6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/parse.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subtypenode.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "utilities/bitMap.inline.hpp"
49 #include "utilities/powerOfTwo.hpp"
50
51 //----------------------------GraphKit-----------------------------------------
52 // Main utility constructor.
53 GraphKit::GraphKit(JVMState* jvms)
54 : Phase(Phase::Parser),
55 _env(C->env()),
56 _gvn(*C->initial_gvn()),
57 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
58 {
59 _exceptions = jvms->map()->next_exception();
60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
61 set_jvms(jvms);
62 }
63
64 // Private constructor for parser.
65 GraphKit::GraphKit()
66 : Phase(Phase::Parser),
67 _env(C->env()),
68 _gvn(*C->initial_gvn()),
69 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
70 {
71 _exceptions = NULL;
72 set_map(NULL);
73 debug_only(_sp = -99);
74 debug_only(set_bci(-99));
75 }
76
77
78
79 //---------------------------clean_stack---------------------------------------
80 // Clear away rubbish from the stack area of the JVM state.
81 // This destroys any arguments that may be waiting on the stack.
810 tty->print_cr("Zombie local %d: ", local);
811 jvms->dump();
812 }
813 return false;
814 }
815 }
816 }
817 return true;
818 }
819
820 #endif //ASSERT
821
822 // Helper function for enforcing certain bytecodes to reexecute if
823 // deoptimization happens
824 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
825 ciMethod* cur_method = jvms->method();
826 int cur_bci = jvms->bci();
827 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
828 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
829 return Interpreter::bytecode_should_reexecute(code) ||
830 (is_anewarray && code == Bytecodes::_multianewarray);
831 // Reexecute _multianewarray bytecode which was replaced with
832 // sequence of [a]newarray. See Parse::do_multianewarray().
833 //
834 // Note: interpreter should not have it set since this optimization
835 // is limited by dimensions and guarded by flag so in some cases
836 // multianewarray() runtime calls will be generated and
837 // the bytecode should not be reexecutes (stack will not be reset).
838 } else
839 return false;
840 }
841
842 // Helper function for adding JVMState and debug information to node
843 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
844 // Add the safepoint edges to the call (or other safepoint).
845
846 // Make sure dead locals are set to top. This
847 // should help register allocation time and cut down on the size
848 // of the deoptimization information.
849 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
850
851 // Walk the inline list to fill in the correct set of JVMState's
852 // Also fill in the associated edges for each JVMState.
853
854 // If the bytecode needs to be reexecuted we need to put
855 // the arguments back on the stack.
856 const bool should_reexecute = jvms()->should_reexecute();
857 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
858
859 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
1063 ciSignature* declared_signature = NULL;
1064 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1065 assert(declared_signature != NULL, "cannot be null");
1066 inputs = declared_signature->arg_size_for_bc(code);
1067 int size = declared_signature->return_type()->size();
1068 depth = size - inputs;
1069 }
1070 break;
1071
1072 case Bytecodes::_multianewarray:
1073 {
1074 ciBytecodeStream iter(method());
1075 iter.reset_to_bci(bci());
1076 iter.next();
1077 inputs = iter.get_dimensions();
1078 assert(rsize == 1, "");
1079 depth = rsize - inputs;
1080 }
1081 break;
1082
1083 case Bytecodes::_ireturn:
1084 case Bytecodes::_lreturn:
1085 case Bytecodes::_freturn:
1086 case Bytecodes::_dreturn:
1087 case Bytecodes::_areturn:
1088 assert(rsize == -depth, "");
1089 inputs = rsize;
1090 break;
1091
1092 case Bytecodes::_jsr:
1093 case Bytecodes::_jsr_w:
1094 inputs = 0;
1095 depth = 1; // S.B. depth=1, not zero
1096 break;
1097
1098 default:
1099 // bytecode produces a typed result
1100 inputs = rsize - depth;
1101 assert(inputs >= 0, "");
1102 break;
1145 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1146 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1147 return _gvn.transform( new AndLNode(conv, mask) );
1148 }
1149
1150 Node* GraphKit::ConvL2I(Node* offset) {
1151 // short-circuit a common case
1152 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1153 if (offset_con != (jlong)Type::OffsetBot) {
1154 return intcon((int) offset_con);
1155 }
1156 return _gvn.transform( new ConvL2INode(offset));
1157 }
1158
1159 //-------------------------load_object_klass-----------------------------------
1160 Node* GraphKit::load_object_klass(Node* obj) {
1161 // Special-case a fresh allocation to avoid building nodes:
1162 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1163 if (akls != NULL) return akls;
1164 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1165 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1166 }
1167
1168 //-------------------------load_array_length-----------------------------------
1169 Node* GraphKit::load_array_length(Node* array) {
1170 // Special-case a fresh allocation to avoid building nodes:
1171 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1172 Node *alen;
1173 if (alloc == NULL) {
1174 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1175 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1176 } else {
1177 alen = alloc->Ideal_length();
1178 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1179 if (ccast != alen) {
1180 alen = _gvn.transform(ccast);
1181 }
1182 }
1183 return alen;
1184 }
1185
1188 // the incoming address with NULL casted away. You are allowed to use the
1189 // not-null value only if you are control dependent on the test.
