1 /*
2 * Copyright (c) 1999, 2025, 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 "c1/c1_Compilation.hpp"
26 #include "c1/c1_FrameMap.hpp"
27 #include "c1/c1_GraphBuilder.hpp"
28 #include "c1/c1_InstructionPrinter.hpp"
29 #include "c1/c1_IR.hpp"
30 #include "c1/c1_Optimizer.hpp"
31 #include "compiler/oopMap.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 == nullptr) 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 nullptr;
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 == nullptr ? 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 _wrote_stable = false;
149 _start = nullptr;
150
151 if (osr_bci != -1) {
152 // selective creation of phi functions is not possibel in osr-methods
153 _requires_phi_function.set_range(0, method->max_locals());
154 }
155
156 assert(method->holder()->is_loaded() , "method holder must be loaded");
157
158 // build graph if monitor pairing is ok
159 if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci);
160 }
161
162
163 int IRScope::max_stack() const {
164 int my_max = method()->max_stack();
165 int callee_max = 0;
166 for (int i = 0; i < number_of_callees(); i++) {
167 callee_max = MAX2(callee_max, callee_no(i)->max_stack());
168 }
169 return my_max + callee_max;
170 }
171
172
173 bool IRScopeDebugInfo::should_reexecute() {
174 ciMethod* cur_method = scope()->method();
175 int cur_bci = bci();
176 if (cur_method != nullptr && cur_bci != SynchronizationEntryBCI) {
177 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
178 return Interpreter::bytecode_should_reexecute(code);
179 } else
180 return false;
181 }
182
183
184 // Implementation of CodeEmitInfo
185
186 // Stack must be NON-null
187 CodeEmitInfo::CodeEmitInfo(ValueStack* stack, XHandlers* exception_handlers, bool deoptimize_on_exception)
188 : _scope_debug_info(nullptr)
189 , _scope(stack->scope())
190 , _exception_handlers(exception_handlers)
191 , _oop_map(nullptr)
192 , _stack(stack)
193 , _deoptimize_on_exception(deoptimize_on_exception)
194 , _force_reexecute(false) {
195 assert(_stack != nullptr, "must be non null");
196 }
197
198
199 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, ValueStack* stack)
200 : _scope_debug_info(nullptr)
201 , _scope(info->_scope)
202 , _exception_handlers(nullptr)
203 , _oop_map(nullptr)
204 , _stack(stack == nullptr ? info->_stack : stack)
205 , _deoptimize_on_exception(info->_deoptimize_on_exception)
206 , _force_reexecute(info->_force_reexecute) {
207
208 // deep copy of exception handlers
209 if (info->_exception_handlers != nullptr) {
210 _exception_handlers = new XHandlers(info->_exception_handlers);
211 }
212 }
213
214
215 void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) {
216 // record the safepoint before recording the debug info for enclosing scopes
217 recorder->add_safepoint(pc_offset, _oop_map->deep_copy());
218 bool reexecute = _force_reexecute || _scope_debug_info->should_reexecute();
219 _scope_debug_info->record_debug_info(recorder, pc_offset, reexecute);
220 recorder->end_safepoint(pc_offset);
221 }
222
223
224 void CodeEmitInfo::add_register_oop(LIR_Opr opr) {
225 assert(_oop_map != nullptr, "oop map must already exist");
226 assert(opr->is_single_cpu(), "should not call otherwise");
227
228 VMReg name = frame_map()->regname(opr);
229 _oop_map->set_oop(name);
230 }
231
232 // Mirror the stack size calculation in the deopt code
233 // How much stack space would we need at this point in the program in
234 // case of deoptimization?
235 int CodeEmitInfo::interpreter_frame_size() const {
236 ValueStack* state = _stack;
237 int size = 0;
238 int callee_parameters = 0;
239 int callee_locals = 0;
240 int extra_args = state->scope()->method()->max_stack() - state->stack_size();
241
242 while (state != nullptr) {
243 int locks = state->locks_size();
244 int temps = state->stack_size();
245 bool is_top_frame = (state == _stack);
246 ciMethod* method = state->scope()->method();
247
248 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
249 temps + callee_parameters,
250 extra_args,
251 locks,
252 callee_parameters,
253 callee_locals,
254 is_top_frame);
255 size += frame_size;
256
257 callee_parameters = method->size_of_parameters();
258 callee_locals = method->max_locals();
259 extra_args = 0;
260 state = state->caller_state();
261 }
262 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
263 }
264
265 // Implementation of IR
266
267 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) :
268 _num_loops(0) {
269 // setup IR fields
270 _compilation = compilation;
271 _top_scope = new IRScope(compilation, nullptr, -1, method, osr_bci, true);
272 _code = nullptr;
273 }
274
275
276 void IR::optimize_blocks() {
277 Optimizer opt(this);
278 if (!compilation()->profile_branches()) {
279 if (DoCEE) {
280 opt.eliminate_conditional_expressions();
281 #ifndef PRODUCT
282 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); }
283 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); }
284 #endif
285 }
286 if (EliminateBlocks) {
287 opt.eliminate_blocks();
288 #ifndef PRODUCT
289 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); }
290 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); }
291 #endif
292 }
293 }
294 }
295
296 void IR::eliminate_null_checks() {
297 Optimizer opt(this);
298 if (EliminateNullChecks) {
299 opt.eliminate_null_checks();
300 #ifndef PRODUCT
301 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); }
302 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); }
303 #endif
304 }
305 }
306
307 // The functionality of this class is to insert a new block between
308 // the 'from' and 'to' block of a critical edge.
309 // It first collects the block pairs, and then processes them.
310 //
311 // Some instructions may introduce more than one edge between two blocks.
