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_non_strict_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 if (_should_reexecute) {
175 return true;
176 }
177 ciMethod* cur_method = scope()->method();
178 int cur_bci = bci();
179 if (cur_method != nullptr && cur_bci != SynchronizationEntryBCI) {
180 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
181 return Interpreter::bytecode_should_reexecute(code);
182 } else
183 return false;
184 }
185
186
187 // Implementation of CodeEmitInfo
188
189 // Stack must be NON-null
190 CodeEmitInfo::CodeEmitInfo(ValueStack* stack, XHandlers* exception_handlers, bool deoptimize_on_exception)
191 : _scope_debug_info(nullptr)
192 , _scope(stack->scope())
193 , _exception_handlers(exception_handlers)
194 , _oop_map(nullptr)
195 , _stack(stack)
196 , _deoptimize_on_exception(deoptimize_on_exception)
197 , _force_reexecute(false) {
198 assert(_stack != nullptr, "must be non null");
199 }
200
201
202 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, ValueStack* stack)
203 : _scope_debug_info(nullptr)
204 , _scope(info->_scope)
205 , _exception_handlers(nullptr)
206 , _oop_map(nullptr)
207 , _stack(stack == nullptr ? info->_stack : stack)
208 , _deoptimize_on_exception(info->_deoptimize_on_exception)
209 , _force_reexecute(info->_force_reexecute) {
210
211 // deep copy of exception handlers
212 if (info->_exception_handlers != nullptr) {
213 _exception_handlers = new XHandlers(info->_exception_handlers);
214 }
215 }
216
217
218 void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset, bool maybe_return_as_fields) {
219 // record the safepoint before recording the debug info for enclosing scopes
220 recorder->add_safepoint(pc_offset, _oop_map->deep_copy());
221 bool reexecute = _force_reexecute || _scope_debug_info->should_reexecute();
222 _scope_debug_info->record_debug_info(recorder, pc_offset, reexecute, maybe_return_as_fields);
223 recorder->end_safepoint(pc_offset);
224 }
225
226
227 void CodeEmitInfo::add_register_oop(LIR_Opr opr) {
228 assert(_oop_map != nullptr, "oop map must already exist");
229 assert(opr->is_single_cpu(), "should not call otherwise");
230
231 VMReg name = frame_map()->regname(opr);
232 _oop_map->set_oop(name);
233 }
234
235 // Mirror the stack size calculation in the deopt code
236 // How much stack space would we need at this point in the program in
237 // case of deoptimization?
238 int CodeEmitInfo::interpreter_frame_size() const {
239 ValueStack* state = _stack;
240 int size = 0;
241 int callee_parameters = 0;
242 int callee_locals = 0;
243 int extra_args = state->scope()->method()->max_stack() - state->stack_size();
244
245 while (state != nullptr) {
246 int locks = state->locks_size();
247 int temps = state->stack_size();
248 bool is_top_frame = (state == _stack);
249 ciMethod* method = state->scope()->method();
250
251 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
252 temps + callee_parameters,
253 extra_args,
254 locks,
255 callee_parameters,
256 callee_locals,
257 is_top_frame);
258 size += frame_size;
259
260 callee_parameters = method->size_of_parameters();
261 callee_locals = method->max_locals();
262 extra_args = 0;
263 state = state->caller_state();
264 }
265 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
266 }
267
268 // Implementation of IR
269
270 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) :
271 _num_loops(0) {
272 // setup IR fields
273 _compilation = compilation;
274 _top_scope = new IRScope(compilation, nullptr, -1, method, osr_bci, true);
275 _code = nullptr;
276 }
277
278
279 void IR::optimize_blocks() {
280 Optimizer opt(this);
281 if (!compilation()->profile_branches()) {
282 if (DoCEE) {
283 opt.eliminate_conditional_expressions();
284 #ifndef PRODUCT
285 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); }
286 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); }
287 #endif
288 }
289 if (EliminateBlocks) {
290 opt.eliminate_blocks();
291 #ifndef PRODUCT
292 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); }
293 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); }
294 #endif
295 }
296 }
297 }
298
299 void IR::eliminate_null_checks() {
300 Optimizer opt(this);
301 if (EliminateNullChecks) {
302 opt.eliminate_null_checks();
303 #ifndef PRODUCT
304 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); }
305 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); }
306 #endif
307 }
308 }
309
310 // The functionality of this class is to insert a new block between
311 // the 'from' and 'to' block of a critical edge.
312 // It first collects the block pairs, and then processes them.
313 //
314 // Some instructions may introduce more than one edge between two blocks.
315 // By checking if the current 'to' block sets critical_edge_split_flag
316 // (all new blocks set this flag) we can avoid repeated processing.
317 // This is why BlockPair contains the index rather than the original 'to' block.
