/src/php-src/Zend/Optimizer/zend_cfg.c
Line | Count | Source |
1 | | /* |
2 | | +----------------------------------------------------------------------+ |
3 | | | Zend Engine, CFG - Control Flow Graph | |
4 | | +----------------------------------------------------------------------+ |
5 | | | Copyright © The PHP Group and Contributors. | |
6 | | +----------------------------------------------------------------------+ |
7 | | | This source file is subject to the Modified BSD License that is | |
8 | | | bundled with this package in the file LICENSE, and is available | |
9 | | | through the World Wide Web at <https://www.php.net/license/>. | |
10 | | | | |
11 | | | SPDX-License-Identifier: BSD-3-Clause | |
12 | | +----------------------------------------------------------------------+ |
13 | | | Authors: Dmitry Stogov <dmitry@php.net> | |
14 | | +----------------------------------------------------------------------+ |
15 | | */ |
16 | | |
17 | | #include "zend_compile.h" |
18 | | #include "zend_cfg.h" |
19 | | #include "zend_func_info.h" |
20 | | #include "zend_worklist.h" |
21 | | #include "zend_optimizer.h" |
22 | | #include "zend_optimizer_internal.h" |
23 | | #include "zend_sort.h" |
24 | | |
25 | | static void zend_mark_reachable(zend_op *opcodes, zend_cfg *cfg, zend_basic_block *b) /* {{{ */ |
26 | 358k | { |
27 | 358k | zend_basic_block *blocks = cfg->blocks; |
28 | | |
29 | 358k | zend_worklist work; |
30 | 358k | ALLOCA_FLAG(list_use_heap) |
31 | 358k | ZEND_WORKLIST_ALLOCA(&work, cfg->blocks_count, list_use_heap); |
32 | | |
33 | 358k | zend_worklist_push(&work, b - cfg->blocks); |
34 | | |
35 | 1.43M | while (zend_worklist_len(&work)) { |
36 | 1.08M | b = cfg->blocks + zend_worklist_pop(&work); |
37 | | |
38 | 1.08M | b->flags |= ZEND_BB_REACHABLE; |
39 | 1.08M | if (b->successors_count == 0) { |
40 | 306k | b->flags |= ZEND_BB_EXIT; |
41 | 306k | continue; |
42 | 306k | } |
43 | | |
44 | 1.86M | for (uint32_t i = 0; i < b->successors_count; i++) { |
45 | 1.08M | zend_basic_block *succ = blocks + b->successors[i]; |
46 | | |
47 | 1.08M | if (b->len != 0) { |
48 | 1.08M | uint8_t opcode = opcodes[b->start + b->len - 1].opcode; |
49 | 1.08M | if (opcode == ZEND_MATCH) { |
50 | 2.88k | succ->flags |= ZEND_BB_TARGET; |
51 | 1.08M | } else if (opcode == ZEND_SWITCH_LONG || opcode == ZEND_SWITCH_STRING) { |
52 | 1.30k | if (i == b->successors_count - 1) { |
53 | 241 | succ->flags |= ZEND_BB_FOLLOW | ZEND_BB_TARGET; |
54 | 1.06k | } else { |
55 | 1.06k | succ->flags |= ZEND_BB_TARGET; |
56 | 1.06k | } |
57 | 1.08M | } else if (b->successors_count == 1) { |
58 | 462k | if (opcode == ZEND_JMP) { |
59 | 157k | succ->flags |= ZEND_BB_TARGET; |
60 | 304k | } else { |
61 | 304k | succ->flags |= ZEND_BB_FOLLOW; |
62 | | |
63 | 304k | if ((cfg->flags & ZEND_CFG_STACKLESS)) { |
64 | 0 | if (opcode == ZEND_INCLUDE_OR_EVAL || |
65 | 0 | opcode == ZEND_GENERATOR_CREATE || |
66 | 0 | opcode == ZEND_YIELD || |
67 | 0 | opcode == ZEND_YIELD_FROM || |
68 | 0 | opcode == ZEND_DO_FCALL || |
69 | 0 | opcode == ZEND_DO_UCALL || |
70 | 0 | opcode == ZEND_DO_FCALL_BY_NAME) { |
71 | 0 | succ->flags |= ZEND_BB_ENTRY; |
72 | 0 | } |
73 | 0 | } |
74 | 304k | if ((cfg->flags & ZEND_CFG_RECV_ENTRY)) { |
75 | 0 | if (opcode == ZEND_RECV || |
76 | 0 | opcode == ZEND_RECV_INIT) { |
77 | 0 | succ->flags |= ZEND_BB_RECV_ENTRY; |
78 | 0 | } |
79 | 0 | } |
80 | 304k | } |
81 | 621k | } else { |
82 | 621k | ZEND_ASSERT(b->successors_count == 2); |
83 | 621k | if (i == 0) { |
84 | 310k | succ->flags |= ZEND_BB_TARGET; |
85 | 310k | } else { |
86 | 310k | succ->flags |= ZEND_BB_FOLLOW; |
87 | 310k | } |
88 | 621k | } |
89 | 1.08M | } else { |
90 | 763 | succ->flags |= ZEND_BB_FOLLOW; |
91 | 763 | } |
92 | | |
93 | | /* Check reachability of successor */ |
94 | 1.08M | if (!(succ->flags & ZEND_BB_REACHABLE)) { |
95 | 899k | zend_worklist_push(&work, succ - cfg->blocks); |
96 | 899k | } |
97 | 1.08M | } |
98 | 774k | } |
99 | | |
100 | 358k | ZEND_WORKLIST_FREE_ALLOCA(&work, list_use_heap); |
101 | 358k | } |
102 | | /* }}} */ |
103 | | |
104 | | static void zend_mark_reachable_blocks(const zend_op_array *op_array, zend_cfg *cfg, uint32_t start) /* {{{ */ |
105 | 284k | { |
106 | 284k | zend_basic_block *blocks = cfg->blocks; |
107 | | |
108 | 284k | blocks[start].flags = ZEND_BB_START; |
109 | 284k | zend_mark_reachable(op_array->opcodes, cfg, blocks + start); |
110 | | |
111 | 284k | if (op_array->last_try_catch) { |
112 | 56.1k | zend_basic_block *b; |