1190 #ifndef PRODUCT
1191 extern int explicit_null_checks_inserted,
1192 explicit_null_checks_elided;
1193 #endif
1194 Node* GraphKit::null_check_common(Node* value, BasicType type,
1195 // optional arguments for variations:
1196 bool assert_null,
1197 Node* *null_control,
1198 bool speculative) {
1199 assert(!assert_null || null_control == NULL, "not both at once");
1200 if (stopped()) return top();
1201 NOT_PRODUCT(explicit_null_checks_inserted++);
1202
1203 // Construct NULL check
1204 Node *chk = NULL;
1205 switch(type) {
1206 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1207 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1208 case T_ARRAY : // fall through
1209 type = T_OBJECT; // simplify further tests
1210 case T_OBJECT : {
1211 const Type *t = _gvn.type( value );
1212
1213 const TypeOopPtr* tp = t->isa_oopptr();
1214 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1215 // Only for do_null_check, not any of its siblings:
1216 && !assert_null && null_control == NULL) {
1217 // Usually, any field access or invocation on an unloaded oop type
1218 // will simply fail to link, since the statically linked class is
1219 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1220 // the static class is loaded but the sharper oop type is not.
1221 // Rather than checking for this obscure case in lots of places,
1222 // we simply observe that a null check on an unloaded class
1223 // will always be followed by a nonsense operation, so we
1224 // can just issue the uncommon trap here.
1225 // Our access to the unloaded class will only be correct
1226 // after it has been loaded and initialized, which requires
1227 // a trip through the interpreter.
1359 }
1360
1361 if (assert_null) {
1362 // Cast obj to null on this path.
1363 replace_in_map(value, zerocon(type));
1364 return zerocon(type);
1365 }
1366
1367 // Cast obj to not-null on this path, if there is no null_control.
1368 // (If there is a null_control, a non-null value may come back to haunt us.)
1369 if (type == T_OBJECT) {
1370 Node* cast = cast_not_null(value, false);
1371 if (null_control == NULL || (*null_control) == top())
1372 replace_in_map(value, cast);
1373 value = cast;
1374 }
1375
1376 return value;
1377 }
1378
1379
1380 //------------------------------cast_not_null----------------------------------
1381 // Cast obj to not-null on this path
1382 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1383 const Type *t = _gvn.type(obj);
1384 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1385 // Object is already not-null?
1386 if( t == t_not_null ) return obj;
1387
1388 Node *cast = new CastPPNode(obj,t_not_null);
1389 cast->init_req(0, control());
1390 cast = _gvn.transform( cast );
1391
1392 // Scan for instances of 'obj' in the current JVM mapping.
1393 // These instances are known to be not-null after the test.
1394 if (do_replace_in_map)
1395 replace_in_map(obj, cast);
1396
1397 return cast; // Return casted value
1398 }
1399
1400 // Sometimes in intrinsics, we implicitly know an object is not null
1401 // (there's no actual null check) so we can cast it to not null. In
1402 // the course of optimizations, the input to the cast can become null.
1403 // In that case that data path will die and we need the control path
1404 // to become dead as well to keep the graph consistent. So we have to
1405 // add a check for null for which one branch can't be taken. It uses
1406 // an Opaque4 node that will cause the check to be removed after loop
1407 // opts so the test goes away and the compiled code doesn't execute a
1408 // useless check.
1409 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1410 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1411 return value;
1495 MemNode::MemOrd mo,
1496 LoadNode::ControlDependency control_dependency,
1497 bool require_atomic_access,
1498 bool unaligned,
1499 bool mismatched,
1500 bool unsafe,
1501 uint8_t barrier_data) {
1502 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1503 const TypePtr* adr_type = NULL; // debug-mode-only argument
1504 debug_only(adr_type = C->get_adr_type(adr_idx));
1505 Node* mem = memory(adr_idx);
1506 Node* ld;
1507 if (require_atomic_access && bt == T_LONG) {
1508 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1509 } else if (require_atomic_access && bt == T_DOUBLE) {
1510 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1511 } else {
1512 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1513 }
1514 ld = _gvn.transform(ld);
1515 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1516 // Improve graph before escape analysis and boxing elimination.
1517 record_for_igvn(ld);
1518 }
1519 return ld;
1520 }
1521
1522 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1523 int adr_idx,
1524 MemNode::MemOrd mo,
1525 bool require_atomic_access,
1526 bool unaligned,
1527 bool mismatched,
1528 bool unsafe) {
1529 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1530 const TypePtr* adr_type = NULL;
1531 debug_only(adr_type = C->get_adr_type(adr_idx));
1532 Node *mem = memory(adr_idx);
1533 Node* st;
1534 if (require_atomic_access && bt == T_LONG) {
1535 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1546 }
1547 if (unsafe) {
1548 st->as_Store()->set_unsafe_access();
1549 }
1550 st = _gvn.transform(st);
1551 set_memory(st, adr_idx);
1552 // Back-to-back stores can only remove intermediate store with DU info
1553 // so push on worklist for optimizer.
