ |
2774 | unsigned int max_order = 0; |
2775 | unsigned first_nonmergeable_int_order = ~0U; | |
2776 | unsigned HOST_WIDE_INT this_end = end; | |
2777 | k = i; | |
2778 | first_nonmergeable_order = ~0U; | |
2779 | for (unsigned int j = i + 1; j < len; ++j) | |
6cd4c66e | 2780 | { |
18e0c3d1 JJ |
2781 | store_immediate_info *info2 = m_store_info[j]; |
2782 | if (info2->bitpos >= this_end) | |
2783 | break; | |
2784 | if (info2->order < try_order) | |
6cd4c66e | 2785 | { |
4119cd69 JJ |
2786 | if (info2->rhs_code != INTEGER_CST |
2787 | || info2->lp_nr != merged_store->lp_nr) | |
18e0c3d1 JJ |
2788 | { |
2789 | /* Normally check_no_overlap makes sure this | |
2790 | doesn't happen, but if end grows below, | |
2791 | then we need to process more stores than | |
2792 | check_no_overlap verified. Example: | |
2793 | MEM[(int *)p_5] = 0; | |
2794 | MEM[(short *)p_5 + 3B] = 1; | |
2795 | MEM[(char *)p_5 + 4B] = _9; | |
2796 | MEM[(char *)p_5 + 2B] = 2; */ | |
2797 | k = 0; | |
2798 | break; | |
2799 | } | |
2800 | k = j; | |
2801 | this_end = MAX (this_end, | |
2802 | info2->bitpos + info2->bitsize); | |
6cd4c66e | 2803 | } |
18e0c3d1 | 2804 | else if (info2->rhs_code == INTEGER_CST |
4119cd69 | 2805 | && info2->lp_nr == merged_store->lp_nr |
18e0c3d1 JJ |
2806 | && !last_iter) |
2807 | { | |
2808 | max_order = MAX (max_order, info2->order + 1); | |
2809 | first_nonmergeable_int_order | |
2810 | = MIN (first_nonmergeable_int_order, | |
2811 | info2->order); | |
2812 | } | |
2813 | else | |
2814 | first_nonmergeable_order | |
2815 | = MIN (first_nonmergeable_order, info2->order); | |
6cd4c66e | 2816 | } |
18e0c3d1 JJ |
2817 | if (k == 0) |
2818 | { | |
2819 | if (last_order == try_order) | |
2820 | break; | |
2821 | /* If this failed, but only because we grew | |
2822 | try_order, retry with the last working one, | |
2823 | so that we merge at least something. */ | |
2824 | try_order = last_order; | |
2825 | last_iter = true; | |
2826 | continue; | |
2827 | } | |
2828 | last_order = try_order; | |
2829 | /* Retry with a larger try_order to see if we could | |
2830 | merge some further INTEGER_CST stores. */ | |
2831 | if (max_order | |
2832 | && (first_nonmergeable_int_order | |
2833 | < first_nonmergeable_order)) | |
2834 | { | |
2835 | try_order = MIN (max_order, | |
2836 | first_nonmergeable_order); | |
2837 | try_order | |
2838 | = MIN (try_order, | |
2839 | merged_store->first_nonmergeable_order); | |
2840 | if (try_order > last_order && ++attempts < 16) | |
2841 | continue; | |
2842 | } | |
2843 | first_nonmergeable_order | |
2844 | = MIN (first_nonmergeable_order, | |
2845 | first_nonmergeable_int_order); | |
2846 | end = this_end; | |
2847 | break; | |
6cd4c66e | 2848 | } |
18e0c3d1 | 2849 | while (1); |
6cd4c66e JJ |
2850 | |
2851 | if (k != 0) | |
2852 | { | |
2853 | merged_store->merge_overlapping (info); | |
2854 | ||
18e0c3d1 JJ |
2855 | merged_store->first_nonmergeable_order |
2856 | = MIN (merged_store->first_nonmergeable_order, | |
2857 | first_nonmergeable_order); | |
2858 | ||
6cd4c66e JJ |
2859 | for (unsigned int j = i + 1; j <= k; j++) |
2860 | { | |
2861 | store_immediate_info *info2 = m_store_info[j]; | |
2862 | gcc_assert (info2->bitpos < end); | |
2863 | if (info2->order < last_order) | |
2864 | { | |
2865 | gcc_assert (info2->rhs_code == INTEGER_CST); | |
18e0c3d1 JJ |
2866 | if (info != info2) |
2867 | merged_store->merge_overlapping (info2); | |
6cd4c66e JJ |
2868 | } |
2869 | /* Other stores are kept and not merged in any | |
2870 | way. */ | |
2871 | } | |
2872 | ignore = k; | |
2873 | goto done; | |
2874 | } | |
2875 | } | |
245f6de1 | 2876 | } |
f663d9ad | 2877 | } |
245f6de1 JJ |
2878 | /* |---store 1---||---store 2---| |
2879 | This store is consecutive to the previous one. | |
2880 | Merge it into the current store group. There can be gaps in between | |
2881 | the stores, but there can't be gaps in between bitregions. */ | |
c94c3532 | 2882 | else if (info->bitregion_start <= merged_store->bitregion_end |
7f5a3982 | 2883 | && merged_store->can_be_merged_into (info)) |
f663d9ad | 2884 | { |
245f6de1 JJ |
2885 | store_immediate_info *infof = merged_store->stores[0]; |
2886 | ||
2887 | /* All the rhs_code ops that take 2 operands are commutative, | |
2888 | swap the operands if it could make the operands compatible. */ | |
2889 | if (infof->ops[0].base_addr | |
2890 | && infof->ops[1].base_addr | |
2891 | && info->ops[0].base_addr | |
2892 | && info->ops[1].base_addr | |
8a91d545 RS |
2893 | && known_eq (info->ops[1].bitpos - infof->ops[0].bitpos, |
2894 | info->bitpos - infof->bitpos) | |
245f6de1 JJ |
2895 | && operand_equal_p (info->ops[1].base_addr, |
2896 | infof->ops[0].base_addr, 0)) | |
127ef369 JJ |
2897 | { |
2898 | std::swap (info->ops[0], info->ops[1]); | |
2899 | info->ops_swapped_p = true; | |
2900 | } | |
4d213bf6 | 2901 | if (check_no_overlap (m_store_info, i, false, |
a7fe6482 JJ |
2902 | MAX (merged_store->last_order, info->order), |
2903 | MAX (merged_store->start + merged_store->width, | |
7f5a3982 | 2904 | info->bitpos + info->bitsize))) |
245f6de1 | 2905 | { |
7f5a3982 EB |
2906 | /* Turn MEM_REF into BIT_INSERT_EXPR for bit-field stores. */ |
2907 | if (info->rhs_code == MEM_REF && infof->rhs_code != MEM_REF) | |
2908 | { | |
2909 | info->rhs_code = BIT_INSERT_EXPR; | |
2910 | info->ops[0].val = gimple_assign_rhs1 (info->stmt); | |
2911 | info->ops[0].base_addr = NULL_TREE; | |
2912 | } | |
2913 | else if (infof->rhs_code == MEM_REF && info->rhs_code != MEM_REF) | |
2914 | { | |
2915 | store_immediate_info *infoj; | |
2916 | unsigned int j; | |
2917 | FOR_EACH_VEC_ELT (merged_store->stores, j, infoj) | |
2918 | { | |
2919 | infoj->rhs_code = BIT_INSERT_EXPR; | |
2920 | infoj->ops[0].val = gimple_assign_rhs1 (infoj->stmt); | |
2921 | infoj->ops[0].base_addr = NULL_TREE; | |
2922 | } | |
2923 | } | |
2924 | if ((infof->ops[0].base_addr | |
2925 | ? compatible_load_p (merged_store, info, base_addr, 0) | |
2926 | : !info->ops[0].base_addr) | |
2927 | && (infof->ops[1].base_addr | |
2928 | ? compatible_load_p (merged_store, info, base_addr, 1) | |
2929 | : !info->ops[1].base_addr)) | |
2930 | { | |
2931 | merged_store->merge_into (info); | |
2932 | goto done; | |
2933 | } | |
245f6de1 JJ |
2934 | } |
2935 | } | |
f663d9ad | 2936 | |
245f6de1 JJ |
2937 | /* |---store 1---| <gap> |---store 2---|. |
2938 | Gap between stores or the rhs not compatible. Start a new group. */ | |
f663d9ad | 2939 | |
245f6de1 JJ |
2940 | /* Try to apply all the stores recorded for the group to determine |
2941 | the bitpattern they write and discard it if that fails. | |
2942 | This will also reject single-store groups. */ | |
c94c3532 | 2943 | if (merged_store->apply_stores ()) |
245f6de1 | 2944 | m_merged_store_groups.safe_push (merged_store); |
c94c3532 EB |
2945 | else |
2946 | delete merged_store; | |
f663d9ad | 2947 | |
245f6de1 | 2948 | merged_store = new merged_store_group (info); |
c94c3532 EB |
2949 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2950 | fputs ("New store group\n", dump_file); | |
2951 | ||
2952 | done: | |
2953 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2954 | { | |
2955 | fprintf (dump_file, "Store %u:\nbitsize:" HOST_WIDE_INT_PRINT_DEC | |
2956 | " bitpos:" HOST_WIDE_INT_PRINT_DEC " val:", | |
2957 | i, info->bitsize, info->bitpos); | |
2958 | print_generic_expr (dump_file, gimple_assign_rhs1 (info->stmt)); | |
2959 | fputc ('\n', dump_file); | |
2960 | } | |
f663d9ad KT |
2961 | } |
2962 | ||
a62b3dc5 | 2963 | /* Record or discard the last store group. */ |
4b84d9b8 JJ |
2964 | if (merged_store) |
2965 | { | |
c94c3532 | 2966 | if (merged_store->apply_stores ()) |
4b84d9b8 | 2967 | m_merged_store_groups.safe_push (merged_store); |
c94c3532 EB |
2968 | else |
2969 | delete merged_store; | |
4b84d9b8 | 2970 | } |
f663d9ad KT |
2971 | |
2972 | gcc_assert (m_merged_store_groups.length () <= m_store_info.length ()); | |
c94c3532 | 2973 | |
f663d9ad KT |
2974 | bool success |
2975 | = !m_merged_store_groups.is_empty () | |
2976 | && m_merged_store_groups.length () < m_store_info.length (); | |
2977 | ||
2978 | if (success && dump_file) | |
c94c3532 | 2979 | fprintf (dump_file, "Coalescing successful!\nMerged into %u stores\n", |
a62b3dc5 | 2980 | m_merged_store_groups.length ()); |
f663d9ad KT |
2981 | |
2982 | return success; | |
2983 | } | |
2984 | ||
245f6de1 JJ |
2985 | /* Return the type to use for the merged stores or loads described by STMTS. |
2986 | This is needed to get the alias sets right. If IS_LOAD, look for rhs, | |
2987 | otherwise lhs. Additionally set *CLIQUEP and *BASEP to MR_DEPENDENCE_* | |
2988 | of the MEM_REFs if any. */ | |
f663d9ad KT |
2989 | |
2990 | static tree | |
245f6de1 JJ |
2991 | get_alias_type_for_stmts (vec<gimple *> &stmts, bool is_load, |
2992 | unsigned short *cliquep, unsigned short *basep) | |
f663d9ad KT |
2993 | { |
2994 | gimple *stmt; | |
2995 | unsigned int i; | |
245f6de1 JJ |
2996 | tree type = NULL_TREE; |
2997 | tree ret = NULL_TREE; | |
2998 | *cliquep = 0; | |
2999 | *basep = 0; | |
f663d9ad KT |
3000 | |
3001 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
3002 | { | |
245f6de1 JJ |
3003 | tree ref = is_load ? gimple_assign_rhs1 (stmt) |
3004 | : gimple_assign_lhs (stmt); | |
3005 | tree type1 = reference_alias_ptr_type (ref); | |
3006 | tree base = get_base_address (ref); | |
f663d9ad | 3007 | |
245f6de1 JJ |
3008 | if (i == 0) |
3009 | { | |
3010 | if (TREE_CODE (base) == MEM_REF) | |
3011 | { | |
3012 | *cliquep = MR_DEPENDENCE_CLIQUE (base); | |
3013 | *basep = MR_DEPENDENCE_BASE (base); | |
3014 | } | |
3015 | ret = type = type1; | |
3016 | continue; | |
3017 | } | |
f663d9ad | 3018 | if (!alias_ptr_types_compatible_p (type, type1)) |
245f6de1 JJ |
3019 | ret = ptr_type_node; |
3020 | if (TREE_CODE (base) != MEM_REF | |
3021 | || *cliquep != MR_DEPENDENCE_CLIQUE (base) | |
3022 | || *basep != MR_DEPENDENCE_BASE (base)) | |
3023 | { | |
3024 | *cliquep = 0; | |
3025 | *basep = 0; | |
3026 | } | |
f663d9ad | 3027 | } |
245f6de1 | 3028 | return ret; |
f663d9ad KT |
3029 | } |
3030 | ||
3031 | /* Return the location_t information we can find among the statements | |
3032 | in STMTS. */ | |
3033 | ||
3034 | static location_t | |
245f6de1 | 3035 | get_location_for_stmts (vec<gimple *> &stmts) |
f663d9ad KT |
3036 | { |
3037 | gimple *stmt; | |
3038 | unsigned int i; | |
3039 | ||
3040 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
3041 | if (gimple_has_location (stmt)) | |
3042 | return gimple_location (stmt); | |
3043 | ||
3044 | return UNKNOWN_LOCATION; | |
3045 | } | |
3046 | ||
3047 | /* Used to decribe a store resulting from splitting a wide store in smaller | |
3048 | regularly-sized stores in split_group. */ | |
3049 | ||
6c1dae73 | 3050 | class split_store |
f663d9ad | 3051 | { |
6c1dae73 | 3052 | public: |
f663d9ad KT |
3053 | unsigned HOST_WIDE_INT bytepos; |
3054 | unsigned HOST_WIDE_INT size; | |
3055 | unsigned HOST_WIDE_INT align; | |
245f6de1 | 3056 | auto_vec<store_immediate_info *> orig_stores; |
a62b3dc5 JJ |
3057 | /* True if there is a single orig stmt covering the whole split store. */ |
3058 | bool orig; | |
f663d9ad KT |
3059 | split_store (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, |
3060 | unsigned HOST_WIDE_INT); | |
3061 | }; | |
3062 | ||
3063 | /* Simple constructor. */ | |
3064 | ||
3065 | split_store::split_store (unsigned HOST_WIDE_INT bp, | |
3066 | unsigned HOST_WIDE_INT sz, | |
3067 | unsigned HOST_WIDE_INT al) | |
a62b3dc5 | 3068 | : bytepos (bp), size (sz), align (al), orig (false) |
f663d9ad | 3069 | { |
245f6de1 | 3070 | orig_stores.create (0); |
f663d9ad KT |
3071 | } |
3072 | ||
245f6de1 JJ |
3073 | /* Record all stores in GROUP that write to the region starting at BITPOS and |
3074 | is of size BITSIZE. Record infos for such statements in STORES if | |
3075 | non-NULL. The stores in GROUP must be sorted by bitposition. Return INFO | |
5384a802 JJ |
3076 | if there is exactly one original store in the range (in that case ignore |
3077 | clobber stmts, unless there are only clobber stmts). */ | |
