Mercurial > hg > truffle
annotate src/share/vm/c1/c1_IR.cpp @ 1339:09ac706c2623
Merge
author | asaha |
---|---|
date | Wed, 24 Mar 2010 17:16:33 -0700 |
parents | 89e0543e1737 |
children | 3cf667df43ef |
rev | line source |
---|---|
0 | 1 /* |
948 | 2 * Copyright 1999-2009 Sun Microsystems, Inc. All Rights Reserved. |
0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 # include "incls/_precompiled.incl" | |
26 # include "incls/_c1_IR.cpp.incl" | |
27 | |
28 | |
29 // Implementation of XHandlers | |
30 // | |
31 // Note: This code could eventually go away if we are | |
32 // just using the ciExceptionHandlerStream. | |
33 | |
34 XHandlers::XHandlers(ciMethod* method) : _list(method->exception_table_length()) { | |
35 ciExceptionHandlerStream s(method); | |
36 while (!s.is_done()) { | |
37 _list.append(new XHandler(s.handler())); | |
38 s.next(); | |
39 } | |
40 assert(s.count() == method->exception_table_length(), "exception table lengths inconsistent"); | |
41 } | |
42 | |
43 // deep copy of all XHandler contained in list | |
44 XHandlers::XHandlers(XHandlers* other) : | |
45 _list(other->length()) | |
46 { | |
47 for (int i = 0; i < other->length(); i++) { | |
48 _list.append(new XHandler(other->handler_at(i))); | |
49 } | |
50 } | |
51 | |
52 // Returns whether a particular exception type can be caught. Also | |
53 // returns true if klass is unloaded or any exception handler | |
54 // classes are unloaded. type_is_exact indicates whether the throw | |
55 // is known to be exactly that class or it might throw a subtype. | |
56 bool XHandlers::could_catch(ciInstanceKlass* klass, bool type_is_exact) const { | |
57 // the type is unknown so be conservative | |
58 if (!klass->is_loaded()) { | |
59 return true; | |
60 } | |
61 | |
62 for (int i = 0; i < length(); i++) { | |
63 XHandler* handler = handler_at(i); | |
64 if (handler->is_catch_all()) { | |
65 // catch of ANY | |
66 return true; | |
67 } | |
68 ciInstanceKlass* handler_klass = handler->catch_klass(); | |
69 // if it's unknown it might be catchable | |
70 if (!handler_klass->is_loaded()) { | |
71 return true; | |
72 } | |
73 // if the throw type is definitely a subtype of the catch type | |
74 // then it can be caught. | |
75 if (klass->is_subtype_of(handler_klass)) { | |
76 return true; | |
77 } | |
78 if (!type_is_exact) { | |
79 // If the type isn't exactly known then it can also be caught by | |
80 // catch statements where the inexact type is a subtype of the | |
81 // catch type. | |
82 // given: foo extends bar extends Exception | |
83 // throw bar can be caught by catch foo, catch bar, and catch | |
84 // Exception, however it can't be caught by any handlers without | |
85 // bar in its type hierarchy. | |
86 if (handler_klass->is_subtype_of(klass)) { | |
87 return true; | |
88 } | |
89 } | |
90 } | |
91 | |
92 return false; | |
93 } | |
94 | |
95 | |
96 bool XHandlers::equals(XHandlers* others) const { | |
97 if (others == NULL) return false; | |
98 if (length() != others->length()) return false; | |
99 | |
100 for (int i = 0; i < length(); i++) { | |
101 if (!handler_at(i)->equals(others->handler_at(i))) return false; | |
102 } | |
103 return true; | |
104 } | |
105 | |
106 bool XHandler::equals(XHandler* other) const { | |
107 assert(entry_pco() != -1 && other->entry_pco() != -1, "must have entry_pco"); | |
108 | |
109 if (entry_pco() != other->entry_pco()) return false; | |
110 if (scope_count() != other->scope_count()) return false; | |
111 if (_desc != other->_desc) return false; | |
112 | |
113 assert(entry_block() == other->entry_block(), "entry_block must be equal when entry_pco is equal"); | |
114 return true; | |
115 } | |
116 | |
117 | |
118 // Implementation of IRScope | |
119 | |
120 BlockBegin* IRScope::header_block(BlockBegin* entry, BlockBegin::Flag f, ValueStack* state) { | |
121 if (entry == NULL) return NULL; | |
122 assert(entry->is_set(f), "entry/flag mismatch"); | |
123 // create header block | |
124 BlockBegin* h = new BlockBegin(entry->bci()); | |
125 BlockEnd* g = new Goto(entry, false); | |
126 h->set_next(g, entry->bci()); | |
127 h->set_end(g); | |
128 h->set(f); | |
129 // setup header block end state | |
130 ValueStack* s = state->copy(); // can use copy since stack is empty (=> no phis) | |
131 assert(s->stack_is_empty(), "must have empty stack at entry point"); | |
132 g->set_state(s); | |
133 return h; | |
134 } | |
135 | |
136 | |
137 BlockBegin* IRScope::build_graph(Compilation* compilation, int osr_bci) { | |
138 GraphBuilder gm(compilation, this); | |
139 NOT_PRODUCT(if (PrintValueNumbering && Verbose) gm.print_stats()); | |
140 if (compilation->bailed_out()) return NULL; | |
141 return gm.start(); | |
142 } | |
143 | |
144 | |
145 IRScope::IRScope(Compilation* compilation, IRScope* caller, int caller_bci, ciMethod* method, int osr_bci, bool create_graph) | |
146 : _callees(2) | |
147 , _compilation(compilation) | |
148 , _lock_stack_size(-1) | |
149 , _requires_phi_function(method->max_locals()) | |
150 { | |
151 _caller = caller; | |
152 _caller_bci = caller == NULL ? -1 : caller_bci; | |
153 _caller_state = NULL; // Must be set later if needed | |
154 _level = caller == NULL ? 0 : caller->level() + 1; | |
155 _method = method; | |
156 _xhandlers = new XHandlers(method); | |
157 _number_of_locks = 0; | |
158 _monitor_pairing_ok = method->has_balanced_monitors(); | |
159 _start = NULL; | |
160 | |
161 if (osr_bci == -1) { | |
162 _requires_phi_function.clear(); | |
163 } else { | |
164 // selective creation of phi functions is not possibel in osr-methods | |
165 _requires_phi_function.set_range(0, method->max_locals()); | |
166 } | |
167 | |
168 assert(method->holder()->is_loaded() , "method holder must be loaded"); | |
169 | |
170 // build graph if monitor pairing is ok | |
171 if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci); | |
172 } | |
173 | |
174 | |
175 int IRScope::max_stack() const { | |
176 int my_max = method()->max_stack(); | |
177 int callee_max = 0; | |
178 for (int i = 0; i < number_of_callees(); i++) { | |
179 callee_max = MAX2(callee_max, callee_no(i)->max_stack()); | |
180 } | |
181 return my_max + callee_max; | |
182 } | |