1554 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1555 record_for_igvn(st);
1556
1557 return st;
1558 }
1559
1560 Node* GraphKit::access_store_at(Node* obj,
1561 Node* adr,
1562 const TypePtr* adr_type,
1563 Node* val,
1564 const Type* val_type,
1565 BasicType bt,
1566 DecoratorSet decorators) {
1567 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1568 // could be delayed during Parse (for example, in adjust_map_after_if()).
1569 // Execute transformation here to avoid barrier generation in such case.
1570 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1571 val = _gvn.makecon(TypePtr::NULL_PTR);
1572 }
1573
1574 if (stopped()) {
1575 return top(); // Dead path ?
1576 }
1577
1578 assert(val != NULL, "not dead path");
1579
1580 C2AccessValuePtr addr(adr, adr_type);
1581 C2AccessValue value(val, val_type);
1582 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1583 if (access.is_raw()) {
1584 return _barrier_set->BarrierSetC2::store_at(access, value);
1585 } else {
1586 return _barrier_set->store_at(access, value);
1587 }
1588 }
1589
1590 Node* GraphKit::access_load_at(Node* obj, // containing obj
1591 Node* adr, // actual adress to store val at
1592 const TypePtr* adr_type,
1593 const Type* val_type,
1594 BasicType bt,
1595 DecoratorSet decorators) {
1596 if (stopped()) {
1597 return top(); // Dead path ?
1598 }
1599
1600 C2AccessValuePtr addr(adr, adr_type);
1601 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1602 if (access.is_raw()) {
1603 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1604 } else {
1605 return _barrier_set->load_at(access, val_type);
1606 }
1607 }
1608
1609 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1610 const Type* val_type,
1611 BasicType bt,
1612 DecoratorSet decorators) {
1613 if (stopped()) {
1614 return top(); // Dead path ?
1615 }
1616
1617 C2AccessValuePtr addr(adr, NULL);
1618 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1619 if (access.is_raw()) {
1620 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1621 } else {
1679 }
1680 }
1681
1682 Node* GraphKit::access_atomic_add_at(Node* obj,
1683 Node* adr,
1684 const TypePtr* adr_type,
1685 int alias_idx,
1686 Node* new_val,
1687 const Type* value_type,
1688 BasicType bt,
1689 DecoratorSet decorators) {
1690 C2AccessValuePtr addr(adr, adr_type);
1691 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1692 if (access.is_raw()) {
1693 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1694 } else {
1695 return _barrier_set->atomic_add_at(access, new_val, value_type);
1696 }
1697 }
1698
1699 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1700 return _barrier_set->clone(this, src, dst, size, is_array);
1701 }
1702
1703 //-------------------------array_element_address-------------------------
1704 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1705 const TypeInt* sizetype, Node* ctrl) {
1706 uint shift = exact_log2(type2aelembytes(elembt));
1707 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1708
1709 // short-circuit a common case (saves lots of confusing waste motion)
1710 jint idx_con = find_int_con(idx, -1);
1711 if (idx_con >= 0) {
1712 intptr_t offset = header + ((intptr_t)idx_con << shift);
1713 return basic_plus_adr(ary, offset);
1714 }
1715
1716 // must be correct type for alignment purposes
1717 Node* base = basic_plus_adr(ary, header);
1718 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1719 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1720 return basic_plus_adr(ary, base, scale);
1721 }
1722
1723 //-------------------------load_array_element-------------------------
1724 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1725 const Type* elemtype = arytype->elem();
1726 BasicType elembt = elemtype->array_element_basic_type();
1727 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1728 if (elembt == T_NARROWOOP) {
1729 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1730 }
1731 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1732 return ld;
1733 }
1734
1735 //-------------------------set_arguments_for_java_call-------------------------
1736 // Arguments (pre-popped from the stack) are taken from the JVMS.
1737 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1738 // Add the call arguments:
1739 uint nargs = call->method()->arg_size();
1740 for (uint i = 0; i < nargs; i++) {
1741 Node* arg = argument(i);
1742 call->init_req(i + TypeFunc::Parms, arg);
1743 }
1744 }
1745
1746 //---------------------------set_edges_for_java_call---------------------------
1747 // Connect a newly created call into the current JVMS.
1748 // A return value node (if any) is returned from set_edges_for_java_call.