f663d9ad | 3078 | |
a62b3dc5 | 3079 | static store_immediate_info * |
99b1c316 | 3080 | find_constituent_stores (class merged_store_group *group, |
245f6de1 JJ |
3081 | vec<store_immediate_info *> *stores, |
3082 | unsigned int *first, | |
3083 | unsigned HOST_WIDE_INT bitpos, | |
3084 | unsigned HOST_WIDE_INT bitsize) | |
f663d9ad | 3085 | { |
a62b3dc5 | 3086 | store_immediate_info *info, *ret = NULL; |
f663d9ad | 3087 | unsigned int i; |
a62b3dc5 JJ |
3088 | bool second = false; |
3089 | bool update_first = true; | |
f663d9ad | 3090 | unsigned HOST_WIDE_INT end = bitpos + bitsize; |
a62b3dc5 | 3091 | for (i = *first; group->stores.iterate (i, &info); ++i) |
f663d9ad KT |
3092 | { |
3093 | unsigned HOST_WIDE_INT stmt_start = info->bitpos; | |
3094 | unsigned HOST_WIDE_INT stmt_end = stmt_start + info->bitsize; | |
a62b3dc5 JJ |
3095 | if (stmt_end <= bitpos) |
3096 | { | |
3097 | /* BITPOS passed to this function never decreases from within the | |
3098 | same split_group call, so optimize and don't scan info records | |
3099 | which are known to end before or at BITPOS next time. | |
3100 | Only do it if all stores before this one also pass this. */ | |
3101 | if (update_first) | |
3102 | *first = i + 1; | |
3103 | continue; | |
3104 | } | |
3105 | else | |
3106 | update_first = false; | |
3107 | ||
f663d9ad | 3108 | /* The stores in GROUP are ordered by bitposition so if we're past |
a62b3dc5 JJ |
3109 | the region for this group return early. */ |
3110 | if (stmt_start >= end) | |
3111 | return ret; | |
3112 | ||
5384a802 JJ |
3113 | if (gimple_clobber_p (info->stmt)) |
3114 | { | |
3115 | if (stores) | |
3116 | stores->safe_push (info); | |
3117 | if (ret == NULL) | |
3118 | ret = info; | |
3119 | continue; | |
3120 | } | |
245f6de1 | 3121 | if (stores) |
a62b3dc5 | 3122 | { |
245f6de1 | 3123 | stores->safe_push (info); |
5384a802 | 3124 | if (ret && !gimple_clobber_p (ret->stmt)) |
a62b3dc5 JJ |
3125 | { |
3126 | ret = NULL; | |
3127 | second = true; | |
3128 | } | |
3129 | } | |
5384a802 | 3130 | else if (ret && !gimple_clobber_p (ret->stmt)) |
a62b3dc5 JJ |
3131 | return NULL; |
3132 | if (!second) | |
3133 | ret = info; | |
f663d9ad | 3134 | } |
a62b3dc5 | 3135 | return ret; |
f663d9ad KT |
3136 | } |
3137 | ||
d7a9512e JJ |
3138 | /* Return how many SSA_NAMEs used to compute value to store in the INFO |
3139 | store have multiple uses. If any SSA_NAME has multiple uses, also | |
3140 | count statements needed to compute it. */ | |
3141 | ||
3142 | static unsigned | |
3143 | count_multiple_uses (store_immediate_info *info) | |
3144 | { | |
3145 | gimple *stmt = info->stmt; | |
3146 | unsigned ret = 0; | |
3147 | switch (info->rhs_code) | |
3148 | { | |
3149 | case INTEGER_CST: | |
3150 | return 0; | |
3151 | case BIT_AND_EXPR: | |
3152 | case BIT_IOR_EXPR: | |
3153 | case BIT_XOR_EXPR: | |
d60edaba JJ |
3154 | if (info->bit_not_p) |
3155 | { | |
3156 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
3157 | ret = 1; /* Fall through below to return | |
3158 | the BIT_NOT_EXPR stmt and then | |
3159 | BIT_{AND,IOR,XOR}_EXPR and anything it | |
3160 | uses. */ | |
3161 | else | |
3162 | /* stmt is after this the BIT_NOT_EXPR. */ | |
3163 | stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3164 | } | |
d7a9512e JJ |
3165 | if (!has_single_use (gimple_assign_rhs1 (stmt))) |
3166 | { | |
3167 | ret += 1 + info->ops[0].bit_not_p; | |
3168 | if (info->ops[1].base_addr) | |
3169 | ret += 1 + info->ops[1].bit_not_p; | |
3170 | return ret + 1; | |
3171 | } | |
3172 | stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3173 | /* stmt is now the BIT_*_EXPR. */ | |
3174 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
127ef369 JJ |
3175 | ret += 1 + info->ops[info->ops_swapped_p].bit_not_p; |
3176 | else if (info->ops[info->ops_swapped_p].bit_not_p) | |
d7a9512e JJ |
3177 | { |
3178 | gimple *stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3179 | if (!has_single_use (gimple_assign_rhs1 (stmt2))) | |
3180 | ++ret; | |
3181 | } | |
3182 | if (info->ops[1].base_addr == NULL_TREE) | |
127ef369 JJ |
3183 | { |
3184 | gcc_checking_assert (!info->ops_swapped_p); | |
3185 | return ret; | |
3186 | } | |
d7a9512e | 3187 | if (!has_single_use (gimple_assign_rhs2 (stmt))) |
127ef369 JJ |
3188 | ret += 1 + info->ops[1 - info->ops_swapped_p].bit_not_p; |
3189 | else if (info->ops[1 - info->ops_swapped_p].bit_not_p) | |
d7a9512e JJ |
3190 | { |
3191 | gimple *stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs2 (stmt)); | |
3192 | if (!has_single_use (gimple_assign_rhs1 (stmt2))) | |
3193 | ++ret; | |
3194 | } | |
3195 | return ret; | |
3196 | case MEM_REF: | |
3197 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
3198 | return 1 + info->ops[0].bit_not_p; | |
3199 | else if (info->ops[0].bit_not_p) | |
3200 | { | |
3201 | stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
3202 | if (!has_single_use (gimple_assign_rhs1 (stmt))) | |
3203 | return 1; | |
3204 | } | |
3205 | return 0; | |
c94c3532 EB |
3206 | case BIT_INSERT_EXPR: |
3207 | return has_single_use (gimple_assign_rhs1 (stmt)) ? 0 : 1; | |
d7a9512e JJ |
3208 | default: |
3209 | gcc_unreachable (); | |
3210 | } | |
3211 | } | |
3212 | ||
f663d9ad | 3213 | /* Split a merged store described by GROUP by populating the SPLIT_STORES |
a62b3dc5 JJ |
3214 | vector (if non-NULL) with split_store structs describing the byte offset |
3215 | (from the base), the bit size and alignment of each store as well as the | |
3216 | original statements involved in each such split group. | |
f663d9ad KT |
3217 | This is to separate the splitting strategy from the statement |
3218 | building/emission/linking done in output_merged_store. | |
a62b3dc5 | 3219 | Return number of new stores. |
245f6de1 JJ |
3220 | If ALLOW_UNALIGNED_STORE is false, then all stores must be aligned. |
3221 | If ALLOW_UNALIGNED_LOAD is false, then all loads must be aligned. | |
3afd514b JJ |
3222 | BZERO_FIRST may be true only when the first store covers the whole group |
3223 | and clears it; if BZERO_FIRST is true, keep that first store in the set | |
3224 | unmodified and emit further stores for the overrides only. | |
a62b3dc5 JJ |
3225 | If SPLIT_STORES is NULL, it is just a dry run to count number of |
3226 | new stores. */ | |
f663d9ad | 3227 | |
a62b3dc5 | 3228 | static unsigned int |
245f6de1 | 3229 | split_group (merged_store_group *group, bool allow_unaligned_store, |
3afd514b | 3230 | bool allow_unaligned_load, bool bzero_first, |
99b1c316 | 3231 | vec<split_store *> *split_stores, |
d7a9512e JJ |
3232 | unsigned *total_orig, |
3233 | unsigned *total_new) | |
f663d9ad | 3234 | { |
a62b3dc5 JJ |
3235 | unsigned HOST_WIDE_INT pos = group->bitregion_start; |
3236 | unsigned HOST_WIDE_INT size = group->bitregion_end - pos; | |
f663d9ad | 3237 | unsigned HOST_WIDE_INT bytepos = pos / BITS_PER_UNIT; |
a62b3dc5 JJ |
3238 | unsigned HOST_WIDE_INT group_align = group->align; |
3239 | unsigned HOST_WIDE_INT align_base = group->align_base; | |
245f6de1 | 3240 | unsigned HOST_WIDE_INT group_load_align = group_align; |
d7a9512e | 3241 | bool any_orig = false; |
f663d9ad | 3242 | |
f663d9ad KT |
3243 | gcc_assert ((size % BITS_PER_UNIT == 0) && (pos % BITS_PER_UNIT == 0)); |
3244 | ||
4b84d9b8 JJ |
3245 | if (group->stores[0]->rhs_code == LROTATE_EXPR |
3246 | || group->stores[0]->rhs_code == NOP_EXPR) | |
3247 | { | |
3afd514b | 3248 | gcc_assert (!bzero_first); |
4b84d9b8 JJ |
3249 | /* For bswap framework using sets of stores, all the checking |
3250 | has been done earlier in try_coalesce_bswap and needs to be | |
3251 | emitted as a single store. */ | |
3252 | if (total_orig) | |
3253 | { | |
3254 | /* Avoid the old/new stmt count heuristics. It should be | |
3255 | always beneficial. */ | |
3256 | total_new[0] = 1; | |
3257 | total_orig[0] = 2; | |
3258 | } | |
3259 | ||
3260 | if (split_stores) | |
3261 | { | |
3262 | unsigned HOST_WIDE_INT align_bitpos | |
3263 | = (group->start - align_base) & (group_align - 1); | |
3264 | unsigned HOST_WIDE_INT align = group_align; | |
3265 | if (align_bitpos) | |
3266 | align = least_bit_hwi (align_bitpos); | |
3267 | bytepos = group->start / BITS_PER_UNIT; | |
99b1c316 | 3268 | split_store *store |
4b84d9b8 JJ |
3269 | = new split_store (bytepos, group->width, align); |
3270 | unsigned int first = 0; | |
3271 | find_constituent_stores (group, &store->orig_stores, | |
3272 | &first, group->start, group->width); | |
3273 | split_stores->safe_push (store); | |
3274 | } | |
3275 | ||
3276 | return 1; | |
3277 | } | |
3278 | ||
a62b3dc5 | 3279 | unsigned int ret = 0, first = 0; |
f663d9ad | 3280 | unsigned HOST_WIDE_INT try_pos = bytepos; |
f663d9ad | 3281 | |
d7a9512e JJ |
3282 | if (total_orig) |
3283 | { | |
3284 | unsigned int i; | |
3285 | store_immediate_info *info = group->stores[0]; | |
3286 | ||
3287 | total_new[0] = 0; | |
3288 | total_orig[0] = 1; /* The orig store. */ | |
3289 | info = group->stores[0]; | |
3290 | if (info->ops[0].base_addr) | |
a6fbd154 | 3291 | total_orig[0]++; |
d7a9512e | 3292 | if (info->ops[1].base_addr) |
a6fbd154 | 3293 | total_orig[0]++; |
d7a9512e JJ |
3294 | switch (info->rhs_code) |
3295 | { | |
3296 | case BIT_AND_EXPR: | |
3297 | case BIT_IOR_EXPR: | |
3298 | case BIT_XOR_EXPR: | |
3299 | total_orig[0]++; /* The orig BIT_*_EXPR stmt. */ | |
3300 | break; | |
3301 | default: | |
3302 | break; | |
3303 | } | |
3304 | total_orig[0] *= group->stores.length (); | |
3305 | ||
3306 | FOR_EACH_VEC_ELT (group->stores, i, info) | |
a6fbd154 JJ |
3307 | { |
3308 | total_new[0] += count_multiple_uses (info); | |
3309 | total_orig[0] += (info->bit_not_p | |
3310 | + info->ops[0].bit_not_p | |
3311 | + info->ops[1].bit_not_p); | |
3312 | } | |
d7a9512e JJ |
3313 | } |
3314 | ||
245f6de1 JJ |
3315 | if (!allow_unaligned_load) |
3316 | for (int i = 0; i < 2; ++i) | |
3317 | if (group->load_align[i]) | |
3318 | group_load_align = MIN (group_load_align, group->load_align[i]); | |
3319 | ||
3afd514b JJ |
3320 | if (bzero_first) |
3321 | { | |
5384a802 JJ |
3322 | store_immediate_info *gstore; |
3323 | FOR_EACH_VEC_ELT (group->stores, first, gstore) | |
3324 | if (!gimple_clobber_p (gstore->stmt)) | |
3325 | break; | |
3326 | ++first; | |
3afd514b JJ |
3327 | ret = 1; |
3328 | if (split_stores) | |
3329 | { | |
99b1c316 | 3330 | split_store *store |
5384a802 JJ |
3331 | = new split_store (bytepos, gstore->bitsize, align_base); |
3332 | store->orig_stores.safe_push (gstore); | |
3afd514b JJ |
3333 | store->orig = true; |
3334 | any_orig = true; | |
3335 | split_stores->safe_push (store); | |
3336 | } | |
3337 | } | |
3338 | ||
f663d9ad KT |
3339 | while (size > 0) |
3340 | { | |
245f6de1 | 3341 | if ((allow_unaligned_store || group_align <= BITS_PER_UNIT) |
3afd514b JJ |
3342 | && (group->mask[try_pos - bytepos] == (unsigned char) ~0U |
3343 | || (bzero_first && group->val[try_pos - bytepos] == 0))) | |
a62b3dc5 JJ |
3344 | { |
3345 | /* Skip padding bytes. */ | |
3346 | ++try_pos; | |
3347 | size -= BITS_PER_UNIT; | |
3348 | continue; | |
3349 | } | |
3350 | ||
f663d9ad | 3351 | unsigned HOST_WIDE_INT try_bitpos = try_pos * BITS_PER_UNIT; |
a62b3dc5 JJ |
3352 | unsigned int try_size = MAX_STORE_BITSIZE, nonmasked; |
3353 | unsigned HOST_WIDE_INT align_bitpos | |
3354 | = (try_bitpos - align_base) & (group_align - 1); | |
3355 | unsigned HOST_WIDE_INT align = group_align; | |
5384a802 | 3356 | bool found_orig = false; |
a62b3dc5 JJ |
3357 | if (align_bitpos) |
3358 | align = least_bit_hwi (align_bitpos); | |
245f6de1 | 3359 | if (!allow_unaligned_store) |
a62b3dc5 | 3360 | try_size = MIN (try_size, align); |
245f6de1 JJ |
3361 | if (!allow_unaligned_load) |
3362 | { | |
3363 | /* If we can't do or don't want to do unaligned stores | |
3364 | as well as loads, we need to take the loads into account | |
3365 | as well. */ | |
3366 | unsigned HOST_WIDE_INT load_align = group_load_align; | |
3367 | align_bitpos = (try_bitpos - align_base) & (load_align - 1); | |
3368 | if (align_bitpos) | |
3369 | load_align = least_bit_hwi (align_bitpos); | |