113 | 56.1k | int changed; |
114 | 56.1k | uint32_t *block_map = cfg->map; |
115 | | |
116 | 111k | do { |
117 | 111k | changed = 0; |
118 | | |
119 | | /* Add exception paths */ |
120 | 261k | for (uint32_t j = 0; j < op_array->last_try_catch; j++) { |
121 | | |
122 | | /* check for jumps into the middle of try block */ |
123 | 149k | b = blocks + block_map[op_array->try_catch_array[j].try_op]; |
124 | 149k | if (!(b->flags & ZEND_BB_REACHABLE)) { |
125 | 64 | zend_basic_block *end; |
126 | | |
127 | 64 | if (op_array->try_catch_array[j].catch_op) { |
128 | 64 | end = blocks + block_map[op_array->try_catch_array[j].catch_op]; |
129 | 200 | while (b != end) { |
130 | 144 | if (b->flags & ZEND_BB_REACHABLE) { |
131 | 8 | op_array->try_catch_array[j].try_op = b->start; |
132 | 8 | break; |
133 | 8 | } |
134 | 136 | b++; |
135 | 136 | } |
136 | 64 | } |
137 | 64 | b = blocks + block_map[op_array->try_catch_array[j].try_op]; |
138 | 64 | if (!(b->flags & ZEND_BB_REACHABLE)) { |
139 | 56 | if (op_array->try_catch_array[j].finally_op) { |
140 | 12 | end = blocks + block_map[op_array->try_catch_array[j].finally_op]; |
141 | 44 | while (b != end) { |
142 | 44 | if (b->flags & ZEND_BB_REACHABLE) { |
143 | | /* In case we get here, there is no live try block but there is a live finally block. |
144 | | * If we do have catch_op set, we need to set it to the first catch block to satisfy |
145 | | * the constraint try_op <= catch_op <= finally_op */ |
146 | 12 | op_array->try_catch_array[j].try_op = |
147 | 12 | op_array->try_catch_array[j].catch_op ? op_array->try_catch_array[j].catch_op : b->start; |
148 | 12 | changed = 1; |
149 | 12 | zend_mark_reachable(op_array->opcodes, cfg, blocks + block_map[op_array->try_catch_array[j].try_op]); |
150 | 12 | break; |
151 | 12 | } |
152 | 32 | b++; |
153 | 32 | } |
154 | 12 | } |
155 | 56 | } |
156 | 64 | } |
157 | | |
158 | 149k | b = blocks + block_map[op_array->try_catch_array[j].try_op]; |
159 | 149k | if (b->flags & ZEND_BB_REACHABLE) { |
160 | 149k | b->flags |= ZEND_BB_TRY; |
161 | 149k | if (op_array->try_catch_array[j].catch_op) { |
162 | 147k | b = blocks + block_map[op_array->try_catch_array[j].catch_op]; |
163 | 147k | b->flags |= ZEND_BB_CATCH; |
164 | 147k | if (!(b->flags & ZEND_BB_REACHABLE)) { |
165 | 73.4k | changed = 1; |
166 | 73.4k | zend_mark_reachable(op_array->opcodes, cfg, b); |
167 | 73.4k | } |
168 | 147k | } |
169 | 149k | if (op_array->try_catch_array[j].finally_op) { |
170 | 3.18k | b = blocks + block_map[op_array->try_catch_array[j].finally_op]; |
171 | 3.18k | b->flags |= ZEND_BB_FINALLY; |
172 | 3.18k | if (!(b->flags & ZEND_BB_REACHABLE)) { |
173 | 372 | changed = 1; |
174 | 372 | zend_mark_reachable(op_array->opcodes, cfg, b); |
175 | 372 | } |
176 | 3.18k | } |
177 | 149k | if (op_array->try_catch_array[j].finally_end) { |
178 | 3.18k | b = blocks + block_map[op_array->try_catch_array[j].finally_end]; |
179 | 3.18k | b->flags |= ZEND_BB_FINALLY_END; |
180 | 3.18k | if (!(b->flags & ZEND_BB_REACHABLE)) { |
181 | 580 | changed = 1; |
182 | 580 | zend_mark_reachable(op_array->opcodes, cfg, b); |
183 | 580 | } |
184 | 3.18k | } |
185 | 149k | } else { |
186 | 44 | if (op_array->try_catch_array[j].catch_op) { |
187 | 44 | ZEND_ASSERT(!(blocks[block_map[op_array->try_catch_array[j].catch_op]].flags & ZEND_BB_REACHABLE)); |
188 | 44 | } |
189 | 44 | if (op_array->try_catch_array[j].finally_op) { |
190 | 0 | ZEND_ASSERT(!(blocks[block_map[op_array->try_catch_array[j].finally_op]].flags & ZEND_BB_REACHABLE)); |
191 | 0 | } |
192 | 44 | if (op_array->try_catch_array[j].finally_end) { |
193 | 0 | ZEND_ASSERT(!(blocks[block_map[op_array->try_catch_array[j].finally_end]].flags & ZEND_BB_REACHABLE)); |
194 | 0 | } |
195 | 44 | } |
196 | 149k | } |
197 | 111k | } while (changed); |
198 | 56.1k | } |
199 | | |
200 | 284k | if (cfg->flags & ZEND_FUNC_FREE_LOOP_VAR) { |
201 | 38.5k | zend_basic_block *b; |
202 | 38.5k | uint32_t *block_map = cfg->map; |
203 | | |
204 | | /* Mark blocks that are unreachable, but free a loop var created in a reachable block. */ |
205 | 406k | for (b = blocks; b < blocks + cfg->blocks_count; b++) { |
206 | 368k | if (b->flags & ZEND_BB_REACHABLE) { |
207 | 354k | continue; |
208 | 354k | } |
209 | | |
210 | 17.4k | for (uint32_t j = b->start; j < b->start + b->len; j++) { |
211 | 3.84k | zend_op *opline = &op_array->opcodes[j]; |
212 | 3.84k | if (zend_optimizer_is_loop_var_free(opline)) { |
213 | 110 | zend_op *def_opline = zend_optimizer_get_loop_var_def(op_array, opline); |