1749 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1750
1751 // Add the predefined inputs:
1752 call->init_req( TypeFunc::Control, control() );
1753 call->init_req( TypeFunc::I_O , i_o() );
1754 call->init_req( TypeFunc::Memory , reset_memory() );
1755 call->init_req( TypeFunc::FramePtr, frameptr() );
1756 call->init_req( TypeFunc::ReturnAdr, top() );
1757
1758 add_safepoint_edges(call, must_throw);
1759
1760 Node* xcall = _gvn.transform(call);
1761
1762 if (xcall == top()) {
1763 set_control(top());
1764 return;
1765 }
1766 assert(xcall == call, "call identity is stable");
1767
1768 // Re-use the current map to produce the result.
1769
1770 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1771 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1772 set_all_memory_call(xcall, separate_io_proj);
1773
1774 //return xcall; // no need, caller already has it
1775 }
1776
1777 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1778 if (stopped()) return top(); // maybe the call folded up?
1779
1780 // Capture the return value, if any.
1781 Node* ret;
1782 if (call->method() == NULL ||
1783 call->method()->return_type()->basic_type() == T_VOID)
1784 ret = top();
1785 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1786
1787 // Note: Since any out-of-line call can produce an exception,
1788 // we always insert an I_O projection from the call into the result.
1789
1790 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1791
1792 if (separate_io_proj) {
1793 // The caller requested separate projections be used by the fall
1794 // through and exceptional paths, so replace the projections for
1795 // the fall through path.
1796 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1797 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1798 }
1799 return ret;
1800 }
1801
1802 //--------------------set_predefined_input_for_runtime_call--------------------
1803 // Reading and setting the memory state is way conservative here.
1804 // The real problem is that I am not doing real Type analysis on memory,
1805 // so I cannot distinguish card mark stores from other stores. Across a GC
1806 // point the Store Barrier and the card mark memory has to agree. I cannot
1807 // have a card mark store and its barrier split across the GC point from
1808 // either above or below. Here I get that to happen by reading ALL of memory.
1809 // A better answer would be to separate out card marks from other memory.
1810 // For now, return the input memory state, so that it can be reused
1811 // after the call, if this call has restricted memory effects.
1812 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1813 // Set fixed predefined input arguments
1814 Node* memory = reset_memory();
1815 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1816 call->init_req( TypeFunc::Control, control() );
1817 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1818 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1869 if (use->is_MergeMem()) {
1870 wl.push(use);
1871 }
1872 }
1873 }
1874
1875 // Replace the call with the current state of the kit.
1876 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1877 JVMState* ejvms = NULL;
1878 if (has_exceptions()) {
1879 ejvms = transfer_exceptions_into_jvms();
1880 }
1881
1882 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1883 ReplacedNodes replaced_nodes_exception;
1884 Node* ex_ctl = top();
1885
1886 SafePointNode* final_state = stop();
1887
1888 // Find all the needed outputs of this call
1889 CallProjections callprojs;
1890 call->extract_projections(&callprojs, true);
1891
1892 Unique_Node_List wl;
1893 Node* init_mem = call->in(TypeFunc::Memory);
1894 Node* final_mem = final_state->in(TypeFunc::Memory);
1895 Node* final_ctl = final_state->in(TypeFunc::Control);
1896 Node* final_io = final_state->in(TypeFunc::I_O);
1897
1898 // Replace all the old call edges with the edges from the inlining result
1899 if (callprojs.fallthrough_catchproj != NULL) {
1900 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1901 }
1902 if (callprojs.fallthrough_memproj != NULL) {
1903 if (final_mem->is_MergeMem()) {
1904 // Parser's exits MergeMem was not transformed but may be optimized
1905 final_mem = _gvn.transform(final_mem);
1906 }
1907 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1908 add_mergemem_users_to_worklist(wl, final_mem);
1909 }
1910 if (callprojs.fallthrough_ioproj != NULL) {
1911 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1912 }
1913
1914 // Replace the result with the new result if it exists and is used
1915 if (callprojs.resproj != NULL && result != NULL) {
1916 C->gvn_replace_by(callprojs.resproj, result);
1917 }
1918
1919 if (ejvms == NULL) {
1920 // No exception edges to simply kill off those paths
1921 if (callprojs.catchall_catchproj != NULL) {
1922 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1923 }
1924 if (callprojs.catchall_memproj != NULL) {
1925 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1926 }
1927 if (callprojs.catchall_ioproj != NULL) {
1928 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1929 }
1930 // Replace the old exception object with top
1931 if (callprojs.exobj != NULL) {
1932 C->gvn_replace_by(callprojs.exobj, C->top());
1933 }
1934 } else {
1935 GraphKit ekit(ejvms);
1936
1937 // Load my combined exception state into the kit, with all phis transformed:
1938 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1939 replaced_nodes_exception = ex_map->replaced_nodes();
1940
1941 Node* ex_oop = ekit.use_exception_state(ex_map);
1942
1943 if (callprojs.catchall_catchproj != NULL) {
1944 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1945 ex_ctl = ekit.control();
1946 }
1947 if (callprojs.catchall_memproj != NULL) {
1948 Node* ex_mem = ekit.reset_memory();
1949 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
1950 add_mergemem_users_to_worklist(wl, ex_mem);
1951 }
1952 if (callprojs.catchall_ioproj != NULL) {
1953 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1954 }
1955
1956 // Replace the old exception object with the newly created one
1957 if (callprojs.exobj != NULL) {
1958 C->gvn_replace_by(callprojs.exobj, ex_oop);
1959 }
1960 }
1961
1962 // Disconnect the call from the graph
1963 call->disconnect_inputs(NULL, C);
1964 C->gvn_replace_by(call, C->top());
1965
1966 // Clean up any MergeMems that feed other MergeMems since the
1967 // optimizer doesn't like that.