312 // By checking if the current 'to' block sets critical_edge_split_flag
313 // (all new blocks set this flag) we can avoid repeated processing.
314 // This is why BlockPair contains the index rather than the original 'to' block.
315 class CriticalEdgeFinder: public BlockClosure {
316 BlockPairList blocks;
317
318 public:
319 CriticalEdgeFinder(IR* ir) {
320 ir->iterate_preorder(this);
321 }
322
323 void block_do(BlockBegin* bb) {
324 BlockEnd* be = bb->end();
325 int nos = be->number_of_sux();
326 if (nos >= 2) {
327 for (int i = 0; i < nos; i++) {
328 BlockBegin* sux = be->sux_at(i);
329 if (sux->number_of_preds() >= 2) {
330 blocks.append(new BlockPair(bb, i));
331 }
332 }
333 }
334 }
335
336 void split_edges() {
337 for (int i = 0; i < blocks.length(); i++) {
338 BlockPair* pair = blocks.at(i);
339 BlockBegin* from = pair->from();
340 int index = pair->index();
341 BlockBegin* to = from->end()->sux_at(index);
342 if (to->is_set(BlockBegin::critical_edge_split_flag)) {
343 // inserted
344 continue;
345 }
346 BlockBegin* split = from->insert_block_between(to);
347 #ifndef PRODUCT
348 if ((PrintIR || PrintIR1) && Verbose) {
349 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)",
350 from->block_id(), to->block_id(), split->block_id());
351 }
352 #endif
353 }
354 }
355 };
356
357 void IR::split_critical_edges() {
358 CriticalEdgeFinder cef(this);
359 cef.split_edges();
360 }
361
362
363 class UseCountComputer: public ValueVisitor, BlockClosure {
364 private:
365 void visit(Value* n) {
366 // Local instructions and Phis for expression stack values at the
367 // start of basic blocks are not added to the instruction list
368 if (!(*n)->is_linked() && (*n)->can_be_linked()) {
369 assert(false, "a node was not appended to the graph");
370 Compilation::current()->bailout("a node was not appended to the graph");
371 }
372 // use n's input if not visited before
373 if (!(*n)->is_pinned() && !(*n)->has_uses()) {
374 // note: a) if the instruction is pinned, it will be handled by compute_use_count
375 // b) if the instruction has uses, it was touched before
376 // => in both cases we don't need to update n's values
377 uses_do(n);
378 }
379 // use n
380 (*n)->_use_count++;
381 }
382
383 Values* worklist;
384 int depth;
385 enum {
386 max_recurse_depth = 20
387 };
388
389 void uses_do(Value* n) {
390 depth++;
391 if (depth > max_recurse_depth) {
392 // don't allow the traversal to recurse too deeply
393 worklist->push(*n);
394 } else {
395 (*n)->input_values_do(this);
396 // special handling for some instructions
397 if ((*n)->as_BlockEnd() != nullptr) {
398 // note on BlockEnd:
399 // must 'use' the stack only if the method doesn't
400 // terminate, however, in those cases stack is empty
401 (*n)->state_values_do(this);
402 }
403 }
404 depth--;
405 }
406
407 void block_do(BlockBegin* b) {
408 depth = 0;
409 // process all pinned nodes as the roots of expression trees
410 for (Instruction* n = b; n != nullptr; n = n->next()) {
411 if (n->is_pinned()) uses_do(&n);
412 }
413 assert(depth == 0, "should have counted back down");
414
415 // now process any unpinned nodes which recursed too deeply
416 while (worklist->length() > 0) {
417 Value t = worklist->pop();
418 if (!t->is_pinned()) {
419 // compute the use count
420 uses_do(&t);
421
422 // pin the instruction so that LIRGenerator doesn't recurse
423 // too deeply during it's evaluation.
424 t->pin();
425 }
426 }
427 assert(depth == 0, "should have counted back down");
428 }
429
430 UseCountComputer() {
431 worklist = new Values();
432 depth = 0;
433 }
434
435 public:
436 static void compute(BlockList* blocks) {
437 UseCountComputer ucc;
438 blocks->iterate_backward(&ucc);
439 }
440 };
441
442
443 // helper macro for short definition of trace-output inside code
444 #ifdef ASSERT
445 #define TRACE_LINEAR_SCAN(level, code) \
446 if (TraceLinearScanLevel >= level) { \
447 code; \
448 }
449 #else
450 #define TRACE_LINEAR_SCAN(level, code)
451 #endif
452
453 class ComputeLinearScanOrder : public StackObj {
454 private:
455 int _max_block_id; // the highest block_id of a block
456 int _num_blocks; // total number of blocks (smaller than _max_block_id)
457 int _num_loops; // total number of loops
458 bool _iterative_dominators;// method requires iterative computation of dominatiors
459
460 BlockList* _linear_scan_order; // the resulting list of blocks in correct order
461
462 ResourceBitMap _visited_blocks; // used for recursive processing of blocks
463 ResourceBitMap _active_blocks; // used for recursive processing of blocks
464 ResourceBitMap _dominator_blocks; // temporary BitMap used for computation of dominator
465 intArray _forward_branches; // number of incoming forward branches for each block
466 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges
467 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop
468 BlockList _work_list; // temporary list (used in mark_loops and compute_order)
469 BlockList _loop_headers;
470
471 Compilation* _compilation;
472
473 // accessors for _visited_blocks and _active_blocks
474 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); }
475 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); }
476 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); }
477 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); }
478 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); }
