Print this page
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/c1/c1_IR.cpp
+++ new/src/share/vm/c1/c1_IR.cpp
1 1 /*
2 2 * Copyright 1999-2010 Sun Microsystems, Inc. All Rights Reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 21 * have any questions.
22 22 *
23 23 */
24 24
25 25 # include "incls/_precompiled.incl"
26 26 # include "incls/_c1_IR.cpp.incl"
27 27
28 28
29 29 // Implementation of XHandlers
30 30 //
31 31 // Note: This code could eventually go away if we are
32 32 // just using the ciExceptionHandlerStream.
33 33
34 34 XHandlers::XHandlers(ciMethod* method) : _list(method->exception_table_length()) {
35 35 ciExceptionHandlerStream s(method);
36 36 while (!s.is_done()) {
37 37 _list.append(new XHandler(s.handler()));
38 38 s.next();
39 39 }
40 40 assert(s.count() == method->exception_table_length(), "exception table lengths inconsistent");
41 41 }
42 42
43 43 // deep copy of all XHandler contained in list
44 44 XHandlers::XHandlers(XHandlers* other) :
45 45 _list(other->length())
46 46 {
47 47 for (int i = 0; i < other->length(); i++) {
48 48 _list.append(new XHandler(other->handler_at(i)));
49 49 }
50 50 }
51 51
52 52 // Returns whether a particular exception type can be caught. Also
53 53 // returns true if klass is unloaded or any exception handler
54 54 // classes are unloaded. type_is_exact indicates whether the throw
55 55 // is known to be exactly that class or it might throw a subtype.
56 56 bool XHandlers::could_catch(ciInstanceKlass* klass, bool type_is_exact) const {
57 57 // the type is unknown so be conservative
58 58 if (!klass->is_loaded()) {
59 59 return true;
60 60 }
61 61
62 62 for (int i = 0; i < length(); i++) {
63 63 XHandler* handler = handler_at(i);
64 64 if (handler->is_catch_all()) {
65 65 // catch of ANY
66 66 return true;
67 67 }
68 68 ciInstanceKlass* handler_klass = handler->catch_klass();
69 69 // if it's unknown it might be catchable
70 70 if (!handler_klass->is_loaded()) {
71 71 return true;
72 72 }
73 73 // if the throw type is definitely a subtype of the catch type
74 74 // then it can be caught.
75 75 if (klass->is_subtype_of(handler_klass)) {
76 76 return true;
77 77 }
78 78 if (!type_is_exact) {
79 79 // If the type isn't exactly known then it can also be caught by
80 80 // catch statements where the inexact type is a subtype of the
81 81 // catch type.
82 82 // given: foo extends bar extends Exception
83 83 // throw bar can be caught by catch foo, catch bar, and catch
84 84 // Exception, however it can't be caught by any handlers without
85 85 // bar in its type hierarchy.
86 86 if (handler_klass->is_subtype_of(klass)) {
87 87 return true;
88 88 }
89 89 }
90 90 }
91 91
92 92 return false;
93 93 }
94 94
95 95
96 96 bool XHandlers::equals(XHandlers* others) const {
97 97 if (others == NULL) return false;
98 98 if (length() != others->length()) return false;
99 99
100 100 for (int i = 0; i < length(); i++) {
101 101 if (!handler_at(i)->equals(others->handler_at(i))) return false;
102 102 }
103 103 return true;
104 104 }
105 105
106 106 bool XHandler::equals(XHandler* other) const {
107 107 assert(entry_pco() != -1 && other->entry_pco() != -1, "must have entry_pco");
108 108
109 109 if (entry_pco() != other->entry_pco()) return false;
110 110 if (scope_count() != other->scope_count()) return false;
111 111 if (_desc != other->_desc) return false;
112 112
113 113 assert(entry_block() == other->entry_block(), "entry_block must be equal when entry_pco is equal");
114 114 return true;
115 115 }
116 116
117 117
118 118 // Implementation of IRScope
119 119
120 120 BlockBegin* IRScope::header_block(BlockBegin* entry, BlockBegin::Flag f, ValueStack* state) {
121 121 if (entry == NULL) return NULL;
122 122 assert(entry->is_set(f), "entry/flag mismatch");
123 123 // create header block
124 124 BlockBegin* h = new BlockBegin(entry->bci());
125 125 BlockEnd* g = new Goto(entry, false);
126 126 h->set_next(g, entry->bci());
127 127 h->set_end(g);
128 128 h->set(f);
129 129 // setup header block end state
130 130 ValueStack* s = state->copy(); // can use copy since stack is empty (=> no phis)
131 131 assert(s->stack_is_empty(), "must have empty stack at entry point");
132 132 g->set_state(s);
133 133 return h;
134 134 }
135 135
136 136
137 137 BlockBegin* IRScope::build_graph(Compilation* compilation, int osr_bci) {
138 138 GraphBuilder gm(compilation, this);
139 139 NOT_PRODUCT(if (PrintValueNumbering && Verbose) gm.print_stats());
140 140 if (compilation->bailed_out()) return NULL;
141 141 return gm.start();
142 142 }
143 143
144 144
145 145 IRScope::IRScope(Compilation* compilation, IRScope* caller, int caller_bci, ciMethod* method, int osr_bci, bool create_graph)
146 146 : _callees(2)
147 147 , _compilation(compilation)
148 148 , _lock_stack_size(-1)
149 149 , _requires_phi_function(method->max_locals())
150 150 {
151 151 _caller = caller;
152 152 _caller_bci = caller == NULL ? -1 : caller_bci;
153 153 _caller_state = NULL; // Must be set later if needed
154 154 _level = caller == NULL ? 0 : caller->level() + 1;
155 155 _method = method;
156 156 _xhandlers = new XHandlers(method);
157 157 _number_of_locks = 0;
158 158 _monitor_pairing_ok = method->has_balanced_monitors();
159 159 _start = NULL;
160 160
161 161 if (osr_bci == -1) {
162 162 _requires_phi_function.clear();
163 163 } else {
164 164 // selective creation of phi functions is not possibel in osr-methods
165 165 _requires_phi_function.set_range(0, method->max_locals());
166 166 }
167 167
168 168 assert(method->holder()->is_loaded() , "method holder must be loaded");
169 169
170 170 // build graph if monitor pairing is ok
171 171 if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci);
172 172 }
173 173
174 174
175 175 int IRScope::max_stack() const {
176 176 int my_max = method()->max_stack();
177 177 int callee_max = 0;
178 178 for (int i = 0; i < number_of_callees(); i++) {
179 179 callee_max = MAX2(callee_max, callee_no(i)->max_stack());
180 180 }
181 181 return my_max + callee_max;
182 182 }
183 183
184 184
185 185 void IRScope::compute_lock_stack_size() {
186 186 if (!InlineMethodsWithExceptionHandlers) {
187 187 _lock_stack_size = 0;
188 188 return;
189 189 }
190 190
191 191 // Figure out whether we have to preserve expression stack elements
192 192 // for parent scopes, and if so, how many
193 193 IRScope* cur_scope = this;
