39 #include "gc/shared/barrierSet.hpp"
40 #include "gc/shared/c2/barrierSetC2.hpp"
41 #include "jfr/jfrEvents.hpp"
42 #include "jvm_io.h"
43 #include "memory/allocation.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "opto/addnode.hpp"
46 #include "opto/block.hpp"
47 #include "opto/c2compiler.hpp"
48 #include "opto/callGenerator.hpp"
49 #include "opto/callnode.hpp"
50 #include "opto/castnode.hpp"
51 #include "opto/cfgnode.hpp"
52 #include "opto/chaitin.hpp"
53 #include "opto/compile.hpp"
54 #include "opto/connode.hpp"
55 #include "opto/convertnode.hpp"
56 #include "opto/divnode.hpp"
57 #include "opto/escape.hpp"
58 #include "opto/idealGraphPrinter.hpp"
59 #include "opto/loopnode.hpp"
60 #include "opto/machnode.hpp"
61 #include "opto/macro.hpp"
62 #include "opto/matcher.hpp"
63 #include "opto/mathexactnode.hpp"
64 #include "opto/memnode.hpp"
65 #include "opto/mulnode.hpp"
66 #include "opto/narrowptrnode.hpp"
67 #include "opto/node.hpp"
68 #include "opto/opcodes.hpp"
69 #include "opto/output.hpp"
70 #include "opto/parse.hpp"
71 #include "opto/phaseX.hpp"
72 #include "opto/rootnode.hpp"
73 #include "opto/runtime.hpp"
74 #include "opto/stringopts.hpp"
75 #include "opto/type.hpp"
76 #include "opto/vector.hpp"
77 #include "opto/vectornode.hpp"
78 #include "runtime/globals_extension.hpp"
383 // as dead to be conservative about the dead node count at any
384 // given time.
385 if (!dead->is_Con()) {
386 record_dead_node(dead->_idx);
387 }
388 if (dead->is_macro()) {
389 remove_macro_node(dead);
390 }
391 if (dead->is_expensive()) {
392 remove_expensive_node(dead);
393 }
394 if (dead->Opcode() == Op_Opaque4) {
395 remove_template_assertion_predicate_opaq(dead);
396 }
397 if (dead->is_ParsePredicate()) {
398 remove_parse_predicate(dead->as_ParsePredicate());
399 }
400 if (dead->for_post_loop_opts_igvn()) {
401 remove_from_post_loop_opts_igvn(dead);
402 }
403 if (dead->is_Call()) {
404 remove_useless_late_inlines( &_late_inlines, dead);
405 remove_useless_late_inlines( &_string_late_inlines, dead);
406 remove_useless_late_inlines( &_boxing_late_inlines, dead);
407 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
408
409 if (dead->is_CallStaticJava()) {
410 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
411 }
412 }
413 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
414 bs->unregister_potential_barrier_node(dead);
415 }
416
417 // Disconnect all useless nodes by disconnecting those at the boundary.
418 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
419 uint next = 0;
420 while (next < useful.size()) {
421 Node *n = useful.at(next++);
422 if (n->is_SafePoint()) {
424 // beyond that point.
425 n->as_SafePoint()->delete_replaced_nodes();
426 }
427 // Use raw traversal of out edges since this code removes out edges
428 int max = n->outcnt();
429 for (int j = 0; j < max; ++j) {
430 Node* child = n->raw_out(j);
431 if (!useful.member(child)) {
432 assert(!child->is_top() || child != top(),
433 "If top is cached in Compile object it is in useful list");
434 // Only need to remove this out-edge to the useless node
435 n->raw_del_out(j);
436 --j;
437 --max;
438 }
439 }
440 if (n->outcnt() == 1 && n->has_special_unique_user()) {
441 assert(useful.member(n->unique_out()), "do not push a useless node");
442 worklist.push(n->unique_out());
443 }
444 }
445
446 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
447 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
448 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
449 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
450 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
451 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
452 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
453 #ifdef ASSERT
454 if (_modified_nodes != nullptr) {
455 _modified_nodes->remove_useless_nodes(useful.member_set());
456 }
457 #endif
458
459 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
460 bs->eliminate_useless_gc_barriers(useful, this);
461 // clean up the late inline lists
462 remove_useless_late_inlines( &_late_inlines, useful);
463 remove_useless_late_inlines( &_string_late_inlines, useful);
464 remove_useless_late_inlines( &_boxing_late_inlines, useful);
465 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
466 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
467 }
468
469 // ============================================================================
470 //------------------------------CompileWrapper---------------------------------
609 // the continuation bci for on stack replacement.
610
611
612 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
613 Options options, DirectiveSet* directive)
614 : Phase(Compiler),
615 _compile_id(ci_env->compile_id()),
616 _options(options),
617 _method(target),
618 _entry_bci(osr_bci),
619 _ilt(nullptr),
620 _stub_function(nullptr),
621 _stub_name(nullptr),
622 _stub_entry_point(nullptr),
623 _max_node_limit(MaxNodeLimit),
624 _post_loop_opts_phase(false),
625 _inlining_progress(false),
626 _inlining_incrementally(false),
627 _do_cleanup(false),
628 _has_reserved_stack_access(target->has_reserved_stack_access()),
629 #ifndef PRODUCT
630 _igv_idx(0),
631 _trace_opto_output(directive->TraceOptoOutputOption),
632 #endif
633 _has_method_handle_invokes(false),
634 _clinit_barrier_on_entry(false),
635 _stress_seed(0),
636 _comp_arena(mtCompiler),
637 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
638 _env(ci_env),
639 _directive(directive),
640 _log(ci_env->log()),
641 _failure_reason(nullptr),
642 _first_failure_details(nullptr),
643 _intrinsics (comp_arena(), 0, 0, nullptr),
644 _macro_nodes (comp_arena(), 8, 0, nullptr),
645 _parse_predicates (comp_arena(), 8, 0, nullptr),
646 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
647 _expensive_nodes (comp_arena(), 8, 0, nullptr),
648 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
649 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
650 _coarsened_locks (comp_arena(), 8, 0, nullptr),
651 _congraph(nullptr),
652 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
653 _unique(0),
654 _dead_node_count(0),
655 _dead_node_list(comp_arena()),
656 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
657 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
658 _node_arena(&_node_arena_one),
659 _mach_constant_base_node(nullptr),
660 _Compile_types(mtCompiler),
661 _initial_gvn(nullptr),
662 _igvn_worklist(nullptr),
663 _types(nullptr),
664 _node_hash(nullptr),
665 _late_inlines(comp_arena(), 2, 0, nullptr),
666 _string_late_inlines(comp_arena(), 2, 0, nullptr),
667 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
668 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
731
732 // GVN that will be run immediately on new nodes
733 uint estimated_size = method()->code_size()*4+64;
734 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
735 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
736 _types = new (comp_arena()) Type_Array(comp_arena());
737 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
738 PhaseGVN gvn;
739 set_initial_gvn(&gvn);
740
741 print_inlining_init();
742 { // Scope for timing the parser
743 TracePhase tp("parse", &timers[_t_parser]);
744
745 // Put top into the hash table ASAP.
746 initial_gvn()->transform(top());
747
748 // Set up tf(), start(), and find a CallGenerator.
749 CallGenerator* cg = nullptr;
750 if (is_osr_compilation()) {
751 const TypeTuple *domain = StartOSRNode::osr_domain();
752 const TypeTuple *range = TypeTuple::make_range(method()->signature());
753 init_tf(TypeFunc::make(domain, range));
754 StartNode* s = new StartOSRNode(root(), domain);
755 initial_gvn()->set_type_bottom(s);
756 init_start(s);
757 cg = CallGenerator::for_osr(method(), entry_bci());
758 } else {
759 // Normal case.
760 init_tf(TypeFunc::make(method()));
761 StartNode* s = new StartNode(root(), tf()->domain());
762 initial_gvn()->set_type_bottom(s);
763 init_start(s);
764 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
765 // With java.lang.ref.reference.get() we must go through the
766 // intrinsic - even when get() is the root
767 // method of the compile - so that, if necessary, the value in
768 // the referent field of the reference object gets recorded by
769 // the pre-barrier code.
770 cg = find_intrinsic(method(), false);
771 }
772 if (cg == nullptr) {
773 float past_uses = method()->interpreter_invocation_count();
774 float expected_uses = past_uses;
775 cg = CallGenerator::for_inline(method(), expected_uses);
776 }
777 }
778 if (failing()) return;
779 if (cg == nullptr) {
780 const char* reason = InlineTree::check_can_parse(method());
781 assert(reason != nullptr, "expect reason for parse failure");
867 print_ideal_ir("print_ideal");
868 }
869 #endif
870
871 #ifdef ASSERT
872 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
873 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
874 #endif
875
876 // Dump compilation data to replay it.
877 if (directive->DumpReplayOption) {
878 env()->dump_replay_data(_compile_id);
879 }
880 if (directive->DumpInlineOption && (ilt() != nullptr)) {
881 env()->dump_inline_data(_compile_id);
882 }
883
884 // Now that we know the size of all the monitors we can add a fixed slot
885 // for the original deopt pc.
886 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
887 set_fixed_slots(next_slot);
888
889 // Compute when to use implicit null checks. Used by matching trap based
890 // nodes and NullCheck optimization.
891 set_allowed_deopt_reasons();
892
893 // Now generate code
894 Code_Gen();
895 }
896
897 //------------------------------Compile----------------------------------------
898 // Compile a runtime stub
899 Compile::Compile( ciEnv* ci_env,
900 TypeFunc_generator generator,
901 address stub_function,
902 const char *stub_name,
903 int is_fancy_jump,
904 bool pass_tls,
905 bool return_pc,
906 DirectiveSet* directive)
907 : Phase(Compiler),
908 _compile_id(0),
909 _options(Options::for_runtime_stub()),
910 _method(nullptr),
911 _entry_bci(InvocationEntryBci),
912 _stub_function(stub_function),
913 _stub_name(stub_name),
914 _stub_entry_point(nullptr),
915 _max_node_limit(MaxNodeLimit),
916 _post_loop_opts_phase(false),
917 _inlining_progress(false),
918 _inlining_incrementally(false),
919 _has_reserved_stack_access(false),
920 #ifndef PRODUCT
921 _igv_idx(0),
922 _trace_opto_output(directive->TraceOptoOutputOption),
923 #endif
924 _has_method_handle_invokes(false),
925 _clinit_barrier_on_entry(false),
926 _stress_seed(0),
927 _comp_arena(mtCompiler),
928 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
929 _env(ci_env),
930 _directive(directive),
931 _log(ci_env->log()),
932 _failure_reason(nullptr),
933 _first_failure_details(nullptr),
934 _congraph(nullptr),
935 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
936 _unique(0),
937 _dead_node_count(0),
938 _dead_node_list(comp_arena()),
939 _node_arena_one(mtCompiler),
1040
1041 _fixed_slots = 0;
1042 set_has_split_ifs(false);
1043 set_has_loops(false); // first approximation
1044 set_has_stringbuilder(false);
1045 set_has_boxed_value(false);
1046 _trap_can_recompile = false; // no traps emitted yet
1047 _major_progress = true; // start out assuming good things will happen
1048 set_has_unsafe_access(false);
1049 set_max_vector_size(0);
1050 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1051 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1052 set_decompile_count(0);
1053
1054 #ifndef PRODUCT
1055 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1056 #endif
1057
1058 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1059 _loop_opts_cnt = LoopOptsCount;
1060 set_do_inlining(Inline);
1061 set_max_inline_size(MaxInlineSize);
1062 set_freq_inline_size(FreqInlineSize);
1063 set_do_scheduling(OptoScheduling);
1064
1065 set_do_vector_loop(false);
1066 set_has_monitors(false);
1067
1068 if (AllowVectorizeOnDemand) {
1069 if (has_method() && _directive->VectorizeOption) {
1070 set_do_vector_loop(true);
1071 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());})
1072 } else if (has_method() && method()->name() != 0 &&
1073 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1074 set_do_vector_loop(true);
1075 }
1076 }
1077 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1078 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());})
1079
1328 // If this method has already thrown a range-check,
1329 // assume it was because we already tried range smearing
1330 // and it failed.