479 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); }
480
481 // accessors for _forward_branches
482 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); }
483 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()); }
484
485 // accessors for _loop_map
486 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); }
487 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); }
488 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); }
489
490 // count edges between blocks
491 void count_edges(BlockBegin* cur, BlockBegin* parent);
492
493 // loop detection
494 void mark_loops();
495 void clear_non_natural_loops(BlockBegin* start_block);
496 void assign_loop_depth(BlockBegin* start_block);
497
498 // computation of final block order
499 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b);
500 void compute_dominator(BlockBegin* cur, BlockBegin* parent);
501 void compute_dominator_impl(BlockBegin* cur, BlockBegin* parent);
502 int compute_weight(BlockBegin* cur);
503 bool ready_for_processing(BlockBegin* cur);
504 void sort_into_work_list(BlockBegin* b);
505 void append_block(BlockBegin* cur);
506 void compute_order(BlockBegin* start_block);
507
508 // fixup of dominators for non-natural loops
509 bool compute_dominators_iter();
510 void compute_dominators();
511
512 // debug functions
513 DEBUG_ONLY(void print_blocks();)
514 DEBUG_ONLY(void verify();)
515
516 Compilation* compilation() const { return _compilation; }
517 public:
518 ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block);
519
520 // accessors for final result
521 BlockList* linear_scan_order() const { return _linear_scan_order; }
522 int num_loops() const { return _num_loops; }
523 };
524
525
526 ComputeLinearScanOrder::ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block) :
527 _max_block_id(BlockBegin::number_of_blocks()),
528 _num_blocks(0),
529 _num_loops(0),
530 _iterative_dominators(false),
531 _linear_scan_order(nullptr), // initialized later with correct size
532 _visited_blocks(_max_block_id),
533 _active_blocks(_max_block_id),
534 _dominator_blocks(_max_block_id),
535 _forward_branches(_max_block_id, _max_block_id, 0),
536 _loop_end_blocks(8),
537 _loop_map(0), // initialized later with correct size
538 _work_list(8),
539 _compilation(c)
540 {
541 TRACE_LINEAR_SCAN(2, tty->print_cr("***** computing linear-scan block order"));
542
543 count_edges(start_block, nullptr);
544
545 if (compilation()->is_profiling()) {
546 ciMethod *method = compilation()->method();
547 if (!method->is_accessor()) {
548 ciMethodData* md = method->method_data_or_null();
549 assert(md != nullptr, "Sanity");
550 md->set_compilation_stats(_num_loops, _num_blocks);
551 }
552 }
553
554 if (_num_loops > 0) {
555 mark_loops();
556 clear_non_natural_loops(start_block);
557 assign_loop_depth(start_block);
558 }
559
560 compute_order(start_block);
561 compute_dominators();
562
563 DEBUG_ONLY(print_blocks());
564 DEBUG_ONLY(verify());
565 }
566
567
568 // Traverse the CFG:
569 // * count total number of blocks
570 // * count all incoming edges and backward incoming edges
571 // * number loop header blocks
572 // * create a list with all loop end blocks
573 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) {
574 TRACE_LINEAR_SCAN(3, tty->print_cr("Enter count_edges for block B%d coming from B%d", cur->block_id(), parent != nullptr ? parent->block_id() : -1));
575 assert(cur->dominator() == nullptr, "dominator already initialized");
576
577 if (is_active(cur)) {
578 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch"));
579 assert(is_visited(cur), "block must be visisted when block is active");
580 assert(parent != nullptr, "must have parent");
581
582 cur->set(BlockBegin::backward_branch_target_flag);
583
584 // When a loop header is also the start of an exception handler, then the backward branch is
585 // an exception edge. Because such edges are usually critical edges which cannot be split, the
586 // loop must be excluded here from processing.
587 if (cur->is_set(BlockBegin::exception_entry_flag)) {
588 // Make sure that dominators are correct in this weird situation
589 _iterative_dominators = true;
590 return;
591 }
592
593 cur->set(BlockBegin::linear_scan_loop_header_flag);
594 parent->set(BlockBegin::linear_scan_loop_end_flag);
595
596 assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur,
597 "loop end blocks must have one successor (critical edges are split)");
598
599 _loop_end_blocks.append(parent);
600 return;
601 }
602
603 // increment number of incoming forward branches
604 inc_forward_branches(cur);
605
606 if (is_visited(cur)) {
607 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited"));
608 return;
609 }
610
611 _num_blocks++;
612 set_visited(cur);
613 set_active(cur);
614
615 // recursive call for all successors
616 int i;
617 for (i = cur->number_of_sux() - 1; i >= 0; i--) {
618 count_edges(cur->sux_at(i), cur);
619 }
620 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
621 count_edges(cur->exception_handler_at(i), cur);
622 }
623
624 clear_active(cur);
625
626 // Each loop has a unique number.
627 // When multiple loops are nested, assign_loop_depth assumes that the
628 // innermost loop has the lowest number. This is guaranteed by setting
629 // the loop number after the recursive calls for the successors above
630 // have returned.