183 | |
184 | |
185 void IRScope::compute_lock_stack_size() { | |
186 if (!InlineMethodsWithExceptionHandlers) { | |
187 _lock_stack_size = 0; | |
188 return; | |
189 } | |
190 | |
191 // Figure out whether we have to preserve expression stack elements | |
192 // for parent scopes, and if so, how many | |
193 IRScope* cur_scope = this; | |
194 while (cur_scope != NULL && !cur_scope->xhandlers()->has_handlers()) { | |
195 cur_scope = cur_scope->caller(); | |
196 } | |
197 _lock_stack_size = (cur_scope == NULL ? 0 : | |
198 (cur_scope->caller_state() == NULL ? 0 : | |
199 cur_scope->caller_state()->stack_size())); | |
200 } | |
201 | |
202 int IRScope::top_scope_bci() const { | |
203 assert(!is_top_scope(), "no correct answer for top scope possible"); | |
204 const IRScope* scope = this; | |
205 while (!scope->caller()->is_top_scope()) { | |
206 scope = scope->caller(); | |
207 } | |
208 return scope->caller_bci(); | |
209 } | |
210 | |
900
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
211 bool IRScopeDebugInfo::should_reexecute() { |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
212 ciMethod* cur_method = scope()->method(); |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
213 int cur_bci = bci(); |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
214 if (cur_method != NULL && cur_bci != SynchronizationEntryBCI) { |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
215 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
216 return Interpreter::bytecode_should_reexecute(code); |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
217 } else |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
218 return false; |
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
219 } |
0 | 220 |
221 | |
222 // Implementation of CodeEmitInfo | |
223 | |
224 // Stack must be NON-null | |
225 CodeEmitInfo::CodeEmitInfo(int bci, ValueStack* stack, XHandlers* exception_handlers) | |
226 : _scope(stack->scope()) | |
227 , _bci(bci) | |
228 , _scope_debug_info(NULL) | |
229 , _oop_map(NULL) | |
230 , _stack(stack) | |
231 , _exception_handlers(exception_handlers) | |
232 , _next(NULL) | |
233 , _id(-1) { | |
234 assert(_stack != NULL, "must be non null"); | |
235 assert(_bci == SynchronizationEntryBCI || Bytecodes::is_defined(scope()->method()->java_code_at_bci(_bci)), "make sure bci points at a real bytecode"); | |
236 } | |
237 | |
238 | |
239 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, bool lock_stack_only) | |
240 : _scope(info->_scope) | |
241 , _exception_handlers(NULL) | |
242 , _bci(info->_bci) | |
243 , _scope_debug_info(NULL) | |
244 , _oop_map(NULL) { | |
245 if (lock_stack_only) { | |
246 if (info->_stack != NULL) { | |
247 _stack = info->_stack->copy_locks(); | |
248 } else { | |
249 _stack = NULL; | |
250 } | |
251 } else { | |
252 _stack = info->_stack; | |
253 } | |
254 | |
255 // deep copy of exception handlers | |
256 if (info->_exception_handlers != NULL) { | |
257 _exception_handlers = new XHandlers(info->_exception_handlers); | |
258 } | |
259 } | |
260 | |
261 | |
262 void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) { | |
263 // record the safepoint before recording the debug info for enclosing scopes | |
264 recorder->add_safepoint(pc_offset, _oop_map->deep_copy()); | |
900
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
265 _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/); |
0 | 266 recorder->end_safepoint(pc_offset); |
267 } | |
268 | |
269 | |
270 void CodeEmitInfo::add_register_oop(LIR_Opr opr) { | |
271 assert(_oop_map != NULL, "oop map must already exist"); | |
272 assert(opr->is_single_cpu(), "should not call otherwise"); | |
273 | |
274 int frame_size = frame_map()->framesize(); | |
275 int arg_count = frame_map()->oop_map_arg_count(); | |
276 VMReg name = frame_map()->regname(opr); | |
277 _oop_map->set_oop(name); | |
278 } | |
279 | |
280 | |
281 | |
282 | |
283 // Implementation of IR | |
284 | |
285 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) : | |
286 _locals_size(in_WordSize(-1)) | |
287 , _num_loops(0) { | |
288 // initialize data structures | |
289 ValueType::initialize(); | |
290 Instruction::initialize(); | |
291 BlockBegin::initialize(); | |
292 GraphBuilder::initialize(); | |
293 // setup IR fields | |
294 _compilation = compilation; | |
295 _top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true); | |
296 _code = NULL; | |
297 } | |
298 | |
299 | |
300 void IR::optimize() { | |
301 Optimizer opt(this); | |
302 if (DoCEE) { | |
303 opt.eliminate_conditional_expressions(); | |
304 #ifndef PRODUCT | |
305 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); } | |
306 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); } | |
307 #endif | |
308 } | |
309 if (EliminateBlocks) { | |
310 opt.eliminate_blocks(); | |
311 #ifndef PRODUCT | |
312 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); } | |
313 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); } | |
314 #endif | |
315 } | |
316 if (EliminateNullChecks) { | |
317 opt.eliminate_null_checks(); | |
318 #ifndef PRODUCT | |
319 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); } | |
320 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); } | |
321 #endif | |
322 } | |
323 } | |
324 | |
325 | |
326 static int sort_pairs(BlockPair** a, BlockPair** b) { | |
327 if ((*a)->from() == (*b)->from()) { | |
328 return (*a)->to()->block_id() - (*b)->to()->block_id(); | |
329 } else { | |
330 return (*a)->from()->block_id() - (*b)->from()->block_id(); | |
331 } | |
332 } | |
333 | |
334 | |
335 class CriticalEdgeFinder: public BlockClosure { | |
336 BlockPairList blocks; | |
337 IR* _ir; | |
338 | |
339 public: | |
340 CriticalEdgeFinder(IR* ir): _ir(ir) {} | |
341 void block_do(BlockBegin* bb) { | |
342 BlockEnd* be = bb->end(); | |
343 int nos = be->number_of_sux(); | |
344 if (nos >= 2) { | |
345 for (int i = 0; i < nos; i++) { | |
346 BlockBegin* sux = be->sux_at(i); | |
347 if (sux->number_of_preds() >= 2) { | |
348 blocks.append(new BlockPair(bb, sux)); | |
349 } | |
350 } | |
351 } | |
352 } | |
353 | |
354 void split_edges() { | |
355 BlockPair* last_pair = NULL; | |