318 class CriticalEdgeFinder: public BlockClosure {
319 BlockPairList blocks;
320
321 public:
322 CriticalEdgeFinder(IR* ir) {
323 ir->iterate_preorder(this);
324 }
325
326 void block_do(BlockBegin* bb) {
327 BlockEnd* be = bb->end();
328 int nos = be->number_of_sux();
329 if (nos >= 2) {
330 for (int i = 0; i < nos; i++) {
331 BlockBegin* sux = be->sux_at(i);
332 if (sux->number_of_preds() >= 2) {
333 blocks.append(new BlockPair(bb, i));
334 }
335 }
336 }
337 }
338
339 void split_edges() {
340 for (int i = 0; i < blocks.length(); i++) {
341 BlockPair* pair = blocks.at(i);
342 BlockBegin* from = pair->from();
343 int index = pair->index();
344 BlockBegin* to = from->end()->sux_at(index);
345 if (to->is_set(BlockBegin::critical_edge_split_flag)) {
346 // inserted
347 continue;
348 }
349 BlockBegin* split = from->insert_block_between(to);
350 #ifndef PRODUCT
351 if ((PrintIR || PrintIR1) && Verbose) {
352 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)",
353 from->block_id(), to->block_id(), split->block_id());
354 }
355 #endif
356 }
357 }
358 };
359
360 void IR::split_critical_edges() {
361 CriticalEdgeFinder cef(this);
362 cef.split_edges();
363 }
364
365
366 class UseCountComputer: public ValueVisitor, BlockClosure {
367 private:
368 void visit(Value* n) {
369 // Local instructions and Phis for expression stack values at the
370 // start of basic blocks are not added to the instruction list
371 if (!(*n)->is_linked() && (*n)->can_be_linked()) {
372 assert(false, "a node was not appended to the graph");
373 Compilation::current()->bailout("a node was not appended to the graph");
374 }
375 // use n's input if not visited before
376 if (!(*n)->is_pinned() && !(*n)->has_uses()) {
377 // note: a) if the instruction is pinned, it will be handled by compute_use_count
378 // b) if the instruction has uses, it was touched before
379 // => in both cases we don't need to update n's values
380 uses_do(n);
381 }
382 // use n
383 (*n)->_use_count++;
384 }
385
386 Values* worklist;
387 int depth;
388 enum {
389 max_recurse_depth = 20
390 };
391
392 void uses_do(Value* n) {
393 depth++;
394 if (depth > max_recurse_depth) {
395 // don't allow the traversal to recurse too deeply
396 worklist->push(*n);
397 } else {
398 (*n)->input_values_do(this);
399 // special handling for some instructions
400 if ((*n)->as_BlockEnd() != nullptr) {
401 // note on BlockEnd:
402 // must 'use' the stack only if the method doesn't
403 // terminate, however, in those cases stack is empty
404 (*n)->state_values_do(this);
405 }
406 }
407 depth--;
408 }
409
410 void block_do(BlockBegin* b) {
411 depth = 0;
412 // process all pinned nodes as the roots of expression trees
413 for (Instruction* n = b; n != nullptr; n = n->next()) {
414 if (n->is_pinned()) uses_do(&n);
415 }
416 assert(depth == 0, "should have counted back down");
417
418 // now process any unpinned nodes which recursed too deeply
419 while (worklist->length() > 0) {
420 Value t = worklist->pop();
421 if (!t->is_pinned()) {
422 // compute the use count
423 uses_do(&t);
424
425 // pin the instruction so that LIRGenerator doesn't recurse
426 // too deeply during it's evaluation.
427 t->pin();
428 }
429 }
430 assert(depth == 0, "should have counted back down");
431 }
432
433 UseCountComputer() {
434 worklist = new Values();
435 depth = 0;
436 }
437
438 public:
439 static void compute(BlockList* blocks) {
440 UseCountComputer ucc;
441 blocks->iterate_backward(&ucc);
442 }
443 };
444
445
446 // helper macro for short definition of trace-output inside code
447 #ifdef ASSERT
448 #define TRACE_LINEAR_SCAN(level, code) \
449 if (TraceLinearScanLevel >= level) { \
450 code; \
451 }
452 #else
453 #define TRACE_LINEAR_SCAN(level, code)
454 #endif
455
456 class ComputeLinearScanOrder : public StackObj {
457 private:
458 int _max_block_id; // the highest block_id of a block
459 int _num_blocks; // total number of blocks (smaller than _max_block_id)
460 int _num_loops; // total number of loops
461 bool _iterative_dominators;// method requires iterative computation of dominatiors
462
463 BlockList* _linear_scan_order; // the resulting list of blocks in correct order
464
465 ResourceBitMap _visited_blocks; // used for recursive processing of blocks
466 ResourceBitMap _active_blocks; // used for recursive processing of blocks
467 ResourceBitMap _dominator_blocks; // temporary BitMap used for computation of dominator
468 intArray _forward_branches; // number of incoming forward branches for each block
469 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges
470 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop
471 BlockList _work_list; // temporary list (used in mark_loops and compute_order)
472 BlockList _loop_headers;
473
474 Compilation* _compilation;
475
476 // accessors for _visited_blocks and _active_blocks
477 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); }
478 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); }
479 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); }
480 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); }
481 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); }
482 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); }
483
484 // accessors for _forward_branches
485 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); }
486 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()); }
487
488 // accessors for _loop_map
489 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); }
490 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); }
491 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); }
492
493 // count edges between blocks
494 void count_edges(BlockBegin* cur, BlockBegin* parent);
495
496 // loop detection
497 void mark_loops();
498 void clear_non_natural_loops(BlockBegin* start_block);
499 void assign_loop_depth(BlockBegin* start_block);
500
501 // computation of final block order
502 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b);
503 void compute_dominator(BlockBegin* cur, BlockBegin* parent);
504 void compute_dominator_impl(BlockBegin* cur, BlockBegin* parent);
505 int compute_weight(BlockBegin* cur);
506 bool ready_for_processing(BlockBegin* cur);
507 void sort_into_work_list(BlockBegin* b);
508 void append_block(BlockBegin* cur);
509 void compute_order(BlockBegin* start_block);
510
511 // fixup of dominators for non-natural loops
