1 | /* |
2 | * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. |
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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "c1/c1_Compilation.hpp" |
27 | #include "c1/c1_FrameMap.hpp" |
28 | #include "c1/c1_GraphBuilder.hpp" |
29 | #include "c1/c1_IR.hpp" |
30 | #include "c1/c1_InstructionPrinter.hpp" |
31 | #include "c1/c1_Optimizer.hpp" |
32 | #include "memory/resourceArea.hpp" |
33 | #include "utilities/bitMap.inline.hpp" |
34 | |
35 | |
36 | // Implementation of XHandlers |
37 | // |
38 | // Note: This code could eventually go away if we are |
39 | // just using the ciExceptionHandlerStream. |
40 | |
41 | XHandlers::XHandlers(ciMethod* method) : _list(method->exception_table_length()) { |
42 | ciExceptionHandlerStream s(method); |
43 | while (!s.is_done()) { |
44 | _list.append(new XHandler(s.handler())); |
45 | s.next(); |
46 | } |
47 | assert(s.count() == method->exception_table_length(), "exception table lengths inconsistent" ); |
48 | } |
49 | |
50 | // deep copy of all XHandler contained in list |
51 | XHandlers::XHandlers(XHandlers* other) : |
52 | _list(other->length()) |
53 | { |
54 | for (int i = 0; i < other->length(); i++) { |
55 | _list.append(new XHandler(other->handler_at(i))); |
56 | } |
57 | } |
58 | |
59 | // Returns whether a particular exception type can be caught. Also |
60 | // returns true if klass is unloaded or any exception handler |
61 | // classes are unloaded. type_is_exact indicates whether the throw |
62 | // is known to be exactly that class or it might throw a subtype. |
63 | bool XHandlers::could_catch(ciInstanceKlass* klass, bool type_is_exact) const { |
64 | // the type is unknown so be conservative |
65 | if (!klass->is_loaded()) { |
66 | return true; |
67 | } |
68 | |
69 | for (int i = 0; i < length(); i++) { |
70 | XHandler* handler = handler_at(i); |
71 | if (handler->is_catch_all()) { |
72 | // catch of ANY |
73 | return true; |
74 | } |
75 | ciInstanceKlass* handler_klass = handler->catch_klass(); |
76 | // if it's unknown it might be catchable |
77 | if (!handler_klass->is_loaded()) { |
78 | return true; |
79 | } |
80 | // if the throw type is definitely a subtype of the catch type |
81 | // then it can be caught. |
82 | if (klass->is_subtype_of(handler_klass)) { |
83 | return true; |
84 | } |
85 | if (!type_is_exact) { |
86 | // If the type isn't exactly known then it can also be caught by |
87 | // catch statements where the inexact type is a subtype of the |
88 | // catch type. |
89 | // given: foo extends bar extends Exception |
90 | // throw bar can be caught by catch foo, catch bar, and catch |
91 | // Exception, however it can't be caught by any handlers without |
92 | // bar in its type hierarchy. |
93 | if (handler_klass->is_subtype_of(klass)) { |
94 | return true; |
95 | } |
96 | } |
97 | } |
98 | |
99 | return false; |
100 | } |
101 | |
102 | |
103 | bool XHandlers::equals(XHandlers* others) const { |
104 | if (others == NULL) return false; |
105 | if (length() != others->length()) return false; |
106 | |
107 | for (int i = 0; i < length(); i++) { |
108 | if (!handler_at(i)->equals(others->handler_at(i))) return false; |
109 | } |
110 | return true; |
111 | } |
112 | |
113 | bool XHandler::equals(XHandler* other) const { |
114 | assert(entry_pco() != -1 && other->entry_pco() != -1, "must have entry_pco" ); |
115 | |
116 | if (entry_pco() != other->entry_pco()) return false; |
117 | if (scope_count() != other->scope_count()) return false; |
118 | if (_desc != other->_desc) return false; |
119 | |
120 | assert(entry_block() == other->entry_block(), "entry_block must be equal when entry_pco is equal" ); |
121 | return true; |
122 | } |
123 | |
124 | |
125 | // Implementation of IRScope |
126 | BlockBegin* IRScope::build_graph(Compilation* compilation, int osr_bci) { |
127 | GraphBuilder gm(compilation, this); |
128 | NOT_PRODUCT(if (PrintValueNumbering && Verbose) gm.print_stats()); |
129 | if (compilation->bailed_out()) return NULL; |
130 | return gm.start(); |
131 | } |
132 | |
133 | |
134 | IRScope::IRScope(Compilation* compilation, IRScope* caller, int caller_bci, ciMethod* method, int osr_bci, bool create_graph) |
135 | : _compilation(compilation) |
136 | , _callees(2) |
137 | , _requires_phi_function(method->max_locals()) |
138 | { |
139 | _caller = caller; |
140 | _level = caller == NULL ? 0 : caller->level() + 1; |
141 | _method = method; |
142 | _xhandlers = new XHandlers(method); |
143 | _number_of_locks = 0; |
144 | _monitor_pairing_ok = method->has_balanced_monitors(); |
145 | _wrote_final = false; |
146 | _wrote_fields = false; |
147 | _wrote_volatile = false; |
148 | _start = NULL; |
149 | |
150 | if (osr_bci != -1) { |
151 | // selective creation of phi functions is not possibel in osr-methods |
152 | _requires_phi_function.set_range(0, method->max_locals()); |
153 | } |
154 | |
155 | assert(method->holder()->is_loaded() , "method holder must be loaded" ); |
156 | |
157 | // build graph if monitor pairing is ok |
158 | if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci); |
159 | } |
160 | |
161 | |
162 | int IRScope::max_stack() const { |
163 | int my_max = method()->max_stack(); |
164 | int callee_max = 0; |
165 | for (int i = 0; i < number_of_callees(); i++) { |
166 | callee_max = MAX2(callee_max, callee_no(i)->max_stack()); |
167 | } |
168 | return my_max + callee_max; |
169 | } |
170 | |
171 | |
172 | bool IRScopeDebugInfo::should_reexecute() { |
173 | ciMethod* cur_method = scope()->method(); |
174 | int cur_bci = bci(); |
175 | if (cur_method != NULL && cur_bci != SynchronizationEntryBCI) { |
176 | Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); |
177 | return Interpreter::bytecode_should_reexecute(code); |
178 | } else |
179 | return false; |
180 | } |
181 | |
182 | |
183 | // Implementation of CodeEmitInfo |
184 | |
185 | // Stack must be NON-null |
186 | CodeEmitInfo::CodeEmitInfo(ValueStack* stack, XHandlers* exception_handlers, bool deoptimize_on_exception) |
187 | : _scope_debug_info(NULL) |
188 | , _scope(stack->scope()) |
189 | , _exception_handlers(exception_handlers) |
190 | , _oop_map(NULL) |
191 | , _stack(stack) |
192 | , _is_method_handle_invoke(false) |
193 | , _deoptimize_on_exception(deoptimize_on_exception) { |
194 | assert(_stack != NULL, "must be non null" ); |
195 | } |
196 | |
197 | |
198 | CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, ValueStack* stack) |
199 | : _scope_debug_info(NULL) |
200 | , _scope(info->_scope) |