631 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) {
632 assert(cur->loop_index() == -1, "cannot set loop-index twice");
633 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops));
634
635 cur->set_loop_index(_num_loops);
636 _loop_headers.append(cur);
637 _num_loops++;
638 }
639
640 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id()));
641 }
642
643
644 void ComputeLinearScanOrder::mark_loops() {
645 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops"));
646
647 _loop_map = BitMap2D(_num_loops, _max_block_id);
648
649 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) {
650 BlockBegin* loop_end = _loop_end_blocks.at(i);
651 BlockBegin* loop_start = loop_end->sux_at(0);
652 int loop_idx = loop_start->loop_index();
653
654 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));
655 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set");
656 assert(loop_end->number_of_sux() == 1, "incorrect number of successors");
657 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set");
658 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set");
659 assert(_work_list.is_empty(), "work list must be empty before processing");
660
661 // add the end-block of the loop to the working list
662 _work_list.push(loop_end);
663 set_block_in_loop(loop_idx, loop_end);
664 do {
665 BlockBegin* cur = _work_list.pop();
666
667 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id()));
668 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list");
669
670 // recursive processing of all predecessors ends when start block of loop is reached
671 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) {
672 for (int j = cur->number_of_preds() - 1; j >= 0; j--) {
673 BlockBegin* pred = cur->pred_at(j);
674
675 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) {
676 // this predecessor has not been processed yet, so add it to work list
677 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id()));
678 _work_list.push(pred);
679 set_block_in_loop(loop_idx, pred);
680 }
681 }
682 }
683 } while (!_work_list.is_empty());
684 }
685 }
686
687
688 // check for non-natural loops (loops where the loop header does not dominate
689 // all other loop blocks = loops with multiple entries).
690 // such loops are ignored
691 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) {
692 for (int i = _num_loops - 1; i >= 0; i--) {
693 if (is_block_in_loop(i, start_block)) {
694 // loop i contains the entry block of the method
695 // -> this is not a natural loop, so ignore it
696 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i));
697
698 BlockBegin *loop_header = _loop_headers.at(i);
699 assert(loop_header->is_set(BlockBegin::linear_scan_loop_header_flag), "Must be loop header");
700
701 for (int j = 0; j < loop_header->number_of_preds(); j++) {
702 BlockBegin *pred = loop_header->pred_at(j);
703 pred->clear(BlockBegin::linear_scan_loop_end_flag);
704 }
705
706 loop_header->clear(BlockBegin::linear_scan_loop_header_flag);
707
708 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) {
709 clear_block_in_loop(i, block_id);
710 }
711 _iterative_dominators = true;
712 }
713 }
714 }
715
716 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) {
717 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing loop-depth and weight"));
718 init_visited();
719
720 assert(_work_list.is_empty(), "work list must be empty before processing");
721 _work_list.append(start_block);
722
723 do {
724 BlockBegin* cur = _work_list.pop();
725
726 if (!is_visited(cur)) {
727 set_visited(cur);
728 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id()));
729
730 // compute loop-depth and loop-index for the block
731 assert(cur->loop_depth() == 0, "cannot set loop-depth twice");
732 int i;
733 int loop_depth = 0;
734 int min_loop_idx = -1;
735 for (i = _num_loops - 1; i >= 0; i--) {
736 if (is_block_in_loop(i, cur)) {
737 loop_depth++;
738 min_loop_idx = i;
739 }
740 }
741 cur->set_loop_depth(loop_depth);
742 cur->set_loop_index(min_loop_idx);
743
744 // append all unvisited successors to work list
745 for (i = cur->number_of_sux() - 1; i >= 0; i--) {
746 _work_list.append(cur->sux_at(i));
747 }
748 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
749 _work_list.append(cur->exception_handler_at(i));
750 }
751 }
752 } while (!_work_list.is_empty());
753 }
754
755
756 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) {
757 assert(a != nullptr && b != nullptr, "must have input blocks");
758
759 _dominator_blocks.clear();
760 while (a != nullptr) {
761 _dominator_blocks.set_bit(a->block_id());
762 assert(a->dominator() != nullptr || a == _linear_scan_order->at(0), "dominator must be initialized");
763 a = a->dominator();
764 }
765 while (b != nullptr && !_dominator_blocks.at(b->block_id())) {
766 assert(b->dominator() != nullptr || b == _linear_scan_order->at(0), "dominator must be initialized");
767 b = b->dominator();
768 }
769
770 assert(b != nullptr, "could not find dominator");
771 return b;
772 }
773
774 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) {
775 init_visited();
776 compute_dominator_impl(cur, parent);
777 }
778
779 void ComputeLinearScanOrder::compute_dominator_impl(BlockBegin* cur, BlockBegin* parent) {
780 // Mark as visited to avoid recursive calls with same parent
781 set_visited(cur);
782
783 if (cur->dominator() == nullptr) {
784 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id()));
785 cur->set_dominator(parent);
786
787 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) {
788 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()));
789 // Does not hold for exception blocks
790 assert(cur->number_of_preds() > 1 || cur->is_set(BlockBegin::exception_entry_flag), "");
791 cur->set_dominator(common_dominator(cur->dominator(), parent));
792 }
793
794 // Additional edge to xhandler of all our successors
795 // range check elimination needs that the state at the end of a
796 // block be valid in every block it dominates so cur must dominate
797 // the exception handlers of its successors.
798 int num_cur_xhandler = cur->number_of_exception_handlers();
799 for (int j = 0; j < num_cur_xhandler; j++) {
800 BlockBegin* xhandler = cur->exception_handler_at(j);
801 if (!is_visited(xhandler)) {
802 compute_dominator_impl(xhandler, parent);
803 }
804 }
805 }
806
807
808 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) {
809 BlockBegin* single_sux = nullptr;
810 if (cur->number_of_sux() == 1) {
811 single_sux = cur->sux_at(0);
812 }
813
814 // limit loop-depth to 15 bit (only for security reason, it will never be so big)
815 int weight = (cur->loop_depth() & 0x7FFF) << 16;
816
817 // general macro for short definition of weight flags
818 // the first instance of INC_WEIGHT_IF has the highest priority
819 int cur_bit = 15;
820 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--;
821
822 // this is necessary for the (very rare) case that two successive blocks have
823 // the same loop depth, but a different loop index (can happen for endless loops
824 // with exception handlers)
825 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag));
826
827 // loop end blocks (blocks that end with a backward branch) are added
828 // after all other blocks of the loop.