3370 | for (int i = 0; i < 2; ++i) | |
3371 | if (group->load_align[i]) | |
3372 | { | |
8a91d545 RS |
3373 | align_bitpos |
3374 | = known_alignment (try_bitpos | |
3375 | - group->stores[0]->bitpos | |
3376 | + group->stores[0]->ops[i].bitpos | |
3377 | - group->load_align_base[i]); | |
3378 | if (align_bitpos & (group_load_align - 1)) | |
245f6de1 JJ |
3379 | { |
3380 | unsigned HOST_WIDE_INT a = least_bit_hwi (align_bitpos); | |
3381 | load_align = MIN (load_align, a); | |
3382 | } | |
3383 | } | |
3384 | try_size = MIN (try_size, load_align); | |
3385 | } | |
a62b3dc5 | 3386 | store_immediate_info *info |
245f6de1 | 3387 | = find_constituent_stores (group, NULL, &first, try_bitpos, try_size); |
5384a802 | 3388 | if (info && !gimple_clobber_p (info->stmt)) |
a62b3dc5 JJ |
3389 | { |
3390 | /* If there is just one original statement for the range, see if | |
3391 | we can just reuse the original store which could be even larger | |
3392 | than try_size. */ | |
3393 | unsigned HOST_WIDE_INT stmt_end | |
3394 | = ROUND_UP (info->bitpos + info->bitsize, BITS_PER_UNIT); | |
245f6de1 JJ |
3395 | info = find_constituent_stores (group, NULL, &first, try_bitpos, |
3396 | stmt_end - try_bitpos); | |
a62b3dc5 JJ |
3397 | if (info && info->bitpos >= try_bitpos) |
3398 | { | |
5384a802 JJ |
3399 | store_immediate_info *info2 = NULL; |
3400 | unsigned int first_copy = first; | |
3401 | if (info->bitpos > try_bitpos | |
3402 | && stmt_end - try_bitpos <= try_size) | |
3403 | { | |
3404 | info2 = find_constituent_stores (group, NULL, &first_copy, | |
3405 | try_bitpos, | |
3406 | info->bitpos - try_bitpos); | |
3407 | gcc_assert (info2 == NULL || gimple_clobber_p (info2->stmt)); | |
3408 | } | |
3409 | if (info2 == NULL && stmt_end - try_bitpos < try_size) | |
3410 | { | |
3411 | info2 = find_constituent_stores (group, NULL, &first_copy, | |
3412 | stmt_end, | |
3413 | (try_bitpos + try_size) | |
3414 | - stmt_end); | |
3415 | gcc_assert (info2 == NULL || gimple_clobber_p (info2->stmt)); | |
3416 | } | |
3417 | if (info2 == NULL) | |
3418 | { | |
3419 | try_size = stmt_end - try_bitpos; | |
3420 | found_orig = true; | |
3421 | goto found; | |
3422 | } | |
a62b3dc5 JJ |
3423 | } |
3424 | } | |
f663d9ad | 3425 | |
a62b3dc5 JJ |
3426 | /* Approximate store bitsize for the case when there are no padding |
3427 | bits. */ | |
3428 | while (try_size > size) | |
3429 | try_size /= 2; | |
3430 | /* Now look for whole padding bytes at the end of that bitsize. */ | |
3431 | for (nonmasked = try_size / BITS_PER_UNIT; nonmasked > 0; --nonmasked) | |
3432 | if (group->mask[try_pos - bytepos + nonmasked - 1] | |
3afd514b JJ |
3433 | != (unsigned char) ~0U |
3434 | && (!bzero_first | |
3435 | || group->val[try_pos - bytepos + nonmasked - 1] != 0)) | |
a62b3dc5 | 3436 | break; |
5384a802 | 3437 | if (nonmasked == 0 || (info && gimple_clobber_p (info->stmt))) |
a62b3dc5 JJ |
3438 | { |
3439 | /* If entire try_size range is padding, skip it. */ | |
3440 | try_pos += try_size / BITS_PER_UNIT; | |
3441 | size -= try_size; | |
3442 | continue; | |
3443 | } | |
3444 | /* Otherwise try to decrease try_size if second half, last 3 quarters | |
3445 | etc. are padding. */ | |
3446 | nonmasked *= BITS_PER_UNIT; | |
3447 | while (nonmasked <= try_size / 2) | |
3448 | try_size /= 2; | |
245f6de1 | 3449 | if (!allow_unaligned_store && group_align > BITS_PER_UNIT) |
a62b3dc5 JJ |
3450 | { |
3451 | /* Now look for whole padding bytes at the start of that bitsize. */ | |
3452 | unsigned int try_bytesize = try_size / BITS_PER_UNIT, masked; | |
3453 | for (masked = 0; masked < try_bytesize; ++masked) | |
3afd514b JJ |
3454 | if (group->mask[try_pos - bytepos + masked] != (unsigned char) ~0U |
3455 | && (!bzero_first | |
3456 | || group->val[try_pos - bytepos + masked] != 0)) | |
a62b3dc5 JJ |
3457 | break; |
3458 | masked *= BITS_PER_UNIT; | |
3459 | gcc_assert (masked < try_size); | |
3460 | if (masked >= try_size / 2) | |
3461 | { | |
3462 | while (masked >= try_size / 2) | |
3463 | { | |
3464 | try_size /= 2; | |
3465 | try_pos += try_size / BITS_PER_UNIT; | |
3466 | size -= try_size; | |
3467 | masked -= try_size; | |
3468 | } | |
3469 | /* Need to recompute the alignment, so just retry at the new | |
3470 | position. */ | |
3471 | continue; | |
3472 | } | |
3473 | } | |
3474 | ||
3475 | found: | |
3476 | ++ret; | |
f663d9ad | 3477 | |
a62b3dc5 JJ |
3478 | if (split_stores) |
3479 | { | |
99b1c316 | 3480 | split_store *store |
a62b3dc5 | 3481 | = new split_store (try_pos, try_size, align); |
245f6de1 JJ |
3482 | info = find_constituent_stores (group, &store->orig_stores, |
3483 | &first, try_bitpos, try_size); | |
a62b3dc5 | 3484 | if (info |
5384a802 | 3485 | && !gimple_clobber_p (info->stmt) |
a62b3dc5 | 3486 | && info->bitpos >= try_bitpos |
5384a802 JJ |
3487 | && info->bitpos + info->bitsize <= try_bitpos + try_size |
3488 | && (store->orig_stores.length () == 1 | |
3489 | || found_orig | |
3490 | || (info->bitpos == try_bitpos | |
3491 | && (info->bitpos + info->bitsize | |
3492 | == try_bitpos + try_size)))) | |
d7a9512e JJ |
3493 | { |
3494 | store->orig = true; | |
3495 | any_orig = true; | |
3496 | } | |
a62b3dc5 JJ |
3497 | split_stores->safe_push (store); |
3498 | } | |
3499 | ||
3500 | try_pos += try_size / BITS_PER_UNIT; | |
f663d9ad | 3501 | size -= try_size; |
f663d9ad | 3502 | } |
a62b3dc5 | 3503 | |
d7a9512e JJ |
3504 | if (total_orig) |
3505 | { | |
a6fbd154 | 3506 | unsigned int i; |
99b1c316 | 3507 | split_store *store; |
d7a9512e JJ |
3508 | /* If we are reusing some original stores and any of the |
3509 | original SSA_NAMEs had multiple uses, we need to subtract | |
3510 | those now before we add the new ones. */ | |
3511 | if (total_new[0] && any_orig) | |
3512 | { | |
d7a9512e JJ |
3513 | FOR_EACH_VEC_ELT (*split_stores, i, store) |
3514 | if (store->orig) | |
3515 | total_new[0] -= count_multiple_uses (store->orig_stores[0]); | |
3516 | } | |
3517 | total_new[0] += ret; /* The new store. */ | |
3518 | store_immediate_info *info = group->stores[0]; | |
3519 | if (info->ops[0].base_addr) | |
a6fbd154 | 3520 | total_new[0] += ret; |
d7a9512e | 3521 | if (info->ops[1].base_addr) |
a6fbd154 | 3522 | total_new[0] += ret; |
d7a9512e JJ |
3523 | switch (info->rhs_code) |
3524 | { | |
3525 | case BIT_AND_EXPR: | |
3526 | case BIT_IOR_EXPR: | |
3527 | case BIT_XOR_EXPR: | |
3528 | total_new[0] += ret; /* The new BIT_*_EXPR stmt. */ | |
3529 | break; | |
3530 | default: | |
3531 | break; | |
3532 | } | |
a6fbd154 JJ |
3533 | FOR_EACH_VEC_ELT (*split_stores, i, store) |
3534 | { | |
3535 | unsigned int j; | |
3536 | bool bit_not_p[3] = { false, false, false }; | |
3537 | /* If all orig_stores have certain bit_not_p set, then | |
3538 | we'd use a BIT_NOT_EXPR stmt and need to account for it. | |
3539 | If some orig_stores have certain bit_not_p set, then | |
3540 | we'd use a BIT_XOR_EXPR with a mask and need to account for | |
3541 | it. */ | |
3542 | FOR_EACH_VEC_ELT (store->orig_stores, j, info) | |
3543 | { | |
3544 | if (info->ops[0].bit_not_p) | |
3545 | bit_not_p[0] = true; | |
3546 | if (info->ops[1].bit_not_p) | |
3547 | bit_not_p[1] = true; | |
3548 | if (info->bit_not_p) | |
3549 | bit_not_p[2] = true; | |
3550 | } | |
3551 | total_new[0] += bit_not_p[0] + bit_not_p[1] + bit_not_p[2]; | |
3552 | } | |
3553 | ||
d7a9512e JJ |
3554 | } |
3555 | ||
a62b3dc5 | 3556 | return ret; |
f663d9ad KT |
3557 | } |
3558 | ||
a6fbd154 JJ |
3559 | /* Return the operation through which the operand IDX (if < 2) or |
3560 | result (IDX == 2) should be inverted. If NOP_EXPR, no inversion | |
3561 | is done, if BIT_NOT_EXPR, all bits are inverted, if BIT_XOR_EXPR, | |
3562 | the bits should be xored with mask. */ | |
3563 | ||
3564 | static enum tree_code | |
3565 | invert_op (split_store *split_store, int idx, tree int_type, tree &mask) | |
3566 | { | |
3567 | unsigned int i; | |
3568 | store_immediate_info *info; | |
3569 | unsigned int cnt = 0; | |
e215422f | 3570 | bool any_paddings = false; |
a6fbd154 JJ |
3571 | FOR_EACH_VEC_ELT (split_store->orig_stores, i, info) |
3572 | { | |
3573 | bool bit_not_p = idx < 2 ? info->ops[idx].bit_not_p : info->bit_not_p; | |
3574 | if (bit_not_p) | |
e215422f JJ |
3575 | { |
3576 | ++cnt; | |
3577 | tree lhs = gimple_assign_lhs (info->stmt); | |
3578 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)) | |
3579 | && TYPE_PRECISION (TREE_TYPE (lhs)) < info->bitsize) | |
3580 | any_paddings = true; | |
3581 | } | |
a6fbd154 JJ |
3582 | } |
3583 | mask = NULL_TREE; | |
3584 | if (cnt == 0) | |
3585 | return NOP_EXPR; | |
e215422f | 3586 | if (cnt == split_store->orig_stores.length () && !any_paddings) |
a6fbd154 JJ |
3587 | return BIT_NOT_EXPR; |
3588 | ||
3589 | unsigned HOST_WIDE_INT try_bitpos = split_store->bytepos * BITS_PER_UNIT; | |
3590 | unsigned buf_size = split_store->size / BITS_PER_UNIT; | |
3591 | unsigned char *buf | |
3592 | = XALLOCAVEC (unsigned char, buf_size); | |
3593 | memset (buf, ~0U, buf_size); | |
3594 | FOR_EACH_VEC_ELT (split_store->orig_stores, i, info) | |
3595 | { | |
3596 | bool bit_not_p = idx < 2 ? info->ops[idx].bit_not_p : info->bit_not_p; | |
3597 | if (!bit_not_p) | |
3598 | continue; | |
3599 | /* Clear regions with bit_not_p and invert afterwards, rather than | |
3600 | clear regions with !bit_not_p, so that gaps in between stores aren't | |
3601 | set in the mask. */ | |
3602 | unsigned HOST_WIDE_INT bitsize = info->bitsize; | |
e215422f | 3603 | unsigned HOST_WIDE_INT prec = bitsize; |
a6fbd154 | 3604 | unsigned int pos_in_buffer = 0; |
e215422f JJ |
3605 | if (any_paddings) |
3606 | { | |
3607 | tree lhs = gimple_assign_lhs (info->stmt); | |
3608 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)) | |
3609 | && TYPE_PRECISION (TREE_TYPE (lhs)) < bitsize) | |
3610 | prec = TYPE_PRECISION (TREE_TYPE (lhs)); | |
3611 | } | |
a6fbd154 JJ |
3612 | if (info->bitpos < try_bitpos) |
3613 | { | |
3614 | gcc_assert (info->bitpos + bitsize > try_bitpos); | |
e215422f JJ |
3615 | if (!BYTES_BIG_ENDIAN) |
3616 | { | |
3617 | if (prec <= try_bitpos - info->bitpos) | |
3618 | continue; | |
3619 | prec -= try_bitpos - info->bitpos; | |
3620 | } | |
3621 | bitsize -= try_bitpos - info->bitpos; | |
3622 | if (BYTES_BIG_ENDIAN && prec > bitsize) | |
3623 | prec = bitsize; | |
a6fbd154 JJ |
3624 | } |
3625 | else | |
3626 | pos_in_buffer = info->bitpos - try_bitpos; | |
e215422f JJ |
3627 | if (prec < bitsize) |
3628 | { | |
3629 | /* If this is a bool inversion, invert just the least significant | |
3630 | prec bits rather than all bits of it. */ | |
3631 | if (BYTES_BIG_ENDIAN) | |
3632 | { | |
3633 | pos_in_buffer += bitsize - prec; | |
3634 | if (pos_in_buffer >= split_store->size) | |
3635 | continue; | |
3636 | } | |
3637 | bitsize = prec; | |
3638 | } | |
a6fbd154 JJ |
3639 | if (pos_in_buffer + bitsize > split_store->size) |
3640 | bitsize = split_store->size - pos_in_buffer; | |
3641 | unsigned char *p = buf + (pos_in_buffer / BITS_PER_UNIT); | |
3642 | if (BYTES_BIG_ENDIAN) | |
3643 | clear_bit_region_be (p, (BITS_PER_UNIT - 1 | |
3644 | - (pos_in_buffer % BITS_PER_UNIT)), bitsize); | |
3645 | else | |
3646 | clear_bit_region (p, pos_in_buffer % BITS_PER_UNIT, bitsize); | |
3647 | } | |
3648 | for (unsigned int i = 0; i < buf_size; ++i) | |
3649 | buf[i] = ~buf[i]; | |
3650 | mask = native_interpret_expr (int_type, buf, buf_size); | |
3651 | return BIT_XOR_EXPR; | |
3652 | } | |
3653 | ||
f663d9ad KT |
3654 | /* Given a merged store group GROUP output the widened version of it. |
3655 | The store chain is against the base object BASE. | |
3656 | Try store sizes of at most MAX_STORE_BITSIZE bits wide and don't output | |
3657 | unaligned stores for STRICT_ALIGNMENT targets or if it's too expensive. | |
3658 | Make sure that the number of statements output is less than the number of | |
3659 | original statements. If a better sequence is possible emit it and | |
3660 | return true. */ | |
3661 | ||
3662 | bool | |
b5926e23 | 3663 | imm_store_chain_info::output_merged_store (merged_store_group *group) |
f663d9ad | 3664 | { |