214 | 110 | if (def_opline) { |
215 | 110 | uint32_t def_block = block_map[def_opline - op_array->opcodes]; |
216 | 110 | if (blocks[def_block].flags & ZEND_BB_REACHABLE) { |
217 | 90 | b->flags |= ZEND_BB_UNREACHABLE_FREE; |
218 | 90 | break; |
219 | 90 | } |
220 | 110 | } |
221 | 110 | } |
222 | 3.84k | } |
223 | 13.6k | } |
224 | 38.5k | } |
225 | 284k | } |
226 | | /* }}} */ |
227 | | |
228 | | void zend_cfg_remark_reachable_blocks(const zend_op_array *op_array, zend_cfg *cfg) /* {{{ */ |
229 | 113k | { |
230 | 113k | zend_basic_block *blocks = cfg->blocks; |
231 | 113k | uint32_t i; |
232 | 113k | uint32_t start = 0; |
233 | | |
234 | 113k | for (i = 0; i < cfg->blocks_count; i++) { |
235 | 113k | if (blocks[i].flags & ZEND_BB_REACHABLE) { |
236 | 113k | start = i; |
237 | 113k | i++; |
238 | 113k | break; |
239 | 113k | } |
240 | 113k | } |
241 | | |
242 | | /* clear all flags */ |
243 | 668k | for (i = 0; i < cfg->blocks_count; i++) { |
244 | 554k | blocks[i].flags = 0; |
245 | 554k | } |
246 | | |
247 | 113k | zend_mark_reachable_blocks(op_array, cfg, start); |
248 | 113k | } |
249 | | /* }}} */ |
250 | | |
251 | 598k | static void initialize_block(zend_basic_block *block) { |
252 | 598k | block->flags = 0; |
253 | 598k | block->successors = block->successors_storage; |
254 | 598k | block->successors_count = 0; |
255 | 598k | block->predecessors_count = 0; |
256 | 598k | block->predecessor_offset = -1; |
257 | 598k | block->idom = -1; |
258 | 598k | block->loop_header = -1; |
259 | 598k | block->level = -1; |
260 | 598k | block->children = -1; |
261 | 598k | block->next_child = -1; |
262 | 598k | } |
263 | | |
264 | 774k | #define BB_START(i) do { \ |
265 | 774k | if (!block_map[i]) { blocks_count++;} \ |
266 | 774k | block_map[i]++; \ |
267 | 774k | } while (0) |
268 | | |
269 | | ZEND_API void zend_build_cfg(zend_arena **arena, const zend_op_array *op_array, uint32_t build_flags, zend_cfg *cfg) /* {{{ */ |
270 | 170k | { |
271 | 170k | uint32_t flags = 0; |
272 | 170k | uint32_t i; |
273 | 170k | uint32_t *block_map; |
274 | 170k | zend_function *fn; |
275 | 170k | int blocks_count = 0; |
276 | 170k | zend_basic_block *blocks; |
277 | 170k | zval *zv; |
278 | 170k | bool extra_entry_block = false; |
279 | | |
280 | 170k | cfg->flags = build_flags & (ZEND_CFG_STACKLESS|ZEND_CFG_RECV_ENTRY); |
281 | | |
282 | 170k | cfg->map = block_map = zend_arena_calloc(arena, op_array->last, sizeof(uint32_t)); |
283 | | |
284 | | /* Build CFG, Step 1: Find basic blocks starts, calculate number of blocks */ |
285 | 170k | BB_START(0); |
286 | 4.09M | for (i = 0; i < op_array->last; i++) { |
287 | 3.92M | zend_op *opline = op_array->opcodes + i; |
288 | 3.92M | switch (opline->opcode) { |
289 | 46.2k | case ZEND_RECV: |
290 | 52.8k | case ZEND_RECV_INIT: |
291 | 52.8k | if (build_flags & ZEND_CFG_RECV_ENTRY) { |
292 | 0 | BB_START(i + 1); |
293 | 0 | } |
294 | 52.8k | break; |
295 | 183k | case ZEND_RETURN: |
296 | 186k | case ZEND_RETURN_BY_REF: |
297 | 190k | case ZEND_GENERATOR_RETURN: |
298 | 190k | case ZEND_VERIFY_NEVER_TYPE: |
299 | 190k | if (i + 1 < op_array->last) { |
300 | 22.4k | BB_START(i + 1); |
301 | 22.4k | } |
302 | 190k | break; |
303 | 792 | case ZEND_MATCH_ERROR: |
304 | 4.79k | case ZEND_THROW: |
305 | | /* Don't treat THROW as terminator if it's used in expression context, |
306 | | * as we may lose live ranges when eliminating unreachable code. */ |
307 | 4.79k | if (opline->extended_value != ZEND_THROW_IS_EXPR && i + 1 < op_array->last) { |
308 | 3.21k | BB_START(i + 1); |
309 | 3.21k | } |
310 | 4.79k | break; |
311 | 9.52k | case ZEND_INCLUDE_OR_EVAL: |
312 | 9.52k | flags |= ZEND_FUNC_INDIRECT_VAR_ACCESS; |
313 | 9.52k | ZEND_FALLTHROUGH; |
314 | 13.2k | case ZEND_GENERATOR_CREATE: |
315 | 17.4k | case ZEND_YIELD: |
316 | 18.6k | case ZEND_YIELD_FROM: |
317 | 18.6k | if (build_flags & ZEND_CFG_STACKLESS) { |
318 | 0 | BB_START(i + 1); |
319 | 0 | } |
320 | 18.6k | break; |
321 | 336k | case ZEND_DO_FCALL: |
322 | 357k | case ZEND_DO_UCALL: |
323 | 357k | case ZEND_DO_FCALL_BY_NAME: |
324 | 357k | flags |= ZEND_FUNC_HAS_CALLS; |
325 | 357k | if (build_flags & ZEND_CFG_STACKLESS) { |
326 | 0 | BB_START(i + 1); |
327 | 0 | } |
328 | 357k | break; |
329 | 0 | case ZEND_DO_ICALL: |
330 | 0 | flags |= ZEND_FUNC_HAS_CALLS; |
331 | 0 | break; |
332 | 191k | case ZEND_INIT_FCALL: |
333 | 196k | case ZEND_INIT_NS_FCALL_BY_NAME: |
334 | 196k | zv = CRT_CONSTANT(opline->op2); |