1968 while (wl.size() > 0) {
1969 _gvn.transform(wl.pop());
1970 }
1971
1972 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1973 replaced_nodes.apply(C, final_ctl);
1974 }
1975 if (!ex_ctl->is_top() && do_replaced_nodes) {
1976 replaced_nodes_exception.apply(C, ex_ctl);
1977 }
1978 }
1979
1980
1981 //------------------------------increment_counter------------------------------
1982 // for statistics: increment a VM counter by 1
1983
1984 void GraphKit::increment_counter(address counter_addr) {
1985 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1986 increment_counter(adr1);
1987 }
1988
1989 void GraphKit::increment_counter(Node* counter_addr) {
1990 int adr_type = Compile::AliasIdxRaw;
1991 Node* ctrl = control();
1992 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2173 speculative = speculative->with_inline_depth(jvms()->depth());
2174 } else if (current_type->would_improve_ptr(ptr_kind)) {
2175 // Profiling report that null was never seen so we can change the
2176 // speculative type to non null ptr.
2177 if (ptr_kind == ProfileAlwaysNull) {
2178 speculative = TypePtr::NULL_PTR;
2179 } else {
2180 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2181 const TypePtr* ptr = TypePtr::NOTNULL;
2182 if (speculative != NULL) {
2183 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2184 } else {
2185 speculative = ptr;
2186 }
2187 }
2188 }
2189
2190 if (speculative != current_type->speculative()) {
2191 // Build a type with a speculative type (what we think we know
2192 // about the type but will need a guard when we use it)
2193 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2194 // We're changing the type, we need a new CheckCast node to carry
2195 // the new type. The new type depends on the control: what
2196 // profiling tells us is only valid from here as far as we can
2197 // tell.
2198 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2199 cast = _gvn.transform(cast);
2200 replace_in_map(n, cast);
2201 n = cast;
2202 }
2203
2204 return n;
2205 }
2206
2207 /**
2208 * Record profiling data from receiver profiling at an invoke with the
2209 * type system so that it can propagate it (speculation)
2210 *
2211 * @param n receiver node
2212 *
2213 * @return node with improved type
2214 */
2215 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2216 if (!UseTypeSpeculation) {
2217 return n;
2218 }
2219 ciKlass* exact_kls = profile_has_unique_klass();
2220 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2221 if ((java_bc() == Bytecodes::_checkcast ||
2222 java_bc() == Bytecodes::_instanceof ||
2223 java_bc() == Bytecodes::_aastore) &&
2224 method()->method_data()->is_mature()) {
2225 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2226 if (data != NULL) {
2227 if (!data->as_BitData()->null_seen()) {
2228 ptr_kind = ProfileNeverNull;
2229 } else {
2230 assert(data->is_ReceiverTypeData(), "bad profile data type");
2231 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2232 uint i = 0;
2233 for (; i < call->row_limit(); i++) {
2234 ciKlass* receiver = call->receiver(i);
2235 if (receiver != NULL) {
2236 break;
2237 }
2238 }
2239 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2240 }
2241 }
2242 }
2243 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2244 }
2245
2246 /**
2247 * Record profiling data from argument profiling at an invoke with the
2248 * type system so that it can propagate it (speculation)
2249 *
2250 * @param dest_method target method for the call
2251 * @param bc what invoke bytecode is this?
2252 */
2253 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2254 if (!UseTypeSpeculation) {
2255 return;
2256 }
2257 const TypeFunc* tf = TypeFunc::make(dest_method);
2258 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2259 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2260 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2261 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2262 if (is_reference_type(targ->basic_type())) {
2263 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2264 ciKlass* better_type = NULL;
2265 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2266 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2267 }
2268 i++;
2269 }
2270 }
2271 }
2272
2273 /**
2274 * Record profiling data from parameter profiling at an invoke with
2275 * the type system so that it can propagate it (speculation)
2276 */
2277 void GraphKit::record_profiled_parameters_for_speculation() {
2278 if (!UseTypeSpeculation) {
2279 return;
2280 }
2281 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2312 if (Matcher::strict_fp_requires_explicit_rounding) {
2313 // If a strict caller invokes a non-strict callee, round a double result.
2314 // A non-strict method may return a double value which has an extended exponent,
2315 // but this must not be visible in a caller which is strict.