829 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag));
830
831 // critical edge split blocks are preferred because than they have a bigger
832 // proability to be completely empty
833 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag));
834
835 // exceptions should not be thrown in normal control flow, so these blocks
836 // are added as late as possible
837 INC_WEIGHT_IF(cur->end()->as_Throw() == nullptr && (single_sux == nullptr || single_sux->end()->as_Throw() == nullptr));
838 INC_WEIGHT_IF(cur->end()->as_Return() == nullptr && (single_sux == nullptr || single_sux->end()->as_Return() == nullptr));
839
840 // exceptions handlers are added as late as possible
841 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag));
842
843 // guarantee that weight is > 0
844 weight |= 1;
845
846 #undef INC_WEIGHT_IF
847 assert(cur_bit >= 0, "too many flags");
848 assert(weight > 0, "weight cannot become negative");
849
850 return weight;
851 }
852
853 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) {
854 // Discount the edge just traveled.
855 // When the number drops to zero, all forward branches were processed
856 if (dec_forward_branches(cur) != 0) {
857 return false;
858 }
859
860 assert(_linear_scan_order->find(cur) == -1, "block already processed (block can be ready only once)");
861 assert(_work_list.find(cur) == -1, "block already in work-list (block can be ready only once)");
862 return true;
863 }
864
865 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) {
866 assert(_work_list.find(cur) == -1, "block already in work list");
867
868 int cur_weight = compute_weight(cur);
869
870 // the linear_scan_number is used to cache the weight of a block
871 cur->set_linear_scan_number(cur_weight);
872
873 #ifndef PRODUCT
874 if (StressLinearScan) {
875 _work_list.insert_before(0, cur);
876 return;
877 }
878 #endif
879
880 _work_list.append(nullptr); // provide space for new element
881
882 int insert_idx = _work_list.length() - 1;
883 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) {
884 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1));
885 insert_idx--;
886 }
887 _work_list.at_put(insert_idx, cur);
888
889 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id()));
890 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()));
891
892 #ifdef ASSERT
893 for (int i = 0; i < _work_list.length(); i++) {
894 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set");
895 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist");
896 }
897 #endif
898 }
899
900 void ComputeLinearScanOrder::append_block(BlockBegin* cur) {
901 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()));
902 assert(_linear_scan_order->find(cur) == -1, "cannot add the same block twice");
903
904 // currently, the linear scan order and code emit order are equal.
905 // therefore the linear_scan_number and the weight of a block must also
906 // be equal.
907 cur->set_linear_scan_number(_linear_scan_order->length());
908 _linear_scan_order->append(cur);
909 }
910
911 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) {
912 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing final block order"));
913
914 // the start block is always the first block in the linear scan order
915 _linear_scan_order = new BlockList(_num_blocks);
916 append_block(start_block);
917
918 assert(start_block->end()->as_Base() != nullptr, "start block must end with Base-instruction");
919 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry();
920 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry();
921
922 BlockBegin* sux_of_osr_entry = nullptr;
923 if (osr_entry != nullptr) {
924 // special handling for osr entry:
925 // ignore the edge between the osr entry and its successor for processing
926 // the osr entry block is added manually below
927 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor");
928 assert(osr_entry->sux_at(0)->number_of_preds() >= 2, "successor of osr entry must have two predecessors (otherwise it is not present in normal control flow");
929
930 sux_of_osr_entry = osr_entry->sux_at(0);
931 dec_forward_branches(sux_of_osr_entry);
932
933 compute_dominator(osr_entry, start_block);
934 _iterative_dominators = true;
935 }
936 compute_dominator(std_entry, start_block);
937
938 // start processing with standard entry block
939 assert(_work_list.is_empty(), "list must be empty before processing");
940
941 if (ready_for_processing(std_entry)) {
942 sort_into_work_list(std_entry);
943 } else {
944 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)");
945 }
946
947 do {
948 BlockBegin* cur = _work_list.pop();
949
950 if (cur == sux_of_osr_entry) {
951 // the osr entry block is ignored in normal processing, it is never added to the
952 // work list. Instead, it is added as late as possible manually here.
953 append_block(osr_entry);
954 compute_dominator(cur, osr_entry);
955 }
956 append_block(cur);
957
958 int i;
959 int num_sux = cur->number_of_sux();
960 // changed loop order to get "intuitive" order of if- and else-blocks
961 for (i = 0; i < num_sux; i++) {
962 BlockBegin* sux = cur->sux_at(i);
963 compute_dominator(sux, cur);
964 if (ready_for_processing(sux)) {
965 sort_into_work_list(sux);
966 }
967 }
968 num_sux = cur->number_of_exception_handlers();
969 for (i = 0; i < num_sux; i++) {
970 BlockBegin* sux = cur->exception_handler_at(i);
971 if (ready_for_processing(sux)) {
972 sort_into_work_list(sux);
973 }
974 }
975 } while (_work_list.length() > 0);
976 }
977
978
979 bool ComputeLinearScanOrder::compute_dominators_iter() {
980 bool changed = false;
981 int num_blocks = _linear_scan_order->length();
982
983 assert(_linear_scan_order->at(0)->dominator() == nullptr, "must not have dominator");
984 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors");
985 for (int i = 1; i < num_blocks; i++) {
986 BlockBegin* block = _linear_scan_order->at(i);
987
988 BlockBegin* dominator = block->pred_at(0);
989 int num_preds = block->number_of_preds();
990
991 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: Processing B%d", block->block_id()));
992
993 for (int j = 0; j < num_preds; j++) {
994
995 BlockBegin *pred = block->pred_at(j);
996 TRACE_LINEAR_SCAN(4, tty->print_cr(" DOM: Subrocessing B%d", pred->block_id()));
997
998 if (block->is_set(BlockBegin::exception_entry_flag)) {
999 dominator = common_dominator(dominator, pred);
1000 int num_pred_preds = pred->number_of_preds();
1001 for (int k = 0; k < num_pred_preds; k++) {
1002 dominator = common_dominator(dominator, pred->pred_at(k));
1003 }
1004 } else {
1005 dominator = common_dominator(dominator, pred);
1006 }
1007 }
1008
1009 if (dominator != block->dominator()) {
1010 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()));
1011
1012 block->set_dominator(dominator);
1013 changed = true;
1014 }
1015 }
1016 return changed;
1017 }
1018
1019 void ComputeLinearScanOrder::compute_dominators() {
1020 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators));
1021
1022 // iterative computation of dominators is only required for methods with non-natural loops
1023 // and OSR-methods. For all other methods, the dominators computed when generating the
1024 // linear scan block order are correct.