335 | 196k | if (opline->opcode == ZEND_INIT_NS_FCALL_BY_NAME) { |
336 | | /* The third literal is the lowercased unqualified name */ |
337 | 4.25k | zv += 2; |
338 | 4.25k | } |
339 | 196k | if ((fn = zend_hash_find_ptr(EG(function_table), Z_STR_P(zv))) != NULL) { |
340 | 154k | if (fn->type == ZEND_INTERNAL_FUNCTION) { |
341 | 154k | flags |= zend_optimizer_classify_function( |
342 | 154k | Z_STR_P(zv), opline->extended_value); |
343 | 154k | } |
344 | 154k | } |
345 | 196k | break; |
346 | 1.06k | case ZEND_FAST_CALL: |
347 | 1.06k | BB_START(OP_JMP_ADDR(opline, opline->op1) - op_array->opcodes); |
348 | 1.06k | BB_START(i + 1); |
349 | 1.06k | break; |
350 | 788 | case ZEND_FAST_RET: |
351 | 788 | if (i + 1 < op_array->last) { |
352 | 788 | BB_START(i + 1); |
353 | 788 | } |
354 | 788 | break; |
355 | 71.4k | case ZEND_JMP: |
356 | 71.4k | BB_START(OP_JMP_ADDR(opline, opline->op1) - op_array->opcodes); |
357 | 71.4k | if (i + 1 < op_array->last) { |
358 | 70.5k | BB_START(i + 1); |
359 | 70.5k | } |
360 | 71.4k | break; |
361 | 26.6k | case ZEND_JMPZ: |
362 | 46.1k | case ZEND_JMPNZ: |
363 | 50.0k | case ZEND_JMPZ_EX: |
364 | 54.3k | case ZEND_JMPNZ_EX: |
365 | 58.0k | case ZEND_JMP_SET: |
366 | 66.1k | case ZEND_COALESCE: |
367 | 68.5k | case ZEND_ASSERT_CHECK: |
368 | 127k | case ZEND_JMP_NULL: |
369 | 128k | case ZEND_BIND_INIT_STATIC_OR_JMP: |
370 | 128k | case ZEND_JMP_FRAMELESS: |
371 | 128k | BB_START(OP_JMP_ADDR(opline, opline->op2) - op_array->opcodes); |
372 | 128k | BB_START(i + 1); |
373 | 128k | break; |
374 | 28.6k | case ZEND_CATCH: |
375 | 28.6k | if (!(opline->extended_value & ZEND_LAST_CATCH)) { |
376 | 4.07k | BB_START(OP_JMP_ADDR(opline, opline->op2) - op_array->opcodes); |
377 | 4.07k | } |
378 | 28.6k | BB_START(i + 1); |
379 | 28.6k | break; |
380 | 16.1k | case ZEND_FE_FETCH_R: |
381 | 17.8k | case ZEND_FE_FETCH_RW: |
382 | 17.8k | BB_START(ZEND_OFFSET_TO_OPLINE_NUM(op_array, opline, opline->extended_value)); |
383 | 17.8k | BB_START(i + 1); |
384 | 17.8k | break; |
385 | 16.1k | case ZEND_FE_RESET_R: |
386 | 17.8k | case ZEND_FE_RESET_RW: |
387 | 17.8k | BB_START(OP_JMP_ADDR(opline, opline->op2) - op_array->opcodes); |
388 | 17.8k | BB_START(i + 1); |
389 | 17.8k | break; |
390 | 20 | case ZEND_SWITCH_LONG: |
391 | 152 | case ZEND_SWITCH_STRING: |
392 | 530 | case ZEND_MATCH: |
393 | 530 | { |
394 | 530 | HashTable *jumptable = Z_ARRVAL_P(CRT_CONSTANT(opline->op2)); |
395 | 530 | zval *zv; |
396 | 4.82k | ZEND_HASH_FOREACH_VAL(jumptable, zv) { |
397 | 4.82k | BB_START(ZEND_OFFSET_TO_OPLINE_NUM(op_array, opline, Z_LVAL_P(zv))); |
398 | 4.82k | } ZEND_HASH_FOREACH_END(); |
399 | 530 | BB_START(ZEND_OFFSET_TO_OPLINE_NUM(op_array, opline, opline->extended_value)); |
400 | 530 | BB_START(i + 1); |
401 | 530 | break; |
402 | 152 | } |
403 | 9.45k | case ZEND_FETCH_R: |
404 | 12.9k | case ZEND_FETCH_W: |
405 | 15.5k | case ZEND_FETCH_RW: |
406 | 15.6k | case ZEND_FETCH_FUNC_ARG: |
407 | 15.8k | case ZEND_FETCH_IS: |
408 | 15.9k | case ZEND_FETCH_UNSET: |
409 | 16.2k | case ZEND_UNSET_VAR: |
410 | 16.6k | case ZEND_ISSET_ISEMPTY_VAR: |
411 | 16.6k | if (opline->extended_value & ZEND_FETCH_LOCAL) { |
412 | 14.0k | flags |= ZEND_FUNC_INDIRECT_VAR_ACCESS; |
413 | 14.0k | } else if ((opline->extended_value & (ZEND_FETCH_GLOBAL | ZEND_FETCH_GLOBAL_LOCK)) && |
414 | 2.58k | !op_array->function_name) { |
415 | 796 | flags |= ZEND_FUNC_INDIRECT_VAR_ACCESS; |
416 | 796 | } |
417 | 16.6k | break; |
418 | 184 | case ZEND_FUNC_GET_ARGS: |
419 | 184 | flags |= ZEND_FUNC_VARARG; |
420 | 184 | break; |
421 | 0 | case ZEND_EXT_STMT: |
422 | 0 | flags |= ZEND_FUNC_HAS_EXTENDED_STMT; |
423 | 0 | break; |
424 | 0 | case ZEND_EXT_FCALL_BEGIN: |
425 | 0 | case ZEND_EXT_FCALL_END: |
426 | 0 | flags |= ZEND_FUNC_HAS_EXTENDED_FCALL; |
427 | 0 | break; |
428 | 51.1k | case ZEND_FREE: |
429 | 69.4k | case ZEND_FE_FREE: |
430 | 69.4k | if (zend_optimizer_is_loop_var_free(opline) |
431 | 18.3k | && ((opline-1)->opcode != ZEND_MATCH_ERROR |
432 | 18.3k | || (opline-1)->extended_value != ZEND_THROW_IS_EXPR)) { |
433 | 18.3k | BB_START(i); |
434 | 18.3k | flags |= ZEND_FUNC_FREE_LOOP_VAR; |
435 | 18.3k | } |
436 | 69.4k | break; |
437 | 3.92M | } |
438 | 3.92M | } |
439 | | |
440 | | /* If the entry block has predecessors, we may need to split it */ |
441 | 170k | if ((build_flags & ZEND_CFG_NO_ENTRY_PREDECESSORS) |
442 | 75.7k | && op_array->last > 0 && block_map[0] > 1) { |
443 | 118 | extra_entry_block = true; |
444 | 118 | } |
445 | | |
446 | 170k | if (op_array->last_try_catch) { |