dd172744 RB |
3665 | split_store *split_store; |
3666 | unsigned int i; | |
a62b3dc5 JJ |
3667 | unsigned HOST_WIDE_INT start_byte_pos |
3668 | = group->bitregion_start / BITS_PER_UNIT; | |
f663d9ad KT |
3669 | |
3670 | unsigned int orig_num_stmts = group->stores.length (); | |
3671 | if (orig_num_stmts < 2) | |
3672 | return false; | |
3673 | ||
99b1c316 | 3674 | auto_vec<class split_store *, 32> split_stores; |
245f6de1 | 3675 | bool allow_unaligned_store |
028d4092 | 3676 | = !STRICT_ALIGNMENT && param_store_merging_allow_unaligned; |
245f6de1 | 3677 | bool allow_unaligned_load = allow_unaligned_store; |
3afd514b | 3678 | bool bzero_first = false; |
5384a802 JJ |
3679 | store_immediate_info *store; |
3680 | unsigned int num_clobber_stmts = 0; | |
3681 | if (group->stores[0]->rhs_code == INTEGER_CST) | |
3682 | { | |
3683 | FOR_EACH_VEC_ELT (group->stores, i, store) | |
3684 | if (gimple_clobber_p (store->stmt)) | |
3685 | num_clobber_stmts++; | |
3686 | else if (TREE_CODE (gimple_assign_rhs1 (store->stmt)) == CONSTRUCTOR | |
3687 | && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (store->stmt)) == 0 | |
3688 | && group->start == store->bitpos | |
3689 | && group->width == store->bitsize | |
3690 | && (group->start % BITS_PER_UNIT) == 0 | |
3691 | && (group->width % BITS_PER_UNIT) == 0) | |
3692 | { | |
3693 | bzero_first = true; | |
3694 | break; | |
3695 | } | |
3696 | else | |
3697 | break; | |
3698 | FOR_EACH_VEC_ELT_FROM (group->stores, i, store, i) | |
3699 | if (gimple_clobber_p (store->stmt)) | |
3700 | num_clobber_stmts++; | |
3701 | if (num_clobber_stmts == orig_num_stmts) | |
3702 | return false; | |
3703 | orig_num_stmts -= num_clobber_stmts; | |
3704 | } | |
3afd514b | 3705 | if (allow_unaligned_store || bzero_first) |
a62b3dc5 JJ |
3706 | { |
3707 | /* If unaligned stores are allowed, see how many stores we'd emit | |
3708 | for unaligned and how many stores we'd emit for aligned stores. | |
3afd514b JJ |
3709 | Only use unaligned stores if it allows fewer stores than aligned. |
3710 | Similarly, if there is a whole region clear first, prefer expanding | |
3711 | it together compared to expanding clear first followed by merged | |
3712 | further stores. */ | |
3713 | unsigned cnt[4] = { ~0, ~0, ~0, ~0 }; | |
3714 | int pass_min = 0; | |
3715 | for (int pass = 0; pass < 4; ++pass) | |
3716 | { | |
3717 | if (!allow_unaligned_store && (pass & 1) != 0) | |
3718 | continue; | |
3719 | if (!bzero_first && (pass & 2) != 0) | |
3720 | continue; | |
3721 | cnt[pass] = split_group (group, (pass & 1) != 0, | |
3722 | allow_unaligned_load, (pass & 2) != 0, | |
3723 | NULL, NULL, NULL); | |
3724 | if (cnt[pass] < cnt[pass_min]) | |
3725 | pass_min = pass; | |
3726 | } | |
3727 | if ((pass_min & 1) == 0) | |
245f6de1 | 3728 | allow_unaligned_store = false; |
3afd514b JJ |
3729 | if ((pass_min & 2) == 0) |
3730 | bzero_first = false; | |
a62b3dc5 | 3731 | } |
d7a9512e | 3732 | unsigned total_orig, total_new; |
3afd514b | 3733 | split_group (group, allow_unaligned_store, allow_unaligned_load, bzero_first, |
d7a9512e | 3734 | &split_stores, &total_orig, &total_new); |
a62b3dc5 | 3735 | |
5384a802 JJ |
3736 | /* Determine if there is a clobber covering the whole group at the start, |
3737 | followed by proposed split stores that cover the whole group. In that | |
3738 | case, prefer the transformation even if | |
3739 | split_stores.length () == orig_num_stmts. */ | |
3740 | bool clobber_first = false; | |
3741 | if (num_clobber_stmts | |
3742 | && gimple_clobber_p (group->stores[0]->stmt) | |
3743 | && group->start == group->stores[0]->bitpos | |
3744 | && group->width == group->stores[0]->bitsize | |
3745 | && (group->start % BITS_PER_UNIT) == 0 | |
3746 | && (group->width % BITS_PER_UNIT) == 0) | |
3747 | { | |
3748 | clobber_first = true; | |
3749 | unsigned HOST_WIDE_INT pos = group->start / BITS_PER_UNIT; | |
3750 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
3751 | if (split_store->bytepos != pos) | |
3752 | { | |
3753 | clobber_first = false; | |
3754 | break; | |
3755 | } | |
3756 | else | |
3757 | pos += split_store->size / BITS_PER_UNIT; | |
3758 | if (pos != (group->start + group->width) / BITS_PER_UNIT) | |
3759 | clobber_first = false; | |
3760 | } | |
3761 | ||
3762 | if (split_stores.length () >= orig_num_stmts + clobber_first) | |
a62b3dc5 | 3763 | { |
5384a802 | 3764 | |
a62b3dc5 JJ |
3765 | /* We didn't manage to reduce the number of statements. Bail out. */ |
3766 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
d7a9512e JJ |
3767 | fprintf (dump_file, "Exceeded original number of stmts (%u)." |
3768 | " Not profitable to emit new sequence.\n", | |
3769 | orig_num_stmts); | |
dd172744 RB |
3770 | FOR_EACH_VEC_ELT (split_stores, i, split_store) |
3771 | delete split_store; | |
a62b3dc5 JJ |
3772 | return false; |
3773 | } | |
d7a9512e JJ |
3774 | if (total_orig <= total_new) |
3775 | { | |
3776 | /* If number of estimated new statements is above estimated original | |
3777 | statements, bail out too. */ | |
3778 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3779 | fprintf (dump_file, "Estimated number of original stmts (%u)" | |
3780 | " not larger than estimated number of new" | |
3781 | " stmts (%u).\n", | |
3782 | total_orig, total_new); | |
dd172744 RB |
3783 | FOR_EACH_VEC_ELT (split_stores, i, split_store) |
3784 | delete split_store; | |
4b84d9b8 | 3785 | return false; |
d7a9512e | 3786 | } |
5384a802 JJ |
3787 | if (group->stores[0]->rhs_code == INTEGER_CST) |
3788 | { | |
3789 | bool all_orig = true; | |
3790 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
3791 | if (!split_store->orig) | |
3792 | { | |
3793 | all_orig = false; | |
3794 | break; | |
3795 | } | |
3796 | if (all_orig) | |
3797 | { | |
3798 | unsigned int cnt = split_stores.length (); | |
3799 | store_immediate_info *store; | |
3800 | FOR_EACH_VEC_ELT (group->stores, i, store) | |
3801 | if (gimple_clobber_p (store->stmt)) | |
3802 | ++cnt; | |
3803 | /* Punt if we wouldn't make any real changes, i.e. keep all | |
3804 | orig stmts + all clobbers. */ | |
3805 | if (cnt == group->stores.length ()) | |
3806 | { | |
3807 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3808 | fprintf (dump_file, "Exceeded original number of stmts (%u)." | |
3809 | " Not profitable to emit new sequence.\n", | |
3810 | orig_num_stmts); | |
3811 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
3812 | delete split_store; | |
3813 | return false; | |
3814 | } | |
3815 | } | |
3816 | } | |
f663d9ad KT |
3817 | |
3818 | gimple_stmt_iterator last_gsi = gsi_for_stmt (group->last_stmt); | |
3819 | gimple_seq seq = NULL; | |
f663d9ad KT |
3820 | tree last_vdef, new_vuse; |
3821 | last_vdef = gimple_vdef (group->last_stmt); | |
3822 | new_vuse = gimple_vuse (group->last_stmt); | |
4b84d9b8 JJ |
3823 | tree bswap_res = NULL_TREE; |
3824 | ||
5384a802 JJ |
3825 | /* Clobbers are not removed. */ |
3826 | if (gimple_clobber_p (group->last_stmt)) | |
3827 | { | |
3828 | new_vuse = make_ssa_name (gimple_vop (cfun), group->last_stmt); | |
3829 | gimple_set_vdef (group->last_stmt, new_vuse); | |
3830 | } | |
3831 | ||
4b84d9b8 JJ |
3832 | if (group->stores[0]->rhs_code == LROTATE_EXPR |
3833 | || group->stores[0]->rhs_code == NOP_EXPR) | |
3834 | { | |
3835 | tree fndecl = NULL_TREE, bswap_type = NULL_TREE, load_type; | |
3836 | gimple *ins_stmt = group->stores[0]->ins_stmt; | |
3837 | struct symbolic_number *n = &group->stores[0]->n; | |
3838 | bool bswap = group->stores[0]->rhs_code == LROTATE_EXPR; | |
3839 | ||
3840 | switch (n->range) | |
3841 | { | |
3842 | case 16: | |
3843 | load_type = bswap_type = uint16_type_node; | |
3844 | break; | |
3845 | case 32: | |
3846 | load_type = uint32_type_node; | |
3847 | if (bswap) | |
3848 | { | |
3849 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP32); | |
3850 | bswap_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
3851 | } | |
3852 | break; | |
3853 | case 64: | |
3854 | load_type = uint64_type_node; | |
3855 | if (bswap) | |
3856 | { | |
3857 | fndecl = builtin_decl_explicit (BUILT_IN_BSWAP64); | |
3858 | bswap_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl))); | |
3859 | } | |
3860 | break; | |
3861 | default: | |
3862 | gcc_unreachable (); | |
3863 | } | |
3864 | ||
3865 | /* If the loads have each vuse of the corresponding store, | |
3866 | we've checked the aliasing already in try_coalesce_bswap and | |
3867 | we want to sink the need load into seq. So need to use new_vuse | |
3868 | on the load. */ | |
30fa8e9c | 3869 | if (n->base_addr) |
4b84d9b8 | 3870 | { |
30fa8e9c JJ |
3871 | if (n->vuse == NULL) |
3872 | { | |
3873 | n->vuse = new_vuse; | |
3874 | ins_stmt = NULL; | |
3875 | } | |
3876 | else | |
3877 | /* Update vuse in case it has changed by output_merged_stores. */ | |
3878 | n->vuse = gimple_vuse (ins_stmt); | |
4b84d9b8 JJ |
3879 | } |
3880 | bswap_res = bswap_replace (gsi_start (seq), ins_stmt, fndecl, | |
3881 | bswap_type, load_type, n, bswap); | |
3882 | gcc_assert (bswap_res); | |
3883 | } | |
f663d9ad KT |
3884 | |
3885 | gimple *stmt = NULL; | |
245f6de1 | 3886 | auto_vec<gimple *, 32> orig_stmts; |
4b84d9b8 JJ |
3887 | gimple_seq this_seq; |
3888 | tree addr = force_gimple_operand_1 (unshare_expr (base_addr), &this_seq, | |
aa55dc0c | 3889 | is_gimple_mem_ref_addr, NULL_TREE); |
4b84d9b8 | 3890 | gimple_seq_add_seq_without_update (&seq, this_seq); |
245f6de1 JJ |
3891 | |
3892 | tree load_addr[2] = { NULL_TREE, NULL_TREE }; | |
3893 | gimple_seq load_seq[2] = { NULL, NULL }; | |
3894 | gimple_stmt_iterator load_gsi[2] = { gsi_none (), gsi_none () }; | |
3895 | for (int j = 0; j < 2; ++j) | |
3896 | { | |
3897 | store_operand_info &op = group->stores[0]->ops[j]; | |
3898 | if (op.base_addr == NULL_TREE) | |
3899 | continue; | |
3900 | ||
3901 | store_immediate_info *infol = group->stores.last (); | |
3902 | if (gimple_vuse (op.stmt) == gimple_vuse (infol->ops[j].stmt)) | |
3903 | { | |
97031af7 JJ |
3904 | /* We can't pick the location randomly; while we've verified |
3905 | all the loads have the same vuse, they can be still in different | |
3906 | basic blocks and we need to pick the one from the last bb: | |
3907 | int x = q[0]; | |
3908 | if (x == N) return; | |
3909 | int y = q[1]; | |
3910 | p[0] = x; | |
3911 | p[1] = y; | |
3912 | otherwise if we put the wider load at the q[0] load, we might | |
3913 | segfault if q[1] is not mapped. */ | |
3914 | basic_block bb = gimple_bb (op.stmt); | |
3915 | gimple *ostmt = op.stmt; | |
3916 | store_immediate_info *info; | |
3917 | FOR_EACH_VEC_ELT (group->stores, i, info) | |
3918 | { | |
3919 | gimple *tstmt = info->ops[j].stmt; | |
3920 | basic_block tbb = gimple_bb (tstmt); | |
3921 | if (dominated_by_p (CDI_DOMINATORS, tbb, bb)) | |
3922 | { | |
3923 | ostmt = tstmt; | |
3924 | bb = tbb; | |
3925 | } | |
3926 | } | |
3927 | load_gsi[j] = gsi_for_stmt (ostmt); | |
245f6de1 JJ |
3928 | load_addr[j] |
3929 | = force_gimple_operand_1 (unshare_expr (op.base_addr), | |
3930 | &load_seq[j], is_gimple_mem_ref_addr, | |
3931 | NULL_TREE); | |
3932 | } | |
3933 | else if (operand_equal_p (base_addr, op.base_addr, 0)) | |
3934 | load_addr[j] = addr; | |
3935 | else | |
3e2927a1 | 3936 | { |
3e2927a1 JJ |
3937 | load_addr[j] |
3938 | = force_gimple_operand_1 (unshare_expr (op.base_addr), | |
3939 | &this_seq, is_gimple_mem_ref_addr, | |
3940 | NULL_TREE); | |
3941 | gimple_seq_add_seq_without_update (&seq, this_seq); | |
3942 | } | |
245f6de1 JJ |
3943 | } |
3944 | ||
f663d9ad KT |
3945 | FOR_EACH_VEC_ELT (split_stores, i, split_store) |
3946 | { | |
3947 | unsigned HOST_WIDE_INT try_size = split_store->size; | |
3948 | unsigned HOST_WIDE_INT try_pos = split_store->bytepos; | |
c94c3532 | 3949 | unsigned HOST_WIDE_INT try_bitpos = try_pos * BITS_PER_UNIT; |
f663d9ad | 3950 | unsigned HOST_WIDE_INT align = split_store->align; |
a62b3dc5 JJ |
3951 | tree dest, src; |
3952 | location_t loc; | |
3953 | if (split_store->orig) | |
3954 | { | |
5384a802 JJ |
3955 | /* If there is just a single non-clobber constituent store |