356 blocks.sort(sort_pairs); | |
357 for (int i = 0; i < blocks.length(); i++) { | |
358 BlockPair* pair = blocks.at(i); | |
359 if (last_pair != NULL && pair->is_same(last_pair)) continue; | |
360 BlockBegin* from = pair->from(); | |
361 BlockBegin* to = pair->to(); | |
362 BlockBegin* split = from->insert_block_between(to); | |
363 #ifndef PRODUCT | |
364 if ((PrintIR || PrintIR1) && Verbose) { | |
365 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)", | |
366 from->block_id(), to->block_id(), split->block_id()); | |
367 } | |
368 #endif | |
369 last_pair = pair; | |
370 } | |
371 } | |
372 }; | |
373 | |
374 void IR::split_critical_edges() { | |
375 CriticalEdgeFinder cef(this); | |
376 | |
377 iterate_preorder(&cef); | |
378 cef.split_edges(); | |
379 } | |
380 | |
381 | |
382 class UseCountComputer: public AllStatic { | |
383 private: | |
384 static void update_use_count(Value* n) { | |
385 // Local instructions and Phis for expression stack values at the | |
386 // start of basic blocks are not added to the instruction list | |
387 if ((*n)->bci() == -99 && (*n)->as_Local() == NULL && | |
388 (*n)->as_Phi() == NULL) { | |
389 assert(false, "a node was not appended to the graph"); | |
390 Compilation::current_compilation()->bailout("a node was not appended to the graph"); | |
391 } | |
392 // use n's input if not visited before | |
393 if (!(*n)->is_pinned() && !(*n)->has_uses()) { | |
394 // note: a) if the instruction is pinned, it will be handled by compute_use_count | |
395 // b) if the instruction has uses, it was touched before | |
396 // => in both cases we don't need to update n's values | |
397 uses_do(n); | |
398 } | |
399 // use n | |
400 (*n)->_use_count++; | |
401 } | |
402 | |
403 static Values* worklist; | |
404 static int depth; | |
405 enum { | |
406 max_recurse_depth = 20 | |
407 }; | |
408 | |
409 static void uses_do(Value* n) { | |
410 depth++; | |
411 if (depth > max_recurse_depth) { | |
412 // don't allow the traversal to recurse too deeply | |
413 worklist->push(*n); | |
414 } else { | |
415 (*n)->input_values_do(update_use_count); | |
416 // special handling for some instructions | |
417 if ((*n)->as_BlockEnd() != NULL) { | |
418 // note on BlockEnd: | |
419 // must 'use' the stack only if the method doesn't | |
420 // terminate, however, in those cases stack is empty | |
421 (*n)->state_values_do(update_use_count); | |
422 } | |
423 } | |
424 depth--; | |
425 } | |
426 | |
427 static void basic_compute_use_count(BlockBegin* b) { | |
428 depth = 0; | |
429 // process all pinned nodes as the roots of expression trees | |
430 for (Instruction* n = b; n != NULL; n = n->next()) { | |
431 if (n->is_pinned()) uses_do(&n); | |
432 } | |
433 assert(depth == 0, "should have counted back down"); | |
434 | |
435 // now process any unpinned nodes which recursed too deeply | |
436 while (worklist->length() > 0) { | |
437 Value t = worklist->pop(); | |
438 if (!t->is_pinned()) { | |
439 // compute the use count | |
440 uses_do(&t); | |
441 | |
442 // pin the instruction so that LIRGenerator doesn't recurse | |
443 // too deeply during it's evaluation. | |
444 t->pin(); | |
445 } | |
446 } | |
447 assert(depth == 0, "should have counted back down"); | |
448 } | |
449 | |
450 public: | |
451 static void compute(BlockList* blocks) { | |
452 worklist = new Values(); | |
453 blocks->blocks_do(basic_compute_use_count); | |
454 worklist = NULL; | |
455 } | |
456 }; | |
457 | |
458 | |
459 Values* UseCountComputer::worklist = NULL; | |
460 int UseCountComputer::depth = 0; | |
461 | |
462 // helper macro for short definition of trace-output inside code | |
463 #ifndef PRODUCT | |
464 #define TRACE_LINEAR_SCAN(level, code) \ | |
465 if (TraceLinearScanLevel >= level) { \ | |
466 code; \ | |
467 } | |
468 #else | |
469 #define TRACE_LINEAR_SCAN(level, code) | |
470 #endif | |
471 | |
472 class ComputeLinearScanOrder : public StackObj { | |
473 private: | |
474 int _max_block_id; // the highest block_id of a block | |
475 int _num_blocks; // total number of blocks (smaller than _max_block_id) | |
476 int _num_loops; // total number of loops | |
477 bool _iterative_dominators;// method requires iterative computation of dominatiors | |
478 | |
479 BlockList* _linear_scan_order; // the resulting list of blocks in correct order | |
480 | |
481 BitMap _visited_blocks; // used for recursive processing of blocks | |
482 BitMap _active_blocks; // used for recursive processing of blocks | |
483 BitMap _dominator_blocks; // temproary BitMap used for computation of dominator | |
484 intArray _forward_branches; // number of incoming forward branches for each block | |
485 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges | |
486 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop | |
487 BlockList _work_list; // temporary list (used in mark_loops and compute_order) | |
488 | |
489 // accessors for _visited_blocks and _active_blocks | |
490 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); } | |
491 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); } | |
492 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); } | |
493 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); } | |
494 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); } | |
495 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); } | |
496 | |
497 // accessors for _forward_branches | |
498 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); } | |
499 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 | |
501 // accessors for _loop_map | |
502 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); } | |
503 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); } | |
504 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); } | |
505 | |
506 // count edges between blocks | |
507 void count_edges(BlockBegin* cur, BlockBegin* parent); | |
508 | |
509 // loop detection | |
510 void mark_loops(); | |
511 void clear_non_natural_loops(BlockBegin* start_block); | |
512 void assign_loop_depth(BlockBegin* start_block); | |
513 | |
514 // computation of final block order | |
515 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b); | |
516 void compute_dominator(BlockBegin* cur, BlockBegin* parent); | |