194 194 while (cur_scope != NULL && !cur_scope->xhandlers()->has_handlers()) {
195 195 cur_scope = cur_scope->caller();
196 196 }
197 197 _lock_stack_size = (cur_scope == NULL ? 0 :
198 198 (cur_scope->caller_state() == NULL ? 0 :
199 199 cur_scope->caller_state()->stack_size()));
200 200 }
201 201
202 202 int IRScope::top_scope_bci() const {
203 203 assert(!is_top_scope(), "no correct answer for top scope possible");
204 204 const IRScope* scope = this;
205 205 while (!scope->caller()->is_top_scope()) {
206 206 scope = scope->caller();
207 207 }
208 208 return scope->caller_bci();
209 209 }
210 210
211 211 bool IRScopeDebugInfo::should_reexecute() {
212 212 ciMethod* cur_method = scope()->method();
213 213 int cur_bci = bci();
214 214 if (cur_method != NULL && cur_bci != SynchronizationEntryBCI) {
215 215 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
216 216 return Interpreter::bytecode_should_reexecute(code);
217 217 } else
218 218 return false;
219 219 }
220 220
221 221
222 222 // Implementation of CodeEmitInfo
↓ open down ↓ |
222 lines elided |
↑ open up ↑ |
223 223
224 224 // Stack must be NON-null
225 225 CodeEmitInfo::CodeEmitInfo(int bci, ValueStack* stack, XHandlers* exception_handlers)
226 226 : _scope(stack->scope())
227 227 , _bci(bci)
228 228 , _scope_debug_info(NULL)
229 229 , _oop_map(NULL)
230 230 , _stack(stack)
231 231 , _exception_handlers(exception_handlers)
232 232 , _next(NULL)
233 - , _id(-1) {
233 + , _id(-1)
234 + , _is_method_handle_invoke(false) {
234 235 assert(_stack != NULL, "must be non null");
235 236 assert(_bci == SynchronizationEntryBCI || Bytecodes::is_defined(scope()->method()->java_code_at_bci(_bci)), "make sure bci points at a real bytecode");
236 237 }
237 238
238 239
239 240 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, bool lock_stack_only)
240 241 : _scope(info->_scope)
241 242 , _exception_handlers(NULL)
242 243 , _bci(info->_bci)
243 244 , _scope_debug_info(NULL)
244 - , _oop_map(NULL) {
245 + , _oop_map(NULL)
246 + , _is_method_handle_invoke(info->_is_method_handle_invoke) {
245 247 if (lock_stack_only) {
246 248 if (info->_stack != NULL) {
247 249 _stack = info->_stack->copy_locks();
248 250 } else {
249 251 _stack = NULL;
250 252 }
251 253 } else {
252 254 _stack = info->_stack;
253 255 }
254 256
255 257 // deep copy of exception handlers
256 258 if (info->_exception_handlers != NULL) {
257 259 _exception_handlers = new XHandlers(info->_exception_handlers);
258 260 }
259 261 }
260 262
261 263
262 -void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset, bool is_method_handle_invoke) {
264 +void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) {
263 265 // record the safepoint before recording the debug info for enclosing scopes
264 266 recorder->add_safepoint(pc_offset, _oop_map->deep_copy());
265 - _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, is_method_handle_invoke);
267 + _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, _is_method_handle_invoke);
266 268 recorder->end_safepoint(pc_offset);
267 269 }
268 270
269 271
270 272 void CodeEmitInfo::add_register_oop(LIR_Opr opr) {
271 273 assert(_oop_map != NULL, "oop map must already exist");
272 274 assert(opr->is_single_cpu(), "should not call otherwise");
273 275
274 276 int frame_size = frame_map()->framesize();
275 277 int arg_count = frame_map()->oop_map_arg_count();
276 278 VMReg name = frame_map()->regname(opr);
277 279 _oop_map->set_oop(name);
278 280 }
279 281
280 282
281 283
282 284
283 285 // Implementation of IR
284 286
285 287 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) :
286 288 _locals_size(in_WordSize(-1))
287 289 , _num_loops(0) {
288 290 // initialize data structures
289 291 ValueType::initialize();
290 292 Instruction::initialize();
291 293 BlockBegin::initialize();
292 294 GraphBuilder::initialize();
293 295 // setup IR fields
294 296 _compilation = compilation;
295 297 _top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true);
296 298 _code = NULL;
297 299 }
298 300
299 301
300 302 void IR::optimize() {
301 303 Optimizer opt(this);
302 304 if (DoCEE) {
303 305 opt.eliminate_conditional_expressions();
304 306 #ifndef PRODUCT
305 307 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); }
306 308 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); }
307 309 #endif
308 310 }
309 311 if (EliminateBlocks) {
310 312 opt.eliminate_blocks();
311 313 #ifndef PRODUCT
312 314 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); }
313 315 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); }
314 316 #endif
315 317 }
316 318 if (EliminateNullChecks) {
317 319 opt.eliminate_null_checks();
318 320 #ifndef PRODUCT
319 321 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); }
320 322 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); }
321 323 #endif
322 324 }
323 325 }
324 326
325 327
326 328 static int sort_pairs(BlockPair** a, BlockPair** b) {
327 329 if ((*a)->from() == (*b)->from()) {
328 330 return (*a)->to()->block_id() - (*b)->to()->block_id();
329 331 } else {
330 332 return (*a)->from()->block_id() - (*b)->from()->block_id();
331 333 }
332 334 }
333 335
334 336
335 337 class CriticalEdgeFinder: public BlockClosure {
336 338 BlockPairList blocks;
337 339 IR* _ir;
338 340
339 341 public:
340 342 CriticalEdgeFinder(IR* ir): _ir(ir) {}
341 343 void block_do(BlockBegin* bb) {
342 344 BlockEnd* be = bb->end();
343 345 int nos = be->number_of_sux();
344 346 if (nos >= 2) {
345 347 for (int i = 0; i < nos; i++) {
346 348 BlockBegin* sux = be->sux_at(i);
347 349 if (sux->number_of_preds() >= 2) {
348 350 blocks.append(new BlockPair(bb, sux));
349 351 }
350 352 }
351 353 }
352 354 }
353 355
354 356 void split_edges() {
355 357 BlockPair* last_pair = NULL;
356 358 blocks.sort(sort_pairs);
357 359 for (int i = 0; i < blocks.length(); i++) {
358 360 BlockPair* pair = blocks.at(i);
359 361 if (last_pair != NULL && pair->is_same(last_pair)) continue;
360 362 BlockBegin* from = pair->from();
361 363 BlockBegin* to = pair->to();
362 364 BlockBegin* split = from->insert_block_between(to);
363 365 #ifndef PRODUCT
364 366 if ((PrintIR || PrintIR1) && Verbose) {
365 367 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)",
366 368 from->block_id(), to->block_id(), split->block_id());
367 369 }
368 370 #endif
369 371 last_pair = pair;
