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 "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/compilerDefinitions.hpp"
38 #include "compiler/compilerOracle.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/locknode.hpp"
63 #include "opto/loopnode.hpp"
64 #include "opto/machnode.hpp"
65 #include "opto/macro.hpp"
66 #include "opto/matcher.hpp"
67 #include "opto/mathexactnode.hpp"
68 #include "opto/memnode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/reachability.hpp"
78 #include "opto/rootnode.hpp"
79 #include "opto/runtime.hpp"
80 #include "opto/stringopts.hpp"
81 #include "opto/type.hpp"
82 #include "opto/vector.hpp"
83 #include "opto/vectornode.hpp"
84 #include "runtime/globals_extension.hpp"
85 #include "runtime/sharedRuntime.hpp"
86 #include "runtime/signature.hpp"
87 #include "runtime/stubRoutines.hpp"
88 #include "runtime/timer.hpp"
89 #include "utilities/align.hpp"
90 #include "utilities/copy.hpp"
91 #include "utilities/hashTable.hpp"
92 #include "utilities/macros.hpp"
93
94 // -------------------- Compile::mach_constant_base_node -----------------------
95 // Constant table base node singleton.
96 MachConstantBaseNode* Compile::mach_constant_base_node() {
97 if (_mach_constant_base_node == nullptr) {
98 _mach_constant_base_node = new MachConstantBaseNode();
99 _mach_constant_base_node->add_req(C->root());
100 }
101 return _mach_constant_base_node;
102 }
103
392 record_dead_node(dead->_idx);
393 }
394 if (dead->is_macro()) {
395 remove_macro_node(dead);
396 }
397 if (dead->is_expensive()) {
398 remove_expensive_node(dead);
399 }
400 if (dead->is_ReachabilityFence()) {
401 remove_reachability_fence(dead->as_ReachabilityFence());
402 }
403 if (dead->is_OpaqueTemplateAssertionPredicate()) {
404 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
405 }
406 if (dead->is_ParsePredicate()) {
407 remove_parse_predicate(dead->as_ParsePredicate());
408 }
409 if (dead->for_post_loop_opts_igvn()) {
410 remove_from_post_loop_opts_igvn(dead);
411 }
412 if (dead->for_merge_stores_igvn()) {
413 remove_from_merge_stores_igvn(dead);
414 }
415 if (dead->is_Call()) {
416 remove_useless_late_inlines( &_late_inlines, dead);
417 remove_useless_late_inlines( &_string_late_inlines, dead);
418 remove_useless_late_inlines( &_boxing_late_inlines, dead);
419 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
420
421 if (dead->is_CallStaticJava()) {
422 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
423 }
424 }
425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
426 bs->unregister_potential_barrier_node(dead);
427 }
428
429 // Disconnect all useless nodes by disconnecting those at the boundary.
430 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
431 uint next = 0;
439 // Use raw traversal of out edges since this code removes out edges
440 int max = n->outcnt();
441 for (int j = 0; j < max; ++j) {
442 Node* child = n->raw_out(j);
443 if (!useful.member(child)) {
444 assert(!child->is_top() || child != top(),
445 "If top is cached in Compile object it is in useful list");
446 // Only need to remove this out-edge to the useless node
447 n->raw_del_out(j);
448 --j;
449 --max;
450 if (child->is_data_proj_of_pure_function(n)) {
451 worklist.push(n);
452 }
453 }
454 }
455 if (n->outcnt() == 1 && n->has_special_unique_user()) {
456 assert(useful.member(n->unique_out()), "do not push a useless node");
457 worklist.push(n->unique_out());
458 }
459 }
460
461 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
462 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
463 // Remove useless Template Assertion Predicate opaque nodes
464 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
465 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
466 remove_useless_nodes(_reachability_fences, useful); // remove useless node recorded for post loop opts IGVN pass
467 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
468 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
469 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
470 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
471 #ifdef ASSERT
472 if (_modified_nodes != nullptr) {
473 _modified_nodes->remove_useless_nodes(useful.member_set());
474 }
475 #endif
476
477 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
478 bs->eliminate_useless_gc_barriers(useful, this);
479 // clean up the late inline lists
480 remove_useless_late_inlines( &_late_inlines, useful);
481 remove_useless_late_inlines( &_string_late_inlines, useful);
482 remove_useless_late_inlines( &_boxing_late_inlines, useful);
483 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
484 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
485 }
486
487 // ============================================================================
635
636 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
637 Options options, DirectiveSet* directive)
638 : Phase(Compiler),
639 _compile_id(ci_env->compile_id()),
640 _options(options),
641 _method(target),
642 _entry_bci(osr_bci),
643 _ilt(nullptr),
644 _stub_function(nullptr),
645 _stub_name(nullptr),
646 _stub_id(StubId::NO_STUBID),
647 _stub_entry_point(nullptr),
648 _max_node_limit(MaxNodeLimit),
649 _node_count_inlining_cutoff(NodeCountInliningCutoff),
650 _post_loop_opts_phase(false),
651 _merge_stores_phase(false),
652 _allow_macro_nodes(true),
653 _inlining_progress(false),
654 _inlining_incrementally(false),
655 _do_cleanup(false),
656 _has_reserved_stack_access(target->has_reserved_stack_access()),
657 #ifndef PRODUCT
658 _igv_idx(0),
659 _trace_opto_output(directive->TraceOptoOutputOption),
660 #endif
661 _clinit_barrier_on_entry(false),
662 _stress_seed(0),
663 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
664 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
665 _env(ci_env),
666 _directive(directive),
667 _log(ci_env->log()),
668 _first_failure_details(nullptr),
669 _intrinsics(comp_arena(), 0, 0, nullptr),
670 _macro_nodes(comp_arena(), 8, 0, nullptr),
671 _parse_predicates(comp_arena(), 8, 0, nullptr),
672 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
673 _expensive_nodes(comp_arena(), 8, 0, nullptr),
674 _reachability_fences(comp_arena(), 8, 0, nullptr),
675 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
676 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
677 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
678 _coarsened_locks(comp_arena(), 8, 0, nullptr),
679 _congraph(nullptr),
680 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
681 _unique(0),
682 _dead_node_count(0),
683 _dead_node_list(comp_arena()),
684 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
685 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
686 _node_arena(&_node_arena_one),
687 _mach_constant_base_node(nullptr),
688 _Compile_types(mtCompiler, Arena::Tag::tag_type),
689 _initial_gvn(nullptr),
690 _igvn_worklist(nullptr),
691 _types(nullptr),
692 _node_hash(nullptr),
693 _late_inlines(comp_arena(), 2, 0, nullptr),
694 _string_late_inlines(comp_arena(), 2, 0, nullptr),
695 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
764 #define MINIMUM_NODE_HASH 1023
765
766 // GVN that will be run immediately on new nodes
767 uint estimated_size = method()->code_size()*4+64;
768 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
769 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
770 _types = new (comp_arena()) Type_Array(comp_arena());
771 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
772 PhaseGVN gvn;
773 set_initial_gvn(&gvn);
774
775 { // Scope for timing the parser
776 TracePhase tp(_t_parser);
777
778 // Put top into the hash table ASAP.
779 initial_gvn()->transform(top());
780
781 // Set up tf(), start(), and find a CallGenerator.
782 CallGenerator* cg = nullptr;
783 if (is_osr_compilation()) {
784 const TypeTuple *domain = StartOSRNode::osr_domain();
785 const TypeTuple *range = TypeTuple::make_range(method()->signature());
786 init_tf(TypeFunc::make(domain, range));
787 StartNode* s = new StartOSRNode(root(), domain);
788 initial_gvn()->set_type_bottom(s);
789 verify_start(s);
790 cg = CallGenerator::for_osr(method(), entry_bci());
791 } else {
792 // Normal case.
793 init_tf(TypeFunc::make(method()));
794 StartNode* s = new StartNode(root(), tf()->domain());
795 initial_gvn()->set_type_bottom(s);
796 verify_start(s);
797 float past_uses = method()->interpreter_invocation_count();
798 float expected_uses = past_uses;
799 cg = CallGenerator::for_inline(method(), expected_uses);
800 }
801 if (failing()) return;
802 if (cg == nullptr) {
803 const char* reason = InlineTree::check_can_parse(method());
804 assert(reason != nullptr, "expect reason for parse failure");
805 stringStream ss;
806 ss.print("cannot parse method: %s", reason);
807 record_method_not_compilable(ss.as_string());
808 return;
809 }
810
811 gvn.set_type(root(), root()->bottom_type());
812
813 JVMState* jvms = build_start_state(start(), tf());
814 if ((jvms = cg->generate(jvms)) == nullptr) {
872
873 #ifndef PRODUCT
874 if (should_print_ideal()) {
875 print_ideal_ir("PrintIdeal");
876 }
877 #endif
878
879 #ifdef ASSERT
880 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
881 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
882 #endif
883
884 // Dump compilation data to replay it.
885 if (directive->DumpReplayOption) {
886 env()->dump_replay_data(_compile_id);
887 }
888 if (directive->DumpInlineOption && (ilt() != nullptr)) {
889 env()->dump_inline_data(_compile_id);
890 }
891
892 // Now that we know the size of all the monitors we can add a fixed slot
893 // for the original deopt pc.
894 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
895 set_fixed_slots(next_slot);
896
897 // Compute when to use implicit null checks. Used by matching trap based
898 // nodes and NullCheck optimization.
899 set_allowed_deopt_reasons();
900
901 // Now generate code
902 Code_Gen();
903 }
904
905 //------------------------------Compile----------------------------------------
906 // Compile a runtime stub
907 Compile::Compile(ciEnv* ci_env,
908 TypeFunc_generator generator,
909 address stub_function,
910 const char* stub_name,
911 StubId stub_id,
912 int is_fancy_jump,
913 bool pass_tls,
914 bool return_pc,
915 DirectiveSet* directive)
916 : Phase(Compiler),
917 _compile_id(0),
918 _options(Options::for_runtime_stub()),
919 _method(nullptr),
920 _entry_bci(InvocationEntryBci),
921 _stub_function(stub_function),
922 _stub_name(stub_name),
923 _stub_id(stub_id),
924 _stub_entry_point(nullptr),
925 _max_node_limit(MaxNodeLimit),
926 _node_count_inlining_cutoff(NodeCountInliningCutoff),
927 _post_loop_opts_phase(false),
928 _merge_stores_phase(false),
929 _allow_macro_nodes(true),
930 _inlining_progress(false),
931 _inlining_incrementally(false),
932 _has_reserved_stack_access(false),
933 #ifndef PRODUCT
934 _igv_idx(0),
935 _trace_opto_output(directive->TraceOptoOutputOption),
936 #endif
937 _clinit_barrier_on_entry(false),
938 _stress_seed(0),
939 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
940 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
941 _env(ci_env),
942 _directive(directive),
943 _log(ci_env->log()),
944 _first_failure_details(nullptr),
945 _reachability_fences(comp_arena(), 8, 0, nullptr),
946 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
947 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
948 _congraph(nullptr),
949 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
950 _unique(0),
951 _dead_node_count(0),
952 _dead_node_list(comp_arena()),
1068 _fixed_slots = 0;
1069 set_has_split_ifs(false);
1070 set_has_loops(false); // first approximation
1071 set_has_stringbuilder(false);
1072 set_has_boxed_value(false);
1073 _trap_can_recompile = false; // no traps emitted yet
1074 _major_progress = true; // start out assuming good things will happen
1075 set_has_unsafe_access(false);
1076 set_max_vector_size(0);
1077 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1078 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1079 set_decompile_count(0);
1080
1081 #ifndef PRODUCT
1082 _phase_counter = 0;
1083 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1084 #endif
1085
1086 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1087 _loop_opts_cnt = LoopOptsCount;
1088 set_do_inlining(Inline);
1089 set_max_inline_size(MaxInlineSize);
1090 set_freq_inline_size(FreqInlineSize);
1091 set_do_scheduling(OptoScheduling);
1092
1093 set_do_vector_loop(false);
1094 set_has_monitors(false);
1095 set_has_scoped_access(false);
1096
1097 if (AllowVectorizeOnDemand) {
1098 if (has_method() && _directive->VectorizeOption) {
1099 set_do_vector_loop(true);
1100 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());})
1101 } else if (has_method() && method()->name() != nullptr &&
1102 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1103 set_do_vector_loop(true);
1104 }
1105 }
1106 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1107 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());})
1339 // If this method has already thrown a range-check,
1340 // assume it was because we already tried range smearing
1341 // and it failed.
1342 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1343 return !already_trapped;
1344 }
1345
1346
1347 //------------------------------flatten_alias_type-----------------------------
1348 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1349 assert(do_aliasing(), "Aliasing should be enabled");
1350 int offset = tj->offset();
1351 TypePtr::PTR ptr = tj->ptr();
1352
1353 // Known instance (scalarizable allocation) alias only with itself.
1354 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1355 tj->is_oopptr()->is_known_instance();
1356
1357 // Process weird unsafe references.
1358 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1359 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1360 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1361 tj = TypeOopPtr::BOTTOM;
1362 ptr = tj->ptr();
1363 offset = tj->offset();
1364 }
1365
1366 // Array pointers need some flattening
1367 const TypeAryPtr* ta = tj->isa_aryptr();
1368 if (ta && ta->is_stable()) {
1369 // Erase stability property for alias analysis.
1370 tj = ta = ta->cast_to_stable(false);
1371 }
1372 if( ta && is_known_inst ) {
1373 if ( offset != Type::OffsetBot &&
1374 offset > arrayOopDesc::length_offset_in_bytes() ) {
1375 offset = Type::OffsetBot; // Flatten constant access into array body only
1376 tj = ta = ta->
1377 remove_speculative()->
1378 cast_to_ptr_type(ptr)->
1379 with_offset(offset);
1380 }
1381 } else if (ta != nullptr) {
1382 // Common slices
1383 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1384 return TypeAryPtr::RANGE;
1385 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1386 return TypeInstPtr::KLASS;
1387 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1388 return TypeInstPtr::MARK;
1389 }
1390
1391 // Remove size and stability
1392 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false);
1393 // Remove ptr, const_oop, and offset
1394 if (ta->elem() == Type::BOTTOM) {
1395 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1396 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1397 // incorrect, see JDK-8331133).
1398 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::OffsetBot);
1399 } else if (ta->elem()->make_oopptr() != nullptr) {
1400 // Object arrays, all of them share the same slice
1401 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false);
1402 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::OffsetBot);
1403 } else {
1404 // Primitive arrays
1405 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::OffsetBot);
1406 }
1407
1408 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1409 // cannot be distinguished by bytecode alone.