1025 if (_iterative_dominators) {
1026 do {
1027 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation"));
1028 } while (compute_dominators_iter());
1029 }
1030
1031 // check that dominators are correct
1032 assert(!compute_dominators_iter(), "fix point not reached");
1033
1034 // Add Blocks to dominates-Array
1035 int num_blocks = _linear_scan_order->length();
1036 for (int i = 0; i < num_blocks; i++) {
1037 BlockBegin* block = _linear_scan_order->at(i);
1038
1039 BlockBegin *dom = block->dominator();
1040 if (dom) {
1041 assert(dom->dominator_depth() != -1, "Dominator must have been visited before");
1042 dom->dominates()->append(block);
1043 block->set_dominator_depth(dom->dominator_depth() + 1);
1044 } else {
1045 block->set_dominator_depth(0);
1046 }
1047 }
1048 }
1049
1050
1051 #ifdef ASSERT
1052 void ComputeLinearScanOrder::print_blocks() {
1053 if (TraceLinearScanLevel >= 2) {
1054 tty->print_cr("----- loop information:");
1055 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
1056 BlockBegin* cur = _linear_scan_order->at(block_idx);
1057
1058 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id());
1059 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
1060 tty->print ("%d ", is_block_in_loop(loop_idx, cur));
1061 }
1062 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth());
1063 }
1064 }
1065
1066 if (TraceLinearScanLevel >= 1) {
1067 tty->print_cr("----- linear-scan block order:");
1068 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
1069 BlockBegin* cur = _linear_scan_order->at(block_idx);
1070 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth());
1071
1072 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " ");
1073 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " ");
1074 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " ");
1075 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " ");
1076
1077 if (cur->dominator() != nullptr) {
1078 tty->print(" dom: B%d ", cur->dominator()->block_id());
1079 } else {
1080 tty->print(" dom: null ");
1081 }
1082
1083 if (cur->number_of_preds() > 0) {
1084 tty->print(" preds: ");
1085 for (int j = 0; j < cur->number_of_preds(); j++) {
1086 BlockBegin* pred = cur->pred_at(j);
1087 tty->print("B%d ", pred->block_id());
1088 }
1089 }
1090 if (cur->number_of_sux() > 0) {
1091 tty->print(" sux: ");
1092 for (int j = 0; j < cur->number_of_sux(); j++) {
1093 BlockBegin* sux = cur->sux_at(j);
1094 tty->print("B%d ", sux->block_id());
1095 }
1096 }
1097 if (cur->number_of_exception_handlers() > 0) {
1098 tty->print(" ex: ");
1099 for (int j = 0; j < cur->number_of_exception_handlers(); j++) {
1100 BlockBegin* ex = cur->exception_handler_at(j);
1101 tty->print("B%d ", ex->block_id());
1102 }
1103 }
1104 tty->cr();
1105 }
1106 }
1107 }
1108
1109 void ComputeLinearScanOrder::verify() {
1110 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list");
1111
1112 if (StressLinearScan) {
1113 // blocks are scrambled when StressLinearScan is used
1114 return;
1115 }
1116
1117 // check that all successors of a block have a higher linear-scan-number
1118 // and that all predecessors of a block have a lower linear-scan-number
1119 // (only backward branches of loops are ignored)
1120 int i;
1121 for (i = 0; i < _linear_scan_order->length(); i++) {
1122 BlockBegin* cur = _linear_scan_order->at(i);
1123
1124 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number");
1125 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->find(cur), "incorrect linear_scan_number");
1126
1127 int j;
1128 for (j = cur->number_of_sux() - 1; j >= 0; j--) {
1129 BlockBegin* sux = cur->sux_at(j);
1130
1131 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->find(sux), "incorrect linear_scan_number");
1132 if (!sux->is_set(BlockBegin::backward_branch_target_flag)) {
1133 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order");
1134 }
1135 if (cur->loop_depth() == sux->loop_depth()) {
1136 assert(cur->loop_index() == sux->loop_index() || sux->is_set(BlockBegin::linear_scan_loop_header_flag), "successive blocks with same loop depth must have same loop index");
1137 }
1138 }
1139
1140 for (j = cur->number_of_preds() - 1; j >= 0; j--) {
1141 BlockBegin* pred = cur->pred_at(j);
1142
1143 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->find(pred), "incorrect linear_scan_number");
1144 if (!cur->is_set(BlockBegin::backward_branch_target_flag)) {
1145 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order");
1146 }
1147 if (cur->loop_depth() == pred->loop_depth()) {
1148 assert(cur->loop_index() == pred->loop_index() || cur->is_set(BlockBegin::linear_scan_loop_header_flag), "successive blocks with same loop depth must have same loop index");
1149 }
1150
1151 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors");
1152 }
1153
1154 // check dominator
1155 if (i == 0) {
1156 assert(cur->dominator() == nullptr, "first block has no dominator");
1157 } else {
1158 assert(cur->dominator() != nullptr, "all but first block must have dominator");
1159 }
1160 // Assertion does not hold for exception handlers
1161 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");
1162 }
1163
1164 // check that all loops are continuous
1165 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
1166 int block_idx = 0;
1167 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop");
1168
1169 // skip blocks before the loop
1170 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
1171 block_idx++;
1172 }
1173 // skip blocks of loop
1174 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
1175 block_idx++;
1176 }
1177 // after the first non-loop block, there must not be another loop-block
1178 while (block_idx < _num_blocks) {
1179 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order");
1180 block_idx++;
1181 }
1182 }
1183 }
1184 #endif // ASSERT
1185
1186
1187 void IR::compute_code() {
1188 assert(is_valid(), "IR must be valid");
1189
1190 ComputeLinearScanOrder compute_order(compilation(), start());
1191 _num_loops = compute_order.num_loops();
1192 _code = compute_order.linear_scan_order();
1193 }
1194
1195
1196 void IR::compute_use_counts() {
1197 // make sure all values coming out of this block get evaluated.