1331 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1332 return !already_trapped;
1333 }
1334
1335
1336 //------------------------------flatten_alias_type-----------------------------
1337 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1338 assert(do_aliasing(), "Aliasing should be enabled");
1339 int offset = tj->offset();
1340 TypePtr::PTR ptr = tj->ptr();
1341
1342 // Known instance (scalarizable allocation) alias only with itself.
1343 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1344 tj->is_oopptr()->is_known_instance();
1345
1346 // Process weird unsafe references.
1347 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1348 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1349 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1350 tj = TypeOopPtr::BOTTOM;
1351 ptr = tj->ptr();
1352 offset = tj->offset();
1353 }
1354
1355 // Array pointers need some flattening
1356 const TypeAryPtr* ta = tj->isa_aryptr();
1357 if (ta && ta->is_stable()) {
1358 // Erase stability property for alias analysis.
1359 tj = ta = ta->cast_to_stable(false);
1360 }
1361 if( ta && is_known_inst ) {
1362 if ( offset != Type::OffsetBot &&
1363 offset > arrayOopDesc::length_offset_in_bytes() ) {
1364 offset = Type::OffsetBot; // Flatten constant access into array body only
1365 tj = ta = ta->
1366 remove_speculative()->
1367 cast_to_ptr_type(ptr)->
1368 with_offset(offset);
1369 }
1370 } else if (ta) {
1371 // For arrays indexed by constant indices, we flatten the alias
1372 // space to include all of the array body. Only the header, klass
1373 // and array length can be accessed un-aliased.
1374 if( offset != Type::OffsetBot ) {
1375 if( ta->const_oop() ) { // MethodData* or Method*
1376 offset = Type::OffsetBot; // Flatten constant access into array body
1377 tj = ta = ta->
1378 remove_speculative()->
1379 cast_to_ptr_type(ptr)->
1380 cast_to_exactness(false)->
1381 with_offset(offset);
1382 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1383 // range is OK as-is.
1384 tj = ta = TypeAryPtr::RANGE;
1385 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1386 tj = TypeInstPtr::KLASS; // all klass loads look alike
1387 ta = TypeAryPtr::RANGE; // generic ignored junk
1388 ptr = TypePtr::BotPTR;
1389 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1390 tj = TypeInstPtr::MARK;
1391 ta = TypeAryPtr::RANGE; // generic ignored junk
1392 ptr = TypePtr::BotPTR;
1393 } else { // Random constant offset into array body
1394 offset = Type::OffsetBot; // Flatten constant access into array body
1395 tj = ta = ta->
1396 remove_speculative()->
1397 cast_to_ptr_type(ptr)->
1398 cast_to_exactness(false)->
1399 with_offset(offset);
1400 }
1401 }
1402 // Arrays of fixed size alias with arrays of unknown size.
1403 if (ta->size() != TypeInt::POS) {
1404 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1405 tj = ta = ta->
1406 remove_speculative()->
1407 cast_to_ptr_type(ptr)->
1408 with_ary(tary)->
1409 cast_to_exactness(false);
1410 }
1411 // Arrays of known objects become arrays of unknown objects.
1412 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1413 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1414 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1415 }
1416 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1417 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1419 }
1420 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1421 // cannot be distinguished by bytecode alone.
1422 if (ta->elem() == TypeInt::BOOL) {
1423 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1424 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1425 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1426 }
1427 // During the 2nd round of IterGVN, NotNull castings are removed.
1428 // Make sure the Bottom and NotNull variants alias the same.
1429 // Also, make sure exact and non-exact variants alias the same.
1430 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1431 tj = ta = ta->
1432 remove_speculative()->
1433 cast_to_ptr_type(TypePtr::BotPTR)->
1434 cast_to_exactness(false)->
1435 with_offset(offset);
1436 }
1437 }
1438
1439 // Oop pointers need some flattening
1440 const TypeInstPtr *to = tj->isa_instptr();
1441 if (to && to != TypeOopPtr::BOTTOM) {
1442 ciInstanceKlass* ik = to->instance_klass();
1443 if( ptr == TypePtr::Constant ) {
1444 if (ik != ciEnv::current()->Class_klass() ||
1445 offset < ik->layout_helper_size_in_bytes()) {
1455 } else if( is_known_inst ) {
1456 tj = to; // Keep NotNull and klass_is_exact for instance type
1457 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1458 // During the 2nd round of IterGVN, NotNull castings are removed.
1459 // Make sure the Bottom and NotNull variants alias the same.
1460 // Also, make sure exact and non-exact variants alias the same.
1461 tj = to = to->
1462 remove_speculative()->
1463 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1464 cast_to_ptr_type(TypePtr::BotPTR)->
1465 cast_to_exactness(false);
1466 }
1467 if (to->speculative() != nullptr) {
1468 tj = to = to->remove_speculative();
1469 }
1470 // Canonicalize the holder of this field
1471 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1472 // First handle header references such as a LoadKlassNode, even if the
1473 // object's klass is unloaded at compile time (4965979).
1474 if (!is_known_inst) { // Do it only for non-instance types
1475 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1476 }
1477 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1478 // Static fields are in the space above the normal instance
1479 // fields in the java.lang.Class instance.
1480 if (ik != ciEnv::current()->Class_klass()) {
1481 to = nullptr;
1482 tj = TypeOopPtr::BOTTOM;
1483 offset = tj->offset();
1484 }
1485 } else {
1486 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1487 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1488 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1489 if( is_known_inst ) {
1490 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1491 } else {
1492 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1493 }
1494 }
1495 }
1496 }
1497
1498 // Klass pointers to object array klasses need some flattening
1499 const TypeKlassPtr *tk = tj->isa_klassptr();
1500 if( tk ) {
1501 // If we are referencing a field within a Klass, we need
1502 // to assume the worst case of an Object. Both exact and
1503 // inexact types must flatten to the same alias class so
1504 // use NotNull as the PTR.
1505 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1506 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1507 env()->Object_klass(),
1508 offset);
1509 }
1510
1511 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1512 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1513 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1514 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1515 } else {
1516 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1517 }
1518 }
1519
1520 // Check for precise loads from the primary supertype array and force them
1521 // to the supertype cache alias index. Check for generic array loads from
1522 // the primary supertype array and also force them to the supertype cache
1523 // alias index. Since the same load can reach both, we need to merge
1524 // these 2 disparate memories into the same alias class. Since the
1525 // primary supertype array is read-only, there's no chance of confusion
1526 // where we bypass an array load and an array store.
1527 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1528 if (offset == Type::OffsetBot ||
1529 (offset >= primary_supers_offset &&
1530 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1531 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1532 offset = in_bytes(Klass::secondary_super_cache_offset());
1533 tj = tk = tk->with_offset(offset);
1534 }
1535 }
1536
1537 // Flatten all Raw pointers together.
1538 if (tj->base() == Type::RawPtr)
1539 tj = TypeRawPtr::BOTTOM;
1629 intptr_t key = (intptr_t) adr_type;
1630 key ^= key >> logAliasCacheSize;
1631 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1632 }
1633
1634
1635 //-----------------------------grow_alias_types--------------------------------
1636 void Compile::grow_alias_types() {
1637 const int old_ats = _max_alias_types; // how many before?
1638 const int new_ats = old_ats; // how many more?
1639 const int grow_ats = old_ats+new_ats; // how many now?
1640 _max_alias_types = grow_ats;
1641 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1642 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1643 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1644 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1645 }
1646
1647
1648 //--------------------------------find_alias_type------------------------------
1649 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1650 if (!do_aliasing()) {
1651 return alias_type(AliasIdxBot);
1652 }
1653
1654 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1655 if (ace->_adr_type == adr_type) {
1656 return alias_type(ace->_index);
1657 }
1658
1659 // Handle special cases.
1660 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1661 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1662
1663 // Do it the slow way.
1664 const TypePtr* flat = flatten_alias_type(adr_type);
1665
1666 #ifdef ASSERT
1667 {
1668 ResourceMark rm;
1669 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1670 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1671 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1672 Type::str(adr_type));
1673 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1674 const TypeOopPtr* foop = flat->is_oopptr();
1675 // Scalarizable allocations have exact klass always.
1676 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1686 if (alias_type(i)->adr_type() == flat) {
1687 idx = i;
1688 break;
1689 }
1690 }
1691
1692 if (idx == AliasIdxTop) {
1693 if (no_create) return nullptr;
1694 // Grow the array if necessary.
1695 if (_num_alias_types == _max_alias_types) grow_alias_types();
1696 // Add a new alias type.
1697 idx = _num_alias_types++;
1698 _alias_types[idx]->Init(idx, flat);
1699 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1700 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1701 if (flat->isa_instptr()) {
1702 if (flat->offset() == java_lang_Class::klass_offset()
1703 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1704 alias_type(idx)->set_rewritable(false);
1705 }
1706 if (flat->isa_aryptr()) {
1707 #ifdef ASSERT
1708 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1709 // (T_BYTE has the weakest alignment and size restrictions...)
1710 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1711 #endif
1712 if (flat->offset() == TypePtr::OffsetBot) {
1713 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1714 }
1715 }
1716 if (flat->isa_klassptr()) {
1717 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1718 alias_type(idx)->set_rewritable(false);
1719 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1720 alias_type(idx)->set_rewritable(false);
1721 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1722 alias_type(idx)->set_rewritable(false);
1723 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1724 alias_type(idx)->set_rewritable(false);
1725 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1726 alias_type(idx)->set_rewritable(false);
1727 }
1728 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1729 // but the base pointer type is not distinctive enough to identify
1730 // references into JavaThread.)
1731
1732 // Check for final fields.
1733 const TypeInstPtr* tinst = flat->isa_instptr();
1734 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1735 ciField* field;
1736 if (tinst->const_oop() != nullptr &&
1737 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1738 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1739 // static field
1740 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1741 field = k->get_field_by_offset(tinst->offset(), true);
1742 } else {
1743 ciInstanceKlass *k = tinst->instance_klass();
1744 field = k->get_field_by_offset(tinst->offset(), false);
1745 }
1746 assert(field == nullptr ||
1747 original_field == nullptr ||
1748 (field->holder() == original_field->holder() &&
1749 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1750 field->is_static() == original_field->is_static()), "wrong field?");
1751 // Set field() and is_rewritable() attributes.
1752 if (field != nullptr) alias_type(idx)->set_field(field);
1753 }
1754 }
1755
1756 // Fill the cache for next time.
1757 ace->_adr_type = adr_type;
1758 ace->_index = idx;
1759 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1760
1761 // Might as well try to fill the cache for the flattened version, too.
1762 AliasCacheEntry* face = probe_alias_cache(flat);
1763 if (face->_adr_type == nullptr) {
1764 face->_adr_type = flat;
1765 face->_index = idx;
1766 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1767 }
1768
1769 return alias_type(idx);
1770 }
1771
1772
1773 Compile::AliasType* Compile::alias_type(ciField* field) {
1774 const TypeOopPtr* t;
1775 if (field->is_static())
1776 t = TypeInstPtr::make(field->holder()->java_mirror());
1777 else
1778 t = TypeOopPtr::make_from_klass_raw(field->holder());
1779 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1780 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1781 return atp;
1782 }
1783
1784
1785 //------------------------------have_alias_type--------------------------------
1786 bool Compile::have_alias_type(const TypePtr* adr_type) {
1866 assert(!C->major_progress(), "not cleared");
1867
1868 if (_for_post_loop_igvn.length() > 0) {
1869 while (_for_post_loop_igvn.length() > 0) {
1870 Node* n = _for_post_loop_igvn.pop();
1871 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1872 igvn._worklist.push(n);
1873 }
1874 igvn.optimize();
1875 if (failing()) return;
1876 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1877 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1878
1879 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1880 if (C->major_progress()) {
1881 C->clear_major_progress(); // ensure that major progress is now clear
1882 }
1883 }
1884 }
1885
1886 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1887 if (OptimizeUnstableIf) {
1888 _unstable_if_traps.append(trap);
1889 }
1890 }
1891
1892 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1893 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1894 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1895 Node* n = trap->uncommon_trap();
1896 if (!useful.member(n)) {
1897 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1898 }
1899 }
1900 }
1901
1902 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1903 // or fold-compares case. Return true if succeed or not found.
