1 /*
2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm_io.h"
27 #include "asm/macroAssembler.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "ci/ciReplay.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "code/exceptionHandlerTable.hpp"
32 #include "code/nmethod.hpp"
33 #include "compiler/compileBroker.hpp"
34 #include "compiler/compileLog.hpp"
35 #include "compiler/disassembler.hpp"
36 #include "compiler/oopMap.hpp"
37 #include "gc/shared/barrierSet.hpp"
38 #include "gc/shared/c2/barrierSetC2.hpp"
39 #include "jfr/jfrEvents.hpp"
40 #include "memory/resourceArea.hpp"
41 #include "opto/addnode.hpp"
42 #include "opto/block.hpp"
43 #include "opto/c2compiler.hpp"
44 #include "opto/callGenerator.hpp"
45 #include "opto/callnode.hpp"
46 #include "opto/castnode.hpp"
47 #include "opto/cfgnode.hpp"
48 #include "opto/chaitin.hpp"
49 #include "opto/compile.hpp"
50 #include "opto/connode.hpp"
51 #include "opto/convertnode.hpp"
52 #include "opto/divnode.hpp"
53 #include "opto/escape.hpp"
54 #include "opto/idealGraphPrinter.hpp"
55 #include "opto/loopnode.hpp"
56 #include "opto/machnode.hpp"
57 #include "opto/macro.hpp"
58 #include "opto/matcher.hpp"
59 #include "opto/mathexactnode.hpp"
60 #include "opto/memnode.hpp"
61 #include "opto/mulnode.hpp"
62 #include "opto/narrowptrnode.hpp"
63 #include "opto/node.hpp"
64 #include "opto/opcodes.hpp"
65 #include "opto/output.hpp"
66 #include "opto/parse.hpp"
67 #include "opto/phaseX.hpp"
68 #include "opto/rootnode.hpp"
69 #include "opto/runtime.hpp"
70 #include "opto/stringopts.hpp"
71 #include "opto/type.hpp"
72 #include "opto/vector.hpp"
73 #include "opto/vectornode.hpp"
74 #include "runtime/globals_extension.hpp"
75 #include "runtime/sharedRuntime.hpp"
76 #include "runtime/signature.hpp"
77 #include "runtime/stubRoutines.hpp"
78 #include "runtime/timer.hpp"
79 #include "utilities/align.hpp"
80 #include "utilities/copy.hpp"
81 #include "utilities/macros.hpp"
82 #include "utilities/resourceHash.hpp"
83
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == nullptr) {
89 _mach_constant_base_node = new MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 class IntrinsicDescPair {
101 private:
102 ciMethod* _m;
103 bool _is_virtual;
104 public:
105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
107 ciMethod* m= elt->method();
108 ciMethod* key_m = key->_m;
109 if (key_m < m) return -1;
110 else if (key_m > m) return 1;
111 else {
112 bool is_virtual = elt->is_virtual();
113 bool key_virtual = key->_is_virtual;
114 if (key_virtual < is_virtual) return -1;
115 else if (key_virtual > is_virtual) return 1;
116 else return 0;
117 }
118 }
119 };
120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
121 #ifdef ASSERT
122 for (int i = 1; i < _intrinsics.length(); i++) {
123 CallGenerator* cg1 = _intrinsics.at(i-1);
124 CallGenerator* cg2 = _intrinsics.at(i);
125 assert(cg1->method() != cg2->method()
126 ? cg1->method() < cg2->method()
127 : cg1->is_virtual() < cg2->is_virtual(),
128 "compiler intrinsics list must stay sorted");
129 }
130 #endif
131 IntrinsicDescPair pair(m, is_virtual);
132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 bool found = false;
137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
138 assert(!found, "registering twice");
139 _intrinsics.insert_before(index, cg);
140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
141 }
142
143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
144 assert(m->is_loaded(), "don't try this on unloaded methods");
145 if (_intrinsics.length() > 0) {
146 bool found = false;
147 int index = intrinsic_insertion_index(m, is_virtual, found);
148 if (found) {
149 return _intrinsics.at(index);
150 }
151 }
152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
153 if (m->intrinsic_id() != vmIntrinsics::_none &&
154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
156 if (cg != nullptr) {
157 // Save it for next time:
158 register_intrinsic(cg);
159 return cg;
160 } else {
161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
162 }
163 }
164 return nullptr;
165 }
166
167 // Compile::make_vm_intrinsic is defined in library_call.cpp.
168
169 #ifndef PRODUCT
170 // statistics gathering...
171
172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
174
175 inline int as_int(vmIntrinsics::ID id) {
176 return vmIntrinsics::as_int(id);
177 }
178
179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
181 int oflags = _intrinsic_hist_flags[as_int(id)];
182 assert(flags != 0, "what happened?");
183 if (is_virtual) {
184 flags |= _intrinsic_virtual;
185 }
186 bool changed = (flags != oflags);
187 if ((flags & _intrinsic_worked) != 0) {
188 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
189 if (count == 1) {
190 changed = true; // first time
191 }
192 // increment the overall count also:
193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
194 }
195 if (changed) {
196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
197 // Something changed about the intrinsic's virtuality.
198 if ((flags & _intrinsic_virtual) != 0) {
199 // This is the first use of this intrinsic as a virtual call.
200 if (oflags != 0) {
201 // We already saw it as a non-virtual, so note both cases.
202 flags |= _intrinsic_both;
203 }
204 } else if ((oflags & _intrinsic_both) == 0) {
205 // This is the first use of this intrinsic as a non-virtual
206 flags |= _intrinsic_both;
207 }
208 }
209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
210 }
211 // update the overall flags also:
212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
213 return changed;
214 }
215
216 static char* format_flags(int flags, char* buf) {
217 buf[0] = 0;
218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
223 if (buf[0] == 0) strcat(buf, ",");
224 assert(buf[0] == ',', "must be");
225 return &buf[1];
226 }
227
228 void Compile::print_intrinsic_statistics() {
229 char flagsbuf[100];
230 ttyLocker ttyl;
231 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
232 tty->print_cr("Compiler intrinsic usage:");
233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
234 if (total == 0) total = 1; // avoid div0 in case of no successes
235 #define PRINT_STAT_LINE(name, c, f) \
236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
237 for (auto id : EnumRange<vmIntrinsicID>{}) {
238 int flags = _intrinsic_hist_flags[as_int(id)];
239 juint count = _intrinsic_hist_count[as_int(id)];
240 if ((flags | count) != 0) {
241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
242 }
243 }
244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
245 if (xtty != nullptr) xtty->tail("statistics");
246 }
247
248 void Compile::print_statistics() {
249 { ttyLocker ttyl;
250 if (xtty != nullptr) xtty->head("statistics type='opto'");
251 Parse::print_statistics();
252 PhaseCCP::print_statistics();
253 PhaseRegAlloc::print_statistics();
254 PhaseOutput::print_statistics();
255 PhasePeephole::print_statistics();
256 PhaseIdealLoop::print_statistics();
257 if (xtty != nullptr) xtty->tail("statistics");
258 }
259 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
260 // put this under its own <statistics> element.
261 print_intrinsic_statistics();
262 }
263 }
264 #endif //PRODUCT
265
266 void Compile::gvn_replace_by(Node* n, Node* nn) {
267 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
268 Node* use = n->last_out(i);
269 bool is_in_table = initial_gvn()->hash_delete(use);
270 uint uses_found = 0;
271 for (uint j = 0; j < use->len(); j++) {
272 if (use->in(j) == n) {
273 if (j < use->req())
274 use->set_req(j, nn);
275 else
276 use->set_prec(j, nn);
277 uses_found++;
278 }
279 }
280 if (is_in_table) {
281 // reinsert into table
282 initial_gvn()->hash_find_insert(use);
283 }
284 record_for_igvn(use);
285 i -= uses_found; // we deleted 1 or more copies of this edge
286 }
287 }
288
289
290 // Identify all nodes that are reachable from below, useful.
291 // Use breadth-first pass that records state in a Unique_Node_List,
292 // recursive traversal is slower.
293 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
294 int estimated_worklist_size = live_nodes();
295 useful.map( estimated_worklist_size, nullptr ); // preallocate space
296
297 // Initialize worklist
298 if (root() != nullptr) { useful.push(root()); }
299 // If 'top' is cached, declare it useful to preserve cached node
300 if (cached_top_node()) { useful.push(cached_top_node()); }
301
302 // Push all useful nodes onto the list, breadthfirst
303 for( uint next = 0; next < useful.size(); ++next ) {
304 assert( next < unique(), "Unique useful nodes < total nodes");
305 Node *n = useful.at(next);
306 uint max = n->len();
307 for( uint i = 0; i < max; ++i ) {
308 Node *m = n->in(i);
309 if (not_a_node(m)) continue;
310 useful.push(m);
311 }
312 }
313 }
314
315 // Update dead_node_list with any missing dead nodes using useful
316 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
317 void Compile::update_dead_node_list(Unique_Node_List &useful) {
318 uint max_idx = unique();
319 VectorSet& useful_node_set = useful.member_set();
320
321 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
322 // If node with index node_idx is not in useful set,
323 // mark it as dead in dead node list.
324 if (!useful_node_set.test(node_idx)) {
325 record_dead_node(node_idx);
326 }
327 }
328 }
329
330 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
331 int shift = 0;
332 for (int i = 0; i < inlines->length(); i++) {
333 CallGenerator* cg = inlines->at(i);
334 if (useful.member(cg->call_node())) {
335 if (shift > 0) {
336 inlines->at_put(i - shift, cg);
337 }
338 } else {
339 shift++; // skip over the dead element
340 }
341 }
342 if (shift > 0) {
343 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
344 }
345 }
346
347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
348 assert(dead != nullptr && dead->is_Call(), "sanity");
349 int found = 0;
350 for (int i = 0; i < inlines->length(); i++) {
351 if (inlines->at(i)->call_node() == dead) {
352 inlines->remove_at(i);
353 found++;
354 NOT_DEBUG( break; ) // elements are unique, so exit early
355 }
356 }
357 assert(found <= 1, "not unique");
358 }
359
360 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
361 for (int i = node_list.length() - 1; i >= 0; i--) {
362 Node* n = node_list.at(i);
363 if (!useful.member(n)) {
364 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
365 }
366 }
367 }
368
369 void Compile::remove_useless_node(Node* dead) {
370 remove_modified_node(dead);
371
372 // Constant node that has no out-edges and has only one in-edge from
373 // root is usually dead. However, sometimes reshaping walk makes
374 // it reachable by adding use edges. So, we will NOT count Con nodes
375 // as dead to be conservative about the dead node count at any
376 // given time.
377 if (!dead->is_Con()) {
378 record_dead_node(dead->_idx);
379 }
380 if (dead->is_macro()) {
381 remove_macro_node(dead);
382 }
383 if (dead->is_expensive()) {
384 remove_expensive_node(dead);
385 }
386 if (dead->Opcode() == Op_Opaque4) {
387 remove_skeleton_predicate_opaq(dead);
388 }
389 if (dead->for_post_loop_opts_igvn()) {
390 remove_from_post_loop_opts_igvn(dead);
391 }
392 if (dead->is_Call()) {
393 remove_useless_late_inlines( &_late_inlines, dead);
394 remove_useless_late_inlines( &_string_late_inlines, dead);
395 remove_useless_late_inlines( &_boxing_late_inlines, dead);
396 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
397 }
398 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
399 bs->unregister_potential_barrier_node(dead);
400 }
401
402 // Disconnect all useless nodes by disconnecting those at the boundary.
403 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
404 uint next = 0;
405 while (next < useful.size()) {
406 Node *n = useful.at(next++);
407 if (n->is_SafePoint()) {
408 // We're done with a parsing phase. Replaced nodes are not valid
409 // beyond that point.
410 n->as_SafePoint()->delete_replaced_nodes();
411 }
412 // Use raw traversal of out edges since this code removes out edges
413 int max = n->outcnt();
414 for (int j = 0; j < max; ++j) {
415 Node* child = n->raw_out(j);
416 if (!useful.member(child)) {
417 assert(!child->is_top() || child != top(),
418 "If top is cached in Compile object it is in useful list");
419 // Only need to remove this out-edge to the useless node
420 n->raw_del_out(j);
421 --j;
422 --max;
423 }
424 }
425 if (n->outcnt() == 1 && n->has_special_unique_user()) {
426 worklist->push(n->unique_out());
427 }
428 }
429
430 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
431 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
432 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
433 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
434 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
435 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
436 #ifdef ASSERT
437 if (_modified_nodes != nullptr) {
438 _modified_nodes->remove_useless_nodes(useful.member_set());
439 }
440 #endif
441
442 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
443 bs->eliminate_useless_gc_barriers(useful, this);
444 // clean up the late inline lists
445 remove_useless_late_inlines( &_late_inlines, useful);
446 remove_useless_late_inlines( &_string_late_inlines, useful);
447 remove_useless_late_inlines( &_boxing_late_inlines, useful);
448 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
449 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
450 }
451
452 // ============================================================================
453 //------------------------------CompileWrapper---------------------------------
454 class CompileWrapper : public StackObj {
455 Compile *const _compile;
456 public:
457 CompileWrapper(Compile* compile);
458
459 ~CompileWrapper();
460 };
461
462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
463 // the Compile* pointer is stored in the current ciEnv:
464 ciEnv* env = compile->env();
465 assert(env == ciEnv::current(), "must already be a ciEnv active");
466 assert(env->compiler_data() == nullptr, "compile already active?");
467 env->set_compiler_data(compile);
468 assert(compile == Compile::current(), "sanity");
469
470 compile->set_type_dict(nullptr);
471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
472 compile->clone_map().set_clone_idx(0);
473 compile->set_type_last_size(0);
474 compile->set_last_tf(nullptr, nullptr);
475 compile->set_indexSet_arena(nullptr);
476 compile->set_indexSet_free_block_list(nullptr);
477 compile->init_type_arena();
478 Type::Initialize(compile);
479 _compile->begin_method();
480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
481 }
482 CompileWrapper::~CompileWrapper() {
483 _compile->end_method();
484 _compile->env()->set_compiler_data(nullptr);
485 }
486
487
488 //----------------------------print_compile_messages---------------------------
489 void Compile::print_compile_messages() {
490 #ifndef PRODUCT
491 // Check if recompiling
492 if (_subsume_loads == false && PrintOpto) {
493 // Recompiling without allowing machine instructions to subsume loads
494 tty->print_cr("*********************************************************");
495 tty->print_cr("** Bailout: Recompile without subsuming loads **");
496 tty->print_cr("*********************************************************");
497 }
498 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
499 // Recompiling without escape analysis
500 tty->print_cr("*********************************************************");
501 tty->print_cr("** Bailout: Recompile without escape analysis **");
502 tty->print_cr("*********************************************************");
503 }
504 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
505 // Recompiling without boxing elimination
506 tty->print_cr("*********************************************************");
507 tty->print_cr("** Bailout: Recompile without boxing elimination **");
508 tty->print_cr("*********************************************************");
509 }
510 if ((_do_locks_coarsening != EliminateLocks) && PrintOpto) {
511 // Recompiling without locks coarsening
512 tty->print_cr("*********************************************************");
513 tty->print_cr("** Bailout: Recompile without locks coarsening **");
514 tty->print_cr("*********************************************************");
515 }
516 if (env()->break_at_compile()) {
517 // Open the debugger when compiling this method.
518 tty->print("### Breaking when compiling: ");
519 method()->print_short_name();
520 tty->cr();
521 BREAKPOINT;
522 }
523
524 if( PrintOpto ) {
525 if (is_osr_compilation()) {
526 tty->print("[OSR]%3d", _compile_id);
527 } else {
528 tty->print("%3d", _compile_id);
529 }
530 }
531 #endif
532 }
533
534 // ============================================================================
535 //------------------------------Compile standard-------------------------------
536 debug_only( int Compile::_debug_idx = 100000; )
537
538 // Compile a method. entry_bci is -1 for normal compilations and indicates
539 // the continuation bci for on stack replacement.
540
541
542 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
543 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing,
544 bool do_locks_coarsening, bool install_code, DirectiveSet* directive)
545 : Phase(Compiler),
546 _compile_id(ci_env->compile_id()),
547 _subsume_loads(subsume_loads),
548 _do_escape_analysis(do_escape_analysis),
549 _install_code(install_code),
550 _eliminate_boxing(eliminate_boxing),
551 _do_locks_coarsening(do_locks_coarsening),
552 _method(target),
553 _entry_bci(osr_bci),
554 _stub_function(nullptr),
555 _stub_name(nullptr),
556 _stub_entry_point(nullptr),
557 _max_node_limit(MaxNodeLimit),
558 _post_loop_opts_phase(false),
559 _inlining_progress(false),
560 _inlining_incrementally(false),
561 _do_cleanup(false),
562 _has_reserved_stack_access(target->has_reserved_stack_access()),
563 #ifndef PRODUCT
564 _igv_idx(0),
565 _trace_opto_output(directive->TraceOptoOutputOption),
566 _print_ideal(directive->PrintIdealOption),
567 #endif
568 _has_method_handle_invokes(false),
569 _clinit_barrier_on_entry(false),
570 _stress_seed(0),
571 _comp_arena(mtCompiler),
572 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
573 _env(ci_env),
574 _directive(directive),
575 _log(ci_env->log()),
576 _failure_reason(nullptr),
577 _intrinsics (comp_arena(), 0, 0, nullptr),
578 _macro_nodes (comp_arena(), 8, 0, nullptr),
579 _predicate_opaqs (comp_arena(), 8, 0, nullptr),
580 _skeleton_predicate_opaqs (comp_arena(), 8, 0, nullptr),
581 _expensive_nodes (comp_arena(), 8, 0, nullptr),
582 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
583 _coarsened_locks (comp_arena(), 8, 0, nullptr),
584 _congraph(nullptr),
585 NOT_PRODUCT(_printer(nullptr) COMMA)
586 _dead_node_list(comp_arena()),
587 _dead_node_count(0),
588 _node_arena(mtCompiler),
589 _old_arena(mtCompiler),
590 _mach_constant_base_node(nullptr),
591 _Compile_types(mtCompiler),
592 _initial_gvn(nullptr),
593 _for_igvn(nullptr),
594 _late_inlines(comp_arena(), 2, 0, nullptr),
595 _string_late_inlines(comp_arena(), 2, 0, nullptr),
596 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
597 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
598 _late_inlines_pos(0),
599 _number_of_mh_late_inlines(0),
600 _native_invokers(comp_arena(), 1, 0, NULL),
601 _print_inlining_stream(NULL),
602 _print_inlining_list(nullptr),
603 _print_inlining_idx(0),
604 _print_inlining_output(nullptr),
605 _replay_inline_data(nullptr),
606 _java_calls(0),
607 _inner_loops(0),
608 _interpreter_frame_size(0)
609 #ifndef PRODUCT
610 , _in_dump_cnt(0)
611 #endif
612 {
613 C = this;
614 CompileWrapper cw(this);
615
616 if (CITimeVerbose) {
617 tty->print(" ");
618 target->holder()->name()->print();
619 tty->print(".");
620 target->print_short_name();
621 tty->print(" ");
622 }
623 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
624 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
625
626 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
627 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
628 // We can always print a disassembly, either abstract (hex dump) or
629 // with the help of a suitable hsdis library. Thus, we should not
630 // couple print_assembly and print_opto_assembly controls.
631 // But: always print opto and regular assembly on compile command 'print'.
632 bool print_assembly = directive->PrintAssemblyOption;
633 set_print_assembly(print_opto_assembly || print_assembly);
634 #else
635 set_print_assembly(false); // must initialize.