1410 if (ta->elem() == TypeInt::BOOL) {
1411 tj = ta = TypeAryPtr::BYTES;
1412 }
1413 }
1414
1415 // Oop pointers need some flattening
1416 const TypeInstPtr *to = tj->isa_instptr();
1417 if (to && to != TypeOopPtr::BOTTOM) {
1418 ciInstanceKlass* ik = to->instance_klass();
1419 if( ptr == TypePtr::Constant ) {
1420 if (ik != ciEnv::current()->Class_klass() ||
1421 offset < ik->layout_helper_size_in_bytes()) {
1422 // No constant oop pointers (such as Strings); they alias with
1423 // unknown strings.
1424 assert(!is_known_inst, "not scalarizable allocation");
1425 tj = to = to->
1426 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1427 remove_speculative()->
1428 cast_to_ptr_type(TypePtr::BotPTR)->
1429 cast_to_exactness(false);
1430 }
1431 } else if( is_known_inst ) {
1432 tj = to; // Keep NotNull and klass_is_exact for instance type
1433 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1434 // During the 2nd round of IterGVN, NotNull castings are removed.
1435 // Make sure the Bottom and NotNull variants alias the same.
1436 // Also, make sure exact and non-exact variants alias the same.
1437 tj = to = to->
1438 remove_speculative()->
1439 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1440 cast_to_ptr_type(TypePtr::BotPTR)->
1441 cast_to_exactness(false);
1442 }
1443 if (to->speculative() != nullptr) {
1444 tj = to = to->remove_speculative();
1445 }
1446 // Canonicalize the holder of this field
1447 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1448 // First handle header references such as a LoadKlassNode, even if the
1449 // object's klass is unloaded at compile time (4965979).
1450 if (!is_known_inst) { // Do it only for non-instance types
1451 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1452 }
1453 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1454 // Static fields are in the space above the normal instance
1455 // fields in the java.lang.Class instance.
1456 if (ik != ciEnv::current()->Class_klass()) {
1457 to = nullptr;
1458 tj = TypeOopPtr::BOTTOM;
1459 offset = tj->offset();
1460 }
1461 } else {
1462 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1463 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1464 assert(tj->offset() == offset, "no change to offset expected");
1465 bool xk = to->klass_is_exact();
1466 int instance_id = to->instance_id();
1467
1468 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1469 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1470 // its interfaces are included.
1471 if (xk && ik->equals(canonical_holder)) {
1472 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1473 } else {
1474 assert(xk || !is_known_inst, "Known instance should be exact type");
1475 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1476 }
1477 }
1478 }
1479
1480 // Klass pointers to object array klasses need some flattening
1481 const TypeKlassPtr *tk = tj->isa_klassptr();
1482 if( tk ) {
1483 // If we are referencing a field within a Klass, we need
1484 // to assume the worst case of an Object. Both exact and
1485 // inexact types must flatten to the same alias class so
1486 // use NotNull as the PTR.
1487 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1488 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1489 env()->Object_klass(),
1490 offset);
1491 }
1492
1493 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1494 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1495 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1496 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1497 } else {
1498 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1499 }
1500 }
1501
1502 // Check for precise loads from the primary supertype array and force them
1503 // to the supertype cache alias index. Check for generic array loads from
1504 // the primary supertype array and also force them to the supertype cache
1505 // alias index. Since the same load can reach both, we need to merge
1506 // these 2 disparate memories into the same alias class. Since the
1507 // primary supertype array is read-only, there's no chance of confusion
1508 // where we bypass an array load and an array store.
1509 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1510 if (offset == Type::OffsetBot ||
1511 (offset >= primary_supers_offset &&
1512 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1513 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1514 offset = in_bytes(Klass::secondary_super_cache_offset());
1515 tj = tk = tk->with_offset(offset);
1516 }
1517 }
1518
1519 // Flatten all Raw pointers together.
1520 if (tj->base() == Type::RawPtr)
1521 tj = TypeRawPtr::BOTTOM;
1611 intptr_t key = (intptr_t) adr_type;
1612 key ^= key >> logAliasCacheSize;
1613 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1614 }
1615
1616
1617 //-----------------------------grow_alias_types--------------------------------
1618 void Compile::grow_alias_types() {
1619 const int old_ats = _max_alias_types; // how many before?
1620 const int new_ats = old_ats; // how many more?
1621 const int grow_ats = old_ats+new_ats; // how many now?
1622 _max_alias_types = grow_ats;
1623 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), _alias_types, old_ats, grow_ats);
1624 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1625 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1626 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1627 }
1628
1629
1630 //--------------------------------find_alias_type------------------------------
1631 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1632 if (!do_aliasing()) {
1633 return alias_type(AliasIdxBot);
1634 }
1635
1636 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1637 if (ace->_adr_type == adr_type) {
1638 return alias_type(ace->_index);
1639 }
1640
1641 // Handle special cases.
1642 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1643 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1644
1645 // Do it the slow way.
1646 const TypePtr* flat = flatten_alias_type(adr_type);
1647
1648 #ifdef ASSERT
1649 {
1650 ResourceMark rm;
1651 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1652 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1653 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1654 Type::str(adr_type));
1655 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1656 const TypeOopPtr* foop = flat->is_oopptr();
1657 // Scalarizable allocations have exact klass always.
1658 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1668 if (alias_type(i)->adr_type() == flat) {
1669 idx = i;
1670 break;
1671 }
1672 }
1673
1674 if (idx == AliasIdxTop) {
1675 if (no_create) return nullptr;
1676 // Grow the array if necessary.
1677 if (_num_alias_types == _max_alias_types) grow_alias_types();
1678 // Add a new alias type.
1679 idx = _num_alias_types++;
1680 _alias_types[idx]->Init(idx, flat);
1681 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1682 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1683 if (flat->isa_instptr()) {
1684 if (flat->offset() == java_lang_Class::klass_offset()
1685 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1686 alias_type(idx)->set_rewritable(false);
1687 }
1688 if (flat->isa_aryptr()) {
1689 #ifdef ASSERT
1690 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1691 // (T_BYTE has the weakest alignment and size restrictions...)
1692 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1693 #endif
1694 if (flat->offset() == TypePtr::OffsetBot) {
1695 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1696 }
1697 }
1698 if (flat->isa_klassptr()) {
1699 if (UseCompactObjectHeaders) {
1700 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1701 alias_type(idx)->set_rewritable(false);
1702 }
1703 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1704 alias_type(idx)->set_rewritable(false);
1705 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1708 alias_type(idx)->set_rewritable(false);
1709 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1710 alias_type(idx)->set_rewritable(false);
1711 }
1712
1713 if (flat->isa_instklassptr()) {
1714 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1715 alias_type(idx)->set_rewritable(false);
1716 }
1717 }
1718 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1719 // but the base pointer type is not distinctive enough to identify
1720 // references into JavaThread.)
1721
1722 // Check for final fields.
1723 const TypeInstPtr* tinst = flat->isa_instptr();
1724 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1725 ciField* field;
1726 if (tinst->const_oop() != nullptr &&
1727 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1728 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1729 // static field
1730 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1731 field = k->get_field_by_offset(tinst->offset(), true);
1732 } else {
1733 ciInstanceKlass *k = tinst->instance_klass();
1734 field = k->get_field_by_offset(tinst->offset(), false);
1735 }
1736 assert(field == nullptr ||
1737 original_field == nullptr ||
1738 (field->holder() == original_field->holder() &&
1739 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1740 field->is_static() == original_field->is_static()), "wrong field?");
1741 // Set field() and is_rewritable() attributes.
1742 if (field != nullptr) alias_type(idx)->set_field(field);
1743 }
1744 }
1745
1746 // Fill the cache for next time.
1747 ace->_adr_type = adr_type;
1748 ace->_index = idx;
1749 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1750
1751 // Might as well try to fill the cache for the flattened version, too.
1752 AliasCacheEntry* face = probe_alias_cache(flat);
1753 if (face->_adr_type == nullptr) {
1754 face->_adr_type = flat;
1755 face->_index = idx;
1756 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1757 }
1758
1759 return alias_type(idx);
1760 }
1761
1762
1763 Compile::AliasType* Compile::alias_type(ciField* field) {
1764 const TypeOopPtr* t;
1765 if (field->is_static())
1766 t = TypeInstPtr::make(field->holder()->java_mirror());
1767 else
1768 t = TypeOopPtr::make_from_klass_raw(field->holder());
1769 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1770 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1771 return atp;
1772 }
1773
1774
1775 //------------------------------have_alias_type--------------------------------
1776 bool Compile::have_alias_type(const TypePtr* adr_type) {
1858 assert(!C->major_progress(), "not cleared");
1859
1860 if (_for_post_loop_igvn.length() > 0) {
1861 while (_for_post_loop_igvn.length() > 0) {
1862 Node* n = _for_post_loop_igvn.pop();
1863 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1864 igvn._worklist.push(n);
1865 }
1866 igvn.optimize();
1867 if (failing()) return;
1868 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1869 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1870
1871 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1872 if (C->major_progress()) {
1873 C->clear_major_progress(); // ensure that major progress is now clear
1874 }
1875 }
1876 }
1877
1878 void Compile::record_for_merge_stores_igvn(Node* n) {
1879 if (!n->for_merge_stores_igvn()) {
1880 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1881 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1882 _for_merge_stores_igvn.append(n);
1883 }
1884 }
1885
1886 void Compile::remove_from_merge_stores_igvn(Node* n) {
1887 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1888 _for_merge_stores_igvn.remove(n);
1889 }
1890
1891 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1892 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1893 // the stores, and we merge the wrong sequence of stores.
1894 // Example:
1895 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1896 // Apply MergeStores:
1897 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1976 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1977 Bytecodes::Code c = iter.cur_bc();
1978 Node* lhs = nullptr;
1979 Node* rhs = nullptr;
1980 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1981 lhs = unc->peek_operand(0);
1982 rhs = unc->peek_operand(1);
1983 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1984 lhs = unc->peek_operand(0);
1985 }
1986
1987 ResourceMark rm;
1988 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1989 assert(live_locals.is_valid(), "broken liveness info");
1990 int len = (int)live_locals.size();
1991
1992 for (int i = 0; i < len; i++) {
1993 Node* local = unc->local(jvms, i);
1994 // kill local using the liveness of next_bci.
1995 // give up when the local looks like an operand to secure reexecution.
1996 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1997 uint idx = jvms->locoff() + i;
1998 #ifdef ASSERT
1999 if (PrintOpto && Verbose) {
2000 tty->print("[unstable_if] kill local#%d: ", idx);
2001 local->dump();
2002 tty->cr();
2003 }
2004 #endif
2005 igvn.replace_input_of(unc, idx, top());
2006 modified = true;
2007 }
2008 }
2009 }
2010
2011 // keep the mondified trap for late query
2012 if (modified) {
2013 trap->set_modified();
2014 } else {
2015 _unstable_if_traps.delete_at(i);
2016 }
2017 }
2018 igvn.optimize();
2019 }
2020
2021 // StringOpts and late inlining of string methods
2022 void Compile::inline_string_calls(bool parse_time) {
2023 {
2024 // remove useless nodes to make the usage analysis simpler
2025 ResourceMark rm;
2026 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2027 }
2028
2029 {
2030 ResourceMark rm;
2031 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2223
2224 if (_string_late_inlines.length() > 0) {
2225 assert(has_stringbuilder(), "inconsistent");
2226
2227 inline_string_calls(false);
2228
2229 if (failing()) return;
2230
2231 inline_incrementally_cleanup(igvn);
2232 }
2233
2234 set_inlining_incrementally(false);
2235 }
2236
2237 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2238 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2239 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2240 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2241 // as if "inlining_incrementally() == true" were set.
2242 assert(inlining_incrementally() == false, "not allowed");
2243 assert(_modified_nodes == nullptr, "not allowed");
2244 assert(_late_inlines.length() > 0, "sanity");
2245
2246 if (StressIncrementalInlining) {
2247 shuffle_late_inlines();
2248 }
2249
2250 while (_late_inlines.length() > 0) {
2251 igvn_worklist()->ensure_empty(); // should be done with igvn
2252
2253 while (inline_incrementally_one()) {
2254 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2255 }
2256 if (failing()) return;
2257
2258 inline_incrementally_cleanup(igvn);
2259 }
2260 }
2261
2262 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2263 if (_loop_opts_cnt > 0) {
2264 while (major_progress() && (_loop_opts_cnt > 0)) {
2265 TracePhase tp(_t_idealLoop);
2266 PhaseIdealLoop::optimize(igvn, mode);
2267 _loop_opts_cnt--;
2268 if (failing()) return false;
2269 if (major_progress()) {
2270 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2271 }
2272 }
2273 }
2274 return true;
2275 }
2276
2277 // Remove edges from "root" to each SafePoint at a backward branch.
2278 // They were inserted during parsing (see add_safepoint()) to make
2279 // infinite loops without calls or exceptions visible to root, i.e.,
2385 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2386 }
2387 assert(!has_vbox_nodes(), "sanity");
2388
2389 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2390 Compile::TracePhase tp(_t_renumberLive);
2391 igvn_worklist()->ensure_empty(); // should be done with igvn
2392 {
2393 ResourceMark rm;
2394 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2395 }
2396 igvn.reset();
2397 igvn.optimize(true);
2398 if (failing()) return;
2399 }
2400
2401 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2402 // safepoints
2403 remove_root_to_sfpts_edges(igvn);
2404
2405 if (failing()) return;
2406
2407 _print_phase_loop_opts = has_loops();
2408 if (_print_phase_loop_opts) {
2409 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2410 }
2411
2412 // Perform escape analysis
2413 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2414 if (has_loops()) {
2415 // Cleanup graph (remove dead nodes).
2416 TracePhase tp(_t_idealLoop);
2417 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2418 if (failing()) return;
2419 }
2420 bool progress;
2421 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2422 do {
2423 ConnectionGraph::do_analysis(this, &igvn);
2424
2425 if (failing()) return;
2426
2427 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2428
2429 // Optimize out fields loads from scalar replaceable allocations.
2430 igvn.optimize(true);
2431 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2432
2433 if (failing()) return;
2434
2435 if (congraph() != nullptr && macro_count() > 0) {
2436 TracePhase tp(_t_macroEliminate);
2437 PhaseMacroExpand mexp(igvn);
2438 mexp.eliminate_macro_nodes();
2439 if (failing()) return;
2440 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2441
2442 igvn.set_delay_transform(false);
2443 igvn.optimize();
2444 if (failing()) return;
2445
2446 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2447 }
2448
2449 ConnectionGraph::verify_ram_nodes(this, root());
2450 if (failing()) return;
2451
2452 progress = do_iterative_escape_analysis() &&
2453 (macro_count() < mcount) &&
2454 ConnectionGraph::has_candidates(this);
2455 // Try again if candidates exist and made progress
2456 // by removing some allocations and/or locks.