370 372 }
371 373 }
372 374 };
373 375
374 376 void IR::split_critical_edges() {
375 377 CriticalEdgeFinder cef(this);
376 378
377 379 iterate_preorder(&cef);
378 380 cef.split_edges();
379 381 }
380 382
381 383
382 384 class UseCountComputer: public AllStatic {
383 385 private:
384 386 static void update_use_count(Value* n) {
385 387 // Local instructions and Phis for expression stack values at the
386 388 // start of basic blocks are not added to the instruction list
387 389 if ((*n)->bci() == -99 && (*n)->as_Local() == NULL &&
388 390 (*n)->as_Phi() == NULL) {
389 391 assert(false, "a node was not appended to the graph");
390 392 Compilation::current_compilation()->bailout("a node was not appended to the graph");
391 393 }
392 394 // use n's input if not visited before
393 395 if (!(*n)->is_pinned() && !(*n)->has_uses()) {
394 396 // note: a) if the instruction is pinned, it will be handled by compute_use_count
395 397 // b) if the instruction has uses, it was touched before
396 398 // => in both cases we don't need to update n's values
397 399 uses_do(n);
398 400 }
399 401 // use n
400 402 (*n)->_use_count++;
401 403 }
402 404
403 405 static Values* worklist;
404 406 static int depth;
405 407 enum {
406 408 max_recurse_depth = 20
407 409 };
408 410
409 411 static void uses_do(Value* n) {
410 412 depth++;
411 413 if (depth > max_recurse_depth) {
412 414 // don't allow the traversal to recurse too deeply
413 415 worklist->push(*n);
414 416 } else {
415 417 (*n)->input_values_do(update_use_count);
416 418 // special handling for some instructions
417 419 if ((*n)->as_BlockEnd() != NULL) {
418 420 // note on BlockEnd:
419 421 // must 'use' the stack only if the method doesn't
420 422 // terminate, however, in those cases stack is empty
421 423 (*n)->state_values_do(update_use_count);
422 424 }
423 425 }
424 426 depth--;
425 427 }
426 428
427 429 static void basic_compute_use_count(BlockBegin* b) {
428 430 depth = 0;
429 431 // process all pinned nodes as the roots of expression trees
430 432 for (Instruction* n = b; n != NULL; n = n->next()) {
431 433 if (n->is_pinned()) uses_do(&n);
432 434 }
433 435 assert(depth == 0, "should have counted back down");
434 436
435 437 // now process any unpinned nodes which recursed too deeply
436 438 while (worklist->length() > 0) {
437 439 Value t = worklist->pop();
438 440 if (!t->is_pinned()) {
439 441 // compute the use count
440 442 uses_do(&t);
441 443
442 444 // pin the instruction so that LIRGenerator doesn't recurse
443 445 // too deeply during it's evaluation.
444 446 t->pin();
445 447 }
446 448 }
447 449 assert(depth == 0, "should have counted back down");
448 450 }
449 451
450 452 public:
451 453 static void compute(BlockList* blocks) {
452 454 worklist = new Values();
453 455 blocks->blocks_do(basic_compute_use_count);
454 456 worklist = NULL;
455 457 }
456 458 };
457 459
458 460
459 461 Values* UseCountComputer::worklist = NULL;
460 462 int UseCountComputer::depth = 0;
461 463
462 464 // helper macro for short definition of trace-output inside code
463 465 #ifndef PRODUCT
464 466 #define TRACE_LINEAR_SCAN(level, code) \
465 467 if (TraceLinearScanLevel >= level) { \
466 468 code; \
467 469 }
468 470 #else
469 471 #define TRACE_LINEAR_SCAN(level, code)
470 472 #endif
471 473
472 474 class ComputeLinearScanOrder : public StackObj {
473 475 private:
474 476 int _max_block_id; // the highest block_id of a block
475 477 int _num_blocks; // total number of blocks (smaller than _max_block_id)
476 478 int _num_loops; // total number of loops
477 479 bool _iterative_dominators;// method requires iterative computation of dominatiors
478 480
479 481 BlockList* _linear_scan_order; // the resulting list of blocks in correct order
480 482
481 483 BitMap _visited_blocks; // used for recursive processing of blocks
482 484 BitMap _active_blocks; // used for recursive processing of blocks
483 485 BitMap _dominator_blocks; // temproary BitMap used for computation of dominator
484 486 intArray _forward_branches; // number of incoming forward branches for each block
485 487 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges
486 488 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop
487 489 BlockList _work_list; // temporary list (used in mark_loops and compute_order)
488 490
489 491 // accessors for _visited_blocks and _active_blocks
490 492 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); }
491 493 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); }
492 494 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); }
493 495 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); }
494 496 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); }
495 497 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); }
496 498
497 499 // accessors for _forward_branches
498 500 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); }
499 501 int dec_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) - 1); return _forward_branches.at(b->block_id()); }
500 502
501 503 // accessors for _loop_map
502 504 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); }
503 505 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); }
504 506 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); }
505 507
506 508 // count edges between blocks
507 509 void count_edges(BlockBegin* cur, BlockBegin* parent);
508 510
509 511 // loop detection
510 512 void mark_loops();
511 513 void clear_non_natural_loops(BlockBegin* start_block);
512 514 void assign_loop_depth(BlockBegin* start_block);
513 515
514 516 // computation of final block order
515 517 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b);
516 518 void compute_dominator(BlockBegin* cur, BlockBegin* parent);
517 519 int compute_weight(BlockBegin* cur);
518 520 bool ready_for_processing(BlockBegin* cur);
519 521 void sort_into_work_list(BlockBegin* b);
520 522 void append_block(BlockBegin* cur);
521 523 void compute_order(BlockBegin* start_block);
522 524
523 525 // fixup of dominators for non-natural loops
524 526 bool compute_dominators_iter();
525 527 void compute_dominators();
526 528
527 529 // debug functions
528 530 NOT_PRODUCT(void print_blocks();)
529 531 DEBUG_ONLY(void verify();)
530 532
531 533 public:
532 534 ComputeLinearScanOrder(BlockBegin* start_block);
533 535
534 536 // accessors for final result