636 #endif
637
638 #ifndef PRODUCT
639 set_parsed_irreducible_loop(false);
640
641 if (directive->ReplayInlineOption) {
642 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
643 }
644 #endif
645 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
646 set_print_intrinsics(directive->PrintIntrinsicsOption);
647 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
648
649 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
650 // Make sure the method being compiled gets its own MDO,
651 // so we can at least track the decompile_count().
652 // Need MDO to record RTM code generation state.
653 method()->ensure_method_data();
654 }
655
656 Init(::AliasLevel);
657
658
659 print_compile_messages();
660
661 _ilt = InlineTree::build_inline_tree_root();
662
663 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
664 assert(num_alias_types() >= AliasIdxRaw, "");
665
666 #define MINIMUM_NODE_HASH 1023
667 // Node list that Iterative GVN will start with
668 Unique_Node_List for_igvn(comp_arena());
669 set_for_igvn(&for_igvn);
670
671 // GVN that will be run immediately on new nodes
672 uint estimated_size = method()->code_size()*4+64;
673 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
674 PhaseGVN gvn(node_arena(), estimated_size);
675 set_initial_gvn(&gvn);
676
677 print_inlining_init();
678 { // Scope for timing the parser
679 TracePhase tp("parse", &timers[_t_parser]);
680
681 // Put top into the hash table ASAP.
682 initial_gvn()->transform_no_reclaim(top());
683
684 // Set up tf(), start(), and find a CallGenerator.
685 CallGenerator* cg = nullptr;
686 if (is_osr_compilation()) {
687 const TypeTuple *domain = StartOSRNode::osr_domain();
688 const TypeTuple *range = TypeTuple::make_range(method()->signature());
689 init_tf(TypeFunc::make(domain, range));
690 StartNode* s = new StartOSRNode(root(), domain);
691 initial_gvn()->set_type_bottom(s);
692 init_start(s);
693 cg = CallGenerator::for_osr(method(), entry_bci());
694 } else {
695 // Normal case.
696 init_tf(TypeFunc::make(method()));
697 StartNode* s = new StartNode(root(), tf()->domain());
698 initial_gvn()->set_type_bottom(s);
699 init_start(s);
700 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
701 // With java.lang.ref.reference.get() we must go through the
702 // intrinsic - even when get() is the root
703 // method of the compile - so that, if necessary, the value in
704 // the referent field of the reference object gets recorded by
705 // the pre-barrier code.
706 cg = find_intrinsic(method(), false);
707 }
708 if (cg == nullptr) {
709 float past_uses = method()->interpreter_invocation_count();
710 float expected_uses = past_uses;
711 cg = CallGenerator::for_inline(method(), expected_uses);
712 }
713 }
714 if (failing()) return;
715 if (cg == nullptr) {
716 record_method_not_compilable("cannot parse method");
717 return;
718 }
719 JVMState* jvms = build_start_state(start(), tf());
720 if ((jvms = cg->generate(jvms)) == nullptr) {
721 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
722 record_method_not_compilable("method parse failed");
723 }
724 return;
725 }
726 GraphKit kit(jvms);
727
728 if (!kit.stopped()) {
729 // Accept return values, and transfer control we know not where.
730 // This is done by a special, unique ReturnNode bound to root.
731 return_values(kit.jvms());
732 }
733
734 if (kit.has_exceptions()) {
735 // Any exceptions that escape from this call must be rethrown
736 // to whatever caller is dynamically above us on the stack.
737 // This is done by a special, unique RethrowNode bound to root.
738 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
739 }
740
741 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
742
743 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
744 inline_string_calls(true);
745 }
746
747 if (failing()) return;
748
749 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
750
751 // Remove clutter produced by parsing.
752 if (!failing()) {
753 ResourceMark rm;
754 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
755 }
756 }
757
758 // Note: Large methods are capped off in do_one_bytecode().
759 if (failing()) return;
760
761 // After parsing, node notes are no longer automagic.
762 // They must be propagated by register_new_node_with_optimizer(),
763 // clone(), or the like.
764 set_default_node_notes(nullptr);
765
766 #ifndef PRODUCT
767 if (should_print(1)) {
768 _printer->print_inlining();
769 }
770 #endif
771
772 if (failing()) return;
773 NOT_PRODUCT( verify_graph_edges(); )
774
775 // If any phase is randomized for stress testing, seed random number
776 // generation and log the seed for repeatability.
777 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
778 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
779 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
780 FLAG_SET_ERGO(StressSeed, _stress_seed);
781 } else {
782 _stress_seed = StressSeed;
783 }
784 if (_log != nullptr) {
785 _log->elem("stress_test seed='%u'", _stress_seed);
786 }
787 }
788
789 // Now optimize
790 Optimize();
791 if (failing()) return;
792 NOT_PRODUCT( verify_graph_edges(); )
793
794 #ifndef PRODUCT
795 if (print_ideal()) {
796 ttyLocker ttyl; // keep the following output all in one block
797 // This output goes directly to the tty, not the compiler log.
798 // To enable tools to match it up with the compilation activity,
799 // be sure to tag this tty output with the compile ID.
800 if (xtty != nullptr) {
801 xtty->head("ideal compile_id='%d'%s", compile_id(),
802 is_osr_compilation() ? " compile_kind='osr'" :
803 "");
804 }
805 root()->dump(9999);
806 if (xtty != nullptr) {
807 xtty->tail("ideal");
808 }
809 }
810 #endif
811
812 #ifdef ASSERT
813 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
814 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
815 #endif
816
817 // Dump compilation data to replay it.
818 if (directive->DumpReplayOption) {
819 env()->dump_replay_data(_compile_id);
820 }
821 if (directive->DumpInlineOption && (ilt() != nullptr)) {
822 env()->dump_inline_data(_compile_id);
823 }
824
825 // Now that we know the size of all the monitors we can add a fixed slot
826 // for the original deopt pc.
827 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
828 set_fixed_slots(next_slot);
829
830 // Compute when to use implicit null checks. Used by matching trap based
831 // nodes and NullCheck optimization.
832 set_allowed_deopt_reasons();
833
834 // Now generate code
835 Code_Gen();
836 }
837
838 //------------------------------Compile----------------------------------------
839 // Compile a runtime stub
840 Compile::Compile( ciEnv* ci_env,
841 TypeFunc_generator generator,
842 address stub_function,
843 const char *stub_name,
844 int is_fancy_jump,
845 bool pass_tls,
846 bool return_pc,
847 DirectiveSet* directive)
848 : Phase(Compiler),
849 _compile_id(0),
850 _subsume_loads(true),
851 _do_escape_analysis(false),
852 _install_code(true),
853 _eliminate_boxing(false),
854 _do_locks_coarsening(false),
855 _method(nullptr),
856 _entry_bci(InvocationEntryBci),
857 _stub_function(stub_function),
858 _stub_name(stub_name),
859 _stub_entry_point(nullptr),
860 _max_node_limit(MaxNodeLimit),
861 _post_loop_opts_phase(false),
862 _inlining_progress(false),
863 _inlining_incrementally(false),
864 _has_reserved_stack_access(false),
865 #ifndef PRODUCT
866 _igv_idx(0),
867 _trace_opto_output(directive->TraceOptoOutputOption),
868 _print_ideal(directive->PrintIdealOption),
869 #endif
870 _has_method_handle_invokes(false),
871 _clinit_barrier_on_entry(false),
872 _stress_seed(0),
873 _comp_arena(mtCompiler),
874 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
875 _env(ci_env),
876 _directive(directive),
877 _log(ci_env->log()),
878 _failure_reason(nullptr),
879 _congraph(nullptr),
880 NOT_PRODUCT(_printer(nullptr) COMMA)
881 _dead_node_list(comp_arena()),
882 _dead_node_count(0),
883 _node_arena(mtCompiler),
884 _old_arena(mtCompiler),
885 _mach_constant_base_node(nullptr),
886 _Compile_types(mtCompiler),
887 _initial_gvn(nullptr),
888 _for_igvn(nullptr),
889 _number_of_mh_late_inlines(0),
890 _native_invokers(),
891 _print_inlining_stream(NULL),
892 _print_inlining_list(nullptr),
893 _print_inlining_idx(0),
894 _print_inlining_output(nullptr),
895 _replay_inline_data(nullptr),
896 _java_calls(0),
897 _inner_loops(0),
898 _interpreter_frame_size(0),
899 #ifndef PRODUCT
900 _in_dump_cnt(0),
901 #endif
902 _allowed_reasons(0) {
903 C = this;
904
905 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
906 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
907
908 #ifndef PRODUCT
909 set_print_assembly(PrintFrameConverterAssembly);
910 set_parsed_irreducible_loop(false);
911 #else
912 set_print_assembly(false); // Must initialize.
913 #endif
914 set_has_irreducible_loop(false); // no loops
915
916 CompileWrapper cw(this);
917 Init(/*AliasLevel=*/ 0);
918 init_tf((*generator)());
919
920 {
921 // The following is a dummy for the sake of GraphKit::gen_stub
922 Unique_Node_List for_igvn(comp_arena());
923 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
924 PhaseGVN gvn(Thread::current()->resource_area(),255);
925 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
926 gvn.transform_no_reclaim(top());
927
928 GraphKit kit;
929 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
930 }
931
932 NOT_PRODUCT( verify_graph_edges(); )
933
934 Code_Gen();
935 }
936
937 //------------------------------Init-------------------------------------------
938 // Prepare for a single compilation
939 void Compile::Init(int aliaslevel) {
940 _unique = 0;
941 _regalloc = nullptr;
942
943 _tf = nullptr; // filled in later
944 _top = nullptr; // cached later
945 _matcher = nullptr; // filled in later
946 _cfg = nullptr; // filled in later
947
948 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
949
950 _node_note_array = nullptr;
951 _default_node_notes = nullptr;
952 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
953
954 _immutable_memory = nullptr; // filled in at first inquiry
955
956 #ifdef ASSERT
957 _type_verify_symmetry = true;
958 _phase_optimize_finished = false;
959 _exception_backedge = false;
960 #endif
961
962 // Globally visible Nodes
963 // First set TOP to null to give safe behavior during creation of RootNode
964 set_cached_top_node(nullptr);
965 set_root(new RootNode());
966 // Now that you have a Root to point to, create the real TOP
967 set_cached_top_node( new ConNode(Type::TOP) );
968 set_recent_alloc(nullptr, nullptr);
969
970 // Create Debug Information Recorder to record scopes, oopmaps, etc.
971 env()->set_oop_recorder(new OopRecorder(env()->arena()));
972 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
973 env()->set_dependencies(new Dependencies(env()));
974
975 _fixed_slots = 0;
976 set_has_split_ifs(false);
977 set_has_loops(false); // first approximation
978 set_has_stringbuilder(false);
979 set_has_boxed_value(false);
980 _trap_can_recompile = false; // no traps emitted yet
981 _major_progress = true; // start out assuming good things will happen
982 set_has_unsafe_access(false);
983 set_max_vector_size(0);
984 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
985 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
986 set_decompile_count(0);
987
988 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
989 _loop_opts_cnt = LoopOptsCount;
990 set_do_inlining(Inline);
991 set_max_inline_size(MaxInlineSize);
992 set_freq_inline_size(FreqInlineSize);
993 set_do_scheduling(OptoScheduling);
994
995 set_do_vector_loop(false);
996
997 if (AllowVectorizeOnDemand) {
998 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
999 set_do_vector_loop(true);
1000 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1001 } else if (has_method() && method()->name() != 0 &&
1002 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1003 set_do_vector_loop(true);
1004 }
1005 }
1006 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1007 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1008
1009 set_age_code(has_method() && method()->profile_aging());
1010 set_rtm_state(NoRTM); // No RTM lock eliding by default
1011 _max_node_limit = _directive->MaxNodeLimitOption;
1012
1013 #if INCLUDE_RTM_OPT
1014 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) {
1015 int rtm_state = method()->method_data()->rtm_state();
1016 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1017 // Don't generate RTM lock eliding code.
1018 set_rtm_state(NoRTM);
1019 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1020 // Generate RTM lock eliding code without abort ratio calculation code.
1021 set_rtm_state(UseRTM);
1022 } else if (UseRTMDeopt) {
1023 // Generate RTM lock eliding code and include abort ratio calculation
1024 // code if UseRTMDeopt is on.
1025 set_rtm_state(ProfileRTM);
1026 }
1027 }
1028 #endif
1029 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1030 set_clinit_barrier_on_entry(true);
1031 }
1032 if (debug_info()->recording_non_safepoints()) {
1033 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1034 (comp_arena(), 8, 0, nullptr));
1035 set_default_node_notes(Node_Notes::make(this));
1036 }
1037
1038 // // -- Initialize types before each compile --
1039 // // Update cached type information
1040 // if( _method && _method->constants() )
1041 // Type::update_loaded_types(_method, _method->constants());
1042
1043 // Init alias_type map.
1044 if (!_do_escape_analysis && aliaslevel == 3)
1045 aliaslevel = 2; // No unique types without escape analysis
1046 _AliasLevel = aliaslevel;
1047 const int grow_ats = 16;
1048 _max_alias_types = grow_ats;
1049 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1050 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1051 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1052 {
1053 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1054 }
1055 // Initialize the first few types.
1056 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1057 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1058 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1059 _num_alias_types = AliasIdxRaw+1;
1060 // Zero out the alias type cache.
1061 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1062 // A null adr_type hits in the cache right away. Preload the right answer.
1063 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1064 }
1065
1066 //---------------------------init_start----------------------------------------
1067 // Install the StartNode on this compile object.
1068 void Compile::init_start(StartNode* s) {
1069 if (failing())
1070 return; // already failing
1071 assert(s == start(), "");
1072 }
1073
1074 /**
1075 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1076 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1077 * the ideal graph.
1078 */
1079 StartNode* Compile::start() const {
1080 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1081 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1082 Node* start = root()->fast_out(i);
1083 if (start->is_Start()) {
1084 return start->as_Start();
1085 }
1086 }
1087 fatal("Did not find Start node!");
1088 return nullptr;
1089 }
1090
1091 //-------------------------------immutable_memory-------------------------------------
1092 // Access immutable memory
1093 Node* Compile::immutable_memory() {
1094 if (_immutable_memory != nullptr) {
1095 return _immutable_memory;
1096 }
1097 StartNode* s = start();
1098 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1099 Node *p = s->fast_out(i);
1100 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1101 _immutable_memory = p;
1102 return _immutable_memory;
1103 }
1104 }
1105 ShouldNotReachHere();
1106 return nullptr;
1107 }
1108
1109 //----------------------set_cached_top_node------------------------------------
1110 // Install the cached top node, and make sure Node::is_top works correctly.
1111 void Compile::set_cached_top_node(Node* tn) {
1112 if (tn != nullptr) verify_top(tn);
1113 Node* old_top = _top;
1114 _top = tn;
1115 // Calling Node::setup_is_top allows the nodes the chance to adjust
1116 // their _out arrays.
1117 if (_top != nullptr) _top->setup_is_top();
1118 if (old_top != nullptr) old_top->setup_is_top();
1119 assert(_top == nullptr || top()->is_top(), "");
1120 }
1121
1122 #ifdef ASSERT
1123 uint Compile::count_live_nodes_by_graph_walk() {
1124 Unique_Node_List useful(comp_arena());
1125 // Get useful node list by walking the graph.
1126 identify_useful_nodes(useful);
1127 return useful.size();
1128 }
1129
1130 void Compile::print_missing_nodes() {
1131
1132 // Return if CompileLog is null and PrintIdealNodeCount is false.
1133 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1134 return;
1135 }
1136
1137 // This is an expensive function. It is executed only when the user
1138 // specifies VerifyIdealNodeCount option or otherwise knows the
1139 // additional work that needs to be done to identify reachable nodes
1140 // by walking the flow graph and find the missing ones using
1141 // _dead_node_list.
1142
1143 Unique_Node_List useful(comp_arena());
1144 // Get useful node list by walking the graph.
1145 identify_useful_nodes(useful);
1146
1147 uint l_nodes = C->live_nodes();
1148 uint l_nodes_by_walk = useful.size();
1149
1150 if (l_nodes != l_nodes_by_walk) {
1151 if (_log != nullptr) {
1152 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1153 _log->stamp();
1154 _log->end_head();
1155 }
1156 VectorSet& useful_member_set = useful.member_set();
1157 int last_idx = l_nodes_by_walk;
1158 for (int i = 0; i < last_idx; i++) {
1159 if (useful_member_set.test(i)) {
1160 if (_dead_node_list.test(i)) {
1161 if (_log != nullptr) {
1162 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1163 }
1164 if (PrintIdealNodeCount) {
1165 // Print the log message to tty
1166 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1167 useful.at(i)->dump();
1168 }
1169 }
1170 }
1171 else if (! _dead_node_list.test(i)) {
1172 if (_log != nullptr) {
1173 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1174 }
1175 if (PrintIdealNodeCount) {
1176 // Print the log message to tty
1177 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1178 }
1179 }
1180 }
1181 if (_log != nullptr) {
1182 _log->tail("mismatched_nodes");
1183 }
1184 }
1185 }
1186 void Compile::record_modified_node(Node* n) {
1187 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1188 _modified_nodes->push(n);
1189 }
1190 }
1191
1192 void Compile::remove_modified_node(Node* n) {
1193 if (_modified_nodes != nullptr) {
1194 _modified_nodes->remove(n);
1195 }
1196 }
1197 #endif
1198
1199 #ifndef PRODUCT
1200 void Compile::verify_top(Node* tn) const {
1201 if (tn != nullptr) {
1202 assert(tn->is_Con(), "top node must be a constant");
1203 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1204 assert(tn->in(0) != nullptr, "must have live top node");
1205 }
1206 }
1207 #endif
1208
1209
1210 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1211
1212 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1213 guarantee(arr != nullptr, "");
1214 int num_blocks = arr->length();
1215 if (grow_by < num_blocks) grow_by = num_blocks;
1216 int num_notes = grow_by * _node_notes_block_size;
1217 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1218 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1219 while (num_notes > 0) {
1220 arr->append(notes);
1221 notes += _node_notes_block_size;
1222 num_notes -= _node_notes_block_size;
1223 }
1224 assert(num_notes == 0, "exact multiple, please");
1225 }
1226
1227 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1228 if (source == nullptr || dest == nullptr) return false;
1229
1230 if (dest->is_Con())
1231 return false; // Do not push debug info onto constants.
1232
1233 #ifdef ASSERT
1234 // Leave a bread crumb trail pointing to the original node:
1235 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1236 dest->set_debug_orig(source);
1237 }
1238 #endif
1239
1240 if (node_note_array() == nullptr)
1241 return false; // Not collecting any notes now.
1242
1243 // This is a copy onto a pre-existing node, which may already have notes.
1244 // If both nodes have notes, do not overwrite any pre-existing notes.
1245 Node_Notes* source_notes = node_notes_at(source->_idx);
1246 if (source_notes == nullptr || source_notes->is_clear()) return false;
1247 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1248 if (dest_notes == nullptr || dest_notes->is_clear()) {
1249 return set_node_notes_at(dest->_idx, source_notes);
1250 }
1251
1252 Node_Notes merged_notes = (*source_notes);
1253 // The order of operations here ensures that dest notes will win...
1254 merged_notes.update_from(dest_notes);
1255 return set_node_notes_at(dest->_idx, &merged_notes);
1256 }
1257
1258
1259 //--------------------------allow_range_check_smearing-------------------------
1260 // Gating condition for coalescing similar range checks.
1261 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1262 // single covering check that is at least as strong as any of them.