517 int compute_weight(BlockBegin* cur); | |
518 bool ready_for_processing(BlockBegin* cur); | |
519 void sort_into_work_list(BlockBegin* b); | |
520 void append_block(BlockBegin* cur); | |
521 void compute_order(BlockBegin* start_block); | |
522 | |
523 // fixup of dominators for non-natural loops | |
524 bool compute_dominators_iter(); | |
525 void compute_dominators(); | |
526 | |
527 // debug functions | |
528 NOT_PRODUCT(void print_blocks();) | |
529 DEBUG_ONLY(void verify();) | |
530 | |
531 public: | |
532 ComputeLinearScanOrder(BlockBegin* start_block); | |
533 | |
534 // accessors for final result | |
535 BlockList* linear_scan_order() const { return _linear_scan_order; } | |
536 int num_loops() const { return _num_loops; } | |
537 }; | |
538 | |
539 | |
540 ComputeLinearScanOrder::ComputeLinearScanOrder(BlockBegin* start_block) : | |
541 _max_block_id(BlockBegin::number_of_blocks()), | |
542 _num_blocks(0), | |
543 _num_loops(0), | |
544 _iterative_dominators(false), | |
545 _visited_blocks(_max_block_id), | |
546 _active_blocks(_max_block_id), | |
547 _dominator_blocks(_max_block_id), | |
548 _forward_branches(_max_block_id, 0), | |
549 _loop_end_blocks(8), | |
550 _work_list(8), | |
551 _linear_scan_order(NULL), // initialized later with correct size | |
552 _loop_map(0, 0) // initialized later with correct size | |
553 { | |
554 TRACE_LINEAR_SCAN(2, "***** computing linear-scan block order"); | |
555 | |
556 init_visited(); | |
557 count_edges(start_block, NULL); | |
558 | |
559 if (_num_loops > 0) { | |
560 mark_loops(); | |
561 clear_non_natural_loops(start_block); | |
562 assign_loop_depth(start_block); | |
563 } | |
564 | |
565 compute_order(start_block); | |
566 compute_dominators(); | |
567 | |
568 NOT_PRODUCT(print_blocks()); | |
569 DEBUG_ONLY(verify()); | |
570 } | |
571 | |
572 | |
573 // Traverse the CFG: | |
574 // * count total number of blocks | |
575 // * count all incoming edges and backward incoming edges | |
576 // * number loop header blocks | |
577 // * create a list with all loop end blocks | |
578 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) { | |
579 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 assert(cur->dominator() == NULL, "dominator already initialized"); | |
581 | |
582 if (is_active(cur)) { | |
583 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch")); | |
584 assert(is_visited(cur), "block must be visisted when block is active"); | |
585 assert(parent != NULL, "must have parent"); | |
586 | |
587 cur->set(BlockBegin::linear_scan_loop_header_flag); | |
588 cur->set(BlockBegin::backward_branch_target_flag); | |
589 | |
590 parent->set(BlockBegin::linear_scan_loop_end_flag); | |
428
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
591 |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
592 // When a loop header is also the start of an exception handler, then the backward branch is |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
593 // an exception edge. Because such edges are usually critical edges which cannot be split, the |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
594 // loop must be excluded here from processing. |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
595 if (cur->is_set(BlockBegin::exception_entry_flag)) { |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
596 // Make sure that dominators are correct in this weird situation |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
597 _iterative_dominators = true; |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
598 return; |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
599 } |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
600 assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur, |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
601 "loop end blocks must have one successor (critical edges are split)"); |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
602 |
0 | 603 _loop_end_blocks.append(parent); |
604 return; | |
605 } | |
606 | |
607 // increment number of incoming forward branches | |
608 inc_forward_branches(cur); | |
609 | |
610 if (is_visited(cur)) { | |
611 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited")); | |
612 return; | |
613 } | |
614 | |
615 _num_blocks++; | |
616 set_visited(cur); | |
617 set_active(cur); | |
618 | |
619 // recursive call for all successors | |
620 int i; | |
621 for (i = cur->number_of_sux() - 1; i >= 0; i--) { | |
622 count_edges(cur->sux_at(i), cur); | |
623 } | |
624 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { | |
625 count_edges(cur->exception_handler_at(i), cur); | |
626 } | |
627 | |
628 clear_active(cur); | |
629 | |
630 // Each loop has a unique number. | |
631 // When multiple loops are nested, assign_loop_depth assumes that the | |
632 // innermost loop has the lowest number. This is guaranteed by setting | |
633 // the loop number after the recursive calls for the successors above | |
634 // have returned. | |
635 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
636 assert(cur->loop_index() == -1, "cannot set loop-index twice"); | |
637 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops)); | |
638 | |
639 cur->set_loop_index(_num_loops); | |
640 _num_loops++; | |
641 } | |
642 | |
643 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id())); | |
644 } | |
645 | |
646 | |
647 void ComputeLinearScanOrder::mark_loops() { | |
648 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops")); | |
649 | |
650 _loop_map = BitMap2D(_num_loops, _max_block_id); | |
651 _loop_map.clear(); | |
652 | |
653 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) { | |
654 BlockBegin* loop_end = _loop_end_blocks.at(i); | |
655 BlockBegin* loop_start = loop_end->sux_at(0); | |
656 int loop_idx = loop_start->loop_index(); | |
657 | |
658 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 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set"); | |
660 assert(loop_end->number_of_sux() == 1, "incorrect number of successors"); | |
661 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set"); | |
662 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set"); | |