201 | , _exception_handlers(NULL) |
202 | , _oop_map(NULL) |
203 | , _stack(stack == NULL ? info->_stack : stack) |
204 | , _is_method_handle_invoke(info->_is_method_handle_invoke) |
205 | , _deoptimize_on_exception(info->_deoptimize_on_exception) { |
206 | |
207 | // deep copy of exception handlers |
208 | if (info->_exception_handlers != NULL) { |
209 | _exception_handlers = new XHandlers(info->_exception_handlers); |
210 | } |
211 | } |
212 | |
213 | |
214 | void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) { |
215 | // record the safepoint before recording the debug info for enclosing scopes |
216 | recorder->add_safepoint(pc_offset, _oop_map->deep_copy()); |
217 | _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, _is_method_handle_invoke); |
218 | recorder->end_safepoint(pc_offset); |
219 | } |
220 | |
221 | |
222 | void CodeEmitInfo::add_register_oop(LIR_Opr opr) { |
223 | assert(_oop_map != NULL, "oop map must already exist" ); |
224 | assert(opr->is_single_cpu(), "should not call otherwise" ); |
225 | |
226 | VMReg name = frame_map()->regname(opr); |
227 | _oop_map->set_oop(name); |
228 | } |
229 | |
230 | // Mirror the stack size calculation in the deopt code |
231 | // How much stack space would we need at this point in the program in |
232 | // case of deoptimization? |
233 | int CodeEmitInfo::interpreter_frame_size() const { |
234 | ValueStack* state = _stack; |
235 | int size = 0; |
236 | int callee_parameters = 0; |
237 | int callee_locals = 0; |
238 | int = state->scope()->method()->max_stack() - state->stack_size(); |
239 | |
240 | while (state != NULL) { |
241 | int locks = state->locks_size(); |
242 | int temps = state->stack_size(); |
243 | bool is_top_frame = (state == _stack); |
244 | ciMethod* method = state->scope()->method(); |
245 | |
246 | int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(), |
247 | temps + callee_parameters, |
248 | extra_args, |
249 | locks, |
250 | callee_parameters, |
251 | callee_locals, |
252 | is_top_frame); |
253 | size += frame_size; |
254 | |
255 | callee_parameters = method->size_of_parameters(); |
256 | callee_locals = method->max_locals(); |
257 | extra_args = 0; |
258 | state = state->caller_state(); |
259 | } |
260 | return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord; |
261 | } |
262 | |
263 | // Implementation of IR |
264 | |
265 | IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) : |
266 | _num_loops(0) { |
267 | // setup IR fields |
268 | _compilation = compilation; |
269 | _top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true); |
270 | _code = NULL; |
271 | } |
272 | |
273 | |
274 | void IR::optimize_blocks() { |
275 | Optimizer opt(this); |
276 | if (!compilation()->profile_branches()) { |
277 | if (DoCEE) { |
278 | opt.eliminate_conditional_expressions(); |
279 | #ifndef PRODUCT |
280 | if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE" ); print(true); } |
281 | if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE" ); print(false); } |
282 | #endif |
283 | } |
284 | if (EliminateBlocks) { |
285 | opt.eliminate_blocks(); |
286 | #ifndef PRODUCT |
287 | if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination" ); print(true); } |
288 | if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination" ); print(false); } |
289 | #endif |
290 | } |
291 | } |
292 | } |
293 | |
294 | void IR::eliminate_null_checks() { |
295 | Optimizer opt(this); |
296 | if (EliminateNullChecks) { |
297 | opt.eliminate_null_checks(); |
298 | #ifndef PRODUCT |
299 | if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination" ); print(true); } |
300 | if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination" ); print(false); } |
301 | #endif |
302 | } |
303 | } |
304 | |
305 | |
306 | static int sort_pairs(BlockPair** a, BlockPair** b) { |
307 | if ((*a)->from() == (*b)->from()) { |
308 | return (*a)->to()->block_id() - (*b)->to()->block_id(); |
309 | } else { |
310 | return (*a)->from()->block_id() - (*b)->from()->block_id(); |
311 | } |
312 | } |
313 | |
314 | |
315 | class CriticalEdgeFinder: public BlockClosure { |
316 | BlockPairList blocks; |
317 | IR* _ir; |
318 | |
319 | public: |
320 | CriticalEdgeFinder(IR* ir): _ir(ir) {} |
321 | void block_do(BlockBegin* bb) { |
322 | BlockEnd* be = bb->end(); |
323 | int nos = be->number_of_sux(); |
324 | if (nos >= 2) { |
325 | for (int i = 0; i < nos; i++) { |
326 | BlockBegin* sux = be->sux_at(i); |
327 | if (sux->number_of_preds() >= 2) { |
328 | blocks.append(new BlockPair(bb, sux)); |
329 | } |
330 | } |
331 | } |
332 | } |
333 | |
334 | void split_edges() { |
335 | BlockPair* last_pair = NULL; |
336 | blocks.sort(sort_pairs); |
337 | for (int i = 0; i < blocks.length(); i++) { |
338 | BlockPair* pair = blocks.at(i); |
339 | if (last_pair != NULL && pair->is_same(last_pair)) continue; |
340 | BlockBegin* from = pair->from(); |
341 | BlockBegin* to = pair->to(); |
342 | BlockBegin* split = from->insert_block_between(to); |
343 | #ifndef PRODUCT |
344 | if ((PrintIR || PrintIR1) && Verbose) { |
345 | tty->print_cr("Split critical edge B%d -> B%d (new block B%d)" , |
346 | from->block_id(), to->block_id(), split->block_id()); |
347 | } |
348 | #endif |
349 | last_pair = pair; |
350 | } |
351 | } |
352 | }; |
353 | |
354 | void IR::split_critical_edges() { |
355 | CriticalEdgeFinder cef(this); |
356 | |
357 | iterate_preorder(&cef); |
358 | cef.split_edges(); |
359 | } |
360 | |
361 | |
362 | class UseCountComputer: public ValueVisitor, BlockClosure { |
363 | private: |
364 | void visit(Value* n) { |
365 | // Local instructions and Phis for expression stack values at the |
366 | // start of basic blocks are not added to the instruction list |
367 | if (!(*n)->is_linked() && (*n)->can_be_linked()) { |
368 | assert(false, "a node was not appended to the graph" ); |
369 | Compilation::current()->bailout("a node was not appended to the graph" ); |
370 | } |
371 | // use n's input if not visited before |
372 | if (!(*n)->is_pinned() && !(*n)->has_uses()) { |
373 | // note: a) if the instruction is pinned, it will be handled by compute_use_count |
374 | // b) if the instruction has uses, it was touched before |
375 | // => in both cases we don't need to update n's values |
376 | uses_do(n); |
377 | } |
378 | // use n |
379 | (*n)->_use_count++; |
380 | } |
381 | |
382 | Values* worklist; |
383 | int depth; |
384 | enum { |
385 | max_recurse_depth = 20 |
386 | }; |
387 | |
388 | void uses_do(Value* n) { |
389 | depth++; |