535 537 BlockList* linear_scan_order() const { return _linear_scan_order; }
536 538 int num_loops() const { return _num_loops; }
537 539 };
538 540
539 541
540 542 ComputeLinearScanOrder::ComputeLinearScanOrder(BlockBegin* start_block) :
541 543 _max_block_id(BlockBegin::number_of_blocks()),
542 544 _num_blocks(0),
543 545 _num_loops(0),
544 546 _iterative_dominators(false),
545 547 _visited_blocks(_max_block_id),
546 548 _active_blocks(_max_block_id),
547 549 _dominator_blocks(_max_block_id),
548 550 _forward_branches(_max_block_id, 0),
549 551 _loop_end_blocks(8),
550 552 _work_list(8),
551 553 _linear_scan_order(NULL), // initialized later with correct size
552 554 _loop_map(0, 0) // initialized later with correct size
553 555 {
554 556 TRACE_LINEAR_SCAN(2, "***** computing linear-scan block order");
555 557
556 558 init_visited();
557 559 count_edges(start_block, NULL);
558 560
559 561 if (_num_loops > 0) {
560 562 mark_loops();
561 563 clear_non_natural_loops(start_block);
562 564 assign_loop_depth(start_block);
563 565 }
564 566
565 567 compute_order(start_block);
566 568 compute_dominators();
567 569
568 570 NOT_PRODUCT(print_blocks());
569 571 DEBUG_ONLY(verify());
570 572 }
571 573
572 574
573 575 // Traverse the CFG:
574 576 // * count total number of blocks
575 577 // * count all incoming edges and backward incoming edges
576 578 // * number loop header blocks
577 579 // * create a list with all loop end blocks
578 580 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) {
579 581 TRACE_LINEAR_SCAN(3, tty->print_cr("Enter count_edges for block B%d coming from B%d", cur->block_id(), parent != NULL ? parent->block_id() : -1));
580 582 assert(cur->dominator() == NULL, "dominator already initialized");
581 583
582 584 if (is_active(cur)) {
583 585 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch"));
584 586 assert(is_visited(cur), "block must be visisted when block is active");
585 587 assert(parent != NULL, "must have parent");
586 588
587 589 cur->set(BlockBegin::linear_scan_loop_header_flag);
588 590 cur->set(BlockBegin::backward_branch_target_flag);
589 591
590 592 parent->set(BlockBegin::linear_scan_loop_end_flag);
591 593
592 594 // When a loop header is also the start of an exception handler, then the backward branch is
593 595 // an exception edge. Because such edges are usually critical edges which cannot be split, the
594 596 // loop must be excluded here from processing.
595 597 if (cur->is_set(BlockBegin::exception_entry_flag)) {
596 598 // Make sure that dominators are correct in this weird situation
597 599 _iterative_dominators = true;
598 600 return;
599 601 }
600 602 assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur,
601 603 "loop end blocks must have one successor (critical edges are split)");
602 604
603 605 _loop_end_blocks.append(parent);
604 606 return;
605 607 }
606 608
607 609 // increment number of incoming forward branches
608 610 inc_forward_branches(cur);
609 611
610 612 if (is_visited(cur)) {
611 613 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited"));
612 614 return;
613 615 }
614 616
615 617 _num_blocks++;
616 618 set_visited(cur);
617 619 set_active(cur);
618 620
619 621 // recursive call for all successors
620 622 int i;
621 623 for (i = cur->number_of_sux() - 1; i >= 0; i--) {
622 624 count_edges(cur->sux_at(i), cur);
623 625 }
624 626 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
625 627 count_edges(cur->exception_handler_at(i), cur);
626 628 }
627 629
628 630 clear_active(cur);
629 631
630 632 // Each loop has a unique number.
631 633 // When multiple loops are nested, assign_loop_depth assumes that the
632 634 // innermost loop has the lowest number. This is guaranteed by setting
633 635 // the loop number after the recursive calls for the successors above
634 636 // have returned.
635 637 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) {
636 638 assert(cur->loop_index() == -1, "cannot set loop-index twice");
637 639 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops));
638 640
639 641 cur->set_loop_index(_num_loops);
640 642 _num_loops++;
641 643 }
642 644
643 645 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id()));
644 646 }
645 647
646 648
647 649 void ComputeLinearScanOrder::mark_loops() {
648 650 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops"));
649 651
650 652 _loop_map = BitMap2D(_num_loops, _max_block_id);
651 653 _loop_map.clear();
652 654
653 655 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) {
654 656 BlockBegin* loop_end = _loop_end_blocks.at(i);
655 657 BlockBegin* loop_start = loop_end->sux_at(0);
656 658 int loop_idx = loop_start->loop_index();
657 659
658 660 TRACE_LINEAR_SCAN(3, tty->print_cr("Processing loop from B%d to B%d (loop %d):", loop_start->block_id(), loop_end->block_id(), loop_idx));
659 661 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set");
660 662 assert(loop_end->number_of_sux() == 1, "incorrect number of successors");
661 663 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set");
662 664 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set");
663 665 assert(_work_list.is_empty(), "work list must be empty before processing");
664 666
665 667 // add the end-block of the loop to the working list
666 668 _work_list.push(loop_end);
667 669 set_block_in_loop(loop_idx, loop_end);
668 670 do {
669 671 BlockBegin* cur = _work_list.pop();
670 672
671 673 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id()));
672 674 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list");
673 675
674 676 // recursive processing of all predecessors ends when start block of loop is reached
675 677 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) {
676 678 for (int j = cur->number_of_preds() - 1; j >= 0; j--) {
677 679 BlockBegin* pred = cur->pred_at(j);
678 680
679 681 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) {
680 682 // this predecessor has not been processed yet, so add it to work list
681 683 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id()));
682 684 _work_list.push(pred);
683 685 set_block_in_loop(loop_idx, pred);
684 686 }
685 687 }
686 688 }
687 689 } while (!_work_list.is_empty());
688 690 }
689 691 }
690 692
691 693
692 694 // check for non-natural loops (loops where the loop header does not dominate
693 695 // all other loop blocks = loops with mulitple entries).