663 assert(_work_list.is_empty(), "work list must be empty before processing"); | |
664 | |
665 // add the end-block of the loop to the working list | |
666 _work_list.push(loop_end); | |
667 set_block_in_loop(loop_idx, loop_end); | |
668 do { | |
669 BlockBegin* cur = _work_list.pop(); | |
670 | |
671 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id())); | |
672 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list"); | |
673 | |
674 // recursive processing of all predecessors ends when start block of loop is reached | |
675 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) { | |
676 for (int j = cur->number_of_preds() - 1; j >= 0; j--) { | |
677 BlockBegin* pred = cur->pred_at(j); | |
678 | |
679 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) { | |
680 // this predecessor has not been processed yet, so add it to work list | |
681 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id())); | |
682 _work_list.push(pred); | |
683 set_block_in_loop(loop_idx, pred); | |
684 } | |
685 } | |
686 } | |
687 } while (!_work_list.is_empty()); | |
688 } | |
689 } | |
690 | |
691 | |
692 // check for non-natural loops (loops where the loop header does not dominate | |
693 // all other loop blocks = loops with mulitple entries). | |
694 // such loops are ignored | |
695 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) { | |
696 for (int i = _num_loops - 1; i >= 0; i--) { | |
697 if (is_block_in_loop(i, start_block)) { | |
698 // loop i contains the entry block of the method | |
699 // -> this is not a natural loop, so ignore it | |
700 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i)); | |
701 | |
702 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) { | |
703 clear_block_in_loop(i, block_id); | |
704 } | |
705 _iterative_dominators = true; | |
706 } | |
707 } | |
708 } | |
709 | |
710 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) { | |
711 TRACE_LINEAR_SCAN(3, "----- computing loop-depth and weight"); | |
712 init_visited(); | |
713 | |
714 assert(_work_list.is_empty(), "work list must be empty before processing"); | |
715 _work_list.append(start_block); | |
716 | |
717 do { | |
718 BlockBegin* cur = _work_list.pop(); | |
719 | |
720 if (!is_visited(cur)) { | |
721 set_visited(cur); | |
722 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id())); | |
723 | |
724 // compute loop-depth and loop-index for the block | |
725 assert(cur->loop_depth() == 0, "cannot set loop-depth twice"); | |
726 int i; | |
727 int loop_depth = 0; | |
728 int min_loop_idx = -1; | |
729 for (i = _num_loops - 1; i >= 0; i--) { | |
730 if (is_block_in_loop(i, cur)) { | |
731 loop_depth++; | |
732 min_loop_idx = i; | |
733 } | |
734 } | |
735 cur->set_loop_depth(loop_depth); | |
736 cur->set_loop_index(min_loop_idx); | |
737 | |
738 // append all unvisited successors to work list | |
739 for (i = cur->number_of_sux() - 1; i >= 0; i--) { | |
740 _work_list.append(cur->sux_at(i)); | |
741 } | |
742 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { | |
743 _work_list.append(cur->exception_handler_at(i)); | |
744 } | |
745 } | |
746 } while (!_work_list.is_empty()); | |
747 } | |
748 | |
749 | |
750 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) { | |
751 assert(a != NULL && b != NULL, "must have input blocks"); | |
752 | |
753 _dominator_blocks.clear(); | |
754 while (a != NULL) { | |
755 _dominator_blocks.set_bit(a->block_id()); | |
756 assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized"); | |
757 a = a->dominator(); | |
758 } | |
759 while (b != NULL && !_dominator_blocks.at(b->block_id())) { | |
760 assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized"); | |
761 b = b->dominator(); | |
762 } | |
763 | |
764 assert(b != NULL, "could not find dominator"); | |
765 return b; | |
766 } | |
767 | |
768 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) { | |
769 if (cur->dominator() == NULL) { | |
770 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id())); | |
771 cur->set_dominator(parent); | |
772 | |
773 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) { | |
774 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 assert(cur->number_of_preds() > 1, ""); | |
776 cur->set_dominator(common_dominator(cur->dominator(), parent)); | |
777 } | |
778 } | |
779 | |
780 | |
781 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) { | |
782 BlockBegin* single_sux = NULL; | |
783 if (cur->number_of_sux() == 1) { | |
784 single_sux = cur->sux_at(0); | |
785 } | |
786 | |
787 // limit loop-depth to 15 bit (only for security reason, it will never be so big) | |
788 int weight = (cur->loop_depth() & 0x7FFF) << 16; | |
789 | |
790 // general macro for short definition of weight flags | |
791 // the first instance of INC_WEIGHT_IF has the highest priority | |
792 int cur_bit = 15; | |
793 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--; | |
794 | |
795 // this is necessery for the (very rare) case that two successing blocks have | |
796 // the same loop depth, but a different loop index (can happen for endless loops | |
797 // with exception handlers) | |
798 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag)); | |
799 | |
800 // loop end blocks (blocks that end with a backward branch) are added | |
801 // after all other blocks of the loop. | |
802 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag)); | |
803 | |
804 // critical edge split blocks are prefered because than they have a bigger | |
805 // proability to be completely empty | |
806 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag)); | |
807 | |
808 // exceptions should not be thrown in normal control flow, so these blocks | |
809 // are added as late as possible | |
810 INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL)); | |
811 INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL)); | |
812 | |
813 // exceptions handlers are added as late as possible | |
814 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag)); | |
815 | |
816 // guarantee that weight is > 0 | |
817 weight |= 1; | |
818 | |
819 #undef INC_WEIGHT_IF | |
820 assert(cur_bit >= 0, "too many flags"); | |
821 assert(weight > 0, "weight cannot become negative"); | |