390 | if (depth > max_recurse_depth) { |
391 | // don't allow the traversal to recurse too deeply |
392 | worklist->push(*n); |
393 | } else { |
394 | (*n)->input_values_do(this); |
395 | // special handling for some instructions |
396 | if ((*n)->as_BlockEnd() != NULL) { |
397 | // note on BlockEnd: |
398 | // must 'use' the stack only if the method doesn't |
399 | // terminate, however, in those cases stack is empty |
400 | (*n)->state_values_do(this); |
401 | } |
402 | } |
403 | depth--; |
404 | } |
405 | |
406 | void block_do(BlockBegin* b) { |
407 | depth = 0; |
408 | // process all pinned nodes as the roots of expression trees |
409 | for (Instruction* n = b; n != NULL; n = n->next()) { |
410 | if (n->is_pinned()) uses_do(&n); |
411 | } |
412 | assert(depth == 0, "should have counted back down" ); |
413 | |
414 | // now process any unpinned nodes which recursed too deeply |
415 | while (worklist->length() > 0) { |
416 | Value t = worklist->pop(); |
417 | if (!t->is_pinned()) { |
418 | // compute the use count |
419 | uses_do(&t); |
420 | |
421 | // pin the instruction so that LIRGenerator doesn't recurse |
422 | // too deeply during it's evaluation. |
423 | t->pin(); |
424 | } |
425 | } |
426 | assert(depth == 0, "should have counted back down" ); |
427 | } |
428 | |
429 | UseCountComputer() { |
430 | worklist = new Values(); |
431 | depth = 0; |
432 | } |
433 | |
434 | public: |
435 | static void compute(BlockList* blocks) { |
436 | UseCountComputer ucc; |
437 | blocks->iterate_backward(&ucc); |
438 | } |
439 | }; |
440 | |
441 | |
442 | // helper macro for short definition of trace-output inside code |
443 | #ifndef PRODUCT |
444 | #define TRACE_LINEAR_SCAN(level, code) \ |
445 | if (TraceLinearScanLevel >= level) { \ |
446 | code; \ |
447 | } |
448 | #else |
449 | #define TRACE_LINEAR_SCAN(level, code) |
450 | #endif |
451 | |
452 | class ComputeLinearScanOrder : public StackObj { |
453 | private: |
454 | int _max_block_id; // the highest block_id of a block |
455 | int _num_blocks; // total number of blocks (smaller than _max_block_id) |
456 | int _num_loops; // total number of loops |
457 | bool _iterative_dominators;// method requires iterative computation of dominatiors |
458 | |
459 | BlockList* _linear_scan_order; // the resulting list of blocks in correct order |
460 | |
461 | ResourceBitMap _visited_blocks; // used for recursive processing of blocks |
462 | ResourceBitMap _active_blocks; // used for recursive processing of blocks |
463 | ResourceBitMap _dominator_blocks; // temproary BitMap used for computation of dominator |
464 | intArray _forward_branches; // number of incoming forward branches for each block |
465 | BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges |
466 | BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop |
467 | BlockList _work_list; // temporary list (used in mark_loops and compute_order) |
468 | BlockList ; |
469 | |
470 | Compilation* _compilation; |
471 | |
472 | // accessors for _visited_blocks and _active_blocks |
473 | void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); } |
474 | bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); } |
475 | bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); } |
476 | void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set" ); _visited_blocks.set_bit(b->block_id()); } |
477 | void set_active(BlockBegin* b) { assert(!is_active(b), "already set" ); _active_blocks.set_bit(b->block_id()); } |
478 | void clear_active(BlockBegin* b) { assert(is_active(b), "not already" ); _active_blocks.clear_bit(b->block_id()); } |
479 | |
480 | // accessors for _forward_branches |
481 | void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); } |
482 | 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()); } |
483 | |
484 | // accessors for _loop_map |
485 | bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); } |
486 | void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); } |
487 | void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); } |
488 | |
489 | // count edges between blocks |
490 | void count_edges(BlockBegin* cur, BlockBegin* parent); |
491 | |
492 | // loop detection |
493 | void mark_loops(); |
494 | void clear_non_natural_loops(BlockBegin* start_block); |
495 | void assign_loop_depth(BlockBegin* start_block); |
496 | |
497 | // computation of final block order |
498 | BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b); |
499 | void compute_dominator(BlockBegin* cur, BlockBegin* parent); |
500 | void compute_dominator_impl(BlockBegin* cur, BlockBegin* parent); |
501 | int compute_weight(BlockBegin* cur); |
502 | bool ready_for_processing(BlockBegin* cur); |
503 | void sort_into_work_list(BlockBegin* b); |
504 | void append_block(BlockBegin* cur); |
505 | void compute_order(BlockBegin* start_block); |
506 | |
507 | // fixup of dominators for non-natural loops |
508 | bool compute_dominators_iter(); |
509 | void compute_dominators(); |
510 | |
511 | // debug functions |
512 | NOT_PRODUCT(void print_blocks();) |
513 | DEBUG_ONLY(void verify();) |
514 | |
515 | Compilation* compilation() const { return _compilation; } |
516 | public: |
517 | ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block); |
518 | |
519 | // accessors for final result |
520 | BlockList* linear_scan_order() const { return _linear_scan_order; } |
521 | int num_loops() const { return _num_loops; } |
522 | }; |
523 | |
524 | |
525 | ComputeLinearScanOrder::ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block) : |
526 | _max_block_id(BlockBegin::number_of_blocks()), |
527 | _num_blocks(0), |
528 | _num_loops(0), |
529 | _iterative_dominators(false), |
530 | _linear_scan_order(NULL), // initialized later with correct size |
531 | _visited_blocks(_max_block_id), |
532 | _active_blocks(_max_block_id), |
533 | _dominator_blocks(_max_block_id), |
534 | _forward_branches(_max_block_id, _max_block_id, 0), |
535 | _loop_end_blocks(8), |
536 | _loop_map(0), // initialized later with correct size |
537 | _work_list(8), |
538 | _compilation(c) |
539 | { |
540 | TRACE_LINEAR_SCAN(2, tty->print_cr("***** computing linear-scan block order" )); |
541 | |
542 | count_edges(start_block, NULL); |
543 | |
544 | if (compilation()->is_profiling()) { |
545 | ciMethod *method = compilation()->method(); |
546 | if (!method->is_accessor()) { |
547 | ciMethodData* md = method->method_data_or_null(); |
548 | assert(md != NULL, "Sanity" ); |
549 | md->set_compilation_stats(_num_loops, _num_blocks); |
550 | } |