3956 | which covers the whole area, just reuse the lhs and rhs. */ | |
3957 | gimple *orig_stmt = NULL; | |
3958 | store_immediate_info *store; | |
3959 | unsigned int j; | |
3960 | FOR_EACH_VEC_ELT (split_store->orig_stores, j, store) | |
3961 | if (!gimple_clobber_p (store->stmt)) | |
3962 | { | |
3963 | orig_stmt = store->stmt; | |
3964 | break; | |
3965 | } | |
245f6de1 JJ |
3966 | dest = gimple_assign_lhs (orig_stmt); |
3967 | src = gimple_assign_rhs1 (orig_stmt); | |
3968 | loc = gimple_location (orig_stmt); | |
a62b3dc5 JJ |
3969 | } |
3970 | else | |
3971 | { | |
245f6de1 JJ |
3972 | store_immediate_info *info; |
3973 | unsigned short clique, base; | |
3974 | unsigned int k; | |
3975 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
3976 | orig_stmts.safe_push (info->stmt); | |
a62b3dc5 | 3977 | tree offset_type |
245f6de1 JJ |
3978 | = get_alias_type_for_stmts (orig_stmts, false, &clique, &base); |
3979 | loc = get_location_for_stmts (orig_stmts); | |
3980 | orig_stmts.truncate (0); | |
a62b3dc5 JJ |
3981 | |
3982 | tree int_type = build_nonstandard_integer_type (try_size, UNSIGNED); | |
3983 | int_type = build_aligned_type (int_type, align); | |
3984 | dest = fold_build2 (MEM_REF, int_type, addr, | |
3985 | build_int_cst (offset_type, try_pos)); | |
245f6de1 JJ |
3986 | if (TREE_CODE (dest) == MEM_REF) |
3987 | { | |
3988 | MR_DEPENDENCE_CLIQUE (dest) = clique; | |
3989 | MR_DEPENDENCE_BASE (dest) = base; | |
3990 | } | |
3991 | ||
c94c3532 EB |
3992 | tree mask; |
3993 | if (bswap_res) | |
3994 | mask = integer_zero_node; | |
3995 | else | |
4b84d9b8 JJ |
3996 | mask = native_interpret_expr (int_type, |
3997 | group->mask + try_pos | |
3998 | - start_byte_pos, | |
3999 | group->buf_size); | |
245f6de1 JJ |
4000 | |
4001 | tree ops[2]; | |
4002 | for (int j = 0; | |
4003 | j < 1 + (split_store->orig_stores[0]->ops[1].val != NULL_TREE); | |
4004 | ++j) | |
4005 | { | |
4006 | store_operand_info &op = split_store->orig_stores[0]->ops[j]; | |
4b84d9b8 JJ |
4007 | if (bswap_res) |
4008 | ops[j] = bswap_res; | |
4009 | else if (op.base_addr) | |
245f6de1 JJ |
4010 | { |
4011 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4012 | orig_stmts.safe_push (info->ops[j].stmt); | |
4013 | ||
4014 | offset_type = get_alias_type_for_stmts (orig_stmts, true, | |
4015 | &clique, &base); | |
4016 | location_t load_loc = get_location_for_stmts (orig_stmts); | |
4017 | orig_stmts.truncate (0); | |
4018 | ||
4019 | unsigned HOST_WIDE_INT load_align = group->load_align[j]; | |
4020 | unsigned HOST_WIDE_INT align_bitpos | |
c94c3532 | 4021 | = known_alignment (try_bitpos |
8a91d545 RS |
4022 | - split_store->orig_stores[0]->bitpos |
4023 | + op.bitpos); | |
4024 | if (align_bitpos & (load_align - 1)) | |
245f6de1 JJ |
4025 | load_align = least_bit_hwi (align_bitpos); |
4026 | ||
4027 | tree load_int_type | |
4028 | = build_nonstandard_integer_type (try_size, UNSIGNED); | |
4029 | load_int_type | |
4030 | = build_aligned_type (load_int_type, load_align); | |
4031 | ||
8a91d545 | 4032 | poly_uint64 load_pos |
c94c3532 | 4033 | = exact_div (try_bitpos |
8a91d545 RS |
4034 | - split_store->orig_stores[0]->bitpos |
4035 | + op.bitpos, | |
4036 | BITS_PER_UNIT); | |
245f6de1 JJ |
4037 | ops[j] = fold_build2 (MEM_REF, load_int_type, load_addr[j], |
4038 | build_int_cst (offset_type, load_pos)); | |
4039 | if (TREE_CODE (ops[j]) == MEM_REF) | |
4040 | { | |
4041 | MR_DEPENDENCE_CLIQUE (ops[j]) = clique; | |
4042 | MR_DEPENDENCE_BASE (ops[j]) = base; | |
4043 | } | |
4044 | if (!integer_zerop (mask)) | |
4045 | /* The load might load some bits (that will be masked off | |
4046 | later on) uninitialized, avoid -W*uninitialized | |
4047 | warnings in that case. */ | |
4048 | TREE_NO_WARNING (ops[j]) = 1; | |
4049 | ||
4050 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
4051 | ops[j]); | |
4052 | gimple_set_location (stmt, load_loc); | |
4053 | if (gsi_bb (load_gsi[j])) | |
4054 | { | |
4055 | gimple_set_vuse (stmt, gimple_vuse (op.stmt)); | |
4056 | gimple_seq_add_stmt_without_update (&load_seq[j], stmt); | |
4057 | } | |
4058 | else | |
4059 | { | |
4060 | gimple_set_vuse (stmt, new_vuse); | |
4061 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4062 | } | |
4063 | ops[j] = gimple_assign_lhs (stmt); | |
a6fbd154 JJ |
4064 | tree xor_mask; |
4065 | enum tree_code inv_op | |
4066 | = invert_op (split_store, j, int_type, xor_mask); | |
4067 | if (inv_op != NOP_EXPR) | |
383ac8dc JJ |
4068 | { |
4069 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
a6fbd154 | 4070 | inv_op, ops[j], xor_mask); |
383ac8dc JJ |
4071 | gimple_set_location (stmt, load_loc); |
4072 | ops[j] = gimple_assign_lhs (stmt); | |
4073 | ||
4074 | if (gsi_bb (load_gsi[j])) | |
4075 | gimple_seq_add_stmt_without_update (&load_seq[j], | |
4076 | stmt); | |
4077 | else | |
4078 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4079 | } | |
245f6de1 JJ |
4080 | } |
4081 | else | |
4082 | ops[j] = native_interpret_expr (int_type, | |
4083 | group->val + try_pos | |
4084 | - start_byte_pos, | |
4085 | group->buf_size); | |
4086 | } | |
4087 | ||
4088 | switch (split_store->orig_stores[0]->rhs_code) | |
4089 | { | |
4090 | case BIT_AND_EXPR: | |
4091 | case BIT_IOR_EXPR: | |
4092 | case BIT_XOR_EXPR: | |
4093 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4094 | { | |
4095 | tree rhs1 = gimple_assign_rhs1 (info->stmt); | |
4096 | orig_stmts.safe_push (SSA_NAME_DEF_STMT (rhs1)); | |
4097 | } | |
4098 | location_t bit_loc; | |
4099 | bit_loc = get_location_for_stmts (orig_stmts); | |
4100 | orig_stmts.truncate (0); | |
4101 | ||
4102 | stmt | |
4103 | = gimple_build_assign (make_ssa_name (int_type), | |
4104 | split_store->orig_stores[0]->rhs_code, | |
4105 | ops[0], ops[1]); | |
4106 | gimple_set_location (stmt, bit_loc); | |
4107 | /* If there is just one load and there is a separate | |
4108 | load_seq[0], emit the bitwise op right after it. */ | |
4109 | if (load_addr[1] == NULL_TREE && gsi_bb (load_gsi[0])) | |
4110 | gimple_seq_add_stmt_without_update (&load_seq[0], stmt); | |
4111 | /* Otherwise, if at least one load is in seq, we need to | |
4112 | emit the bitwise op right before the store. If there | |
4113 | are two loads and are emitted somewhere else, it would | |
4114 | be better to emit the bitwise op as early as possible; | |
4115 | we don't track where that would be possible right now | |
4116 | though. */ | |
4117 | else | |
4118 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4119 | src = gimple_assign_lhs (stmt); | |
a6fbd154 JJ |
4120 | tree xor_mask; |
4121 | enum tree_code inv_op; | |
4122 | inv_op = invert_op (split_store, 2, int_type, xor_mask); | |
4123 | if (inv_op != NOP_EXPR) | |
d60edaba JJ |
4124 | { |
4125 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
a6fbd154 | 4126 | inv_op, src, xor_mask); |
d60edaba JJ |
4127 | gimple_set_location (stmt, bit_loc); |
4128 | if (load_addr[1] == NULL_TREE && gsi_bb (load_gsi[0])) | |
4129 | gimple_seq_add_stmt_without_update (&load_seq[0], stmt); | |
4130 | else | |
4131 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4132 | src = gimple_assign_lhs (stmt); | |
4133 | } | |
245f6de1 | 4134 | break; |
4b84d9b8 JJ |
4135 | case LROTATE_EXPR: |
4136 | case NOP_EXPR: | |
4137 | src = ops[0]; | |
4138 | if (!is_gimple_val (src)) | |
4139 | { | |
4140 | stmt = gimple_build_assign (make_ssa_name (TREE_TYPE (src)), | |
4141 | src); | |
4142 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4143 | src = gimple_assign_lhs (stmt); | |
4144 | } | |
4145 | if (!useless_type_conversion_p (int_type, TREE_TYPE (src))) | |
4146 | { | |
4147 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
4148 | NOP_EXPR, src); | |
4149 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4150 | src = gimple_assign_lhs (stmt); | |
4151 | } | |
be52ac73 JJ |
4152 | inv_op = invert_op (split_store, 2, int_type, xor_mask); |
4153 | if (inv_op != NOP_EXPR) | |
4154 | { | |
4155 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
4156 | inv_op, src, xor_mask); | |
4157 | gimple_set_location (stmt, loc); | |
4158 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4159 | src = gimple_assign_lhs (stmt); | |
4160 | } | |
4b84d9b8 | 4161 | break; |
245f6de1 JJ |
4162 | default: |
4163 | src = ops[0]; | |
4164 | break; | |
4165 | } | |
4166 | ||
c94c3532 EB |
4167 | /* If bit insertion is required, we use the source as an accumulator |
4168 | into which the successive bit-field values are manually inserted. | |
4169 | FIXME: perhaps use BIT_INSERT_EXPR instead in some cases? */ | |
4170 | if (group->bit_insertion) | |
4171 | FOR_EACH_VEC_ELT (split_store->orig_stores, k, info) | |
4172 | if (info->rhs_code == BIT_INSERT_EXPR | |
4173 | && info->bitpos < try_bitpos + try_size | |
4174 | && info->bitpos + info->bitsize > try_bitpos) | |
4175 | { | |
4176 | /* Mask, truncate, convert to final type, shift and ior into | |
4177 | the accumulator. Note that every step can be a no-op. */ | |
4178 | const HOST_WIDE_INT start_gap = info->bitpos - try_bitpos; | |
4179 | const HOST_WIDE_INT end_gap | |
4180 | = (try_bitpos + try_size) - (info->bitpos + info->bitsize); | |
4181 | tree tem = info->ops[0].val; | |
ed01d707 EB |
4182 | if (!INTEGRAL_TYPE_P (TREE_TYPE (tem))) |
4183 | { | |
4184 | const unsigned HOST_WIDE_INT size | |
4185 | = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (tem))); | |
4186 | tree integer_type | |
4187 | = build_nonstandard_integer_type (size, UNSIGNED); | |
4188 | tem = gimple_build (&seq, loc, VIEW_CONVERT_EXPR, | |
4189 | integer_type, tem); | |
4190 | } | |
c14add82 EB |
4191 | if (TYPE_PRECISION (TREE_TYPE (tem)) <= info->bitsize) |
4192 | { | |
4193 | tree bitfield_type | |
4194 | = build_nonstandard_integer_type (info->bitsize, | |
4195 | UNSIGNED); | |
4196 | tem = gimple_convert (&seq, loc, bitfield_type, tem); | |
4197 | } | |
4198 | else if ((BYTES_BIG_ENDIAN ? start_gap : end_gap) > 0) | |
c94c3532 EB |
4199 | { |
4200 | const unsigned HOST_WIDE_INT imask | |
4201 | = (HOST_WIDE_INT_1U << info->bitsize) - 1; | |
4202 | tem = gimple_build (&seq, loc, | |
4203 | BIT_AND_EXPR, TREE_TYPE (tem), tem, | |
4204 | build_int_cst (TREE_TYPE (tem), | |
4205 | imask)); | |
4206 | } | |
4207 | const HOST_WIDE_INT shift | |
4208 | = (BYTES_BIG_ENDIAN ? end_gap : start_gap); | |
4209 | if (shift < 0) | |
4210 | tem = gimple_build (&seq, loc, | |
4211 | RSHIFT_EXPR, TREE_TYPE (tem), tem, | |
4212 | build_int_cst (NULL_TREE, -shift)); | |
4213 | tem = gimple_convert (&seq, loc, int_type, tem); | |
4214 | if (shift > 0) | |
4215 | tem = gimple_build (&seq, loc, | |
4216 | LSHIFT_EXPR, int_type, tem, | |
4217 | build_int_cst (NULL_TREE, shift)); | |
4218 | src = gimple_build (&seq, loc, | |
4219 | BIT_IOR_EXPR, int_type, tem, src); | |
4220 | } | |
4221 | ||
a62b3dc5 JJ |
4222 | if (!integer_zerop (mask)) |
4223 | { | |
4224 | tree tem = make_ssa_name (int_type); | |
4225 | tree load_src = unshare_expr (dest); | |
4226 | /* The load might load some or all bits uninitialized, | |
4227 | avoid -W*uninitialized warnings in that case. | |
4228 | As optimization, it would be nice if all the bits are | |
4229 | provably uninitialized (no stores at all yet or previous | |
4230 | store a CLOBBER) we'd optimize away the load and replace | |
4231 | it e.g. with 0. */ | |
4232 | TREE_NO_WARNING (load_src) = 1; | |
4233 | stmt = gimple_build_assign (tem, load_src); | |
4234 | gimple_set_location (stmt, loc); | |
4235 | gimple_set_vuse (stmt, new_vuse); | |
4236 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4237 | ||
4238 | /* FIXME: If there is a single chunk of zero bits in mask, | |
4239 | perhaps use BIT_INSERT_EXPR instead? */ | |
4240 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
4241 | BIT_AND_EXPR, tem, mask); | |
4242 | gimple_set_location (stmt, loc); | |
4243 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4244 | tem = gimple_assign_lhs (stmt); | |
4245 | ||
245f6de1 JJ |
4246 | if (TREE_CODE (src) == INTEGER_CST) |
4247 | src = wide_int_to_tree (int_type, | |
4248 | wi::bit_and_not (wi::to_wide (src), | |