822 | |
823 return weight; | |
824 } | |
825 | |
826 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) { | |
827 // Discount the edge just traveled. | |
828 // When the number drops to zero, all forward branches were processed | |
829 if (dec_forward_branches(cur) != 0) { | |
830 return false; | |
831 } | |
832 | |
833 assert(_linear_scan_order->index_of(cur) == -1, "block already processed (block can be ready only once)"); | |
834 assert(_work_list.index_of(cur) == -1, "block already in work-list (block can be ready only once)"); | |
835 return true; | |
836 } | |
837 | |
838 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) { | |
839 assert(_work_list.index_of(cur) == -1, "block already in work list"); | |
840 | |
841 int cur_weight = compute_weight(cur); | |
842 | |
843 // the linear_scan_number is used to cache the weight of a block | |
844 cur->set_linear_scan_number(cur_weight); | |
845 | |
846 #ifndef PRODUCT | |
847 if (StressLinearScan) { | |
848 _work_list.insert_before(0, cur); | |
849 return; | |
850 } | |
851 #endif | |
852 | |
853 _work_list.append(NULL); // provide space for new element | |
854 | |
855 int insert_idx = _work_list.length() - 1; | |
856 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) { | |
857 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1)); | |
858 insert_idx--; | |
859 } | |
860 _work_list.at_put(insert_idx, cur); | |
861 | |
862 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id())); | |
863 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 | |
865 #ifdef ASSERT | |
866 for (int i = 0; i < _work_list.length(); i++) { | |
867 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set"); | |
868 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist"); | |
869 } | |
870 #endif | |
871 } | |
872 | |
873 void ComputeLinearScanOrder::append_block(BlockBegin* cur) { | |
874 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 assert(_linear_scan_order->index_of(cur) == -1, "cannot add the same block twice"); | |
876 | |
877 // currently, the linear scan order and code emit order are equal. | |
878 // therefore the linear_scan_number and the weight of a block must also | |
879 // be equal. | |
880 cur->set_linear_scan_number(_linear_scan_order->length()); | |
881 _linear_scan_order->append(cur); | |
882 } | |
883 | |
884 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) { | |
885 TRACE_LINEAR_SCAN(3, "----- computing final block order"); | |
886 | |
887 // the start block is always the first block in the linear scan order | |
888 _linear_scan_order = new BlockList(_num_blocks); | |
889 append_block(start_block); | |
890 | |
891 assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction"); | |
892 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry(); | |
893 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry(); | |
894 | |
895 BlockBegin* sux_of_osr_entry = NULL; | |
896 if (osr_entry != NULL) { | |
897 // special handling for osr entry: | |
898 // ignore the edge between the osr entry and its successor for processing | |
899 // the osr entry block is added manually below | |
900 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor"); | |
901 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 | |
903 sux_of_osr_entry = osr_entry->sux_at(0); | |
904 dec_forward_branches(sux_of_osr_entry); | |
905 | |
906 compute_dominator(osr_entry, start_block); | |
907 _iterative_dominators = true; | |
908 } | |
909 compute_dominator(std_entry, start_block); | |
910 | |
911 // start processing with standard entry block | |
912 assert(_work_list.is_empty(), "list must be empty before processing"); | |
913 | |
914 if (ready_for_processing(std_entry)) { | |
915 sort_into_work_list(std_entry); | |
916 } else { | |
917 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)"); | |
918 } | |
919 | |
920 do { | |
921 BlockBegin* cur = _work_list.pop(); | |
922 | |
923 if (cur == sux_of_osr_entry) { | |
924 // the osr entry block is ignored in normal processing, it is never added to the | |
925 // work list. Instead, it is added as late as possible manually here. | |
926 append_block(osr_entry); | |
927 compute_dominator(cur, osr_entry); | |
928 } | |
929 append_block(cur); | |
930 | |
931 int i; | |
932 int num_sux = cur->number_of_sux(); | |
933 // changed loop order to get "intuitive" order of if- and else-blocks | |
934 for (i = 0; i < num_sux; i++) { | |
935 BlockBegin* sux = cur->sux_at(i); | |
936 compute_dominator(sux, cur); | |
937 if (ready_for_processing(sux)) { | |
938 sort_into_work_list(sux); | |
939 } | |
940 } | |
941 num_sux = cur->number_of_exception_handlers(); | |
942 for (i = 0; i < num_sux; i++) { | |
943 BlockBegin* sux = cur->exception_handler_at(i); | |
944 compute_dominator(sux, cur); | |
945 if (ready_for_processing(sux)) { | |
946 sort_into_work_list(sux); | |
947 } | |
948 } | |
949 } while (_work_list.length() > 0); | |
950 } | |
951 | |
952 | |
953 bool ComputeLinearScanOrder::compute_dominators_iter() { | |
954 bool changed = false; | |
955 int num_blocks = _linear_scan_order->length(); | |
956 | |
957 assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator"); | |
958 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors"); | |
959 for (int i = 1; i < num_blocks; i++) { | |
960 BlockBegin* block = _linear_scan_order->at(i); | |
961 | |
962 BlockBegin* dominator = block->pred_at(0); | |
963 int num_preds = block->number_of_preds(); | |
964 for (int i = 1; i < num_preds; i++) { | |
965 dominator = common_dominator(dominator, block->pred_at(i)); | |
966 } | |
967 | |
968 if (dominator != block->dominator()) { | |
969 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 | |
971 block->set_dominator(dominator); | |
972 changed = true; | |
973 } | |
974 } | |
975 return changed; | |
976 } | |
977 | |
978 void ComputeLinearScanOrder::compute_dominators() { | |
979 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators)); | |
980 | |
981 // iterative computation of dominators is only required for methods with non-natural loops | |
982 // and OSR-methods. For all other methods, the dominators computed when generating the | |