2457 } while (progress);
2458 }
2459
2460 // Loop transforms on the ideal graph. Range Check Elimination,
2461 // peeling, unrolling, etc.
2462
2463 // Set loop opts counter
2464 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2465 {
2466 TracePhase tp(_t_idealLoop);
2467 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2468 _loop_opts_cnt--;
2469 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2470 if (failing()) return;
2471 }
2472 // Loop opts pass if partial peeling occurred in previous pass
2473 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2474 TracePhase tp(_t_idealLoop);
2475 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2476 _loop_opts_cnt--;
2477 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2478 if (failing()) return;
2479 }
2527
2528 // Once loop optimizations are over, it is safe to get rid of all reachability fence nodes and
2529 // migrate reachability edges to safepoints.
2530 if (OptimizeReachabilityFences && _reachability_fences.length() > 0) {
2531 TracePhase tp1(_t_idealLoop);
2532 TracePhase tp2(_t_reachability);
2533 PhaseIdealLoop::optimize(igvn, PostLoopOptsExpandReachabilityFences);
2534 print_method(PHASE_EXPAND_REACHABILITY_FENCES, 2);
2535 if (failing()) return;
2536 assert(_reachability_fences.length() == 0 || PreserveReachabilityFencesOnConstants, "no RF nodes allowed");
2537 }
2538
2539 process_for_merge_stores_igvn(igvn);
2540
2541 if (failing()) return;
2542
2543 #ifdef ASSERT
2544 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2545 #endif
2546
2547 {
2548 TracePhase tp(_t_macroExpand);
2549 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2550 PhaseMacroExpand mex(igvn);
2551 // Do not allow new macro nodes once we start to eliminate and expand
2552 C->reset_allow_macro_nodes();
2553 // Last attempt to eliminate macro nodes before expand
2554 mex.eliminate_macro_nodes();
2555 if (failing()) {
2556 return;
2557 }
2558 mex.eliminate_opaque_looplimit_macro_nodes();
2559 if (failing()) {
2560 return;
2561 }
2562 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2563 if (mex.expand_macro_nodes()) {
2564 assert(failing(), "must bail out w/ explicit message");
2565 return;
2566 }
2567 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2568 }
2569
2570 {
2571 TracePhase tp(_t_barrierExpand);
2572 if (bs->expand_barriers(this, igvn)) {
2573 assert(failing(), "must bail out w/ explicit message");
2574 return;
2575 }
2576 print_method(PHASE_BARRIER_EXPANSION, 2);
2577 }
2578
2579 if (C->max_vector_size() > 0) {
2580 C->optimize_logic_cones(igvn);
2581 igvn.optimize();
2582 if (failing()) return;
2583 }
2584
2585 DEBUG_ONLY( _modified_nodes = nullptr; )
2586
2587 assert(igvn._worklist.size() == 0, "not empty");
2588
2589 if (_late_inlines.length() > 0) {
2590 // More opportunities to optimize virtual and MH calls.
2591 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2592 process_late_inline_calls_no_inline(igvn);
2593 if (failing()) return;
2594 }
2595 assert(_late_inlines.length() == 0, "late inline queue must be drained");
2596 } // (End scope of igvn; run destructor if necessary for asserts.)
2597
2598 check_no_dead_use();
2599
2600 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2601 // to remove hashes to unlock nodes for modifications.
2602 C->node_hash()->clear();
2603
2604 // A method with only infinite loops has no edges entering loops from root
2605 {
2606 TracePhase tp(_t_graphReshaping);
2607 if (final_graph_reshaping()) {
2608 assert(failing(), "must bail out w/ explicit message");
2609 return;
2610 }
2611 }
2612
2613 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2614 DEBUG_ONLY(set_phase_optimize_finished();)
2615 }
3274 n->subsume_by(sub, this);
3275 }
3276 }
3277
3278 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3279 switch( nop ) {
3280 case Op_Opaque1: // Remove Opaque Nodes before matching
3281 n->subsume_by(n->in(1), this);
3282 break;
3283 case Op_CallLeafPure: {
3284 // If the pure call is not supported, then lower to a CallLeaf.
3285 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3286 CallNode* call = n->as_Call();
3287 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3288 call->_name, TypeRawPtr::BOTTOM);
3289 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3290 new_call->init_req(TypeFunc::I_O, C->top());
3291 new_call->init_req(TypeFunc::Memory, C->top());
3292 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3293 new_call->init_req(TypeFunc::FramePtr, C->top());
3294 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3295 new_call->init_req(i, call->in(i));
3296 }
3297 n->subsume_by(new_call, this);
3298 }
3299 break;
3300 }
3301 case Op_CallStaticJava:
3302 case Op_CallJava:
3303 case Op_CallDynamicJava:
3304 frc.inc_java_call_count(); // Count java call site;
3305 case Op_CallRuntime:
3306 case Op_CallLeaf:
3307 case Op_CallLeafVector:
3308 case Op_CallLeafNoFP: {
3309 assert (n->is_Call(), "");
3310 CallNode *call = n->as_Call();
3311 // See if uncommon argument is shared
3312 if (call->is_CallStaticJava() && call->as_CallStaticJava()->_name) {
3313 Node *n = call->in(TypeFunc::Parms);
3314 int nop = n->Opcode();
3321 nop != Op_DecodeNKlass &&
3322 !n->is_Mem() &&
3323 !n->is_Phi()) {
3324 Node *x = n->clone();
3325 call->set_req(TypeFunc::Parms, x);
3326 }
3327 }
3328 break;
3329 }
3330
3331 // Mem nodes need explicit cases to satisfy assert(!n->is_Mem()) in default.
3332 case Op_StoreF:
3333 case Op_LoadF:
3334 case Op_StoreD:
3335 case Op_LoadD:
3336 case Op_LoadD_unaligned:
3337 case Op_StoreB:
3338 case Op_StoreC:
3339 case Op_StoreI:
3340 case Op_StoreL:
3341 case Op_CompareAndSwapB:
3342 case Op_CompareAndSwapS:
3343 case Op_CompareAndSwapI:
3344 case Op_CompareAndSwapL:
3345 case Op_CompareAndSwapP:
3346 case Op_CompareAndSwapN:
3347 case Op_WeakCompareAndSwapB:
3348 case Op_WeakCompareAndSwapS:
3349 case Op_WeakCompareAndSwapI:
3350 case Op_WeakCompareAndSwapL:
3351 case Op_WeakCompareAndSwapP:
3352 case Op_WeakCompareAndSwapN:
3353 case Op_CompareAndExchangeB:
3354 case Op_CompareAndExchangeS:
3355 case Op_CompareAndExchangeI:
3356 case Op_CompareAndExchangeL:
3357 case Op_CompareAndExchangeP:
3358 case Op_CompareAndExchangeN:
3359 case Op_GetAndAddS:
3360 case Op_GetAndAddB:
3852 k->subsume_by(m, this);
3853 }
3854 }
3855 }
3856 break;
3857 }
3858 case Op_CmpUL: {
3859 if (!Matcher::has_match_rule(Op_CmpUL)) {
3860 // No support for unsigned long comparisons
3861 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3862 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3863 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3864 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3865 Node* andl = new AndLNode(orl, remove_sign_mask);
3866 Node* cmp = new CmpLNode(andl, n->in(2));
3867 n->subsume_by(cmp, this);
3868 }
3869 break;
3870 }
3871 #ifdef ASSERT
3872 case Op_ConNKlass: {
3873 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3874 ciKlass* klass = tp->is_klassptr()->exact_klass();
3875 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3876 break;
3877 }
3878 #endif
3879 default:
3880 assert(!n->is_Call(), "");
3881 assert(!n->is_Mem(), "");
3882 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3883 break;
3884 }
3885 }
3886
3887 //------------------------------final_graph_reshaping_walk---------------------
3888 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3889 // requires that the walk visits a node's inputs before visiting the node.
3890 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3891 Unique_Node_List sfpt;
4224 }
4225 }
4226
4227 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4228 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4229 }
4230
4231 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4232 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4233 }
4234
4235 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4236 if (holder->is_initialized()) {
4237 return false;
4238 }
4239 if (holder->is_being_initialized()) {
4240 if (accessing_method->holder() == holder) {
4241 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4242 // <init>, or a static method. In all those cases, there was an initialization
4243 // barrier on the holder klass passed.
4244 if (accessing_method->is_static_initializer() ||
4245 accessing_method->is_object_initializer() ||
4246 accessing_method->is_static()) {
4247 return false;
4248 }
4249 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4250 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4251 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4252 // child class can become fully initialized while its parent class is still being initialized.
4253 if (accessing_method->is_static_initializer()) {
4254 return false;
4255 }
4256 }
4257 ciMethod* root = method(); // the root method of compilation
4258 if (root != accessing_method) {
4259 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4260 }
4261 }
4262 return true;
4263 }
4264
4265 #ifndef PRODUCT
4266 //------------------------------verify_bidirectional_edges---------------------
4267 // For each input edge to a node (ie - for each Use-Def edge), verify that
4268 // there is a corresponding Def-Use edge.
4269 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4270 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4271 uint stack_size = live_nodes() >> 4;
4272 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4273 if (root_and_safepoints != nullptr) {
4303 if (in != nullptr && !in->is_top()) {
4304 // Count instances of `next`
4305 int cnt = 0;
4306 for (uint idx = 0; idx < in->_outcnt; idx++) {
4307 if (in->_out[idx] == n) {
4308 cnt++;
4309 }
4310 }
4311 assert(cnt > 0, "Failed to find Def-Use edge.");
4312 // Check for duplicate edges
4313 // walk the input array downcounting the input edges to n
4314 for (uint j = 0; j < length; j++) {
4315 if (n->in(j) == in) {
4316 cnt--;
4317 }
4318 }
4319 assert(cnt == 0, "Mismatched edge count.");
4320 } else if (in == nullptr) {
4321 assert(i == 0 || i >= n->req() ||
4322 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4323 (n->is_Unlock() && i == (n->req() - 1)) ||
4324 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4325 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4326 } else {
4327 assert(in->is_top(), "sanity");
4328 // Nothing to check.
4329 }
4330 }
4331 }
4332 }
4333
4334 //------------------------------verify_graph_edges---------------------------
4335 // Walk the Graph and verify that there is a one-to-one correspondence
4336 // between Use-Def edges and Def-Use edges in the graph.
4337 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4338 if (VerifyGraphEdges) {
4339 Unique_Node_List visited;
4340
4341 // Call graph walk to check edges
4342 verify_bidirectional_edges(visited, root_and_safepoints);
4343 if (no_dead_code) {
4344 // Now make sure that no visited node is used by an unvisited node.
4345 bool dead_nodes = false;
4456 // (1) subklass is already limited to a subtype of superklass => always ok
4457 // (2) subklass does not overlap with superklass => always fail
4458 // (3) superklass has NO subtypes and we can check with a simple compare.
4459 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4460 if (skip) {
4461 return SSC_full_test; // Let caller generate the general case.
4462 }
4463
4464 if (subk->is_java_subtype_of(superk)) {
4465 return SSC_always_true; // (0) and (1) this test cannot fail
4466 }
4467
4468 if (!subk->maybe_java_subtype_of(superk)) {
4469 return SSC_always_false; // (2) true path dead; no dynamic test needed
4470 }
4471
4472 const Type* superelem = superk;
4473 if (superk->isa_aryklassptr()) {
4474 int ignored;
4475 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4476 }
4477
4478 if (superelem->isa_instklassptr()) {
4479 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4480 if (!ik->has_subklass()) {
4481 if (!ik->is_final()) {
4482 // Add a dependency if there is a chance of a later subclass.
4483 dependencies()->assert_leaf_type(ik);
4484 }
4485 if (!superk->maybe_java_subtype_of(subk)) {
4486 return SSC_always_false;
4487 }
4488 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4489 }
4490 } else {
4491 // A primitive array type has no subtypes.
4492 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4493 }
4494
4495 return SSC_full_test;
5288 } else {
5289 _debug_network_printer->update_compiled_method(C->method());
5290 }
5291 tty->print_cr("Method printed over network stream to IGV");
5292 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5293 }
5294 #endif // !PRODUCT
5295
5296 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5297 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5298 return value;
5299 }
5300 Node* result = nullptr;
5301 if (bt == T_BYTE) {
5302 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5303 result = new RShiftINode(result, phase->intcon(24));
5304 } else if (bt == T_BOOLEAN) {
5305 result = new AndINode(value, phase->intcon(0xFF));
5306 } else if (bt == T_CHAR) {
5307 result = new AndINode(value,phase->intcon(0xFFFF));
5308 } else {
5309 assert(bt == T_SHORT, "unexpected narrow type");
5310 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5311 result = new RShiftINode(result, phase->intcon(16));
5312 }
5313 if (transform_res) {
5314 result = phase->transform(result);
5315 }
5316 return result;
5317 }
5318
5319 void Compile::record_method_not_compilable_oom() {
5320 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5321 }
5322
5323 #ifndef PRODUCT
5324 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5325 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5326 // Follow non-control edges until we reach CFG nodes
5327 for (uint i = 0; i < worklist.size(); i++) {
|
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 "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciFlatArray.hpp"
28 #include "ci/ciInlineKlass.hpp"
29 #include "ci/ciReplay.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "code/aotCodeCache.hpp"
32 #include "code/exceptionHandlerTable.hpp"
33 #include "code/nmethod.hpp"
34 #include "compiler/compilationFailureInfo.hpp"
35 #include "compiler/compilationMemoryStatistic.hpp"
36 #include "compiler/compileBroker.hpp"
37 #include "compiler/compileLog.hpp"
38 #include "compiler/compiler_globals.hpp"
39 #include "compiler/compilerDefinitions.hpp"
40 #include "compiler/compilerOracle.hpp"
41 #include "compiler/disassembler.hpp"
42 #include "compiler/oopMap.hpp"
43 #include "gc/shared/barrierSet.hpp"
44 #include "gc/shared/c2/barrierSetC2.hpp"
45 #include "jfr/jfrEvents.hpp"
46 #include "jvm_io.h"
47 #include "memory/allocation.hpp"
48 #include "memory/arena.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "opto/addnode.hpp"
51 #include "opto/block.hpp"
52 #include "opto/c2compiler.hpp"
53 #include "opto/callGenerator.hpp"
54 #include "opto/callnode.hpp"
55 #include "opto/castnode.hpp"
56 #include "opto/cfgnode.hpp"
57 #include "opto/chaitin.hpp"
58 #include "opto/compile.hpp"
59 #include "opto/connode.hpp"
60 #include "opto/convertnode.hpp"
61 #include "opto/divnode.hpp"
62 #include "opto/escape.hpp"
63 #include "opto/idealGraphPrinter.hpp"
64 #include "opto/inlinetypenode.hpp"
65 #include "opto/locknode.hpp"
66 #include "opto/loopnode.hpp"
67 #include "opto/machnode.hpp"
68 #include "opto/macro.hpp"
69 #include "opto/matcher.hpp"
70 #include "opto/mathexactnode.hpp"
71 #include "opto/memnode.hpp"
72 #include "opto/movenode.hpp"
73 #include "opto/mulnode.hpp"
74 #include "opto/multnode.hpp"
75 #include "opto/narrowptrnode.hpp"
76 #include "opto/node.hpp"
77 #include "opto/opaquenode.hpp"
78 #include "opto/opcodes.hpp"
79 #include "opto/output.hpp"
80 #include "opto/parse.hpp"
81 #include "opto/phaseX.hpp"
82 #include "opto/reachability.hpp"
83 #include "opto/rootnode.hpp"
84 #include "opto/runtime.hpp"
85 #include "opto/stringopts.hpp"
86 #include "opto/type.hpp"
87 #include "opto/vector.hpp"
88 #include "opto/vectornode.hpp"
89 #include "runtime/arguments.hpp"
90 #include "runtime/globals_extension.hpp"
91 #include "runtime/sharedRuntime.hpp"
92 #include "runtime/signature.hpp"
93 #include "runtime/stubRoutines.hpp"
94 #include "runtime/timer.hpp"
95 #include "utilities/align.hpp"
96 #include "utilities/copy.hpp"
97 #include "utilities/hashTable.hpp"
98 #include "utilities/macros.hpp"
99
100 // -------------------- Compile::mach_constant_base_node -----------------------
101 // Constant table base node singleton.