694 696 // such loops are ignored
695 697 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) {
696 698 for (int i = _num_loops - 1; i >= 0; i--) {
697 699 if (is_block_in_loop(i, start_block)) {
698 700 // loop i contains the entry block of the method
699 701 // -> this is not a natural loop, so ignore it
700 702 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i));
701 703
702 704 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) {
703 705 clear_block_in_loop(i, block_id);
704 706 }
705 707 _iterative_dominators = true;
706 708 }
707 709 }
708 710 }
709 711
710 712 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) {
711 713 TRACE_LINEAR_SCAN(3, "----- computing loop-depth and weight");
712 714 init_visited();
713 715
714 716 assert(_work_list.is_empty(), "work list must be empty before processing");
715 717 _work_list.append(start_block);
716 718
717 719 do {
718 720 BlockBegin* cur = _work_list.pop();
719 721
720 722 if (!is_visited(cur)) {
721 723 set_visited(cur);
722 724 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id()));
723 725
724 726 // compute loop-depth and loop-index for the block
725 727 assert(cur->loop_depth() == 0, "cannot set loop-depth twice");
726 728 int i;
727 729 int loop_depth = 0;
728 730 int min_loop_idx = -1;
729 731 for (i = _num_loops - 1; i >= 0; i--) {
730 732 if (is_block_in_loop(i, cur)) {
731 733 loop_depth++;
732 734 min_loop_idx = i;
733 735 }
734 736 }
735 737 cur->set_loop_depth(loop_depth);
736 738 cur->set_loop_index(min_loop_idx);
737 739
738 740 // append all unvisited successors to work list
739 741 for (i = cur->number_of_sux() - 1; i >= 0; i--) {
740 742 _work_list.append(cur->sux_at(i));
741 743 }
742 744 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
743 745 _work_list.append(cur->exception_handler_at(i));
744 746 }
745 747 }
746 748 } while (!_work_list.is_empty());
747 749 }
748 750
749 751
750 752 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) {
751 753 assert(a != NULL && b != NULL, "must have input blocks");
752 754
753 755 _dominator_blocks.clear();
754 756 while (a != NULL) {
755 757 _dominator_blocks.set_bit(a->block_id());
756 758 assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized");
757 759 a = a->dominator();
758 760 }
759 761 while (b != NULL && !_dominator_blocks.at(b->block_id())) {
760 762 assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized");
761 763 b = b->dominator();
762 764 }
763 765
764 766 assert(b != NULL, "could not find dominator");
765 767 return b;
766 768 }
767 769
768 770 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) {
769 771 if (cur->dominator() == NULL) {
770 772 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id()));
771 773 cur->set_dominator(parent);
772 774
773 775 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) {
774 776 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: computing dominator of B%d: common dominator of B%d and B%d is B%d", cur->block_id(), parent->block_id(), cur->dominator()->block_id(), common_dominator(cur->dominator(), parent)->block_id()));
775 777 assert(cur->number_of_preds() > 1, "");
776 778 cur->set_dominator(common_dominator(cur->dominator(), parent));
777 779 }
778 780 }
779 781
780 782
781 783 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) {
782 784 BlockBegin* single_sux = NULL;
783 785 if (cur->number_of_sux() == 1) {
784 786 single_sux = cur->sux_at(0);
785 787 }
786 788
787 789 // limit loop-depth to 15 bit (only for security reason, it will never be so big)
788 790 int weight = (cur->loop_depth() & 0x7FFF) << 16;
789 791
790 792 // general macro for short definition of weight flags
791 793 // the first instance of INC_WEIGHT_IF has the highest priority
792 794 int cur_bit = 15;
793 795 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--;
794 796
795 797 // this is necessery for the (very rare) case that two successing blocks have
796 798 // the same loop depth, but a different loop index (can happen for endless loops
797 799 // with exception handlers)
798 800 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag));
799 801
800 802 // loop end blocks (blocks that end with a backward branch) are added
801 803 // after all other blocks of the loop.
802 804 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag));
803 805
804 806 // critical edge split blocks are prefered because than they have a bigger
805 807 // proability to be completely empty
806 808 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag));
807 809
808 810 // exceptions should not be thrown in normal control flow, so these blocks
809 811 // are added as late as possible
810 812 INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL));
811 813 INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL));
812 814
813 815 // exceptions handlers are added as late as possible
814 816 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag));
815 817
816 818 // guarantee that weight is > 0
817 819 weight |= 1;
818 820
819 821 #undef INC_WEIGHT_IF
820 822 assert(cur_bit >= 0, "too many flags");
821 823 assert(weight > 0, "weight cannot become negative");
822 824
823 825 return weight;
824 826 }
825 827
826 828 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) {
827 829 // Discount the edge just traveled.
828 830 // When the number drops to zero, all forward branches were processed
829 831 if (dec_forward_branches(cur) != 0) {
830 832 return false;
831 833 }
832 834
833 835 assert(_linear_scan_order->index_of(cur) == -1, "block already processed (block can be ready only once)");
834 836 assert(_work_list.index_of(cur) == -1, "block already in work-list (block can be ready only once)");
835 837 return true;
836 838 }
837 839
838 840 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) {
839 841 assert(_work_list.index_of(cur) == -1, "block already in work list");
840 842
841 843 int cur_weight = compute_weight(cur);
842 844
843 845 // the linear_scan_number is used to cache the weight of a block
844 846 cur->set_linear_scan_number(cur_weight);
845 847
846 848 #ifndef PRODUCT
847 849 if (StressLinearScan) {
848 850 _work_list.insert_before(0, cur);
849 851 return;
850 852 }
851 853 #endif
852 854
853 855 _work_list.append(NULL); // provide space for new element
854 856
855 857 int insert_idx = _work_list.length() - 1;
856 858 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) {
857 859 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1));
858 860 insert_idx--;
859 861 }
860 862 _work_list.at_put(insert_idx, cur);
861 863
862 864 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id()));
863 865 TRACE_LINEAR_SCAN(3, for (int i = 0; i < _work_list.length(); i++) tty->print_cr("%8d B%2d weight:%6x", i, _work_list.at(i)->block_id(), _work_list.at(i)->linear_scan_number()));
864 866
865 867 #ifdef ASSERT
866 868 for (int i = 0; i < _work_list.length(); i++) {
867 869 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set");
868 870 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist");
869 871 }
870 872 #endif
871 873 }
872 874
873 875 void ComputeLinearScanOrder::append_block(BlockBegin* cur) {
874 876 TRACE_LINEAR_SCAN(3, tty->print_cr("appending block B%d (weight 0x%6x) to linear-scan order", cur->block_id(), cur->linear_scan_number()));
875 877 assert(_linear_scan_order->index_of(cur) == -1, "cannot add the same block twice");
876 878
877 879 // currently, the linear scan order and code emit order are equal.