1263 // If the optimization succeeds, the simplified (strengthened) range check
1264 // will always succeed. If it fails, we will deopt, and then give up
1265 // on the optimization.
1266 bool Compile::allow_range_check_smearing() const {
1267 // If this method has already thrown a range-check,
1268 // assume it was because we already tried range smearing
1269 // and it failed.
1270 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1271 return !already_trapped;
1272 }
1273
1274
1275 //------------------------------flatten_alias_type-----------------------------
1276 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1277 int offset = tj->offset();
1278 TypePtr::PTR ptr = tj->ptr();
1279
1280 // Known instance (scalarizable allocation) alias only with itself.
1281 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1282 tj->is_oopptr()->is_known_instance();
1283
1284 // Process weird unsafe references.
1285 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1286 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1287 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1288 tj = TypeOopPtr::BOTTOM;
1289 ptr = tj->ptr();
1290 offset = tj->offset();
1291 }
1292
1293 // Array pointers need some flattening
1294 const TypeAryPtr *ta = tj->isa_aryptr();
1295 if (ta && ta->is_stable()) {
1296 // Erase stability property for alias analysis.
1297 tj = ta = ta->cast_to_stable(false);
1298 }
1299 if( ta && is_known_inst ) {
1300 if ( offset != Type::OffsetBot &&
1301 offset > arrayOopDesc::length_offset_in_bytes() ) {
1302 offset = Type::OffsetBot; // Flatten constant access into array body only
1303 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1304 }
1305 } else if( ta && _AliasLevel >= 2 ) {
1306 // For arrays indexed by constant indices, we flatten the alias
1307 // space to include all of the array body. Only the header, klass
1308 // and array length can be accessed un-aliased.
1309 if( offset != Type::OffsetBot ) {
1310 if( ta->const_oop() ) { // MethodData* or Method*
1311 offset = Type::OffsetBot; // Flatten constant access into array body
1312 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1313 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1314 // range is OK as-is.
1315 tj = ta = TypeAryPtr::RANGE;
1316 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1317 tj = TypeInstPtr::KLASS; // all klass loads look alike
1318 ta = TypeAryPtr::RANGE; // generic ignored junk
1319 ptr = TypePtr::BotPTR;
1320 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1321 tj = TypeInstPtr::MARK;
1322 ta = TypeAryPtr::RANGE; // generic ignored junk
1323 ptr = TypePtr::BotPTR;
1324 } else { // Random constant offset into array body
1325 offset = Type::OffsetBot; // Flatten constant access into array body
1326 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1327 }
1328 }
1329 // Arrays of fixed size alias with arrays of unknown size.
1330 if (ta->size() != TypeInt::POS) {
1331 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1332 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1333 }
1334 // Arrays of known objects become arrays of unknown objects.
1335 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1336 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1337 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1338 }
1339 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1340 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1341 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1342 }
1343 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1344 // cannot be distinguished by bytecode alone.
1345 if (ta->elem() == TypeInt::BOOL) {
1346 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1347 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1348 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1349 }
1350 // During the 2nd round of IterGVN, NotNull castings are removed.
1351 // Make sure the Bottom and NotNull variants alias the same.
1352 // Also, make sure exact and non-exact variants alias the same.
1353 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1354 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1355 }
1356 }
1357
1358 // Oop pointers need some flattening
1359 const TypeInstPtr *to = tj->isa_instptr();
1360 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1361 ciInstanceKlass *k = to->klass()->as_instance_klass();
1362 if( ptr == TypePtr::Constant ) {
1363 if (to->klass() != ciEnv::current()->Class_klass() ||
1364 offset < k->layout_helper_size_in_bytes()) {
1365 // No constant oop pointers (such as Strings); they alias with
1366 // unknown strings.
1367 assert(!is_known_inst, "not scalarizable allocation");
1368 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1369 }
1370 } else if( is_known_inst ) {
1371 tj = to; // Keep NotNull and klass_is_exact for instance type
1372 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1373 // During the 2nd round of IterGVN, NotNull castings are removed.
1374 // Make sure the Bottom and NotNull variants alias the same.
1375 // Also, make sure exact and non-exact variants alias the same.
1376 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1377 }
1378 if (to->speculative() != nullptr) {
1379 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1380 }
1381 // Canonicalize the holder of this field
1382 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1383 // First handle header references such as a LoadKlassNode, even if the
1384 // object's klass is unloaded at compile time (4965979).
1385 if (!is_known_inst) { // Do it only for non-instance types
1386 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1387 }
1388 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) {
1389 // Static fields are in the space above the normal instance
1390 // fields in the java.lang.Class instance.
1391 if (to->klass() != ciEnv::current()->Class_klass()) {
1392 to = nullptr;
1393 tj = TypeOopPtr::BOTTOM;
1394 offset = tj->offset();
1395 }
1396 } else {
1397 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1398 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1399 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1400 if( is_known_inst ) {
1401 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1402 } else {
1403 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1404 }
1405 }
1406 }
1407 }
1408
1409 // Klass pointers to object array klasses need some flattening
1410 const TypeKlassPtr *tk = tj->isa_klassptr();
1411 if( tk ) {
1412 // If we are referencing a field within a Klass, we need
1413 // to assume the worst case of an Object. Both exact and
1414 // inexact types must flatten to the same alias class so
1415 // use NotNull as the PTR.
1416 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1417
1418 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1419 TypeKlassPtr::OBJECT->klass(),
1420 offset);
1421 }
1422
1423 ciKlass* klass = tk->klass();
1424 if( klass->is_obj_array_klass() ) {
1425 ciKlass* k = TypeAryPtr::OOPS->klass();
1426 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1427 k = TypeInstPtr::BOTTOM->klass();
1428 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1429 }
1430
1431 // Check for precise loads from the primary supertype array and force them
1432 // to the supertype cache alias index. Check for generic array loads from
1433 // the primary supertype array and also force them to the supertype cache
1434 // alias index. Since the same load can reach both, we need to merge
1435 // these 2 disparate memories into the same alias class. Since the
1436 // primary supertype array is read-only, there's no chance of confusion
1437 // where we bypass an array load and an array store.
1438 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1439 if (offset == Type::OffsetBot ||
1440 (offset >= primary_supers_offset &&
1441 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1442 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1443 offset = in_bytes(Klass::secondary_super_cache_offset());
1444 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1445 }
1446 }
1447
1448 // Flatten all Raw pointers together.
1449 if (tj->base() == Type::RawPtr)
1450 tj = TypeRawPtr::BOTTOM;
1451
1452 if (tj->base() == Type::AnyPtr)
1453 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1454
1455 // Flatten all to bottom for now
1456 switch( _AliasLevel ) {
1457 case 0:
1458 tj = TypePtr::BOTTOM;
1459 break;
1460 case 1: // Flatten to: oop, static, field or array
1461 switch (tj->base()) {
1462 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1463 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1464 case Type::AryPtr: // do not distinguish arrays at all
1465 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1466 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1467 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1468 default: ShouldNotReachHere();
1469 }
1470 break;
1471 case 2: // No collapsing at level 2; keep all splits
1472 case 3: // No collapsing at level 3; keep all splits
1473 break;
1474 default:
1475 Unimplemented();
1476 }
1477
1478 offset = tj->offset();
1479 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1480
1481 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1482 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1483 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1484 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1485 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1486 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1487 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1488 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1489 assert( tj->ptr() != TypePtr::TopPTR &&
1490 tj->ptr() != TypePtr::AnyNull &&
1491 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1492 // assert( tj->ptr() != TypePtr::Constant ||
1493 // tj->base() == Type::RawPtr ||
1494 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1495
1496 return tj;
1497 }
1498
1499 void Compile::AliasType::Init(int i, const TypePtr* at) {
1500 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1501 _index = i;
1502 _adr_type = at;
1503 _field = nullptr;
1504 _element = nullptr;
1505 _is_rewritable = true; // default
1506 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1507 if (atoop != nullptr && atoop->is_known_instance()) {
1508 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1509 _general_index = Compile::current()->get_alias_index(gt);
1510 } else {
1511 _general_index = 0;
1512 }
1513 }
1514
1515 BasicType Compile::AliasType::basic_type() const {
1516 if (element() != nullptr) {
1517 const Type* element = adr_type()->is_aryptr()->elem();
1518 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1519 } if (field() != nullptr) {
1520 return field()->layout_type();
1521 } else {
1522 return T_ILLEGAL; // unknown
1523 }
1524 }
1525
1526 //---------------------------------print_on------------------------------------
1527 #ifndef PRODUCT
1528 void Compile::AliasType::print_on(outputStream* st) {
1529 if (index() < 10)
1530 st->print("@ <%d> ", index());
1531 else st->print("@ <%d>", index());
1532 st->print(is_rewritable() ? " " : " RO");
1533 int offset = adr_type()->offset();
1534 if (offset == Type::OffsetBot)
1535 st->print(" +any");
1536 else st->print(" +%-3d", offset);
1537 st->print(" in ");
1538 adr_type()->dump_on(st);
1539 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1540 if (field() != nullptr && tjp) {
1541 if (tjp->klass() != field()->holder() ||
1542 tjp->offset() != field()->offset_in_bytes()) {
1543 st->print(" != ");
1544 field()->print();
1545 st->print(" ***");
1546 }
1547 }
1548 }
1549
1550 void print_alias_types() {
1551 Compile* C = Compile::current();
1552 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1553 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1554 C->alias_type(idx)->print_on(tty);
1555 tty->cr();
1556 }
1557 }
1558 #endif
1559
1560
1561 //----------------------------probe_alias_cache--------------------------------
1562 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1563 intptr_t key = (intptr_t) adr_type;
1564 key ^= key >> logAliasCacheSize;
1565 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1566 }
1567
1568
1569 //-----------------------------grow_alias_types--------------------------------
1570 void Compile::grow_alias_types() {
1571 const int old_ats = _max_alias_types; // how many before?
1572 const int new_ats = old_ats; // how many more?
1573 const int grow_ats = old_ats+new_ats; // how many now?
1574 _max_alias_types = grow_ats;
1575 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1576 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1577 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1578 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1579 }
1580
1581
1582 //--------------------------------find_alias_type------------------------------
1583 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1584 if (_AliasLevel == 0)
1585 return alias_type(AliasIdxBot);
1586
1587 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1588 if (ace->_adr_type == adr_type) {
1589 return alias_type(ace->_index);
1590 }
1591
1592 // Handle special cases.
1593 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1594 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1595
1596 // Do it the slow way.
1597 const TypePtr* flat = flatten_alias_type(adr_type);
1598
1599 #ifdef ASSERT
1600 {
1601 ResourceMark rm;
1602 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1603 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1604 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1605 Type::str(adr_type));
1606 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1607 const TypeOopPtr* foop = flat->is_oopptr();
1608 // Scalarizable allocations have exact klass always.
1609 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1610 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1611 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1612 Type::str(foop), Type::str(xoop));
1613 }
1614 }
1615 #endif
1616
1617 int idx = AliasIdxTop;
1618 for (int i = 0; i < num_alias_types(); i++) {
1619 if (alias_type(i)->adr_type() == flat) {
1620 idx = i;
1621 break;
1622 }
1623 }
1624
1625 if (idx == AliasIdxTop) {
1626 if (no_create) return nullptr;
1627 // Grow the array if necessary.
1628 if (_num_alias_types == _max_alias_types) grow_alias_types();
1629 // Add a new alias type.
1630 idx = _num_alias_types++;
1631 _alias_types[idx]->Init(idx, flat);
1632 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1633 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1634 if (flat->isa_instptr()) {
1635 if (flat->offset() == java_lang_Class::klass_offset()
1636 && flat->is_instptr()->klass() == env()->Class_klass())
1637 alias_type(idx)->set_rewritable(false);
1638 }
1639 if (flat->isa_aryptr()) {
1640 #ifdef ASSERT
1641 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1642 // (T_BYTE has the weakest alignment and size restrictions...)
1643 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1644 #endif
1645 if (flat->offset() == TypePtr::OffsetBot) {
1646 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1647 }
1648 }
1649 if (flat->isa_klassptr()) {
1650 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1651 alias_type(idx)->set_rewritable(false);
1652 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1653 alias_type(idx)->set_rewritable(false);
1654 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1655 alias_type(idx)->set_rewritable(false);
1656 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1657 alias_type(idx)->set_rewritable(false);
1658 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1659 alias_type(idx)->set_rewritable(false);
1660 }
1661 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1662 // but the base pointer type is not distinctive enough to identify
1663 // references into JavaThread.)
1664
1665 // Check for final fields.
1666 const TypeInstPtr* tinst = flat->isa_instptr();
1667 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1668 ciField* field;
1669 if (tinst->const_oop() != nullptr &&
1670 tinst->klass() == ciEnv::current()->Class_klass() &&
1671 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) {
1672 // static field
1673 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1674 field = k->get_field_by_offset(tinst->offset(), true);
1675 } else {
1676 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1677 field = k->get_field_by_offset(tinst->offset(), false);
1678 }
1679 assert(field == nullptr ||
1680 original_field == nullptr ||
1681 (field->holder() == original_field->holder() &&
1682 field->offset() == original_field->offset() &&
1683 field->is_static() == original_field->is_static()), "wrong field?");
1684 // Set field() and is_rewritable() attributes.
1685 if (field != nullptr) alias_type(idx)->set_field(field);
1686 }
1687 }
1688
1689 // Fill the cache for next time.
1690 ace->_adr_type = adr_type;
1691 ace->_index = idx;
1692 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1693
1694 // Might as well try to fill the cache for the flattened version, too.
1695 AliasCacheEntry* face = probe_alias_cache(flat);
1696 if (face->_adr_type == nullptr) {
1697 face->_adr_type = flat;
1698 face->_index = idx;
1699 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1700 }
1701
1702 return alias_type(idx);
1703 }
1704
1705
1706 Compile::AliasType* Compile::alias_type(ciField* field) {
1707 const TypeOopPtr* t;
1708 if (field->is_static())
1709 t = TypeInstPtr::make(field->holder()->java_mirror());
1710 else
1711 t = TypeOopPtr::make_from_klass_raw(field->holder());
1712 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1713 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1714 return atp;
1715 }
1716
1717
1718 //------------------------------have_alias_type--------------------------------
1719 bool Compile::have_alias_type(const TypePtr* adr_type) {
1720 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1721 if (ace->_adr_type == adr_type) {
1722 return true;
1723 }
1724
1725 // Handle special cases.
1726 if (adr_type == nullptr) return true;
1727 if (adr_type == TypePtr::BOTTOM) return true;
1728
1729 return find_alias_type(adr_type, true, nullptr) != nullptr;
1730 }
1731
1732 //-----------------------------must_alias--------------------------------------
1733 // True if all values of the given address type are in the given alias category.
1734 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1735 if (alias_idx == AliasIdxBot) return true; // the universal category
1736 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1737 if (alias_idx == AliasIdxTop) return false; // the empty category
1738 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1739
1740 // the only remaining possible overlap is identity
1741 int adr_idx = get_alias_index(adr_type);
1742 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1743 assert(adr_idx == alias_idx ||
1744 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1745 && adr_type != TypeOopPtr::BOTTOM),
1746 "should not be testing for overlap with an unsafe pointer");
1747 return adr_idx == alias_idx;
1748 }
1749
1750 //------------------------------can_alias--------------------------------------
1751 // True if any values of the given address type are in the given alias category.
1752 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1753 if (alias_idx == AliasIdxTop) return false; // the empty category
1754 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1755 // Known instance doesn't alias with bottom memory
1756 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1757 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1758
1759 // the only remaining possible overlap is identity
1760 int adr_idx = get_alias_index(adr_type);
1761 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1762 return adr_idx == alias_idx;
1763 }
1764
1765 //---------------------cleanup_loop_predicates-----------------------
1766 // Remove the opaque nodes that protect the predicates so that all unused
1767 // checks and uncommon_traps will be eliminated from the ideal graph
1768 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1769 if (predicate_count()==0) return;
1770 for (int i = predicate_count(); i > 0; i--) {
1771 Node * n = predicate_opaque1_node(i-1);
1772 assert(n->Opcode() == Op_Opaque1, "must be");
1773 igvn.replace_node(n, n->in(1));
1774 }
1775 assert(predicate_count()==0, "should be clean!");
1776 }
1777
1778 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1779 if (!n->for_post_loop_opts_igvn()) {
1780 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1781 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1782 _for_post_loop_igvn.append(n);
1783 }
1784 }
1785
1786 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1787 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1788 _for_post_loop_igvn.remove(n);
1789 }
1790
1791 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1792 // Verify that all previous optimizations produced a valid graph
1793 // at least to this point, even if no loop optimizations were done.
1794 PhaseIdealLoop::verify(igvn);
1795
1796 C->set_post_loop_opts_phase(); // no more loop opts allowed
1797
1798 assert(!C->major_progress(), "not cleared");
1799
1800 if (_for_post_loop_igvn.length() > 0) {
1801 while (_for_post_loop_igvn.length() > 0) {
1802 Node* n = _for_post_loop_igvn.pop();
1803 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1804 igvn._worklist.push(n);
1805 }
1806 igvn.optimize();
1807 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1808
1809 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1810 if (C->major_progress()) {
1811 C->clear_major_progress(); // ensure that major progress is now clear
1812 }
1813 }
1814 }
1815
1816 // StringOpts and late inlining of string methods
1817 void Compile::inline_string_calls(bool parse_time) {
1818 {
1819 // remove useless nodes to make the usage analysis simpler
1820 ResourceMark rm;
1821 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1822 }
1823
1824 {
1825 ResourceMark rm;
1826 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1827 PhaseStringOpts pso(initial_gvn(), for_igvn());
1828 print_method(PHASE_AFTER_STRINGOPTS, 3);
1829 }
1830
1831 // now inline anything that we skipped the first time around
1832 if (!parse_time) {
1833 _late_inlines_pos = _late_inlines.length();
1834 }
1835
1836 while (_string_late_inlines.length() > 0) {
1837 CallGenerator* cg = _string_late_inlines.pop();
1838 cg->do_late_inline();
1839 if (failing()) return;
1840 }
1841 _string_late_inlines.trunc_to(0);
1842 }
1843
1844 // Late inlining of boxing methods
1845 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1846 if (_boxing_late_inlines.length() > 0) {
1847 assert(has_boxed_value(), "inconsistent");
1848
1849 PhaseGVN* gvn = initial_gvn();
1850 set_inlining_incrementally(true);
1851
1852 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1853 for_igvn()->clear();
1854 gvn->replace_with(&igvn);
1855
1856 _late_inlines_pos = _late_inlines.length();
1857
1858 while (_boxing_late_inlines.length() > 0) {
1859 CallGenerator* cg = _boxing_late_inlines.pop();
1860 cg->do_late_inline();
1861 if (failing()) return;
1862 }
1863 _boxing_late_inlines.trunc_to(0);
1864
1865 inline_incrementally_cleanup(igvn);
1866
1867 set_inlining_incrementally(false);
1868 }
1869 }
1870
1871 bool Compile::inline_incrementally_one() {
1872 assert(IncrementalInline, "incremental inlining should be on");
1873
1874 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1875
1876 set_inlining_progress(false);
1877 set_do_cleanup(false);
1878
1879 for (int i = 0; i < _late_inlines.length(); i++) {
1880 _late_inlines_pos = i+1;
1881 CallGenerator* cg = _late_inlines.at(i);
1882 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
1883 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
1884 cg->do_late_inline();
1885 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
1886 if (failing()) {
1887 return false;
1888 } else if (inlining_progress()) {
1889 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
1890 print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3);
1891 break; // process one call site at a time
1892 }
1893 } else {
1894 // Ignore late inline direct calls when inlining is not allowed.