551 | } |
552 | |
553 | if (_num_loops > 0) { |
554 | mark_loops(); |
555 | clear_non_natural_loops(start_block); |
556 | assign_loop_depth(start_block); |
557 | } |
558 | |
559 | compute_order(start_block); |
560 | compute_dominators(); |
561 | |
562 | NOT_PRODUCT(print_blocks()); |
563 | DEBUG_ONLY(verify()); |
564 | } |
565 | |
566 | |
567 | // Traverse the CFG: |
568 | // * count total number of blocks |
569 | // * count all incoming edges and backward incoming edges |
570 | // * number loop header blocks |
571 | // * create a list with all loop end blocks |
572 | void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) { |
573 | 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)); |
574 | assert(cur->dominator() == NULL, "dominator already initialized" ); |
575 | |
576 | if (is_active(cur)) { |
577 | TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch" )); |
578 | assert(is_visited(cur), "block must be visisted when block is active" ); |
579 | assert(parent != NULL, "must have parent" ); |
580 | |
581 | cur->set(BlockBegin::backward_branch_target_flag); |
582 | |
583 | // When a loop header is also the start of an exception handler, then the backward branch is |
584 | // an exception edge. Because such edges are usually critical edges which cannot be split, the |
585 | // loop must be excluded here from processing. |
586 | if (cur->is_set(BlockBegin::exception_entry_flag)) { |
587 | // Make sure that dominators are correct in this weird situation |
588 | _iterative_dominators = true; |
589 | return; |
590 | } |
591 | |
592 | cur->set(BlockBegin::linear_scan_loop_header_flag); |
593 | parent->set(BlockBegin::linear_scan_loop_end_flag); |
594 | |
595 | assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur, |
596 | "loop end blocks must have one successor (critical edges are split)" ); |
597 | |
598 | _loop_end_blocks.append(parent); |
599 | return; |
600 | } |
601 | |
602 | // increment number of incoming forward branches |
603 | inc_forward_branches(cur); |
604 | |
605 | if (is_visited(cur)) { |
606 | TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited" )); |
607 | return; |
608 | } |
609 | |
610 | _num_blocks++; |
611 | set_visited(cur); |
612 | set_active(cur); |
613 | |
614 | // recursive call for all successors |
615 | int i; |
616 | for (i = cur->number_of_sux() - 1; i >= 0; i--) { |
617 | count_edges(cur->sux_at(i), cur); |
618 | } |
619 | for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { |
620 | count_edges(cur->exception_handler_at(i), cur); |
621 | } |
622 | |
623 | clear_active(cur); |
624 | |
625 | // Each loop has a unique number. |
626 | // When multiple loops are nested, assign_loop_depth assumes that the |
627 | // innermost loop has the lowest number. This is guaranteed by setting |
628 | // the loop number after the recursive calls for the successors above |
629 | // have returned. |
630 | if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { |
631 | assert(cur->loop_index() == -1, "cannot set loop-index twice" ); |
632 | TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d" , cur->block_id(), _num_loops)); |
633 | |
634 | cur->set_loop_index(_num_loops); |
635 | _loop_headers.append(cur); |
636 | _num_loops++; |
637 | } |
638 | |
639 | TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d" , cur->block_id())); |
640 | } |
641 | |
642 | |
643 | void ComputeLinearScanOrder::mark_loops() { |
644 | TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops" )); |
645 | |
646 | _loop_map = BitMap2D(_num_loops, _max_block_id); |
647 | |
648 | for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) { |
649 | BlockBegin* loop_end = _loop_end_blocks.at(i); |
650 | BlockBegin* loop_start = loop_end->sux_at(0); |
651 | int loop_idx = loop_start->loop_index(); |
652 | |
653 | 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)); |
654 | assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set" ); |
655 | assert(loop_end->number_of_sux() == 1, "incorrect number of successors" ); |
656 | assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set" ); |
657 | assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set" ); |
658 | assert(_work_list.is_empty(), "work list must be empty before processing" ); |
659 | |
660 | // add the end-block of the loop to the working list |
661 | _work_list.push(loop_end); |
662 | set_block_in_loop(loop_idx, loop_end); |
663 | do { |
664 | BlockBegin* cur = _work_list.pop(); |
665 | |
666 | TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d" , cur->block_id())); |
667 | assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list" ); |
668 | |
669 | // recursive processing of all predecessors ends when start block of loop is reached |
670 | if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) { |
671 | for (int j = cur->number_of_preds() - 1; j >= 0; j--) { |
672 | BlockBegin* pred = cur->pred_at(j); |
673 | |
674 | if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) { |
675 | // this predecessor has not been processed yet, so add it to work list |
676 | TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d" , pred->block_id())); |
677 | _work_list.push(pred); |
678 | set_block_in_loop(loop_idx, pred); |
679 | } |
680 | } |
681 | } |
682 | } while (!_work_list.is_empty()); |
683 | } |
684 | } |
685 | |
686 | |
687 | // check for non-natural loops (loops where the loop header does not dominate |
688 | // all other loop blocks = loops with mulitple entries). |
689 | // such loops are ignored |
690 | void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) { |
691 | for (int i = _num_loops - 1; i >= 0; i--) { |
692 | if (is_block_in_loop(i, start_block)) { |
693 | // loop i contains the entry block of the method |
694 | // -> this is not a natural loop, so ignore it |
695 | TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored" , i)); |
696 | |
697 | BlockBegin * = _loop_headers.at(i); |
698 | assert(loop_header->is_set(BlockBegin::linear_scan_loop_header_flag), "Must be loop header" ); |
699 | |
700 | for (int j = 0; j < loop_header->number_of_preds(); j++) { |
701 | BlockBegin *pred = loop_header->pred_at(j); |
702 | pred->clear(BlockBegin::linear_scan_loop_end_flag); |
703 | } |
704 | |
705 | loop_header->clear(BlockBegin::linear_scan_loop_header_flag); |
706 | |
707 | for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) { |
708 | clear_block_in_loop(i, block_id); |
709 | } |
710 | _iterative_dominators = true; |
711 | } |
712 | } |
713 | } |
714 | |