983 // linear scan block order are correct. | |
984 if (_iterative_dominators) { | |
985 do { | |
986 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation")); | |
987 } while (compute_dominators_iter()); | |
988 } | |
989 | |
990 // check that dominators are correct | |
991 assert(!compute_dominators_iter(), "fix point not reached"); | |
992 } | |
993 | |
994 | |
995 #ifndef PRODUCT | |
996 void ComputeLinearScanOrder::print_blocks() { | |
997 if (TraceLinearScanLevel >= 2) { | |
998 tty->print_cr("----- loop information:"); | |
999 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { | |
1000 BlockBegin* cur = _linear_scan_order->at(block_idx); | |
1001 | |
1002 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id()); | |
1003 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { | |
1004 tty->print ("%d ", is_block_in_loop(loop_idx, cur)); | |
1005 } | |
1006 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth()); | |
1007 } | |
1008 } | |
1009 | |
1010 if (TraceLinearScanLevel >= 1) { | |
1011 tty->print_cr("----- linear-scan block order:"); | |
1012 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { | |
1013 BlockBegin* cur = _linear_scan_order->at(block_idx); | |
1014 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth()); | |
1015 | |
1016 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " "); | |
1017 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " "); | |
1018 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " "); | |
1019 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " "); | |
1020 | |
1021 if (cur->dominator() != NULL) { | |
1022 tty->print(" dom: B%d ", cur->dominator()->block_id()); | |
1023 } else { | |
1024 tty->print(" dom: NULL "); | |
1025 } | |
1026 | |
1027 if (cur->number_of_preds() > 0) { | |
1028 tty->print(" preds: "); | |
1029 for (int j = 0; j < cur->number_of_preds(); j++) { | |
1030 BlockBegin* pred = cur->pred_at(j); | |
1031 tty->print("B%d ", pred->block_id()); | |
1032 } | |
1033 } | |
1034 if (cur->number_of_sux() > 0) { | |
1035 tty->print(" sux: "); | |
1036 for (int j = 0; j < cur->number_of_sux(); j++) { | |
1037 BlockBegin* sux = cur->sux_at(j); | |
1038 tty->print("B%d ", sux->block_id()); | |
1039 } | |
1040 } | |
1041 if (cur->number_of_exception_handlers() > 0) { | |
1042 tty->print(" ex: "); | |
1043 for (int j = 0; j < cur->number_of_exception_handlers(); j++) { | |
1044 BlockBegin* ex = cur->exception_handler_at(j); | |
1045 tty->print("B%d ", ex->block_id()); | |
1046 } | |
1047 } | |
1048 tty->cr(); | |
1049 } | |
1050 } | |
1051 } | |
1052 #endif | |
1053 | |
1054 #ifdef ASSERT | |
1055 void ComputeLinearScanOrder::verify() { | |
1056 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list"); | |
1057 | |
1058 if (StressLinearScan) { | |
1059 // blocks are scrambled when StressLinearScan is used | |
1060 return; | |
1061 } | |
1062 | |
1063 // check that all successors of a block have a higher linear-scan-number | |
1064 // and that all predecessors of a block have a lower linear-scan-number | |
1065 // (only backward branches of loops are ignored) | |
1066 int i; | |
1067 for (i = 0; i < _linear_scan_order->length(); i++) { | |
1068 BlockBegin* cur = _linear_scan_order->at(i); | |
1069 | |
1070 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number"); | |
1071 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->index_of(cur), "incorrect linear_scan_number"); | |
1072 | |
1073 int j; | |
1074 for (j = cur->number_of_sux() - 1; j >= 0; j--) { | |
1075 BlockBegin* sux = cur->sux_at(j); | |
1076 | |
1077 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->index_of(sux), "incorrect linear_scan_number"); | |
1078 if (!cur->is_set(BlockBegin::linear_scan_loop_end_flag)) { | |
1079 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order"); | |
1080 } | |
1081 if (cur->loop_depth() == sux->loop_depth()) { | |
1082 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 } | |
1084 } | |
1085 | |
1086 for (j = cur->number_of_preds() - 1; j >= 0; j--) { | |
1087 BlockBegin* pred = cur->pred_at(j); | |
1088 | |
1089 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->index_of(pred), "incorrect linear_scan_number"); | |
1090 if (!cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
1091 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order"); | |
1092 } | |
1093 if (cur->loop_depth() == pred->loop_depth()) { | |
1094 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 } | |
1096 | |
1097 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors"); | |
1098 } | |
1099 | |
1100 // check dominator | |
1101 if (i == 0) { | |
1102 assert(cur->dominator() == NULL, "first block has no dominator"); | |
1103 } else { | |
1104 assert(cur->dominator() != NULL, "all but first block must have dominator"); | |
1105 } | |
1106 assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0), "Single predecessor must also be dominator"); | |
1107 } | |
1108 | |
1109 // check that all loops are continuous | |
1110 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { | |
1111 int block_idx = 0; | |
1112 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop"); | |
1113 | |
1114 // skip blocks before the loop | |
1115 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { | |
1116 block_idx++; | |
1117 } | |
1118 // skip blocks of loop | |
1119 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { | |
1120 block_idx++; | |
1121 } | |
1122 // after the first non-loop block, there must not be another loop-block | |
1123 while (block_idx < _num_blocks) { | |
1124 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order"); | |
1125 block_idx++; | |
1126 } | |
1127 } | |
1128 } | |
1129 #endif | |
1130 | |
1131 | |
1132 void IR::compute_code() { | |
1133 assert(is_valid(), "IR must be valid"); | |
1134 | |
1135 ComputeLinearScanOrder compute_order(start()); | |
1136 _num_loops = compute_order.num_loops(); | |
1137 _code = compute_order.linear_scan_order(); | |
1138 } | |
1139 | |
1140 | |
1141 void IR::compute_use_counts() { | |
1142 // make sure all values coming out of this block get evaluated. | |
1143 int num_blocks = _code->length(); | |
1144 for (int i = 0; i < num_blocks; i++) { | |