447 | 44.1k | for (uint32_t j = 0; j < op_array->last_try_catch; j++) { |
448 | 25.2k | BB_START(op_array->try_catch_array[j].try_op); |
449 | 25.2k | if (op_array->try_catch_array[j].catch_op) { |
450 | 24.6k | BB_START(op_array->try_catch_array[j].catch_op); |
451 | 24.6k | } |
452 | 25.2k | if (op_array->try_catch_array[j].finally_op) { |
453 | 788 | BB_START(op_array->try_catch_array[j].finally_op); |
454 | 788 | } |
455 | 25.2k | if (op_array->try_catch_array[j].finally_end) { |
456 | 788 | BB_START(op_array->try_catch_array[j].finally_end); |
457 | 788 | } |
458 | 25.2k | } |
459 | 18.9k | } |
460 | | |
461 | 170k | blocks_count += extra_entry_block; |
462 | 170k | cfg->blocks_count = blocks_count; |
463 | | |
464 | | /* Build CFG, Step 2: Build Array of Basic Blocks */ |
465 | 170k | cfg->blocks = blocks = zend_arena_calloc(arena, sizeof(zend_basic_block), blocks_count); |
466 | | |
467 | 170k | blocks_count = -1; |
468 | | |
469 | 170k | if (extra_entry_block) { |
470 | 118 | initialize_block(&blocks[0]); |
471 | 118 | blocks[0].start = 0; |
472 | 118 | blocks[0].len = 0; |
473 | 118 | blocks_count++; |
474 | 118 | } |
475 | | |
476 | 4.09M | for (i = 0; i < op_array->last; i++) { |
477 | 3.92M | if (block_map[i]) { |
478 | 598k | if (blocks_count >= 0) { |
479 | 428k | blocks[blocks_count].len = i - blocks[blocks_count].start; |
480 | 428k | } |
481 | 598k | blocks_count++; |
482 | 598k | initialize_block(&blocks[blocks_count]); |
483 | 598k | blocks[blocks_count].start = i; |
484 | 598k | } |
485 | 3.92M | block_map[i] = blocks_count; |
486 | 3.92M | } |
487 | | |
488 | 170k | blocks[blocks_count].len = i - blocks[blocks_count].start; |
489 | 170k | blocks_count++; |
490 | | |
491 | | /* Build CFG, Step 3: Calculate successors */ |
492 | 768k | for (int j = 0; j < blocks_count; j++) { |
493 | 598k | zend_basic_block *block = &blocks[j]; |
494 | 598k | zend_op *opline; |
495 | 598k | if (block->len == 0) { |
496 | 118 | block->successors_count = 1; |
497 | 118 | block->successors[0] = j + 1; |
498 | 118 | continue; |
499 | 118 | } |
500 | | |
501 | 598k | opline = op_array->opcodes + block->start + block->len - 1; |
502 | 598k | switch (opline->opcode) { |
503 | 788 | case ZEND_FAST_RET: |
504 | 183k | case ZEND_RETURN: |
505 | 187k | case ZEND_RETURN_BY_REF: |
506 | 191k | case ZEND_GENERATOR_RETURN: |
507 | 195k | case ZEND_THROW: |
508 | 195k | case ZEND_MATCH_ERROR: |
509 | 196k | case ZEND_VERIFY_NEVER_TYPE: |
510 | 196k | break; |
511 | 71.4k | case ZEND_JMP: |
512 | 71.4k | block->successors_count = 1; |
513 | 71.4k | block->successors[0] = block_map[OP_JMP_ADDR(opline, opline->op1) - op_array->opcodes]; |
514 | 71.4k | break; |
515 | 26.6k | case ZEND_JMPZ: |
516 | 46.1k | case ZEND_JMPNZ: |
517 | 50.0k | case ZEND_JMPZ_EX: |
518 | 54.3k | case ZEND_JMPNZ_EX: |
519 | 58.0k | case ZEND_JMP_SET: |
520 | 66.1k | case ZEND_COALESCE: |
521 | 68.5k | case ZEND_ASSERT_CHECK: |
522 | 127k | case ZEND_JMP_NULL: |
523 | 128k | case ZEND_BIND_INIT_STATIC_OR_JMP: |
524 | 128k | case ZEND_JMP_FRAMELESS: |
525 | 128k | block->successors_count = 2; |
526 | 128k | block->successors[0] = block_map[OP_JMP_ADDR(opline, opline->op2) - op_array->opcodes]; |
527 | 128k | block->successors[1] = j + 1; |
528 | 128k | break; |
529 | 28.6k | case ZEND_CATCH: |
530 | 28.6k | if (!(opline->extended_value & ZEND_LAST_CATCH)) { |
531 | 4.07k | block->successors_count = 2; |
532 | 4.07k | block->successors[0] = block_map[OP_JMP_ADDR(opline, opline->op2) - op_array->opcodes]; |
533 | 4.07k | block->successors[1] = j + 1; |
534 | 24.6k | } else { |
535 | 24.6k | block->successors_count = 1; |
536 | 24.6k | block->successors[0] = j + 1; |
537 | 24.6k | } |
538 | 28.6k | break; |
539 | 16.1k | case ZEND_FE_FETCH_R: |
540 | 17.8k | case ZEND_FE_FETCH_RW: |
541 | 17.8k | block->successors_count = 2; |
542 | 17.8k | block->successors[0] = block_map[ZEND_OFFSET_TO_OPLINE_NUM(op_array, opline, opline->extended_value)]; |
543 | 17.8k | block->successors[1] = j + 1; |
544 | 17.8k | break; |
545 | 16.1k | case ZEND_FE_RESET_R: |
546 | 17.8k | case ZEND_FE_RESET_RW: |
547 | 17.8k | block->successors_count = 2; |
548 | 17.8k | block->successors[0] = block_map[OP_JMP_ADDR(opline, opline->op2) - op_array->opcodes]; |
549 | 17.8k | block->successors[1] = j + 1; |
550 | 17.8k | break; |
551 | 1.06k | case ZEND_FAST_CALL: |
552 | 1.06k | block->successors_count = 2; |
553 | 1.06k | block->successors[0] = block_map[OP_JMP_ADDR(opline, opline->op1) - op_array->opcodes]; |