512 bool compute_dominators_iter();
513 void compute_dominators();
514
515 // debug functions
516 DEBUG_ONLY(void print_blocks();)
517 DEBUG_ONLY(void verify();)
518
519 Compilation* compilation() const { return _compilation; }
520 public:
521 ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block);
522
523 // accessors for final result
524 BlockList* linear_scan_order() const { return _linear_scan_order; }
525 int num_loops() const { return _num_loops; }
526 };
527
528
529 ComputeLinearScanOrder::ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block) :
530 _max_block_id(BlockBegin::number_of_blocks()),
531 _num_blocks(0),
532 _num_loops(0),
533 _iterative_dominators(false),
534 _linear_scan_order(nullptr), // initialized later with correct size
535 _visited_blocks(_max_block_id),
536 _active_blocks(_max_block_id),
537 _dominator_blocks(_max_block_id),
538 _forward_branches(_max_block_id, _max_block_id, 0),
539 _loop_end_blocks(8),
540 _loop_map(0), // initialized later with correct size
541 _work_list(8),
542 _compilation(c)
543 {
544 TRACE_LINEAR_SCAN(2, tty->print_cr("***** computing linear-scan block order"));
545
546 count_edges(start_block, nullptr);
547
548 if (compilation()->is_profiling()) {
549 ciMethod *method = compilation()->method();
550 if (!method->is_accessor()) {
551 ciMethodData* md = method->method_data_or_null();
552 assert(md != nullptr, "Sanity");
553 md->set_compilation_stats(_num_loops, _num_blocks);
554 }
555 }
556
557 if (_num_loops > 0) {
558 mark_loops();
559 clear_non_natural_loops(start_block);
560 assign_loop_depth(start_block);
561 }
562
563 compute_order(start_block);
564 compute_dominators();
565
566 DEBUG_ONLY(print_blocks());
567 DEBUG_ONLY(verify());
568 }
569
570
571 // Traverse the CFG:
572 // * count total number of blocks
573 // * count all incoming edges and backward incoming edges
574 // * number loop header blocks
575 // * create a list with all loop end blocks
576 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) {
577 TRACE_LINEAR_SCAN(3, tty->print_cr("Enter count_edges for block B%d coming from B%d", cur->block_id(), parent != nullptr ? parent->block_id() : -1));
578 assert(cur->dominator() == nullptr, "dominator already initialized");
579
580 if (is_active(cur)) {
581 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch"));
582 assert(is_visited(cur), "block must be visisted when block is active");
583 assert(parent != nullptr, "must have parent");
584
585 cur->set(BlockBegin::backward_branch_target_flag);
586
587 // When a loop header is also the start of an exception handler, then the backward branch is
588 // an exception edge. Because such edges are usually critical edges which cannot be split, the
589 // loop must be excluded here from processing.
590 if (cur->is_set(BlockBegin::exception_entry_flag)) {
591 // Make sure that dominators are correct in this weird situation
592 _iterative_dominators = true;
593 return;
594 }
595
596 cur->set(BlockBegin::linear_scan_loop_header_flag);
597 parent->set(BlockBegin::linear_scan_loop_end_flag);
598
599 assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur,
600 "loop end blocks must have one successor (critical edges are split)");
601
602 _loop_end_blocks.append(parent);
603 return;
604 }
605
606 // increment number of incoming forward branches
607 inc_forward_branches(cur);
608
609 if (is_visited(cur)) {
610 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited"));
611 return;
612 }
613
614 _num_blocks++;
615 set_visited(cur);
616 set_active(cur);
617
618 // recursive call for all successors
619 int i;
620 for (i = cur->number_of_sux() - 1; i >= 0; i--) {
621 count_edges(cur->sux_at(i), cur);
622 }
623 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
624 count_edges(cur->exception_handler_at(i), cur);
625 }
626
627 clear_active(cur);
628
629 // Each loop has a unique number.
630 // When multiple loops are nested, assign_loop_depth assumes that the
631 // innermost loop has the lowest number. This is guaranteed by setting
632 // the loop number after the recursive calls for the successors above
633 // have returned.
634 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) {
635 assert(cur->loop_index() == -1, "cannot set loop-index twice");
636 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops));
637
638 cur->set_loop_index(_num_loops);
639 _loop_headers.append(cur);
640 _num_loops++;
641 }
642
643 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id()));
644 }
645
646
647 void ComputeLinearScanOrder::mark_loops() {
648 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops"));
649
650 _loop_map = BitMap2D(_num_loops, _max_block_id);
651
652 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) {
653 BlockBegin* loop_end = _loop_end_blocks.at(i);
654 BlockBegin* loop_start = loop_end->sux_at(0);
655 int loop_idx = loop_start->loop_index();
656
657 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));
658 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set");
659 assert(loop_end->number_of_sux() == 1, "incorrect number of successors");
660 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set");
661 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set");
662 assert(_work_list.is_empty(), "work list must be empty before processing");
663
664 // add the end-block of the loop to the working list
665 _work_list.push(loop_end);
666 set_block_in_loop(loop_idx, loop_end);
667 do {
668 BlockBegin* cur = _work_list.pop();
669
670 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id()));
671 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list");
672
673 // recursive processing of all predecessors ends when start block of loop is reached
674 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) {
675 for (int j = cur->number_of_preds() - 1; j >= 0; j--) {
676 BlockBegin* pred = cur->pred_at(j);
677
678 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) {
679 // this predecessor has not been processed yet, so add it to work list
680 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id()));
681 _work_list.push(pred);
682 set_block_in_loop(loop_idx, pred);
683 }
684 }
685 }
686 } while (!_work_list.is_empty());
687 }
688 }
689
690
691 // check for non-natural loops (loops where the loop header does not dominate
692 // all other loop blocks = loops with multiple entries).