1904 //
1905 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1941 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1942 Bytecodes::Code c = iter.cur_bc();
1943 Node* lhs = nullptr;
1944 Node* rhs = nullptr;
1945 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1946 lhs = unc->peek_operand(0);
1947 rhs = unc->peek_operand(1);
1948 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1949 lhs = unc->peek_operand(0);
1950 }
1951
1952 ResourceMark rm;
1953 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1954 assert(live_locals.is_valid(), "broken liveness info");
1955 int len = (int)live_locals.size();
1956
1957 for (int i = 0; i < len; i++) {
1958 Node* local = unc->local(jvms, i);
1959 // kill local using the liveness of next_bci.
1960 // give up when the local looks like an operand to secure reexecution.
1961 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1962 uint idx = jvms->locoff() + i;
1963 #ifdef ASSERT
1964 if (PrintOpto && Verbose) {
1965 tty->print("[unstable_if] kill local#%d: ", idx);
1966 local->dump();
1967 tty->cr();
1968 }
1969 #endif
1970 igvn.replace_input_of(unc, idx, top());
1971 modified = true;
1972 }
1973 }
1974 }
1975
1976 // keep the mondified trap for late query
1977 if (modified) {
1978 trap->set_modified();
1979 } else {
1980 _unstable_if_traps.delete_at(i);
1981 }
1982 }
1983 igvn.optimize();
1984 }
1985
1986 // StringOpts and late inlining of string methods
1987 void Compile::inline_string_calls(bool parse_time) {
1988 {
1989 // remove useless nodes to make the usage analysis simpler
1990 ResourceMark rm;
1991 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1992 }
1993
1994 {
1995 ResourceMark rm;
1996 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2151
2152 if (_string_late_inlines.length() > 0) {
2153 assert(has_stringbuilder(), "inconsistent");
2154
2155 inline_string_calls(false);
2156
2157 if (failing()) return;
2158
2159 inline_incrementally_cleanup(igvn);
2160 }
2161
2162 set_inlining_incrementally(false);
2163 }
2164
2165 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2166 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2167 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2168 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2169 // as if "inlining_incrementally() == true" were set.
2170 assert(inlining_incrementally() == false, "not allowed");
2171 assert(_modified_nodes == nullptr, "not allowed");
2172 assert(_late_inlines.length() > 0, "sanity");
2173
2174 while (_late_inlines.length() > 0) {
2175 igvn_worklist()->ensure_empty(); // should be done with igvn
2176
2177 while (inline_incrementally_one()) {
2178 assert(!failing(), "inconsistent");
2179 }
2180 if (failing()) return;
2181
2182 inline_incrementally_cleanup(igvn);
2183 }
2184 }
2185
2186 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2187 if (_loop_opts_cnt > 0) {
2188 while (major_progress() && (_loop_opts_cnt > 0)) {
2189 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2190 PhaseIdealLoop::optimize(igvn, mode);
2191 _loop_opts_cnt--;
2192 if (failing()) return false;
2193 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2194 }
2195 }
2196 return true;
2197 }
2198
2199 // Remove edges from "root" to each SafePoint at a backward branch.
2200 // They were inserted during parsing (see add_safepoint()) to make
2201 // infinite loops without calls or exceptions visible to root, i.e.,
2202 // useful.
2203 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2310 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2311 }
2312 assert(!has_vbox_nodes(), "sanity");
2313
2314 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2315 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2316 igvn_worklist()->ensure_empty(); // should be done with igvn
2317 {
2318 ResourceMark rm;
2319 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2320 }
2321 igvn.reset_from_gvn(initial_gvn());
2322 igvn.optimize();
2323 if (failing()) return;
2324 }
2325
2326 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2327 // safepoints
2328 remove_root_to_sfpts_edges(igvn);
2329
2330 if (failing()) return;
2331
2332 // Perform escape analysis
2333 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2334 if (has_loops()) {
2335 // Cleanup graph (remove dead nodes).
2336 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2337 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2338 if (failing()) return;
2339 }
2340 bool progress;
2341 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2342 do {
2343 ConnectionGraph::do_analysis(this, &igvn);
2344
2345 if (failing()) return;
2346
2347 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2348
2349 // Optimize out fields loads from scalar replaceable allocations.
2433 if (failing()) return;
2434
2435 // Loop transforms on the ideal graph. Range Check Elimination,
2436 // peeling, unrolling, etc.
2437 if (!optimize_loops(igvn, LoopOptsDefault)) {
2438 return;
2439 }
2440
2441 if (failing()) return;
2442
2443 C->clear_major_progress(); // ensure that major progress is now clear
2444
2445 process_for_post_loop_opts_igvn(igvn);
2446
2447 if (failing()) return;
2448
2449 #ifdef ASSERT
2450 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2451 #endif
2452
2453 {
2454 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2455 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2456 PhaseMacroExpand mex(igvn);
2457 if (mex.expand_macro_nodes()) {
2458 assert(failing(), "must bail out w/ explicit message");
2459 return;
2460 }
2461 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2462 }
2463
2464 {
2465 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2466 if (bs->expand_barriers(this, igvn)) {
2467 assert(failing(), "must bail out w/ explicit message");
2468 return;
2469 }
2470 print_method(PHASE_BARRIER_EXPANSION, 2);
2471 }
2472
2473 if (C->max_vector_size() > 0) {
2474 C->optimize_logic_cones(igvn);
2475 igvn.optimize();
2476 if (failing()) return;
2477 }
2478
2479 DEBUG_ONLY( _modified_nodes = nullptr; )
2480
2481 assert(igvn._worklist.size() == 0, "not empty");
2482
2483 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2484
2485 if (_late_inlines.length() > 0) {
2486 // More opportunities to optimize virtual and MH calls.
2487 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2488 process_late_inline_calls_no_inline(igvn);
2489 if (failing()) return;
2490 }
2491 } // (End scope of igvn; run destructor if necessary for asserts.)
2492
2493 check_no_dead_use();
2494
2495 process_print_inlining();
2496
2497 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2498 // to remove hashes to unlock nodes for modifications.
2499 C->node_hash()->clear();
2500
2501 // A method with only infinite loops has no edges entering loops from root
2502 {
2503 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2504 if (final_graph_reshaping()) {
2505 assert(failing(), "must bail out w/ explicit message");
2506 return;
2507 }
2508 }
2509
2510 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3098 // Accumulate any precedence edges
3099 if (mem->in(i) != nullptr) {
3100 n->add_prec(mem->in(i));
3101 }
3102 }
3103 // Everything above this point has been processed.
3104 done = true;
3105 }
3106 // Eliminate the previous StoreCM
3107 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3108 assert(mem->outcnt() == 0, "should be dead");
3109 mem->disconnect_inputs(this);
3110 } else {
3111 prev = mem;
3112 }
3113 mem = prev->in(MemNode::Memory);
3114 }
3115 }
3116 }
3117
3118 //------------------------------final_graph_reshaping_impl----------------------
3119 // Implement items 1-5 from final_graph_reshaping below.
3120 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3121
3122 if ( n->outcnt() == 0 ) return; // dead node
3123 uint nop = n->Opcode();
3124
3125 // Check for 2-input instruction with "last use" on right input.
3126 // Swap to left input. Implements item (2).
3127 if( n->req() == 3 && // two-input instruction
3128 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3129 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3130 n->in(2)->outcnt() == 1 &&// right use IS a last use
3131 !n->in(2)->is_Con() ) { // right use is not a constant
3132 // Check for commutative opcode
3133 switch( nop ) {
3134 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3135 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3136 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3137 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3250 if (n->outcnt() > 1 &&
3251 !n->is_Proj() &&
3252 nop != Op_CreateEx &&
3253 nop != Op_CheckCastPP &&
3254 nop != Op_DecodeN &&
3255 nop != Op_DecodeNKlass &&
3256 !n->is_Mem() &&
3257 !n->is_Phi()) {
3258 Node *x = n->clone();
3259 call->set_req(TypeFunc::Parms, x);
3260 }
3261 }
3262 break;
3263 }
3264
3265 case Op_StoreCM:
3266 {
3267 // Convert OopStore dependence into precedence edge
3268 Node* prec = n->in(MemNode::OopStore);
3269 n->del_req(MemNode::OopStore);
3270 n->add_prec(prec);
3271 eliminate_redundant_card_marks(n);
3272 }
3273
3274 // fall through
3275
3276 case Op_StoreB:
3277 case Op_StoreC:
3278 case Op_StoreI:
3279 case Op_StoreL:
3280 case Op_CompareAndSwapB:
3281 case Op_CompareAndSwapS:
3282 case Op_CompareAndSwapI:
3283 case Op_CompareAndSwapL:
3284 case Op_CompareAndSwapP:
3285 case Op_CompareAndSwapN:
3286 case Op_WeakCompareAndSwapB:
3287 case Op_WeakCompareAndSwapS:
3288 case Op_WeakCompareAndSwapI:
3289 case Op_WeakCompareAndSwapL:
3290 case Op_WeakCompareAndSwapP:
3871 // Replace all nodes with identical edges as m with m
3872 k->subsume_by(m, this);
3873 }
3874 }
3875 }
3876 break;
3877 }
3878 case Op_CmpUL: {
3879 if (!Matcher::has_match_rule(Op_CmpUL)) {
3880 // No support for unsigned long comparisons
3881 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3882 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3883 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3884 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3885 Node* andl = new AndLNode(orl, remove_sign_mask);
3886 Node* cmp = new CmpLNode(andl, n->in(2));
3887 n->subsume_by(cmp, this);
3888 }
3889 break;
3890 }
3891 default:
3892 assert(!n->is_Call(), "");
3893 assert(!n->is_Mem(), "");
3894 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3895 break;
3896 }
3897 }
3898
3899 //------------------------------final_graph_reshaping_walk---------------------
3900 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3901 // requires that the walk visits a node's inputs before visiting the node.
3902 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3903 Unique_Node_List sfpt;
3904
3905 frc._visited.set(root->_idx); // first, mark node as visited
3906 uint cnt = root->req();
3907 Node *n = root;
3908 uint i = 0;
3909 while (true) {
3910 if (i < cnt) {
4250 }
4251 }
4252
4253 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4254 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4255 }
4256
4257 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4258 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4259 }
4260
4261 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4262 if (holder->is_initialized()) {
4263 return false;
4264 }
4265 if (holder->is_being_initialized()) {
4266 if (accessing_method->holder() == holder) {
4267 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4268 // <init>, or a static method. In all those cases, there was an initialization
4269 // barrier on the holder klass passed.
4270 if (accessing_method->is_static_initializer() ||
4271 accessing_method->is_object_initializer() ||
4272 accessing_method->is_static()) {
4273 return false;
4274 }
4275 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4276 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4277 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4278 // child class can become fully initialized while its parent class is still being initialized.
4279 if (accessing_method->is_static_initializer()) {
4280 return false;
4281 }
4282 }
4283 ciMethod* root = method(); // the root method of compilation
4284 if (root != accessing_method) {
4285 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4286 }
4287 }
4288 return true;
4289 }
4290
4291 #ifndef PRODUCT
4292 //------------------------------verify_bidirectional_edges---------------------
4293 // For each input edge to a node (ie - for each Use-Def edge), verify that
4294 // there is a corresponding Def-Use edge.
4295 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4296 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4297 uint stack_size = live_nodes() >> 4;
4298 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4299 nstack.push(_root);
4315 if (in != nullptr && !in->is_top()) {
4316 // Count instances of `next`
4317 int cnt = 0;
4318 for (uint idx = 0; idx < in->_outcnt; idx++) {
4319 if (in->_out[idx] == n) {
4320 cnt++;
4321 }
4322 }
4323 assert(cnt > 0, "Failed to find Def-Use edge.");
4324 // Check for duplicate edges
4325 // walk the input array downcounting the input edges to n
4326 for (uint j = 0; j < length; j++) {
4327 if (n->in(j) == in) {
4328 cnt--;
4329 }
4330 }
4331 assert(cnt == 0, "Mismatched edge count.");
4332 } else if (in == nullptr) {
4333 assert(i == 0 || i >= n->req() ||
4334 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4335 (n->is_Unlock() && i == (n->req() - 1)) ||
4336 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4337 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4338 } else {
4339 assert(in->is_top(), "sanity");
4340 // Nothing to check.
4341 }
4342 }
4343 }
4344 }
4345
4346 //------------------------------verify_graph_edges---------------------------
4347 // Walk the Graph and verify that there is a one-to-one correspondence
4348 // between Use-Def edges and Def-Use edges in the graph.