554 | 1.06k | block->successors[1] = j + 1; |
555 | 1.06k | break; |
556 | 20 | case ZEND_SWITCH_LONG: |
557 | 152 | case ZEND_SWITCH_STRING: |
558 | 530 | case ZEND_MATCH: |
559 | 530 | { |
560 | 530 | HashTable *jumptable = Z_ARRVAL_P(CRT_CONSTANT(opline->op2)); |
561 | 530 | zval *zv; |
562 | 530 | uint32_t s = 0; |
563 | | |
564 | 530 | block->successors_count = (opline->opcode == ZEND_MATCH ? 1 : 2) + zend_hash_num_elements(jumptable); |
565 | 530 | block->successors = zend_arena_calloc(arena, block->successors_count, sizeof(int)); |
566 | | |
567 | 4.82k | ZEND_HASH_FOREACH_VAL(jumptable, zv) { |
568 | 4.82k | block->successors[s++] = block_map[ZEND_OFFSET_TO_OPLINE_NUM(op_array, opline, Z_LVAL_P(zv))]; |
569 | 4.82k | } ZEND_HASH_FOREACH_END(); |
570 | | |
571 | 530 | block->successors[s++] = block_map[ZEND_OFFSET_TO_OPLINE_NUM(op_array, opline, opline->extended_value)]; |
572 | 530 | if (opline->opcode != ZEND_MATCH) { |
573 | 152 | block->successors[s++] = j + 1; |
574 | 152 | } |
575 | 530 | break; |
576 | 152 | } |
577 | 136k | default: |
578 | 136k | block->successors_count = 1; |
579 | 136k | block->successors[0] = j + 1; |
580 | 136k | break; |
581 | 598k | } |
582 | 598k | } |
583 | | |
584 | | /* Build CFG, Step 4, Mark Reachable Basic Blocks */ |
585 | 170k | cfg->flags |= flags; |
586 | 170k | zend_mark_reachable_blocks(op_array, cfg, 0); |
587 | 170k | } |
588 | | /* }}} */ |
589 | | |
590 | | ZEND_API void zend_cfg_build_predecessors(zend_arena **arena, zend_cfg *cfg) /* {{{ */ |
591 | 72.4k | { |
592 | 72.4k | zend_basic_block *b; |
593 | 72.4k | zend_basic_block *blocks = cfg->blocks; |
594 | 72.4k | zend_basic_block *end = blocks + cfg->blocks_count; |
595 | 72.4k | uint32_t edges = 0; |
596 | 72.4k | int *predecessors; |
597 | | |
598 | 274k | for (b = blocks; b < end; b++) { |
599 | 202k | b->predecessors_count = 0; |
600 | 202k | } |
601 | 274k | for (b = blocks; b < end; b++) { |
602 | 202k | if (!(b->flags & ZEND_BB_REACHABLE)) { |
603 | 28 | b->successors_count = 0; |
604 | 28 | b->predecessors_count = 0; |
605 | 202k | } else { |
606 | 393k | for (uint32_t s = 0; s < b->successors_count; s++) { |
607 | 191k | edges++; |
608 | 191k | blocks[b->successors[s]].predecessors_count++; |
609 | 191k | } |
610 | 202k | } |
611 | 202k | } |
612 | | |
613 | 72.4k | cfg->edges_count = edges; |
614 | 72.4k | cfg->predecessors = predecessors = (int*)zend_arena_calloc(arena, sizeof(int), edges); |
615 | | |
616 | 72.4k | edges = 0; |
617 | 274k | for (b = blocks; b < end; b++) { |
618 | 202k | if (b->flags & ZEND_BB_REACHABLE) { |
619 | 202k | b->predecessor_offset = edges; |
620 | 202k | edges += b->predecessors_count; |
621 | 202k | b->predecessors_count = 0; |
622 | 202k | } |
623 | 202k | } |
624 | | |
625 | 274k | for (uint32_t j = 0; j < cfg->blocks_count; j++) { |
626 | 202k | if (blocks[j].flags & ZEND_BB_REACHABLE) { |
627 | | /* SWITCH_STRING/LONG may have few identical successors */ |
628 | 393k | for (uint32_t s = 0; s < blocks[j].successors_count; s++) { |
629 | 191k | int duplicate = 0; |
630 | | |
631 | 258k | for (uint32_t p = 0; p < s; p++) { |
632 | 67.1k | if (blocks[j].successors[p] == blocks[j].successors[s]) { |
633 | 194 | duplicate = 1; |
634 | 194 | break; |
635 | 194 | } |
636 | 67.1k | } |
637 | 191k | if (!duplicate) { |
638 | 191k | zend_basic_block *b = blocks + blocks[j].successors[s]; |
639 | | |
640 | 191k | predecessors[b->predecessor_offset + b->predecessors_count] = j; |
641 | 191k | b->predecessors_count++; |
642 | 191k | } |
643 | 191k | } |
644 | 202k | } |
645 | 202k | } |
646 | 72.4k | } |
647 | | /* }}} */ |
648 | | |
649 | | /* Computes a postorder numbering of the CFG */ |
650 | | static void compute_postnum_recursive( |
651 | | int *postnum, uint32_t *cur, const zend_cfg *cfg, int block_num) /* {{{ */ |
652 | 205k | { |
653 | 205k | zend_basic_block *block = &cfg->blocks[block_num]; |
654 | 205k | if (postnum[block_num] != -1) { |
655 | 61.8k | return; |
656 | 61.8k | } |
657 | | |
658 | 143k | postnum[block_num] = -2; /* Marker for "currently visiting" */ |
659 | 334k | for (uint32_t s = 0; s < block->successors_count; s++) { |
660 | 191k | compute_postnum_recursive(postnum, cur, cfg, block->successors[s]); |
661 | 191k | } |
662 | 143k | postnum[block_num] = (*cur)++; |
663 | 143k | } |
664 | | /* }}} */ |
665 | | |
666 | | /* Computes dominator tree using algorithm from "A Simple, Fast Dominance Algorithm" by |
667 | | * Cooper, Harvey and Kennedy. */ |