2316 BasicType result_type = dest_method->return_type()->basic_type();
2317 assert(method() != NULL, "must have caller context");
2318 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2319 // Destination method's return value is on top of stack
2320 // dstore_rounding() does gvn.transform
2321 Node *result = pop_pair();
2322 result = dstore_rounding(result);
2323 push_pair(result);
2324 }
2325 }
2326 }
2327
2328 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2329 if (Matcher::strict_fp_requires_explicit_rounding) {
2330 // (Note: TypeFunc::make has a cache that makes this fast.)
2331 const TypeFunc* tf = TypeFunc::make(dest_method);
2332 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2333 for (int j = 0; j < nargs; j++) {
2334 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2335 if (targ->basic_type() == T_DOUBLE) {
2336 // If any parameters are doubles, they must be rounded before
2337 // the call, dstore_rounding does gvn.transform
2338 Node *arg = argument(j);
2339 arg = dstore_rounding(arg);
2340 set_argument(j, arg);
2341 }
2342 }
2343 }
2344 }
2345
2346 // rounding for strict float precision conformance
2347 Node* GraphKit::precision_rounding(Node* n) {
2348 if (Matcher::strict_fp_requires_explicit_rounding) {
2349 #ifdef IA32
2350 if (_method->flags().is_strict() && UseSSE == 0) {
2351 return _gvn.transform(new RoundFloatNode(0, n));
2352 }
2353 #else
2354 Unimplemented();
2794
2795 // Now do a linear scan of the secondary super-klass array. Again, no real
2796 // performance impact (too rare) but it's gotta be done.
2797 // Since the code is rarely used, there is no penalty for moving it
2798 // out of line, and it can only improve I-cache density.
2799 // The decision to inline or out-of-line this final check is platform
2800 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2801 Node* psc = gvn.transform(
2802 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2803
2804 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2805 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2806 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2807
2808 // Return false path; set default control to true path.
2809 *ctrl = gvn.transform(r_ok_subtype);
2810 return gvn.transform(r_not_subtype);
2811 }
2812
2813 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2814 if (ExpandSubTypeCheckAtParseTime) {
2815 MergeMemNode* mem = merged_memory();
2816 Node* ctrl = control();
2817 Node* subklass = obj_or_subklass;
2818 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2819 subklass = load_object_klass(obj_or_subklass);
2820 }
2821
2822 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2823 set_control(ctrl);
2824 return n;
2825 }
2826
2827 const TypePtr* adr_type = TypeKlassPtr::make(TypePtr::NotNull, C->env()->Object_klass(), Type::OffsetBot);
2828 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2829 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2830 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2831 set_control(_gvn.transform(new IfTrueNode(iff)));
2832 return _gvn.transform(new IfFalseNode(iff));
2833 }
2834
2835 // Profile-driven exact type check:
2836 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2837 float prob,
2838 Node* *casted_receiver) {
2839 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2840 Node* recv_klass = load_object_klass(receiver);
2841 Node* want_klass = makecon(tklass);
2842 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2843 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2844 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2845 set_control( _gvn.transform( new IfTrueNode (iff) ));
2846 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2847
2848 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2849 assert(recv_xtype->klass_is_exact(), "");
2850
2851 // Subsume downstream occurrences of receiver with a cast to
2852 // recv_xtype, since now we know what the type will be.
2853 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2854 (*casted_receiver) = _gvn.transform(cast);
2855 // (User must make the replace_in_map call.)
2856
2857 return fail;
2858 }
2859
2860 //------------------------------subtype_check_receiver-------------------------
2861 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2862 Node** casted_receiver) {
2863 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2864 Node* want_klass = makecon(tklass);
2865
2866 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2867
2868 // Cast receiver after successful check
2869 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2870 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2871 (*casted_receiver) = _gvn.transform(cast);
2872
2873 return slow_ctl;
2874 }
2875
2876 //------------------------------seems_never_null-------------------------------
2877 // Use null_seen information if it is available from the profile.
2878 // If we see an unexpected null at a type check we record it and force a
2879 // recompile; the offending check will be recompiled to handle NULLs.
2880 // If we see several offending BCIs, then all checks in the
2881 // method will be recompiled.
2882 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2883 speculating = !_gvn.type(obj)->speculative_maybe_null();
2884 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2885 if (UncommonNullCast // Cutout for this technique
2886 && obj != null() // And not the -Xcomp stupid case?
2887 && !too_many_traps(reason)
2888 ) {
2889 if (speculating) {
2890 return true;
2891 }
2892 if (data == NULL)
2893 // Edge case: no mature data. Be optimistic here.
2894 return true;
2895 // If the profile has not seen a null, assume it won't happen.
2896 assert(java_bc() == Bytecodes::_checkcast ||
2897 java_bc() == Bytecodes::_instanceof ||
2898 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2899 return !data->as_BitData()->null_seen();
2900 }
2901 speculating = false;
2902 return false;
2903 }
2904
2905 void GraphKit::guard_klass_being_initialized(Node* klass) {
2906 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
2907 Node* adr = basic_plus_adr(top(), klass, init_state_off);
2908 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2909 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
2910 T_BYTE, MemNode::unordered);
2911 init_state = _gvn.transform(init_state);
2912
2913 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
2914
2915 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
2916 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2917
2918 { BuildCutout unless(this, tst, PROB_MAX);
2958
2959 //------------------------maybe_cast_profiled_receiver-------------------------
2960 // If the profile has seen exactly one type, narrow to exactly that type.