715 | void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) { |
716 | TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing loop-depth and weight" )); |
717 | init_visited(); |
718 | |
719 | assert(_work_list.is_empty(), "work list must be empty before processing" ); |
720 | _work_list.append(start_block); |
721 | |
722 | do { |
723 | BlockBegin* cur = _work_list.pop(); |
724 | |
725 | if (!is_visited(cur)) { |
726 | set_visited(cur); |
727 | TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d" , cur->block_id())); |
728 | |
729 | // compute loop-depth and loop-index for the block |
730 | assert(cur->loop_depth() == 0, "cannot set loop-depth twice" ); |
731 | int i; |
732 | int loop_depth = 0; |
733 | int min_loop_idx = -1; |
734 | for (i = _num_loops - 1; i >= 0; i--) { |
735 | if (is_block_in_loop(i, cur)) { |
736 | loop_depth++; |
737 | min_loop_idx = i; |
738 | } |
739 | } |
740 | cur->set_loop_depth(loop_depth); |
741 | cur->set_loop_index(min_loop_idx); |
742 | |
743 | // append all unvisited successors to work list |
744 | for (i = cur->number_of_sux() - 1; i >= 0; i--) { |
745 | _work_list.append(cur->sux_at(i)); |
746 | } |
747 | for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { |
748 | _work_list.append(cur->exception_handler_at(i)); |
749 | } |
750 | } |
751 | } while (!_work_list.is_empty()); |
752 | } |
753 | |
754 | |
755 | BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) { |
756 | assert(a != NULL && b != NULL, "must have input blocks" ); |
757 | |
758 | _dominator_blocks.clear(); |
759 | while (a != NULL) { |
760 | _dominator_blocks.set_bit(a->block_id()); |
761 | assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized" ); |
762 | a = a->dominator(); |
763 | } |
764 | while (b != NULL && !_dominator_blocks.at(b->block_id())) { |
765 | assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized" ); |
766 | b = b->dominator(); |
767 | } |
768 | |
769 | assert(b != NULL, "could not find dominator" ); |
770 | return b; |
771 | } |
772 | |
773 | void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) { |
774 | init_visited(); |
775 | compute_dominator_impl(cur, parent); |
776 | } |
777 | |
778 | void ComputeLinearScanOrder::compute_dominator_impl(BlockBegin* cur, BlockBegin* parent) { |
779 | // Mark as visited to avoid recursive calls with same parent |
780 | set_visited(cur); |
781 | |
782 | if (cur->dominator() == NULL) { |
783 | TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d" , cur->block_id(), parent->block_id())); |
784 | cur->set_dominator(parent); |
785 | |
786 | } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) { |
787 | 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())); |
788 | // Does not hold for exception blocks |
789 | assert(cur->number_of_preds() > 1 || cur->is_set(BlockBegin::exception_entry_flag), "" ); |
790 | cur->set_dominator(common_dominator(cur->dominator(), parent)); |
791 | } |
792 | |
793 | // Additional edge to xhandler of all our successors |
794 | // range check elimination needs that the state at the end of a |
795 | // block be valid in every block it dominates so cur must dominate |
796 | // the exception handlers of its successors. |
797 | int num_cur_xhandler = cur->number_of_exception_handlers(); |
798 | for (int j = 0; j < num_cur_xhandler; j++) { |
799 | BlockBegin* xhandler = cur->exception_handler_at(j); |
800 | if (!is_visited(xhandler)) { |
801 | compute_dominator_impl(xhandler, parent); |
802 | } |
803 | } |
804 | } |
805 | |
806 | |
807 | int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) { |
808 | BlockBegin* single_sux = NULL; |
809 | if (cur->number_of_sux() == 1) { |
810 | single_sux = cur->sux_at(0); |
811 | } |
812 | |
813 | // limit loop-depth to 15 bit (only for security reason, it will never be so big) |
814 | int weight = (cur->loop_depth() & 0x7FFF) << 16; |
815 | |
816 | // general macro for short definition of weight flags |
817 | // the first instance of INC_WEIGHT_IF has the highest priority |
818 | int cur_bit = 15; |
819 | #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--; |
820 | |
821 | // this is necessery for the (very rare) case that two successing blocks have |
822 | // the same loop depth, but a different loop index (can happen for endless loops |
823 | // with exception handlers) |
824 | INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag)); |
825 | |
826 | // loop end blocks (blocks that end with a backward branch) are added |
827 | // after all other blocks of the loop. |
828 | INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag)); |
829 | |
830 | // critical edge split blocks are prefered because than they have a bigger |
831 | // proability to be completely empty |
832 | INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag)); |
833 | |
834 | // exceptions should not be thrown in normal control flow, so these blocks |
835 | // are added as late as possible |
836 | INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL)); |
837 | INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL)); |
838 | |
839 | // exceptions handlers are added as late as possible |
840 | INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag)); |
841 | |
842 | // guarantee that weight is > 0 |
843 | weight |= 1; |
844 | |
845 | #undef INC_WEIGHT_IF |
846 | assert(cur_bit >= 0, "too many flags" ); |
847 | assert(weight > 0, "weight cannot become negative" ); |
848 | |
849 | return weight; |
850 | } |
851 | |
852 | bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) { |
853 | // Discount the edge just traveled. |
854 | // When the number drops to zero, all forward branches were processed |
855 | if (dec_forward_branches(cur) != 0) { |
856 | return false; |
857 | } |
858 | |
859 | assert(_linear_scan_order->find(cur) == -1, "block already processed (block can be ready only once)" ); |
860 | assert(_work_list.find(cur) == -1, "block already in work-list (block can be ready only once)" ); |
861 | return true; |
862 | } |
863 | |
864 | void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) { |
865 | assert(_work_list.find(cur) == -1, "block already in work list" ); |
866 | |
867 | int cur_weight = compute_weight(cur); |
868 | |
869 | // the linear_scan_number is used to cache the weight of a block |
870 | cur->set_linear_scan_number(cur_weight); |
871 | |
872 | #ifndef PRODUCT |
873 | if (StressLinearScan) { |
874 | _work_list.insert_before(0, cur); |
875 | return; |
876 | } |
877 | #endif |
878 | |
879 | _work_list.append(NULL); // provide space for new element |