878 880 // therefore the linear_scan_number and the weight of a block must also
879 881 // be equal.
880 882 cur->set_linear_scan_number(_linear_scan_order->length());
881 883 _linear_scan_order->append(cur);
882 884 }
883 885
884 886 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) {
885 887 TRACE_LINEAR_SCAN(3, "----- computing final block order");
886 888
887 889 // the start block is always the first block in the linear scan order
888 890 _linear_scan_order = new BlockList(_num_blocks);
889 891 append_block(start_block);
890 892
891 893 assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction");
892 894 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry();
893 895 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry();
894 896
895 897 BlockBegin* sux_of_osr_entry = NULL;
896 898 if (osr_entry != NULL) {
897 899 // special handling for osr entry:
898 900 // ignore the edge between the osr entry and its successor for processing
899 901 // the osr entry block is added manually below
900 902 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor");
901 903 assert(osr_entry->sux_at(0)->number_of_preds() >= 2, "sucessor of osr entry must have two predecessors (otherwise it is not present in normal control flow");
902 904
903 905 sux_of_osr_entry = osr_entry->sux_at(0);
904 906 dec_forward_branches(sux_of_osr_entry);
905 907
906 908 compute_dominator(osr_entry, start_block);
907 909 _iterative_dominators = true;
908 910 }
909 911 compute_dominator(std_entry, start_block);
910 912
911 913 // start processing with standard entry block
912 914 assert(_work_list.is_empty(), "list must be empty before processing");
913 915
914 916 if (ready_for_processing(std_entry)) {
915 917 sort_into_work_list(std_entry);
916 918 } else {
917 919 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)");
918 920 }
919 921
920 922 do {
921 923 BlockBegin* cur = _work_list.pop();
922 924
923 925 if (cur == sux_of_osr_entry) {
924 926 // the osr entry block is ignored in normal processing, it is never added to the
925 927 // work list. Instead, it is added as late as possible manually here.
926 928 append_block(osr_entry);
927 929 compute_dominator(cur, osr_entry);
928 930 }
929 931 append_block(cur);
930 932
931 933 int i;
932 934 int num_sux = cur->number_of_sux();
933 935 // changed loop order to get "intuitive" order of if- and else-blocks
934 936 for (i = 0; i < num_sux; i++) {
935 937 BlockBegin* sux = cur->sux_at(i);
936 938 compute_dominator(sux, cur);
937 939 if (ready_for_processing(sux)) {
938 940 sort_into_work_list(sux);
939 941 }
940 942 }
941 943 num_sux = cur->number_of_exception_handlers();
942 944 for (i = 0; i < num_sux; i++) {
943 945 BlockBegin* sux = cur->exception_handler_at(i);
944 946 compute_dominator(sux, cur);
945 947 if (ready_for_processing(sux)) {
946 948 sort_into_work_list(sux);
947 949 }
948 950 }
949 951 } while (_work_list.length() > 0);
950 952 }
951 953
952 954
953 955 bool ComputeLinearScanOrder::compute_dominators_iter() {
954 956 bool changed = false;
955 957 int num_blocks = _linear_scan_order->length();
956 958
957 959 assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator");
958 960 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors");
959 961 for (int i = 1; i < num_blocks; i++) {
960 962 BlockBegin* block = _linear_scan_order->at(i);
961 963
962 964 BlockBegin* dominator = block->pred_at(0);
963 965 int num_preds = block->number_of_preds();
964 966 for (int i = 1; i < num_preds; i++) {
965 967 dominator = common_dominator(dominator, block->pred_at(i));
966 968 }
967 969
968 970 if (dominator != block->dominator()) {
969 971 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: updating dominator of B%d from B%d to B%d", block->block_id(), block->dominator()->block_id(), dominator->block_id()));
970 972
971 973 block->set_dominator(dominator);
972 974 changed = true;
973 975 }
974 976 }
975 977 return changed;
976 978 }
977 979
978 980 void ComputeLinearScanOrder::compute_dominators() {
979 981 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators));
980 982
981 983 // iterative computation of dominators is only required for methods with non-natural loops
982 984 // and OSR-methods. For all other methods, the dominators computed when generating the
983 985 // linear scan block order are correct.