2961 // Subsequent type checks will always fold up.
2962 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2963 ciKlass* require_klass,
2964 ciKlass* spec_klass,
2965 bool safe_for_replace) {
2966 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2967
2968 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2969
2970 // Make sure we haven't already deoptimized from this tactic.
2971 if (too_many_traps_or_recompiles(reason))
2972 return NULL;
2973
2974 // (No, this isn't a call, but it's enough like a virtual call
2975 // to use the same ciMethod accessor to get the profile info...)
2976 // If we have a speculative type use it instead of profiling (which
2977 // may not help us)
2978 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2979 if (exact_kls != NULL) {// no cast failures here
2980 if (require_klass == NULL ||
2981 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2982 // If we narrow the type to match what the type profile sees or
2983 // the speculative type, we can then remove the rest of the
2984 // cast.
2985 // This is a win, even if the exact_kls is very specific,
2986 // because downstream operations, such as method calls,
2987 // will often benefit from the sharper type.
2988 Node* exact_obj = not_null_obj; // will get updated in place...
2989 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2990 &exact_obj);
2991 { PreserveJVMState pjvms(this);
2992 set_control(slow_ctl);
2993 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2994 }
2995 if (safe_for_replace) {
2996 replace_in_map(not_null_obj, exact_obj);
2997 }
2998 return exact_obj;
3063 // and the reflective instance-of call.
3064 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3065 kill_dead_locals(); // Benefit all the uncommon traps
3066 assert( !stopped(), "dead parse path should be checked in callers" );
3067 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3068 "must check for not-null not-dead klass in callers");
3069
3070 // Make the merge point
3071 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3072 RegionNode* region = new RegionNode(PATH_LIMIT);
3073 Node* phi = new PhiNode(region, TypeInt::BOOL);
3074 C->set_has_split_ifs(true); // Has chance for split-if optimization
3075
3076 ciProfileData* data = NULL;
3077 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3078 data = method()->method_data()->bci_to_data(bci());
3079 }
3080 bool speculative_not_null = false;
3081 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3082 && seems_never_null(obj, data, speculative_not_null));
3083
3084 // Null check; get casted pointer; set region slot 3
3085 Node* null_ctl = top();
3086 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3087
3088 // If not_null_obj is dead, only null-path is taken
3089 if (stopped()) { // Doing instance-of on a NULL?
3090 set_control(null_ctl);
3091 return intcon(0);
3092 }
3093 region->init_req(_null_path, null_ctl);
3094 phi ->init_req(_null_path, intcon(0)); // Set null path value
3095 if (null_ctl == top()) {
3096 // Do this eagerly, so that pattern matches like is_diamond_phi
3097 // will work even during parsing.
3098 assert(_null_path == PATH_LIMIT-1, "delete last");
3099 region->del_req(_null_path);
3100 phi ->del_req(_null_path);
3101 }
3102
3103 // Do we know the type check always succeed?
3104 bool known_statically = false;
3105 if (_gvn.type(superklass)->singleton()) {
3106 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3107 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3108 if (subk != NULL && subk->is_loaded()) {
3109 int static_res = C->static_subtype_check(superk, subk);
3110 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3111 }
3112 }
3113
3114 if (!known_statically) {
3115 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3116 // We may not have profiling here or it may not help us. If we
3117 // have a speculative type use it to perform an exact cast.
3118 ciKlass* spec_obj_type = obj_type->speculative_type();
3119 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3120 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3121 if (stopped()) { // Profile disagrees with this path.
3122 set_control(null_ctl); // Null is the only remaining possibility.
3123 return intcon(0);
3124 }
3125 if (cast_obj != NULL) {
3126 not_null_obj = cast_obj;
3127 }
3128 }
3129 }
3130
3131 // Generate the subtype check
3132 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3133
3134 // Plug in the success path to the general merge in slot 1.
3135 region->init_req(_obj_path, control());
3136 phi ->init_req(_obj_path, intcon(1));
3137
3138 // Plug in the failing path to the general merge in slot 2.
3139 region->init_req(_fail_path, not_subtype_ctrl);
3140 phi ->init_req(_fail_path, intcon(0));
3141
3142 // Return final merged results
3143 set_control( _gvn.transform(region) );
3144 record_for_igvn(region);
3145
3146 // If we know the type check always succeeds then we don't use the
3147 // profiling data at this bytecode. Don't lose it, feed it to the
3148 // type system as a speculative type.
3149 if (safe_for_replace) {
3150 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3151 replace_in_map(obj, casted_obj);
3152 }
3153
3154 return _gvn.transform(phi);
3155 }
3156
3157 //-------------------------------gen_checkcast---------------------------------
3158 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3159 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3160 // uncommon-trap paths work. Adjust stack after this call.