4349 void Compile::verify_graph_edges(bool no_dead_code) {
4350 if (VerifyGraphEdges) {
4351 Unique_Node_List visited;
4352
4353 // Call graph walk to check edges
4354 verify_bidirectional_edges(visited);
4355 if (no_dead_code) {
4356 // Now make sure that no visited node is used by an unvisited node.
4357 bool dead_nodes = false;
4447 // (1) subklass is already limited to a subtype of superklass => always ok
4448 // (2) subklass does not overlap with superklass => always fail
4449 // (3) superklass has NO subtypes and we can check with a simple compare.
4450 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4451 if (skip) {
4452 return SSC_full_test; // Let caller generate the general case.
4453 }
4454
4455 if (subk->is_java_subtype_of(superk)) {
4456 return SSC_always_true; // (0) and (1) this test cannot fail
4457 }
4458
4459 if (!subk->maybe_java_subtype_of(superk)) {
4460 return SSC_always_false; // (2) true path dead; no dynamic test needed
4461 }
4462
4463 const Type* superelem = superk;
4464 if (superk->isa_aryklassptr()) {
4465 int ignored;
4466 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4467 }
4468
4469 if (superelem->isa_instklassptr()) {
4470 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4471 if (!ik->has_subklass()) {
4472 if (!ik->is_final()) {
4473 // Add a dependency if there is a chance of a later subclass.
4474 dependencies()->assert_leaf_type(ik);
4475 }
4476 if (!superk->maybe_java_subtype_of(subk)) {
4477 return SSC_always_false;
4478 }
4479 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4480 }
4481 } else {
4482 // A primitive array type has no subtypes.
4483 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4484 }
4485
4486 return SSC_full_test;
5007 const Type* t = igvn.type_or_null(n);
5008 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5009 if (n->is_Type()) {
5010 t = n->as_Type()->type();
5011 assert(t == t->remove_speculative(), "no more speculative types");
5012 }
5013 // Iterate over outs - endless loops is unreachable from below
5014 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5015 Node *m = n->fast_out(i);
5016 if (not_a_node(m)) {
5017 continue;
5018 }
5019 worklist.push(m);
5020 }
5021 }
5022 igvn.check_no_speculative_types();
5023 #endif
5024 }
5025 }
5026
5027 // Auxiliary methods to support randomized stressing/fuzzing.
5028
5029 int Compile::random() {
5030 _stress_seed = os::next_random(_stress_seed);
5031 return static_cast<int>(_stress_seed);
5032 }
5033
5034 // This method can be called the arbitrary number of times, with current count
5035 // as the argument. The logic allows selecting a single candidate from the
5036 // running list of candidates as follows:
5037 // int count = 0;
5038 // Cand* selected = null;
5039 // while(cand = cand->next()) {
5040 // if (randomized_select(++count)) {
5041 // selected = cand;
5042 // }
5043 // }
5044 //
5045 // Including count equalizes the chances any candidate is "selected".
5046 // This is useful when we don't have the complete list of candidates to choose
|
39 #include "gc/shared/barrierSet.hpp"
40 #include "gc/shared/c2/barrierSetC2.hpp"
41 #include "jfr/jfrEvents.hpp"
42 #include "jvm_io.h"
43 #include "memory/allocation.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "opto/addnode.hpp"
46 #include "opto/block.hpp"
47 #include "opto/c2compiler.hpp"
48 #include "opto/callGenerator.hpp"
49 #include "opto/callnode.hpp"
50 #include "opto/castnode.hpp"
51 #include "opto/cfgnode.hpp"
52 #include "opto/chaitin.hpp"
53 #include "opto/compile.hpp"
54 #include "opto/connode.hpp"
55 #include "opto/convertnode.hpp"
56 #include "opto/divnode.hpp"
57 #include "opto/escape.hpp"
58 #include "opto/idealGraphPrinter.hpp"
59 #include "opto/inlinetypenode.hpp"
60 #include "opto/loopnode.hpp"
61 #include "opto/machnode.hpp"
62 #include "opto/macro.hpp"
63 #include "opto/matcher.hpp"
64 #include "opto/mathexactnode.hpp"
65 #include "opto/memnode.hpp"
66 #include "opto/mulnode.hpp"
67 #include "opto/narrowptrnode.hpp"
68 #include "opto/node.hpp"
69 #include "opto/opcodes.hpp"
70 #include "opto/output.hpp"
71 #include "opto/parse.hpp"
72 #include "opto/phaseX.hpp"
73 #include "opto/rootnode.hpp"
74 #include "opto/runtime.hpp"
75 #include "opto/stringopts.hpp"
76 #include "opto/type.hpp"
77 #include "opto/vector.hpp"
78 #include "opto/vectornode.hpp"
79 #include "runtime/globals_extension.hpp"
384 // as dead to be conservative about the dead node count at any
385 // given time.
386 if (!dead->is_Con()) {
387 record_dead_node(dead->_idx);
388 }
389 if (dead->is_macro()) {
390 remove_macro_node(dead);
391 }
392 if (dead->is_expensive()) {
393 remove_expensive_node(dead);
394 }
395 if (dead->Opcode() == Op_Opaque4) {
396 remove_template_assertion_predicate_opaq(dead);
397 }
398 if (dead->is_ParsePredicate()) {
399 remove_parse_predicate(dead->as_ParsePredicate());
400 }
401 if (dead->for_post_loop_opts_igvn()) {
402 remove_from_post_loop_opts_igvn(dead);
403 }
404 if (dead->is_InlineType()) {
405 remove_inline_type(dead);
406 }
407 if (dead->is_Call()) {
408 remove_useless_late_inlines( &_late_inlines, dead);
409 remove_useless_late_inlines( &_string_late_inlines, dead);
410 remove_useless_late_inlines( &_boxing_late_inlines, dead);
411 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
412
413 if (dead->is_CallStaticJava()) {
414 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
415 }
416 }
417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
418 bs->unregister_potential_barrier_node(dead);
419 }
420
421 // Disconnect all useless nodes by disconnecting those at the boundary.
422 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
423 uint next = 0;
424 while (next < useful.size()) {
425 Node *n = useful.at(next++);
426 if (n->is_SafePoint()) {
428 // beyond that point.
429 n->as_SafePoint()->delete_replaced_nodes();
430 }
431 // Use raw traversal of out edges since this code removes out edges
432 int max = n->outcnt();
433 for (int j = 0; j < max; ++j) {
434 Node* child = n->raw_out(j);
435 if (!useful.member(child)) {
436 assert(!child->is_top() || child != top(),
437 "If top is cached in Compile object it is in useful list");
438 // Only need to remove this out-edge to the useless node
439 n->raw_del_out(j);
440 --j;
441 --max;
442 }
443 }
444 if (n->outcnt() == 1 && n->has_special_unique_user()) {
445 assert(useful.member(n->unique_out()), "do not push a useless node");
446 worklist.push(n->unique_out());
447 }
448 if (n->outcnt() == 0) {
449 worklist.push(n);
450 }
451 }
452
453 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
454 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
455 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
456 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
457 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
458 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
459 #ifdef ASSERT
460 if (_modified_nodes != nullptr) {
461 _modified_nodes->remove_useless_nodes(useful.member_set());
462 }
463 #endif
464 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
465 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
466 #ifdef ASSERT
467 if (_modified_nodes != nullptr) {
468 _modified_nodes->remove_useless_nodes(useful.member_set());
469 }
470 #endif
471
472 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
473 bs->eliminate_useless_gc_barriers(useful, this);
474 // clean up the late inline lists
475 remove_useless_late_inlines( &_late_inlines, useful);
476 remove_useless_late_inlines( &_string_late_inlines, useful);
477 remove_useless_late_inlines( &_boxing_late_inlines, useful);
478 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
479 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
480 }
481
482 // ============================================================================
483 //------------------------------CompileWrapper---------------------------------
622 // the continuation bci for on stack replacement.
623
624
625 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
626 Options options, DirectiveSet* directive)
627 : Phase(Compiler),
628 _compile_id(ci_env->compile_id()),
629 _options(options),
630 _method(target),
631 _entry_bci(osr_bci),
632 _ilt(nullptr),
633 _stub_function(nullptr),
634 _stub_name(nullptr),
635 _stub_entry_point(nullptr),
636 _max_node_limit(MaxNodeLimit),
637 _post_loop_opts_phase(false),
638 _inlining_progress(false),
639 _inlining_incrementally(false),
640 _do_cleanup(false),
641 _has_reserved_stack_access(target->has_reserved_stack_access()),
642 _has_circular_inline_type(false),
643 #ifndef PRODUCT
644 _igv_idx(0),
645 _trace_opto_output(directive->TraceOptoOutputOption),
646 #endif
647 _has_method_handle_invokes(false),
648 _clinit_barrier_on_entry(false),
649 _stress_seed(0),
650 _comp_arena(mtCompiler),
651 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
652 _env(ci_env),
653 _directive(directive),
654 _log(ci_env->log()),
655 _failure_reason(nullptr),
656 _first_failure_details(nullptr),
657 _intrinsics (comp_arena(), 0, 0, nullptr),
658 _macro_nodes (comp_arena(), 8, 0, nullptr),
659 _parse_predicates (comp_arena(), 8, 0, nullptr),
660 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
661 _expensive_nodes (comp_arena(), 8, 0, nullptr),
662 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
663 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
664 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
665 _coarsened_locks (comp_arena(), 8, 0, nullptr),
666 _congraph(nullptr),
667 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
668 _unique(0),
669 _dead_node_count(0),
670 _dead_node_list(comp_arena()),
671 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
672 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
673 _node_arena(&_node_arena_one),
674 _mach_constant_base_node(nullptr),
675 _Compile_types(mtCompiler),
676 _initial_gvn(nullptr),
677 _igvn_worklist(nullptr),
678 _types(nullptr),
679 _node_hash(nullptr),
680 _late_inlines(comp_arena(), 2, 0, nullptr),
681 _string_late_inlines(comp_arena(), 2, 0, nullptr),
682 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
683 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
746
747 // GVN that will be run immediately on new nodes
748 uint estimated_size = method()->code_size()*4+64;
749 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
750 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
751 _types = new (comp_arena()) Type_Array(comp_arena());
752 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
753 PhaseGVN gvn;
754 set_initial_gvn(&gvn);
755
756 print_inlining_init();
757 { // Scope for timing the parser
758 TracePhase tp("parse", &timers[_t_parser]);
759
760 // Put top into the hash table ASAP.
761 initial_gvn()->transform(top());
762
763 // Set up tf(), start(), and find a CallGenerator.
764 CallGenerator* cg = nullptr;
765 if (is_osr_compilation()) {
766 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
767 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
768 initial_gvn()->set_type_bottom(s);
769 init_start(s);
770 cg = CallGenerator::for_osr(method(), entry_bci());
771 } else {
772 // Normal case.
773 init_tf(TypeFunc::make(method()));
774 StartNode* s = new StartNode(root(), tf()->domain_cc());
775 initial_gvn()->set_type_bottom(s);
776 init_start(s);
777 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
778 // With java.lang.ref.reference.get() we must go through the
779 // intrinsic - even when get() is the root
780 // method of the compile - so that, if necessary, the value in
781 // the referent field of the reference object gets recorded by
782 // the pre-barrier code.
783 cg = find_intrinsic(method(), false);
784 }
785 if (cg == nullptr) {
786 float past_uses = method()->interpreter_invocation_count();
787 float expected_uses = past_uses;
788 cg = CallGenerator::for_inline(method(), expected_uses);
789 }
790 }
791 if (failing()) return;
792 if (cg == nullptr) {
793 const char* reason = InlineTree::check_can_parse(method());
794 assert(reason != nullptr, "expect reason for parse failure");
880 print_ideal_ir("print_ideal");
881 }
882 #endif
883
884 #ifdef ASSERT
885 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
886 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
887 #endif
888
889 // Dump compilation data to replay it.
890 if (directive->DumpReplayOption) {
891 env()->dump_replay_data(_compile_id);
892 }
893 if (directive->DumpInlineOption && (ilt() != nullptr)) {
894 env()->dump_inline_data(_compile_id);
895 }
896
897 // Now that we know the size of all the monitors we can add a fixed slot
898 // for the original deopt pc.