1145 _code->at(i)->end()->state()->pin_stack_for_linear_scan(); | |
1146 } | |
1147 | |
1148 // compute use counts | |
1149 UseCountComputer::compute(_code); | |
1150 } | |
1151 | |
1152 | |
1153 void IR::iterate_preorder(BlockClosure* closure) { | |
1154 assert(is_valid(), "IR must be valid"); | |
1155 start()->iterate_preorder(closure); | |
1156 } | |
1157 | |
1158 | |
1159 void IR::iterate_postorder(BlockClosure* closure) { | |
1160 assert(is_valid(), "IR must be valid"); | |
1161 start()->iterate_postorder(closure); | |
1162 } | |
1163 | |
1164 void IR::iterate_linear_scan_order(BlockClosure* closure) { | |
1165 linear_scan_order()->iterate_forward(closure); | |
1166 } | |
1167 | |
1168 | |
1169 #ifndef PRODUCT | |
1170 class BlockPrinter: public BlockClosure { | |
1171 private: | |
1172 InstructionPrinter* _ip; | |
1173 bool _cfg_only; | |
1174 bool _live_only; | |
1175 | |
1176 public: | |
1177 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) { | |
1178 _ip = ip; | |
1179 _cfg_only = cfg_only; | |
1180 _live_only = live_only; | |
1181 } | |
1182 | |
1183 virtual void block_do(BlockBegin* block) { | |
1184 if (_cfg_only) { | |
1185 _ip->print_instr(block); tty->cr(); | |
1186 } else { | |
1187 block->print_block(*_ip, _live_only); | |
1188 } | |
1189 } | |
1190 }; | |
1191 | |
1192 | |
1193 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) { | |
1194 ttyLocker ttyl; | |
1195 InstructionPrinter ip(!cfg_only); | |
1196 BlockPrinter bp(&ip, cfg_only, live_only); | |
1197 start->iterate_preorder(&bp); | |
1198 tty->cr(); | |
1199 } | |
1200 | |
1201 void IR::print(bool cfg_only, bool live_only) { | |
1202 if (is_valid()) { | |
1203 print(start(), cfg_only, live_only); | |
1204 } else { | |
1205 tty->print_cr("invalid IR"); | |
1206 } | |
1207 } | |
1208 | |
1209 | |
1210 define_array(BlockListArray, BlockList*) | |
1211 define_stack(BlockListList, BlockListArray) | |
1212 | |
1213 class PredecessorValidator : public BlockClosure { | |
1214 private: | |
1215 BlockListList* _predecessors; | |
1216 BlockList* _blocks; | |
1217 | |
1218 static int cmp(BlockBegin** a, BlockBegin** b) { | |
1219 return (*a)->block_id() - (*b)->block_id(); | |
1220 } | |
1221 | |
1222 public: | |
1223 PredecessorValidator(IR* hir) { | |
1224 ResourceMark rm; | |
1225 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), NULL); | |
1226 _blocks = new BlockList(); | |
1227 | |
1228 int i; | |
1229 hir->start()->iterate_preorder(this); | |
1230 if (hir->code() != NULL) { | |
1231 assert(hir->code()->length() == _blocks->length(), "must match"); | |
1232 for (i = 0; i < _blocks->length(); i++) { | |
1233 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists"); | |
1234 } | |
1235 } | |
1236 | |
1237 for (i = 0; i < _blocks->length(); i++) { | |
1238 BlockBegin* block = _blocks->at(i); | |
1239 BlockList* preds = _predecessors->at(block->block_id()); | |
1240 if (preds == NULL) { | |
1241 assert(block->number_of_preds() == 0, "should be the same"); | |
1242 continue; | |
1243 } | |
1244 | |
1245 // clone the pred list so we can mutate it | |
1246 BlockList* pred_copy = new BlockList(); | |
1247 int j; | |
1248 for (j = 0; j < block->number_of_preds(); j++) { | |
1249 pred_copy->append(block->pred_at(j)); | |
1250 } | |
1251 // sort them in the same order | |
1252 preds->sort(cmp); | |
1253 pred_copy->sort(cmp); | |
1254 int length = MIN2(preds->length(), block->number_of_preds()); | |
1255 for (j = 0; j < block->number_of_preds(); j++) { | |
1256 assert(preds->at(j) == pred_copy->at(j), "must match"); | |
1257 } | |
1258 | |
1259 assert(preds->length() == block->number_of_preds(), "should be the same"); | |
1260 } | |
1261 } | |
1262 | |
1263 virtual void block_do(BlockBegin* block) { | |
1264 _blocks->append(block); | |
1265 BlockEnd* be = block->end(); | |
1266 int n = be->number_of_sux(); | |
1267 int i; | |
1268 for (i = 0; i < n; i++) { | |
1269 BlockBegin* sux = be->sux_at(i); | |
1270 assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler"); | |
1271 | |
1272 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); | |
1273 if (preds == NULL) { | |
1274 preds = new BlockList(); | |
1275 _predecessors->at_put(sux->block_id(), preds); | |
1276 } | |
1277 preds->append(block); | |
1278 } | |
1279 | |
1280 n = block->number_of_exception_handlers(); | |
1281 for (i = 0; i < n; i++) { | |
1282 BlockBegin* sux = block->exception_handler_at(i); | |
1283 assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler"); | |
1284 | |
1285 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); | |
1286 if (preds == NULL) { | |
1287 preds = new BlockList(); | |
1288 _predecessors->at_put(sux->block_id(), preds); | |
1289 } | |
1290 preds->append(block); | |
1291 } | |
1292 } | |
1293 }; | |
1294 | |
1295 void IR::verify() { | |
1296 #ifdef ASSERT | |
1297 PredecessorValidator pv(this); | |
1298 #endif | |
1299 } | |
1300 | |
1301 #endif // PRODUCT | |
1302 | |
1303 void SubstitutionResolver::substitute(Value* v) { | |
1304 Value v0 = *v; | |
1305 if (v0) { | |
1306 Value vs = v0->subst(); | |
1307 if (vs != v0) { | |
1308 *v = v0->subst(); | |
1309 } | |
1310 } | |
1311 } | |
1312 | |
1313 #ifdef ASSERT | |
1314 void check_substitute(Value* v) { | |
1315 Value v0 = *v; | |
1316 if (v0) { | |
1317 Value vs = v0->subst(); | |
1318 assert(vs == v0, "missed substitution"); | |
1319 } | |
1320 } | |
1321 #endif | |
1322 | |
1323 | |
1324 void SubstitutionResolver::block_do(BlockBegin* block) { | |
1325 Instruction* last = NULL; | |
1326 for (Instruction* n = block; n != NULL;) { | |
1327 n->values_do(substitute); | |
1328 // need to remove this instruction from the instruction stream | |
1329 if (n->subst() != n) { | |
1330 assert(last != NULL, "must have last"); | |
1331 last->set_next(n->next(), n->next()->bci()); | |
1332 } else { | |
1333 last = n; | |
1334 } | |
1335 n = last->next(); | |
1336 } | |
1337 | |
1338 #ifdef ASSERT | |
1339 if (block->state()) block->state()->values_do(check_substitute); | |
1340 block->block_values_do(check_substitute); | |
1341 if (block->end() && block->end()->state()) block->end()->state()->values_do(check_substitute); | |
1342 #endif | |
1343 } |