693 // such loops are ignored
694 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) {
695 for (int i = _num_loops - 1; i >= 0; i--) {
696 if (is_block_in_loop(i, start_block)) {
697 // loop i contains the entry block of the method
698 // -> this is not a natural loop, so ignore it
699 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i));
700
701 BlockBegin *loop_header = _loop_headers.at(i);
702 assert(loop_header->is_set(BlockBegin::linear_scan_loop_header_flag), "Must be loop header");
703
704 for (int j = 0; j < loop_header->number_of_preds(); j++) {
705 BlockBegin *pred = loop_header->pred_at(j);
706 pred->clear(BlockBegin::linear_scan_loop_end_flag);
707 }
708
709 loop_header->clear(BlockBegin::linear_scan_loop_header_flag);
710
711 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) {
712 clear_block_in_loop(i, block_id);
713 }
714 _iterative_dominators = true;
715 }
716 }
717 }
718
719 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) {
720 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing loop-depth and weight"));
721 init_visited();
722
723 assert(_work_list.is_empty(), "work list must be empty before processing");
724 _work_list.append(start_block);
725
726 do {
727 BlockBegin* cur = _work_list.pop();
728
729 if (!is_visited(cur)) {
730 set_visited(cur);
731 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id()));
732
733 // compute loop-depth and loop-index for the block
734 assert(cur->loop_depth() == 0, "cannot set loop-depth twice");
735 int i;
736 int loop_depth = 0;
737 int min_loop_idx = -1;
738 for (i = _num_loops - 1; i >= 0; i--) {
739 if (is_block_in_loop(i, cur)) {
740 loop_depth++;
741 min_loop_idx = i;
742 }
743 }
744 cur->set_loop_depth(loop_depth);
745 cur->set_loop_index(min_loop_idx);
746
747 // append all unvisited successors to work list
748 for (i = cur->number_of_sux() - 1; i >= 0; i--) {
749 _work_list.append(cur->sux_at(i));
750 }
751 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
752 _work_list.append(cur->exception_handler_at(i));
753 }
754 }
755 } while (!_work_list.is_empty());
756 }
757
758
759 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) {
760 assert(a != nullptr && b != nullptr, "must have input blocks");
761
762 _dominator_blocks.clear();
763 while (a != nullptr) {
764 _dominator_blocks.set_bit(a->block_id());
765 assert(a->dominator() != nullptr || a == _linear_scan_order->at(0), "dominator must be initialized");
766 a = a->dominator();
767 }
768 while (b != nullptr && !_dominator_blocks.at(b->block_id())) {
769 assert(b->dominator() != nullptr || b == _linear_scan_order->at(0), "dominator must be initialized");
770 b = b->dominator();
771 }
772
773 assert(b != nullptr, "could not find dominator");
774 return b;
775 }
776
777 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) {
778 init_visited();
779 compute_dominator_impl(cur, parent);
780 }
781
782 void ComputeLinearScanOrder::compute_dominator_impl(BlockBegin* cur, BlockBegin* parent) {
783 // Mark as visited to avoid recursive calls with same parent
784 set_visited(cur);
785
786 if (cur->dominator() == nullptr) {
787 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id()));
788 cur->set_dominator(parent);
789
790 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) {
791 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()));
792 // Does not hold for exception blocks
793 assert(cur->number_of_preds() > 1 || cur->is_set(BlockBegin::exception_entry_flag), "");
794 cur->set_dominator(common_dominator(cur->dominator(), parent));
795 }
796
797 // Additional edge to xhandler of all our successors
798 // range check elimination needs that the state at the end of a
799 // block be valid in every block it dominates so cur must dominate
800 // the exception handlers of its successors.
801 int num_cur_xhandler = cur->number_of_exception_handlers();
802 for (int j = 0; j < num_cur_xhandler; j++) {
803 BlockBegin* xhandler = cur->exception_handler_at(j);
804 if (!is_visited(xhandler)) {
805 compute_dominator_impl(xhandler, parent);
806 }
807 }
808 }
809
810
811 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) {
812 BlockBegin* single_sux = nullptr;
813 if (cur->number_of_sux() == 1) {
814 single_sux = cur->sux_at(0);
815 }
816
817 // limit loop-depth to 15 bit (only for security reason, it will never be so big)
818 int weight = (cur->loop_depth() & 0x7FFF) << 16;
819
820 // general macro for short definition of weight flags
821 // the first instance of INC_WEIGHT_IF has the highest priority
822 int cur_bit = 15;
823 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--;
824
825 // this is necessary for the (very rare) case that two successive blocks have
826 // the same loop depth, but a different loop index (can happen for endless loops
827 // with exception handlers)
828 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag));
829
830 // loop end blocks (blocks that end with a backward branch) are added
831 // after all other blocks of the loop.