4249 | wi::to_wide (mask))); | |
4250 | else | |
4251 | { | |
4252 | tree nmask | |
4253 | = wide_int_to_tree (int_type, | |
4254 | wi::bit_not (wi::to_wide (mask))); | |
4255 | stmt = gimple_build_assign (make_ssa_name (int_type), | |
4256 | BIT_AND_EXPR, src, nmask); | |
4257 | gimple_set_location (stmt, loc); | |
4258 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4259 | src = gimple_assign_lhs (stmt); | |
4260 | } | |
a62b3dc5 JJ |
4261 | stmt = gimple_build_assign (make_ssa_name (int_type), |
4262 | BIT_IOR_EXPR, tem, src); | |
4263 | gimple_set_location (stmt, loc); | |
4264 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4265 | src = gimple_assign_lhs (stmt); | |
4266 | } | |
4267 | } | |
f663d9ad KT |
4268 | |
4269 | stmt = gimple_build_assign (dest, src); | |
4270 | gimple_set_location (stmt, loc); | |
4271 | gimple_set_vuse (stmt, new_vuse); | |
4272 | gimple_seq_add_stmt_without_update (&seq, stmt); | |
4273 | ||
629387a6 EB |
4274 | if (group->lp_nr && stmt_could_throw_p (cfun, stmt)) |
4275 | add_stmt_to_eh_lp (stmt, group->lp_nr); | |
4276 | ||
f663d9ad KT |
4277 | tree new_vdef; |
4278 | if (i < split_stores.length () - 1) | |
a62b3dc5 | 4279 | new_vdef = make_ssa_name (gimple_vop (cfun), stmt); |
f663d9ad KT |
4280 | else |
4281 | new_vdef = last_vdef; | |
4282 | ||
4283 | gimple_set_vdef (stmt, new_vdef); | |
4284 | SSA_NAME_DEF_STMT (new_vdef) = stmt; | |
4285 | new_vuse = new_vdef; | |
4286 | } | |
4287 | ||
4288 | FOR_EACH_VEC_ELT (split_stores, i, split_store) | |
4289 | delete split_store; | |
4290 | ||
f663d9ad KT |
4291 | gcc_assert (seq); |
4292 | if (dump_file) | |
4293 | { | |
4294 | fprintf (dump_file, | |
c94c3532 | 4295 | "New sequence of %u stores to replace old one of %u stores\n", |
a62b3dc5 | 4296 | split_stores.length (), orig_num_stmts); |
f663d9ad KT |
4297 | if (dump_flags & TDF_DETAILS) |
4298 | print_gimple_seq (dump_file, seq, 0, TDF_VOPS | TDF_MEMSYMS); | |
4299 | } | |
629387a6 | 4300 | |
5384a802 JJ |
4301 | if (gimple_clobber_p (group->last_stmt)) |
4302 | update_stmt (group->last_stmt); | |
4303 | ||
629387a6 EB |
4304 | if (group->lp_nr > 0) |
4305 | { | |
4306 | /* We're going to insert a sequence of (potentially) throwing stores | |
4307 | into an active EH region. This means that we're going to create | |
4308 | new basic blocks with EH edges pointing to the post landing pad | |
4309 | and, therefore, to have to update its PHI nodes, if any. For the | |
4310 | virtual PHI node, we're going to use the VDEFs created above, but | |
4311 | for the other nodes, we need to record the original reaching defs. */ | |
4312 | eh_landing_pad lp = get_eh_landing_pad_from_number (group->lp_nr); | |
4313 | basic_block lp_bb = label_to_block (cfun, lp->post_landing_pad); | |
4314 | basic_block last_bb = gimple_bb (group->last_stmt); | |
4315 | edge last_edge = find_edge (last_bb, lp_bb); | |
4316 | auto_vec<tree, 16> last_defs; | |
4317 | gphi_iterator gpi; | |
4318 | for (gpi = gsi_start_phis (lp_bb); !gsi_end_p (gpi); gsi_next (&gpi)) | |
4319 | { | |
4320 | gphi *phi = gpi.phi (); | |
4321 | tree last_def; | |
4322 | if (virtual_operand_p (gimple_phi_result (phi))) | |
4323 | last_def = NULL_TREE; | |
4324 | else | |
4325 | last_def = gimple_phi_arg_def (phi, last_edge->dest_idx); | |
4326 | last_defs.safe_push (last_def); | |
4327 | } | |
4328 | ||
4329 | /* Do the insertion. Then, if new basic blocks have been created in the | |
4330 | process, rewind the chain of VDEFs create above to walk the new basic | |
4331 | blocks and update the corresponding arguments of the PHI nodes. */ | |
4332 | update_modified_stmts (seq); | |
4333 | if (gimple_find_sub_bbs (seq, &last_gsi)) | |
4334 | while (last_vdef != gimple_vuse (group->last_stmt)) | |
4335 | { | |
4336 | gimple *stmt = SSA_NAME_DEF_STMT (last_vdef); | |
4337 | if (stmt_could_throw_p (cfun, stmt)) | |
4338 | { | |
4339 | edge new_edge = find_edge (gimple_bb (stmt), lp_bb); | |
4340 | unsigned int i; | |
4341 | for (gpi = gsi_start_phis (lp_bb), i = 0; | |
4342 | !gsi_end_p (gpi); | |
4343 | gsi_next (&gpi), i++) | |
4344 | { | |
4345 | gphi *phi = gpi.phi (); | |
4346 | tree new_def; | |
4347 | if (virtual_operand_p (gimple_phi_result (phi))) | |
4348 | new_def = last_vdef; | |
4349 | else | |
4350 | new_def = last_defs[i]; | |
4351 | add_phi_arg (phi, new_def, new_edge, UNKNOWN_LOCATION); | |
4352 | } | |
4353 | } | |
4354 | last_vdef = gimple_vuse (stmt); | |
4355 | } | |
4356 | } | |
4357 | else | |
4358 | gsi_insert_seq_after (&last_gsi, seq, GSI_SAME_STMT); | |
4359 | ||
245f6de1 JJ |
4360 | for (int j = 0; j < 2; ++j) |
4361 | if (load_seq[j]) | |
4362 | gsi_insert_seq_after (&load_gsi[j], load_seq[j], GSI_SAME_STMT); | |
f663d9ad KT |
4363 | |
4364 | return true; | |
4365 | } | |
4366 | ||
4367 | /* Process the merged_store_group objects created in the coalescing phase. | |
4368 | The stores are all against the base object BASE. | |
4369 | Try to output the widened stores and delete the original statements if | |
4370 | successful. Return true iff any changes were made. */ | |
4371 | ||
4372 | bool | |
b5926e23 | 4373 | imm_store_chain_info::output_merged_stores () |
f663d9ad KT |
4374 | { |
4375 | unsigned int i; | |
4376 | merged_store_group *merged_store; | |
4377 | bool ret = false; | |
4378 | FOR_EACH_VEC_ELT (m_merged_store_groups, i, merged_store) | |
4379 | { | |
a95b474a ML |
4380 | if (dbg_cnt (store_merging) |
4381 | && output_merged_store (merged_store)) | |
f663d9ad KT |
4382 | { |
4383 | unsigned int j; | |
4384 | store_immediate_info *store; | |
4385 | FOR_EACH_VEC_ELT (merged_store->stores, j, store) | |
4386 | { | |
4387 | gimple *stmt = store->stmt; | |
4388 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
5384a802 JJ |
4389 | /* Don't remove clobbers, they are still useful even if |
4390 | everything is overwritten afterwards. */ | |
4391 | if (gimple_clobber_p (stmt)) | |
4392 | continue; | |
f663d9ad | 4393 | gsi_remove (&gsi, true); |
629387a6 EB |
4394 | if (store->lp_nr) |
4395 | remove_stmt_from_eh_lp (stmt); | |
f663d9ad KT |
4396 | if (stmt != merged_store->last_stmt) |
4397 | { | |
4398 | unlink_stmt_vdef (stmt); | |
4399 | release_defs (stmt); | |
4400 | } | |
4401 | } | |
4402 | ret = true; | |
4403 | } | |
4404 | } | |
4405 | if (ret && dump_file) | |
4406 | fprintf (dump_file, "Merging successful!\n"); | |
4407 | ||
4408 | return ret; | |
4409 | } | |
4410 | ||
4411 | /* Coalesce the store_immediate_info objects recorded against the base object | |
4412 | BASE in the first phase and output them. | |
4413 | Delete the allocated structures. | |
4414 | Return true if any changes were made. */ | |
4415 | ||
4416 | bool | |
b5926e23 | 4417 | imm_store_chain_info::terminate_and_process_chain () |
f663d9ad KT |
4418 | { |
4419 | /* Process store chain. */ | |
4420 | bool ret = false; | |
4421 | if (m_store_info.length () > 1) | |
4422 | { | |
4423 | ret = coalesce_immediate_stores (); | |
4424 | if (ret) | |
b5926e23 | 4425 | ret = output_merged_stores (); |
f663d9ad KT |
4426 | } |
4427 | ||
4428 | /* Delete all the entries we allocated ourselves. */ | |
4429 | store_immediate_info *info; | |
4430 | unsigned int i; | |
4431 | FOR_EACH_VEC_ELT (m_store_info, i, info) | |
4432 | delete info; | |
4433 | ||
4434 | merged_store_group *merged_info; | |
4435 | FOR_EACH_VEC_ELT (m_merged_store_groups, i, merged_info) | |
4436 | delete merged_info; | |
4437 | ||
4438 | return ret; | |
4439 | } | |
4440 | ||
4441 | /* Return true iff LHS is a destination potentially interesting for | |
4442 | store merging. In practice these are the codes that get_inner_reference | |
4443 | can process. */ | |
4444 | ||
4445 | static bool | |
4446 | lhs_valid_for_store_merging_p (tree lhs) | |
4447 | { | |
629387a6 | 4448 | if (DECL_P (lhs)) |
f663d9ad KT |
4449 | return true; |
4450 | ||
629387a6 EB |
4451 | switch (TREE_CODE (lhs)) |
4452 | { | |
4453 | case ARRAY_REF: | |
4454 | case ARRAY_RANGE_REF: | |
4455 | case BIT_FIELD_REF: | |
4456 | case COMPONENT_REF: | |
4457 | case MEM_REF: | |
4458 | return true; | |
4459 | default: | |
4460 | return false; | |
4461 | } | |
4462 | ||
4463 | gcc_unreachable (); | |
f663d9ad KT |
4464 | } |
4465 | ||
4466 | /* Return true if the tree RHS is a constant we want to consider | |
4467 | during store merging. In practice accept all codes that | |
4468 | native_encode_expr accepts. */ | |
4469 | ||
4470 | static bool | |
4471 | rhs_valid_for_store_merging_p (tree rhs) | |
4472 | { | |
cf098191 | 4473 | unsigned HOST_WIDE_INT size; |
3afd514b | 4474 | if (TREE_CODE (rhs) == CONSTRUCTOR |
3afd514b JJ |
4475 | && CONSTRUCTOR_NELTS (rhs) == 0 |
4476 | && TYPE_SIZE_UNIT (TREE_TYPE (rhs)) | |
4477 | && tree_fits_uhwi_p (TYPE_SIZE_UNIT (TREE_TYPE (rhs)))) | |
4478 | return true; | |
cf098191 RS |
4479 | return (GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (rhs))).is_constant (&size) |
4480 | && native_encode_expr (rhs, NULL, size) != 0); | |
f663d9ad KT |
4481 | } |
4482 | ||
629387a6 EB |
4483 | /* Adjust *PBITPOS, *PBITREGION_START and *PBITREGION_END by BYTE_OFF bytes |
4484 | and return true on success or false on failure. */ | |
4485 | ||
4486 | static bool | |
4487 | adjust_bit_pos (poly_offset_int byte_off, | |
4488 | poly_int64 *pbitpos, | |
4489 | poly_uint64 *pbitregion_start, | |
4490 | poly_uint64 *pbitregion_end) | |
4491 | { | |
4492 | poly_offset_int bit_off = byte_off << LOG2_BITS_PER_UNIT; | |
4493 | bit_off += *pbitpos; | |
4494 | ||
4495 | if (known_ge (bit_off, 0) && bit_off.to_shwi (pbitpos)) | |
4496 | { | |
4497 | if (maybe_ne (*pbitregion_end, 0U)) | |
4498 | { | |
4499 | bit_off = byte_off << LOG2_BITS_PER_UNIT; | |
4500 | bit_off += *pbitregion_start; | |
4501 | if (bit_off.to_uhwi (pbitregion_start)) | |
4502 | { | |
4503 | bit_off = byte_off << LOG2_BITS_PER_UNIT; | |
4504 | bit_off += *pbitregion_end; | |
4505 | if (!bit_off.to_uhwi (pbitregion_end)) | |
4506 | *pbitregion_end = 0; | |
4507 | } | |
4508 | else | |
4509 | *pbitregion_end = 0; | |
4510 | } | |
4511 | return true; | |
4512 | } | |
4513 | else | |
4514 | return false; | |
4515 | } | |
4516 | ||
245f6de1 JJ |
4517 | /* If MEM is a memory reference usable for store merging (either as |
4518 | store destination or for loads), return the non-NULL base_addr | |
4519 | and set *PBITSIZE, *PBITPOS, *PBITREGION_START and *PBITREGION_END. | |
4520 | Otherwise return NULL, *PBITPOS should be still valid even for that | |
4521 | case. */ | |
4522 | ||
4523 | static tree | |
8a91d545 RS |
4524 | mem_valid_for_store_merging (tree mem, poly_uint64 *pbitsize, |
4525 | poly_uint64 *pbitpos, | |
4526 | poly_uint64 *pbitregion_start, | |
4527 | poly_uint64 *pbitregion_end) | |
245f6de1 | 4528 | { |
8a91d545 RS |
4529 | poly_int64 bitsize, bitpos; |
4530 | poly_uint64 bitregion_start = 0, bitregion_end = 0; | |
245f6de1 JJ |
4531 | machine_mode mode; |
4532 | int unsignedp = 0, reversep = 0, volatilep = 0; | |
4533 | tree offset; | |
4534 | tree base_addr = get_inner_reference (mem, &bitsize, &bitpos, &offset, &mode, | |
4535 | &unsignedp, &reversep, &volatilep); | |
4536 | *pbitsize = bitsize; | |
8a91d545 | 4537 | if (known_eq (bitsize, 0)) |
245f6de1 JJ |
4538 | return NULL_TREE; |
4539 | ||
4540 | if (TREE_CODE (mem) == COMPONENT_REF | |
4541 | && DECL_BIT_FIELD_TYPE (TREE_OPERAND (mem, 1))) | |
4542 | { | |
4543 | get_bit_range (&bitregion_start, &bitregion_end, mem, &bitpos, &offset); | |
8a91d545 RS |
4544 | if (maybe_ne (bitregion_end, 0U)) |
4545 | bitregion_end += 1; | |
245f6de1 JJ |
4546 | } |
4547 | ||
4548 | if (reversep) | |
4549 | return NULL_TREE; | |
4550 | ||
4551 | /* We do not want to rewrite TARGET_MEM_REFs. */ | |
4552 | if (TREE_CODE (base_addr) == TARGET_MEM_REF) | |
4553 | return NULL_TREE; | |
4554 | /* In some cases get_inner_reference may return a | |
4555 | MEM_REF [ptr + byteoffset]. For the purposes of this pass | |
4556 | canonicalize the base_addr to MEM_REF [ptr] and take | |
4557 | byteoffset into account in the bitpos. This occurs in | |
4558 | PR 23684 and this way we can catch more chains. */ | |