1895 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
1896 }
1897 }
1898 // Remove processed elements.
1899 _late_inlines.remove_till(_late_inlines_pos);
1900 _late_inlines_pos = 0;
1901
1902 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
1903
1904 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
1905
1906 set_inlining_progress(false);
1907 set_do_cleanup(false);
1908
1909 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
1910 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
1911 }
1912
1913 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
1914 {
1915 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
1916 ResourceMark rm;
1917 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1918 }
1919 {
1920 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
1921 igvn = PhaseIterGVN(initial_gvn());
1922 igvn.optimize();
1923 }
1924 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
1925 }
1926
1927 // Perform incremental inlining until bound on number of live nodes is reached
1928 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1929 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
1930
1931 set_inlining_incrementally(true);
1932 uint low_live_nodes = 0;
1933
1934 while (_late_inlines.length() > 0) {
1935 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1936 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1937 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
1938 // PhaseIdealLoop is expensive so we only try it once we are
1939 // out of live nodes and we only try it again if the previous
1940 // helped got the number of nodes down significantly
1941 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
1942 if (failing()) return;
1943 low_live_nodes = live_nodes();
1944 _major_progress = true;
1945 }
1946
1947 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1948 bool do_print_inlining = print_inlining() || print_intrinsics();
1949 if (do_print_inlining || log() != nullptr) {
1950 // Print inlining message for candidates that we couldn't inline for lack of space.
1951 for (int i = 0; i < _late_inlines.length(); i++) {
1952 CallGenerator* cg = _late_inlines.at(i);
1953 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
1954 if (do_print_inlining) {
1955 cg->print_inlining_late(msg);
1956 }
1957 log_late_inline_failure(cg, msg);
1958 }
1959 }
1960 break; // finish
1961 }
1962 }
1963
1964 for_igvn()->clear();
1965 initial_gvn()->replace_with(&igvn);
1966
1967 while (inline_incrementally_one()) {
1968 assert(!failing(), "inconsistent");
1969 }
1970 if (failing()) return;
1971
1972 inline_incrementally_cleanup(igvn);
1973
1974 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
1975
1976 if (failing()) return;
1977
1978 if (_late_inlines.length() == 0) {
1979 break; // no more progress
1980 }
1981 }
1982 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1983
1984 if (_string_late_inlines.length() > 0) {
1985 assert(has_stringbuilder(), "inconsistent");
1986 for_igvn()->clear();
1987 initial_gvn()->replace_with(&igvn);
1988
1989 inline_string_calls(false);
1990
1991 if (failing()) return;
1992
1993 inline_incrementally_cleanup(igvn);
1994 }
1995
1996 set_inlining_incrementally(false);
1997 }
1998
1999 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2000 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2001 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2002 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2003 // as if "inlining_incrementally() == true" were set.
2004 assert(inlining_incrementally() == false, "not allowed");
2005 assert(_modified_nodes == nullptr, "not allowed");
2006 assert(_late_inlines.length() > 0, "sanity");
2007
2008 while (_late_inlines.length() > 0) {
2009 for_igvn()->clear();
2010 initial_gvn()->replace_with(&igvn);
2011
2012 while (inline_incrementally_one()) {
2013 assert(!failing(), "inconsistent");
2014 }
2015 if (failing()) return;
2016
2017 inline_incrementally_cleanup(igvn);
2018 }
2019 }
2020
2021 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2022 if (_loop_opts_cnt > 0) {
2023 while (major_progress() && (_loop_opts_cnt > 0)) {
2024 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2025 PhaseIdealLoop::optimize(igvn, mode);
2026 _loop_opts_cnt--;
2027 if (failing()) return false;
2028 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2029 }
2030 }
2031 return true;
2032 }
2033
2034 // Remove edges from "root" to each SafePoint at a backward branch.
2035 // They were inserted during parsing (see add_safepoint()) to make
2036 // infinite loops without calls or exceptions visible to root, i.e.,
2037 // useful.
2038 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2039 Node *r = root();
2040 if (r != nullptr) {
2041 for (uint i = r->req(); i < r->len(); ++i) {
2042 Node *n = r->in(i);
2043 if (n != nullptr && n->is_SafePoint()) {
2044 r->rm_prec(i);
2045 if (n->outcnt() == 0) {
2046 igvn.remove_dead_node(n);
2047 }
2048 --i;
2049 }
2050 }
2051 // Parsing may have added top inputs to the root node (Path
2052 // leading to the Halt node proven dead). Make sure we get a
2053 // chance to clean them up.
2054 igvn._worklist.push(r);
2055 igvn.optimize();
2056 }
2057 }
2058
2059 //------------------------------Optimize---------------------------------------
2060 // Given a graph, optimize it.
2061 void Compile::Optimize() {
2062 TracePhase tp("optimizer", &timers[_t_optimizer]);
2063
2064 #ifndef PRODUCT
2065 if (env()->break_at_compile()) {
2066 BREAKPOINT;
2067 }
2068
2069 #endif
2070
2071 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2072 #ifdef ASSERT
2073 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2074 #endif
2075
2076 ResourceMark rm;
2077
2078 print_inlining_reinit();
2079
2080 NOT_PRODUCT( verify_graph_edges(); )
2081
2082 print_method(PHASE_AFTER_PARSING);
2083
2084 {
2085 // Iterative Global Value Numbering, including ideal transforms
2086 // Initialize IterGVN with types and values from parse-time GVN
2087 PhaseIterGVN igvn(initial_gvn());
2088 #ifdef ASSERT
2089 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2090 #endif
2091 {
2092 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2093 igvn.optimize();
2094 }
2095
2096 if (failing()) return;
2097
2098 print_method(PHASE_ITER_GVN1, 2);
2099
2100 inline_incrementally(igvn);
2101
2102 print_method(PHASE_INCREMENTAL_INLINE, 2);
2103
2104 if (failing()) return;
2105
2106 if (eliminate_boxing()) {
2107 // Inline valueOf() methods now.
2108 inline_boxing_calls(igvn);
2109
2110 if (failing()) return;
2111
2112 if (AlwaysIncrementalInline) {
2113 inline_incrementally(igvn);
2114 }
2115
2116 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2117
2118 if (failing()) return;
2119 }
2120
2121 // Remove the speculative part of types and clean up the graph from
2122 // the extra CastPP nodes whose only purpose is to carry them. Do
2123 // that early so that optimizations are not disrupted by the extra
2124 // CastPP nodes.
2125 remove_speculative_types(igvn);
2126
2127 if (failing()) return;
2128
2129 // No more new expensive nodes will be added to the list from here
2130 // so keep only the actual candidates for optimizations.
2131 cleanup_expensive_nodes(igvn);
2132
2133 if (failing()) return;
2134
2135 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2136 if (EnableVectorSupport && has_vbox_nodes()) {
2137 TracePhase tp("", &timers[_t_vector]);
2138 PhaseVector pv(igvn);
2139 pv.optimize_vector_boxes();
2140 if (failing()) return;
2141 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2142 }
2143 assert(!has_vbox_nodes(), "sanity");
2144
2145 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2146 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2147 initial_gvn()->replace_with(&igvn);
2148 for_igvn()->clear();
2149 Unique_Node_List new_worklist(C->comp_arena());
2150 {
2151 ResourceMark rm;
2152 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2153 }
2154 Unique_Node_List* save_for_igvn = for_igvn();
2155 set_for_igvn(&new_worklist);
2156 igvn = PhaseIterGVN(initial_gvn());
2157 igvn.optimize();
2158 set_for_igvn(save_for_igvn);
2159 }
2160
2161 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2162 // safepoints
2163 remove_root_to_sfpts_edges(igvn);
2164
2165 if (failing()) return;
2166
2167 // Perform escape analysis
2168 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2169 if (has_loops()) {
2170 // Cleanup graph (remove dead nodes).
2171 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2172 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2173 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2174 if (failing()) return;
2175 }
2176 ConnectionGraph::do_analysis(this, &igvn);
2177
2178 if (failing()) return;
2179
2180 // Optimize out fields loads from scalar replaceable allocations.
2181 igvn.optimize();
2182 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2183
2184 if (failing()) return;
2185
2186 if (congraph() != nullptr && macro_count() > 0) {
2187 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2188 PhaseMacroExpand mexp(igvn);
2189 mexp.eliminate_macro_nodes();
2190 igvn.set_delay_transform(false);
2191
2192 igvn.optimize();
2193 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2194
2195 if (failing()) return;
2196 }
2197 }
2198
2199 // Loop transforms on the ideal graph. Range Check Elimination,
2200 // peeling, unrolling, etc.
2201
2202 // Set loop opts counter
2203 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2204 {
2205 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2206 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2207 _loop_opts_cnt--;
2208 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2209 if (failing()) return;
2210 }
2211 // Loop opts pass if partial peeling occurred in previous pass
2212 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2213 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2214 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2215 _loop_opts_cnt--;
2216 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2217 if (failing()) return;
2218 }
2219 // Loop opts pass for loop-unrolling before CCP
2220 if(major_progress() && (_loop_opts_cnt > 0)) {
2221 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2222 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2223 _loop_opts_cnt--;
2224 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2225 }
2226 if (!failing()) {
2227 // Verify that last round of loop opts produced a valid graph
2228 PhaseIdealLoop::verify(igvn);
2229 }
2230 }
2231 if (failing()) return;
2232
2233 // Conditional Constant Propagation;
2234 PhaseCCP ccp( &igvn );
2235 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2236 {
2237 TracePhase tp("ccp", &timers[_t_ccp]);
2238 ccp.do_transform();
2239 }
2240 print_method(PHASE_CCP1, 2);
2241
2242 assert( true, "Break here to ccp.dump_old2new_map()");
2243
2244 // Iterative Global Value Numbering, including ideal transforms
2245 {
2246 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2247 igvn = ccp;
2248 igvn.optimize();
2249 }
2250 print_method(PHASE_ITER_GVN2, 2);
2251
2252 if (failing()) return;
2253
2254 // Loop transforms on the ideal graph. Range Check Elimination,
2255 // peeling, unrolling, etc.
2256 if (!optimize_loops(igvn, LoopOptsDefault)) {
2257 return;
2258 }
2259
2260 if (failing()) return;
2261
2262 C->clear_major_progress(); // ensure that major progress is now clear
2263
2264 process_for_post_loop_opts_igvn(igvn);
2265
2266 if (failing()) return;
2267
2268 #ifdef ASSERT
2269 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2270 #endif
2271
2272 {
2273 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2274 PhaseMacroExpand mex(igvn);
2275 if (mex.expand_macro_nodes()) {
2276 assert(failing(), "must bail out w/ explicit message");
2277 return;
2278 }
2279 print_method(PHASE_MACRO_EXPANSION, 2);
2280 }
2281
2282 {
2283 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2284 if (bs->expand_barriers(this, igvn)) {
2285 assert(failing(), "must bail out w/ explicit message");
2286 return;
2287 }
2288 print_method(PHASE_BARRIER_EXPANSION, 2);
2289 }
2290
2291 if (C->max_vector_size() > 0) {
2292 C->optimize_logic_cones(igvn);
2293 igvn.optimize();
2294 }
2295
2296 DEBUG_ONLY( _modified_nodes = nullptr; )
2297
2298 assert(igvn._worklist.size() == 0, "not empty");
2299
2300 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2301
2302 if (_late_inlines.length() > 0) {
2303 // More opportunities to optimize virtual and MH calls.
2304 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2305 process_late_inline_calls_no_inline(igvn);
2306 if (failing()) return;
2307 }
2308 } // (End scope of igvn; run destructor if necessary for asserts.)
2309
2310 check_no_dead_use();
2311
2312 process_print_inlining();
2313
2314 // A method with only infinite loops has no edges entering loops from root
2315 {
2316 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2317 if (final_graph_reshaping()) {
2318 assert(failing(), "must bail out w/ explicit message");
2319 return;
2320 }
2321 }
2322
2323 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2324 DEBUG_ONLY(set_phase_optimize_finished();)
2325 }
2326
2327 #ifdef ASSERT
2328 void Compile::check_no_dead_use() const {
2329 ResourceMark rm;
2330 Unique_Node_List wq;
2331 wq.push(root());
2332 for (uint i = 0; i < wq.size(); ++i) {
2333 Node* n = wq.at(i);
2334 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2335 Node* u = n->fast_out(j);
2336 if (u->outcnt() == 0 && !u->is_Con()) {
2337 u->dump();
2338 fatal("no reachable node should have no use");
2339 }
2340 wq.push(u);
2341 }
2342 }
2343 }
2344 #endif
2345
2346 void Compile::inline_vector_reboxing_calls() {
2347 if (C->_vector_reboxing_late_inlines.length() > 0) {
2348 _late_inlines_pos = C->_late_inlines.length();
2349 while (_vector_reboxing_late_inlines.length() > 0) {
2350 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2351 cg->do_late_inline();
2352 if (failing()) return;
2353 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2354 }
2355 _vector_reboxing_late_inlines.trunc_to(0);
2356 }
2357 }
2358
2359 bool Compile::has_vbox_nodes() {
2360 if (C->_vector_reboxing_late_inlines.length() > 0) {
2361 return true;
2362 }
2363 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2364 Node * n = C->macro_node(macro_idx);
2365 assert(n->is_macro(), "only macro nodes expected here");
2366 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2367 return true;
2368 }
2369 }
2370 return false;
2371 }
2372
2373 //---------------------------- Bitwise operation packing optimization ---------------------------
2374
2375 static bool is_vector_unary_bitwise_op(Node* n) {
2376 return n->Opcode() == Op_XorV &&
2377 VectorNode::is_vector_bitwise_not_pattern(n);
2378 }
2379
2380 static bool is_vector_binary_bitwise_op(Node* n) {
2381 switch (n->Opcode()) {
2382 case Op_AndV:
2383 case Op_OrV:
2384 return true;
2385
2386 case Op_XorV:
2387 return !is_vector_unary_bitwise_op(n);
2388
2389 default:
2390 return false;
2391 }
2392 }
2393
2394 static bool is_vector_ternary_bitwise_op(Node* n) {
2395 return n->Opcode() == Op_MacroLogicV;
2396 }
2397
2398 static bool is_vector_bitwise_op(Node* n) {
2399 return is_vector_unary_bitwise_op(n) ||
2400 is_vector_binary_bitwise_op(n) ||
2401 is_vector_ternary_bitwise_op(n);
2402 }
2403
2404 static bool is_vector_bitwise_cone_root(Node* n) {
2405 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2406 return false;
2407 }
2408 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2409 if (is_vector_bitwise_op(n->fast_out(i))) {
2410 return false;
2411 }
2412 }
2413 return true;
2414 }
2415
2416 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2417 uint cnt = 0;
2418 if (is_vector_bitwise_op(n)) {
2419 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2420 for (uint i = 1; i < n->req(); i++) {
2421 Node* in = n->in(i);
2422 bool skip = VectorNode::is_all_ones_vector(in);
2423 if (!skip && !inputs.member(in)) {
2424 inputs.push(in);
2425 cnt++;
2426 }
2427 }
2428 assert(cnt <= 1, "not unary");
2429 } else {
2430 uint last_req = n->req();
2431 if (is_vector_ternary_bitwise_op(n)) {
2432 last_req = n->req() - 1; // skip last input
2433 }
2434 for (uint i = 1; i < last_req; i++) {
2435 Node* def = n->in(i);
2436 if (!inputs.member(def)) {
2437 inputs.push(def);
2438 cnt++;
2439 }
2440 }
2441 }
2442 partition.push(n);
2443 } else { // not a bitwise operations
2444 if (!inputs.member(n)) {
2445 inputs.push(n);
2446 cnt++;
2447 }
2448 }
2449 return cnt;
2450 }
2451
2452 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2453 Unique_Node_List useful_nodes;
2454 C->identify_useful_nodes(useful_nodes);
2455
2456 for (uint i = 0; i < useful_nodes.size(); i++) {
2457 Node* n = useful_nodes.at(i);
2458 if (is_vector_bitwise_cone_root(n)) {
2459 list.push(n);
2460 }
2461 }
2462 }
2463
2464 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2465 const TypeVect* vt,
2466 Unique_Node_List& partition,
2467 Unique_Node_List& inputs) {
2468 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2469 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2470 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2471
2472 Node* in1 = inputs.at(0);
2473 Node* in2 = inputs.at(1);
2474 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2475
2476 uint func = compute_truth_table(partition, inputs);
2477 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt));
2478 }
2479
2480 static uint extract_bit(uint func, uint pos) {
2481 return (func & (1 << pos)) >> pos;
2482 }
2483
2484 //
2485 // A macro logic node represents a truth table. It has 4 inputs,
2486 // First three inputs corresponds to 3 columns of a truth table
2487 // and fourth input captures the logic function.
2488 //
2489 // eg. fn = (in1 AND in2) OR in3;
2490 //
2491 // MacroNode(in1,in2,in3,fn)
2492 //
2493 // -----------------
2494 // in1 in2 in3 fn
2495 // -----------------
2496 // 0 0 0 0
2497 // 0 0 1 1
2498 // 0 1 0 0
2499 // 0 1 1 1
2500 // 1 0 0 0
2501 // 1 0 1 1
2502 // 1 1 0 1
2503 // 1 1 1 1
2504 //
2505
2506 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2507 int res = 0;
2508 for (int i = 0; i < 8; i++) {
2509 int bit1 = extract_bit(in1, i);
2510 int bit2 = extract_bit(in2, i);
2511 int bit3 = extract_bit(in3, i);
2512
2513 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2514 int func_bit = extract_bit(func, func_bit_pos);
2515
2516 res |= func_bit << i;
2517 }
2518 return res;
2519 }
2520
2521 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2522 assert(n != nullptr, "");
2523 assert(eval_map.contains(n), "absent");
2524 return *(eval_map.get(n));
2525 }
2526
2527 static void eval_operands(Node* n,
2528 uint& func1, uint& func2, uint& func3,
2529 ResourceHashtable<Node*,uint>& eval_map) {
2530 assert(is_vector_bitwise_op(n), "");
2531
2532 if (is_vector_unary_bitwise_op(n)) {
2533 Node* opnd = n->in(1);
2534 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2535 opnd = n->in(2);
2536 }
2537 func1 = eval_operand(opnd, eval_map);
2538 } else if (is_vector_binary_bitwise_op(n)) {
2539 func1 = eval_operand(n->in(1), eval_map);
2540 func2 = eval_operand(n->in(2), eval_map);
2541 } else {
2542 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2543 func1 = eval_operand(n->in(1), eval_map);
2544 func2 = eval_operand(n->in(2), eval_map);
2545 func3 = eval_operand(n->in(3), eval_map);
2546 }
2547 }
2548
2549 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2550 assert(inputs.size() <= 3, "sanity");
2551 ResourceMark rm;
2552 uint res = 0;
2553 ResourceHashtable<Node*,uint> eval_map;
2554
2555 // Populate precomputed functions for inputs.
2556 // Each input corresponds to one column of 3 input truth-table.