832 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag));
833
834 // critical edge split blocks are preferred because than they have a bigger
835 // proability to be completely empty
836 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag));
837
838 // exceptions should not be thrown in normal control flow, so these blocks
839 // are added as late as possible
840 INC_WEIGHT_IF(cur->end()->as_Throw() == nullptr && (single_sux == nullptr || single_sux->end()->as_Throw() == nullptr));
841 INC_WEIGHT_IF(cur->end()->as_Return() == nullptr && (single_sux == nullptr || single_sux->end()->as_Return() == nullptr));
842
843 // exceptions handlers are added as late as possible
844 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag));
845
846 // guarantee that weight is > 0
847 weight |= 1;
848
849 #undef INC_WEIGHT_IF
850 assert(cur_bit >= 0, "too many flags");
851 assert(weight > 0, "weight cannot become negative");
852
853 return weight;
854 }
855
856 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) {
857 // Discount the edge just traveled.
858 // When the number drops to zero, all forward branches were processed
859 if (dec_forward_branches(cur) != 0) {
860 return false;
861 }
862
863 assert(_linear_scan_order->find(cur) == -1, "block already processed (block can be ready only once)");
864 assert(_work_list.find(cur) == -1, "block already in work-list (block can be ready only once)");
865 return true;
866 }
867
868 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) {
869 assert(_work_list.find(cur) == -1, "block already in work list");
870
871 int cur_weight = compute_weight(cur);
872
873 // the linear_scan_number is used to cache the weight of a block
874 cur->set_linear_scan_number(cur_weight);
875
876 #ifndef PRODUCT
877 if (StressLinearScan) {
878 _work_list.insert_before(0, cur);
879 return;
880 }
881 #endif
882
883 _work_list.append(nullptr); // provide space for new element
884
885 int insert_idx = _work_list.length() - 1;
886 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) {
887 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1));
888 insert_idx--;
889 }
890 _work_list.at_put(insert_idx, cur);
891
892 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id()));
893 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()));
894
895 #ifdef ASSERT
896 for (int i = 0; i < _work_list.length(); i++) {
897 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set");
898 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist");
899 }
900 #endif
901 }
902
903 void ComputeLinearScanOrder::append_block(BlockBegin* cur) {
904 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()));
905 assert(_linear_scan_order->find(cur) == -1, "cannot add the same block twice");
906
907 // currently, the linear scan order and code emit order are equal.
908 // therefore the linear_scan_number and the weight of a block must also
909 // be equal.
910 cur->set_linear_scan_number(_linear_scan_order->length());
911 _linear_scan_order->append(cur);
912 }
913
914 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) {
915 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing final block order"));
916
917 // the start block is always the first block in the linear scan order
918 _linear_scan_order = new BlockList(_num_blocks);
919 append_block(start_block);
920
921 assert(start_block->end()->as_Base() != nullptr, "start block must end with Base-instruction");
922 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry();
923 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry();
924
925 BlockBegin* sux_of_osr_entry = nullptr;
926 if (osr_entry != nullptr) {
927 // special handling for osr entry:
928 // ignore the edge between the osr entry and its successor for processing
929 // the osr entry block is added manually below
930 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor");
931 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");
932
933 sux_of_osr_entry = osr_entry->sux_at(0);
934 dec_forward_branches(sux_of_osr_entry);
935
936 compute_dominator(osr_entry, start_block);
937 _iterative_dominators = true;
938 }
939 compute_dominator(std_entry, start_block);
940
941 // start processing with standard entry block
942 assert(_work_list.is_empty(), "list must be empty before processing");
943
944 if (ready_for_processing(std_entry)) {
945 sort_into_work_list(std_entry);
946 } else {
947 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)");
948 }
949
950 do {
951 BlockBegin* cur = _work_list.pop();
952
953 if (cur == sux_of_osr_entry) {
954 // the osr entry block is ignored in normal processing, it is never added to the
955 // work list. Instead, it is added as late as possible manually here.
956 append_block(osr_entry);
957 compute_dominator(cur, osr_entry);
958 }
959 append_block(cur);
960
961 int i;
962 int num_sux = cur->number_of_sux();
963 // changed loop order to get "intuitive" order of if- and else-blocks
964 for (i = 0; i < num_sux; i++) {
965 BlockBegin* sux = cur->sux_at(i);
966 compute_dominator(sux, cur);
967 if (ready_for_processing(sux)) {
968 sort_into_work_list(sux);
969 }
970 }
971 num_sux = cur->number_of_exception_handlers();
972 for (i = 0; i < num_sux; i++) {
973 BlockBegin* sux = cur->exception_handler_at(i);
974 if (ready_for_processing(sux)) {
975 sort_into_work_list(sux);
976 }
977 }
978 } while (_work_list.length() > 0);
979 }
980
981
982 bool ComputeLinearScanOrder::compute_dominators_iter() {
983 bool changed = false;
984 int num_blocks = _linear_scan_order->length();
985
986 assert(_linear_scan_order->at(0)->dominator() == nullptr, "must not have dominator");
987 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors");
988 for (int i = 1; i < num_blocks; i++) {
989 BlockBegin* block = _linear_scan_order->at(i);
990
991 BlockBegin* dominator = block->pred_at(0);
992 int num_preds = block->number_of_preds();
993
994 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: Processing B%d", block->block_id()));
995
996 for (int j = 0; j < num_preds; j++) {
997
998 BlockBegin *pred = block->pred_at(j);
999 TRACE_LINEAR_SCAN(4, tty->print_cr(" DOM: Subrocessing B%d", pred->block_id()));
1000
1001 if (block->is_set(BlockBegin::exception_entry_flag)) {
1002 dominator = common_dominator(dominator, pred);
1003 int num_pred_preds = pred->number_of_preds();
1004 for (int k = 0; k < num_pred_preds; k++) {
1005 dominator = common_dominator(dominator, pred->pred_at(k));
1006 }
1007 } else {
1008 dominator = common_dominator(dominator, pred);
1009 }
1010 }
1011
1012 if (dominator != block->dominator()) {
1013 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()));
1014
1015 block->set_dominator(dominator);
1016 changed = true;
1017 }
1018 }
1019 return changed;
1020 }
1021
1022 void ComputeLinearScanOrder::compute_dominators() {
1023 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators));
1024
1025 // iterative computation of dominators is only required for methods with non-natural loops
1026 // and OSR-methods. For all other methods, the dominators computed when generating the
1027 // linear scan block order are correct.