4559 | else if (TREE_CODE (base_addr) == MEM_REF) | |
4560 | { | |
629387a6 EB |
4561 | if (!adjust_bit_pos (mem_ref_offset (base_addr), &bitpos, |
4562 | &bitregion_start, &bitregion_end)) | |
245f6de1 JJ |
4563 | return NULL_TREE; |
4564 | base_addr = TREE_OPERAND (base_addr, 0); | |
4565 | } | |
4566 | /* get_inner_reference returns the base object, get at its | |
4567 | address now. */ | |
4568 | else | |
4569 | { | |
8a91d545 | 4570 | if (maybe_lt (bitpos, 0)) |
245f6de1 JJ |
4571 | return NULL_TREE; |
4572 | base_addr = build_fold_addr_expr (base_addr); | |
4573 | } | |
4574 | ||
629387a6 | 4575 | if (offset) |
245f6de1 JJ |
4576 | { |
4577 | /* If the access is variable offset then a base decl has to be | |
4578 | address-taken to be able to emit pointer-based stores to it. | |
4579 | ??? We might be able to get away with re-using the original | |
4580 | base up to the first variable part and then wrapping that inside | |
4581 | a BIT_FIELD_REF. */ | |
4582 | tree base = get_base_address (base_addr); | |
629387a6 | 4583 | if (!base || (DECL_P (base) && !TREE_ADDRESSABLE (base))) |
245f6de1 JJ |
4584 | return NULL_TREE; |
4585 | ||
629387a6 EB |
4586 | /* Similarly to above for the base, remove constant from the offset. */ |
4587 | if (TREE_CODE (offset) == PLUS_EXPR | |
4588 | && TREE_CODE (TREE_OPERAND (offset, 1)) == INTEGER_CST | |
4589 | && adjust_bit_pos (wi::to_poly_offset (TREE_OPERAND (offset, 1)), | |
4590 | &bitpos, &bitregion_start, &bitregion_end)) | |
4591 | offset = TREE_OPERAND (offset, 0); | |
4592 | ||
245f6de1 JJ |
4593 | base_addr = build2 (POINTER_PLUS_EXPR, TREE_TYPE (base_addr), |
4594 | base_addr, offset); | |
4595 | } | |
4596 | ||
629387a6 EB |
4597 | if (known_eq (bitregion_end, 0U)) |
4598 | { | |
4599 | bitregion_start = round_down_to_byte_boundary (bitpos); | |
4600 | bitregion_end = round_up_to_byte_boundary (bitpos + bitsize); | |
4601 | } | |
4602 | ||
245f6de1 JJ |
4603 | *pbitsize = bitsize; |
4604 | *pbitpos = bitpos; | |
4605 | *pbitregion_start = bitregion_start; | |
4606 | *pbitregion_end = bitregion_end; | |
4607 | return base_addr; | |
4608 | } | |
4609 | ||
4610 | /* Return true if STMT is a load that can be used for store merging. | |
4611 | In that case fill in *OP. BITSIZE, BITPOS, BITREGION_START and | |
4612 | BITREGION_END are properties of the corresponding store. */ | |
4613 | ||
4614 | static bool | |
4615 | handled_load (gimple *stmt, store_operand_info *op, | |
8a91d545 RS |
4616 | poly_uint64 bitsize, poly_uint64 bitpos, |
4617 | poly_uint64 bitregion_start, poly_uint64 bitregion_end) | |
245f6de1 | 4618 | { |
383ac8dc | 4619 | if (!is_gimple_assign (stmt)) |
245f6de1 | 4620 | return false; |
383ac8dc JJ |
4621 | if (gimple_assign_rhs_code (stmt) == BIT_NOT_EXPR) |
4622 | { | |
4623 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
4624 | if (TREE_CODE (rhs1) == SSA_NAME | |
383ac8dc JJ |
4625 | && handled_load (SSA_NAME_DEF_STMT (rhs1), op, bitsize, bitpos, |
4626 | bitregion_start, bitregion_end)) | |
4627 | { | |
d60edaba JJ |
4628 | /* Don't allow _1 = load; _2 = ~1; _3 = ~_2; which should have |
4629 | been optimized earlier, but if allowed here, would confuse the | |
4630 | multiple uses counting. */ | |
4631 | if (op->bit_not_p) | |
4632 | return false; | |
383ac8dc JJ |
4633 | op->bit_not_p = !op->bit_not_p; |
4634 | return true; | |
4635 | } | |
4636 | return false; | |
4637 | } | |
4638 | if (gimple_vuse (stmt) | |
4639 | && gimple_assign_load_p (stmt) | |
36bbc05d | 4640 | && !stmt_can_throw_internal (cfun, stmt) |
245f6de1 JJ |
4641 | && !gimple_has_volatile_ops (stmt)) |
4642 | { | |
4643 | tree mem = gimple_assign_rhs1 (stmt); | |
4644 | op->base_addr | |
4645 | = mem_valid_for_store_merging (mem, &op->bitsize, &op->bitpos, | |
4646 | &op->bitregion_start, | |
4647 | &op->bitregion_end); | |
4648 | if (op->base_addr != NULL_TREE | |
8a91d545 RS |
4649 | && known_eq (op->bitsize, bitsize) |
4650 | && multiple_p (op->bitpos - bitpos, BITS_PER_UNIT) | |
4651 | && known_ge (op->bitpos - op->bitregion_start, | |
4652 | bitpos - bitregion_start) | |
4653 | && known_ge (op->bitregion_end - op->bitpos, | |
4654 | bitregion_end - bitpos)) | |
245f6de1 JJ |
4655 | { |
4656 | op->stmt = stmt; | |
4657 | op->val = mem; | |
383ac8dc | 4658 | op->bit_not_p = false; |
245f6de1 JJ |
4659 | return true; |
4660 | } | |
4661 | } | |
4662 | return false; | |
4663 | } | |
4664 | ||
629387a6 EB |
4665 | /* Return the index number of the landing pad for STMT, if any. */ |
4666 | ||
4667 | static int | |
4668 | lp_nr_for_store (gimple *stmt) | |
4669 | { | |
4670 | if (!cfun->can_throw_non_call_exceptions || !cfun->eh) | |
4671 | return 0; | |
4672 | ||
4673 | if (!stmt_could_throw_p (cfun, stmt)) | |
4674 | return 0; | |
4675 | ||
4676 | return lookup_stmt_eh_lp (stmt); | |
4677 | } | |
4678 | ||
245f6de1 | 4679 | /* Record the store STMT for store merging optimization if it can be |
629387a6 | 4680 | optimized. Return true if any changes were made. */ |
245f6de1 | 4681 | |
629387a6 | 4682 | bool |
245f6de1 JJ |
4683 | pass_store_merging::process_store (gimple *stmt) |
4684 | { | |
4685 | tree lhs = gimple_assign_lhs (stmt); | |
4686 | tree rhs = gimple_assign_rhs1 (stmt); | |
2c832ffe SSF |
4687 | poly_uint64 bitsize, bitpos = 0; |
4688 | poly_uint64 bitregion_start = 0, bitregion_end = 0; | |
245f6de1 JJ |
4689 | tree base_addr |
4690 | = mem_valid_for_store_merging (lhs, &bitsize, &bitpos, | |
4691 | &bitregion_start, &bitregion_end); | |
8a91d545 | 4692 | if (known_eq (bitsize, 0U)) |
629387a6 | 4693 | return false; |
245f6de1 JJ |
4694 | |
4695 | bool invalid = (base_addr == NULL_TREE | |
8a91d545 RS |
4696 | || (maybe_gt (bitsize, |
4697 | (unsigned int) MAX_BITSIZE_MODE_ANY_INT) | |
3afd514b JJ |
4698 | && TREE_CODE (rhs) != INTEGER_CST |
4699 | && (TREE_CODE (rhs) != CONSTRUCTOR | |
4700 | || CONSTRUCTOR_NELTS (rhs) != 0))); | |
245f6de1 | 4701 | enum tree_code rhs_code = ERROR_MARK; |
d60edaba | 4702 | bool bit_not_p = false; |
4b84d9b8 JJ |
4703 | struct symbolic_number n; |
4704 | gimple *ins_stmt = NULL; | |
245f6de1 JJ |
4705 | store_operand_info ops[2]; |
4706 | if (invalid) | |
4707 | ; | |
4708 | else if (rhs_valid_for_store_merging_p (rhs)) | |
4709 | { | |
4710 | rhs_code = INTEGER_CST; | |
4711 | ops[0].val = rhs; | |
4712 | } | |
d7a9512e | 4713 | else if (TREE_CODE (rhs) != SSA_NAME) |
245f6de1 JJ |
4714 | invalid = true; |
4715 | else | |
4716 | { | |
4717 | gimple *def_stmt = SSA_NAME_DEF_STMT (rhs), *def_stmt1, *def_stmt2; | |
4718 | if (!is_gimple_assign (def_stmt)) | |
4719 | invalid = true; | |
4720 | else if (handled_load (def_stmt, &ops[0], bitsize, bitpos, | |
4721 | bitregion_start, bitregion_end)) | |
4722 | rhs_code = MEM_REF; | |
d60edaba JJ |
4723 | else if (gimple_assign_rhs_code (def_stmt) == BIT_NOT_EXPR) |
4724 | { | |
4725 | tree rhs1 = gimple_assign_rhs1 (def_stmt); | |
4726 | if (TREE_CODE (rhs1) == SSA_NAME | |
4727 | && is_gimple_assign (SSA_NAME_DEF_STMT (rhs1))) | |
4728 | { | |
4729 | bit_not_p = true; | |
4730 | def_stmt = SSA_NAME_DEF_STMT (rhs1); | |
4731 | } | |
4732 | } | |
c94c3532 | 4733 | |
d60edaba | 4734 | if (rhs_code == ERROR_MARK && !invalid) |
245f6de1 JJ |
4735 | switch ((rhs_code = gimple_assign_rhs_code (def_stmt))) |
4736 | { | |
4737 | case BIT_AND_EXPR: | |
4738 | case BIT_IOR_EXPR: | |
4739 | case BIT_XOR_EXPR: | |
4740 | tree rhs1, rhs2; | |
4741 | rhs1 = gimple_assign_rhs1 (def_stmt); | |
4742 | rhs2 = gimple_assign_rhs2 (def_stmt); | |
4743 | invalid = true; | |
d7a9512e | 4744 | if (TREE_CODE (rhs1) != SSA_NAME) |
245f6de1 JJ |
4745 | break; |
4746 | def_stmt1 = SSA_NAME_DEF_STMT (rhs1); | |
4747 | if (!is_gimple_assign (def_stmt1) | |
4748 | || !handled_load (def_stmt1, &ops[0], bitsize, bitpos, | |
4749 | bitregion_start, bitregion_end)) | |
4750 | break; | |
4751 | if (rhs_valid_for_store_merging_p (rhs2)) | |
4752 | ops[1].val = rhs2; | |
d7a9512e | 4753 | else if (TREE_CODE (rhs2) != SSA_NAME) |
245f6de1 JJ |
4754 | break; |
4755 | else | |
4756 | { | |
4757 | def_stmt2 = SSA_NAME_DEF_STMT (rhs2); | |
4758 | if (!is_gimple_assign (def_stmt2)) | |
4759 | break; | |
4760 | else if (!handled_load (def_stmt2, &ops[1], bitsize, bitpos, | |
4761 | bitregion_start, bitregion_end)) | |
4762 | break; | |
4763 | } | |
4764 | invalid = false; | |
4765 | break; | |
4766 | default: | |
4767 | invalid = true; | |
4768 | break; | |
4769 | } | |
c94c3532 | 4770 | |
8a91d545 RS |
4771 | unsigned HOST_WIDE_INT const_bitsize; |
4772 | if (bitsize.is_constant (&const_bitsize) | |
c94c3532 | 4773 | && (const_bitsize % BITS_PER_UNIT) == 0 |
8a91d545 | 4774 | && const_bitsize <= 64 |
c94c3532 | 4775 | && multiple_p (bitpos, BITS_PER_UNIT)) |
4b84d9b8 JJ |
4776 | { |
4777 | ins_stmt = find_bswap_or_nop_1 (def_stmt, &n, 12); | |
4778 | if (ins_stmt) | |
4779 | { | |
4780 | uint64_t nn = n.n; | |
4781 | for (unsigned HOST_WIDE_INT i = 0; | |
8a91d545 RS |
4782 | i < const_bitsize; |
4783 | i += BITS_PER_UNIT, nn >>= BITS_PER_MARKER) | |
4b84d9b8 JJ |
4784 | if ((nn & MARKER_MASK) == 0 |
4785 | || (nn & MARKER_MASK) == MARKER_BYTE_UNKNOWN) | |
4786 | { | |
4787 | ins_stmt = NULL; | |
4788 | break; | |
4789 | } | |
4790 | if (ins_stmt) | |
4791 | { | |
4792 | if (invalid) | |
4793 | { | |
4794 | rhs_code = LROTATE_EXPR; | |
4795 | ops[0].base_addr = NULL_TREE; | |
4796 | ops[1].base_addr = NULL_TREE; | |
4797 | } | |
4798 | invalid = false; | |
4799 | } | |
4800 | } | |
4801 | } | |
c94c3532 EB |
4802 | |
4803 | if (invalid | |
4804 | && bitsize.is_constant (&const_bitsize) | |
4805 | && ((const_bitsize % BITS_PER_UNIT) != 0 | |
4806 | || !multiple_p (bitpos, BITS_PER_UNIT)) | |
ed01d707 | 4807 | && const_bitsize <= MAX_FIXED_MODE_SIZE) |
c94c3532 | 4808 | { |
c14add82 | 4809 | /* Bypass a conversion to the bit-field type. */ |
31a5d8c5 EB |
4810 | if (!bit_not_p |
4811 | && is_gimple_assign (def_stmt) | |
4812 | && CONVERT_EXPR_CODE_P (rhs_code)) | |
c94c3532 EB |
4813 | { |
4814 | tree rhs1 = gimple_assign_rhs1 (def_stmt); | |
4815 | if (TREE_CODE (rhs1) == SSA_NAME | |
c14add82 | 4816 | && INTEGRAL_TYPE_P (TREE_TYPE (rhs1))) |
c94c3532 EB |
4817 | rhs = rhs1; |
4818 | } | |
4819 | rhs_code = BIT_INSERT_EXPR; | |
31a5d8c5 | 4820 | bit_not_p = false; |
c94c3532 EB |
4821 | ops[0].val = rhs; |
4822 | ops[0].base_addr = NULL_TREE; | |
4823 | ops[1].base_addr = NULL_TREE; | |
4824 | invalid = false; | |
4825 | } | |
245f6de1 JJ |
4826 | } |
4827 | ||
8a91d545 RS |
4828 | unsigned HOST_WIDE_INT const_bitsize, const_bitpos; |
4829 | unsigned HOST_WIDE_INT const_bitregion_start, const_bitregion_end; | |
4830 | if (invalid | |
4831 | || !bitsize.is_constant (&const_bitsize) | |
4832 | || !bitpos.is_constant (&const_bitpos) | |
4833 | || !bitregion_start.is_constant (&const_bitregion_start) | |
4834 | || !bitregion_end.is_constant (&const_bitregion_end)) | |
629387a6 | 4835 | return terminate_all_aliasing_chains (NULL, stmt); |
245f6de1 | 4836 | |
4b84d9b8 JJ |
4837 | if (!ins_stmt) |
4838 | memset (&n, 0, sizeof (n)); | |
4839 | ||
99b1c316 | 4840 | class imm_store_chain_info **chain_info = NULL; |
629387a6 | 4841 | bool ret = false; |
383ac8dc JJ |
4842 | if (base_addr) |
4843 | chain_info = m_stores.get (base_addr); | |
4844 | ||
245f6de1 JJ |
4845 | store_immediate_info *info; |
4846 | if (chain_info) | |
4847 | { | |
4848 | unsigned int ord = (*chain_info)->m_store_info.length (); | |
8a91d545 RS |
4849 | info = new store_immediate_info (const_bitsize, const_bitpos, |
4850 | const_bitregion_start, | |
4851 | const_bitregion_end, | |
4852 | stmt, ord, rhs_code, n, ins_stmt, | |
629387a6 EB |
4853 | bit_not_p, lp_nr_for_store (stmt), |
4854 | ops[0], ops[1]); | |
245f6de1 JJ |
4855 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4856 | { | |
4857 | fprintf (dump_file, "Recording immediate store from stmt:\n"); | |
4858 | print_gimple_stmt (dump_file, stmt, 0); | |
4859 | } | |
4860 | (*chain_info)->m_store_info.safe_push (info); | |
629387a6 | 4861 | ret |= terminate_all_aliasing_chains (chain_info, stmt); |
245f6de1 JJ |