880 | |
881 | int insert_idx = _work_list.length() - 1; |
882 | while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) { |
883 | _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1)); |
884 | insert_idx--; |
885 | } |
886 | _work_list.at_put(insert_idx, cur); |
887 | |
888 | TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:" , cur->block_id())); |
889 | 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())); |
890 | |
891 | #ifdef ASSERT |
892 | for (int i = 0; i < _work_list.length(); i++) { |
893 | assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set" ); |
894 | assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist" ); |
895 | } |
896 | #endif |
897 | } |
898 | |
899 | void ComputeLinearScanOrder::append_block(BlockBegin* cur) { |
900 | 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())); |
901 | assert(_linear_scan_order->find(cur) == -1, "cannot add the same block twice" ); |
902 | |
903 | // currently, the linear scan order and code emit order are equal. |
904 | // therefore the linear_scan_number and the weight of a block must also |
905 | // be equal. |
906 | cur->set_linear_scan_number(_linear_scan_order->length()); |
907 | _linear_scan_order->append(cur); |
908 | } |
909 | |
910 | void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) { |
911 | TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing final block order" )); |
912 | |
913 | // the start block is always the first block in the linear scan order |
914 | _linear_scan_order = new BlockList(_num_blocks); |
915 | append_block(start_block); |
916 | |
917 | assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction" ); |
918 | BlockBegin* std_entry = ((Base*)start_block->end())->std_entry(); |
919 | BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry(); |
920 | |
921 | BlockBegin* sux_of_osr_entry = NULL; |
922 | if (osr_entry != NULL) { |
923 | // special handling for osr entry: |
924 | // ignore the edge between the osr entry and its successor for processing |
925 | // the osr entry block is added manually below |
926 | assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor" ); |
927 | 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" ); |
928 | |
929 | sux_of_osr_entry = osr_entry->sux_at(0); |
930 | dec_forward_branches(sux_of_osr_entry); |
931 | |
932 | compute_dominator(osr_entry, start_block); |
933 | _iterative_dominators = true; |
934 | } |
935 | compute_dominator(std_entry, start_block); |
936 | |
937 | // start processing with standard entry block |
938 | assert(_work_list.is_empty(), "list must be empty before processing" ); |
939 | |
940 | if (ready_for_processing(std_entry)) { |
941 | sort_into_work_list(std_entry); |
942 | } else { |
943 | assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)" ); |
944 | } |
945 | |
946 | do { |
947 | BlockBegin* cur = _work_list.pop(); |
948 | |
949 | if (cur == sux_of_osr_entry) { |
950 | // the osr entry block is ignored in normal processing, it is never added to the |
951 | // work list. Instead, it is added as late as possible manually here. |
952 | append_block(osr_entry); |
953 | compute_dominator(cur, osr_entry); |
954 | } |
955 | append_block(cur); |
956 | |
957 | int i; |
958 | int num_sux = cur->number_of_sux(); |
959 | // changed loop order to get "intuitive" order of if- and else-blocks |
960 | for (i = 0; i < num_sux; i++) { |
961 | BlockBegin* sux = cur->sux_at(i); |
962 | compute_dominator(sux, cur); |
963 | if (ready_for_processing(sux)) { |
964 | sort_into_work_list(sux); |
965 | } |
966 | } |
967 | num_sux = cur->number_of_exception_handlers(); |
968 | for (i = 0; i < num_sux; i++) { |
969 | BlockBegin* sux = cur->exception_handler_at(i); |
970 | if (ready_for_processing(sux)) { |
971 | sort_into_work_list(sux); |
972 | } |
973 | } |
974 | } while (_work_list.length() > 0); |
975 | } |
976 | |
977 | |
978 | bool ComputeLinearScanOrder::compute_dominators_iter() { |
979 | bool changed = false; |
980 | int num_blocks = _linear_scan_order->length(); |
981 | |
982 | assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator" ); |
983 | assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors" ); |
984 | for (int i = 1; i < num_blocks; i++) { |
985 | BlockBegin* block = _linear_scan_order->at(i); |
986 | |
987 | BlockBegin* dominator = block->pred_at(0); |
988 | int num_preds = block->number_of_preds(); |
989 | |
990 | TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: Processing B%d" , block->block_id())); |
991 | |
992 | for (int j = 0; j < num_preds; j++) { |
993 | |
994 | BlockBegin *pred = block->pred_at(j); |
995 | TRACE_LINEAR_SCAN(4, tty->print_cr(" DOM: Subrocessing B%d" , pred->block_id())); |
996 | |
997 | if (block->is_set(BlockBegin::exception_entry_flag)) { |
998 | dominator = common_dominator(dominator, pred); |
999 | int num_pred_preds = pred->number_of_preds(); |
1000 | for (int k = 0; k < num_pred_preds; k++) { |
1001 | dominator = common_dominator(dominator, pred->pred_at(k)); |
1002 | } |
1003 | } else { |
1004 | dominator = common_dominator(dominator, pred); |
1005 | } |
1006 | } |
1007 | |
1008 | if (dominator != block->dominator()) { |
1009 | 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())); |
1010 | |
1011 | block->set_dominator(dominator); |
1012 | changed = true; |
1013 | } |
1014 | } |
1015 | return changed; |
1016 | } |
1017 | |
1018 | void ComputeLinearScanOrder::compute_dominators() { |
1019 | TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)" , _iterative_dominators)); |
1020 | |
1021 | // iterative computation of dominators is only required for methods with non-natural loops |
1022 | // and OSR-methods. For all other methods, the dominators computed when generating the |
1023 | // linear scan block order are correct. |
1024 | if (_iterative_dominators) { |
1025 | do { |
1026 | TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation" )); |
1027 | } while (compute_dominators_iter()); |
1028 | } |
1029 | |
1030 | // check that dominators are correct |
1031 | assert(!compute_dominators_iter(), "fix point not reached" ); |
1032 | |
1033 | // Add Blocks to dominates-Array |
1034 | int num_blocks = _linear_scan_order->length(); |
1035 | for (int i = 0; i < num_blocks; i++) { |
1036 | BlockBegin* block = _linear_scan_order->at(i); |
1037 | |
1038 | BlockBegin *dom = block->dominator(); |
1039 | if (dom) { |
1040 | assert(dom->dominator_depth() != -1, "Dominator must have been visited before" ); |
1041 | dom->dominates()->append(block); |