1028 if (_iterative_dominators) {
1029 do {
1030 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation"));
1031 } while (compute_dominators_iter());
1032 }
1033
1034 // check that dominators are correct
1035 assert(!compute_dominators_iter(), "fix point not reached");
1036
1037 // Add Blocks to dominates-Array
1038 int num_blocks = _linear_scan_order->length();
1039 for (int i = 0; i < num_blocks; i++) {
1040 BlockBegin* block = _linear_scan_order->at(i);
1041
1042 BlockBegin *dom = block->dominator();
1043 if (dom) {
1044 assert(dom->dominator_depth() != -1, "Dominator must have been visited before");
1045 dom->dominates()->append(block);
1046 block->set_dominator_depth(dom->dominator_depth() + 1);
1047 } else {
1048 block->set_dominator_depth(0);
1049 }
1050 }
1051 }
1052
1053
1054 #ifdef ASSERT
1055 void ComputeLinearScanOrder::print_blocks() {
1056 if (TraceLinearScanLevel >= 2) {
1057 tty->print_cr("----- loop information:");
1058 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
1059 BlockBegin* cur = _linear_scan_order->at(block_idx);
1060
1061 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id());
1062 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
1063 tty->print ("%d ", is_block_in_loop(loop_idx, cur));
1064 }
1065 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth());
1066 }
1067 }
1068
1069 if (TraceLinearScanLevel >= 1) {
1070 tty->print_cr("----- linear-scan block order:");
1071 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
1072 BlockBegin* cur = _linear_scan_order->at(block_idx);
1073 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth());
1074
1075 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " ");
1076 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " ");
1077 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " ");
1078 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " ");
1079
1080 if (cur->dominator() != nullptr) {
1081 tty->print(" dom: B%d ", cur->dominator()->block_id());
1082 } else {
1083 tty->print(" dom: null ");
1084 }
1085
1086 if (cur->number_of_preds() > 0) {
1087 tty->print(" preds: ");
1088 for (int j = 0; j < cur->number_of_preds(); j++) {
1089 BlockBegin* pred = cur->pred_at(j);
1090 tty->print("B%d ", pred->block_id());
1091 }
1092 }
1093 if (cur->number_of_sux() > 0) {
1094 tty->print(" sux: ");
1095 for (int j = 0; j < cur->number_of_sux(); j++) {
1096 BlockBegin* sux = cur->sux_at(j);
1097 tty->print("B%d ", sux->block_id());
1098 }
1099 }
1100 if (cur->number_of_exception_handlers() > 0) {
1101 tty->print(" ex: ");
1102 for (int j = 0; j < cur->number_of_exception_handlers(); j++) {
1103 BlockBegin* ex = cur->exception_handler_at(j);
1104 tty->print("B%d ", ex->block_id());
1105 }
1106 }
1107 tty->cr();
1108 }
1109 }
1110 }
1111
1112 void ComputeLinearScanOrder::verify() {
1113 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list");
1114
1115 if (StressLinearScan) {
1116 // blocks are scrambled when StressLinearScan is used
1117 return;
1118 }
1119
1120 // check that all successors of a block have a higher linear-scan-number
1121 // and that all predecessors of a block have a lower linear-scan-number
1122 // (only backward branches of loops are ignored)
1123 int i;
1124 for (i = 0; i < _linear_scan_order->length(); i++) {
1125 BlockBegin* cur = _linear_scan_order->at(i);
1126
1127 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number");
1128 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->find(cur), "incorrect linear_scan_number");
1129
1130 int j;
1131 for (j = cur->number_of_sux() - 1; j >= 0; j--) {
1132 BlockBegin* sux = cur->sux_at(j);
1133
1134 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->find(sux), "incorrect linear_scan_number");
1135 if (!sux->is_set(BlockBegin::backward_branch_target_flag)) {
1136 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order");
1137 }
1138 if (cur->loop_depth() == sux->loop_depth()) {
1139 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");
1140 }
1141 }
1142
1143 for (j = cur->number_of_preds() - 1; j >= 0; j--) {
1144 BlockBegin* pred = cur->pred_at(j);
1145
1146 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->find(pred), "incorrect linear_scan_number");
1147 if (!cur->is_set(BlockBegin::backward_branch_target_flag)) {
1148 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order");
1149 }
1150 if (cur->loop_depth() == pred->loop_depth()) {
1151 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");
1152 }
1153
1154 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors");
1155 }
1156
1157 // check dominator
1158 if (i == 0) {
1159 assert(cur->dominator() == nullptr, "first block has no dominator");
1160 } else {
1161 assert(cur->dominator() != nullptr, "all but first block must have dominator");
1162 }
1163 // Assertion does not hold for exception handlers
1164 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");
1165 }
1166
1167 // check that all loops are continuous
1168 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
1169 int block_idx = 0;
1170 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop");
1171
1172 // skip blocks before the loop
1173 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
1174 block_idx++;
1175 }
1176 // skip blocks of loop
1177 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
1178 block_idx++;
1179 }
1180 // after the first non-loop block, there must not be another loop-block
1181 while (block_idx < _num_blocks) {
1182 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order");
1183 block_idx++;
1184 }
1185 }
1186 }
1187 #endif // ASSERT
1188
1189
1190 void IR::compute_code() {
1191 assert(is_valid(), "IR must be valid");
1192
1193 ComputeLinearScanOrder compute_order(compilation(), start());
1194 _num_loops = compute_order.num_loops();
1195 _code = compute_order.linear_scan_order();
1196 }
1197
1198
1199 void IR::compute_use_counts() {
1200 // make sure all values coming out of this block get evaluated.