668 | | ZEND_API void zend_cfg_compute_dominators_tree(const zend_op_array *op_array, zend_cfg *cfg) /* {{{ */ |
669 | 72.4k | { |
670 | 72.4k | zend_basic_block *blocks = cfg->blocks; |
671 | 72.4k | uint32_t blocks_count = cfg->blocks_count; |
672 | 72.4k | uint32_t j; |
673 | 72.4k | int changed; |
674 | | |
675 | 72.4k | if (cfg->blocks_count == 1) { |
676 | 58.9k | blocks[0].level = 0; |
677 | 58.9k | return; |
678 | 58.9k | } |
679 | | |
680 | 13.5k | ALLOCA_FLAG(use_heap) |
681 | 13.5k | int *postnum = do_alloca(sizeof(int) * cfg->blocks_count, use_heap); |
682 | 13.5k | memset(postnum, -1, sizeof(int) * cfg->blocks_count); |
683 | 13.5k | j = 0; |
684 | 13.5k | compute_postnum_recursive(postnum, &j, cfg, 0); |
685 | | |
686 | | /* FIXME: move declarations */ |
687 | 13.5k | blocks[0].idom = 0; |
688 | 32.8k | do { |
689 | 32.8k | changed = 0; |
690 | | /* Iterating in RPO here would converge faster */ |
691 | 350k | for (j = 1; j < blocks_count; j++) { |
692 | 317k | int idom = -1; |
693 | | |
694 | 317k | if ((blocks[j].flags & ZEND_BB_REACHABLE) == 0) { |
695 | 54 | continue; |
696 | 54 | } |
697 | 781k | for (uint32_t k = 0; k < blocks[j].predecessors_count; k++) { |
698 | 464k | int pred = cfg->predecessors[blocks[j].predecessor_offset + k]; |
699 | | |
700 | 464k | if (blocks[pred].idom >= 0) { |
701 | 406k | if (idom < 0) { |
702 | 283k | idom = pred; |
703 | 283k | } else { |
704 | 245k | while (idom != pred) { |
705 | 327k | while (postnum[pred] < postnum[idom]) pred = blocks[pred].idom; |
706 | 133k | while (postnum[idom] < postnum[pred]) idom = blocks[idom].idom; |
707 | 122k | } |
708 | 122k | } |
709 | 406k | } |
710 | 464k | } |
711 | | |
712 | 317k | if (idom >= 0 && blocks[j].idom != idom) { |
713 | 129k | blocks[j].idom = idom; |
714 | 129k | changed = 1; |
715 | 129k | } |
716 | 317k | } |
717 | 32.8k | } while (changed); |
718 | 13.5k | blocks[0].idom = -1; |
719 | | |
720 | 143k | for (j = 1; j < blocks_count; j++) { |
721 | 129k | if ((blocks[j].flags & ZEND_BB_REACHABLE) == 0) { |
722 | 28 | continue; |
723 | 28 | } |
724 | 129k | if (blocks[j].idom >= 0) { |
725 | | /* Sort by block number to traverse children in pre-order */ |
726 | 129k | if (blocks[blocks[j].idom].children < 0 || |
727 | 69.9k | j < blocks[blocks[j].idom].children) { |
728 | 69.9k | blocks[j].next_child = blocks[blocks[j].idom].children; |
729 | 69.9k | blocks[blocks[j].idom].children = j; |
730 | 69.9k | } else { |
731 | 59.8k | int k = blocks[blocks[j].idom].children; |
732 | 67.5k | while (blocks[k].next_child >=0 && j > blocks[k].next_child) { |
733 | 7.68k | k = blocks[k].next_child; |
734 | 7.68k | } |
735 | 59.8k | blocks[j].next_child = blocks[k].next_child; |
736 | 59.8k | blocks[k].next_child = j; |
737 | 59.8k | } |
738 | 129k | } |
739 | 129k | } |
740 | | |
741 | 156k | for (j = 0; j < blocks_count; j++) { |
742 | 143k | int idom = blocks[j].idom, level = 0; |
743 | 143k | if ((blocks[j].flags & ZEND_BB_REACHABLE) == 0) { |
744 | 28 | continue; |
745 | 28 | } |
746 | 149k | while (idom >= 0) { |
747 | 136k | level++; |
748 | 136k | if (blocks[idom].level >= 0) { |
749 | 129k | level += blocks[idom].level; |
750 | 129k | break; |
751 | 129k | } else { |
752 | 6.44k | idom = blocks[idom].idom; |
753 | 6.44k | } |
754 | 136k | } |
755 | 143k | blocks[j].level = level; |
756 | 143k | } |
757 | | |
758 | 13.5k | free_alloca(postnum, use_heap); |
759 | 13.5k | } |
760 | | /* }}} */ |
761 | | |
762 | | static bool dominates(zend_basic_block *blocks, int a, int b) /* {{{ */ |
763 | 65.8k | { |
764 | 108k | while (blocks[b].level > blocks[a].level) { |
765 | 42.9k | b = blocks[b].idom; |
766 | 42.9k | } |
767 | 65.8k | return a == b; |
768 | 65.8k | } |
769 | | /* }}} */ |
770 | | |
771 | | ZEND_API void zend_cfg_identify_loops(const zend_op_array *op_array, zend_cfg *cfg) /* {{{ */ |
772 | 72.4k | { |
773 | 72.4k | int i, j, n; |
774 | 72.4k | int time; |
775 | 72.4k | zend_basic_block *blocks = cfg->blocks; |
776 | 72.4k | int *entry_times, *exit_times; |
777 | 72.4k | zend_worklist work; |
778 | 72.4k | int flag = ZEND_FUNC_NO_LOOPS; |
779 | 72.4k | int *sorted_blocks; |
780 | 72.4k | ALLOCA_FLAG(list_use_heap) |
781 | 72.4k | ALLOCA_FLAG(tree_use_heap) |
782 | | |
783 | 72.4k | if (cfg->blocks_count == 1) { |
784 | 58.9k | cfg->flags |= flag; |
785 | 58.9k | return; |
786 | 58.9k | } |
787 | | |
788 | 13.5k | ZEND_WORKLIST_ALLOCA(&work, cfg->blocks_count, list_use_heap); |
789 | | |