102 MachConstantBaseNode* Compile::mach_constant_base_node() {
103 if (_mach_constant_base_node == nullptr) {
104 _mach_constant_base_node = new MachConstantBaseNode();
105 _mach_constant_base_node->add_req(C->root());
106 }
107 return _mach_constant_base_node;
108 }
109
398 record_dead_node(dead->_idx);
399 }
400 if (dead->is_macro()) {
401 remove_macro_node(dead);
402 }
403 if (dead->is_expensive()) {
404 remove_expensive_node(dead);
405 }
406 if (dead->is_ReachabilityFence()) {
407 remove_reachability_fence(dead->as_ReachabilityFence());
408 }
409 if (dead->is_OpaqueTemplateAssertionPredicate()) {
410 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
411 }
412 if (dead->is_ParsePredicate()) {
413 remove_parse_predicate(dead->as_ParsePredicate());
414 }
415 if (dead->for_post_loop_opts_igvn()) {
416 remove_from_post_loop_opts_igvn(dead);
417 }
418 if (dead->is_InlineType()) {
419 remove_inline_type(dead);
420 }
421 if (dead->is_LoadFlat() || dead->is_StoreFlat()) {
422 remove_flat_access(dead);
423 }
424 if (dead->for_merge_stores_igvn()) {
425 remove_from_merge_stores_igvn(dead);
426 }
427 if (dead->is_Call()) {
428 remove_useless_late_inlines( &_late_inlines, dead);
429 remove_useless_late_inlines( &_string_late_inlines, dead);
430 remove_useless_late_inlines( &_boxing_late_inlines, dead);
431 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
432
433 if (dead->is_CallStaticJava()) {
434 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
435 }
436 }
437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
438 bs->unregister_potential_barrier_node(dead);
439 }
440
441 // Disconnect all useless nodes by disconnecting those at the boundary.
442 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
443 uint next = 0;
451 // Use raw traversal of out edges since this code removes out edges
452 int max = n->outcnt();
453 for (int j = 0; j < max; ++j) {
454 Node* child = n->raw_out(j);
455 if (!useful.member(child)) {
456 assert(!child->is_top() || child != top(),
457 "If top is cached in Compile object it is in useful list");
458 // Only need to remove this out-edge to the useless node
459 n->raw_del_out(j);
460 --j;
461 --max;
462 if (child->is_data_proj_of_pure_function(n)) {
463 worklist.push(n);
464 }
465 }
466 }
467 if (n->outcnt() == 1 && n->has_special_unique_user()) {
468 assert(useful.member(n->unique_out()), "do not push a useless node");
469 worklist.push(n->unique_out());
470 }
471 if (n->outcnt() == 0) {
472 worklist.push(n);
473 }
474 }
475
476 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
477 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
478 // Remove useless Template Assertion Predicate opaque nodes
479 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
480 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
481 remove_useless_nodes(_reachability_fences, useful); // remove useless node recorded for post loop opts IGVN pass
482 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
483 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
484 remove_useless_nodes(_flat_access_nodes, useful); // remove useless flat access nodes
485 #ifdef ASSERT
486 if (_modified_nodes != nullptr) {
487 _modified_nodes->remove_useless_nodes(useful.member_set());
488 }
489 #endif
490 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
491 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
492 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
493 #ifdef ASSERT
494 if (_modified_nodes != nullptr) {
495 _modified_nodes->remove_useless_nodes(useful.member_set());
496 }
497 #endif
498
499 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
500 bs->eliminate_useless_gc_barriers(useful, this);
501 // clean up the late inline lists
502 remove_useless_late_inlines( &_late_inlines, useful);
503 remove_useless_late_inlines( &_string_late_inlines, useful);
504 remove_useless_late_inlines( &_boxing_late_inlines, useful);
505 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
506 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
507 }
508
509 // ============================================================================
657
658 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
659 Options options, DirectiveSet* directive)
660 : Phase(Compiler),
661 _compile_id(ci_env->compile_id()),
662 _options(options),
663 _method(target),
664 _entry_bci(osr_bci),
665 _ilt(nullptr),
666 _stub_function(nullptr),
667 _stub_name(nullptr),
668 _stub_id(StubId::NO_STUBID),
669 _stub_entry_point(nullptr),
670 _max_node_limit(MaxNodeLimit),
671 _node_count_inlining_cutoff(NodeCountInliningCutoff),
672 _post_loop_opts_phase(false),
673 _merge_stores_phase(false),
674 _allow_macro_nodes(true),
675 _inlining_progress(false),
676 _inlining_incrementally(false),
677 _strength_reduction(false),
678 _do_cleanup(false),
679 _has_reserved_stack_access(target->has_reserved_stack_access()),
680 _has_circular_inline_type(false),
681 #ifndef PRODUCT
682 _igv_idx(0),
683 _trace_opto_output(directive->TraceOptoOutputOption),
684 #endif
685 _clinit_barrier_on_entry(false),
686 _stress_seed(0),
687 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
688 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
689 _env(ci_env),
690 _directive(directive),
691 _log(ci_env->log()),
692 _first_failure_details(nullptr),
693 _intrinsics(comp_arena(), 0, 0, nullptr),
694 _macro_nodes(comp_arena(), 8, 0, nullptr),
695 _parse_predicates(comp_arena(), 8, 0, nullptr),
696 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
697 _expensive_nodes(comp_arena(), 8, 0, nullptr),
698 _reachability_fences(comp_arena(), 8, 0, nullptr),
699 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
700 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
701 _flat_access_nodes(comp_arena(), 8, 0, nullptr),
702 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
703 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
704 _coarsened_locks(comp_arena(), 8, 0, nullptr),
705 _congraph(nullptr),
706 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
707 _unique(0),
708 _dead_node_count(0),
709 _dead_node_list(comp_arena()),
710 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
711 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
712 _node_arena(&_node_arena_one),
713 _mach_constant_base_node(nullptr),
714 _Compile_types(mtCompiler, Arena::Tag::tag_type),
715 _initial_gvn(nullptr),
716 _igvn_worklist(nullptr),
717 _types(nullptr),
718 _node_hash(nullptr),
719 _late_inlines(comp_arena(), 2, 0, nullptr),
720 _string_late_inlines(comp_arena(), 2, 0, nullptr),
721 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
790 #define MINIMUM_NODE_HASH 1023
791
792 // GVN that will be run immediately on new nodes
793 uint estimated_size = method()->code_size()*4+64;
794 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
795 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
796 _types = new (comp_arena()) Type_Array(comp_arena());
797 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
798 PhaseGVN gvn;
799 set_initial_gvn(&gvn);
800
801 { // Scope for timing the parser
802 TracePhase tp(_t_parser);
803
804 // Put top into the hash table ASAP.
805 initial_gvn()->transform(top());
806
807 // Set up tf(), start(), and find a CallGenerator.
808 CallGenerator* cg = nullptr;
809 if (is_osr_compilation()) {
810 init_tf(TypeFunc::make(method(), false, /* is_osr_compilation = */ true));
811 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
812 initial_gvn()->set_type_bottom(s);
813 verify_start(s);
814 cg = CallGenerator::for_osr(method(), entry_bci());
815 } else {
816 // Normal case.
817 init_tf(TypeFunc::make(method(), false));
818 StartNode* s = new StartNode(root(), tf()->domain_cc());
819 initial_gvn()->set_type_bottom(s);
820 verify_start(s);
821 float past_uses = method()->interpreter_invocation_count();
822 float expected_uses = past_uses;
823 cg = CallGenerator::for_inline(method(), expected_uses);
824 }
825 if (failing()) return;
826 if (cg == nullptr) {
827 const char* reason = InlineTree::check_can_parse(method());
828 assert(reason != nullptr, "expect reason for parse failure");
829 stringStream ss;
830 ss.print("cannot parse method: %s", reason);
831 record_method_not_compilable(ss.as_string());
832 return;
833 }
834
835 gvn.set_type(root(), root()->bottom_type());
836
837 JVMState* jvms = build_start_state(start(), tf());
838 if ((jvms = cg->generate(jvms)) == nullptr) {
896
897 #ifndef PRODUCT
898 if (should_print_ideal()) {
899 print_ideal_ir("PrintIdeal");
900 }
901 #endif
902
903 #ifdef ASSERT
904 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
905 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
906 #endif
907
908 // Dump compilation data to replay it.
909 if (directive->DumpReplayOption) {
910 env()->dump_replay_data(_compile_id);
911 }
912 if (directive->DumpInlineOption && (ilt() != nullptr)) {
913 env()->dump_inline_data(_compile_id);
914 }
915
916 // Now that we know the size of all the monitors we can add fixed slots:
917 // [...]
918 // rsp+80: saved fp register
919 // rsp+76: Fixed slot 7
920 // rsp+72: Fixed slot 6 (stack increment)
921 // rsp+68: Fixed slot 5
922 // rsp+64: Fixed slot 4 (null marker)
923 // rsp+60: Fixed slot 3
924 // rsp+56: Fixed slot 2 (original deopt pc)
925 // rsp+52: Fixed slot 1
926 // rsp+48: Fixed slot 0 (monitors)
927 // rsp+44: spill
928 // [...]
929
930 // One extra slot for the original deopt pc.
931 int next_slot = fixed_slots();
932 next_slot += VMRegImpl::slots_per_word;
933
934 // One extra slot for the special stack increment value.
935 if (needs_stack_repair()) {
936 next_slot += VMRegImpl::slots_per_word;
937 }
938
939 // One extra slot to hold the null marker at scalarized returns.
940 if (needs_nm_slot()) {
941 next_slot += VMRegImpl::slots_per_word;
942 }
943 set_fixed_slots(next_slot);
944
945 // Compute when to use implicit null checks. Used by matching trap based
946 // nodes and NullCheck optimization.
947 set_allowed_deopt_reasons();
948
949 // Now generate code
950 Code_Gen();
951 }
952
953 //------------------------------Compile----------------------------------------
954 // Compile a runtime stub
955 Compile::Compile(ciEnv* ci_env,
956 TypeFunc_generator generator,
957 address stub_function,
958 const char* stub_name,
959 StubId stub_id,
960 int is_fancy_jump,
961 bool pass_tls,
962 bool return_pc,
963 DirectiveSet* directive)
964 : Phase(Compiler),
965 _compile_id(0),
966 _options(Options::for_runtime_stub()),
967 _method(nullptr),
968 _entry_bci(InvocationEntryBci),
969 _stub_function(stub_function),
970 _stub_name(stub_name),
971 _stub_id(stub_id),
972 _stub_entry_point(nullptr),
973 _max_node_limit(MaxNodeLimit),
974 _node_count_inlining_cutoff(NodeCountInliningCutoff),
975 _post_loop_opts_phase(false),
976 _merge_stores_phase(false),
977 _allow_macro_nodes(true),
978 _inlining_progress(false),
979 _inlining_incrementally(false),
980 _has_reserved_stack_access(false),
981 _has_circular_inline_type(false),
982 #ifndef PRODUCT
983 _igv_idx(0),
984 _trace_opto_output(directive->TraceOptoOutputOption),
985 #endif
986 _clinit_barrier_on_entry(false),
987 _stress_seed(0),
988 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
989 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
990 _env(ci_env),
991 _directive(directive),
992 _log(ci_env->log()),
993 _first_failure_details(nullptr),
994 _reachability_fences(comp_arena(), 8, 0, nullptr),
995 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
996 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
997 _congraph(nullptr),
998 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
999 _unique(0),
1000 _dead_node_count(0),
1001 _dead_node_list(comp_arena()),
1117 _fixed_slots = 0;
1118 set_has_split_ifs(false);
1119 set_has_loops(false); // first approximation
1120 set_has_stringbuilder(false);
1121 set_has_boxed_value(false);
1122 _trap_can_recompile = false; // no traps emitted yet
1123 _major_progress = true; // start out assuming good things will happen
1124 set_has_unsafe_access(false);
1125 set_max_vector_size(0);
1126 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1127 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1128 set_decompile_count(0);
1129
1130 #ifndef PRODUCT
1131 _phase_counter = 0;
1132 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1133 #endif
1134
1135 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1136 _loop_opts_cnt = LoopOptsCount;
1137 _has_flat_accesses = false;
1138 _flat_accesses_share_alias = true;
1139 _scalarize_in_safepoints = false;
1140 _needs_nm_slot = false;
1141
1142 set_do_inlining(Inline);
1143 set_max_inline_size(MaxInlineSize);
1144 set_freq_inline_size(FreqInlineSize);
1145 set_do_scheduling(OptoScheduling);
1146
1147 set_do_vector_loop(false);
1148 set_has_monitors(false);
1149 set_has_scoped_access(false);
1150
1151 if (AllowVectorizeOnDemand) {
1152 if (has_method() && _directive->VectorizeOption) {
1153 set_do_vector_loop(true);
1154 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());})
1155 } else if (has_method() && method()->name() != nullptr &&
1156 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1157 set_do_vector_loop(true);
1158 }
1159 }
1160 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1161 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());})
1393 // If this method has already thrown a range-check,
1394 // assume it was because we already tried range smearing
1395 // and it failed.