1201 int num_blocks = _code->length();
1202 for (int i = 0; i < num_blocks; i++) {
1203 _code->at(i)->end()->state()->pin_stack_for_linear_scan();
1204 }
1205
1206 // compute use counts
1207 UseCountComputer::compute(_code);
1208 }
1209
1210
1211 void IR::iterate_preorder(BlockClosure* closure) {
1212 assert(is_valid(), "IR must be valid");
1213 start()->iterate_preorder(closure);
1214 }
1215
1216
1217 void IR::iterate_postorder(BlockClosure* closure) {
1218 assert(is_valid(), "IR must be valid");
1219 start()->iterate_postorder(closure);
1220 }
1221
1222 void IR::iterate_linear_scan_order(BlockClosure* closure) {
1223 linear_scan_order()->iterate_forward(closure);
1224 }
1225
1226
1227 #ifndef PRODUCT
1228 class BlockPrinter: public BlockClosure {
1229 private:
1230 InstructionPrinter* _ip;
1231 bool _cfg_only;
1232 bool _live_only;
1233
1234 public:
1235 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) {
1236 _ip = ip;
1237 _cfg_only = cfg_only;
1238 _live_only = live_only;
1239 }
1240
1241 virtual void block_do(BlockBegin* block) {
1242 if (_cfg_only) {
1243 _ip->print_instr(block); tty->cr();
1244 } else {
1245 block->print_block(*_ip, _live_only);
1246 }
1247 }
1248 };
1249
1250
1251 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) {
1252 ttyLocker ttyl;
1253 InstructionPrinter ip(!cfg_only);
1254 BlockPrinter bp(&ip, cfg_only, live_only);
1255 start->iterate_preorder(&bp);
1256 tty->cr();
1257 }
1258
1259 void IR::print(bool cfg_only, bool live_only) {
1260 if (is_valid()) {
1261 print(start(), cfg_only, live_only);
1262 } else {
1263 tty->print_cr("invalid IR");
1264 }
1265 }
1266 #endif // PRODUCT
1267
1268 #ifdef ASSERT
1269 class EndNotNullValidator : public BlockClosure {
1270 public:
1271 virtual void block_do(BlockBegin* block) {
1272 assert(block->end() != nullptr, "Expect block end to exist.");
1273 }
1274 };
1275
1276 class XentryFlagValidator : public BlockClosure {
1277 public:
1278 virtual void block_do(BlockBegin* block) {
1279 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1280 assert(!block->end()->sux_at(i)->is_set(BlockBegin::exception_entry_flag), "must not be xhandler");
1281 }
1282 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1283 assert(block->exception_handler_at(i)->is_set(BlockBegin::exception_entry_flag), "must be xhandler");
1284 }
1285 }
1286 };
1287
1288 typedef GrowableArray<BlockList*> BlockListList;
1289
1290 // Validation goals:
1291 // - code() length == blocks length
1292 // - code() contents == blocks content
1293 // - Each block's computed predecessors match sux lists (length)
1294 // - Each block's computed predecessors match sux lists (set content)
1295 class PredecessorAndCodeValidator : public BlockClosure {
1296 private:
1297 BlockListList* _predecessors; // Each index i will hold predecessors of block with id i
1298 BlockList* _blocks;
1299
1300 static int cmp(BlockBegin** a, BlockBegin** b) {
1301 return (*a)->block_id() - (*b)->block_id();
1302 }
1303
1304 public:
1305 PredecessorAndCodeValidator(IR* hir) {
1306 ResourceMark rm;
1307 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), BlockBegin::number_of_blocks(), nullptr);
1308 _blocks = new BlockList(BlockBegin::number_of_blocks());
1309
1310 hir->start()->iterate_preorder(this);
1311 if (hir->code() != nullptr) {
1312 assert(hir->code()->length() == _blocks->length(), "must match");
1313 for (int i = 0; i < _blocks->length(); i++) {
1314 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists");
1315 }
1316 }
1317
1318 for (int i = 0; i < _blocks->length(); i++) {
1319 BlockBegin* block = _blocks->at(i);
1320 verify_block_preds_against_collected_preds(block);
1321 }
1322 }
1323
1324 virtual void block_do(BlockBegin* block) {
1325 _blocks->append(block);
1326 collect_predecessors(block);
1327 }
1328
1329 private:
1330 void collect_predecessors(BlockBegin* block) {
1331 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1332 collect_predecessor(block, block->end()->sux_at(i));
1333 }
1334 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1335 collect_predecessor(block, block->exception_handler_at(i));
1336 }
1337 }
1338
1339 void collect_predecessor(BlockBegin* const pred, const BlockBegin* sux) {
1340 BlockList* preds = _predecessors->at_grow(sux->block_id(), nullptr);
1341 if (preds == nullptr) {
1342 preds = new BlockList();
1343 _predecessors->at_put(sux->block_id(), preds);
1344 }
1345 preds->append(pred);
1346 }
1347
1348 void verify_block_preds_against_collected_preds(const BlockBegin* block) const {