1042 | block->set_dominator_depth(dom->dominator_depth() + 1); |
1043 | } else { |
1044 | block->set_dominator_depth(0); |
1045 | } |
1046 | } |
1047 | } |
1048 | |
1049 | |
1050 | #ifndef PRODUCT |
1051 | void ComputeLinearScanOrder::print_blocks() { |
1052 | if (TraceLinearScanLevel >= 2) { |
1053 | tty->print_cr("----- loop information:" ); |
1054 | for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { |
1055 | BlockBegin* cur = _linear_scan_order->at(block_idx); |
1056 | |
1057 | tty->print("%4d: B%2d: " , cur->linear_scan_number(), cur->block_id()); |
1058 | for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { |
1059 | tty->print ("%d " , is_block_in_loop(loop_idx, cur)); |
1060 | } |
1061 | tty->print_cr(" -> loop_index: %2d, loop_depth: %2d" , cur->loop_index(), cur->loop_depth()); |
1062 | } |
1063 | } |
1064 | |
1065 | if (TraceLinearScanLevel >= 1) { |
1066 | tty->print_cr("----- linear-scan block order:" ); |
1067 | for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { |
1068 | BlockBegin* cur = _linear_scan_order->at(block_idx); |
1069 | tty->print("%4d: B%2d loop: %2d depth: %2d" , cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth()); |
1070 | |
1071 | tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " " ); |
1072 | tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " " ); |
1073 | tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " " ); |
1074 | tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " " ); |
1075 | |
1076 | if (cur->dominator() != NULL) { |
1077 | tty->print(" dom: B%d " , cur->dominator()->block_id()); |
1078 | } else { |
1079 | tty->print(" dom: NULL " ); |
1080 | } |
1081 | |
1082 | if (cur->number_of_preds() > 0) { |
1083 | tty->print(" preds: " ); |
1084 | for (int j = 0; j < cur->number_of_preds(); j++) { |
1085 | BlockBegin* pred = cur->pred_at(j); |
1086 | tty->print("B%d " , pred->block_id()); |
1087 | } |
1088 | } |
1089 | if (cur->number_of_sux() > 0) { |
1090 | tty->print(" sux: " ); |
1091 | for (int j = 0; j < cur->number_of_sux(); j++) { |
1092 | BlockBegin* sux = cur->sux_at(j); |
1093 | tty->print("B%d " , sux->block_id()); |
1094 | } |
1095 | } |
1096 | if (cur->number_of_exception_handlers() > 0) { |
1097 | tty->print(" ex: " ); |
1098 | for (int j = 0; j < cur->number_of_exception_handlers(); j++) { |
1099 | BlockBegin* ex = cur->exception_handler_at(j); |
1100 | tty->print("B%d " , ex->block_id()); |
1101 | } |
1102 | } |
1103 | tty->cr(); |
1104 | } |
1105 | } |
1106 | } |
1107 | #endif |
1108 | |
1109 | #ifdef ASSERT |
1110 | void ComputeLinearScanOrder::verify() { |
1111 | assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list" ); |
1112 | |
1113 | if (StressLinearScan) { |
1114 | // blocks are scrambled when StressLinearScan is used |
1115 | return; |
1116 | } |
1117 | |
1118 | // check that all successors of a block have a higher linear-scan-number |
1119 | // and that all predecessors of a block have a lower linear-scan-number |
1120 | // (only backward branches of loops are ignored) |
1121 | int i; |
1122 | for (i = 0; i < _linear_scan_order->length(); i++) { |
1123 | BlockBegin* cur = _linear_scan_order->at(i); |
1124 | |
1125 | assert(cur->linear_scan_number() == i, "incorrect linear_scan_number" ); |
1126 | assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->find(cur), "incorrect linear_scan_number" ); |
1127 | |
1128 | int j; |
1129 | for (j = cur->number_of_sux() - 1; j >= 0; j--) { |
1130 | BlockBegin* sux = cur->sux_at(j); |
1131 | |
1132 | assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->find(sux), "incorrect linear_scan_number" ); |
1133 | if (!sux->is_set(BlockBegin::backward_branch_target_flag)) { |
1134 | assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order" ); |
1135 | } |
1136 | if (cur->loop_depth() == sux->loop_depth()) { |
1137 | 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" ); |
1138 | } |
1139 | } |
1140 | |
1141 | for (j = cur->number_of_preds() - 1; j >= 0; j--) { |
1142 | BlockBegin* pred = cur->pred_at(j); |
1143 | |
1144 | assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->find(pred), "incorrect linear_scan_number" ); |
1145 | if (!cur->is_set(BlockBegin::backward_branch_target_flag)) { |
1146 | assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order" ); |
1147 | } |
1148 | if (cur->loop_depth() == pred->loop_depth()) { |
1149 | 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" ); |
1150 | } |
1151 | |
1152 | assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors" ); |
1153 | } |
1154 | |
1155 | // check dominator |
1156 | if (i == 0) { |
1157 | assert(cur->dominator() == NULL, "first block has no dominator" ); |
1158 | } else { |
1159 | assert(cur->dominator() != NULL, "all but first block must have dominator" ); |
1160 | } |
1161 | // Assertion does not hold for exception handlers |
1162 | assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0) || cur->is_set(BlockBegin::exception_entry_flag), "Single predecessor must also be dominator" ); |
1163 | } |
1164 | |
1165 | // check that all loops are continuous |
1166 | for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { |
1167 | int block_idx = 0; |
1168 | assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop" ); |
1169 | |
1170 | // skip blocks before the loop |
1171 | while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { |
1172 | block_idx++; |
1173 | } |
1174 | // skip blocks of loop |
1175 | while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { |
1176 | block_idx++; |
1177 | } |
1178 | // after the first non-loop block, there must not be another loop-block |
1179 | while (block_idx < _num_blocks) { |
1180 | assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order" ); |
1181 | block_idx++; |
1182 | } |
1183 | } |
1184 | } |
1185 | #endif |
1186 | |
1187 | |
1188 | void IR::compute_code() { |
1189 | assert(is_valid(), "IR must be valid" ); |
1190 | |
1191 | ComputeLinearScanOrder compute_order(compilation(), start()); |
1192 | _num_loops = compute_order.num_loops(); |
1193 | _code = compute_order.linear_scan_order(); |
1194 | } |
1195 | |
1196 | |
1197 | void IR::compute_use_counts() { |
1198 | // make sure all values coming out of this block get evaluated. |
1199 | int num_blocks = _code->length(); |
1200 | for (int i = 0; i < num_blocks; i++) { |
1201 | _code->at(i)->end()->state()->pin_stack_for_linear_scan(); |
1202 | } |
1203 | |
1204 | // compute use counts |
1205 | UseCountComputer::compute(_code); |
1206 | } |
1207 | |
1208 | |