1396 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1397 return !already_trapped;
1398 }
1399
1400
1401 //------------------------------flatten_alias_type-----------------------------
1402 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1403 assert(do_aliasing(), "Aliasing should be enabled");
1404 int offset = tj->offset();
1405 TypePtr::PTR ptr = tj->ptr();
1406
1407 // Known instance (scalarizable allocation) alias only with itself.
1408 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1409 tj->is_oopptr()->is_known_instance();
1410
1411 // Process weird unsafe references.
1412 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1413 assert(InlineUnsafeOps || StressReflectiveCode || UseAcmpFastPath, "indeterminate pointers come only from unsafe ops");
1414 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1415 tj = TypeOopPtr::BOTTOM;
1416 ptr = tj->ptr();
1417 offset = tj->offset();
1418 }
1419
1420 // Array pointers need some flattening
1421 const TypeAryPtr* ta = tj->isa_aryptr();
1422 if( ta && is_known_inst ) {
1423 if ( offset != Type::OffsetBot &&
1424 offset > arrayOopDesc::length_offset_in_bytes() ) {
1425 offset = Type::OffsetBot; // Flatten constant access into array body only
1426 tj = ta = ta->
1427 remove_speculative()->
1428 cast_to_ptr_type(ptr)->
1429 with_offset(offset);
1430 }
1431 } else if (ta != nullptr) {
1432 // Common slices
1433 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1434 return TypeAryPtr::RANGE;
1435 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1436 return TypeInstPtr::KLASS;
1437 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1438 return TypeInstPtr::MARK;
1439 }
1440
1441 // Remove size and stability
1442 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false, ta->is_flat(), ta->is_not_flat(), ta->is_not_null_free(), ta->is_atomic());
1443 // Remove ptr, const_oop, and offset
1444 if (ta->elem() == Type::BOTTOM) {
1445 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1446 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1447 // incorrect, see JDK-8331133).
1448 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::Offset::bottom);
1449 } else if (ta->elem()->make_oopptr() != nullptr) {
1450 // Object arrays, keep field_offset
1451 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, ta->klass_is_exact(), Type::Offset::bottom, Type::Offset(ta->field_offset()));
1452 } else {
1453 // Primitive arrays
1454 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::Offset::bottom);
1455 }
1456
1457 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1458 // cannot be distinguished by bytecode alone.
1459 if (ta->elem() == TypeInt::BOOL) {
1460 tj = ta = TypeAryPtr::BYTES;
1461 }
1462
1463 // All arrays of references share the same slice
1464 if (!ta->is_flat() && ta->elem()->make_oopptr() != nullptr) {
1465 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false, false, true, true, true);
1466 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::Offset::bottom);
1467 }
1468
1469 if (ta->is_flat()) {
1470 if (_flat_accesses_share_alias) {
1471 // Initially all flattened array accesses share a single slice
1472 tj = ta = TypeAryPtr::INLINES;
1473 } else {
1474 // Flat accesses are always exact
1475 tj = ta = ta->cast_to_exactness(true);
1476 }
1477 }
1478 }
1479
1480 // Oop pointers need some flattening
1481 const TypeInstPtr *to = tj->isa_instptr();
1482 if (to && to != TypeOopPtr::BOTTOM) {
1483 ciInstanceKlass* ik = to->instance_klass();
1484 tj = to = to->cast_to_maybe_flat_in_array(); // flatten to maybe flat in array
1485 if( ptr == TypePtr::Constant ) {
1486 if (ik != ciEnv::current()->Class_klass() ||
1487 offset < ik->layout_helper_size_in_bytes()) {
1488 // No constant oop pointers (such as Strings); they alias with
1489 // unknown strings.
1490 assert(!is_known_inst, "not scalarizable allocation");
1491 tj = to = to->
1492 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1493 remove_speculative()->
1494 cast_to_ptr_type(TypePtr::BotPTR)->
1495 cast_to_exactness(false);
1496 }
1497 } else if( is_known_inst ) {
1498 tj = to; // Keep NotNull and klass_is_exact for instance type
1499 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1500 // During the 2nd round of IterGVN, NotNull castings are removed.
1501 // Make sure the Bottom and NotNull variants alias the same.
1502 // Also, make sure exact and non-exact variants alias the same.
1503 tj = to = to->
1504 remove_speculative()->
1505 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1506 cast_to_ptr_type(TypePtr::BotPTR)->
1507 cast_to_exactness(false);
1508 }
1509 if (to->speculative() != nullptr) {
1510 tj = to = to->remove_speculative();
1511 }
1512 // Canonicalize the holder of this field
1513 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1514 // First handle header references such as a LoadKlassNode, even if the
1515 // object's klass is unloaded at compile time (4965979).
1516 if (!is_known_inst) { // Do it only for non-instance types
1517 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1518 }
1519 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1520 // Static fields are in the space above the normal instance
1521 // fields in the java.lang.Class instance.
1522 if (ik != ciEnv::current()->Class_klass()) {
1523 to = nullptr;
1524 tj = TypeOopPtr::BOTTOM;
1525 offset = tj->offset();
1526 }
1527 } else {
1528 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1529 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1530 assert(tj->offset() == offset, "no change to offset expected");
1531 bool xk = to->klass_is_exact();
1532 int instance_id = to->instance_id();
1533
1534 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1535 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1536 // its interfaces are included.
1537 if (xk && ik->equals(canonical_holder)) {
1538 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1539 TypePtr::MaybeFlat), "exact type should be canonical type");
1540 } else {
1541 assert(xk || !is_known_inst, "Known instance should be exact type");
1542 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1543 TypePtr::MaybeFlat);
1544 }
1545 }
1546 }
1547
1548 // Klass pointers to object array klasses need some flattening
1549 const TypeKlassPtr *tk = tj->isa_klassptr();
1550 if( tk ) {
1551 // If we are referencing a field within a Klass, we need
1552 // to assume the worst case of an Object. Both exact and
1553 // inexact types must flatten to the same alias class so
1554 // use NotNull as the PTR.
1555 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1556 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1557 env()->Object_klass(),
1558 Type::Offset(offset),
1559 TypePtr::MaybeFlat);
1560 }
1561
1562 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1563 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1564 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1565 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset), TypePtr::MaybeFlat);
1566 } else {
1567 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_flat(), tk->is_null_free(), tk->is_atomic(), tk->is_aryklassptr()->is_refined_type());
1568 }
1569 }
1570 // Check for precise loads from the primary supertype array and force them
1571 // to the supertype cache alias index. Check for generic array loads from
1572 // the primary supertype array and also force them to the supertype cache
1573 // alias index. Since the same load can reach both, we need to merge
1574 // these 2 disparate memories into the same alias class. Since the
1575 // primary supertype array is read-only, there's no chance of confusion
1576 // where we bypass an array load and an array store.
1577 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1578 if (offset == Type::OffsetBot ||
1579 (offset >= primary_supers_offset &&
1580 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1581 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1582 offset = in_bytes(Klass::secondary_super_cache_offset());
1583 tj = tk = tk->with_offset(offset);
1584 }
1585 }
1586
1587 // Flatten all Raw pointers together.
1588 if (tj->base() == Type::RawPtr)
1589 tj = TypeRawPtr::BOTTOM;
1679 intptr_t key = (intptr_t) adr_type;
1680 key ^= key >> logAliasCacheSize;
1681 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1682 }
1683
1684
1685 //-----------------------------grow_alias_types--------------------------------
1686 void Compile::grow_alias_types() {
1687 const int old_ats = _max_alias_types; // how many before?
1688 const int new_ats = old_ats; // how many more?
1689 const int grow_ats = old_ats+new_ats; // how many now?
1690 _max_alias_types = grow_ats;
1691 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), _alias_types, old_ats, grow_ats);
1692 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1693 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1694 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1695 }
1696
1697
1698 //--------------------------------find_alias_type------------------------------
1699 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1700 if (!do_aliasing()) {
1701 return alias_type(AliasIdxBot);
1702 }
1703
1704 AliasCacheEntry* ace = nullptr;
1705 if (!uncached) {
1706 ace = probe_alias_cache(adr_type);
1707 if (ace->_adr_type == adr_type) {
1708 return alias_type(ace->_index);
1709 }
1710 }
1711
1712 // Handle special cases.
1713 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1714 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1715
1716 // Do it the slow way.
1717 const TypePtr* flat = flatten_alias_type(adr_type);
1718
1719 #ifdef ASSERT
1720 {
1721 ResourceMark rm;
1722 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1723 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1724 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1725 Type::str(adr_type));
1726 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1727 const TypeOopPtr* foop = flat->is_oopptr();
1728 // Scalarizable allocations have exact klass always.
1729 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1739 if (alias_type(i)->adr_type() == flat) {
1740 idx = i;
1741 break;
1742 }
1743 }
1744
1745 if (idx == AliasIdxTop) {
1746 if (no_create) return nullptr;
1747 // Grow the array if necessary.
1748 if (_num_alias_types == _max_alias_types) grow_alias_types();
1749 // Add a new alias type.
1750 idx = _num_alias_types++;
1751 _alias_types[idx]->Init(idx, flat);
1752 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1753 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1754 if (flat->isa_instptr()) {
1755 if (flat->offset() == java_lang_Class::klass_offset()
1756 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1757 alias_type(idx)->set_rewritable(false);
1758 }
1759 ciField* field = nullptr;
1760 if (flat->isa_aryptr()) {
1761 #ifdef ASSERT
1762 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1763 // (T_BYTE has the weakest alignment and size restrictions...)
1764 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1765 #endif
1766 const Type* elemtype = flat->is_aryptr()->elem();
1767 if (flat->offset() == TypePtr::OffsetBot) {
1768 alias_type(idx)->set_element(elemtype);
1769 }
1770 int field_offset = flat->is_aryptr()->field_offset().get();
1771 if (flat->is_flat() &&
1772 field_offset != Type::OffsetBot) {
1773 ciInlineKlass* vk = elemtype->inline_klass();
1774 field_offset += vk->payload_offset();
1775 field = vk->get_field_by_offset(field_offset, false);
1776 }
1777 }
1778 if (flat->isa_klassptr()) {
1779 if (UseCompactObjectHeaders) {
1780 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1781 alias_type(idx)->set_rewritable(false);
1782 }
1783 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1784 alias_type(idx)->set_rewritable(false);
1785 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1786 alias_type(idx)->set_rewritable(false);
1787 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1788 alias_type(idx)->set_rewritable(false);
1789 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1790 alias_type(idx)->set_rewritable(false);
1791 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1792 alias_type(idx)->set_rewritable(false);
1793 }
1794
1795 if (flat->isa_instklassptr()) {
1796 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1797 alias_type(idx)->set_rewritable(false);
1798 }
1799 }
1800 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1801 // but the base pointer type is not distinctive enough to identify
1802 // references into JavaThread.)
1803
1804 // Check for final fields.
1805 const TypeInstPtr* tinst = flat->isa_instptr();
1806 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1807 if (tinst->const_oop() != nullptr &&
1808 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1809 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1810 // static field
1811 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1812 field = k->get_field_by_offset(tinst->offset(), true);
1813 } else if (tinst->is_inlinetypeptr()) {
1814 // Inline type field
1815 ciInlineKlass* vk = tinst->inline_klass();
1816 field = vk->get_field_by_offset(tinst->offset(), false);
1817 } else {
1818 ciInstanceKlass *k = tinst->instance_klass();
1819 field = k->get_field_by_offset(tinst->offset(), false);
1820 }
1821 }
1822 assert(field == nullptr ||
1823 original_field == nullptr ||
1824 (field->holder() == original_field->holder() &&
1825 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1826 field->is_static() == original_field->is_static()), "wrong field?");
1827 // Set field() and is_rewritable() attributes.
1828 if (field != nullptr) {
1829 alias_type(idx)->set_field(field);
1830 if (flat->isa_aryptr()) {
1831 // Fields of flat arrays are rewritable although they are declared final
1832 assert(flat->is_flat(), "must be a flat array");
1833 alias_type(idx)->set_rewritable(true);
1834 }
1835 }
1836 }
1837
1838 // Fill the cache for next time.
1839 if (!uncached) {
1840 ace->_adr_type = adr_type;
1841 ace->_index = idx;
1842 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1843
1844 // Might as well try to fill the cache for the flattened version, too.
1845 AliasCacheEntry* face = probe_alias_cache(flat);
1846 if (face->_adr_type == nullptr) {
1847 face->_adr_type = flat;
1848 face->_index = idx;
1849 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1850 }
1851 }
1852
1853 return alias_type(idx);
1854 }
1855
1856
1857 Compile::AliasType* Compile::alias_type(ciField* field) {
1858 const TypeOopPtr* t;
1859 if (field->is_static())
1860 t = TypeInstPtr::make(field->holder()->java_mirror());
1861 else
1862 t = TypeOopPtr::make_from_klass_raw(field->holder());
1863 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1864 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1865 return atp;
1866 }
1867
1868
1869 //------------------------------have_alias_type--------------------------------
1870 bool Compile::have_alias_type(const TypePtr* adr_type) {
1952 assert(!C->major_progress(), "not cleared");
1953
1954 if (_for_post_loop_igvn.length() > 0) {
1955 while (_for_post_loop_igvn.length() > 0) {
1956 Node* n = _for_post_loop_igvn.pop();
1957 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1958 igvn._worklist.push(n);
1959 }
1960 igvn.optimize();
1961 if (failing()) return;
1962 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1963 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1964
1965 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1966 if (C->major_progress()) {
1967 C->clear_major_progress(); // ensure that major progress is now clear
1968 }
1969 }
1970 }
1971
1972 void Compile::add_inline_type(Node* n) {
1973 assert(n->is_InlineType(), "unexpected node");
1974 _inline_type_nodes.push(n);
1975 }
1976
1977 void Compile::remove_inline_type(Node* n) {
1978 assert(n->is_InlineType(), "unexpected node");
1979 if (_inline_type_nodes.contains(n)) {
1980 _inline_type_nodes.remove(n);
1981 }
1982 }
1983
1984 // Does the return value keep otherwise useless inline type allocations alive?