899 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
900 if (needs_stack_repair()) {
901 // One extra slot for the special stack increment value
902 next_slot += 2;
903 }
904 // TODO 8284443 Only reserve extra slot if needed
905 if (InlineTypeReturnedAsFields) {
906 // One extra slot to hold the IsInit information for a nullable
907 // inline type return if we run out of registers.
908 next_slot += 2;
909 }
910 set_fixed_slots(next_slot);
911
912 // Compute when to use implicit null checks. Used by matching trap based
913 // nodes and NullCheck optimization.
914 set_allowed_deopt_reasons();
915
916 // Now generate code
917 Code_Gen();
918 }
919
920 //------------------------------Compile----------------------------------------
921 // Compile a runtime stub
922 Compile::Compile( ciEnv* ci_env,
923 TypeFunc_generator generator,
924 address stub_function,
925 const char *stub_name,
926 int is_fancy_jump,
927 bool pass_tls,
928 bool return_pc,
929 DirectiveSet* directive)
930 : Phase(Compiler),
931 _compile_id(0),
932 _options(Options::for_runtime_stub()),
933 _method(nullptr),
934 _entry_bci(InvocationEntryBci),
935 _stub_function(stub_function),
936 _stub_name(stub_name),
937 _stub_entry_point(nullptr),
938 _max_node_limit(MaxNodeLimit),
939 _post_loop_opts_phase(false),
940 _inlining_progress(false),
941 _inlining_incrementally(false),
942 _has_reserved_stack_access(false),
943 _has_circular_inline_type(false),
944 #ifndef PRODUCT
945 _igv_idx(0),
946 _trace_opto_output(directive->TraceOptoOutputOption),
947 #endif
948 _has_method_handle_invokes(false),
949 _clinit_barrier_on_entry(false),
950 _stress_seed(0),
951 _comp_arena(mtCompiler),
952 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
953 _env(ci_env),
954 _directive(directive),
955 _log(ci_env->log()),
956 _failure_reason(nullptr),
957 _first_failure_details(nullptr),
958 _congraph(nullptr),
959 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
960 _unique(0),
961 _dead_node_count(0),
962 _dead_node_list(comp_arena()),
963 _node_arena_one(mtCompiler),
1064
1065 _fixed_slots = 0;
1066 set_has_split_ifs(false);
1067 set_has_loops(false); // first approximation
1068 set_has_stringbuilder(false);
1069 set_has_boxed_value(false);
1070 _trap_can_recompile = false; // no traps emitted yet
1071 _major_progress = true; // start out assuming good things will happen
1072 set_has_unsafe_access(false);
1073 set_max_vector_size(0);
1074 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1075 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1076 set_decompile_count(0);
1077
1078 #ifndef PRODUCT
1079 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1080 #endif
1081
1082 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1083 _loop_opts_cnt = LoopOptsCount;
1084 _has_flat_accesses = false;
1085 _flat_accesses_share_alias = true;
1086 _scalarize_in_safepoints = false;
1087
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
1096 if (AllowVectorizeOnDemand) {
1097 if (has_method() && _directive->VectorizeOption) {
1098 set_do_vector_loop(true);
1099 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());})
1100 } else if (has_method() && method()->name() != 0 &&
1101 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1102 set_do_vector_loop(true);
1103 }
1104 }
1105 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1106 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());})
1107
1356 // If this method has already thrown a range-check,
1357 // assume it was because we already tried range smearing
1358 // and it failed.
1359 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1360 return !already_trapped;
1361 }
1362
1363
1364 //------------------------------flatten_alias_type-----------------------------
1365 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1366 assert(do_aliasing(), "Aliasing should be enabled");
1367 int offset = tj->offset();
1368 TypePtr::PTR ptr = tj->ptr();
1369
1370 // Known instance (scalarizable allocation) alias only with itself.
1371 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1372 tj->is_oopptr()->is_known_instance();
1373
1374 // Process weird unsafe references.
1375 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1376 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1377 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1378 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1379 tj = TypeOopPtr::BOTTOM;
1380 ptr = tj->ptr();
1381 offset = tj->offset();
1382 }
1383
1384 // Array pointers need some flattening
1385 const TypeAryPtr* ta = tj->isa_aryptr();
1386 if (ta && ta->is_stable()) {
1387 // Erase stability property for alias analysis.
1388 tj = ta = ta->cast_to_stable(false);
1389 }
1390 if (ta && ta->is_not_flat()) {
1391 // Erase not flat property for alias analysis.
1392 tj = ta = ta->cast_to_not_flat(false);
1393 }
1394 if (ta && ta->is_not_null_free()) {
1395 // Erase not null free property for alias analysis.
1396 tj = ta = ta->cast_to_not_null_free(false);
1397 }
1398
1399 if( ta && is_known_inst ) {
1400 if ( offset != Type::OffsetBot &&
1401 offset > arrayOopDesc::length_offset_in_bytes() ) {
1402 offset = Type::OffsetBot; // Flatten constant access into array body only
1403 tj = ta = ta->
1404 remove_speculative()->
1405 cast_to_ptr_type(ptr)->
1406 with_offset(offset);
1407 }
1408 } else if (ta) {
1409 // For arrays indexed by constant indices, we flatten the alias
1410 // space to include all of the array body. Only the header, klass
1411 // and array length can be accessed un-aliased.
1412 // For flat inline type array, each field has its own slice so
1413 // we must include the field offset.
1414 if( offset != Type::OffsetBot ) {
1415 if( ta->const_oop() ) { // MethodData* or Method*
1416 offset = Type::OffsetBot; // Flatten constant access into array body
1417 tj = ta = ta->
1418 remove_speculative()->
1419 cast_to_ptr_type(ptr)->
1420 cast_to_exactness(false)->
1421 with_offset(offset);
1422 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1423 // range is OK as-is.
1424 tj = ta = TypeAryPtr::RANGE;
1425 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1426 tj = TypeInstPtr::KLASS; // all klass loads look alike
1427 ta = TypeAryPtr::RANGE; // generic ignored junk
1428 ptr = TypePtr::BotPTR;
1429 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1430 tj = TypeInstPtr::MARK;
1431 ta = TypeAryPtr::RANGE; // generic ignored junk
1432 ptr = TypePtr::BotPTR;
1433 } else { // Random constant offset into array body
1434 offset = Type::OffsetBot; // Flatten constant access into array body
1435 tj = ta = ta->
1436 remove_speculative()->
1437 cast_to_ptr_type(ptr)->
1438 cast_to_exactness(false)->
1439 with_offset(offset);
1440 }
1441 }
1442 // Arrays of fixed size alias with arrays of unknown size.
1443 if (ta->size() != TypeInt::POS) {
1444 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1445 tj = ta = ta->
1446 remove_speculative()->
1447 cast_to_ptr_type(ptr)->
1448 with_ary(tary)->
1449 cast_to_exactness(false);
1450 }
1451 // Arrays of known objects become arrays of unknown objects.
1452 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1453 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1455 }
1456 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1457 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1458 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1459 }
1460 // Initially all flattened array accesses share a single slice
1461 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1462 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1463 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1464 }
1465 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1466 // cannot be distinguished by bytecode alone.
1467 if (ta->elem() == TypeInt::BOOL) {
1468 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1469 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1470 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1471 }
1472 // During the 2nd round of IterGVN, NotNull castings are removed.
1473 // Make sure the Bottom and NotNull variants alias the same.
1474 // Also, make sure exact and non-exact variants alias the same.
1475 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1476 tj = ta = ta->
1477 remove_speculative()->
1478 cast_to_ptr_type(TypePtr::BotPTR)->
1479 cast_to_exactness(false)->
1480 with_offset(offset);
1481 }
1482 }
1483
1484 // Oop pointers need some flattening
1485 const TypeInstPtr *to = tj->isa_instptr();
1486 if (to && to != TypeOopPtr::BOTTOM) {
1487 ciInstanceKlass* ik = to->instance_klass();
1488 if( ptr == TypePtr::Constant ) {
1489 if (ik != ciEnv::current()->Class_klass() ||
1490 offset < ik->layout_helper_size_in_bytes()) {
1500 } else if( is_known_inst ) {
1501 tj = to; // Keep NotNull and klass_is_exact for instance type
1502 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1503 // During the 2nd round of IterGVN, NotNull castings are removed.
1504 // Make sure the Bottom and NotNull variants alias the same.
1505 // Also, make sure exact and non-exact variants alias the same.
1506 tj = to = to->
1507 remove_speculative()->
1508 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1509 cast_to_ptr_type(TypePtr::BotPTR)->
1510 cast_to_exactness(false);
1511 }
1512 if (to->speculative() != nullptr) {
1513 tj = to = to->remove_speculative();
1514 }
1515 // Canonicalize the holder of this field
1516 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1517 // First handle header references such as a LoadKlassNode, even if the
1518 // object's klass is unloaded at compile time (4965979).
1519 if (!is_known_inst) { // Do it only for non-instance types
1520 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1521 }
1522 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1523 // Static fields are in the space above the normal instance
1524 // fields in the java.lang.Class instance.
1525 if (ik != ciEnv::current()->Class_klass()) {
1526 to = nullptr;
1527 tj = TypeOopPtr::BOTTOM;
1528 offset = tj->offset();
1529 }
1530 } else {
1531 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1532 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1533 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1534 if( is_known_inst ) {
1535 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, Type::Offset(offset), to->instance_id());
1536 } else {
1537 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, Type::Offset(offset));
1538 }
1539 }
1540 }
1541 }
1542
1543 // Klass pointers to object array klasses need some flattening
1544 const TypeKlassPtr *tk = tj->isa_klassptr();
1545 if( tk ) {
1546 // If we are referencing a field within a Klass, we need
1547 // to assume the worst case of an Object. Both exact and
1548 // inexact types must flatten to the same alias class so
1549 // use NotNull as the PTR.
1550 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1551 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1552 env()->Object_klass(),
1553 Type::Offset(offset));
1554 }
1555
1556 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1557 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1558 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1559 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1560 } else {
1561 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_null_free());
1562 }
1563 }
1564 // Check for precise loads from the primary supertype array and force them
1565 // to the supertype cache alias index. Check for generic array loads from
1566 // the primary supertype array and also force them to the supertype cache
1567 // alias index. Since the same load can reach both, we need to merge
1568 // these 2 disparate memories into the same alias class. Since the
1569 // primary supertype array is read-only, there's no chance of confusion
1570 // where we bypass an array load and an array store.
1571 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1572 if (offset == Type::OffsetBot ||
1573 (offset >= primary_supers_offset &&
1574 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1575 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1576 offset = in_bytes(Klass::secondary_super_cache_offset());
1577 tj = tk = tk->with_offset(offset);
1578 }
1579 }
1580
1581 // Flatten all Raw pointers together.
1582 if (tj->base() == Type::RawPtr)
1583 tj = TypeRawPtr::BOTTOM;
1673 intptr_t key = (intptr_t) adr_type;
1674 key ^= key >> logAliasCacheSize;
1675 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1676 }
1677
1678
1679 //-----------------------------grow_alias_types--------------------------------
1680 void Compile::grow_alias_types() {
1681 const int old_ats = _max_alias_types; // how many before?
1682 const int new_ats = old_ats; // how many more?
1683 const int grow_ats = old_ats+new_ats; // how many now?
1684 _max_alias_types = grow_ats;
1685 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1686 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1687 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1688 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1689 }
1690
1691
1692 //--------------------------------find_alias_type------------------------------
1693 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1694 if (!do_aliasing()) {
1695 return alias_type(AliasIdxBot);
1696 }
1697
1698 AliasCacheEntry* ace = nullptr;
1699 if (!uncached) {
1700 ace = probe_alias_cache(adr_type);
1701 if (ace->_adr_type == adr_type) {
1702 return alias_type(ace->_index);
1703 }
1704 }
1705
1706 // Handle special cases.
1707 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1708 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1709
1710 // Do it the slow way.
1711 const TypePtr* flat = flatten_alias_type(adr_type);
1712
1713 #ifdef ASSERT
1714 {
1715 ResourceMark rm;
1716 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1717 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1718 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1719 Type::str(adr_type));
1720 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1721 const TypeOopPtr* foop = flat->is_oopptr();
1722 // Scalarizable allocations have exact klass always.
1723 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1733 if (alias_type(i)->adr_type() == flat) {
1734 idx = i;
1735 break;
1736 }
1737 }
1738
1739 if (idx == AliasIdxTop) {
1740 if (no_create) return nullptr;
1741 // Grow the array if necessary.