4862 | /* If we reach the limit of stores to merge in a chain terminate and |
4863 | process the chain now. */ | |
4864 | if ((*chain_info)->m_store_info.length () | |
028d4092 | 4865 | == (unsigned int) param_max_stores_to_merge) |
245f6de1 JJ |
4866 | { |
4867 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4868 | fprintf (dump_file, | |
4869 | "Reached maximum number of statements to merge:\n"); | |
629387a6 | 4870 | ret |= terminate_and_process_chain (*chain_info); |
245f6de1 | 4871 | } |
629387a6 | 4872 | return ret; |
245f6de1 JJ |
4873 | } |
4874 | ||
4875 | /* Store aliases any existing chain? */ | |
629387a6 | 4876 | ret |= terminate_all_aliasing_chains (NULL, stmt); |
245f6de1 | 4877 | /* Start a new chain. */ |
99b1c316 | 4878 | class imm_store_chain_info *new_chain |
245f6de1 | 4879 | = new imm_store_chain_info (m_stores_head, base_addr); |
8a91d545 RS |
4880 | info = new store_immediate_info (const_bitsize, const_bitpos, |
4881 | const_bitregion_start, | |
4882 | const_bitregion_end, | |
4883 | stmt, 0, rhs_code, n, ins_stmt, | |
629387a6 EB |
4884 | bit_not_p, lp_nr_for_store (stmt), |
4885 | ops[0], ops[1]); | |
245f6de1 JJ |
4886 | new_chain->m_store_info.safe_push (info); |
4887 | m_stores.put (base_addr, new_chain); | |
4888 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4889 | { | |
4890 | fprintf (dump_file, "Starting new chain with statement:\n"); | |
4891 | print_gimple_stmt (dump_file, stmt, 0); | |
4892 | fprintf (dump_file, "The base object is:\n"); | |
4893 | print_generic_expr (dump_file, base_addr); | |
4894 | fprintf (dump_file, "\n"); | |
4895 | } | |
629387a6 EB |
4896 | return ret; |
4897 | } | |
4898 | ||
4899 | /* Return true if STMT is a store valid for store merging. */ | |
4900 | ||
4901 | static bool | |
4902 | store_valid_for_store_merging_p (gimple *stmt) | |
4903 | { | |
4904 | return gimple_assign_single_p (stmt) | |
4905 | && gimple_vdef (stmt) | |
4906 | && lhs_valid_for_store_merging_p (gimple_assign_lhs (stmt)) | |
5384a802 | 4907 | && (!gimple_has_volatile_ops (stmt) || gimple_clobber_p (stmt)); |
629387a6 EB |
4908 | } |
4909 | ||
4910 | enum basic_block_status { BB_INVALID, BB_VALID, BB_EXTENDED_VALID }; | |
4911 | ||
4912 | /* Return the status of basic block BB wrt store merging. */ | |
4913 | ||
4914 | static enum basic_block_status | |
4915 | get_status_for_store_merging (basic_block bb) | |
4916 | { | |
4917 | unsigned int num_statements = 0; | |
4918 | gimple_stmt_iterator gsi; | |
4919 | edge e; | |
4920 | ||
4921 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4922 | { | |
4923 | gimple *stmt = gsi_stmt (gsi); | |
4924 | ||
4925 | if (is_gimple_debug (stmt)) | |
4926 | continue; | |
4927 | ||
4928 | if (store_valid_for_store_merging_p (stmt) && ++num_statements >= 2) | |
4929 | break; | |
4930 | } | |
4931 | ||
4932 | if (num_statements == 0) | |
4933 | return BB_INVALID; | |
4934 | ||
4935 | if (cfun->can_throw_non_call_exceptions && cfun->eh | |
4936 | && store_valid_for_store_merging_p (gimple_seq_last_stmt (bb_seq (bb))) | |
4937 | && (e = find_fallthru_edge (bb->succs)) | |
4938 | && e->dest == bb->next_bb) | |
4939 | return BB_EXTENDED_VALID; | |
4940 | ||
4941 | return num_statements >= 2 ? BB_VALID : BB_INVALID; | |
245f6de1 JJ |
4942 | } |
4943 | ||
f663d9ad | 4944 | /* Entry point for the pass. Go over each basic block recording chains of |
245f6de1 JJ |
4945 | immediate stores. Upon encountering a terminating statement (as defined |
4946 | by stmt_terminates_chain_p) process the recorded stores and emit the widened | |
4947 | variants. */ | |
f663d9ad KT |
4948 | |
4949 | unsigned int | |
4950 | pass_store_merging::execute (function *fun) | |
4951 | { | |
4952 | basic_block bb; | |
4953 | hash_set<gimple *> orig_stmts; | |
629387a6 EB |
4954 | bool changed = false, open_chains = false; |
4955 | ||
4956 | /* If the function can throw and catch non-call exceptions, we'll be trying | |
4957 | to merge stores across different basic blocks so we need to first unsplit | |
4958 | the EH edges in order to streamline the CFG of the function. */ | |
4959 | if (cfun->can_throw_non_call_exceptions && cfun->eh) | |
4960 | unsplit_eh_edges (); | |
f663d9ad | 4961 | |
4b84d9b8 JJ |
4962 | calculate_dominance_info (CDI_DOMINATORS); |
4963 | ||
f663d9ad KT |
4964 | FOR_EACH_BB_FN (bb, fun) |
4965 | { | |
629387a6 | 4966 | const basic_block_status bb_status = get_status_for_store_merging (bb); |
f663d9ad | 4967 | gimple_stmt_iterator gsi; |
f663d9ad | 4968 | |
629387a6 EB |
4969 | if (open_chains && (bb_status == BB_INVALID || !single_pred_p (bb))) |
4970 | { | |
4971 | changed |= terminate_and_process_all_chains (); | |
4972 | open_chains = false; | |
f663d9ad KT |
4973 | } |
4974 | ||
629387a6 | 4975 | if (bb_status == BB_INVALID) |
f663d9ad KT |
4976 | continue; |
4977 | ||
4978 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4979 | fprintf (dump_file, "Processing basic block <%d>:\n", bb->index); | |
4980 | ||
4981 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4982 | { | |
4983 | gimple *stmt = gsi_stmt (gsi); | |
4984 | ||
50b6d676 AO |
4985 | if (is_gimple_debug (stmt)) |
4986 | continue; | |
4987 | ||
5384a802 | 4988 | if (gimple_has_volatile_ops (stmt) && !gimple_clobber_p (stmt)) |
f663d9ad KT |
4989 | { |
4990 | /* Terminate all chains. */ | |
4991 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4992 | fprintf (dump_file, "Volatile access terminates " | |
4993 | "all chains\n"); | |
629387a6 EB |
4994 | changed |= terminate_and_process_all_chains (); |
4995 | open_chains = false; | |
f663d9ad KT |
4996 | continue; |
4997 | } | |
4998 | ||
629387a6 EB |
4999 | if (store_valid_for_store_merging_p (stmt)) |
5000 | changed |= process_store (stmt); | |
245f6de1 | 5001 | else |
629387a6 EB |
5002 | changed |= terminate_all_aliasing_chains (NULL, stmt); |
5003 | } | |
5004 | ||
5005 | if (bb_status == BB_EXTENDED_VALID) | |
5006 | open_chains = true; | |
5007 | else | |
5008 | { | |
5009 | changed |= terminate_and_process_all_chains (); | |
5010 | open_chains = false; | |
f663d9ad | 5011 | } |
f663d9ad | 5012 | } |
629387a6 EB |
5013 | |
5014 | if (open_chains) | |
5015 | changed |= terminate_and_process_all_chains (); | |
5016 | ||
5017 | /* If the function can throw and catch non-call exceptions and something | |
5018 | changed during the pass, then the CFG has (very likely) changed too. */ | |
5019 | if (cfun->can_throw_non_call_exceptions && cfun->eh && changed) | |
5020 | { | |
5021 | free_dominance_info (CDI_DOMINATORS); | |
5022 | return TODO_cleanup_cfg; | |
5023 | } | |
5024 | ||
f663d9ad KT |
5025 | return 0; |
5026 | } | |
5027 | ||
5028 | } // anon namespace | |
5029 | ||
5030 | /* Construct and return a store merging pass object. */ | |
5031 | ||
5032 | gimple_opt_pass * | |
5033 | make_pass_store_merging (gcc::context *ctxt) | |
5034 | { | |
5035 | return new pass_store_merging (ctxt); | |
5036 | } | |
c22d8787 KT |
5037 | |
5038 | #if CHECKING_P | |
5039 | ||
5040 | namespace selftest { | |
5041 | ||
5042 | /* Selftests for store merging helpers. */ | |
5043 | ||
5044 | /* Assert that all elements of the byte arrays X and Y, both of length N | |
5045 | are equal. */ | |
5046 | ||
5047 | static void | |
5048 | verify_array_eq (unsigned char *x, unsigned char *y, unsigned int n) | |
5049 | { | |
5050 | for (unsigned int i = 0; i < n; i++) | |
5051 | { | |
5052 | if (x[i] != y[i]) | |
5053 | { | |
5054 | fprintf (stderr, "Arrays do not match. X:\n"); | |
5055 | dump_char_array (stderr, x, n); | |
5056 | fprintf (stderr, "Y:\n"); | |
5057 | dump_char_array (stderr, y, n); | |
5058 | } | |
5059 | ASSERT_EQ (x[i], y[i]); | |
5060 | } | |
5061 | } | |
5062 | ||
8aba425f | 5063 | /* Test shift_bytes_in_array_left and that it carries bits across between |
c22d8787 KT |
5064 | bytes correctly. */ |
5065 | ||
5066 | static void | |
8aba425f | 5067 | verify_shift_bytes_in_array_left (void) |
c22d8787 KT |
5068 | { |
5069 | /* byte 1 | byte 0 | |
5070 | 00011111 | 11100000. */ | |
5071 | unsigned char orig[2] = { 0xe0, 0x1f }; | |
5072 | unsigned char in[2]; | |
5073 | memcpy (in, orig, sizeof orig); | |
5074 | ||
5075 | unsigned char expected[2] = { 0x80, 0x7f }; | |
8aba425f | 5076 | shift_bytes_in_array_left (in, sizeof (in), 2); |
c22d8787 KT |
5077 | verify_array_eq (in, expected, sizeof (in)); |
5078 | ||
5079 | memcpy (in, orig, sizeof orig); | |
5080 | memcpy (expected, orig, sizeof orig); | |
5081 | /* Check that shifting by zero doesn't change anything. */ | |
8aba425f | 5082 | shift_bytes_in_array_left (in, sizeof (in), 0); |
c22d8787 KT |
5083 | verify_array_eq (in, expected, sizeof (in)); |
5084 | ||
5085 | } | |
5086 | ||
5087 | /* Test shift_bytes_in_array_right and that it carries bits across between | |
5088 | bytes correctly. */ | |
5089 | ||
5090 | static void | |
5091 | verify_shift_bytes_in_array_right (void) | |
5092 | { | |
5093 | /* byte 1 | byte 0 | |
5094 | 00011111 | 11100000. */ | |
5095 | unsigned char orig[2] = { 0x1f, 0xe0}; | |
5096 | unsigned char in[2]; | |
5097 | memcpy (in, orig, sizeof orig); | |
5098 | unsigned char expected[2] = { 0x07, 0xf8}; | |
5099 | shift_bytes_in_array_right (in, sizeof (in), 2); | |
5100 | verify_array_eq (in, expected, sizeof (in)); | |
5101 | ||
5102 | memcpy (in, orig, sizeof orig); | |
5103 | memcpy (expected, orig, sizeof orig); | |
5104 | /* Check that shifting by zero doesn't change anything. */ | |
5105 | shift_bytes_in_array_right (in, sizeof (in), 0); | |
5106 | verify_array_eq (in, expected, sizeof (in)); | |
5107 | } | |
5108 | ||
5109 | /* Test clear_bit_region that it clears exactly the bits asked and | |
5110 | nothing more. */ | |
5111 | ||
5112 | static void | |
5113 | verify_clear_bit_region (void) | |
5114 | { | |
5115 | /* Start with all bits set and test clearing various patterns in them. */ | |
5116 | unsigned char orig[3] = { 0xff, 0xff, 0xff}; | |
5117 | unsigned char in[3]; | |
5118 | unsigned char expected[3]; | |
5119 | memcpy (in, orig, sizeof in); | |
5120 | ||
5121 | /* Check zeroing out all the bits. */ | |
5122 | clear_bit_region (in, 0, 3 * BITS_PER_UNIT); | |
5123 | expected[0] = expected[1] = expected[2] = 0; | |
5124 | verify_array_eq (in, expected, sizeof in); | |
5125 | ||
5126 | memcpy (in, orig, sizeof in); | |
5127 | /* Leave the first and last bits intact. */ | |
5128 | clear_bit_region (in, 1, 3 * BITS_PER_UNIT - 2); | |
5129 | expected[0] = 0x1; | |
5130 | expected[1] = 0; | |
5131 | expected[2] = 0x80; | |
5132 | verify_array_eq (in, expected, sizeof in); | |
5133 | } | |
5134 | ||
5384a802 | 5135 | /* Test clear_bit_region_be that it clears exactly the bits asked and |
c22d8787 KT |
5136 | nothing more. */ |
5137 | ||
5138 | static void | |
5139 | verify_clear_bit_region_be (void) | |
5140 | { | |
5141 | /* Start with all bits set and test clearing various patterns in them. */ | |
5142 | unsigned char orig[3] = { 0xff, 0xff, 0xff}; | |
5143 | unsigned char in[3]; | |
5144 | unsigned char expected[3]; | |
5145 | memcpy (in, orig, sizeof in); | |
5146 | ||
5147 | /* Check zeroing out all the bits. */ | |
5148 | clear_bit_region_be (in, BITS_PER_UNIT - 1, 3 * BITS_PER_UNIT); | |
5149 | expected[0] = expected[1] = expected[2] = 0; | |
5150 | verify_array_eq (in, expected, sizeof in); | |
5151 | ||
5152 | memcpy (in, orig, sizeof in); | |
5153 | /* Leave the first and last bits intact. */ | |
5154 | clear_bit_region_be (in, BITS_PER_UNIT - 2, 3 * BITS_PER_UNIT - 2); | |
5155 | expected[0] = 0x80; | |
5156 | expected[1] = 0; | |
5157 | expected[2] = 0x1; | |
5158 | verify_array_eq (in, expected, sizeof in); | |
5159 | } | |
5160 | ||
5161 | ||
5162 | /* Run all of the selftests within this file. */ | |
5163 | ||
5164 | void | |
5165 | store_merging_c_tests (void) | |
5166 | { | |
8aba425f | 5167 | verify_shift_bytes_in_array_left (); |
c22d8787 KT |
5168 | verify_shift_bytes_in_array_right (); |
5169 | verify_clear_bit_region (); | |
5170 | verify_clear_bit_region_be (); | |
5171 | } | |
5172 | ||
5173 | } // namespace selftest | |
5174 | #endif /* CHECKING_P. */ |