1985 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1986 ResourceMark rm;
1987 Unique_Node_List wq;
1988 wq.push(ret_val);
1989 bool some_allocations = false;
1990 for (uint i = 0; i < wq.size(); i++) {
1991 Node* n = wq.at(i);
1992 if (n->outcnt() > 1) {
1993 // Some other use for the allocation
1994 return false;
1995 } else if (n->is_InlineType()) {
1996 wq.push(n->in(1));
1997 } else if (n->is_Phi()) {
1998 for (uint j = 1; j < n->req(); j++) {
1999 wq.push(n->in(j));
2000 }
2001 } else if (n->is_CheckCastPP() &&
2002 n->in(1)->is_Proj() &&
2003 n->in(1)->in(0)->is_Allocate()) {
2004 some_allocations = true;
2005 } else if (n->is_CheckCastPP()) {
2006 wq.push(n->in(1));
2007 }
2008 }
2009 return some_allocations;
2010 }
2011
2012 bool Compile::clear_argument_if_only_used_as_buffer_at_calls(Node* result_cast, PhaseIterGVN& igvn) {
2013 ResourceMark rm;
2014 Unique_Node_List wq;
2015 wq.push(result_cast);
2016 Node_List calls;
2017 for (uint i = 0; i < wq.size(); ++i) {
2018 Node* n = wq.at(i);
2019 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2020 Node* u = n->fast_out(j);
2021 if (u->is_Phi()) {
2022 wq.push(u);
2023 } else if (u->is_InlineType() && u->as_InlineType()->get_oop() == n) {
2024 wq.push(u);
2025 } else if (u->is_CallJava()) {
2026 CallJavaNode* call = u->as_CallJava();
2027 if (call->method() != nullptr && call->method()->mismatch()) {
2028 return false;
2029 }
2030 uint nargs = call->tf()->domain_cc()->cnt();
2031 for (uint k = TypeFunc::Parms; k < nargs; k++) {
2032 Node* in = call->in(k);
2033 if (in == n && (call->method() == nullptr || !call->method()->is_scalarized_buffer_arg(k - TypeFunc::Parms))) {
2034 return false;
2035 }
2036 }
2037 calls.push(call);
2038 } else if (u->Opcode() == Op_EncodeP) {
2039 wq.push(u);
2040 } else if (u->is_AddP()) {
2041 wq.push(u);
2042 } else if (u->is_Store() && u->in(MemNode::Address) == n) {
2043 // storing to the buffer is fine
2044 } else if (u->is_SafePoint()) {
2045 SafePointNode* sfpt = u->as_SafePoint();
2046 int input = u->find_edge(n);
2047 JVMState* jvms = sfpt->jvms();
2048 if (jvms != nullptr) {
2049 if (input < (int)jvms->debug_start()) {
2050 return false;
2051 }
2052 }
2053 } else {
2054 return false;
2055 }
2056 }
2057 }
2058 for (uint i = 0; i < calls.size(); ++i) {
2059 CallJavaNode* call = calls.at(i)->as_CallJava();
2060 uint nargs = call->tf()->domain_cc()->cnt();
2061 for (uint k = TypeFunc::Parms; k < nargs; k++) {
2062 Node* in = call->in(k);
2063 if (wq.member(in)) {
2064 assert(call->method()->is_scalarized_buffer_arg(k - TypeFunc::Parms), "only buffer argument removed here");
2065 igvn.replace_input_of(call, k, igvn.zerocon(T_OBJECT));
2066 }
2067 }
2068 }
2069 return true;
2070 }
2071
2072 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
2073 // Make sure that the return value does not keep an otherwise unused allocation alive
2074 if (tf()->returns_inline_type_as_fields()) {
2075 Node* ret = nullptr;
2076 for (uint i = 1; i < root()->req(); i++) {
2077 Node* in = root()->in(i);
2078 if (in->Opcode() == Op_Return) {
2079 assert(ret == nullptr, "only one return");
2080 ret = in;
2081 }
2082 }
2083 if (ret != nullptr) {
2084 Node* ret_val = ret->in(TypeFunc::Parms);
2085 if (igvn.type(ret_val)->isa_oopptr() &&
2086 return_val_keeps_allocations_alive(ret_val)) {
2087 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2088 assert(ret_val->outcnt() == 0, "should be dead now");
2089 igvn.remove_dead_node(ret_val, PhaseIterGVN::NodeOrigin::Graph);
2090 }
2091 }
2092 }
2093 // if a newly allocated object is a value that's only passed as argument to calls as (possibly null) buffers, then
2094 // clear the call argument inputs so the allocation node can be removed
2095 for (int i = 0; i < C->macro_count(); ++i) {
2096 Node* macro_node = C->macro_node(i);
2097 if (macro_node->Opcode() == Op_Allocate) {
2098 AllocateNode* allocate = macro_node->as_Allocate();
2099 Node* result_cast = allocate->result_cast();
2100 if (result_cast != nullptr) {
2101 const Type* result_type = igvn.type(result_cast);
2102 if (result_type->is_inlinetypeptr()) {
2103 clear_argument_if_only_used_as_buffer_at_calls(result_cast, igvn);
2104 }
2105 }
2106 }
2107 }
2108
2109 if (_inline_type_nodes.length() == 0) {
2110 // keep the graph canonical
2111 igvn.optimize();
2112 return;
2113 }
2114 // Scalarize inline types in safepoint debug info.
2115 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2116 set_scalarize_in_safepoints(true);
2117 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2118 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2119 vt->make_scalar_in_safepoints(&igvn);
2120 igvn.record_for_igvn(vt);
2121 }
2122 if (remove) {
2123 // Remove inline type nodes by replacing them with their oop input
2124 while (_inline_type_nodes.length() > 0) {
2125 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2126 if (vt->outcnt() == 0) {
2127 igvn.remove_dead_node(vt, PhaseIterGVN::NodeOrigin::Graph);
2128 continue;
2129 }
2130 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2131 DEBUG_ONLY(bool must_be_buffered = false);
2132 Node* u = vt->out(i);
2133 // Check if any users are blackholes. If so, rewrite them to use either the
2134 // allocated buffer, or individual components, instead of the inline type node
2135 // that goes away.
2136 if (u->is_Blackhole()) {
2137 BlackholeNode* bh = u->as_Blackhole();
2138
2139 // Unlink the old input
2140 int idx = bh->find_edge(vt);
2141 assert(idx != -1, "The edge should be there");
2142 bh->del_req(idx);
2143 --i;
2144
2145 if (vt->is_allocated(&igvn)) {
2146 // Already has the allocated instance, blackhole that
2147 bh->add_req(vt->get_oop());
2148 } else {
2149 // Not allocated yet, blackhole the components
2150 for (uint c = 0; c < vt->field_count(); c++) {
2151 bh->add_req(vt->field_value(c));
2152 }
2153 }
2154
2155 // Node modified, record for IGVN
2156 igvn.record_for_igvn(bh);
2157 }
2158 #ifdef ASSERT
2159 // Verify that inline type is buffered when replacing by oop
2160 else if (u->is_InlineType()) {
2161 // InlineType uses don't need buffering because they are about to be replaced as well
2162 } else if (u->is_Phi()) {
2163 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2164 } else {
2165 must_be_buffered = true;
2166 }
2167 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2168 vt->dump(0);
2169 u->dump(0);
2170 assert(false, "Should have been buffered");
2171 }
2172 #endif
2173 }
2174 igvn.replace_node(vt, vt->get_oop());
2175 }
2176 }
2177 igvn.optimize();
2178 }
2179
2180 void Compile::add_flat_access(Node* n) {
2181 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2182 assert(!_flat_access_nodes.contains(n), "duplicate insertion");
2183 _flat_access_nodes.push(n);
2184 }
2185
2186 void Compile::remove_flat_access(Node* n) {
2187 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2188 _flat_access_nodes.remove_if_existing(n);
2189 }
2190
2191 void Compile::process_flat_accesses(PhaseIterGVN& igvn) {
2192 assert(igvn._worklist.size() == 0, "should be empty");
2193 igvn.set_delay_transform(true);
2194 for (int i = _flat_access_nodes.length() - 1; i >= 0; i--) {
2195 Node* n = _flat_access_nodes.at(i);
2196 assert(n != nullptr, "unexpected nullptr");
2197 if (n->is_LoadFlat()) {
2198 LoadFlatNode* loadn = n->as_LoadFlat();
2199 // Expending a flat load atomically means that we get a chunk of memory spanning multiple fields
2200 // that we chop with bitwise operations. That is too subtle for some optimizations, especially
2201 // constant folding when fields are constant. If we can get a constant object from which we are
2202 // flat-loading, we can simply replace the loads at compilation-time by the field of the constant
2203 // object.
2204 ciInstance* loaded_from = nullptr;
2205 if (FoldStableValues) {
2206 const TypeOopPtr* base_type = igvn.type(loadn->base())->is_oopptr();
2207 ciObject* oop = base_type->const_oop();
2208 int off = igvn.type(loadn->ptr())->isa_ptr()->offset();
2209
2210 if (oop != nullptr && oop->is_instance()) {
2211 ciInstance* holder = oop->as_instance();
2212 ciKlass* klass = holder->klass();
2213 ciInstanceKlass* iklass = klass->as_instance_klass();
2214 ciField* field = iklass->get_non_flat_field_by_offset(off);
2215
2216 if (field->is_stable()) {
2217 ciConstant fv = holder->field_value(field);
2218 if (is_reference_type(fv.basic_type()) && fv.as_object()->is_instance()) {
2219 // The field value is an object, not null. We can use stability.
2220 loaded_from = fv.as_object()->as_instance();
2221 }
2222 }
2223 } else if (oop != nullptr && oop->is_array() && off != Type::OffsetBot) {
2224 ciArray* array = oop->as_array();
2225 ciConstant elt = array->element_value_by_offset(off);
2226 const TypeAryPtr* aryptr = base_type->is_aryptr();
2227 if (aryptr->is_stable() && aryptr->is_atomic() && is_reference_type(elt.basic_type()) && elt.as_object()->is_instance()) {
2228 loaded_from = elt.as_object()->as_instance();
2229 }
2230 }
2231 }
2232
2233 if (loaded_from != nullptr) {
2234 loadn->expand_constant(igvn, loaded_from);
2235 } else {
2236 loadn->expand_atomic(igvn);
2237 }
2238 } else {
2239 n->as_StoreFlat()->expand_atomic(igvn);
2240 }
2241 }
2242 _flat_access_nodes.clear_and_deallocate();
2243 igvn.set_delay_transform(false);
2244 igvn.optimize();
2245 }
2246
2247 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2248 DEBUG_ONLY(igvn.verify_empty_worklist(nullptr));
2249 if (!_has_flat_accesses) {
2250 return;
2251 }
2252 // Initially, all flat array accesses share the same slice to
2253 // keep dependencies with Object[] array accesses (that could be
2254 // to a flat array) correct. We're done with parsing so we
2255 // now know all flat array accesses in this compile
2256 // unit. Let's move flat array accesses to their own slice,
2257 // one per element field. This should help memory access
2258 // optimizations.
2259 ResourceMark rm;
2260 Unique_Node_List wq;
2261 wq.push(root());
2262
2263 Node_List mergememnodes;
2264 Node_List memnodes;
2265
2266 // Alias index currently shared by all flat memory accesses
2267 int index = get_alias_index(TypeAryPtr::INLINES);
2268
2269 // Find MergeMem nodes and flat array accesses
2270 for (uint i = 0; i < wq.size(); i++) {
2271 Node* n = wq.at(i);
2272 if (n->is_Mem()) {
2273 const TypePtr* adr_type = nullptr;
2274 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2275 if (adr_type == TypeAryPtr::INLINES) {
2276 memnodes.push(n);
2277 }
2278 } else if (n->is_MergeMem()) {
2279 MergeMemNode* mm = n->as_MergeMem();
2280 if (mm->memory_at(index) != mm->base_memory()) {
2281 mergememnodes.push(n);
2282 }
2283 }
2284 for (uint j = 0; j < n->req(); j++) {
2285 Node* m = n->in(j);
2286 if (m != nullptr) {
2287 wq.push(m);
2288 }
2289 }
2290 }
2291
2292 _flat_accesses_share_alias = false;
2293
2294 // We are going to change the slice for the flat array
2295 // accesses so we need to clear the cache entries that refer to
2296 // them.
2297 for (uint i = 0; i < AliasCacheSize; i++) {
2298 AliasCacheEntry* ace = &_alias_cache[i];
2299 if (ace->_adr_type != nullptr &&
2300 ace->_adr_type->is_flat()) {
2301 ace->_adr_type = nullptr;
2302 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2303 }
2304 }
2305
2306 #ifdef ASSERT
2307 for (uint i = 0; i < memnodes.size(); i++) {
2308 Node* m = memnodes.at(i);
2309 const TypePtr* adr_type = m->adr_type();
2310 m->as_Mem()->set_adr_type(adr_type);
2311 }
2312 #endif // ASSERT
2313
2314 int start_alias = num_alias_types(); // Start of new aliases
2315 Node_Stack stack(0);
2316 #ifdef ASSERT
2317 VectorSet seen(Thread::current()->resource_area());
2318 #endif
2319 // Now let's fix the memory graph so each flat array access
2320 // is moved to the right slice. Start from the MergeMem nodes.
2321 uint last = unique();
2322 for (uint i = 0; i < mergememnodes.size(); i++) {
2323 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2324 if (current->outcnt() == 0) {
2325 // This node is killed by a previous iteration
2326 continue;
2327 }
2328
2329 Node* n = current->memory_at(index);
2330 MergeMemNode* mm = nullptr;
2331 do {
2332 // Follow memory edges through memory accesses, phis and
2333 // narrow membars and push nodes on the stack. Once we hit
2334 // bottom memory, we pop element off the stack one at a
2335 // time, in reverse order, and move them to the right slice
2336 // by changing their memory edges.
2337 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() ||
2338 (n->adr_type() == TypeAryPtr::INLINES && !n->is_NarrowMemProj())) {
2339 assert(!seen.test_set(n->_idx), "");
2340 // Uses (a load for instance) will need to be moved to the
2341 // right slice as well and will get a new memory state
2342 // that we don't know yet. The use could also be the
2343 // backedge of a loop. We put a place holder node between
2344 // the memory node and its uses. We replace that place
2345 // holder with the correct memory state once we know it,
2346 // i.e. when nodes are popped off the stack. Using the
2347 // place holder make the logic work in the presence of
2348 // loops.