1742 if (_num_alias_types == _max_alias_types) grow_alias_types();
1743 // Add a new alias type.
1744 idx = _num_alias_types++;
1745 _alias_types[idx]->Init(idx, flat);
1746 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1747 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1748 if (flat->isa_instptr()) {
1749 if (flat->offset() == java_lang_Class::klass_offset()
1750 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1751 alias_type(idx)->set_rewritable(false);
1752 }
1753 ciField* field = nullptr;
1754 if (flat->isa_aryptr()) {
1755 #ifdef ASSERT
1756 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1757 // (T_BYTE has the weakest alignment and size restrictions...)
1758 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1759 #endif
1760 const Type* elemtype = flat->is_aryptr()->elem();
1761 if (flat->offset() == TypePtr::OffsetBot) {
1762 alias_type(idx)->set_element(elemtype);
1763 }
1764 int field_offset = flat->is_aryptr()->field_offset().get();
1765 if (flat->is_flat() &&
1766 field_offset != Type::OffsetBot) {
1767 ciInlineKlass* vk = elemtype->inline_klass();
1768 field_offset += vk->first_field_offset();
1769 field = vk->get_field_by_offset(field_offset, false);
1770 }
1771 }
1772 if (flat->isa_klassptr()) {
1773 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1774 alias_type(idx)->set_rewritable(false);
1775 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1776 alias_type(idx)->set_rewritable(false);
1777 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1778 alias_type(idx)->set_rewritable(false);
1779 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1780 alias_type(idx)->set_rewritable(false);
1781 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1782 alias_type(idx)->set_rewritable(false);
1783 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1784 alias_type(idx)->set_rewritable(false);
1785 }
1786 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1787 // but the base pointer type is not distinctive enough to identify
1788 // references into JavaThread.)
1789
1790 // Check for final fields.
1791 const TypeInstPtr* tinst = flat->isa_instptr();
1792 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1793 if (tinst->const_oop() != nullptr &&
1794 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1795 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1796 // static field
1797 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1798 field = k->get_field_by_offset(tinst->offset(), true);
1799 } else if (tinst->is_inlinetypeptr()) {
1800 // Inline type field
1801 ciInlineKlass* vk = tinst->inline_klass();
1802 field = vk->get_field_by_offset(tinst->offset(), false);
1803 } else {
1804 ciInstanceKlass *k = tinst->instance_klass();
1805 field = k->get_field_by_offset(tinst->offset(), false);
1806 }
1807 }
1808 assert(field == nullptr ||
1809 original_field == nullptr ||
1810 (field->holder() == original_field->holder() &&
1811 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1812 field->is_static() == original_field->is_static()), "wrong field?");
1813 // Set field() and is_rewritable() attributes.
1814 if (field != nullptr) {
1815 alias_type(idx)->set_field(field);
1816 if (flat->isa_aryptr()) {
1817 // Fields of flat arrays are rewritable although they are declared final
1818 assert(flat->is_flat(), "must be a flat array");
1819 alias_type(idx)->set_rewritable(true);
1820 }
1821 }
1822 }
1823
1824 // Fill the cache for next time.
1825 if (!uncached) {
1826 ace->_adr_type = adr_type;
1827 ace->_index = idx;
1828 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1829
1830 // Might as well try to fill the cache for the flattened version, too.
1831 AliasCacheEntry* face = probe_alias_cache(flat);
1832 if (face->_adr_type == nullptr) {
1833 face->_adr_type = flat;
1834 face->_index = idx;
1835 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1836 }
1837 }
1838
1839 return alias_type(idx);
1840 }
1841
1842
1843 Compile::AliasType* Compile::alias_type(ciField* field) {
1844 const TypeOopPtr* t;
1845 if (field->is_static())
1846 t = TypeInstPtr::make(field->holder()->java_mirror());
1847 else
1848 t = TypeOopPtr::make_from_klass_raw(field->holder());
1849 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1850 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1851 return atp;
1852 }
1853
1854
1855 //------------------------------have_alias_type--------------------------------
1856 bool Compile::have_alias_type(const TypePtr* adr_type) {
1936 assert(!C->major_progress(), "not cleared");
1937
1938 if (_for_post_loop_igvn.length() > 0) {
1939 while (_for_post_loop_igvn.length() > 0) {
1940 Node* n = _for_post_loop_igvn.pop();
1941 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1942 igvn._worklist.push(n);
1943 }
1944 igvn.optimize();
1945 if (failing()) return;
1946 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1947 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1948
1949 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1950 if (C->major_progress()) {
1951 C->clear_major_progress(); // ensure that major progress is now clear
1952 }
1953 }
1954 }
1955
1956 void Compile::add_inline_type(Node* n) {
1957 assert(n->is_InlineType(), "unexpected node");
1958 _inline_type_nodes.push(n);
1959 }
1960
1961 void Compile::remove_inline_type(Node* n) {
1962 assert(n->is_InlineType(), "unexpected node");
1963 if (_inline_type_nodes.contains(n)) {
1964 _inline_type_nodes.remove(n);
1965 }
1966 }
1967
1968 // Does the return value keep otherwise useless inline type allocations alive?
1969 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1970 ResourceMark rm;
1971 Unique_Node_List wq;
1972 wq.push(ret_val);
1973 bool some_allocations = false;
1974 for (uint i = 0; i < wq.size(); i++) {
1975 Node* n = wq.at(i);
1976 if (n->outcnt() > 1) {
1977 // Some other use for the allocation
1978 return false;
1979 } else if (n->is_InlineType()) {
1980 wq.push(n->in(1));
1981 } else if (n->is_Phi()) {
1982 for (uint j = 1; j < n->req(); j++) {
1983 wq.push(n->in(j));
1984 }
1985 } else if (n->is_CheckCastPP() &&
1986 n->in(1)->is_Proj() &&
1987 n->in(1)->in(0)->is_Allocate()) {
1988 some_allocations = true;
1989 } else if (n->is_CheckCastPP()) {
1990 wq.push(n->in(1));
1991 }
1992 }
1993 return some_allocations;
1994 }
1995
1996 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1997 // Make sure that the return value does not keep an otherwise unused allocation alive
1998 if (tf()->returns_inline_type_as_fields()) {
1999 Node* ret = nullptr;
2000 for (uint i = 1; i < root()->req(); i++) {
2001 Node* in = root()->in(i);
2002 if (in->Opcode() == Op_Return) {
2003 assert(ret == nullptr, "only one return");
2004 ret = in;
2005 }
2006 }
2007 if (ret != nullptr) {
2008 Node* ret_val = ret->in(TypeFunc::Parms);
2009 if (igvn.type(ret_val)->isa_oopptr() &&
2010 return_val_keeps_allocations_alive(ret_val)) {
2011 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2012 assert(ret_val->outcnt() == 0, "should be dead now");
2013 igvn.remove_dead_node(ret_val);
2014 }
2015 }
2016 }
2017 if (_inline_type_nodes.length() == 0) {
2018 return;
2019 }
2020 // Scalarize inline types in safepoint debug info.
2021 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2022 set_scalarize_in_safepoints(true);
2023 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2024 _inline_type_nodes.at(i)->as_InlineType()->make_scalar_in_safepoints(&igvn);
2025 }
2026 if (remove) {
2027 // Remove inline type nodes by replacing them with their oop input
2028 while (_inline_type_nodes.length() > 0) {
2029 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2030 if (vt->outcnt() == 0) {
2031 igvn.remove_dead_node(vt);
2032 continue;
2033 }
2034 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2035 DEBUG_ONLY(bool must_be_buffered = false);
2036 Node* u = vt->out(i);
2037 // Check if any users are blackholes. If so, rewrite them to use either the
2038 // allocated buffer, or individual components, instead of the inline type node
2039 // that goes away.
2040 if (u->is_Blackhole()) {
2041 BlackholeNode* bh = u->as_Blackhole();
2042
2043 // Unlink the old input
2044 int idx = bh->find_edge(vt);
2045 assert(idx != -1, "The edge should be there");
2046 bh->del_req(idx);
2047 --i;
2048
2049 if (vt->is_allocated(&igvn)) {
2050 // Already has the allocated instance, blackhole that
2051 bh->add_req(vt->get_oop());
2052 } else {
2053 // Not allocated yet, blackhole the components
2054 for (uint c = 0; c < vt->field_count(); c++) {
2055 bh->add_req(vt->field_value(c));
2056 }
2057 }
2058
2059 // Node modified, record for IGVN
2060 igvn.record_for_igvn(bh);
2061 }
2062 #ifdef ASSERT
2063 // Verify that inline type is buffered when replacing by oop
2064 else if (u->is_InlineType()) {
2065 // InlineType uses don't need buffering because they are about to be replaced as well
2066 } else if (u->is_Phi()) {
2067 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2068 } else {
2069 must_be_buffered = true;
2070 }
2071 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2072 vt->dump(0);
2073 u->dump(0);
2074 assert(false, "Should have been buffered");
2075 }
2076 #endif
2077 }
2078 igvn.replace_node(vt, vt->get_oop());
2079 }
2080 }
2081 igvn.optimize();
2082 }
2083
2084 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2085 if (!_has_flat_accesses) {
2086 return;
2087 }
2088 // Initially, all flat array accesses share the same slice to
2089 // keep dependencies with Object[] array accesses (that could be
2090 // to a flat array) correct. We're done with parsing so we
2091 // now know all flat array accesses in this compile
2092 // unit. Let's move flat array accesses to their own slice,
2093 // one per element field. This should help memory access
2094 // optimizations.
2095 ResourceMark rm;
2096 Unique_Node_List wq;
2097 wq.push(root());
2098
2099 Node_List mergememnodes;
2100 Node_List memnodes;
2101
2102 // Alias index currently shared by all flat memory accesses
2103 int index = get_alias_index(TypeAryPtr::INLINES);
2104
2105 // Find MergeMem nodes and flat array accesses
2106 for (uint i = 0; i < wq.size(); i++) {
2107 Node* n = wq.at(i);
2108 if (n->is_Mem()) {
2109 const TypePtr* adr_type = nullptr;
2110 if (n->Opcode() == Op_StoreCM) {
2111 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2112 } else {
2113 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2114 }
2115 if (adr_type == TypeAryPtr::INLINES) {
2116 memnodes.push(n);
2117 }
2118 } else if (n->is_MergeMem()) {
2119 MergeMemNode* mm = n->as_MergeMem();
2120 if (mm->memory_at(index) != mm->base_memory()) {
2121 mergememnodes.push(n);
2122 }
2123 }
2124 for (uint j = 0; j < n->req(); j++) {
2125 Node* m = n->in(j);
2126 if (m != nullptr) {
2127 wq.push(m);
2128 }
2129 }
2130 }
2131
2132 if (memnodes.size() > 0) {
2133 _flat_accesses_share_alias = false;
2134
2135 // We are going to change the slice for the flat array
2136 // accesses so we need to clear the cache entries that refer to
2137 // them.
2138 for (uint i = 0; i < AliasCacheSize; i++) {
2139 AliasCacheEntry* ace = &_alias_cache[i];
2140 if (ace->_adr_type != nullptr &&
2141 ace->_adr_type->is_flat()) {
2142 ace->_adr_type = nullptr;
2143 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2144 }
2145 }
2146
2147 // Find what aliases we are going to add
2148 int start_alias = num_alias_types()-1;
2149 int stop_alias = 0;
2150
2151 for (uint i = 0; i < memnodes.size(); i++) {
2152 Node* m = memnodes.at(i);
2153 const TypePtr* adr_type = nullptr;
2154 if (m->Opcode() == Op_StoreCM) {
2155 adr_type = m->in(MemNode::OopStore)->adr_type();
2156 if (adr_type != TypeAryPtr::INLINES) {
2157 // store was optimized out and we lost track of the adr_type
2158 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2159 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2160 get_alias_index(adr_type));
2161 igvn.register_new_node_with_optimizer(clone);
2162 igvn.replace_node(m, clone);
2163 }
2164 } else {
2165 adr_type = m->adr_type();
2166 #ifdef ASSERT
2167 m->as_Mem()->set_adr_type(adr_type);
2168 #endif
2169 }
2170 int idx = get_alias_index(adr_type);
2171 start_alias = MIN2(start_alias, idx);
2172 stop_alias = MAX2(stop_alias, idx);
2173 }
2174
2175 assert(stop_alias >= start_alias, "should have expanded aliases");
2176
2177 Node_Stack stack(0);
2178 #ifdef ASSERT
2179 VectorSet seen(Thread::current()->resource_area());
2180 #endif
2181 // Now let's fix the memory graph so each flat array access
2182 // is moved to the right slice. Start from the MergeMem nodes.