984 986 if (_iterative_dominators) {
985 987 do {
986 988 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation"));
987 989 } while (compute_dominators_iter());
988 990 }
989 991
990 992 // check that dominators are correct
991 993 assert(!compute_dominators_iter(), "fix point not reached");
992 994 }
993 995
994 996
995 997 #ifndef PRODUCT
996 998 void ComputeLinearScanOrder::print_blocks() {
997 999 if (TraceLinearScanLevel >= 2) {
998 1000 tty->print_cr("----- loop information:");
999 1001 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
1000 1002 BlockBegin* cur = _linear_scan_order->at(block_idx);
1001 1003
1002 1004 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id());
1003 1005 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
1004 1006 tty->print ("%d ", is_block_in_loop(loop_idx, cur));
1005 1007 }
1006 1008 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth());
1007 1009 }
1008 1010 }
1009 1011
1010 1012 if (TraceLinearScanLevel >= 1) {
1011 1013 tty->print_cr("----- linear-scan block order:");
1012 1014 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
1013 1015 BlockBegin* cur = _linear_scan_order->at(block_idx);
1014 1016 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth());
1015 1017
1016 1018 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " ");
1017 1019 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " ");
1018 1020 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " ");
1019 1021 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " ");
1020 1022
1021 1023 if (cur->dominator() != NULL) {
1022 1024 tty->print(" dom: B%d ", cur->dominator()->block_id());
1023 1025 } else {
1024 1026 tty->print(" dom: NULL ");
1025 1027 }
1026 1028
1027 1029 if (cur->number_of_preds() > 0) {
1028 1030 tty->print(" preds: ");
1029 1031 for (int j = 0; j < cur->number_of_preds(); j++) {
1030 1032 BlockBegin* pred = cur->pred_at(j);
1031 1033 tty->print("B%d ", pred->block_id());
1032 1034 }
1033 1035 }
1034 1036 if (cur->number_of_sux() > 0) {
1035 1037 tty->print(" sux: ");
1036 1038 for (int j = 0; j < cur->number_of_sux(); j++) {
1037 1039 BlockBegin* sux = cur->sux_at(j);
1038 1040 tty->print("B%d ", sux->block_id());
1039 1041 }
1040 1042 }
1041 1043 if (cur->number_of_exception_handlers() > 0) {
1042 1044 tty->print(" ex: ");
1043 1045 for (int j = 0; j < cur->number_of_exception_handlers(); j++) {
1044 1046 BlockBegin* ex = cur->exception_handler_at(j);
1045 1047 tty->print("B%d ", ex->block_id());
1046 1048 }
1047 1049 }
1048 1050 tty->cr();
1049 1051 }
1050 1052 }
1051 1053 }
1052 1054 #endif
1053 1055
1054 1056 #ifdef ASSERT
1055 1057 void ComputeLinearScanOrder::verify() {
1056 1058 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list");
1057 1059
1058 1060 if (StressLinearScan) {
1059 1061 // blocks are scrambled when StressLinearScan is used
1060 1062 return;
1061 1063 }
1062 1064
1063 1065 // check that all successors of a block have a higher linear-scan-number
1064 1066 // and that all predecessors of a block have a lower linear-scan-number
1065 1067 // (only backward branches of loops are ignored)
1066 1068 int i;
1067 1069 for (i = 0; i < _linear_scan_order->length(); i++) {
1068 1070 BlockBegin* cur = _linear_scan_order->at(i);
1069 1071
1070 1072 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number");
1071 1073 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->index_of(cur), "incorrect linear_scan_number");
1072 1074
1073 1075 int j;
1074 1076 for (j = cur->number_of_sux() - 1; j >= 0; j--) {
1075 1077 BlockBegin* sux = cur->sux_at(j);
1076 1078
1077 1079 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->index_of(sux), "incorrect linear_scan_number");
1078 1080 if (!cur->is_set(BlockBegin::linear_scan_loop_end_flag)) {
1079 1081 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order");
1080 1082 }
1081 1083 if (cur->loop_depth() == sux->loop_depth()) {
1082 1084 assert(cur->loop_index() == sux->loop_index() || sux->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index");
1083 1085 }
1084 1086 }
1085 1087
1086 1088 for (j = cur->number_of_preds() - 1; j >= 0; j--) {
1087 1089 BlockBegin* pred = cur->pred_at(j);
1088 1090
1089 1091 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->index_of(pred), "incorrect linear_scan_number");
1090 1092 if (!cur->is_set(BlockBegin::linear_scan_loop_header_flag)) {
1091 1093 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order");
1092 1094 }
1093 1095 if (cur->loop_depth() == pred->loop_depth()) {
1094 1096 assert(cur->loop_index() == pred->loop_index() || cur->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index");
1095 1097 }
1096 1098
1097 1099 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors");
1098 1100 }
1099 1101
1100 1102 // check dominator
1101 1103 if (i == 0) {
1102 1104 assert(cur->dominator() == NULL, "first block has no dominator");
1103 1105 } else {
1104 1106 assert(cur->dominator() != NULL, "all but first block must have dominator");
1105 1107 }
1106 1108 assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0), "Single predecessor must also be dominator");
1107 1109 }
1108 1110
1109 1111 // check that all loops are continuous
1110 1112 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
1111 1113 int block_idx = 0;
1112 1114 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop");
1113 1115
1114 1116 // skip blocks before the loop
1115 1117 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
1116 1118 block_idx++;
1117 1119 }
1118 1120 // skip blocks of loop
1119 1121 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
1120 1122 block_idx++;
1121 1123 }
1122 1124 // after the first non-loop block, there must not be another loop-block
1123 1125 while (block_idx < _num_blocks) {
1124 1126 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order");
1125 1127 block_idx++;
1126 1128 }
1127 1129 }
1128 1130 }
1129 1131 #endif
1130 1132
1131 1133
1132 1134 void IR::compute_code() {
1133 1135 assert(is_valid(), "IR must be valid");
1134 1136
1135 1137 ComputeLinearScanOrder compute_order(start());
1136 1138 _num_loops = compute_order.num_loops();
1137 1139 _code = compute_order.linear_scan_order();
1138 1140 }
1139 1141
1140 1142
1141 1143 void IR::compute_use_counts() {
1142 1144 // make sure all values coming out of this block get evaluated.