2349 if (n->outcnt() > 1) {
2350 Node* place_holder = nullptr;
2351 assert(!n->has_out_with(Op_Node), "");
2352 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2353 Node* u = n->out(k);
2354 if (u != current && u->_idx < last) {
2355 bool success = false;
2356 for (uint l = 0; l < u->req(); l++) {
2357 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2358 continue;
2359 }
2360 Node* in = u->in(l);
2361 if (in == n) {
2362 if (place_holder == nullptr) {
2363 place_holder = new Node(1);
2364 place_holder->init_req(0, n);
2365 }
2366 igvn.replace_input_of(u, l, place_holder);
2367 success = true;
2368 }
2369 }
2370 if (success) {
2371 --k;
2372 }
2373 }
2374 }
2375 }
2376 if (n->is_Phi()) {
2377 stack.push(n, 1);
2378 n = n->in(1);
2379 } else if (n->is_Mem()) {
2380 stack.push(n, n->req());
2381 n = n->in(MemNode::Memory);
2382 } else {
2383 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2384 stack.push(n, n->req());
2385 n = n->in(0)->in(TypeFunc::Memory);
2386 }
2387 } else {
2388 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || n->is_NarrowMemProj(), "");
2389 // Build a new MergeMem node to carry the new memory state
2390 // as we build it. IGVN should fold extraneous MergeMem
2391 // nodes.
2392 if (n->is_NarrowMemProj()) {
2393 // We need 1 NarrowMemProj for each slice of this array
2394 InitializeNode* init = n->in(0)->as_Initialize();
2395 AllocateNode* alloc = init->allocation();
2396 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2397 const TypeAryKlassPtr* klass_type = klass_node->bottom_type()->isa_aryklassptr();
2398 assert(klass_type != nullptr, "must be an array");
2399 assert(klass_type->klass_is_exact(), "must be an exact klass");
2400 ciArrayKlass* klass = klass_type->exact_klass()->as_array_klass();
2401 assert(klass->is_flat_array_klass(), "must be a flat array");
2402 ciInlineKlass* elem_klass = klass->element_klass()->as_inline_klass();
2403 const TypeAryPtr* oop_type = klass_type->as_instance_type()->is_aryptr();
2404 assert(oop_type->klass_is_exact(), "must be an exact klass");
2405
2406 Node* base = alloc->in(TypeFunc::Memory);
2407 assert(base->bottom_type() == Type::MEMORY, "the memory input of AllocateNode must be a memory");
2408 assert(base->adr_type() == TypePtr::BOTTOM, "the memory input of AllocateNode must be a bottom memory");
2409 // Must create a MergeMem with base as the base memory, do not clone if base is a
2410 // MergeMem because it may not be processed yet
2411 mm = MergeMemNode::make(nullptr);
2412 mm->set_base_memory(base);
2413 for (int j = 0; j < elem_klass->nof_nonstatic_fields(); j++) {
2414 int field_offset = elem_klass->nonstatic_field_at(j)->offset_in_bytes() - elem_klass->payload_offset();
2415 const TypeAryPtr* field_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(field_offset);
2416 int field_alias_idx = get_alias_index(field_ptr);
2417 assert(field_ptr == get_adr_type(field_alias_idx), "must match");
2418 Node* new_proj = new NarrowMemProjNode(init, field_ptr);
2419 igvn.register_new_node_with_optimizer(new_proj);
2420 mm->set_memory_at(field_alias_idx, new_proj);
2421 }
2422 if (!klass->is_elem_null_free()) {
2423 int nm_offset = elem_klass->null_marker_offset_in_payload();
2424 const TypeAryPtr* nm_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(nm_offset);
2425 int nm_alias_idx = get_alias_index(nm_ptr);
2426 assert(nm_ptr == get_adr_type(nm_alias_idx), "must match");
2427 Node* new_proj = new NarrowMemProjNode(init, nm_ptr);
2428 igvn.register_new_node_with_optimizer(new_proj);
2429 mm->set_memory_at(nm_alias_idx, new_proj);
2430 }
2431
2432 // Replace all uses of the old NarrowMemProj with the correct state
2433 MergeMemNode* new_n = MergeMemNode::make(mm);
2434 igvn.register_new_node_with_optimizer(new_n);
2435 igvn.replace_node(n, new_n);
2436 } else {
2437 // Must create a MergeMem with n as the base memory, do not clone if n is a MergeMem
2438 // because it may not be processed yet
2439 mm = MergeMemNode::make(nullptr);
2440 mm->set_base_memory(n);
2441 }
2442
2443 igvn.register_new_node_with_optimizer(mm);
2444 while (stack.size() > 0) {
2445 Node* m = stack.node();
2446 uint idx = stack.index();
2447 if (m->is_Mem()) {
2448 // Move memory node to its new slice
2449 const TypePtr* adr_type = m->adr_type();
2450 int alias = get_alias_index(adr_type);
2451 Node* prev = mm->memory_at(alias);
2452 igvn.replace_input_of(m, MemNode::Memory, prev);
2453 mm->set_memory_at(alias, m);
2454 } else if (m->is_Phi()) {
2455 // We need as many new phis as there are new aliases
2456 Node* new_phi_in = MergeMemNode::make(mm);
2457 igvn.register_new_node_with_optimizer(new_phi_in);
2458 igvn.replace_input_of(m, idx, new_phi_in);
2459 if (idx == m->req()-1) {
2460 Node* r = m->in(0);
2461 for (int j = start_alias; j < num_alias_types(); j++) {
2462 const TypePtr* adr_type = get_adr_type(j);
2463 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2464 continue;
2465 }
2466 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2467 igvn.register_new_node_with_optimizer(phi);
2468 for (uint k = 1; k < m->req(); k++) {
2469 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2470 }
2471 mm->set_memory_at(j, phi);
2472 }
2473 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2474 igvn.register_new_node_with_optimizer(base_phi);
2475 for (uint k = 1; k < m->req(); k++) {
2476 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2477 }
2478 mm->set_base_memory(base_phi);
2479 }
2480 } else {
2481 // This is a MemBarCPUOrder node from
2482 // Parse::array_load()/Parse::array_store(), in the
2483 // branch that handles flat arrays hidden under
2484 // an Object[] array. We also need one new membar per
2485 // new alias to keep the unknown access that the
2486 // membars protect properly ordered with accesses to
2487 // known flat array.
2488 assert(m->is_Proj(), "projection expected");
2489 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2490 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2491 for (int j = start_alias; j < num_alias_types(); j++) {
2492 const TypePtr* adr_type = get_adr_type(j);
2493 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2494 continue;
2495 }
2496 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2497 igvn.register_new_node_with_optimizer(mb);
2498 Node* mem = mm->memory_at(j);
2499 mb->init_req(TypeFunc::Control, ctrl);
2500 mb->init_req(TypeFunc::Memory, mem);
2501 ctrl = new ProjNode(mb, TypeFunc::Control);
2502 igvn.register_new_node_with_optimizer(ctrl);
2503 mem = new ProjNode(mb, TypeFunc::Memory);
2504 igvn.register_new_node_with_optimizer(mem);
2505 mm->set_memory_at(j, mem);
2506 }
2507 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2508 }
2509 if (idx < m->req()-1) {
2510 idx += 1;
2511 stack.set_index(idx);
2512 n = m->in(idx);
2513 break;
2514 }
2515 // Take care of place holder nodes
2516 if (m->has_out_with(Op_Node)) {
2517 Node* place_holder = m->find_out_with(Op_Node);
2518 if (place_holder != nullptr) {
2519 Node* mm_clone = mm->clone();
2520 igvn.register_new_node_with_optimizer(mm_clone);
2521 Node* hook = new Node(1);
2522 hook->init_req(0, mm);
2523 igvn.replace_node(place_holder, mm_clone);
2524 hook->destruct(&igvn);
2525 }
2526 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2527 }
2528 stack.pop();
2529 }
2530 }
2531 } while(stack.size() > 0);
2532 // Fix the memory state at the MergeMem we started from
2533 igvn.rehash_node_delayed(current);
2534 for (int j = start_alias; j < num_alias_types(); j++) {
2535 const TypePtr* adr_type = get_adr_type(j);
2536 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2537 continue;
2538 }
2539 current->set_memory_at(j, mm);
2540 }
2541 current->set_memory_at(index, current->base_memory());
2542 }
2543 igvn.optimize();
2544
2545 #ifdef ASSERT
2546 wq.clear();
2547 wq.push(root());
2548 for (uint i = 0; i < wq.size(); i++) {
2549 Node* n = wq.at(i);
2550 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2551 for (uint j = 0; j < n->req(); j++) {
2552 Node* m = n->in(j);
2553 if (m != nullptr) {
2554 wq.push(m);
2555 }
2556 }
2557 }
2558 #endif
2559
2560 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2561 }
2562
2563 void Compile::record_for_merge_stores_igvn(Node* n) {
2564 if (!n->for_merge_stores_igvn()) {
2565 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2566 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2567 _for_merge_stores_igvn.append(n);
2568 }
2569 }
2570
2571 void Compile::remove_from_merge_stores_igvn(Node* n) {
2572 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2573 _for_merge_stores_igvn.remove(n);
2574 }
2575
2576 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2577 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2578 // the stores, and we merge the wrong sequence of stores.
2579 // Example:
2580 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2581 // Apply MergeStores:
2582 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2661 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2662 Bytecodes::Code c = iter.cur_bc();
2663 Node* lhs = nullptr;
2664 Node* rhs = nullptr;
2665 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2666 lhs = unc->peek_operand(0);
2667 rhs = unc->peek_operand(1);
2668 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2669 lhs = unc->peek_operand(0);
2670 }
2671
2672 ResourceMark rm;
2673 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2674 assert(live_locals.is_valid(), "broken liveness info");
2675 int len = (int)live_locals.size();
2676
2677 for (int i = 0; i < len; i++) {
2678 Node* local = unc->local(jvms, i);
2679 // kill local using the liveness of next_bci.
2680 // give up when the local looks like an operand to secure reexecution.
2681 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2682 uint idx = jvms->locoff() + i;
2683 #ifdef ASSERT
2684 if (PrintOpto && Verbose) {
2685 tty->print("[unstable_if] kill local#%d: ", idx);
2686 local->dump();
2687 tty->cr();
2688 }
2689 #endif
2690 igvn.replace_input_of(unc, idx, top());
2691 modified = true;
2692 }
2693 }
2694 }
2695
2696 // keep the modified trap for late query
2697 if (modified) {
2698 trap->set_modified();
2699 } else {
2700 _unstable_if_traps.delete_at(i);
2701 }
2702 }
2703 igvn.optimize();
2704 }
2705
2706 // StringOpts and late inlining of string methods
2707 void Compile::inline_string_calls(bool parse_time) {
2708 {
2709 // remove useless nodes to make the usage analysis simpler
2710 ResourceMark rm;
2711 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2712 }
2713
2714 {
2715 ResourceMark rm;
2716 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2908
2909 if (_string_late_inlines.length() > 0) {
2910 assert(has_stringbuilder(), "inconsistent");
2911
2912 inline_string_calls(false);
2913
2914 if (failing()) return;
2915
2916 inline_incrementally_cleanup(igvn);
2917 }
2918
2919 set_inlining_incrementally(false);
2920 }
2921
2922 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2923 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2924 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2925 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2926 // as if "inlining_incrementally() == true" were set.
2927 assert(inlining_incrementally() == false, "not allowed");
2928 set_strength_reduction(true);
2929 #ifdef ASSERT
2930 Unique_Node_List* modified_nodes = _modified_nodes;
2931 _modified_nodes = nullptr;
2932 #endif
2933 assert(_late_inlines.length() > 0, "sanity");
2934
2935 if (StressIncrementalInlining) {
2936 shuffle_late_inlines();
2937 }
2938
2939 while (_late_inlines.length() > 0) {
2940 igvn_worklist()->ensure_empty(); // should be done with igvn
2941
2942 while (inline_incrementally_one()) {
2943 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2944 }
2945 if (failing()) return;
2946
2947 inline_incrementally_cleanup(igvn);
2948 }
2949 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2950 set_strength_reduction(false);
2951 }
2952
2953 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2954 if (_loop_opts_cnt > 0) {
2955 while (major_progress() && (_loop_opts_cnt > 0)) {
2956 TracePhase tp(_t_idealLoop);
2957 PhaseIdealLoop::optimize(igvn, mode);
2958 _loop_opts_cnt--;
2959 if (failing()) return false;
2960 if (major_progress()) {
2961 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2962 }
2963 }
2964 }
2965 return true;
2966 }
2967
2968 // Remove edges from "root" to each SafePoint at a backward branch.
2969 // They were inserted during parsing (see add_safepoint()) to make
2970 // infinite loops without calls or exceptions visible to root, i.e.,
3076 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
3077 }
3078 assert(!has_vbox_nodes(), "sanity");
3079
3080 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
3081 Compile::TracePhase tp(_t_renumberLive);
3082 igvn_worklist()->ensure_empty(); // should be done with igvn
3083 {
3084 ResourceMark rm;
3085 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
3086 }
3087 igvn.reset();
3088 igvn.optimize(true);
3089 if (failing()) return;
3090 }
3091
3092 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
3093 // safepoints
3094 remove_root_to_sfpts_edges(igvn);
3095
3096 // Process inline type nodes now that all inlining is over
3097 process_inline_types(igvn);
3098
3099 adjust_flat_array_access_aliases(igvn);
3100
3101 if (failing()) return;
3102
3103 if (C->macro_count() > 0) {
3104 // Eliminate some macro nodes before EA to reduce analysis pressure
3105 PhaseMacroExpand mexp(igvn);
3106 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3107 if (failing()) {
3108 return;
3109 }
3110 igvn.set_delay_transform(false);
3111 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3112 }
3113
3114 _print_phase_loop_opts = has_loops();
3115 if (_print_phase_loop_opts) {
3116 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
3117 }
3118
3119 // Perform escape analysis
3120 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
3121 if (has_loops()) {
3122 // Cleanup graph (remove dead nodes).
3123 TracePhase tp(_t_idealLoop);
3124 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
3125 if (failing()) {
3126 return;
3127 }
3128 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
3129 if (C->macro_count() > 0) {
3130 // Eliminate some macro nodes before EA to reduce analysis pressure
3131 PhaseMacroExpand mexp(igvn);
3132 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3133 if (failing()) {
3134 return;
3135 }
3136 igvn.set_delay_transform(false);
3137 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3138 }
3139 }
3140
3141 bool progress;
3142 do {
3143 ConnectionGraph::do_analysis(this, &igvn);
3144
3145 if (failing()) return;
3146
3147 int mcount = macro_count(); // Record number of allocations and locks before IGVN
3148
3149 // Optimize out fields loads from scalar replaceable allocations.
3150 igvn.optimize(true);
3151 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
3152
3153 if (failing()) return;
3154
3155 if (congraph() != nullptr && macro_count() > 0) {
3156 TracePhase tp(_t_macroEliminate);
3157 PhaseMacroExpand mexp(igvn);
3158 mexp.eliminate_macro_nodes();
3159 if (failing()) {
3160 return;
3161 }
3162 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3163
3164 igvn.set_delay_transform(false);
3165 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3166 }
3167
3168 ConnectionGraph::verify_ram_nodes(this, root());
3169 if (failing()) return;
3170
3171 progress = do_iterative_escape_analysis() &&
3172 (macro_count() < mcount) &&
3173 ConnectionGraph::has_candidates(this);
3174 // Try again if candidates exist and made progress
3175 // by removing some allocations and/or locks.