1198 int num_blocks = _code->length();
1199 for (int i = 0; i < num_blocks; i++) {
1200 _code->at(i)->end()->state()->pin_stack_for_linear_scan();
1201 }
1202
1203 // compute use counts
1204 UseCountComputer::compute(_code);
1205 }
1206
1207
1208 void IR::iterate_preorder(BlockClosure* closure) {
1209 assert(is_valid(), "IR must be valid");
1210 start()->iterate_preorder(closure);
1211 }
1212
1213
1214 void IR::iterate_postorder(BlockClosure* closure) {
1215 assert(is_valid(), "IR must be valid");
1216 start()->iterate_postorder(closure);
1217 }
1218
1219 void IR::iterate_linear_scan_order(BlockClosure* closure) {
1220 linear_scan_order()->iterate_forward(closure);
1221 }
1222
1223
1224 #ifndef PRODUCT
1225 class BlockPrinter: public BlockClosure {
1226 private:
1227 InstructionPrinter* _ip;
1228 bool _cfg_only;
1229 bool _live_only;
1230
1231 public:
1232 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) {
1233 _ip = ip;
1234 _cfg_only = cfg_only;
1235 _live_only = live_only;
1236 }
1237
1238 virtual void block_do(BlockBegin* block) {
1239 if (_cfg_only) {
1240 _ip->print_instr(block); tty->cr();
1241 } else {
1242 block->print_block(*_ip, _live_only);
1243 }
1244 }
1245 };
1246
1247
1248 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) {
1249 ttyLocker ttyl;
1250 InstructionPrinter ip(!cfg_only);
1251 BlockPrinter bp(&ip, cfg_only, live_only);
1252 start->iterate_preorder(&bp);
1253 tty->cr();
1254 }
1255
1256 void IR::print(bool cfg_only, bool live_only) {
1257 if (is_valid()) {
1258 print(start(), cfg_only, live_only);
1259 } else {
1260 tty->print_cr("invalid IR");
1261 }
1262 }
1263 #endif // PRODUCT
1264
1265 #ifdef ASSERT
1266 class EndNotNullValidator : public BlockClosure {
1267 public:
1268 virtual void block_do(BlockBegin* block) {
1269 assert(block->end() != nullptr, "Expect block end to exist.");
1270 }
1271 };
1272
1273 class XentryFlagValidator : public BlockClosure {
1274 public:
1275 virtual void block_do(BlockBegin* block) {
1276 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1277 assert(!block->end()->sux_at(i)->is_set(BlockBegin::exception_entry_flag), "must not be xhandler");
1278 }
1279 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1280 assert(block->exception_handler_at(i)->is_set(BlockBegin::exception_entry_flag), "must be xhandler");
1281 }
1282 }
1283 };
1284
1285 typedef GrowableArray<BlockList*> BlockListList;
1286
1287 // Validation goals:
1288 // - code() length == blocks length
1289 // - code() contents == blocks content
1290 // - Each block's computed predecessors match sux lists (length)
1291 // - Each block's computed predecessors match sux lists (set content)
1292 class PredecessorAndCodeValidator : public BlockClosure {
1293 private:
1294 BlockListList* _predecessors; // Each index i will hold predecessors of block with id i
1295 BlockList* _blocks;
1296
1297 static int cmp(BlockBegin** a, BlockBegin** b) {
1298 return (*a)->block_id() - (*b)->block_id();
1299 }
1300
1301 public:
1302 PredecessorAndCodeValidator(IR* hir) {
1303 ResourceMark rm;
1304 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), BlockBegin::number_of_blocks(), nullptr);
1305 _blocks = new BlockList(BlockBegin::number_of_blocks());
1306
1307 hir->start()->iterate_preorder(this);
1308 if (hir->code() != nullptr) {
1309 assert(hir->code()->length() == _blocks->length(), "must match");
1310 for (int i = 0; i < _blocks->length(); i++) {
1311 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists");
1312 }
1313 }
1314
1315 for (int i = 0; i < _blocks->length(); i++) {
1316 BlockBegin* block = _blocks->at(i);
1317 verify_block_preds_against_collected_preds(block);
1318 }
1319 }
1320
1321 virtual void block_do(BlockBegin* block) {
1322 _blocks->append(block);
1323 collect_predecessors(block);
1324 }
1325
1326 private:
1327 void collect_predecessors(BlockBegin* block) {
1328 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1329 collect_predecessor(block, block->end()->sux_at(i));
1330 }
1331 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1332 collect_predecessor(block, block->exception_handler_at(i));
1333 }
1334 }
1335
1336 void collect_predecessor(BlockBegin* const pred, const BlockBegin* sux) {
1337 BlockList* preds = _predecessors->at_grow(sux->block_id(), nullptr);
1338 if (preds == nullptr) {
1339 preds = new BlockList();
1340 _predecessors->at_put(sux->block_id(), preds);
1341 }
1342 preds->append(pred);
1343 }
1344
1345 void verify_block_preds_against_collected_preds(const BlockBegin* block) const {