1349 BlockList* preds = _predecessors->at(block->block_id());
1350 if (preds == nullptr) {
1351 assert(block->number_of_preds() == 0, "should be the same");
1352 return;
1353 }
1354 assert(preds->length() == block->number_of_preds(), "should be the same");
1355
1356 // clone the pred list so we can mutate it
1357 BlockList* pred_copy = new BlockList();
1358 for (int j = 0; j < block->number_of_preds(); j++) {
1359 pred_copy->append(block->pred_at(j));
1360 }
1361 // sort them in the same order
1362 preds->sort(cmp);
1363 pred_copy->sort(cmp);
1364 for (int j = 0; j < block->number_of_preds(); j++) {
1365 assert(preds->at(j) == pred_copy->at(j), "must match");
1366 }
1367 }
1368 };
1369
1370 class VerifyBlockBeginField : public BlockClosure {
1371 public:
1372 virtual void block_do(BlockBegin* block) {
1373 for (Instruction* cur = block; cur != nullptr; cur = cur->next()) {
1374 assert(cur->block() == block, "Block begin is not correct");
1375 }
1376 }
1377 };
1378
1379 class ValidateEdgeMutuality : public BlockClosure {
1380 public:
1381 virtual void block_do(BlockBegin* block) {
1382 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1383 assert(block->end()->sux_at(i)->is_predecessor(block), "Block's successor should have it as predecessor");
1384 }
1385
1386 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1387 assert(block->exception_handler_at(i)->is_predecessor(block), "Block's exception handler should have it as predecessor");
1388 }
1389
1390 for (int i = 0; i < block->number_of_preds(); i++) {
1391 assert(block->pred_at(i) != nullptr, "Predecessor must exist");
1392 assert(block->pred_at(i)->end() != nullptr, "Predecessor end must exist");
1393 bool is_sux = block->pred_at(i)->end()->is_sux(block);
1394 bool is_xhandler = block->pred_at(i)->is_exception_handler(block);
1395 assert(is_sux || is_xhandler, "Block's predecessor should have it as successor or xhandler");
1396 }
1397 }
1398 };
1399
1400 void IR::expand_with_neighborhood(BlockList& blocks) {
1401 int original_size = blocks.length();
1402 for (int h = 0; h < original_size; h++) {
1403 BlockBegin* block = blocks.at(h);
1404
1405 for (int i = 0; i < block->end()->number_of_sux(); i++) {
1406 if (!blocks.contains(block->end()->sux_at(i))) {
1407 blocks.append(block->end()->sux_at(i));
1408 }
1409 }
1410
1411 for (int i = 0; i < block->number_of_preds(); i++) {
1412 if (!blocks.contains(block->pred_at(i))) {
1413 blocks.append(block->pred_at(i));
1414 }
1415 }
1416
1417 for (int i = 0; i < block->number_of_exception_handlers(); i++) {
1418 if (!blocks.contains(block->exception_handler_at(i))) {
1419 blocks.append(block->exception_handler_at(i));
1420 }
1421 }
1422 }
1423 }
1424
1425 void IR::verify_local(BlockList& blocks) {
1426 EndNotNullValidator ennv;
1427 blocks.iterate_forward(&ennv);
1428
1429 ValidateEdgeMutuality vem;
1430 blocks.iterate_forward(&vem);
1431
1432 VerifyBlockBeginField verifier;
1433 blocks.iterate_forward(&verifier);
1434 }
1435
1436 void IR::verify() {
1437 XentryFlagValidator xe;
1438 iterate_postorder(&xe);
1439
1440 PredecessorAndCodeValidator pv(this);
1441
1442 EndNotNullValidator ennv;
1443 iterate_postorder(&ennv);
1444
1445 ValidateEdgeMutuality vem;
1446 iterate_postorder(&vem);
1447
1448 VerifyBlockBeginField verifier;
1449 iterate_postorder(&verifier);
1450 }
1451 #endif // ASSERT
1452
1453 void SubstitutionResolver::visit(Value* v) {
1454 Value v0 = *v;
1455 if (v0) {
1456 Value vs = v0->subst();
1457 if (vs != v0) {
1458 *v = v0->subst();
1459 }
1460 }
1461 }
1462
1463 #ifdef ASSERT
1464 class SubstitutionChecker: public ValueVisitor {
1465 void visit(Value* v) {
1466 Value v0 = *v;
1467 if (v0) {
1468 Value vs = v0->subst();
1469 assert(vs == v0, "missed substitution");
1470 }
1471 }
1472 };
1473 #endif
1474
1475
1476 void SubstitutionResolver::block_do(BlockBegin* block) {
1477 Instruction* last = nullptr;
1478 for (Instruction* n = block; n != nullptr;) {
1479 n->values_do(this);
1480 // need to remove this instruction from the instruction stream
1481 if (n->subst() != n) {
1482 guarantee(last != nullptr, "must have last");
1483 last->set_next(n->next());
1484 } else {
1485 last = n;
1486 }
1487 n = last->next();
1488 }
1489
1490 #ifdef ASSERT
1491 SubstitutionChecker check_substitute;
1492 if (block->state()) block->state()->values_do(&check_substitute);
1493 block->block_values_do(&check_substitute);
1494 if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute);
1495 #endif
1496 }