1209 | void IR::iterate_preorder(BlockClosure* closure) { |
1210 | assert(is_valid(), "IR must be valid" ); |
1211 | start()->iterate_preorder(closure); |
1212 | } |
1213 | |
1214 | |
1215 | void IR::iterate_postorder(BlockClosure* closure) { |
1216 | assert(is_valid(), "IR must be valid" ); |
1217 | start()->iterate_postorder(closure); |
1218 | } |
1219 | |
1220 | void IR::iterate_linear_scan_order(BlockClosure* closure) { |
1221 | linear_scan_order()->iterate_forward(closure); |
1222 | } |
1223 | |
1224 | |
1225 | #ifndef PRODUCT |
1226 | class BlockPrinter: public BlockClosure { |
1227 | private: |
1228 | InstructionPrinter* _ip; |
1229 | bool _cfg_only; |
1230 | bool _live_only; |
1231 | |
1232 | public: |
1233 | BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) { |
1234 | _ip = ip; |
1235 | _cfg_only = cfg_only; |
1236 | _live_only = live_only; |
1237 | } |
1238 | |
1239 | virtual void block_do(BlockBegin* block) { |
1240 | if (_cfg_only) { |
1241 | _ip->print_instr(block); tty->cr(); |
1242 | } else { |
1243 | block->print_block(*_ip, _live_only); |
1244 | } |
1245 | } |
1246 | }; |
1247 | |
1248 | |
1249 | void IR::print(BlockBegin* start, bool cfg_only, bool live_only) { |
1250 | ttyLocker ttyl; |
1251 | InstructionPrinter ip(!cfg_only); |
1252 | BlockPrinter bp(&ip, cfg_only, live_only); |
1253 | start->iterate_preorder(&bp); |
1254 | tty->cr(); |
1255 | } |
1256 | |
1257 | void IR::print(bool cfg_only, bool live_only) { |
1258 | if (is_valid()) { |
1259 | print(start(), cfg_only, live_only); |
1260 | } else { |
1261 | tty->print_cr("invalid IR" ); |
1262 | } |
1263 | } |
1264 | |
1265 | |
1266 | typedef GrowableArray<BlockList*> BlockListList; |
1267 | |
1268 | class PredecessorValidator : public BlockClosure { |
1269 | private: |
1270 | BlockListList* _predecessors; |
1271 | BlockList* _blocks; |
1272 | |
1273 | static int cmp(BlockBegin** a, BlockBegin** b) { |
1274 | return (*a)->block_id() - (*b)->block_id(); |
1275 | } |
1276 | |
1277 | public: |
1278 | PredecessorValidator(IR* hir) { |
1279 | ResourceMark rm; |
1280 | _predecessors = new BlockListList(BlockBegin::number_of_blocks(), BlockBegin::number_of_blocks(), NULL); |
1281 | _blocks = new BlockList(); |
1282 | |
1283 | int i; |
1284 | hir->start()->iterate_preorder(this); |
1285 | if (hir->code() != NULL) { |
1286 | assert(hir->code()->length() == _blocks->length(), "must match" ); |
1287 | for (i = 0; i < _blocks->length(); i++) { |
1288 | assert(hir->code()->contains(_blocks->at(i)), "should be in both lists" ); |
1289 | } |
1290 | } |
1291 | |
1292 | for (i = 0; i < _blocks->length(); i++) { |
1293 | BlockBegin* block = _blocks->at(i); |
1294 | BlockList* preds = _predecessors->at(block->block_id()); |
1295 | if (preds == NULL) { |
1296 | assert(block->number_of_preds() == 0, "should be the same" ); |
1297 | continue; |
1298 | } |
1299 | |
1300 | // clone the pred list so we can mutate it |
1301 | BlockList* pred_copy = new BlockList(); |
1302 | int j; |
1303 | for (j = 0; j < block->number_of_preds(); j++) { |
1304 | pred_copy->append(block->pred_at(j)); |
1305 | } |
1306 | // sort them in the same order |
1307 | preds->sort(cmp); |
1308 | pred_copy->sort(cmp); |
1309 | int length = MIN2(preds->length(), block->number_of_preds()); |
1310 | for (j = 0; j < block->number_of_preds(); j++) { |
1311 | assert(preds->at(j) == pred_copy->at(j), "must match" ); |
1312 | } |
1313 | |
1314 | assert(preds->length() == block->number_of_preds(), "should be the same" ); |
1315 | } |
1316 | } |
1317 | |
1318 | virtual void block_do(BlockBegin* block) { |
1319 | _blocks->append(block); |
1320 | BlockEnd* be = block->end(); |
1321 | int n = be->number_of_sux(); |
1322 | int i; |
1323 | for (i = 0; i < n; i++) { |
1324 | BlockBegin* sux = be->sux_at(i); |
1325 | assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler" ); |
1326 | |
1327 | BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); |
1328 | if (preds == NULL) { |
1329 | preds = new BlockList(); |
1330 | _predecessors->at_put(sux->block_id(), preds); |
1331 | } |
1332 | preds->append(block); |
1333 | } |
1334 | |
1335 | n = block->number_of_exception_handlers(); |
1336 | for (i = 0; i < n; i++) { |
1337 | BlockBegin* sux = block->exception_handler_at(i); |
1338 | assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler" ); |
1339 | |
1340 | BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); |
1341 | if (preds == NULL) { |
1342 | preds = new BlockList(); |
1343 | _predecessors->at_put(sux->block_id(), preds); |
1344 | } |
1345 | preds->append(block); |
1346 | } |
1347 | } |
1348 | }; |
1349 | |
1350 | class VerifyBlockBeginField : public BlockClosure { |
1351 | |
1352 | public: |
1353 | |
1354 | virtual void block_do(BlockBegin *block) { |
1355 | for ( Instruction *cur = block; cur != NULL; cur = cur->next()) { |
1356 | assert(cur->block() == block, "Block begin is not correct" ); |
1357 | } |
1358 | } |
1359 | }; |
1360 | |
1361 | void IR::verify() { |
1362 | #ifdef ASSERT |
1363 | PredecessorValidator pv(this); |
1364 | VerifyBlockBeginField verifier; |
1365 | this->iterate_postorder(&verifier); |
1366 | #endif |
1367 | } |
1368 | |
1369 | #endif // PRODUCT |
1370 | |
1371 | void SubstitutionResolver::visit(Value* v) { |
1372 | Value v0 = *v; |
1373 | if (v0) { |
1374 | Value vs = v0->subst(); |
1375 | if (vs != v0) { |
1376 | *v = v0->subst(); |
1377 | } |
1378 | } |
1379 | } |
1380 | |
1381 | #ifdef ASSERT |
1382 | class SubstitutionChecker: public ValueVisitor { |
1383 | void visit(Value* v) { |
1384 | Value v0 = *v; |
1385 | if (v0) { |
1386 | Value vs = v0->subst(); |
1387 | assert(vs == v0, "missed substitution" ); |
1388 | } |
1389 | } |
1390 | }; |
1391 | #endif |
1392 | |
1393 | |
1394 | void SubstitutionResolver::block_do(BlockBegin* block) { |
1395 | Instruction* last = NULL; |
1396 | for (Instruction* n = block; n != NULL;) { |
1397 | n->values_do(this); |
1398 | // need to remove this instruction from the instruction stream |
1399 | if (n->subst() != n) { |
1400 | guarantee(last != NULL, "must have last" ); |
1401 | last->set_next(n->next()); |
1402 | } else { |
1403 | last = n; |
1404 | } |
1405 | n = last->next(); |
1406 | } |
1407 | |
1408 | #ifdef ASSERT |
1409 | SubstitutionChecker check_substitute; |
1410 | if (block->state()) block->state()->values_do(&check_substitute); |
1411 | block->block_values_do(&check_substitute); |
1412 | if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute); |
1413 | #endif |
1414 | } |
1415 | |