3176 } while (progress);
3177 }
3178
3179 process_flat_accesses(igvn);
3180 if (failing()) {
3181 return;
3182 }
3183
3184 // Loop transforms on the ideal graph. Range Check Elimination,
3185 // peeling, unrolling, etc.
3186
3187 // Set loop opts counter
3188 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
3189 {
3190 TracePhase tp(_t_idealLoop);
3191 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
3192 _loop_opts_cnt--;
3193 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
3194 if (failing()) return;
3195 }
3196 // Loop opts pass if partial peeling occurred in previous pass
3197 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
3198 TracePhase tp(_t_idealLoop);
3199 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3200 _loop_opts_cnt--;
3201 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
3202 if (failing()) return;
3203 }
3251
3252 // Once loop optimizations are over, it is safe to get rid of all reachability fence nodes and
3253 // migrate reachability edges to safepoints.
3254 if (OptimizeReachabilityFences && _reachability_fences.length() > 0) {
3255 TracePhase tp1(_t_idealLoop);
3256 TracePhase tp2(_t_reachability);
3257 PhaseIdealLoop::optimize(igvn, PostLoopOptsExpandReachabilityFences);
3258 print_method(PHASE_EXPAND_REACHABILITY_FENCES, 2);
3259 if (failing()) return;
3260 assert(_reachability_fences.length() == 0 || PreserveReachabilityFencesOnConstants, "no RF nodes allowed");
3261 }
3262
3263 process_for_merge_stores_igvn(igvn);
3264
3265 if (failing()) return;
3266
3267 #ifdef ASSERT
3268 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3269 #endif
3270
3271 if (_late_inlines.length() > 0) {
3272 // More opportunities to optimize virtual and MH calls.
3273 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3274 process_late_inline_calls_no_inline(igvn);
3275 if (failing()) {
3276 return;
3277 }
3278 process_inline_types(igvn);
3279 }
3280 assert(_late_inlines.length() == 0, "late inline queue must be drained");
3281
3282 {
3283 TracePhase tp(_t_macroExpand);
3284 PhaseMacroExpand mex(igvn);
3285 // Last attempt to eliminate macro nodes.
3286 mex.eliminate_macro_nodes();
3287 if (failing()) {
3288 return;
3289 }
3290
3291 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3292 // Do not allow new macro nodes once we start to eliminate and expand
3293 C->reset_allow_macro_nodes();
3294 // Last attempt to eliminate macro nodes before expand
3295 mex.eliminate_macro_nodes();
3296 if (failing()) {
3297 return;
3298 }
3299 mex.eliminate_opaque_looplimit_macro_nodes();
3300 if (failing()) {
3301 return;
3302 }
3303 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3304 if (mex.expand_macro_nodes()) {
3305 assert(failing(), "must bail out w/ explicit message");
3306 return;
3307 }
3308 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3309 }
3310
3311 // Process inline type nodes again and remove them. From here
3312 // on we don't need to keep track of field values anymore.
3313 process_inline_types(igvn, /* remove= */ true);
3314
3315 {
3316 TracePhase tp(_t_barrierExpand);
3317 if (bs->expand_barriers(this, igvn)) {
3318 assert(failing(), "must bail out w/ explicit message");
3319 return;
3320 }
3321 print_method(PHASE_BARRIER_EXPANSION, 2);
3322 }
3323
3324 if (C->max_vector_size() > 0) {
3325 C->optimize_logic_cones(igvn);
3326 igvn.optimize();
3327 if (failing()) return;
3328 }
3329
3330 DEBUG_ONLY( _modified_nodes = nullptr; )
3331 DEBUG_ONLY( _late_inlines.clear(); )
3332
3333 assert(igvn._worklist.size() == 0, "not empty");
3334 } // (End scope of igvn; run destructor if necessary for asserts.)
3335
3336 check_no_dead_use();
3337
3338 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3339 // to remove hashes to unlock nodes for modifications.
3340 C->node_hash()->clear();
3341
3342 // A method with only infinite loops has no edges entering loops from root
3343 {
3344 TracePhase tp(_t_graphReshaping);
3345 if (final_graph_reshaping()) {
3346 assert(failing(), "must bail out w/ explicit message");
3347 return;
3348 }
3349 }
3350
3351 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3352 DEBUG_ONLY(set_phase_optimize_finished();)
3353 }
4012 n->subsume_by(sub, this);
4013 }
4014 }
4015
4016 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
4017 switch( nop ) {
4018 case Op_Opaque1: // Remove Opaque Nodes before matching
4019 n->subsume_by(n->in(1), this);
4020 break;
4021 case Op_CallLeafPure: {
4022 // If the pure call is not supported, then lower to a CallLeaf.
4023 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
4024 CallNode* call = n->as_Call();
4025 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
4026 call->_name, TypeRawPtr::BOTTOM);
4027 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
4028 new_call->init_req(TypeFunc::I_O, C->top());
4029 new_call->init_req(TypeFunc::Memory, C->top());
4030 new_call->init_req(TypeFunc::ReturnAdr, C->top());
4031 new_call->init_req(TypeFunc::FramePtr, C->top());
4032 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain_sig()->cnt(); i++) {
4033 new_call->init_req(i, call->in(i));
4034 }
4035 n->subsume_by(new_call, this);
4036 }
4037 break;
4038 }
4039 case Op_CallStaticJava:
4040 case Op_CallJava:
4041 case Op_CallDynamicJava:
4042 frc.inc_java_call_count(); // Count java call site;
4043 case Op_CallRuntime:
4044 case Op_CallLeaf:
4045 case Op_CallLeafVector:
4046 case Op_CallLeafNoFP: {
4047 assert (n->is_Call(), "");
4048 CallNode *call = n->as_Call();
4049 // See if uncommon argument is shared
4050 if (call->is_CallStaticJava() && call->as_CallStaticJava()->_name) {
4051 Node *n = call->in(TypeFunc::Parms);
4052 int nop = n->Opcode();
4059 nop != Op_DecodeNKlass &&
4060 !n->is_Mem() &&
4061 !n->is_Phi()) {
4062 Node *x = n->clone();
4063 call->set_req(TypeFunc::Parms, x);
4064 }
4065 }
4066 break;
4067 }
4068
4069 // Mem nodes need explicit cases to satisfy assert(!n->is_Mem()) in default.
4070 case Op_StoreF:
4071 case Op_LoadF:
4072 case Op_StoreD:
4073 case Op_LoadD:
4074 case Op_LoadD_unaligned:
4075 case Op_StoreB:
4076 case Op_StoreC:
4077 case Op_StoreI:
4078 case Op_StoreL:
4079 case Op_StoreLSpecial:
4080 case Op_CompareAndSwapB:
4081 case Op_CompareAndSwapS:
4082 case Op_CompareAndSwapI:
4083 case Op_CompareAndSwapL:
4084 case Op_CompareAndSwapP:
4085 case Op_CompareAndSwapN:
4086 case Op_WeakCompareAndSwapB:
4087 case Op_WeakCompareAndSwapS:
4088 case Op_WeakCompareAndSwapI:
4089 case Op_WeakCompareAndSwapL:
4090 case Op_WeakCompareAndSwapP:
4091 case Op_WeakCompareAndSwapN:
4092 case Op_CompareAndExchangeB:
4093 case Op_CompareAndExchangeS:
4094 case Op_CompareAndExchangeI:
4095 case Op_CompareAndExchangeL:
4096 case Op_CompareAndExchangeP:
4097 case Op_CompareAndExchangeN:
4098 case Op_GetAndAddS:
4099 case Op_GetAndAddB:
4591 k->subsume_by(m, this);
4592 }
4593 }
4594 }
4595 break;
4596 }
4597 case Op_CmpUL: {
4598 if (!Matcher::has_match_rule(Op_CmpUL)) {
4599 // No support for unsigned long comparisons
4600 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4601 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4602 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4603 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4604 Node* andl = new AndLNode(orl, remove_sign_mask);
4605 Node* cmp = new CmpLNode(andl, n->in(2));
4606 n->subsume_by(cmp, this);
4607 }
4608 break;
4609 }
4610 #ifdef ASSERT
4611 case Op_InlineType: {
4612 n->dump(-1);
4613 assert(false, "inline type node was not removed");
4614 break;
4615 }
4616 case Op_ConNKlass: {
4617 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4618 ciKlass* klass = tp->is_klassptr()->exact_klass();
4619 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4620 break;
4621 }
4622 #endif
4623 default:
4624 assert(!n->is_Call(), "");
4625 assert(!n->is_Mem(), "");
4626 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4627 break;
4628 }
4629 }
4630
4631 //------------------------------final_graph_reshaping_walk---------------------
4632 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4633 // requires that the walk visits a node's inputs before visiting the node.
4634 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4635 Unique_Node_List sfpt;
4968 }
4969 }
4970
4971 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4972 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4973 }
4974
4975 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4976 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4977 }
4978
4979 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4980 if (holder->is_initialized()) {
4981 return false;
4982 }
4983 if (holder->is_being_initialized()) {
4984 if (accessing_method->holder() == holder) {
4985 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4986 // <init>, or a static method. In all those cases, there was an initialization
4987 // barrier on the holder klass passed.
4988 if (accessing_method->is_class_initializer() ||
4989 accessing_method->is_object_constructor() ||
4990 accessing_method->is_static()) {
4991 return false;
4992 }
4993 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4994 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4995 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4996 // child class can become fully initialized while its parent class is still being initialized.
4997 if (accessing_method->is_class_initializer()) {
4998 return false;
4999 }
5000 }
5001 ciMethod* root = method(); // the root method of compilation
5002 if (root != accessing_method) {
5003 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
5004 }
5005 }
5006 return true;
5007 }
5008
5009 #ifndef PRODUCT
5010 //------------------------------verify_bidirectional_edges---------------------
5011 // For each input edge to a node (ie - for each Use-Def edge), verify that
5012 // there is a corresponding Def-Use edge.
5013 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
5014 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
5015 uint stack_size = live_nodes() >> 4;
5016 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
5017 if (root_and_safepoints != nullptr) {
5047 if (in != nullptr && !in->is_top()) {
5048 // Count instances of `next`
5049 int cnt = 0;
5050 for (uint idx = 0; idx < in->_outcnt; idx++) {
5051 if (in->_out[idx] == n) {
5052 cnt++;
5053 }
5054 }
5055 assert(cnt > 0, "Failed to find Def-Use edge.");
5056 // Check for duplicate edges
5057 // walk the input array downcounting the input edges to n
5058 for (uint j = 0; j < length; j++) {
5059 if (n->in(j) == in) {
5060 cnt--;
5061 }
5062 }
5063 assert(cnt == 0, "Mismatched edge count.");
5064 } else if (in == nullptr) {
5065 assert(i == 0 || i >= n->req() ||
5066 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
5067 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
5068 (n->is_Unlock() && i == (n->req() - 1)) ||
5069 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
5070 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
5071 } else {
5072 assert(in->is_top(), "sanity");
5073 // Nothing to check.
5074 }
5075 }
5076 }
5077 }
5078
5079 //------------------------------verify_graph_edges---------------------------
5080 // Walk the Graph and verify that there is a one-to-one correspondence
5081 // between Use-Def edges and Def-Use edges in the graph.
5082 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
5083 if (VerifyGraphEdges) {
5084 Unique_Node_List visited;
5085
5086 // Call graph walk to check edges
5087 verify_bidirectional_edges(visited, root_and_safepoints);
5088 if (no_dead_code) {
5089 // Now make sure that no visited node is used by an unvisited node.
5090 bool dead_nodes = false;
5201 // (1) subklass is already limited to a subtype of superklass => always ok
5202 // (2) subklass does not overlap with superklass => always fail
5203 // (3) superklass has NO subtypes and we can check with a simple compare.
5204 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
5205 if (skip) {
5206 return SSC_full_test; // Let caller generate the general case.
5207 }
5208
5209 if (subk->is_java_subtype_of(superk)) {
5210 return SSC_always_true; // (0) and (1) this test cannot fail
5211 }
5212
5213 if (!subk->maybe_java_subtype_of(superk)) {
5214 return SSC_always_false; // (2) true path dead; no dynamic test needed
5215 }
5216
5217 const Type* superelem = superk;
5218 if (superk->isa_aryklassptr()) {
5219 int ignored;
5220 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5221
5222 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5223 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5224 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5225 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5226 return SSC_full_test;
5227 }
5228 }
5229
5230 if (superelem->isa_instklassptr()) {
5231 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5232 if (!ik->has_subklass()) {
5233 if (!ik->is_final()) {
5234 // Add a dependency if there is a chance of a later subclass.
5235 dependencies()->assert_leaf_type(ik);
5236 }
5237 if (!superk->maybe_java_subtype_of(subk)) {
5238 return SSC_always_false;
5239 }
5240 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5241 }
5242 } else {
5243 // A primitive array type has no subtypes.
5244 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5245 }
5246
5247 return SSC_full_test;
6040 } else {
6041 _debug_network_printer->update_compiled_method(C->method());
6042 }
6043 tty->print_cr("Method printed over network stream to IGV");
6044 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
6045 }
6046 #endif // !PRODUCT
6047
6048 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
6049 if (type != nullptr && phase->type(value)->higher_equal(type)) {
6050 return value;
6051 }
6052 Node* result = nullptr;
6053 if (bt == T_BYTE) {
6054 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
6055 result = new RShiftINode(result, phase->intcon(24));
6056 } else if (bt == T_BOOLEAN) {
6057 result = new AndINode(value, phase->intcon(0xFF));
6058 } else if (bt == T_CHAR) {
6059 result = new AndINode(value,phase->intcon(0xFFFF));
6060 } else if (bt == T_FLOAT) {
6061 result = new MoveI2FNode(value);
6062 } else {
6063 assert(bt == T_SHORT, "unexpected narrow type");
6064 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
6065 result = new RShiftINode(result, phase->intcon(16));
6066 }
6067 if (transform_res) {
6068 result = phase->transform(result);
6069 }
6070 return result;
6071 }
6072
6073 void Compile::record_method_not_compilable_oom() {
6074 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
6075 }
6076
6077 #ifndef PRODUCT
6078 // Collects all the control inputs from nodes on the worklist and from their data dependencies
6079 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
6080 // Follow non-control edges until we reach CFG nodes
6081 for (uint i = 0; i < worklist.size(); i++) {
|