1143 1145 int num_blocks = _code->length();
1144 1146 for (int i = 0; i < num_blocks; i++) {
1145 1147 _code->at(i)->end()->state()->pin_stack_for_linear_scan();
1146 1148 }
1147 1149
1148 1150 // compute use counts
1149 1151 UseCountComputer::compute(_code);
1150 1152 }
1151 1153
1152 1154
1153 1155 void IR::iterate_preorder(BlockClosure* closure) {
1154 1156 assert(is_valid(), "IR must be valid");
1155 1157 start()->iterate_preorder(closure);
1156 1158 }
1157 1159
1158 1160
1159 1161 void IR::iterate_postorder(BlockClosure* closure) {
1160 1162 assert(is_valid(), "IR must be valid");
1161 1163 start()->iterate_postorder(closure);
1162 1164 }
1163 1165
1164 1166 void IR::iterate_linear_scan_order(BlockClosure* closure) {
1165 1167 linear_scan_order()->iterate_forward(closure);
1166 1168 }
1167 1169
1168 1170
1169 1171 #ifndef PRODUCT
1170 1172 class BlockPrinter: public BlockClosure {
1171 1173 private:
1172 1174 InstructionPrinter* _ip;
1173 1175 bool _cfg_only;
1174 1176 bool _live_only;
1175 1177
1176 1178 public:
1177 1179 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) {
1178 1180 _ip = ip;
1179 1181 _cfg_only = cfg_only;
1180 1182 _live_only = live_only;
1181 1183 }
1182 1184
1183 1185 virtual void block_do(BlockBegin* block) {
1184 1186 if (_cfg_only) {
1185 1187 _ip->print_instr(block); tty->cr();
1186 1188 } else {
1187 1189 block->print_block(*_ip, _live_only);
1188 1190 }
1189 1191 }
1190 1192 };
1191 1193
1192 1194
1193 1195 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) {
1194 1196 ttyLocker ttyl;
1195 1197 InstructionPrinter ip(!cfg_only);
1196 1198 BlockPrinter bp(&ip, cfg_only, live_only);
1197 1199 start->iterate_preorder(&bp);
1198 1200 tty->cr();
1199 1201 }
1200 1202
1201 1203 void IR::print(bool cfg_only, bool live_only) {
1202 1204 if (is_valid()) {
1203 1205 print(start(), cfg_only, live_only);
1204 1206 } else {
1205 1207 tty->print_cr("invalid IR");
1206 1208 }
1207 1209 }
1208 1210
1209 1211
1210 1212 define_array(BlockListArray, BlockList*)
1211 1213 define_stack(BlockListList, BlockListArray)
1212 1214
1213 1215 class PredecessorValidator : public BlockClosure {
1214 1216 private:
1215 1217 BlockListList* _predecessors;
1216 1218 BlockList* _blocks;
1217 1219
1218 1220 static int cmp(BlockBegin** a, BlockBegin** b) {
1219 1221 return (*a)->block_id() - (*b)->block_id();
1220 1222 }
1221 1223
1222 1224 public:
1223 1225 PredecessorValidator(IR* hir) {
1224 1226 ResourceMark rm;
1225 1227 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), NULL);
1226 1228 _blocks = new BlockList();
1227 1229
1228 1230 int i;
1229 1231 hir->start()->iterate_preorder(this);
1230 1232 if (hir->code() != NULL) {
1231 1233 assert(hir->code()->length() == _blocks->length(), "must match");
1232 1234 for (i = 0; i < _blocks->length(); i++) {
1233 1235 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists");
1234 1236 }
1235 1237 }
1236 1238
1237 1239 for (i = 0; i < _blocks->length(); i++) {
1238 1240 BlockBegin* block = _blocks->at(i);
1239 1241 BlockList* preds = _predecessors->at(block->block_id());
1240 1242 if (preds == NULL) {
1241 1243 assert(block->number_of_preds() == 0, "should be the same");
1242 1244 continue;
1243 1245 }
1244 1246
1245 1247 // clone the pred list so we can mutate it
1246 1248 BlockList* pred_copy = new BlockList();
1247 1249 int j;
1248 1250 for (j = 0; j < block->number_of_preds(); j++) {
1249 1251 pred_copy->append(block->pred_at(j));
1250 1252 }
1251 1253 // sort them in the same order
1252 1254 preds->sort(cmp);
1253 1255 pred_copy->sort(cmp);
1254 1256 int length = MIN2(preds->length(), block->number_of_preds());
1255 1257 for (j = 0; j < block->number_of_preds(); j++) {
1256 1258 assert(preds->at(j) == pred_copy->at(j), "must match");
1257 1259 }
1258 1260
1259 1261 assert(preds->length() == block->number_of_preds(), "should be the same");
1260 1262 }
1261 1263 }
1262 1264
1263 1265 virtual void block_do(BlockBegin* block) {
1264 1266 _blocks->append(block);
1265 1267 BlockEnd* be = block->end();
1266 1268 int n = be->number_of_sux();
1267 1269 int i;
1268 1270 for (i = 0; i < n; i++) {
1269 1271 BlockBegin* sux = be->sux_at(i);
1270 1272 assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler");
1271 1273
1272 1274 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL);
1273 1275 if (preds == NULL) {
1274 1276 preds = new BlockList();
1275 1277 _predecessors->at_put(sux->block_id(), preds);
1276 1278 }
1277 1279 preds->append(block);
1278 1280 }
1279 1281
1280 1282 n = block->number_of_exception_handlers();
1281 1283 for (i = 0; i < n; i++) {
1282 1284 BlockBegin* sux = block->exception_handler_at(i);
1283 1285 assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler");
1284 1286
1285 1287 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL);
1286 1288 if (preds == NULL) {
1287 1289 preds = new BlockList();
1288 1290 _predecessors->at_put(sux->block_id(), preds);
1289 1291 }
1290 1292 preds->append(block);
1291 1293 }
1292 1294 }
1293 1295 };
1294 1296
1295 1297 void IR::verify() {
1296 1298 #ifdef ASSERT
1297 1299 PredecessorValidator pv(this);
1298 1300 #endif
1299 1301 }
1300 1302
1301 1303 #endif // PRODUCT
1302 1304
1303 1305 void SubstitutionResolver::substitute(Value* v) {
1304 1306 Value v0 = *v;
1305 1307 if (v0) {
1306 1308 Value vs = v0->subst();
1307 1309 if (vs != v0) {
1308 1310 *v = v0->subst();
1309 1311 }
1310 1312 }
1311 1313 }
1312 1314
1313 1315 #ifdef ASSERT
1314 1316 void check_substitute(Value* v) {
1315 1317 Value v0 = *v;
1316 1318 if (v0) {
1317 1319 Value vs = v0->subst();
1318 1320 assert(vs == v0, "missed substitution");
1319 1321 }
1320 1322 }
1321 1323 #endif
1322 1324
1323 1325
1324 1326 void SubstitutionResolver::block_do(BlockBegin* block) {
1325 1327 Instruction* last = NULL;
1326 1328 for (Instruction* n = block; n != NULL;) {
1327 1329 n->values_do(substitute);
1328 1330 // need to remove this instruction from the instruction stream
1329 1331 if (n->subst() != n) {
1330 1332 assert(last != NULL, "must have last");
1331 1333 last->set_next(n->next(), n->next()->bci());
1332 1334 } else {
1333 1335 last = n;
1334 1336 }
1335 1337 n = last->next();
1336 1338 }
1337 1339
1338 1340 #ifdef ASSERT
1339 1341 if (block->state()) block->state()->values_do(check_substitute);
1340 1342 block->block_values_do(check_substitute);
1341 1343 if (block->end() && block->end()->state()) block->end()->state()->values_do(check_substitute);
1342 1344 #endif
1343 1345 }
↓ open down ↓ |
1068 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX