Mercurial > hg > truffle
annotate src/share/vm/c1/c1_IR.cpp @ 1657:1a1ce2076047
Merge
author | ysr |
---|---|
date | Fri, 16 Jul 2010 10:09:15 -0700 |
parents | b812ff5abc73 |
children | d5d065957597 |
rev | line source |
---|---|
0 | 1 /* |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
2 * Copyright (c) 1999, 2010, Oracle and/or its affiliates. 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
21 * questions. |
0 | 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) | |
1564 | 233 , _id(-1) |
234 , _is_method_handle_invoke(false) { | |
0 | 235 assert(_stack != NULL, "must be non null"); |
236 assert(_bci == SynchronizationEntryBCI || Bytecodes::is_defined(scope()->method()->java_code_at_bci(_bci)), "make sure bci points at a real bytecode"); | |
237 } | |
238 | |
239 | |
240 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, bool lock_stack_only) | |
241 : _scope(info->_scope) | |
242 , _exception_handlers(NULL) | |
243 , _bci(info->_bci) | |
244 , _scope_debug_info(NULL) | |
1564 | 245 , _oop_map(NULL) |
246 , _is_method_handle_invoke(info->_is_method_handle_invoke) { | |
0 | 247 if (lock_stack_only) { |
248 if (info->_stack != NULL) { | |
249 _stack = info->_stack->copy_locks(); | |
250 } else { | |
251 _stack = NULL; | |
252 } | |
253 } else { | |
254 _stack = info->_stack; | |
255 } | |
256 | |
257 // deep copy of exception handlers | |
258 if (info->_exception_handlers != NULL) { | |
259 _exception_handlers = new XHandlers(info->_exception_handlers); | |
260 } | |
261 } | |
262 | |
263 | |
1564 | 264 void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) { |
0 | 265 // record the safepoint before recording the debug info for enclosing scopes |
266 recorder->add_safepoint(pc_offset, _oop_map->deep_copy()); | |
1564 | 267 _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, _is_method_handle_invoke); |
0 | 268 recorder->end_safepoint(pc_offset); |
269 } | |
270 | |
271 | |
272 void CodeEmitInfo::add_register_oop(LIR_Opr opr) { | |
273 assert(_oop_map != NULL, "oop map must already exist"); | |
274 assert(opr->is_single_cpu(), "should not call otherwise"); | |
275 | |
276 int frame_size = frame_map()->framesize(); | |
277 int arg_count = frame_map()->oop_map_arg_count(); | |
278 VMReg name = frame_map()->regname(opr); | |
279 _oop_map->set_oop(name); | |
280 } | |
281 | |
282 | |
283 | |
284 | |
285 // Implementation of IR | |
286 | |
287 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) : | |
288 _locals_size(in_WordSize(-1)) | |
289 , _num_loops(0) { | |
290 // setup IR fields | |
291 _compilation = compilation; | |
292 _top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true); | |
293 _code = NULL; | |
294 } | |
295 | |
296 | |
297 void IR::optimize() { | |
298 Optimizer opt(this); | |
299 if (DoCEE) { | |
300 opt.eliminate_conditional_expressions(); | |
301 #ifndef PRODUCT | |
302 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); } | |
303 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); } | |
304 #endif | |
305 } | |
306 if (EliminateBlocks) { | |
307 opt.eliminate_blocks(); | |
308 #ifndef PRODUCT | |
309 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); } | |
310 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); } | |
311 #endif | |
312 } | |
313 if (EliminateNullChecks) { | |
314 opt.eliminate_null_checks(); | |
315 #ifndef PRODUCT | |
316 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); } | |
317 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); } | |
318 #endif | |
319 } | |
320 } | |
321 | |
322 | |
323 static int sort_pairs(BlockPair** a, BlockPair** b) { | |
324 if ((*a)->from() == (*b)->from()) { | |
325 return (*a)->to()->block_id() - (*b)->to()->block_id(); | |
326 } else { | |
327 return (*a)->from()->block_id() - (*b)->from()->block_id(); | |
328 } | |
329 } | |
330 | |
331 | |
332 class CriticalEdgeFinder: public BlockClosure { | |
333 BlockPairList blocks; | |
334 IR* _ir; | |
335 | |
336 public: | |
337 CriticalEdgeFinder(IR* ir): _ir(ir) {} | |
338 void block_do(BlockBegin* bb) { | |
339 BlockEnd* be = bb->end(); | |
340 int nos = be->number_of_sux(); | |
341 if (nos >= 2) { | |
342 for (int i = 0; i < nos; i++) { | |
343 BlockBegin* sux = be->sux_at(i); | |
344 if (sux->number_of_preds() >= 2) { | |
345 blocks.append(new BlockPair(bb, sux)); | |
346 } | |
347 } | |
348 } | |
349 } | |
350 | |
351 void split_edges() { | |
352 BlockPair* last_pair = NULL; | |
353 blocks.sort(sort_pairs); | |
354 for (int i = 0; i < blocks.length(); i++) { | |
355 BlockPair* pair = blocks.at(i); | |
356 if (last_pair != NULL && pair->is_same(last_pair)) continue; | |
357 BlockBegin* from = pair->from(); | |
358 BlockBegin* to = pair->to(); | |
359 BlockBegin* split = from->insert_block_between(to); | |
360 #ifndef PRODUCT | |
361 if ((PrintIR || PrintIR1) && Verbose) { | |
362 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)", | |
363 from->block_id(), to->block_id(), split->block_id()); | |
364 } | |
365 #endif | |
366 last_pair = pair; | |
367 } | |
368 } | |
369 }; | |
370 | |
371 void IR::split_critical_edges() { | |
372 CriticalEdgeFinder cef(this); | |
373 | |
374 iterate_preorder(&cef); | |
375 cef.split_edges(); | |
376 } | |
377 | |
378 | |
1584 | 379 class UseCountComputer: public ValueVisitor, BlockClosure { |
0 | 380 private: |
1584 | 381 void visit(Value* n) { |
0 | 382 // Local instructions and Phis for expression stack values at the |
383 // start of basic blocks are not added to the instruction list | |
384 if ((*n)->bci() == -99 && (*n)->as_Local() == NULL && | |
385 (*n)->as_Phi() == NULL) { | |
386 assert(false, "a node was not appended to the graph"); | |
1584 | 387 Compilation::current()->bailout("a node was not appended to the graph"); |
0 | 388 } |
389 // use n's input if not visited before | |
390 if (!(*n)->is_pinned() && !(*n)->has_uses()) { | |
391 // note: a) if the instruction is pinned, it will be handled by compute_use_count | |
392 // b) if the instruction has uses, it was touched before | |
393 // => in both cases we don't need to update n's values | |
394 uses_do(n); | |
395 } | |
396 // use n | |
397 (*n)->_use_count++; | |
398 } | |
399 | |
1584 | 400 Values* worklist; |
401 int depth; | |
0 | 402 enum { |
403 max_recurse_depth = 20 | |
404 }; | |
405 | |
1584 | 406 void uses_do(Value* n) { |
0 | 407 depth++; |
408 if (depth > max_recurse_depth) { | |
409 // don't allow the traversal to recurse too deeply | |
410 worklist->push(*n); | |
411 } else { | |
1584 | 412 (*n)->input_values_do(this); |
0 | 413 // special handling for some instructions |
414 if ((*n)->as_BlockEnd() != NULL) { | |
415 // note on BlockEnd: | |
416 // must 'use' the stack only if the method doesn't | |
417 // terminate, however, in those cases stack is empty | |
1584 | 418 (*n)->state_values_do(this); |
0 | 419 } |
420 } | |
421 depth--; | |
422 } | |
423 | |
1584 | 424 void block_do(BlockBegin* b) { |
0 | 425 depth = 0; |
426 // process all pinned nodes as the roots of expression trees | |
427 for (Instruction* n = b; n != NULL; n = n->next()) { | |
428 if (n->is_pinned()) uses_do(&n); | |
429 } | |
430 assert(depth == 0, "should have counted back down"); | |
431 | |
432 // now process any unpinned nodes which recursed too deeply | |
433 while (worklist->length() > 0) { | |
434 Value t = worklist->pop(); | |
435 if (!t->is_pinned()) { | |
436 // compute the use count | |
437 uses_do(&t); | |
438 | |
439 // pin the instruction so that LIRGenerator doesn't recurse | |
440 // too deeply during it's evaluation. | |
441 t->pin(); | |
442 } | |
443 } | |
444 assert(depth == 0, "should have counted back down"); | |
445 } | |
446 | |
1584 | 447 UseCountComputer() { |
448 worklist = new Values(); | |
449 depth = 0; | |
450 } | |
451 | |
0 | 452 public: |
453 static void compute(BlockList* blocks) { | |
1584 | 454 UseCountComputer ucc; |
455 blocks->iterate_backward(&ucc); | |
0 | 456 } |
457 }; | |
458 | |
459 | |
460 // helper macro for short definition of trace-output inside code | |
461 #ifndef PRODUCT | |
462 #define TRACE_LINEAR_SCAN(level, code) \ | |
463 if (TraceLinearScanLevel >= level) { \ | |
464 code; \ | |
465 } | |
466 #else | |
467 #define TRACE_LINEAR_SCAN(level, code) | |
468 #endif | |
469 | |
470 class ComputeLinearScanOrder : public StackObj { | |
471 private: | |
472 int _max_block_id; // the highest block_id of a block | |
473 int _num_blocks; // total number of blocks (smaller than _max_block_id) | |
474 int _num_loops; // total number of loops | |
475 bool _iterative_dominators;// method requires iterative computation of dominatiors | |
476 | |
477 BlockList* _linear_scan_order; // the resulting list of blocks in correct order | |
478 | |
479 BitMap _visited_blocks; // used for recursive processing of blocks | |
480 BitMap _active_blocks; // used for recursive processing of blocks | |
481 BitMap _dominator_blocks; // temproary BitMap used for computation of dominator | |
482 intArray _forward_branches; // number of incoming forward branches for each block | |
483 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges | |
484 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop | |
485 BlockList _work_list; // temporary list (used in mark_loops and compute_order) | |
486 | |
487 // accessors for _visited_blocks and _active_blocks | |
488 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); } | |
489 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); } | |
490 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); } | |
491 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); } | |
492 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); } | |
493 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); } | |
494 | |
495 // accessors for _forward_branches | |
496 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); } | |
497 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()); } | |
498 | |
499 // accessors for _loop_map | |
500 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); } | |
501 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); } | |
502 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); } | |
503 | |
504 // count edges between blocks | |
505 void count_edges(BlockBegin* cur, BlockBegin* parent); | |
506 | |
507 // loop detection | |
508 void mark_loops(); | |
509 void clear_non_natural_loops(BlockBegin* start_block); | |
510 void assign_loop_depth(BlockBegin* start_block); | |
511 | |
512 // computation of final block order | |
513 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b); | |
514 void compute_dominator(BlockBegin* cur, BlockBegin* parent); | |
515 int compute_weight(BlockBegin* cur); | |
516 bool ready_for_processing(BlockBegin* cur); | |
517 void sort_into_work_list(BlockBegin* b); | |
518 void append_block(BlockBegin* cur); | |
519 void compute_order(BlockBegin* start_block); | |
520 | |
521 // fixup of dominators for non-natural loops | |
522 bool compute_dominators_iter(); | |
523 void compute_dominators(); | |
524 | |
525 // debug functions | |
526 NOT_PRODUCT(void print_blocks();) | |
527 DEBUG_ONLY(void verify();) | |
528 | |
529 public: | |
530 ComputeLinearScanOrder(BlockBegin* start_block); | |
531 | |
532 // accessors for final result | |
533 BlockList* linear_scan_order() const { return _linear_scan_order; } | |
534 int num_loops() const { return _num_loops; } | |
535 }; | |
536 | |
537 | |
538 ComputeLinearScanOrder::ComputeLinearScanOrder(BlockBegin* start_block) : | |
539 _max_block_id(BlockBegin::number_of_blocks()), | |
540 _num_blocks(0), | |
541 _num_loops(0), | |
542 _iterative_dominators(false), | |
543 _visited_blocks(_max_block_id), | |
544 _active_blocks(_max_block_id), | |
545 _dominator_blocks(_max_block_id), | |
546 _forward_branches(_max_block_id, 0), | |
547 _loop_end_blocks(8), | |
548 _work_list(8), | |
549 _linear_scan_order(NULL), // initialized later with correct size | |
550 _loop_map(0, 0) // initialized later with correct size | |
551 { | |
552 TRACE_LINEAR_SCAN(2, "***** computing linear-scan block order"); | |
553 | |
554 init_visited(); | |
555 count_edges(start_block, NULL); | |
556 | |
557 if (_num_loops > 0) { | |
558 mark_loops(); | |
559 clear_non_natural_loops(start_block); | |
560 assign_loop_depth(start_block); | |
561 } | |
562 | |
563 compute_order(start_block); | |
564 compute_dominators(); | |
565 | |
566 NOT_PRODUCT(print_blocks()); | |
567 DEBUG_ONLY(verify()); | |
568 } | |
569 | |
570 | |
571 // Traverse the CFG: | |
572 // * count total number of blocks | |
573 // * count all incoming edges and backward incoming edges | |
574 // * number loop header blocks | |
575 // * create a list with all loop end blocks | |
576 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) { | |
577 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)); | |
578 assert(cur->dominator() == NULL, "dominator already initialized"); | |
579 | |
580 if (is_active(cur)) { | |
581 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch")); | |
582 assert(is_visited(cur), "block must be visisted when block is active"); | |
583 assert(parent != NULL, "must have parent"); | |
584 | |
585 cur->set(BlockBegin::linear_scan_loop_header_flag); | |
586 cur->set(BlockBegin::backward_branch_target_flag); | |
587 | |
588 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
|
589 |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
590 // 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
|
591 // 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
|
592 // 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
|
593 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
|
594 // 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
|
595 _iterative_dominators = true; |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
596 return; |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
597 } |
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
598 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
|
599 "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
|
600 |
0 | 601 _loop_end_blocks.append(parent); |
602 return; | |
603 } | |
604 | |
605 // increment number of incoming forward branches | |
606 inc_forward_branches(cur); | |
607 | |
608 if (is_visited(cur)) { | |
609 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited")); | |
610 return; | |
611 } | |
612 | |
613 _num_blocks++; | |
614 set_visited(cur); | |
615 set_active(cur); | |
616 | |
617 // recursive call for all successors | |
618 int i; | |
619 for (i = cur->number_of_sux() - 1; i >= 0; i--) { | |
620 count_edges(cur->sux_at(i), cur); | |
621 } | |
622 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { | |
623 count_edges(cur->exception_handler_at(i), cur); | |
624 } | |
625 | |
626 clear_active(cur); | |
627 | |
628 // Each loop has a unique number. | |
629 // When multiple loops are nested, assign_loop_depth assumes that the | |
630 // innermost loop has the lowest number. This is guaranteed by setting | |
631 // the loop number after the recursive calls for the successors above | |
632 // have returned. | |
633 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
634 assert(cur->loop_index() == -1, "cannot set loop-index twice"); | |
635 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops)); | |
636 | |
637 cur->set_loop_index(_num_loops); | |
638 _num_loops++; | |
639 } | |
640 | |
641 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id())); | |
642 } | |
643 | |
644 | |
645 void ComputeLinearScanOrder::mark_loops() { | |
646 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops")); | |
647 | |
648 _loop_map = BitMap2D(_num_loops, _max_block_id); | |
649 _loop_map.clear(); | |
650 | |
651 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) { | |
652 BlockBegin* loop_end = _loop_end_blocks.at(i); | |
653 BlockBegin* loop_start = loop_end->sux_at(0); | |
654 int loop_idx = loop_start->loop_index(); | |
655 | |
656 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)); | |
657 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set"); | |
658 assert(loop_end->number_of_sux() == 1, "incorrect number of successors"); | |
659 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set"); | |
660 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set"); | |
661 assert(_work_list.is_empty(), "work list must be empty before processing"); | |
662 | |
663 // add the end-block of the loop to the working list | |
664 _work_list.push(loop_end); | |
665 set_block_in_loop(loop_idx, loop_end); | |
666 do { | |
667 BlockBegin* cur = _work_list.pop(); | |
668 | |
669 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id())); | |
670 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list"); | |
671 | |
672 // recursive processing of all predecessors ends when start block of loop is reached | |
673 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) { | |
674 for (int j = cur->number_of_preds() - 1; j >= 0; j--) { | |
675 BlockBegin* pred = cur->pred_at(j); | |
676 | |
677 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) { | |
678 // this predecessor has not been processed yet, so add it to work list | |
679 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id())); | |
680 _work_list.push(pred); | |
681 set_block_in_loop(loop_idx, pred); | |
682 } | |
683 } | |
684 } | |
685 } while (!_work_list.is_empty()); | |
686 } | |
687 } | |
688 | |
689 | |
690 // check for non-natural loops (loops where the loop header does not dominate | |
691 // all other loop blocks = loops with mulitple entries). | |
692 // such loops are ignored | |
693 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) { | |
694 for (int i = _num_loops - 1; i >= 0; i--) { | |
695 if (is_block_in_loop(i, start_block)) { | |
696 // loop i contains the entry block of the method | |
697 // -> this is not a natural loop, so ignore it | |
698 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i)); | |
699 | |
700 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) { | |
701 clear_block_in_loop(i, block_id); | |
702 } | |
703 _iterative_dominators = true; | |
704 } | |
705 } | |
706 } | |
707 | |
708 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) { | |
709 TRACE_LINEAR_SCAN(3, "----- computing loop-depth and weight"); | |
710 init_visited(); | |
711 | |
712 assert(_work_list.is_empty(), "work list must be empty before processing"); | |
713 _work_list.append(start_block); | |
714 | |
715 do { | |
716 BlockBegin* cur = _work_list.pop(); | |
717 | |
718 if (!is_visited(cur)) { | |
719 set_visited(cur); | |
720 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id())); | |
721 | |
722 // compute loop-depth and loop-index for the block | |
723 assert(cur->loop_depth() == 0, "cannot set loop-depth twice"); | |
724 int i; | |
725 int loop_depth = 0; | |
726 int min_loop_idx = -1; | |
727 for (i = _num_loops - 1; i >= 0; i--) { | |
728 if (is_block_in_loop(i, cur)) { | |
729 loop_depth++; | |
730 min_loop_idx = i; | |
731 } | |
732 } | |
733 cur->set_loop_depth(loop_depth); | |
734 cur->set_loop_index(min_loop_idx); | |
735 | |
736 // append all unvisited successors to work list | |
737 for (i = cur->number_of_sux() - 1; i >= 0; i--) { | |
738 _work_list.append(cur->sux_at(i)); | |
739 } | |
740 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { | |
741 _work_list.append(cur->exception_handler_at(i)); | |
742 } | |
743 } | |
744 } while (!_work_list.is_empty()); | |
745 } | |
746 | |
747 | |
748 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) { | |
749 assert(a != NULL && b != NULL, "must have input blocks"); | |
750 | |
751 _dominator_blocks.clear(); | |
752 while (a != NULL) { | |
753 _dominator_blocks.set_bit(a->block_id()); | |
754 assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized"); | |
755 a = a->dominator(); | |
756 } | |
757 while (b != NULL && !_dominator_blocks.at(b->block_id())) { | |
758 assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized"); | |
759 b = b->dominator(); | |
760 } | |
761 | |
762 assert(b != NULL, "could not find dominator"); | |
763 return b; | |
764 } | |
765 | |
766 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) { | |
767 if (cur->dominator() == NULL) { | |
768 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id())); | |
769 cur->set_dominator(parent); | |
770 | |
771 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) { | |
772 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())); | |
773 assert(cur->number_of_preds() > 1, ""); | |
774 cur->set_dominator(common_dominator(cur->dominator(), parent)); | |
775 } | |
776 } | |
777 | |
778 | |
779 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) { | |
780 BlockBegin* single_sux = NULL; | |
781 if (cur->number_of_sux() == 1) { | |
782 single_sux = cur->sux_at(0); | |
783 } | |
784 | |
785 // limit loop-depth to 15 bit (only for security reason, it will never be so big) | |
786 int weight = (cur->loop_depth() & 0x7FFF) << 16; | |
787 | |
788 // general macro for short definition of weight flags | |
789 // the first instance of INC_WEIGHT_IF has the highest priority | |
790 int cur_bit = 15; | |
791 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--; | |
792 | |
793 // this is necessery for the (very rare) case that two successing blocks have | |
794 // the same loop depth, but a different loop index (can happen for endless loops | |
795 // with exception handlers) | |
796 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag)); | |
797 | |
798 // loop end blocks (blocks that end with a backward branch) are added | |
799 // after all other blocks of the loop. | |
800 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag)); | |
801 | |
802 // critical edge split blocks are prefered because than they have a bigger | |
803 // proability to be completely empty | |
804 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag)); | |
805 | |
806 // exceptions should not be thrown in normal control flow, so these blocks | |
807 // are added as late as possible | |
808 INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL)); | |
809 INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL)); | |
810 | |
811 // exceptions handlers are added as late as possible | |
812 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag)); | |
813 | |
814 // guarantee that weight is > 0 | |
815 weight |= 1; | |
816 | |
817 #undef INC_WEIGHT_IF | |
818 assert(cur_bit >= 0, "too many flags"); | |
819 assert(weight > 0, "weight cannot become negative"); | |
820 | |
821 return weight; | |
822 } | |
823 | |
824 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) { | |
825 // Discount the edge just traveled. | |
826 // When the number drops to zero, all forward branches were processed | |
827 if (dec_forward_branches(cur) != 0) { | |
828 return false; | |
829 } | |
830 | |
831 assert(_linear_scan_order->index_of(cur) == -1, "block already processed (block can be ready only once)"); | |
832 assert(_work_list.index_of(cur) == -1, "block already in work-list (block can be ready only once)"); | |
833 return true; | |
834 } | |
835 | |
836 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) { | |
837 assert(_work_list.index_of(cur) == -1, "block already in work list"); | |
838 | |
839 int cur_weight = compute_weight(cur); | |
840 | |
841 // the linear_scan_number is used to cache the weight of a block | |
842 cur->set_linear_scan_number(cur_weight); | |
843 | |
844 #ifndef PRODUCT | |
845 if (StressLinearScan) { | |
846 _work_list.insert_before(0, cur); | |
847 return; | |
848 } | |
849 #endif | |
850 | |
851 _work_list.append(NULL); // provide space for new element | |
852 | |
853 int insert_idx = _work_list.length() - 1; | |
854 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) { | |
855 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1)); | |
856 insert_idx--; | |
857 } | |
858 _work_list.at_put(insert_idx, cur); | |
859 | |
860 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id())); | |
861 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())); | |
862 | |
863 #ifdef ASSERT | |
864 for (int i = 0; i < _work_list.length(); i++) { | |
865 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set"); | |
866 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist"); | |
867 } | |
868 #endif | |
869 } | |
870 | |
871 void ComputeLinearScanOrder::append_block(BlockBegin* cur) { | |
872 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())); | |
873 assert(_linear_scan_order->index_of(cur) == -1, "cannot add the same block twice"); | |
874 | |
875 // currently, the linear scan order and code emit order are equal. | |
876 // therefore the linear_scan_number and the weight of a block must also | |
877 // be equal. | |
878 cur->set_linear_scan_number(_linear_scan_order->length()); | |
879 _linear_scan_order->append(cur); | |
880 } | |
881 | |
882 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) { | |
883 TRACE_LINEAR_SCAN(3, "----- computing final block order"); | |
884 | |
885 // the start block is always the first block in the linear scan order | |
886 _linear_scan_order = new BlockList(_num_blocks); | |
887 append_block(start_block); | |
888 | |
889 assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction"); | |
890 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry(); | |
891 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry(); | |
892 | |
893 BlockBegin* sux_of_osr_entry = NULL; | |
894 if (osr_entry != NULL) { | |
895 // special handling for osr entry: | |
896 // ignore the edge between the osr entry and its successor for processing | |
897 // the osr entry block is added manually below | |
898 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor"); | |
899 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"); | |
900 | |
901 sux_of_osr_entry = osr_entry->sux_at(0); | |
902 dec_forward_branches(sux_of_osr_entry); | |
903 | |
904 compute_dominator(osr_entry, start_block); | |
905 _iterative_dominators = true; | |
906 } | |
907 compute_dominator(std_entry, start_block); | |
908 | |
909 // start processing with standard entry block | |
910 assert(_work_list.is_empty(), "list must be empty before processing"); | |
911 | |
912 if (ready_for_processing(std_entry)) { | |
913 sort_into_work_list(std_entry); | |
914 } else { | |
915 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)"); | |
916 } | |
917 | |
918 do { | |
919 BlockBegin* cur = _work_list.pop(); | |
920 | |
921 if (cur == sux_of_osr_entry) { | |
922 // the osr entry block is ignored in normal processing, it is never added to the | |
923 // work list. Instead, it is added as late as possible manually here. | |
924 append_block(osr_entry); | |
925 compute_dominator(cur, osr_entry); | |
926 } | |
927 append_block(cur); | |
928 | |
929 int i; | |
930 int num_sux = cur->number_of_sux(); | |
931 // changed loop order to get "intuitive" order of if- and else-blocks | |
932 for (i = 0; i < num_sux; i++) { | |
933 BlockBegin* sux = cur->sux_at(i); | |
934 compute_dominator(sux, cur); | |
935 if (ready_for_processing(sux)) { | |
936 sort_into_work_list(sux); | |
937 } | |
938 } | |
939 num_sux = cur->number_of_exception_handlers(); | |
940 for (i = 0; i < num_sux; i++) { | |
941 BlockBegin* sux = cur->exception_handler_at(i); | |
942 compute_dominator(sux, cur); | |
943 if (ready_for_processing(sux)) { | |
944 sort_into_work_list(sux); | |
945 } | |
946 } | |
947 } while (_work_list.length() > 0); | |
948 } | |
949 | |
950 | |
951 bool ComputeLinearScanOrder::compute_dominators_iter() { | |
952 bool changed = false; | |
953 int num_blocks = _linear_scan_order->length(); | |
954 | |
955 assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator"); | |
956 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors"); | |
957 for (int i = 1; i < num_blocks; i++) { | |
958 BlockBegin* block = _linear_scan_order->at(i); | |
959 | |
960 BlockBegin* dominator = block->pred_at(0); | |
961 int num_preds = block->number_of_preds(); | |
962 for (int i = 1; i < num_preds; i++) { | |
963 dominator = common_dominator(dominator, block->pred_at(i)); | |
964 } | |
965 | |
966 if (dominator != block->dominator()) { | |
967 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())); | |
968 | |
969 block->set_dominator(dominator); | |
970 changed = true; | |
971 } | |
972 } | |
973 return changed; | |
974 } | |
975 | |
976 void ComputeLinearScanOrder::compute_dominators() { | |
977 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators)); | |
978 | |
979 // iterative computation of dominators is only required for methods with non-natural loops | |
980 // and OSR-methods. For all other methods, the dominators computed when generating the | |
981 // linear scan block order are correct. | |
982 if (_iterative_dominators) { | |
983 do { | |
984 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation")); | |
985 } while (compute_dominators_iter()); | |
986 } | |
987 | |
988 // check that dominators are correct | |
989 assert(!compute_dominators_iter(), "fix point not reached"); | |
990 } | |
991 | |
992 | |
993 #ifndef PRODUCT | |
994 void ComputeLinearScanOrder::print_blocks() { | |
995 if (TraceLinearScanLevel >= 2) { | |
996 tty->print_cr("----- loop information:"); | |
997 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { | |
998 BlockBegin* cur = _linear_scan_order->at(block_idx); | |
999 | |
1000 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id()); | |
1001 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { | |
1002 tty->print ("%d ", is_block_in_loop(loop_idx, cur)); | |
1003 } | |
1004 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth()); | |
1005 } | |
1006 } | |
1007 | |
1008 if (TraceLinearScanLevel >= 1) { | |
1009 tty->print_cr("----- linear-scan block order:"); | |
1010 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { | |
1011 BlockBegin* cur = _linear_scan_order->at(block_idx); | |
1012 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth()); | |
1013 | |
1014 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " "); | |
1015 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " "); | |
1016 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " "); | |
1017 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " "); | |
1018 | |
1019 if (cur->dominator() != NULL) { | |
1020 tty->print(" dom: B%d ", cur->dominator()->block_id()); | |
1021 } else { | |
1022 tty->print(" dom: NULL "); | |
1023 } | |
1024 | |
1025 if (cur->number_of_preds() > 0) { | |
1026 tty->print(" preds: "); | |
1027 for (int j = 0; j < cur->number_of_preds(); j++) { | |
1028 BlockBegin* pred = cur->pred_at(j); | |
1029 tty->print("B%d ", pred->block_id()); | |
1030 } | |
1031 } | |
1032 if (cur->number_of_sux() > 0) { | |
1033 tty->print(" sux: "); | |
1034 for (int j = 0; j < cur->number_of_sux(); j++) { | |
1035 BlockBegin* sux = cur->sux_at(j); | |
1036 tty->print("B%d ", sux->block_id()); | |
1037 } | |
1038 } | |
1039 if (cur->number_of_exception_handlers() > 0) { | |
1040 tty->print(" ex: "); | |
1041 for (int j = 0; j < cur->number_of_exception_handlers(); j++) { | |
1042 BlockBegin* ex = cur->exception_handler_at(j); | |
1043 tty->print("B%d ", ex->block_id()); | |
1044 } | |
1045 } | |
1046 tty->cr(); | |
1047 } | |
1048 } | |
1049 } | |
1050 #endif | |
1051 | |
1052 #ifdef ASSERT | |
1053 void ComputeLinearScanOrder::verify() { | |
1054 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list"); | |
1055 | |
1056 if (StressLinearScan) { | |
1057 // blocks are scrambled when StressLinearScan is used | |
1058 return; | |
1059 } | |
1060 | |
1061 // check that all successors of a block have a higher linear-scan-number | |
1062 // and that all predecessors of a block have a lower linear-scan-number | |
1063 // (only backward branches of loops are ignored) | |
1064 int i; | |
1065 for (i = 0; i < _linear_scan_order->length(); i++) { | |
1066 BlockBegin* cur = _linear_scan_order->at(i); | |
1067 | |
1068 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number"); | |
1069 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->index_of(cur), "incorrect linear_scan_number"); | |
1070 | |
1071 int j; | |
1072 for (j = cur->number_of_sux() - 1; j >= 0; j--) { | |
1073 BlockBegin* sux = cur->sux_at(j); | |
1074 | |
1075 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->index_of(sux), "incorrect linear_scan_number"); | |
1076 if (!cur->is_set(BlockBegin::linear_scan_loop_end_flag)) { | |
1077 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order"); | |
1078 } | |
1079 if (cur->loop_depth() == sux->loop_depth()) { | |
1080 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"); | |
1081 } | |
1082 } | |
1083 | |
1084 for (j = cur->number_of_preds() - 1; j >= 0; j--) { | |
1085 BlockBegin* pred = cur->pred_at(j); | |
1086 | |
1087 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->index_of(pred), "incorrect linear_scan_number"); | |
1088 if (!cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
1089 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order"); | |
1090 } | |
1091 if (cur->loop_depth() == pred->loop_depth()) { | |
1092 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"); | |
1093 } | |
1094 | |
1095 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors"); | |
1096 } | |
1097 | |
1098 // check dominator | |
1099 if (i == 0) { | |
1100 assert(cur->dominator() == NULL, "first block has no dominator"); | |
1101 } else { | |
1102 assert(cur->dominator() != NULL, "all but first block must have dominator"); | |
1103 } | |
1104 assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0), "Single predecessor must also be dominator"); | |
1105 } | |
1106 | |
1107 // check that all loops are continuous | |
1108 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { | |
1109 int block_idx = 0; | |
1110 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop"); | |
1111 | |
1112 // skip blocks before the loop | |
1113 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { | |
1114 block_idx++; | |
1115 } | |
1116 // skip blocks of loop | |
1117 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { | |
1118 block_idx++; | |
1119 } | |
1120 // after the first non-loop block, there must not be another loop-block | |
1121 while (block_idx < _num_blocks) { | |
1122 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order"); | |
1123 block_idx++; | |
1124 } | |
1125 } | |
1126 } | |
1127 #endif | |
1128 | |
1129 | |
1130 void IR::compute_code() { | |
1131 assert(is_valid(), "IR must be valid"); | |
1132 | |
1133 ComputeLinearScanOrder compute_order(start()); | |
1134 _num_loops = compute_order.num_loops(); | |
1135 _code = compute_order.linear_scan_order(); | |
1136 } | |
1137 | |
1138 | |
1139 void IR::compute_use_counts() { | |
1140 // make sure all values coming out of this block get evaluated. | |
1141 int num_blocks = _code->length(); | |
1142 for (int i = 0; i < num_blocks; i++) { | |
1143 _code->at(i)->end()->state()->pin_stack_for_linear_scan(); | |
1144 } | |
1145 | |
1146 // compute use counts | |
1147 UseCountComputer::compute(_code); | |
1148 } | |
1149 | |
1150 | |
1151 void IR::iterate_preorder(BlockClosure* closure) { | |
1152 assert(is_valid(), "IR must be valid"); | |
1153 start()->iterate_preorder(closure); | |
1154 } | |
1155 | |
1156 | |
1157 void IR::iterate_postorder(BlockClosure* closure) { | |
1158 assert(is_valid(), "IR must be valid"); | |
1159 start()->iterate_postorder(closure); | |
1160 } | |
1161 | |
1162 void IR::iterate_linear_scan_order(BlockClosure* closure) { | |
1163 linear_scan_order()->iterate_forward(closure); | |
1164 } | |
1165 | |
1166 | |
1167 #ifndef PRODUCT | |
1168 class BlockPrinter: public BlockClosure { | |
1169 private: | |
1170 InstructionPrinter* _ip; | |
1171 bool _cfg_only; | |
1172 bool _live_only; | |
1173 | |
1174 public: | |
1175 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) { | |
1176 _ip = ip; | |
1177 _cfg_only = cfg_only; | |
1178 _live_only = live_only; | |
1179 } | |
1180 | |
1181 virtual void block_do(BlockBegin* block) { | |
1182 if (_cfg_only) { | |
1183 _ip->print_instr(block); tty->cr(); | |
1184 } else { | |
1185 block->print_block(*_ip, _live_only); | |
1186 } | |
1187 } | |
1188 }; | |
1189 | |
1190 | |
1191 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) { | |
1192 ttyLocker ttyl; | |
1193 InstructionPrinter ip(!cfg_only); | |
1194 BlockPrinter bp(&ip, cfg_only, live_only); | |
1195 start->iterate_preorder(&bp); | |
1196 tty->cr(); | |
1197 } | |
1198 | |
1199 void IR::print(bool cfg_only, bool live_only) { | |
1200 if (is_valid()) { | |
1201 print(start(), cfg_only, live_only); | |
1202 } else { | |
1203 tty->print_cr("invalid IR"); | |
1204 } | |
1205 } | |
1206 | |
1207 | |
1208 define_array(BlockListArray, BlockList*) | |
1209 define_stack(BlockListList, BlockListArray) | |
1210 | |
1211 class PredecessorValidator : public BlockClosure { | |
1212 private: | |
1213 BlockListList* _predecessors; | |
1214 BlockList* _blocks; | |
1215 | |
1216 static int cmp(BlockBegin** a, BlockBegin** b) { | |
1217 return (*a)->block_id() - (*b)->block_id(); | |
1218 } | |
1219 | |
1220 public: | |
1221 PredecessorValidator(IR* hir) { | |
1222 ResourceMark rm; | |
1223 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), NULL); | |
1224 _blocks = new BlockList(); | |
1225 | |
1226 int i; | |
1227 hir->start()->iterate_preorder(this); | |
1228 if (hir->code() != NULL) { | |
1229 assert(hir->code()->length() == _blocks->length(), "must match"); | |
1230 for (i = 0; i < _blocks->length(); i++) { | |
1231 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists"); | |
1232 } | |
1233 } | |
1234 | |
1235 for (i = 0; i < _blocks->length(); i++) { | |
1236 BlockBegin* block = _blocks->at(i); | |
1237 BlockList* preds = _predecessors->at(block->block_id()); | |
1238 if (preds == NULL) { | |
1239 assert(block->number_of_preds() == 0, "should be the same"); | |
1240 continue; | |
1241 } | |
1242 | |
1243 // clone the pred list so we can mutate it | |
1244 BlockList* pred_copy = new BlockList(); | |
1245 int j; | |
1246 for (j = 0; j < block->number_of_preds(); j++) { | |
1247 pred_copy->append(block->pred_at(j)); | |
1248 } | |
1249 // sort them in the same order | |
1250 preds->sort(cmp); | |
1251 pred_copy->sort(cmp); | |
1252 int length = MIN2(preds->length(), block->number_of_preds()); | |
1253 for (j = 0; j < block->number_of_preds(); j++) { | |
1254 assert(preds->at(j) == pred_copy->at(j), "must match"); | |
1255 } | |
1256 | |
1257 assert(preds->length() == block->number_of_preds(), "should be the same"); | |
1258 } | |
1259 } | |
1260 | |
1261 virtual void block_do(BlockBegin* block) { | |
1262 _blocks->append(block); | |
1263 BlockEnd* be = block->end(); | |
1264 int n = be->number_of_sux(); | |
1265 int i; | |
1266 for (i = 0; i < n; i++) { | |
1267 BlockBegin* sux = be->sux_at(i); | |
1268 assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler"); | |
1269 | |
1270 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); | |
1271 if (preds == NULL) { | |
1272 preds = new BlockList(); | |
1273 _predecessors->at_put(sux->block_id(), preds); | |
1274 } | |
1275 preds->append(block); | |
1276 } | |
1277 | |
1278 n = block->number_of_exception_handlers(); | |
1279 for (i = 0; i < n; i++) { | |
1280 BlockBegin* sux = block->exception_handler_at(i); | |
1281 assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler"); | |
1282 | |
1283 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); | |
1284 if (preds == NULL) { | |
1285 preds = new BlockList(); | |
1286 _predecessors->at_put(sux->block_id(), preds); | |
1287 } | |
1288 preds->append(block); | |
1289 } | |
1290 } | |
1291 }; | |
1292 | |
1293 void IR::verify() { | |
1294 #ifdef ASSERT | |
1295 PredecessorValidator pv(this); | |
1296 #endif | |
1297 } | |
1298 | |
1299 #endif // PRODUCT | |
1300 | |
1584 | 1301 void SubstitutionResolver::visit(Value* v) { |
0 | 1302 Value v0 = *v; |
1303 if (v0) { | |
1304 Value vs = v0->subst(); | |
1305 if (vs != v0) { | |
1306 *v = v0->subst(); | |
1307 } | |
1308 } | |
1309 } | |
1310 | |
1311 #ifdef ASSERT | |
1584 | 1312 class SubstitutionChecker: public ValueVisitor { |
1313 void visit(Value* v) { | |
1314 Value v0 = *v; | |
1315 if (v0) { | |
1316 Value vs = v0->subst(); | |
1317 assert(vs == v0, "missed substitution"); | |
1318 } | |
0 | 1319 } |
1584 | 1320 }; |
0 | 1321 #endif |
1322 | |
1323 | |
1324 void SubstitutionResolver::block_do(BlockBegin* block) { | |
1325 Instruction* last = NULL; | |
1326 for (Instruction* n = block; n != NULL;) { | |
1584 | 1327 n->values_do(this); |
0 | 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 | |
1584 | 1339 SubstitutionChecker check_substitute; |
1340 if (block->state()) block->state()->values_do(&check_substitute); | |
1341 block->block_values_do(&check_substitute); | |
1342 if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute); | |
0 | 1343 #endif |
1344 } |