2183 uint last = unique();
2184 for (uint i = 0; i < mergememnodes.size(); i++) {
2185 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2186 Node* n = current->memory_at(index);
2187 MergeMemNode* mm = nullptr;
2188 do {
2189 // Follow memory edges through memory accesses, phis and
2190 // narrow membars and push nodes on the stack. Once we hit
2191 // bottom memory, we pop element off the stack one at a
2192 // time, in reverse order, and move them to the right slice
2193 // by changing their memory edges.
2194 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2195 assert(!seen.test_set(n->_idx), "");
2196 // Uses (a load for instance) will need to be moved to the
2197 // right slice as well and will get a new memory state
2198 // that we don't know yet. The use could also be the
2199 // backedge of a loop. We put a place holder node between
2200 // the memory node and its uses. We replace that place
2201 // holder with the correct memory state once we know it,
2202 // i.e. when nodes are popped off the stack. Using the
2203 // place holder make the logic work in the presence of
2204 // loops.
2205 if (n->outcnt() > 1) {
2206 Node* place_holder = nullptr;
2207 assert(!n->has_out_with(Op_Node), "");
2208 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2209 Node* u = n->out(k);
2210 if (u != current && u->_idx < last) {
2211 bool success = false;
2212 for (uint l = 0; l < u->req(); l++) {
2213 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2214 continue;
2215 }
2216 Node* in = u->in(l);
2217 if (in == n) {
2218 if (place_holder == nullptr) {
2219 place_holder = new Node(1);
2220 place_holder->init_req(0, n);
2221 }
2222 igvn.replace_input_of(u, l, place_holder);
2223 success = true;
2224 }
2225 }
2226 if (success) {
2227 --k;
2228 }
2229 }
2230 }
2231 }
2232 if (n->is_Phi()) {
2233 stack.push(n, 1);
2234 n = n->in(1);
2235 } else if (n->is_Mem()) {
2236 stack.push(n, n->req());
2237 n = n->in(MemNode::Memory);
2238 } else {
2239 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2240 stack.push(n, n->req());
2241 n = n->in(0)->in(TypeFunc::Memory);
2242 }
2243 } else {
2244 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2245 // Build a new MergeMem node to carry the new memory state
2246 // as we build it. IGVN should fold extraneous MergeMem
2247 // nodes.
2248 mm = MergeMemNode::make(n);
2249 igvn.register_new_node_with_optimizer(mm);
2250 while (stack.size() > 0) {
2251 Node* m = stack.node();
2252 uint idx = stack.index();
2253 if (m->is_Mem()) {
2254 // Move memory node to its new slice
2255 const TypePtr* adr_type = m->adr_type();
2256 int alias = get_alias_index(adr_type);
2257 Node* prev = mm->memory_at(alias);
2258 igvn.replace_input_of(m, MemNode::Memory, prev);
2259 mm->set_memory_at(alias, m);
2260 } else if (m->is_Phi()) {
2261 // We need as many new phis as there are new aliases
2262 igvn.replace_input_of(m, idx, mm);
2263 if (idx == m->req()-1) {
2264 Node* r = m->in(0);
2265 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2266 const TypePtr* adr_type = get_adr_type(j);
2267 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2268 continue;
2269 }
2270 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2271 igvn.register_new_node_with_optimizer(phi);
2272 for (uint k = 1; k < m->req(); k++) {
2273 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2274 }
2275 mm->set_memory_at(j, phi);
2276 }
2277 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2278 igvn.register_new_node_with_optimizer(base_phi);
2279 for (uint k = 1; k < m->req(); k++) {
2280 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2281 }
2282 mm->set_base_memory(base_phi);
2283 }
2284 } else {
2285 // This is a MemBarCPUOrder node from
2286 // Parse::array_load()/Parse::array_store(), in the
2287 // branch that handles flat arrays hidden under
2288 // an Object[] array. We also need one new membar per
2289 // new alias to keep the unknown access that the
2290 // membars protect properly ordered with accesses to
2291 // known flat array.
2292 assert(m->is_Proj(), "projection expected");
2293 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2294 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2295 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2296 const TypePtr* adr_type = get_adr_type(j);
2297 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2298 continue;
2299 }
2300 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2301 igvn.register_new_node_with_optimizer(mb);
2302 Node* mem = mm->memory_at(j);
2303 mb->init_req(TypeFunc::Control, ctrl);
2304 mb->init_req(TypeFunc::Memory, mem);
2305 ctrl = new ProjNode(mb, TypeFunc::Control);
2306 igvn.register_new_node_with_optimizer(ctrl);
2307 mem = new ProjNode(mb, TypeFunc::Memory);
2308 igvn.register_new_node_with_optimizer(mem);
2309 mm->set_memory_at(j, mem);
2310 }
2311 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2312 }
2313 if (idx < m->req()-1) {
2314 idx += 1;
2315 stack.set_index(idx);
2316 n = m->in(idx);
2317 break;
2318 }
2319 // Take care of place holder nodes
2320 if (m->has_out_with(Op_Node)) {
2321 Node* place_holder = m->find_out_with(Op_Node);
2322 if (place_holder != nullptr) {
2323 Node* mm_clone = mm->clone();
2324 igvn.register_new_node_with_optimizer(mm_clone);
2325 Node* hook = new Node(1);
2326 hook->init_req(0, mm);
2327 igvn.replace_node(place_holder, mm_clone);
2328 hook->destruct(&igvn);
2329 }
2330 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2331 }
2332 stack.pop();
2333 }
2334 }
2335 } while(stack.size() > 0);
2336 // Fix the memory state at the MergeMem we started from
2337 igvn.rehash_node_delayed(current);
2338 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2339 const TypePtr* adr_type = get_adr_type(j);
2340 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2341 continue;
2342 }
2343 current->set_memory_at(j, mm);
2344 }
2345 current->set_memory_at(index, current->base_memory());
2346 }
2347 igvn.optimize();
2348 }
2349 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2350 #ifdef ASSERT
2351 if (!_flat_accesses_share_alias) {
2352 wq.clear();
2353 wq.push(root());
2354 for (uint i = 0; i < wq.size(); i++) {
2355 Node* n = wq.at(i);
2356 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2357 for (uint j = 0; j < n->req(); j++) {
2358 Node* m = n->in(j);
2359 if (m != nullptr) {
2360 wq.push(m);
2361 }
2362 }
2363 }
2364 }
2365 #endif
2366 }
2367
2368 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2369 if (OptimizeUnstableIf) {
2370 _unstable_if_traps.append(trap);
2371 }
2372 }
2373
2374 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2375 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2376 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2377 Node* n = trap->uncommon_trap();
2378 if (!useful.member(n)) {
2379 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2380 }
2381 }
2382 }
2383
2384 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2385 // or fold-compares case. Return true if succeed or not found.
2386 //
2387 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2423 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2424 Bytecodes::Code c = iter.cur_bc();
2425 Node* lhs = nullptr;
2426 Node* rhs = nullptr;
2427 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2428 lhs = unc->peek_operand(0);
2429 rhs = unc->peek_operand(1);
2430 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2431 lhs = unc->peek_operand(0);
2432 }
2433
2434 ResourceMark rm;
2435 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2436 assert(live_locals.is_valid(), "broken liveness info");
2437 int len = (int)live_locals.size();
2438
2439 for (int i = 0; i < len; i++) {
2440 Node* local = unc->local(jvms, i);
2441 // kill local using the liveness of next_bci.
2442 // give up when the local looks like an operand to secure reexecution.
2443 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2444 uint idx = jvms->locoff() + i;
2445 #ifdef ASSERT
2446 if (PrintOpto && Verbose) {
2447 tty->print("[unstable_if] kill local#%d: ", idx);
2448 local->dump();
2449 tty->cr();
2450 }
2451 #endif
2452 igvn.replace_input_of(unc, idx, top());
2453 modified = true;
2454 }
2455 }
2456 }
2457
2458 // keep the modified trap for late query
2459 if (modified) {
2460 trap->set_modified();
2461 } else {
2462 _unstable_if_traps.delete_at(i);
2463 }
2464 }
2465 igvn.optimize();
2466 }
2467
2468 // StringOpts and late inlining of string methods
2469 void Compile::inline_string_calls(bool parse_time) {
2470 {
2471 // remove useless nodes to make the usage analysis simpler
2472 ResourceMark rm;
2473 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2474 }
2475
2476 {
2477 ResourceMark rm;
2478 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2633
2634 if (_string_late_inlines.length() > 0) {
2635 assert(has_stringbuilder(), "inconsistent");
2636
2637 inline_string_calls(false);
2638
2639 if (failing()) return;
2640
2641 inline_incrementally_cleanup(igvn);
2642 }
2643
2644 set_inlining_incrementally(false);
2645 }
2646
2647 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2648 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2649 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2650 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2651 // as if "inlining_incrementally() == true" were set.
2652 assert(inlining_incrementally() == false, "not allowed");
2653 #ifdef ASSERT
2654 Unique_Node_List* modified_nodes = _modified_nodes;
2655 _modified_nodes = nullptr;
2656 #endif
2657 assert(_late_inlines.length() > 0, "sanity");
2658
2659 while (_late_inlines.length() > 0) {
2660 igvn_worklist()->ensure_empty(); // should be done with igvn
2661
2662 while (inline_incrementally_one()) {
2663 assert(!failing(), "inconsistent");
2664 }
2665 if (failing()) return;
2666
2667 inline_incrementally_cleanup(igvn);
2668 }
2669 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2670 }
2671
2672 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2673 if (_loop_opts_cnt > 0) {
2674 while (major_progress() && (_loop_opts_cnt > 0)) {
2675 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2676 PhaseIdealLoop::optimize(igvn, mode);
2677 _loop_opts_cnt--;
2678 if (failing()) return false;
2679 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2680 }
2681 }
2682 return true;
2683 }
2684
2685 // Remove edges from "root" to each SafePoint at a backward branch.
2686 // They were inserted during parsing (see add_safepoint()) to make
2687 // infinite loops without calls or exceptions visible to root, i.e.,
2688 // useful.
2689 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2796 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2797 }
2798 assert(!has_vbox_nodes(), "sanity");
2799
2800 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2801 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2802 igvn_worklist()->ensure_empty(); // should be done with igvn
2803 {
2804 ResourceMark rm;
2805 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2806 }
2807 igvn.reset_from_gvn(initial_gvn());
2808 igvn.optimize();
2809 if (failing()) return;
2810 }
2811
2812 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2813 // safepoints
2814 remove_root_to_sfpts_edges(igvn);
2815
2816 // Process inline type nodes now that all inlining is over
2817 process_inline_types(igvn);
2818
2819 adjust_flat_array_access_aliases(igvn);
2820
2821 if (failing()) return;
2822
2823 // Perform escape analysis
2824 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2825 if (has_loops()) {
2826 // Cleanup graph (remove dead nodes).
2827 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2828 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2829 if (failing()) return;
2830 }
2831 bool progress;
2832 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2833 do {
2834 ConnectionGraph::do_analysis(this, &igvn);
2835
2836 if (failing()) return;
2837
2838 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2839
2840 // Optimize out fields loads from scalar replaceable allocations.
2924 if (failing()) return;
2925
2926 // Loop transforms on the ideal graph. Range Check Elimination,
2927 // peeling, unrolling, etc.
2928 if (!optimize_loops(igvn, LoopOptsDefault)) {
2929 return;
2930 }
2931
2932 if (failing()) return;
2933
2934 C->clear_major_progress(); // ensure that major progress is now clear
2935
2936 process_for_post_loop_opts_igvn(igvn);
2937
2938 if (failing()) return;
2939
2940 #ifdef ASSERT
2941 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2942 #endif
2943
2944 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2945
2946 if (_late_inlines.length() > 0) {
2947 // More opportunities to optimize virtual and MH calls.