1346 BlockList* preds = _predecessors->at(block->block_id());
1347 if (preds == nullptr) {
1348 assert(block->number_of_preds() == 0, "should be the same");
1349 return;
1350 }
1351 assert(preds->length() == block->number_of_preds(), "should be the same");
1352
1353 // clone the pred list so we can mutate it
1354 BlockList* pred_copy = new BlockList();
1355 for (int j = 0; j < block->number_of_preds(); j++) {
1356 pred_copy->append(block->pred_at(j));
1357 }
1358 // sort them in the same order
1359 preds->sort(cmp);
1360 pred_copy->sort(cmp);
1361 for (int j = 0; j < block->number_of_preds(); j++) {
1362 assert(preds->at(j) == pred_copy->at(j), "must match");
1363 }
1364 }
1365 };
1366
1367 class VerifyBlockBeginField : public BlockClosure {
1368 public:
1369 virtual void block_do(BlockBegin* block) {
1370 for (Instruction* cur = block; cur != nullptr; cur = cur->next()) {
1371 assert(cur->block() == block, "Block begin is not correct");
1372 }
1373 }
1374 };
1375
1376 class ValidateEdgeMutuality : public BlockClosure {
1377 public:
1378 virtual void block_do(BlockBegin* block) {
1379 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1380 assert(block->end()->sux_at(i)->is_predecessor(block), "Block's successor should have it as predecessor");
1381 }
1382
1383 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1384 assert(block->exception_handler_at(i)->is_predecessor(block), "Block's exception handler should have it as predecessor");
1385 }
1386
1387 for (int i = 0; i < block->number_of_preds(); i++) {
1388 assert(block->pred_at(i) != nullptr, "Predecessor must exist");
1389 assert(block->pred_at(i)->end() != nullptr, "Predecessor end must exist");
1390 bool is_sux = block->pred_at(i)->end()->is_sux(block);
1391 bool is_xhandler = block->pred_at(i)->is_exception_handler(block);
1392 assert(is_sux || is_xhandler, "Block's predecessor should have it as successor or xhandler");
1393 }
1394 }
1395 };
1396
1397 void IR::expand_with_neighborhood(BlockList& blocks) {
1398 int original_size = blocks.length();
1399 for (int h = 0; h < original_size; h++) {
1400 BlockBegin* block = blocks.at(h);
1401
1402 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1403 if (!blocks.contains(block->end()->sux_at(i))) {
1404 blocks.append(block->end()->sux_at(i));
1405 }
1406 }
1407
1408 for (int i = 0; i < block->number_of_preds(); i++) {
1409 if (!blocks.contains(block->pred_at(i))) {
1410 blocks.append(block->pred_at(i));
1411 }
1412 }
1413
1414 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1415 if (!blocks.contains(block->exception_handler_at(i))) {
1416 blocks.append(block->exception_handler_at(i));
1417 }
1418 }
1419 }
1420 }
1421
1422 void IR::verify_local(BlockList& blocks) {
1423 EndNotNullValidator ennv;
1424 blocks.iterate_forward(&ennv);
1425
1426 ValidateEdgeMutuality vem;
1427 blocks.iterate_forward(&vem);
1428
1429 VerifyBlockBeginField verifier;
1430 blocks.iterate_forward(&verifier);
1431 }
1432
1433 void IR::verify() {
1434 XentryFlagValidator xe;
1435 iterate_postorder(&xe);
1436
1437 PredecessorAndCodeValidator pv(this);
1438
1439 EndNotNullValidator ennv;
1440 iterate_postorder(&ennv);
1441
1442 ValidateEdgeMutuality vem;
1443 iterate_postorder(&vem);
1444
1445 VerifyBlockBeginField verifier;
1446 iterate_postorder(&verifier);
1447 }
1448 #endif // ASSERT
1449
1450 void SubstitutionResolver::visit(Value* v) {
1451 Value v0 = *v;
1452 if (v0) {
1453 Value vs = v0->subst();
1454 if (vs != v0) {
1455 *v = v0->subst();
1456 }
1457 }
1458 }
1459
1460 #ifdef ASSERT
1461 class SubstitutionChecker: public ValueVisitor {
1462 void visit(Value* v) {
1463 Value v0 = *v;
1464 if (v0) {
1465 Value vs = v0->subst();
1466 assert(vs == v0, "missed substitution");
1467 }
1468 }
1469 };
1470 #endif
1471
1472
1473 void SubstitutionResolver::block_do(BlockBegin* block) {
1474 Instruction* last = nullptr;
1475 for (Instruction* n = block; n != nullptr;) {
1476 n->values_do(this);
1477 // need to remove this instruction from the instruction stream
1478 if (n->subst() != n) {
1479 guarantee(last != nullptr, "must have last");
1480 last->set_next(n->next());
1481 } else {
1482 last = n;
1483 }
1484 n = last->next();
1485 }
1486
1487 #ifdef ASSERT
1488 SubstitutionChecker check_substitute;
1489 if (block->state()) block->state()->values_do(&check_substitute);
1490 block->block_values_do(&check_substitute);
1491 if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute);
1492 #endif
1493 }