2557 uint input_funcs[] = { 0xAA, // (_, _, a) -> a
2558 0xCC, // (_, b, _) -> b
2559 0xF0 }; // (c, _, _) -> c
2560 for (uint i = 0; i < inputs.size(); i++) {
2561 eval_map.put(inputs.at(i), input_funcs[i]);
2562 }
2563
2564 for (uint i = 0; i < partition.size(); i++) {
2565 Node* n = partition.at(i);
2566
2567 uint func1 = 0, func2 = 0, func3 = 0;
2568 eval_operands(n, func1, func2, func3, eval_map);
2569
2570 switch (n->Opcode()) {
2571 case Op_OrV:
2572 assert(func3 == 0, "not binary");
2573 res = func1 | func2;
2574 break;
2575 case Op_AndV:
2576 assert(func3 == 0, "not binary");
2577 res = func1 & func2;
2578 break;
2579 case Op_XorV:
2580 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2581 assert(func2 == 0 && func3 == 0, "not unary");
2582 res = (~func1) & 0xFF;
2583 } else {
2584 assert(func3 == 0, "not binary");
2585 res = func1 ^ func2;
2586 }
2587 break;
2588 case Op_MacroLogicV:
2589 // Ordering of inputs may change during evaluation of sub-tree
2590 // containing MacroLogic node as a child node, thus a re-evaluation
2591 // makes sure that function is evaluated in context of current
2592 // inputs.
2593 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2594 break;
2595
2596 default: assert(false, "not supported: %s", n->Name());
2597 }
2598 assert(res <= 0xFF, "invalid");
2599 eval_map.put(n, res);
2600 }
2601 return res;
2602 }
2603
2604 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2605 assert(partition.size() == 0, "not empty");
2606 assert(inputs.size() == 0, "not empty");
2607 if (is_vector_ternary_bitwise_op(n)) {
2608 return false;
2609 }
2610
2611 bool is_unary_op = is_vector_unary_bitwise_op(n);
2612 if (is_unary_op) {
2613 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary");
2614 return false; // too few inputs
2615 }
2616
2617 assert(is_vector_binary_bitwise_op(n), "not binary");
2618 Node* in1 = n->in(1);
2619 Node* in2 = n->in(2);
2620
2621 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs);
2622 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs);
2623 partition.push(n);
2624
2625 // Too many inputs?
2626 if (inputs.size() > 3) {
2627 partition.clear();
2628 inputs.clear();
2629 { // Recompute in2 inputs
2630 Unique_Node_List not_used;
2631 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used);
2632 }
2633 // Pick the node with minimum number of inputs.
2634 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) {
2635 return false; // still too many inputs
2636 }
2637 // Recompute partition & inputs.
2638 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2);
2639 collect_unique_inputs(child, partition, inputs);
2640
2641 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1);
2642 inputs.push(other_input);
2643
2644 partition.push(n);
2645 }
2646
2647 return (partition.size() == 2 || partition.size() == 3) &&
2648 (inputs.size() == 2 || inputs.size() == 3);
2649 }
2650
2651
2652 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2653 assert(is_vector_bitwise_op(n), "not a root");
2654
2655 visited.set(n->_idx);
2656
2657 // 1) Do a DFS walk over the logic cone.
2658 for (uint i = 1; i < n->req(); i++) {
2659 Node* in = n->in(i);
2660 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2661 process_logic_cone_root(igvn, in, visited);
2662 }
2663 }
2664
2665 // 2) Bottom up traversal: Merge node[s] with
2666 // the parent to form macro logic node.
2667 Unique_Node_List partition;
2668 Unique_Node_List inputs;
2669 if (compute_logic_cone(n, partition, inputs)) {
2670 const TypeVect* vt = n->bottom_type()->is_vect();
2671 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2672 igvn.replace_node(n, macro_logic);
2673 }
2674 }
2675
2676 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2677 ResourceMark rm;
2678 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2679 Unique_Node_List list;
2680 collect_logic_cone_roots(list);
2681
2682 while (list.size() > 0) {
2683 Node* n = list.pop();
2684 const TypeVect* vt = n->bottom_type()->is_vect();
2685 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2686 if (supported) {
2687 VectorSet visited(comp_arena());
2688 process_logic_cone_root(igvn, n, visited);
2689 }
2690 }
2691 }
2692 }
2693
2694 //------------------------------Code_Gen---------------------------------------
2695 // Given a graph, generate code for it
2696 void Compile::Code_Gen() {
2697 if (failing()) {
2698 return;
2699 }
2700
2701 // Perform instruction selection. You might think we could reclaim Matcher
2702 // memory PDQ, but actually the Matcher is used in generating spill code.
2703 // Internals of the Matcher (including some VectorSets) must remain live
2704 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2705 // set a bit in reclaimed memory.
2706
2707 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2708 // nodes. Mapping is only valid at the root of each matched subtree.
2709 NOT_PRODUCT( verify_graph_edges(); )
2710
2711 Matcher matcher;
2712 _matcher = &matcher;
2713 {
2714 TracePhase tp("matcher", &timers[_t_matcher]);
2715 matcher.match();
2716 if (failing()) {
2717 return;
2718 }
2719 }
2720 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2721 // nodes. Mapping is only valid at the root of each matched subtree.
2722 NOT_PRODUCT( verify_graph_edges(); )
2723
2724 // If you have too many nodes, or if matching has failed, bail out
2725 check_node_count(0, "out of nodes matching instructions");
2726 if (failing()) {
2727 return;
2728 }
2729
2730 print_method(PHASE_MATCHING, 2);
2731
2732 // Build a proper-looking CFG
2733 PhaseCFG cfg(node_arena(), root(), matcher);
2734 _cfg = &cfg;
2735 {
2736 TracePhase tp("scheduler", &timers[_t_scheduler]);
2737 bool success = cfg.do_global_code_motion();
2738 if (!success) {
2739 return;
2740 }
2741
2742 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2743 NOT_PRODUCT( verify_graph_edges(); )
2744 cfg.verify();
2745 }
2746
2747 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2748 _regalloc = ®alloc;
2749 {
2750 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2751 // Perform register allocation. After Chaitin, use-def chains are
2752 // no longer accurate (at spill code) and so must be ignored.
2753 // Node->LRG->reg mappings are still accurate.
2754 _regalloc->Register_Allocate();
2755
2756 // Bail out if the allocator builds too many nodes
2757 if (failing()) {
2758 return;
2759 }
2760 }
2761
2762 // Prior to register allocation we kept empty basic blocks in case the
2763 // the allocator needed a place to spill. After register allocation we
2764 // are not adding any new instructions. If any basic block is empty, we
2765 // can now safely remove it.
2766 {
2767 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2768 cfg.remove_empty_blocks();
2769 if (do_freq_based_layout()) {
2770 PhaseBlockLayout layout(cfg);
2771 } else {
2772 cfg.set_loop_alignment();
2773 }
2774 cfg.fixup_flow();
2775 cfg.remove_unreachable_blocks();
2776 cfg.verify_dominator_tree();
2777 }
2778
2779 // Apply peephole optimizations
2780 if( OptoPeephole ) {
2781 TracePhase tp("peephole", &timers[_t_peephole]);
2782 PhasePeephole peep( _regalloc, cfg);
2783 peep.do_transform();
2784 }
2785
2786 // Do late expand if CPU requires this.
2787 if (Matcher::require_postalloc_expand) {
2788 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2789 cfg.postalloc_expand(_regalloc);
2790 }
2791
2792 // Convert Nodes to instruction bits in a buffer
2793 {
2794 TracePhase tp("output", &timers[_t_output]);
2795 PhaseOutput output;
2796 output.Output();
2797 if (failing()) return;
2798 output.install();
2799 }
2800
2801 print_method(PHASE_FINAL_CODE);
2802
2803 // He's dead, Jim.
2804 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2805 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2806 }
2807
2808 //------------------------------Final_Reshape_Counts---------------------------
2809 // This class defines counters to help identify when a method
2810 // may/must be executed using hardware with only 24-bit precision.
2811 struct Final_Reshape_Counts : public StackObj {
2812 int _call_count; // count non-inlined 'common' calls
2813 int _float_count; // count float ops requiring 24-bit precision
2814 int _double_count; // count double ops requiring more precision
2815 int _java_call_count; // count non-inlined 'java' calls
2816 int _inner_loop_count; // count loops which need alignment
2817 VectorSet _visited; // Visitation flags
2818 Node_List _tests; // Set of IfNodes & PCTableNodes
2819
2820 Final_Reshape_Counts() :
2821 _call_count(0), _float_count(0), _double_count(0),
2822 _java_call_count(0), _inner_loop_count(0) { }
2823
2824 void inc_call_count () { _call_count ++; }
2825 void inc_float_count () { _float_count ++; }
2826 void inc_double_count() { _double_count++; }
2827 void inc_java_call_count() { _java_call_count++; }
2828 void inc_inner_loop_count() { _inner_loop_count++; }
2829
2830 int get_call_count () const { return _call_count ; }
2831 int get_float_count () const { return _float_count ; }
2832 int get_double_count() const { return _double_count; }
2833 int get_java_call_count() const { return _java_call_count; }
2834 int get_inner_loop_count() const { return _inner_loop_count; }
2835 };
2836
2837 // Eliminate trivially redundant StoreCMs and accumulate their
2838 // precedence edges.
2839 void Compile::eliminate_redundant_card_marks(Node* n) {
2840 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2841 if (n->in(MemNode::Address)->outcnt() > 1) {
2842 // There are multiple users of the same address so it might be
2843 // possible to eliminate some of the StoreCMs
2844 Node* mem = n->in(MemNode::Memory);
2845 Node* adr = n->in(MemNode::Address);
2846 Node* val = n->in(MemNode::ValueIn);
2847 Node* prev = n;
2848 bool done = false;
2849 // Walk the chain of StoreCMs eliminating ones that match. As
2850 // long as it's a chain of single users then the optimization is
2851 // safe. Eliminating partially redundant StoreCMs would require
2852 // cloning copies down the other paths.
2853 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2854 if (adr == mem->in(MemNode::Address) &&
2855 val == mem->in(MemNode::ValueIn)) {
2856 // redundant StoreCM
2857 if (mem->req() > MemNode::OopStore) {
2858 // Hasn't been processed by this code yet.
2859 n->add_prec(mem->in(MemNode::OopStore));
2860 } else {
2861 // Already converted to precedence edge
2862 for (uint i = mem->req(); i < mem->len(); i++) {
2863 // Accumulate any precedence edges
2864 if (mem->in(i) != nullptr) {
2865 n->add_prec(mem->in(i));
2866 }
2867 }
2868 // Everything above this point has been processed.
2869 done = true;
2870 }
2871 // Eliminate the previous StoreCM
2872 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2873 assert(mem->outcnt() == 0, "should be dead");
2874 mem->disconnect_inputs(this);
2875 } else {
2876 prev = mem;
2877 }
2878 mem = prev->in(MemNode::Memory);
2879 }
2880 }
2881 }
2882
2883 //------------------------------final_graph_reshaping_impl----------------------
2884 // Implement items 1-5 from final_graph_reshaping below.
2885 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
2886
2887 if ( n->outcnt() == 0 ) return; // dead node
2888 uint nop = n->Opcode();
2889
2890 // Check for 2-input instruction with "last use" on right input.
2891 // Swap to left input. Implements item (2).
2892 if( n->req() == 3 && // two-input instruction
2893 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2894 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2895 n->in(2)->outcnt() == 1 &&// right use IS a last use
2896 !n->in(2)->is_Con() ) { // right use is not a constant
2897 // Check for commutative opcode
2898 switch( nop ) {
2899 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2900 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
2901 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
2902 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2903 case Op_AndL: case Op_XorL: case Op_OrL:
2904 case Op_AndI: case Op_XorI: case Op_OrI: {
2905 // Move "last use" input to left by swapping inputs
2906 n->swap_edges(1, 2);
2907 break;
2908 }
2909 default:
2910 break;
2911 }
2912 }
2913
2914 #ifdef ASSERT
2915 if( n->is_Mem() ) {
2916 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2917 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
2918 // oop will be recorded in oop map if load crosses safepoint
2919 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2920 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2921 "raw memory operations should have control edge");
2922 }
2923 if (n->is_MemBar()) {
2924 MemBarNode* mb = n->as_MemBar();
2925 if (mb->trailing_store() || mb->trailing_load_store()) {
2926 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2927 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2928 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2929 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2930 } else if (mb->leading()) {
2931 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2932 }
2933 }
2934 #endif
2935 // Count FPU ops and common calls, implements item (3)
2936 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
2937 if (!gc_handled) {
2938 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
2939 }
2940
2941 // Collect CFG split points
2942 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2943 frc._tests.push(n);
2944 }
2945 }
2946
2947 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
2948 switch( nop ) {
2949 // Count all float operations that may use FPU
2950 case Op_AddF:
2951 case Op_SubF:
2952 case Op_MulF:
2953 case Op_DivF:
2954 case Op_NegF:
2955 case Op_ModF:
2956 case Op_ConvI2F:
2957 case Op_ConF:
2958 case Op_CmpF:
2959 case Op_CmpF3:
2960 case Op_StoreF:
2961 case Op_LoadF:
2962 // case Op_ConvL2F: // longs are split into 32-bit halves
2963 frc.inc_float_count();
2964 break;
2965
2966 case Op_ConvF2D:
2967 case Op_ConvD2F:
2968 frc.inc_float_count();
2969 frc.inc_double_count();
2970 break;
2971
2972 // Count all double operations that may use FPU
2973 case Op_AddD:
2974 case Op_SubD:
2975 case Op_MulD:
2976 case Op_DivD:
2977 case Op_NegD:
2978 case Op_ModD:
2979 case Op_ConvI2D:
2980 case Op_ConvD2I:
2981 // case Op_ConvL2D: // handled by leaf call
2982 // case Op_ConvD2L: // handled by leaf call
2983 case Op_ConD:
2984 case Op_CmpD:
2985 case Op_CmpD3:
2986 case Op_StoreD:
2987 case Op_LoadD:
2988 case Op_LoadD_unaligned:
2989 frc.inc_double_count();
2990 break;
2991 case Op_Opaque1: // Remove Opaque Nodes before matching
2992 case Op_Opaque2: // Remove Opaque Nodes before matching
2993 case Op_Opaque3:
2994 n->subsume_by(n->in(1), this);
2995 break;
2996 case Op_CallStaticJava:
2997 case Op_CallJava:
2998 case Op_CallDynamicJava:
2999 frc.inc_java_call_count(); // Count java call site;
3000 case Op_CallRuntime:
3001 case Op_CallLeaf:
3002 case Op_CallLeafVector:
3003 case Op_CallNative:
3004 case Op_CallLeafNoFP: {
3005 assert (n->is_Call(), "");
3006 CallNode *call = n->as_Call();
3007 // Count call sites where the FP mode bit would have to be flipped.
3008 // Do not count uncommon runtime calls:
3009 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3010 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3011 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3012 frc.inc_call_count(); // Count the call site
3013 } else { // See if uncommon argument is shared
3014 Node *n = call->in(TypeFunc::Parms);
3015 int nop = n->Opcode();
3016 // Clone shared simple arguments to uncommon calls, item (1).
3017 if (n->outcnt() > 1 &&
3018 !n->is_Proj() &&
3019 nop != Op_CreateEx &&
3020 nop != Op_CheckCastPP &&
3021 nop != Op_DecodeN &&
3022 nop != Op_DecodeNKlass &&
3023 !n->is_Mem() &&
3024 !n->is_Phi()) {
3025 Node *x = n->clone();
3026 call->set_req(TypeFunc::Parms, x);
3027 }
3028 }
3029 break;
3030 }
3031
3032 case Op_StoreCM:
3033 {
3034 // Convert OopStore dependence into precedence edge
3035 Node* prec = n->in(MemNode::OopStore);
3036 n->del_req(MemNode::OopStore);
3037 n->add_prec(prec);
3038 eliminate_redundant_card_marks(n);
3039 }
3040
3041 // fall through
3042
3043 case Op_StoreB:
3044 case Op_StoreC:
3045 case Op_StorePConditional:
3046 case Op_StoreI:
3047 case Op_StoreL:
3048 case Op_StoreIConditional:
3049 case Op_StoreLConditional:
3050 case Op_CompareAndSwapB:
3051 case Op_CompareAndSwapS:
3052 case Op_CompareAndSwapI:
3053 case Op_CompareAndSwapL:
3054 case Op_CompareAndSwapP:
3055 case Op_CompareAndSwapN:
3056 case Op_WeakCompareAndSwapB:
3057 case Op_WeakCompareAndSwapS:
3058 case Op_WeakCompareAndSwapI:
3059 case Op_WeakCompareAndSwapL:
3060 case Op_WeakCompareAndSwapP:
3061 case Op_WeakCompareAndSwapN:
3062 case Op_CompareAndExchangeB:
3063 case Op_CompareAndExchangeS:
3064 case Op_CompareAndExchangeI:
3065 case Op_CompareAndExchangeL:
3066 case Op_CompareAndExchangeP:
3067 case Op_CompareAndExchangeN:
3068 case Op_GetAndAddS:
3069 case Op_GetAndAddB:
3070 case Op_GetAndAddI:
3071 case Op_GetAndAddL:
3072 case Op_GetAndSetS:
3073 case Op_GetAndSetB:
3074 case Op_GetAndSetI:
3075 case Op_GetAndSetL:
3076 case Op_GetAndSetP:
3077 case Op_GetAndSetN:
3078 case Op_StoreP:
3079 case Op_StoreN:
3080 case Op_StoreNKlass:
3081 case Op_LoadB:
3082 case Op_LoadUB:
3083 case Op_LoadUS:
3084 case Op_LoadI:
3085 case Op_LoadKlass:
3086 case Op_LoadNKlass:
3087 case Op_LoadL:
3088 case Op_LoadL_unaligned:
3089 case Op_LoadPLocked:
3090 case Op_LoadP:
3091 case Op_LoadN:
3092 case Op_LoadRange:
3093 case Op_LoadS:
3094 break;
3095
3096 case Op_AddP: { // Assert sane base pointers
3097 Node *addp = n->in(AddPNode::Address);
3098 assert( !addp->is_AddP() ||
3099 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3100 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3101 "Base pointers must match (addp %u)", addp->_idx );
3102 #ifdef _LP64
3103 if ((UseCompressedOops || UseCompressedClassPointers) &&
3104 addp->Opcode() == Op_ConP &&
3105 addp == n->in(AddPNode::Base) &&
3106 n->in(AddPNode::Offset)->is_Con()) {
3107 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3108 // on the platform and on the compressed oops mode.
3109 // Use addressing with narrow klass to load with offset on x86.
3110 // Some platforms can use the constant pool to load ConP.
3111 // Do this transformation here since IGVN will convert ConN back to ConP.
3112 const Type* t = addp->bottom_type();
3113 bool is_oop = t->isa_oopptr() != nullptr;
3114 bool is_klass = t->isa_klassptr() != nullptr;
3115
3116 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3117 (is_klass && Matcher::const_klass_prefer_decode())) {
3118 Node* nn = nullptr;
3119
3120 int op = is_oop ? Op_ConN : Op_ConNKlass;
3121
3122 // Look for existing ConN node of the same exact type.
3123 Node* r = root();
3124 uint cnt = r->outcnt();
3125 for (uint i = 0; i < cnt; i++) {
3126 Node* m = r->raw_out(i);
3127 if (m!= nullptr && m->Opcode() == op &&
3128 m->bottom_type()->make_ptr() == t) {
3129 nn = m;
3130 break;
3131 }
3132 }
3133 if (nn != nullptr) {
3134 // Decode a narrow oop to match address
3135 // [R12 + narrow_oop_reg<<3 + offset]
3136 if (is_oop) {
3137 nn = new DecodeNNode(nn, t);
3138 } else {
3139 nn = new DecodeNKlassNode(nn, t);
3140 }
3141 // Check for succeeding AddP which uses the same Base.
3142 // Otherwise we will run into the assertion above when visiting that guy.