790 | | /* We don't materialize the DJ spanning tree explicitly, as we are only interested in ancestor |
791 | | * queries. These are implemented by checking entry/exit times of the DFS search. */ |
792 | 13.5k | entry_times = do_alloca(3 * sizeof(int) * cfg->blocks_count, tree_use_heap); |
793 | 13.5k | exit_times = entry_times + cfg->blocks_count; |
794 | 13.5k | sorted_blocks = exit_times + cfg->blocks_count; |
795 | 13.5k | memset(entry_times, -1, 2 * sizeof(int) * cfg->blocks_count); |
796 | | |
797 | 13.5k | zend_worklist_push(&work, 0); |
798 | 13.5k | time = 0; |
799 | 156k | while (zend_worklist_len(&work)) { |
800 | 273k | next: |
801 | 273k | i = zend_worklist_peek(&work); |
802 | 273k | if (entry_times[i] == -1) { |
803 | 143k | entry_times[i] = time++; |
804 | 143k | } |
805 | | /* Visit blocks immediately dominated by i. */ |
806 | 421k | for (j = blocks[i].children; j >= 0; j = blocks[j].next_child) { |
807 | 230k | if (zend_worklist_push(&work, j)) { |
808 | 82.6k | goto next; |
809 | 82.6k | } |
810 | 230k | } |
811 | | /* Visit join edges. */ |
812 | 382k | for (j = 0; j < blocks[i].successors_count; j++) { |
813 | 239k | int succ = blocks[i].successors[j]; |
814 | 239k | if (blocks[succ].idom == i) { |
815 | 125k | continue; |
816 | 125k | } else if (zend_worklist_push(&work, succ)) { |
817 | 47.1k | goto next; |
818 | 47.1k | } |
819 | 239k | } |
820 | 143k | exit_times[i] = time++; |
821 | 143k | zend_worklist_pop(&work); |
822 | 143k | } |
823 | | |
824 | | /* Sort blocks by level, which is the opposite order in which we want to process them */ |
825 | 13.5k | sorted_blocks[0] = 0; |
826 | 13.5k | j = 0; |
827 | 13.5k | n = 1; |
828 | 92.5k | while (j != n) { |
829 | 79.0k | i = j; |
830 | 79.0k | j = n; |
831 | 222k | for (; i < j; i++) { |
832 | 143k | int child; |
833 | 273k | for (child = blocks[sorted_blocks[i]].children; child >= 0; child = blocks[child].next_child) { |
834 | 129k | sorted_blocks[n++] = child; |
835 | 129k | } |
836 | 143k | } |
837 | 79.0k | } |
838 | | |
839 | | /* Identify loops. See Sreedhar et al, "Identifying Loops Using DJ Graphs". */ |
840 | 156k | while (n > 0) { |
841 | 143k | i = sorted_blocks[--n]; |
842 | | |
843 | 143k | if (blocks[i].predecessors_count < 2) { |
844 | | /* loop header has at least two input edges */ |
845 | 84.9k | continue; |
846 | 84.9k | } |
847 | | |
848 | 178k | for (j = 0; j < blocks[i].predecessors_count; j++) { |
849 | 119k | int pred = cfg->predecessors[blocks[i].predecessor_offset + j]; |
850 | | |
851 | | /* A join edge is one for which the predecessor does not |
852 | | immediately dominate the successor. */ |
853 | 119k | if (blocks[i].idom == pred) { |
854 | 54.1k | continue; |
855 | 54.1k | } |
856 | | |
857 | | /* In a loop back-edge (back-join edge), the successor dominates |
858 | | the predecessor. */ |
859 | 65.8k | if (dominates(blocks, i, pred)) { |
860 | 12.2k | blocks[i].flags |= ZEND_BB_LOOP_HEADER; |
861 | 12.2k | flag &= ~ZEND_FUNC_NO_LOOPS; |
862 | 12.2k | if (!zend_worklist_len(&work)) { |
863 | 11.1k | zend_bitset_clear(work.visited, zend_bitset_len(cfg->blocks_count)); |
864 | 11.1k | } |
865 | 12.2k | zend_worklist_push(&work, pred); |
866 | 53.6k | } else { |
867 | | /* Otherwise it's a cross-join edge. See if it's a branch |
868 | | to an ancestor on the DJ spanning tree. */ |
869 | 53.6k | if (entry_times[pred] > entry_times[i] && exit_times[pred] < exit_times[i]) { |
870 | 0 | blocks[i].flags |= ZEND_BB_IRREDUCIBLE_LOOP; |
871 | 0 | flag |= ZEND_FUNC_IRREDUCIBLE; |
872 | 0 | flag &= ~ZEND_FUNC_NO_LOOPS; |
873 | 0 | } |
874 | 53.6k | } |
875 | 65.8k | } |
876 | 113k | while (zend_worklist_len(&work)) { |
877 | 55.1k | j = zend_worklist_pop(&work); |
878 | 58.6k | while (blocks[j].loop_header >= 0) { |
879 | 3.47k | j = blocks[j].loop_header; |
880 | 3.47k | } |
881 | 55.1k | if (j != i) { |
882 | 42.2k | if (blocks[j].idom < 0 && j != 0) { |
883 | | /* Ignore blocks that are unreachable or only abnormally reachable. */ |
884 | 0 | continue; |
885 | 0 | } |
886 | 42.2k | blocks[j].loop_header = i; |
887 | 100k | for (uint32_t k = 0; k < blocks[j].predecessors_count; k++) { |
888 | 57.7k | zend_worklist_push(&work, cfg->predecessors[blocks[j].predecessor_offset + k]); |
889 | 57.7k | } |
890 | 42.2k | } |
891 | 55.1k | } |
892 | 58.3k | } |
893 | | |
894 | 13.5k | free_alloca(entry_times, tree_use_heap); |
895 | 13.5k | ZEND_WORKLIST_FREE_ALLOCA(&work, list_use_heap); |
896 | | |
897 | 13.5k | cfg->flags |= flag; |
898 | 13.5k | } |
899 | | /* }}} */ |