2948 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2949 process_late_inline_calls_no_inline(igvn);
2950 }
2951
2952 {
2953 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2954 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2955 PhaseMacroExpand mex(igvn);
2956 if (mex.expand_macro_nodes()) {
2957 assert(failing(), "must bail out w/ explicit message");
2958 return;
2959 }
2960 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2961 }
2962
2963 // Process inline type nodes again and remove them. From here
2964 // on we don't need to keep track of field values anymore.
2965 process_inline_types(igvn, /* remove= */ true);
2966
2967 {
2968 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2969 if (bs->expand_barriers(this, igvn)) {
2970 assert(failing(), "must bail out w/ explicit message");
2971 return;
2972 }
2973 print_method(PHASE_BARRIER_EXPANSION, 2);
2974 }
2975
2976 if (C->max_vector_size() > 0) {
2977 C->optimize_logic_cones(igvn);
2978 igvn.optimize();
2979 if (failing()) return;
2980 }
2981
2982 DEBUG_ONLY( _modified_nodes = nullptr; )
2983 DEBUG_ONLY( _late_inlines.clear(); )
2984
2985 assert(igvn._worklist.size() == 0, "not empty");
2986 } // (End scope of igvn; run destructor if necessary for asserts.)
2987
2988 check_no_dead_use();
2989
2990 process_print_inlining();
2991
2992 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2993 // to remove hashes to unlock nodes for modifications.
2994 C->node_hash()->clear();
2995
2996 // A method with only infinite loops has no edges entering loops from root
2997 {
2998 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2999 if (final_graph_reshaping()) {
3000 assert(failing(), "must bail out w/ explicit message");
3001 return;
3002 }
3003 }
3004
3005 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3593 // Accumulate any precedence edges
3594 if (mem->in(i) != nullptr) {
3595 n->add_prec(mem->in(i));
3596 }
3597 }
3598 // Everything above this point has been processed.
3599 done = true;
3600 }
3601 // Eliminate the previous StoreCM
3602 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3603 assert(mem->outcnt() == 0, "should be dead");
3604 mem->disconnect_inputs(this);
3605 } else {
3606 prev = mem;
3607 }
3608 mem = prev->in(MemNode::Memory);
3609 }
3610 }
3611 }
3612
3613
3614 //------------------------------final_graph_reshaping_impl----------------------
3615 // Implement items 1-5 from final_graph_reshaping below.
3616 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3617
3618 if ( n->outcnt() == 0 ) return; // dead node
3619 uint nop = n->Opcode();
3620
3621 // Check for 2-input instruction with "last use" on right input.
3622 // Swap to left input. Implements item (2).
3623 if( n->req() == 3 && // two-input instruction
3624 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3625 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3626 n->in(2)->outcnt() == 1 &&// right use IS a last use
3627 !n->in(2)->is_Con() ) { // right use is not a constant
3628 // Check for commutative opcode
3629 switch( nop ) {
3630 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3631 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3632 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3633 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3746 if (n->outcnt() > 1 &&
3747 !n->is_Proj() &&
3748 nop != Op_CreateEx &&
3749 nop != Op_CheckCastPP &&
3750 nop != Op_DecodeN &&
3751 nop != Op_DecodeNKlass &&
3752 !n->is_Mem() &&
3753 !n->is_Phi()) {
3754 Node *x = n->clone();
3755 call->set_req(TypeFunc::Parms, x);
3756 }
3757 }
3758 break;
3759 }
3760
3761 case Op_StoreCM:
3762 {
3763 // Convert OopStore dependence into precedence edge
3764 Node* prec = n->in(MemNode::OopStore);
3765 n->del_req(MemNode::OopStore);
3766 if (prec->is_MergeMem()) {
3767 MergeMemNode* mm = prec->as_MergeMem();
3768 Node* base = mm->base_memory();
3769 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3770 const TypePtr* adr_type = get_adr_type(i);
3771 if (adr_type->is_flat()) {
3772 Node* m = mm->memory_at(i);
3773 n->add_prec(m);
3774 }
3775 }
3776 if (mm->outcnt() == 0) {
3777 mm->disconnect_inputs(this);
3778 }
3779 } else {
3780 n->add_prec(prec);
3781 }
3782 eliminate_redundant_card_marks(n);
3783 }
3784
3785 // fall through
3786
3787 case Op_StoreB:
3788 case Op_StoreC:
3789 case Op_StoreI:
3790 case Op_StoreL:
3791 case Op_CompareAndSwapB:
3792 case Op_CompareAndSwapS:
3793 case Op_CompareAndSwapI:
3794 case Op_CompareAndSwapL:
3795 case Op_CompareAndSwapP:
3796 case Op_CompareAndSwapN:
3797 case Op_WeakCompareAndSwapB:
3798 case Op_WeakCompareAndSwapS:
3799 case Op_WeakCompareAndSwapI:
3800 case Op_WeakCompareAndSwapL:
3801 case Op_WeakCompareAndSwapP:
4382 // Replace all nodes with identical edges as m with m
4383 k->subsume_by(m, this);
4384 }
4385 }
4386 }
4387 break;
4388 }
4389 case Op_CmpUL: {
4390 if (!Matcher::has_match_rule(Op_CmpUL)) {
4391 // No support for unsigned long comparisons
4392 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4393 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4394 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4395 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4396 Node* andl = new AndLNode(orl, remove_sign_mask);
4397 Node* cmp = new CmpLNode(andl, n->in(2));
4398 n->subsume_by(cmp, this);
4399 }
4400 break;
4401 }
4402 #ifdef ASSERT
4403 case Op_InlineType: {
4404 n->dump(-1);
4405 assert(false, "inline type node was not removed");
4406 break;
4407 }
4408 #endif
4409 default:
4410 assert(!n->is_Call(), "");
4411 assert(!n->is_Mem(), "");
4412 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4413 break;
4414 }
4415 }
4416
4417 //------------------------------final_graph_reshaping_walk---------------------
4418 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4419 // requires that the walk visits a node's inputs before visiting the node.
4420 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4421 Unique_Node_List sfpt;
4422
4423 frc._visited.set(root->_idx); // first, mark node as visited
4424 uint cnt = root->req();
4425 Node *n = root;
4426 uint i = 0;
4427 while (true) {
4428 if (i < cnt) {
4768 }
4769 }
4770
4771 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4772 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4773 }
4774
4775 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4776 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4777 }
4778
4779 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4780 if (holder->is_initialized()) {
4781 return false;
4782 }
4783 if (holder->is_being_initialized()) {
4784 if (accessing_method->holder() == holder) {
4785 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4786 // <init>, or a static method. In all those cases, there was an initialization
4787 // barrier on the holder klass passed.
4788 if (accessing_method->is_class_initializer() ||
4789 accessing_method->is_object_constructor() ||
4790 accessing_method->is_static()) {
4791 return false;
4792 }
4793 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4794 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4795 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4796 // child class can become fully initialized while its parent class is still being initialized.
4797 if (accessing_method->is_class_initializer()) {
4798 return false;
4799 }
4800 }
4801 ciMethod* root = method(); // the root method of compilation
4802 if (root != accessing_method) {
4803 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4804 }
4805 }
4806 return true;
4807 }
4808
4809 #ifndef PRODUCT
4810 //------------------------------verify_bidirectional_edges---------------------
4811 // For each input edge to a node (ie - for each Use-Def edge), verify that
4812 // there is a corresponding Def-Use edge.
4813 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4814 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4815 uint stack_size = live_nodes() >> 4;
4816 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4817 nstack.push(_root);
4833 if (in != nullptr && !in->is_top()) {
4834 // Count instances of `next`
4835 int cnt = 0;
4836 for (uint idx = 0; idx < in->_outcnt; idx++) {
4837 if (in->_out[idx] == n) {
4838 cnt++;
4839 }
4840 }
4841 assert(cnt > 0, "Failed to find Def-Use edge.");
4842 // Check for duplicate edges
4843 // walk the input array downcounting the input edges to n
4844 for (uint j = 0; j < length; j++) {
4845 if (n->in(j) == in) {
4846 cnt--;
4847 }
4848 }
4849 assert(cnt == 0, "Mismatched edge count.");
4850 } else if (in == nullptr) {
4851 assert(i == 0 || i >= n->req() ||
4852 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4853 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4854 (n->is_Unlock() && i == (n->req() - 1)) ||
4855 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4856 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4857 } else {
4858 assert(in->is_top(), "sanity");
4859 // Nothing to check.
4860 }
4861 }
4862 }
4863 }
4864
4865 //------------------------------verify_graph_edges---------------------------
4866 // Walk the Graph and verify that there is a one-to-one correspondence
4867 // between Use-Def edges and Def-Use edges in the graph.
4868 void Compile::verify_graph_edges(bool no_dead_code) {
4869 if (VerifyGraphEdges) {
4870 Unique_Node_List visited;
4871
4872 // Call graph walk to check edges
4873 verify_bidirectional_edges(visited);
4874 if (no_dead_code) {
4875 // Now make sure that no visited node is used by an unvisited node.
4876 bool dead_nodes = false;
4966 // (1) subklass is already limited to a subtype of superklass => always ok
4967 // (2) subklass does not overlap with superklass => always fail
4968 // (3) superklass has NO subtypes and we can check with a simple compare.
4969 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4970 if (skip) {
4971 return SSC_full_test; // Let caller generate the general case.
4972 }
4973
4974 if (subk->is_java_subtype_of(superk)) {
4975 return SSC_always_true; // (0) and (1) this test cannot fail
4976 }
4977
4978 if (!subk->maybe_java_subtype_of(superk)) {
4979 return SSC_always_false; // (2) true path dead; no dynamic test needed
4980 }
4981
4982 const Type* superelem = superk;
4983 if (superk->isa_aryklassptr()) {
4984 int ignored;
4985 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4986
4987 // TODO 8325106 Fix comment
4988 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4989 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to
4990 // the klass for [LMyValue. Perform a full test.
4991 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4992 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4993 return SSC_full_test;
4994 }
4995 }
4996
4997 if (superelem->isa_instklassptr()) {
4998 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4999 if (!ik->has_subklass()) {
5000 if (!ik->is_final()) {
5001 // Add a dependency if there is a chance of a later subclass.
5002 dependencies()->assert_leaf_type(ik);
5003 }
5004 if (!superk->maybe_java_subtype_of(subk)) {
5005 return SSC_always_false;
5006 }
5007 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5008 }
5009 } else {
5010 // A primitive array type has no subtypes.
5011 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5012 }
5013
5014 return SSC_full_test;
5535 const Type* t = igvn.type_or_null(n);
5536 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5537 if (n->is_Type()) {
5538 t = n->as_Type()->type();
5539 assert(t == t->remove_speculative(), "no more speculative types");
5540 }
5541 // Iterate over outs - endless loops is unreachable from below
5542 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5543 Node *m = n->fast_out(i);
5544 if (not_a_node(m)) {
5545 continue;
5546 }
5547 worklist.push(m);
5548 }
5549 }
5550 igvn.check_no_speculative_types();
5551 #endif
5552 }
5553 }
5554
5555 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5556 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5557 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5558 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5559 ta->is_zero_type() || tb->is_zero_type() ||
5560 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5561 // Use old acmp if one operand is null or not an inline type
5562 return new CmpPNode(a, b);
5563 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5564 // We know that one operand is an inline type. Therefore,
5565 // new acmp will only return true if both operands are nullptr.
5566 // Check if both operands are null by or'ing the oops.
5567 a = phase->transform(new CastP2XNode(nullptr, a));
5568 b = phase->transform(new CastP2XNode(nullptr, b));
5569 a = phase->transform(new OrXNode(a, b));
5570 return new CmpXNode(a, phase->MakeConX(0));
5571 }
5572 // Use new acmp
5573 return nullptr;
5574 }
5575
5576 // Auxiliary methods to support randomized stressing/fuzzing.
5577
5578 int Compile::random() {
5579 _stress_seed = os::next_random(_stress_seed);
5580 return static_cast<int>(_stress_seed);
5581 }
5582
5583 // This method can be called the arbitrary number of times, with current count
5584 // as the argument. The logic allows selecting a single candidate from the
5585 // running list of candidates as follows:
5586 // int count = 0;
5587 // Cand* selected = null;
5588 // while(cand = cand->next()) {
5589 // if (randomized_select(++count)) {
5590 // selected = cand;
5591 // }
5592 // }
5593 //
5594 // Including count equalizes the chances any candidate is "selected".
5595 // This is useful when we don't have the complete list of candidates to choose
|