3143 for (uint i = 0; i < n->outcnt(); ++i) {
3144 Node *out_i = n->raw_out(i);
3145 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3146 out_i->set_req(AddPNode::Base, nn);
3147 #ifdef ASSERT
3148 for (uint j = 0; j < out_i->outcnt(); ++j) {
3149 Node *out_j = out_i->raw_out(j);
3150 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3151 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3152 }
3153 #endif
3154 }
3155 }
3156 n->set_req(AddPNode::Base, nn);
3157 n->set_req(AddPNode::Address, nn);
3158 if (addp->outcnt() == 0) {
3159 addp->disconnect_inputs(this);
3160 }
3161 }
3162 }
3163 }
3164 #endif
3165 break;
3166 }
3167
3168 case Op_CastPP: {
3169 // Remove CastPP nodes to gain more freedom during scheduling but
3170 // keep the dependency they encode as control or precedence edges
3171 // (if control is set already) on memory operations. Some CastPP
3172 // nodes don't have a control (don't carry a dependency): skip
3173 // those.
3174 if (n->in(0) != nullptr) {
3175 ResourceMark rm;
3176 Unique_Node_List wq;
3177 wq.push(n);
3178 for (uint next = 0; next < wq.size(); ++next) {
3179 Node *m = wq.at(next);
3180 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3181 Node* use = m->fast_out(i);
3182 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3183 use->ensure_control_or_add_prec(n->in(0));
3184 } else {
3185 switch(use->Opcode()) {
3186 case Op_AddP:
3187 case Op_DecodeN:
3188 case Op_DecodeNKlass:
3189 case Op_CheckCastPP:
3190 case Op_CastPP:
3191 wq.push(use);
3192 break;
3193 }
3194 }
3195 }
3196 }
3197 }
3198 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3199 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3200 Node* in1 = n->in(1);
3201 const Type* t = n->bottom_type();
3202 Node* new_in1 = in1->clone();
3203 new_in1->as_DecodeN()->set_type(t);
3204
3205 if (!Matcher::narrow_oop_use_complex_address()) {
3206 //
3207 // x86, ARM and friends can handle 2 adds in addressing mode
3208 // and Matcher can fold a DecodeN node into address by using
3209 // a narrow oop directly and do implicit null check in address:
3210 //
3211 // [R12 + narrow_oop_reg<<3 + offset]
3212 // NullCheck narrow_oop_reg
3213 //
3214 // On other platforms (Sparc) we have to keep new DecodeN node and
3215 // use it to do implicit null check in address:
3216 //
3217 // decode_not_null narrow_oop_reg, base_reg
3218 // [base_reg + offset]
3219 // NullCheck base_reg
3220 //
3221 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3222 // to keep the information to which null check the new DecodeN node
3223 // corresponds to use it as value in implicit_null_check().
3224 //
3225 new_in1->set_req(0, n->in(0));
3226 }
3227
3228 n->subsume_by(new_in1, this);
3229 if (in1->outcnt() == 0) {
3230 in1->disconnect_inputs(this);
3231 }
3232 } else {
3233 n->subsume_by(n->in(1), this);
3234 if (n->outcnt() == 0) {
3235 n->disconnect_inputs(this);
3236 }
3237 }
3238 break;
3239 }
3240 #ifdef _LP64
3241 case Op_CmpP:
3242 // Do this transformation here to preserve CmpPNode::sub() and
3243 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3244 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3245 Node* in1 = n->in(1);
3246 Node* in2 = n->in(2);
3247 if (!in1->is_DecodeNarrowPtr()) {
3248 in2 = in1;
3249 in1 = n->in(2);
3250 }
3251 assert(in1->is_DecodeNarrowPtr(), "sanity");
3252
3253 Node* new_in2 = nullptr;
3254 if (in2->is_DecodeNarrowPtr()) {
3255 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3256 new_in2 = in2->in(1);
3257 } else if (in2->Opcode() == Op_ConP) {
3258 const Type* t = in2->bottom_type();
3259 if (t == TypePtr::NULL_PTR) {
3260 assert(in1->is_DecodeN(), "compare klass to null?");
3261 // Don't convert CmpP null check into CmpN if compressed
3262 // oops implicit null check is not generated.
3263 // This will allow to generate normal oop implicit null check.
3264 if (Matcher::gen_narrow_oop_implicit_null_checks())
3265 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3266 //
3267 // This transformation together with CastPP transformation above
3268 // will generated code for implicit null checks for compressed oops.
3269 //
3270 // The original code after Optimize()
3271 //
3272 // LoadN memory, narrow_oop_reg
3273 // decode narrow_oop_reg, base_reg
3274 // CmpP base_reg, nullptr
3275 // CastPP base_reg // NotNull
3276 // Load [base_reg + offset], val_reg
3277 //
3278 // after these transformations will be
3279 //
3280 // LoadN memory, narrow_oop_reg
3281 // CmpN narrow_oop_reg, nullptr
3282 // decode_not_null narrow_oop_reg, base_reg
3283 // Load [base_reg + offset], val_reg
3284 //
3285 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3286 // since narrow oops can be used in debug info now (see the code in
3287 // final_graph_reshaping_walk()).
3288 //
3289 // At the end the code will be matched to
3290 // on x86:
3291 //
3292 // Load_narrow_oop memory, narrow_oop_reg
3293 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3294 // NullCheck narrow_oop_reg
3295 //
3296 // and on sparc:
3297 //
3298 // Load_narrow_oop memory, narrow_oop_reg
3299 // decode_not_null narrow_oop_reg, base_reg
3300 // Load [base_reg + offset], val_reg
3301 // NullCheck base_reg
3302 //
3303 } else if (t->isa_oopptr()) {
3304 new_in2 = ConNode::make(t->make_narrowoop());
3305 } else if (t->isa_klassptr()) {
3306 new_in2 = ConNode::make(t->make_narrowklass());
3307 }
3308 }
3309 if (new_in2 != nullptr) {
3310 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3311 n->subsume_by(cmpN, this);
3312 if (in1->outcnt() == 0) {
3313 in1->disconnect_inputs(this);
3314 }
3315 if (in2->outcnt() == 0) {
3316 in2->disconnect_inputs(this);
3317 }
3318 }
3319 }
3320 break;
3321
3322 case Op_DecodeN:
3323 case Op_DecodeNKlass:
3324 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3325 // DecodeN could be pinned when it can't be fold into
3326 // an address expression, see the code for Op_CastPP above.
3327 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3328 break;
3329
3330 case Op_EncodeP:
3331 case Op_EncodePKlass: {
3332 Node* in1 = n->in(1);
3333 if (in1->is_DecodeNarrowPtr()) {
3334 n->subsume_by(in1->in(1), this);
3335 } else if (in1->Opcode() == Op_ConP) {
3336 const Type* t = in1->bottom_type();
3337 if (t == TypePtr::NULL_PTR) {
3338 assert(t->isa_oopptr(), "null klass?");
3339 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3340 } else if (t->isa_oopptr()) {
3341 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3342 } else if (t->isa_klassptr()) {
3343 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3344 }
3345 }
3346 if (in1->outcnt() == 0) {
3347 in1->disconnect_inputs(this);
3348 }
3349 break;
3350 }
3351
3352 case Op_Proj: {
3353 if (OptimizeStringConcat || IncrementalInline) {
3354 ProjNode* proj = n->as_Proj();
3355 if (proj->_is_io_use) {
3356 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3357 // Separate projections were used for the exception path which
3358 // are normally removed by a late inline. If it wasn't inlined
3359 // then they will hang around and should just be replaced with
3360 // the original one. Merge them.
3361 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3362 if (non_io_proj != nullptr) {
3363 proj->subsume_by(non_io_proj , this);
3364 }
3365 }
3366 }
3367 break;
3368 }
3369
3370 case Op_Phi:
3371 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3372 // The EncodeP optimization may create Phi with the same edges
3373 // for all paths. It is not handled well by Register Allocator.
3374 Node* unique_in = n->in(1);
3375 assert(unique_in != nullptr, "");
3376 uint cnt = n->req();
3377 for (uint i = 2; i < cnt; i++) {
3378 Node* m = n->in(i);
3379 assert(m != nullptr, "");
3380 if (unique_in != m)
3381 unique_in = nullptr;
3382 }
3383 if (unique_in != nullptr) {
3384 n->subsume_by(unique_in, this);
3385 }
3386 }
3387 break;
3388
3389 #endif
3390
3391 #ifdef ASSERT
3392 case Op_CastII:
3393 // Verify that all range check dependent CastII nodes were removed.
3394 if (n->isa_CastII()->has_range_check()) {
3395 n->dump(3);
3396 assert(false, "Range check dependent CastII node was not removed");
3397 }
3398 break;
3399 #endif
3400
3401 case Op_ModI:
3402 if (UseDivMod) {
3403 // Check if a%b and a/b both exist
3404 Node* d = n->find_similar(Op_DivI);
3405 if (d) {
3406 // Replace them with a fused divmod if supported
3407 if (Matcher::has_match_rule(Op_DivModI)) {
3408 DivModINode* divmod = DivModINode::make(n);
3409 d->subsume_by(divmod->div_proj(), this);
3410 n->subsume_by(divmod->mod_proj(), this);
3411 } else {
3412 // replace a%b with a-((a/b)*b)
3413 Node* mult = new MulINode(d, d->in(2));
3414 Node* sub = new SubINode(d->in(1), mult);
3415 n->subsume_by(sub, this);
3416 }
3417 }
3418 }
3419 break;
3420
3421 case Op_ModL:
3422 if (UseDivMod) {
3423 // Check if a%b and a/b both exist
3424 Node* d = n->find_similar(Op_DivL);
3425 if (d) {
3426 // Replace them with a fused divmod if supported
3427 if (Matcher::has_match_rule(Op_DivModL)) {
3428 DivModLNode* divmod = DivModLNode::make(n);
3429 d->subsume_by(divmod->div_proj(), this);
3430 n->subsume_by(divmod->mod_proj(), this);
3431 } else {
3432 // replace a%b with a-((a/b)*b)
3433 Node* mult = new MulLNode(d, d->in(2));
3434 Node* sub = new SubLNode(d->in(1), mult);
3435 n->subsume_by(sub, this);
3436 }
3437 }
3438 }
3439 break;
3440
3441 case Op_LoadVector:
3442 case Op_StoreVector:
3443 case Op_LoadVectorGather:
3444 case Op_StoreVectorScatter:
3445 case Op_VectorCmpMasked:
3446 case Op_VectorMaskGen:
3447 case Op_LoadVectorMasked:
3448 case Op_StoreVectorMasked:
3449 break;
3450
3451 case Op_AddReductionVI:
3452 case Op_AddReductionVL:
3453 case Op_AddReductionVF:
3454 case Op_AddReductionVD:
3455 case Op_MulReductionVI:
3456 case Op_MulReductionVL:
3457 case Op_MulReductionVF:
3458 case Op_MulReductionVD:
3459 case Op_MinReductionV:
3460 case Op_MaxReductionV:
3461 case Op_AndReductionV:
3462 case Op_OrReductionV:
3463 case Op_XorReductionV:
3464 break;
3465
3466 case Op_PackB:
3467 case Op_PackS:
3468 case Op_PackI:
3469 case Op_PackF:
3470 case Op_PackL:
3471 case Op_PackD:
3472 if (n->req()-1 > 2) {
3473 // Replace many operand PackNodes with a binary tree for matching
3474 PackNode* p = (PackNode*) n;
3475 Node* btp = p->binary_tree_pack(1, n->req());
3476 n->subsume_by(btp, this);
3477 }
3478 break;
3479 case Op_Loop:
3480 assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3481 case Op_CountedLoop:
3482 case Op_LongCountedLoop:
3483 case Op_OuterStripMinedLoop:
3484 if (n->as_Loop()->is_inner_loop()) {
3485 frc.inc_inner_loop_count();
3486 }
3487 n->as_Loop()->verify_strip_mined(0);
3488 break;
3489 case Op_LShiftI:
3490 case Op_RShiftI:
3491 case Op_URShiftI:
3492 case Op_LShiftL:
3493 case Op_RShiftL:
3494 case Op_URShiftL:
3495 if (Matcher::need_masked_shift_count) {
3496 // The cpu's shift instructions don't restrict the count to the
3497 // lower 5/6 bits. We need to do the masking ourselves.
3498 Node* in2 = n->in(2);
3499 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3500 const TypeInt* t = in2->find_int_type();
3501 if (t != nullptr && t->is_con()) {
3502 juint shift = t->get_con();
3503 if (shift > mask) { // Unsigned cmp
3504 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3505 }
3506 } else {
3507 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3508 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3509 n->set_req(2, shift);
3510 }
3511 }
3512 if (in2->outcnt() == 0) { // Remove dead node
3513 in2->disconnect_inputs(this);
3514 }
3515 }
3516 break;
3517 case Op_MemBarStoreStore:
3518 case Op_MemBarRelease:
3519 // Break the link with AllocateNode: it is no longer useful and
3520 // confuses register allocation.
3521 if (n->req() > MemBarNode::Precedent) {
3522 n->set_req(MemBarNode::Precedent, top());
3523 }
3524 break;
3525 case Op_MemBarAcquire: {
3526 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3527 // At parse time, the trailing MemBarAcquire for a volatile load
3528 // is created with an edge to the load. After optimizations,
3529 // that input may be a chain of Phis. If those phis have no
3530 // other use, then the MemBarAcquire keeps them alive and
3531 // register allocation can be confused.
3532 dead_nodes.push(n->in(MemBarNode::Precedent));
3533 n->set_req(MemBarNode::Precedent, top());
3534 }
3535 break;
3536 }
3537 case Op_Blackhole:
3538 break;
3539 case Op_RangeCheck: {
3540 RangeCheckNode* rc = n->as_RangeCheck();
3541 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3542 n->subsume_by(iff, this);
3543 frc._tests.push(iff);
3544 break;
3545 }
3546 case Op_ConvI2L: {
3547 if (!Matcher::convi2l_type_required) {
3548 // Code generation on some platforms doesn't need accurate
3549 // ConvI2L types. Widening the type can help remove redundant
3550 // address computations.
3551 n->as_Type()->set_type(TypeLong::INT);
3552 ResourceMark rm;
3553 Unique_Node_List wq;
3554 wq.push(n);
3555 for (uint next = 0; next < wq.size(); next++) {
3556 Node *m = wq.at(next);
3557
3558 for(;;) {
3559 // Loop over all nodes with identical inputs edges as m
3560 Node* k = m->find_similar(m->Opcode());
3561 if (k == nullptr) {
3562 break;
3563 }
3564 // Push their uses so we get a chance to remove node made
3565 // redundant
3566 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3567 Node* u = k->fast_out(i);
3568 if (u->Opcode() == Op_LShiftL ||
3569 u->Opcode() == Op_AddL ||
3570 u->Opcode() == Op_SubL ||
3571 u->Opcode() == Op_AddP) {
3572 wq.push(u);
3573 }
3574 }
3575 // Replace all nodes with identical edges as m with m
3576 k->subsume_by(m, this);
3577 }
3578 }
3579 }
3580 break;
3581 }
3582 case Op_CmpUL: {
3583 if (!Matcher::has_match_rule(Op_CmpUL)) {
3584 // No support for unsigned long comparisons
3585 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3586 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3587 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3588 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3589 Node* andl = new AndLNode(orl, remove_sign_mask);
3590 Node* cmp = new CmpLNode(andl, n->in(2));
3591 n->subsume_by(cmp, this);
3592 }
3593 break;
3594 }
3595 default:
3596 assert(!n->is_Call(), "");
3597 assert(!n->is_Mem(), "");
3598 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3599 break;
3600 }
3601 }
3602
3603 //------------------------------final_graph_reshaping_walk---------------------
3604 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3605 // requires that the walk visits a node's inputs before visiting the node.
3606 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3607 Unique_Node_List sfpt;
3608
3609 frc._visited.set(root->_idx); // first, mark node as visited
3610 uint cnt = root->req();
3611 Node *n = root;
3612 uint i = 0;
3613 while (true) {
3614 if (i < cnt) {
3615 // Place all non-visited non-null inputs onto stack
3616 Node* m = n->in(i);
3617 ++i;
3618 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3619 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3620 // compute worst case interpreter size in case of a deoptimization
3621 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3622
3623 sfpt.push(m);
3624 }
3625 cnt = m->req();
3626 nstack.push(n, i); // put on stack parent and next input's index
3627 n = m;
3628 i = 0;
3629 }
3630 } else {
3631 // Now do post-visit work
3632 final_graph_reshaping_impl(n, frc, dead_nodes);
3633 if (nstack.is_empty())
3634 break; // finished
3635 n = nstack.node(); // Get node from stack
3636 cnt = n->req();
3637 i = nstack.index();
3638 nstack.pop(); // Shift to the next node on stack
3639 }
3640 }
3641
3642 // Skip next transformation if compressed oops are not used.
3643 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3644 (!UseCompressedOops && !UseCompressedClassPointers))
3645 return;
3646
3647 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3648 // It could be done for an uncommon traps or any safepoints/calls
3649 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3650 while (sfpt.size() > 0) {
3651 n = sfpt.pop();
3652 JVMState *jvms = n->as_SafePoint()->jvms();
3653 assert(jvms != nullptr, "sanity");
3654 int start = jvms->debug_start();
3655 int end = n->req();
3656 bool is_uncommon = (n->is_CallStaticJava() &&
3657 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3658 for (int j = start; j < end; j++) {
3659 Node* in = n->in(j);
3660 if (in->is_DecodeNarrowPtr()) {
3661 bool safe_to_skip = true;
3662 if (!is_uncommon ) {
3663 // Is it safe to skip?
3664 for (uint i = 0; i < in->outcnt(); i++) {
3665 Node* u = in->raw_out(i);
3666 if (!u->is_SafePoint() ||
3667 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3668 safe_to_skip = false;
3669 }
3670 }
3671 }
3672 if (safe_to_skip) {
3673 n->set_req(j, in->in(1));
3674 }
3675 if (in->outcnt() == 0) {
3676 in->disconnect_inputs(this);
3677 }
3678 }
3679 }
3680 }
3681 }
3682
3683 //------------------------------final_graph_reshaping--------------------------
3684 // Final Graph Reshaping.
3685 //
3686 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3687 // and not commoned up and forced early. Must come after regular
3688 // optimizations to avoid GVN undoing the cloning. Clone constant
3689 // inputs to Loop Phis; these will be split by the allocator anyways.
3690 // Remove Opaque nodes.
3691 // (2) Move last-uses by commutative operations to the left input to encourage
3692 // Intel update-in-place two-address operations and better register usage
3693 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3694 // calls canonicalizing them back.
3695 // (3) Count the number of double-precision FP ops, single-precision FP ops
3696 // and call sites. On Intel, we can get correct rounding either by
3697 // forcing singles to memory (requires extra stores and loads after each
3698 // FP bytecode) or we can set a rounding mode bit (requires setting and
3699 // clearing the mode bit around call sites). The mode bit is only used
3700 // if the relative frequency of single FP ops to calls is low enough.
3701 // This is a key transform for SPEC mpeg_audio.
3702 // (4) Detect infinite loops; blobs of code reachable from above but not
3703 // below. Several of the Code_Gen algorithms fail on such code shapes,
3704 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3705 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3706 // Detection is by looking for IfNodes where only 1 projection is
3707 // reachable from below or CatchNodes missing some targets.
3708 // (5) Assert for insane oop offsets in debug mode.
3709
3710 bool Compile::final_graph_reshaping() {
3711 // an infinite loop may have been eliminated by the optimizer,
3712 // in which case the graph will be empty.
3713 if (root()->req() == 1) {
3714 record_method_not_compilable("trivial infinite loop");
3715 return true;
3716 }
3717
3718 // Expensive nodes have their control input set to prevent the GVN
3719 // from freely commoning them. There's no GVN beyond this point so
3720 // no need to keep the control input. We want the expensive nodes to
3721 // be freely moved to the least frequent code path by gcm.
3722 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3723 for (int i = 0; i < expensive_count(); i++) {
3724 _expensive_nodes.at(i)->set_req(0, nullptr);
3725 }
3726
3727 Final_Reshape_Counts frc;
3728
3729 // Visit everybody reachable!
3730 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3731 Node_Stack nstack(live_nodes() >> 1);
3732 Unique_Node_List dead_nodes;
3733 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3734
3735 // Check for unreachable (from below) code (i.e., infinite loops).
3736 for( uint i = 0; i < frc._tests.size(); i++ ) {
3737 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3738 // Get number of CFG targets.
3739 // Note that PCTables include exception targets after calls.
3740 uint required_outcnt = n->required_outcnt();
3741 if (n->outcnt() != required_outcnt) {
3742 // Check for a few special cases. Rethrow Nodes never take the
3743 // 'fall-thru' path, so expected kids is 1 less.
3744 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3745 if (n->in(0)->in(0)->is_Call()) {
3746 CallNode* call = n->in(0)->in(0)->as_Call();
3747 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3748 required_outcnt--; // Rethrow always has 1 less kid
3749 } else if (call->req() > TypeFunc::Parms &&
3750 call->is_CallDynamicJava()) {
3751 // Check for null receiver. In such case, the optimizer has
3752 // detected that the virtual call will always result in a null
3753 // pointer exception. The fall-through projection of this CatchNode
3754 // will not be populated.
3755 Node* arg0 = call->in(TypeFunc::Parms);
3756 if (arg0->is_Type() &&
3757 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3758 required_outcnt--;
3759 }
3760 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3761 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3762 // Check for illegal array length. In such case, the optimizer has
3763 // detected that the allocation attempt will always result in an
3764 // exception. There is no fall-through projection of this CatchNode .
3765 assert(call->is_CallStaticJava(), "static call expected");
3766 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3767 uint valid_length_test_input = call->req() - 1;
3768 Node* valid_length_test = call->in(valid_length_test_input);
3769 call->del_req(valid_length_test_input);
3770 if (valid_length_test->find_int_con(1) == 0) {
3771 required_outcnt--;
3772 }
3773 dead_nodes.push(valid_length_test);
3774 assert(n->outcnt() == required_outcnt, "malformed control flow");
3775 continue;
3776 }
3777 }
3778 }
3779 // Recheck with a better notion of 'required_outcnt'
3780 if (n->outcnt() != required_outcnt) {
3781 record_method_not_compilable("malformed control flow");
3782 return true; // Not all targets reachable!
3783 }
3784 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
3785 CallNode* call = n->in(0)->in(0)->as_Call();
3786 if (call->entry_point() == OptoRuntime::new_array_Java() ||
3787 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3788 assert(call->is_CallStaticJava(), "static call expected");
3789 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3790 uint valid_length_test_input = call->req() - 1;
3791 dead_nodes.push(call->in(valid_length_test_input));
3792 call->del_req(valid_length_test_input); // valid length test useless now
3793 }
3794 }
3795 // Check that I actually visited all kids. Unreached kids
3796 // must be infinite loops.
3797 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3798 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3799 record_method_not_compilable("infinite loop");
3800 return true; // Found unvisited kid; must be unreach
3801 }
3802
3803 // Here so verification code in final_graph_reshaping_walk()
3804 // always see an OuterStripMinedLoopEnd
3805 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
3806 IfNode* init_iff = n->as_If();
3807 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3808 n->subsume_by(iff, this);
3809 }
3810 }
3811
3812 while (dead_nodes.size() > 0) {
3813 Node* m = dead_nodes.pop();
3814 if (m->outcnt() == 0 && m != top()) {
3815 for (uint j = 0; j < m->req(); j++) {
3816 Node* in = m->in(j);
3817 if (in != nullptr) {
3818 dead_nodes.push(in);
3819 }
3820 }
3821 m->disconnect_inputs(this);
3822 }
3823 }
3824
3825 #ifdef IA32
3826 // If original bytecodes contained a mixture of floats and doubles
3827 // check if the optimizer has made it homogenous, item (3).
3828 if (UseSSE == 0 &&
3829 frc.get_float_count() > 32 &&
3830 frc.get_double_count() == 0 &&
3831 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3832 set_24_bit_selection_and_mode(false, true);
3833 }
3834 #endif // IA32
3835
3836 set_java_calls(frc.get_java_call_count());
3837 set_inner_loops(frc.get_inner_loop_count());
3838
3839 // No infinite loops, no reason to bail out.
3840 return false;
3841 }
3842
3843 //-----------------------------too_many_traps----------------------------------
3844 // Report if there are too many traps at the current method and bci.
3845 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3846 bool Compile::too_many_traps(ciMethod* method,
3847 int bci,
3848 Deoptimization::DeoptReason reason) {
3849 ciMethodData* md = method->method_data();
3850 if (md->is_empty()) {
3851 // Assume the trap has not occurred, or that it occurred only
3852 // because of a transient condition during start-up in the interpreter.
3853 return false;
3854 }
3855 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
3856 if (md->has_trap_at(bci, m, reason) != 0) {
3857 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3858 // Also, if there are multiple reasons, or if there is no per-BCI record,
3859 // assume the worst.
3860 if (log())
3861 log()->elem("observe trap='%s' count='%d'",
3862 Deoptimization::trap_reason_name(reason),
3863 md->trap_count(reason));
3864 return true;
3865 } else {
3866 // Ignore method/bci and see if there have been too many globally.
3867 return too_many_traps(reason, md);
3868 }
3869 }
3870
3871 // Less-accurate variant which does not require a method and bci.
3872 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3873 ciMethodData* logmd) {
3874 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3875 // Too many traps globally.
3876 // Note that we use cumulative trap_count, not just md->trap_count.
3877 if (log()) {
3878 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
3879 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3880 Deoptimization::trap_reason_name(reason),
3881 mcount, trap_count(reason));
3882 }
3883 return true;
3884 } else {
3885 // The coast is clear.
3886 return false;
3887 }
3888 }
3889
3890 //--------------------------too_many_recompiles--------------------------------
3891 // Report if there are too many recompiles at the current method and bci.
3892 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3893 // Is not eager to return true, since this will cause the compiler to use
3894 // Action_none for a trap point, to avoid too many recompilations.
3895 bool Compile::too_many_recompiles(ciMethod* method,
3896 int bci,
3897 Deoptimization::DeoptReason reason) {
3898 ciMethodData* md = method->method_data();
3899 if (md->is_empty()) {
3900 // Assume the trap has not occurred, or that it occurred only
3901 // because of a transient condition during start-up in the interpreter.
3902 return false;
3903 }
3904 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3905 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3906 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3907 Deoptimization::DeoptReason per_bc_reason
3908 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3909 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
3910 if ((per_bc_reason == Deoptimization::Reason_none
3911 || md->has_trap_at(bci, m, reason) != 0)
3912 // The trap frequency measure we care about is the recompile count:
3913 && md->trap_recompiled_at(bci, m)
3914 && md->overflow_recompile_count() >= bc_cutoff) {
3915 // Do not emit a trap here if it has already caused recompilations.
3916 // Also, if there are multiple reasons, or if there is no per-BCI record,
3917 // assume the worst.
3918 if (log())
3919 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3920 Deoptimization::trap_reason_name(reason),
3921 md->trap_count(reason),
3922 md->overflow_recompile_count());
3923 return true;
3924 } else if (trap_count(reason) != 0
3925 && decompile_count() >= m_cutoff) {
3926 // Too many recompiles globally, and we have seen this sort of trap.
3927 // Use cumulative decompile_count, not just md->decompile_count.
3928 if (log())
3929 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3930 Deoptimization::trap_reason_name(reason),
3931 md->trap_count(reason), trap_count(reason),
3932 md->decompile_count(), decompile_count());
3933 return true;
3934 } else {
3935 // The coast is clear.
3936 return false;
3937 }
3938 }
3939
3940 // Compute when not to trap. Used by matching trap based nodes and
3941 // NullCheck optimization.
3942 void Compile::set_allowed_deopt_reasons() {
3943 _allowed_reasons = 0;
3944 if (is_method_compilation()) {
3945 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3946 assert(rs < BitsPerInt, "recode bit map");
3947 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3948 _allowed_reasons |= nth_bit(rs);
3949 }
3950 }
3951 }
3952 }
3953
3954 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3955 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3956 }
3957
3958 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3959 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3960 }
3961
3962 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3963 if (holder->is_initialized()) {
3964 return false;
3965 }
3966 if (holder->is_being_initialized()) {
3967 if (accessing_method->holder() == holder) {
3968 // Access inside a class. The barrier can be elided when access happens in <clinit>,
3969 // <init>, or a static method. In all those cases, there was an initialization
3970 // barrier on the holder klass passed.
3971 if (accessing_method->is_static_initializer() ||
3972 accessing_method->is_object_initializer() ||
3973 accessing_method->is_static()) {
3974 return false;
3975 }
3976 } else if (accessing_method->holder()->is_subclass_of(holder)) {
3977 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3978 // In case of <init> or a static method, the barrier is on the subclass is not enough:
3979 // child class can become fully initialized while its parent class is still being initialized.
3980 if (accessing_method->is_static_initializer()) {
3981 return false;
3982 }
3983 }
3984 ciMethod* root = method(); // the root method of compilation
3985 if (root != accessing_method) {
3986 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3987 }
3988 }
3989 return true;
3990 }
3991
3992 #ifndef PRODUCT
3993 //------------------------------verify_graph_edges---------------------------
3994 // Walk the Graph and verify that there is a one-to-one correspondence
3995 // between Use-Def edges and Def-Use edges in the graph.
3996 void Compile::verify_graph_edges(bool no_dead_code) {
3997 if (VerifyGraphEdges) {
3998 Unique_Node_List visited;
3999 // Call recursive graph walk to check edges
4000 _root->verify_edges(visited);
4001 if (no_dead_code) {
4002 // Now make sure that no visited node is used by an unvisited node.
4003 bool dead_nodes = false;
4004 Unique_Node_List checked;
4005 while (visited.size() > 0) {
4006 Node* n = visited.pop();
4007 checked.push(n);
4008 for (uint i = 0; i < n->outcnt(); i++) {
4009 Node* use = n->raw_out(i);
4010 if (checked.member(use)) continue; // already checked
4011 if (visited.member(use)) continue; // already in the graph
4012 if (use->is_Con()) continue; // a dead ConNode is OK
4013 // At this point, we have found a dead node which is DU-reachable.
4014 if (!dead_nodes) {
4015 tty->print_cr("*** Dead nodes reachable via DU edges:");
4016 dead_nodes = true;
4017 }
4018 use->dump(2);
4019 tty->print_cr("---");
4020 checked.push(use); // No repeats; pretend it is now checked.
4021 }
4022 }
4023 assert(!dead_nodes, "using nodes must be reachable from root");
4024 }
4025 }
4026 }
4027 #endif
4028
4029 // The Compile object keeps track of failure reasons separately from the ciEnv.
4030 // This is required because there is not quite a 1-1 relation between the
4031 // ciEnv and its compilation task and the Compile object. Note that one
4032 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4033 // to backtrack and retry without subsuming loads. Other than this backtracking
4034 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4035 // by the logic in C2Compiler.
4036 void Compile::record_failure(const char* reason) {
4037 if (log() != nullptr) {
4038 log()->elem("failure reason='%s' phase='compile'", reason);
4039 }
4040 if (_failure_reason == nullptr) {
4041 // Record the first failure reason.
4042 _failure_reason = reason;
4043 }
4044
4045 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4046 C->print_method(PHASE_FAILURE);
4047 }
4048 _root = nullptr; // flush the graph, too
4049 }
4050
4051 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4052 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4053 _phase_name(name), _dolog(CITimeVerbose)
4054 {
4055 if (_dolog) {
4056 C = Compile::current();
4057 _log = C->log();
4058 } else {
4059 C = nullptr;
4060 _log = nullptr;
4061 }
4062 if (_log != nullptr) {
4063 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4064 _log->stamp();
4065 _log->end_head();
4066 }
4067 }
4068
4069 Compile::TracePhase::~TracePhase() {
4070
4071 C = Compile::current();
4072 if (_dolog) {
4073 _log = C->log();
4074 } else {
4075 _log = nullptr;
4076 }
4077
4078 #ifdef ASSERT
4079 if (PrintIdealNodeCount) {
4080 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4081 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4082 }
4083
4084 if (VerifyIdealNodeCount) {
4085 Compile::current()->print_missing_nodes();
4086 }
4087 #endif
4088
4089 if (_log != nullptr) {
4090 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4091 }
4092 }
4093
4094 //----------------------------static_subtype_check-----------------------------
4095 // Shortcut important common cases when superklass is exact:
4096 // (0) superklass is java.lang.Object (can occur in reflective code)
4097 // (1) subklass is already limited to a subtype of superklass => always ok
4098 // (2) subklass does not overlap with superklass => always fail
4099 // (3) superklass has NO subtypes and we can check with a simple compare.
4100 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4101 if (StressReflectiveCode) {
4102 return SSC_full_test; // Let caller generate the general case.
4103 }
4104
4105 if (superk == env()->Object_klass()) {
4106 return SSC_always_true; // (0) this test cannot fail
4107 }
4108
4109 ciType* superelem = superk;
4110 ciType* subelem = subk;
4111 if (superelem->is_array_klass()) {
4112 superelem = superelem->as_array_klass()->base_element_type();
4113 }
4114 if (subelem->is_array_klass()) {
4115 subelem = subelem->as_array_klass()->base_element_type();
4116 }
4117
4118 if (!subk->is_interface()) { // cannot trust static interface types yet
4119 if (subk->is_subtype_of(superk)) {
4120 return SSC_always_true; // (1) false path dead; no dynamic test needed
4121 }
4122 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4123 !(subelem->is_klass() && subelem->as_klass()->is_interface()) &&
4124 !superk->is_subtype_of(subk)) {
4125 return SSC_always_false; // (2) true path dead; no dynamic test needed
4126 }
4127 }
4128
4129 // If casting to an instance klass, it must have no subtypes
4130 if (superk->is_interface()) {
4131 // Cannot trust interfaces yet.
4132 // %%% S.B. superk->nof_implementors() == 1
4133 } else if (superelem->is_instance_klass()) {
4134 ciInstanceKlass* ik = superelem->as_instance_klass();
4135 if (!ik->has_subklass() && !ik->is_interface()) {
4136 if (!ik->is_final()) {
4137 // Add a dependency if there is a chance of a later subclass.
4138 dependencies()->assert_leaf_type(ik);
4139 }
4140 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4141 }
4142 } else {
4143 // A primitive array type has no subtypes.
4144 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4145 }
4146
4147 return SSC_full_test;
4148 }
4149
4150 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4151 #ifdef _LP64
4152 // The scaled index operand to AddP must be a clean 64-bit value.
4153 // Java allows a 32-bit int to be incremented to a negative
4154 // value, which appears in a 64-bit register as a large
4155 // positive number. Using that large positive number as an
4156 // operand in pointer arithmetic has bad consequences.
4157 // On the other hand, 32-bit overflow is rare, and the possibility
4158 // can often be excluded, if we annotate the ConvI2L node with
4159 // a type assertion that its value is known to be a small positive
4160 // number. (The prior range check has ensured this.)
4161 // This assertion is used by ConvI2LNode::Ideal.
4162 int index_max = max_jint - 1; // array size is max_jint, index is one less
4163 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4164 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4165 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4166 #endif
4167 return idx;
4168 }
4169
4170 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4171 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4172 if (ctrl != nullptr) {
4173 // Express control dependency by a CastII node with a narrow type.
4174 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4175 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4176 // node from floating above the range check during loop optimizations. Otherwise, the
4177 // ConvI2L node may be eliminated independently of the range check, causing the data path
4178 // to become TOP while the control path is still there (although it's unreachable).
4179 value->set_req(0, ctrl);
4180 value = phase->transform(value);
4181 }
4182 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4183 return phase->transform(new ConvI2LNode(value, ltype));
4184 }
4185
4186 void Compile::print_inlining_stream_free() {
4187 if (_print_inlining_stream != NULL) {
4188 _print_inlining_stream->~stringStream();
4189 _print_inlining_stream = NULL;
4190 }
4191 }
4192
4193 // The message about the current inlining is accumulated in
4194 // _print_inlining_stream and transfered into the _print_inlining_list
4195 // once we know whether inlining succeeds or not. For regular
4196 // inlining, messages are appended to the buffer pointed by
4197 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4198 // a new buffer is added after _print_inlining_idx in the list. This
4199 // way we can update the inlining message for late inlining call site
4200 // when the inlining is attempted again.
4201 void Compile::print_inlining_init() {
4202 if (print_inlining() || print_intrinsics()) {
4203 // print_inlining_init is actually called several times.
4204 print_inlining_stream_free();
4205 _print_inlining_stream = new stringStream();
4206 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4207 }
4208 }
4209
4210 void Compile::print_inlining_reinit() {
4211 if (print_inlining() || print_intrinsics()) {
4212 print_inlining_stream_free();
4213 // Re allocate buffer when we change ResourceMark
4214 _print_inlining_stream = new stringStream();
4215 }
4216 }
4217
4218 void Compile::print_inlining_reset() {
4219 _print_inlining_stream->reset();
4220 }
4221
4222 void Compile::print_inlining_commit() {
4223 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4224 // Transfer the message from _print_inlining_stream to the current
4225 // _print_inlining_list buffer and clear _print_inlining_stream.
4226 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4227 print_inlining_reset();
4228 }
4229
4230 void Compile::print_inlining_push() {
4231 // Add new buffer to the _print_inlining_list at current position
4232 _print_inlining_idx++;
4233 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4234 }
4235
4236 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4237 return _print_inlining_list->at(_print_inlining_idx);
4238 }
4239
4240 void Compile::print_inlining_update(CallGenerator* cg) {
4241 if (print_inlining() || print_intrinsics()) {
4242 if (cg->is_late_inline()) {
4243 if (print_inlining_current()->cg() != cg &&
4244 (print_inlining_current()->cg() != nullptr ||
4245 print_inlining_current()->ss()->size() != 0)) {
4246 print_inlining_push();
4247 }
4248 print_inlining_commit();
4249 print_inlining_current()->set_cg(cg);
4250 } else {
4251 if (print_inlining_current()->cg() != nullptr) {
4252 print_inlining_push();
4253 }
4254 print_inlining_commit();
4255 }
4256 }
4257 }
4258
4259 void Compile::print_inlining_move_to(CallGenerator* cg) {
4260 // We resume inlining at a late inlining call site. Locate the
4261 // corresponding inlining buffer so that we can update it.
4262 if (print_inlining() || print_intrinsics()) {
4263 for (int i = 0; i < _print_inlining_list->length(); i++) {
4264 if (_print_inlining_list->at(i)->cg() == cg) {
4265 _print_inlining_idx = i;
4266 return;
4267 }
4268 }
4269 ShouldNotReachHere();
4270 }
4271 }
4272
4273 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4274 if (print_inlining() || print_intrinsics()) {
4275 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4276 assert(print_inlining_current()->cg() == cg, "wrong entry");
4277 // replace message with new message
4278 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4279 print_inlining_commit();
4280 print_inlining_current()->set_cg(cg);
4281 }
4282 }
4283
4284 void Compile::print_inlining_assert_ready() {
4285 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4286 }
4287
4288 void Compile::process_print_inlining() {
4289 assert(_late_inlines.length() == 0, "not drained yet");
4290 if (print_inlining() || print_intrinsics()) {
4291 ResourceMark rm;
4292 stringStream ss;
4293 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4294 for (int i = 0; i < _print_inlining_list->length(); i++) {
4295 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4296 ss.print("%s", pib->ss()->as_string());
4297 delete pib;
4298 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4299 }
4300 // Reset _print_inlining_list, it only contains destructed objects.
4301 // It is on the arena, so it will be freed when the arena is reset.
4302 _print_inlining_list = nullptr;
4303 // _print_inlining_stream won't be used anymore, either.
4304 print_inlining_stream_free();
4305 size_t end = ss.size();
4306 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4307 strncpy(_print_inlining_output, ss.base(), end+1);
4308 _print_inlining_output[end] = 0;
4309 }
4310 }
4311
4312 void Compile::dump_print_inlining() {
4313 if (_print_inlining_output != nullptr) {
4314 tty->print_raw(_print_inlining_output);
4315 }
4316 }
4317
4318 void Compile::log_late_inline(CallGenerator* cg) {
4319 if (log() != nullptr) {
4320 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4321 cg->unique_id());
4322 JVMState* p = cg->call_node()->jvms();
4323 while (p != nullptr) {
4324 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4325 p = p->caller();
4326 }
4327 log()->tail("late_inline");
4328 }
4329 }
4330
4331 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4332 log_late_inline(cg);
4333 if (log() != nullptr) {
4334 log()->inline_fail(msg);
4335 }
4336 }
4337
4338 void Compile::log_inline_id(CallGenerator* cg) {
4339 if (log() != nullptr) {
4340 // The LogCompilation tool needs a unique way to identify late
4341 // inline call sites. This id must be unique for this call site in
4342 // this compilation. Try to have it unique across compilations as
4343 // well because it can be convenient when grepping through the log
4344 // file.
4345 // Distinguish OSR compilations from others in case CICountOSR is
4346 // on.
4347 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4348 cg->set_unique_id(id);
4349 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4350 }
4351 }
4352
4353 void Compile::log_inline_failure(const char* msg) {
4354 if (C->log() != nullptr) {
4355 C->log()->inline_fail(msg);
4356 }
4357 }
4358
4359
4360 // Dump inlining replay data to the stream.
4361 // Don't change thread state and acquire any locks.
4362 void Compile::dump_inline_data(outputStream* out) {
4363 InlineTree* inl_tree = ilt();
4364 if (inl_tree != nullptr) {
4365 out->print(" inline %d", inl_tree->count());
4366 inl_tree->dump_replay_data(out);
4367 }
4368 }
4369
4370 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4371 if (n1->Opcode() < n2->Opcode()) return -1;
4372 else if (n1->Opcode() > n2->Opcode()) return 1;
4373
4374 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4375 for (uint i = 1; i < n1->req(); i++) {
4376 if (n1->in(i) < n2->in(i)) return -1;
4377 else if (n1->in(i) > n2->in(i)) return 1;
4378 }
4379
4380 return 0;
4381 }
4382
4383 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4384 Node* n1 = *n1p;
4385 Node* n2 = *n2p;
4386
4387 return cmp_expensive_nodes(n1, n2);
4388 }
4389
4390 void Compile::sort_expensive_nodes() {
4391 if (!expensive_nodes_sorted()) {
4392 _expensive_nodes.sort(cmp_expensive_nodes);
4393 }
4394 }
4395
4396 bool Compile::expensive_nodes_sorted() const {
4397 for (int i = 1; i < _expensive_nodes.length(); i++) {
4398 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4399 return false;
4400 }
4401 }
4402 return true;
4403 }
4404
4405 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4406 if (_expensive_nodes.length() == 0) {
4407 return false;
4408 }
4409
4410 assert(OptimizeExpensiveOps, "optimization off?");
4411
4412 // Take this opportunity to remove dead nodes from the list
4413 int j = 0;
4414 for (int i = 0; i < _expensive_nodes.length(); i++) {
4415 Node* n = _expensive_nodes.at(i);
4416 if (!n->is_unreachable(igvn)) {
4417 assert(n->is_expensive(), "should be expensive");
4418 _expensive_nodes.at_put(j, n);
4419 j++;
4420 }
4421 }
4422 _expensive_nodes.trunc_to(j);
4423
4424 // Then sort the list so that similar nodes are next to each other
4425 // and check for at least two nodes of identical kind with same data
4426 // inputs.
4427 sort_expensive_nodes();
4428
4429 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4430 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4431 return true;
4432 }
4433 }
4434
4435 return false;
4436 }
4437
4438 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4439 if (_expensive_nodes.length() == 0) {
4440 return;
4441 }
4442
4443 assert(OptimizeExpensiveOps, "optimization off?");
4444
4445 // Sort to bring similar nodes next to each other and clear the
4446 // control input of nodes for which there's only a single copy.
4447 sort_expensive_nodes();
4448
4449 int j = 0;
4450 int identical = 0;
4451 int i = 0;
4452 bool modified = false;
4453 for (; i < _expensive_nodes.length()-1; i++) {
4454 assert(j <= i, "can't write beyond current index");
4455 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4456 identical++;
4457 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4458 continue;
4459 }
4460 if (identical > 0) {
4461 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4462 identical = 0;
4463 } else {
4464 Node* n = _expensive_nodes.at(i);
4465 igvn.replace_input_of(n, 0, nullptr);
4466 igvn.hash_insert(n);
4467 modified = true;
4468 }
4469 }
4470 if (identical > 0) {
4471 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4472 } else if (_expensive_nodes.length() >= 1) {
4473 Node* n = _expensive_nodes.at(i);
4474 igvn.replace_input_of(n, 0, nullptr);
4475 igvn.hash_insert(n);
4476 modified = true;
4477 }
4478 _expensive_nodes.trunc_to(j);
4479 if (modified) {
4480 igvn.optimize();
4481 }
4482 }
4483
4484 void Compile::add_expensive_node(Node * n) {
4485 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4486 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4487 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4488 if (OptimizeExpensiveOps) {
4489 _expensive_nodes.append(n);
4490 } else {
4491 // Clear control input and let IGVN optimize expensive nodes if
4492 // OptimizeExpensiveOps is off.
4493 n->set_req(0, nullptr);
4494 }
4495 }
4496
4497 /**
4498 * Track coarsened Lock and Unlock nodes.
4499 */
4500
4501 class Lock_List : public Node_List {
4502 uint _origin_cnt;
4503 public:
4504 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4505 uint origin_cnt() const { return _origin_cnt; }
4506 };
4507
4508 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4509 int length = locks.length();
4510 if (length > 0) {
4511 // Have to keep this list until locks elimination during Macro nodes elimination.
4512 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4513 for (int i = 0; i < length; i++) {
4514 AbstractLockNode* lock = locks.at(i);
4515 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4516 locks_list->push(lock);
4517 }
4518 _coarsened_locks.append(locks_list);
4519 }
4520 }
4521
4522 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4523 int count = coarsened_count();
4524 for (int i = 0; i < count; i++) {
4525 Node_List* locks_list = _coarsened_locks.at(i);
4526 for (uint j = 0; j < locks_list->size(); j++) {
4527 Node* lock = locks_list->at(j);
4528 assert(lock->is_AbstractLock(), "sanity");
4529 if (!useful.member(lock)) {
4530 locks_list->yank(lock);
4531 }
4532 }
4533 }
4534 }
4535
4536 void Compile::remove_coarsened_lock(Node* n) {
4537 if (n->is_AbstractLock()) {
4538 int count = coarsened_count();
4539 for (int i = 0; i < count; i++) {
4540 Node_List* locks_list = _coarsened_locks.at(i);
4541 locks_list->yank(n);
4542 }
4543 }
4544 }
4545
4546 bool Compile::coarsened_locks_consistent() {
4547 int count = coarsened_count();
4548 for (int i = 0; i < count; i++) {
4549 bool unbalanced = false;
4550 bool modified = false; // track locks kind modifications
4551 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4552 uint size = locks_list->size();
4553 if (size == 0) {
4554 unbalanced = false; // All locks were eliminated - good
4555 } else if (size != locks_list->origin_cnt()) {
4556 unbalanced = true; // Some locks were removed from list
4557 } else {
4558 for (uint j = 0; j < size; j++) {
4559 Node* lock = locks_list->at(j);
4560 // All nodes in group should have the same state (modified or not)
4561 if (!lock->as_AbstractLock()->is_coarsened()) {
4562 if (j == 0) {
4563 // first on list was modified, the rest should be too for consistency
4564 modified = true;
4565 } else if (!modified) {
4566 // this lock was modified but previous locks on the list were not
4567 unbalanced = true;
4568 break;
4569 }
4570 } else if (modified) {
4571 // previous locks on list were modified but not this lock
4572 unbalanced = true;
4573 break;
4574 }
4575 }
4576 }
4577 if (unbalanced) {
4578 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4579 #ifdef ASSERT
4580 if (PrintEliminateLocks) {
4581 tty->print_cr("=== unbalanced coarsened locks ===");
4582 for (uint l = 0; l < size; l++) {
4583 locks_list->at(l)->dump();
4584 }
4585 }
4586 #endif
4587 record_failure(C2Compiler::retry_no_locks_coarsening());
4588 return false;
4589 }
4590 }
4591 return true;
4592 }
4593
4594 /**
4595 * Remove the speculative part of types and clean up the graph
4596 */
4597 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4598 if (UseTypeSpeculation) {
4599 Unique_Node_List worklist;
4600 worklist.push(root());
4601 int modified = 0;
4602 // Go over all type nodes that carry a speculative type, drop the
4603 // speculative part of the type and enqueue the node for an igvn
4604 // which may optimize it out.
4605 for (uint next = 0; next < worklist.size(); ++next) {
4606 Node *n = worklist.at(next);
4607 if (n->is_Type()) {
4608 TypeNode* tn = n->as_Type();
4609 const Type* t = tn->type();
4610 const Type* t_no_spec = t->remove_speculative();
4611 if (t_no_spec != t) {
4612 bool in_hash = igvn.hash_delete(n);
4613 assert(in_hash, "node should be in igvn hash table");
4614 tn->set_type(t_no_spec);
4615 igvn.hash_insert(n);
4616 igvn._worklist.push(n); // give it a chance to go away
4617 modified++;
4618 }
4619 }
4620 // Iterate over outs - endless loops is unreachable from below
4621 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4622 Node *m = n->fast_out(i);
4623 if (not_a_node(m)) {
4624 continue;
4625 }
4626 worklist.push(m);
4627 }
4628 }
4629 // Drop the speculative part of all types in the igvn's type table
4630 igvn.remove_speculative_types();
4631 if (modified > 0) {
4632 igvn.optimize();
4633 if (failing()) return;
4634 }
4635 #ifdef ASSERT
4636 // Verify that after the IGVN is over no speculative type has resurfaced
4637 worklist.clear();
4638 worklist.push(root());
4639 for (uint next = 0; next < worklist.size(); ++next) {
4640 Node *n = worklist.at(next);
4641 const Type* t = igvn.type_or_null(n);
4642 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4643 if (n->is_Type()) {
4644 t = n->as_Type()->type();
4645 assert(t == t->remove_speculative(), "no more speculative types");
4646 }
4647 // Iterate over outs - endless loops is unreachable from below
4648 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4649 Node *m = n->fast_out(i);
4650 if (not_a_node(m)) {
4651 continue;
4652 }
4653 worklist.push(m);
4654 }
4655 }
4656 igvn.check_no_speculative_types();
4657 #endif
4658 }
4659 }
4660
4661 // Auxiliary methods to support randomized stressing/fuzzing.
4662
4663 int Compile::random() {
4664 _stress_seed = os::next_random(_stress_seed);
4665 return static_cast<int>(_stress_seed);
4666 }
4667
4668 // This method can be called the arbitrary number of times, with current count
4669 // as the argument. The logic allows selecting a single candidate from the
4670 // running list of candidates as follows:
4671 // int count = 0;
4672 // Cand* selected = null;
4673 // while(cand = cand->next()) {
4674 // if (randomized_select(++count)) {
4675 // selected = cand;
4676 // }
4677 // }
4678 //
4679 // Including count equalizes the chances any candidate is "selected".
4680 // This is useful when we don't have the complete list of candidates to choose
4681 // from uniformly. In this case, we need to adjust the randomicity of the
4682 // selection, or else we will end up biasing the selection towards the latter
4683 // candidates.
4684 //
4685 // Quick back-envelope calculation shows that for the list of n candidates
4686 // the equal probability for the candidate to persist as "best" can be
4687 // achieved by replacing it with "next" k-th candidate with the probability
4688 // of 1/k. It can be easily shown that by the end of the run, the
4689 // probability for any candidate is converged to 1/n, thus giving the
4690 // uniform distribution among all the candidates.
4691 //
4692 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4693 #define RANDOMIZED_DOMAIN_POW 29
4694 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4695 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4696 bool Compile::randomized_select(int count) {
4697 assert(count > 0, "only positive");
4698 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4699 }
4700
4701 CloneMap& Compile::clone_map() { return _clone_map; }
4702 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4703
4704 void NodeCloneInfo::dump() const {
4705 tty->print(" {%d:%d} ", idx(), gen());
4706 }
4707
4708 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4709 uint64_t val = value(old->_idx);
4710 NodeCloneInfo cio(val);
4711 assert(val != 0, "old node should be in the map");
4712 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4713 insert(nnn->_idx, cin.get());
4714 #ifndef PRODUCT
4715 if (is_debug()) {
4716 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4717 }
4718 #endif
4719 }
4720
4721 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4722 NodeCloneInfo cio(value(old->_idx));
4723 if (cio.get() == 0) {
4724 cio.set(old->_idx, 0);
4725 insert(old->_idx, cio.get());
4726 #ifndef PRODUCT
4727 if (is_debug()) {
4728 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4729 }
4730 #endif
4731 }
4732 clone(old, nnn, gen);
4733 }
4734
4735 int CloneMap::max_gen() const {
4736 int g = 0;
4737 DictI di(_dict);
4738 for(; di.test(); ++di) {
4739 int t = gen(di._key);
4740 if (g < t) {
4741 g = t;
4742 #ifndef PRODUCT
4743 if (is_debug()) {
4744 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4745 }
4746 #endif
4747 }
4748 }
4749 return g;
4750 }
4751
4752 void CloneMap::dump(node_idx_t key) const {
4753 uint64_t val = value(key);
4754 if (val != 0) {
4755 NodeCloneInfo ni(val);
4756 ni.dump();
4757 }
4758 }
4759
4760 // Move Allocate nodes to the start of the list
4761 void Compile::sort_macro_nodes() {
4762 int count = macro_count();
4763 int allocates = 0;
4764 for (int i = 0; i < count; i++) {
4765 Node* n = macro_node(i);
4766 if (n->is_Allocate()) {
4767 if (i != allocates) {
4768 Node* tmp = macro_node(allocates);
4769 _macro_nodes.at_put(allocates, n);
4770 _macro_nodes.at_put(i, tmp);
4771 }
4772 allocates++;
4773 }
4774 }
4775 }
4776
4777 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) {
4778 EventCompilerPhase event;
4779 if (event.should_commit()) {
4780 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
4781 }
4782 #ifndef PRODUCT
4783 if (should_print(level)) {
4784 _printer->print_method(name, level);
4785 }
4786 #endif
4787 C->_latest_stage_start_counter.stamp();
4788 }
4789
4790 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) {
4791 char output[1024];
4792 #ifndef PRODUCT
4793 if (idx != 0) {
4794 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx);
4795 } else {
4796 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt));
4797 }
4798 #endif
4799 print_method(cpt, output, level);
4800 }
4801
4802 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
4803 ResourceMark rm;
4804 stringStream ss;
4805 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
4806 if (n != nullptr) {
4807 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
4808 } else {
4809 ss.print_raw(": nullptr");
4810 }
4811 C->print_method(cpt, ss.as_string(), level);
4812 }
4813
4814 void Compile::end_method(int level) {
4815 EventCompilerPhase event;
4816 if (event.should_commit()) {
4817 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level);
4818 }
4819
4820 #ifndef PRODUCT
4821 if (_method != nullptr && should_print(level)) {
4822 _printer->end_method();
4823 }
4824 #endif
4825 }
4826
4827
4828 #ifndef PRODUCT
4829 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
4830 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
4831
4832 // Called from debugger. Prints method to the default file with the default phase name.
4833 // This works regardless of any Ideal Graph Visualizer flags set or not.
4834 void igv_print() {
4835 Compile::current()->igv_print_method_to_file();
4836 }
4837
4838 // Same as igv_print() above but with a specified phase name.
4839 void igv_print(const char* phase_name) {
4840 Compile::current()->igv_print_method_to_file(phase_name);
4841 }
4842
4843 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
4844 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
4845 // This works regardless of any Ideal Graph Visualizer flags set or not.
4846 void igv_print(bool network) {
4847 if (network) {
4848 Compile::current()->igv_print_method_to_network();
4849 } else {
4850 Compile::current()->igv_print_method_to_file();
4851 }
4852 }
4853
4854 // Same as igv_print(bool network) above but with a specified phase name.
4855 void igv_print(bool network, const char* phase_name) {
4856 if (network) {
4857 Compile::current()->igv_print_method_to_network(phase_name);
4858 } else {
4859 Compile::current()->igv_print_method_to_file(phase_name);
4860 }
4861 }
4862
4863 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
4864 void igv_print_default() {
4865 Compile::current()->print_method(PHASE_DEBUG, 0);
4866 }
4867
4868 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
4869 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
4870 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
4871 void igv_append() {
4872 Compile::current()->igv_print_method_to_file("Debug", true);
4873 }
4874
4875 // Same as igv_append() above but with a specified phase name.
4876 void igv_append(const char* phase_name) {
4877 Compile::current()->igv_print_method_to_file(phase_name, true);
4878 }
4879
4880 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
4881 const char* file_name = "custom_debug.xml";
4882 if (_debug_file_printer == nullptr) {
4883 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
4884 } else {
4885 _debug_file_printer->update_compiled_method(C->method());
4886 }
4887 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
4888 _debug_file_printer->print(phase_name, (Node*)C->root());
4889 }
4890
4891 void Compile::igv_print_method_to_network(const char* phase_name) {
4892 if (_debug_network_printer == nullptr) {
4893 _debug_network_printer = new IdealGraphPrinter(C);
4894 } else {
4895 _debug_network_printer->update_compiled_method(C->method());
4896 }
4897 tty->print_cr("Method printed over network stream to IGV");
4898 _debug_network_printer->print(phase_name, (Node*)C->root());
4899 }
4900 #endif
4901
4902 void Compile::add_native_invoker(RuntimeStub* stub) {
4903 _native_invokers.append(stub);
4904 }
4905
4906 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
4907 if (type != nullptr && phase->type(value)->higher_equal(type)) {
4908 return value;
4909 }
4910 Node* result = nullptr;
4911 if (bt == T_BYTE) {
4912 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
4913 result = new RShiftINode(result, phase->intcon(24));
4914 } else if (bt == T_BOOLEAN) {
4915 result = new AndINode(value, phase->intcon(0xFF));
4916 } else if (bt == T_CHAR) {
4917 result = new AndINode(value,phase->intcon(0xFFFF));
4918 } else {
4919 assert(bt == T_SHORT, "unexpected narrow type");
4920 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
4921 result = new RShiftINode(result, phase->intcon(16));
4922 }
4923 if (transform_res) {
4924 result = phase->transform(result);
4925 }
4926 return result;
4927 }
4928