3161 // If failure_control is supplied and not null, it is filled in with
3162 // the control edge for the cast failure. Otherwise, an appropriate
3163 // uncommon trap or exception is thrown.
3164 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3165 Node* *failure_control) {
3166 kill_dead_locals(); // Benefit all the uncommon traps
3167 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3168 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3169
3170 // Fast cutout: Check the case that the cast is vacuously true.
3171 // This detects the common cases where the test will short-circuit
3172 // away completely. We do this before we perform the null check,
3173 // because if the test is going to turn into zero code, we don't
3174 // want a residual null check left around. (Causes a slowdown,
3175 // for example, in some objArray manipulations, such as a[i]=a[j].)
3176 if (tk->singleton()) {
3177 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3178 if (objtp != NULL && objtp->klass() != NULL) {
3179 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3180 case Compile::SSC_always_true:
3181 // If we know the type check always succeed then we don't use
3182 // the profiling data at this bytecode. Don't lose it, feed it
3183 // to the type system as a speculative type.
3184 return record_profiled_receiver_for_speculation(obj);
3185 case Compile::SSC_always_false:
3186 // It needs a null check because a null will *pass* the cast check.
3187 // A non-null value will always produce an exception.
3188 return null_assert(obj);
3189 }
3190 }
3191 }
3192
3193 ciProfileData* data = NULL;
3194 bool safe_for_replace = false;
3195 if (failure_control == NULL) { // use MDO in regular case only
3196 assert(java_bc() == Bytecodes::_aastore ||
3197 java_bc() == Bytecodes::_checkcast,
3198 "interpreter profiles type checks only for these BCs");
3199 data = method()->method_data()->bci_to_data(bci());
3200 safe_for_replace = true;
3201 }
3202
3203 // Make the merge point
3204 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3205 RegionNode* region = new RegionNode(PATH_LIMIT);
3206 Node* phi = new PhiNode(region, toop);
3207 C->set_has_split_ifs(true); // Has chance for split-if optimization
3208
3209 // Use null-cast information if it is available
3210 bool speculative_not_null = false;
3211 bool never_see_null = ((failure_control == NULL) // regular case only
3212 && seems_never_null(obj, data, speculative_not_null));
3213
3214 // Null check; get casted pointer; set region slot 3
3215 Node* null_ctl = top();
3216 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3217
3218 // If not_null_obj is dead, only null-path is taken
3219 if (stopped()) { // Doing instance-of on a NULL?
3220 set_control(null_ctl);
3221 return null();
3222 }
3223 region->init_req(_null_path, null_ctl);
3224 phi ->init_req(_null_path, null()); // Set null path value
3225 if (null_ctl == top()) {
3226 // Do this eagerly, so that pattern matches like is_diamond_phi
3227 // will work even during parsing.
3228 assert(_null_path == PATH_LIMIT-1, "delete last");
3229 region->del_req(_null_path);
3230 phi ->del_req(_null_path);
3231 }
3232
3233 Node* cast_obj = NULL;
3234 if (tk->klass_is_exact()) {
3235 // The following optimization tries to statically cast the speculative type of the object
3236 // (for example obtained during profiling) to the type of the superklass and then do a
3237 // dynamic check that the type of the object is what we expect. To work correctly
3238 // for checkcast and aastore the type of superklass should be exact.
3239 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3240 // We may not have profiling here or it may not help us. If we have
3241 // a speculative type use it to perform an exact cast.
3242 ciKlass* spec_obj_type = obj_type->speculative_type();
3243 if (spec_obj_type != NULL || data != NULL) {
3244 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3245 if (cast_obj != NULL) {
3246 if (failure_control != NULL) // failure is now impossible
3247 (*failure_control) = top();
3248 // adjust the type of the phi to the exact klass:
3249 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3250 }
3251 }
3252 }
3253
3254 if (cast_obj == NULL) {
3255 // Generate the subtype check
3256 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass );
3257
3258 // Plug in success path into the merge
3259 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3260 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3261 if (failure_control == NULL) {
3262 if (not_subtype_ctrl != top()) { // If failure is possible
3263 PreserveJVMState pjvms(this);
3264 set_control(not_subtype_ctrl);
3265 builtin_throw(Deoptimization::Reason_class_check, load_object_klass(not_null_obj));
3266 }
3267 } else {
3268 (*failure_control) = not_subtype_ctrl;
3269 }
3270 }
3271
3272 region->init_req(_obj_path, control());
3273 phi ->init_req(_obj_path, cast_obj);
3274
3275 // A merge of NULL or Casted-NotNull obj
3276 Node* res = _gvn.transform(phi);
3277
3278 // Note I do NOT always 'replace_in_map(obj,result)' here.
3279 // if( tk->klass()->can_be_primary_super() )
3280 // This means that if I successfully store an Object into an array-of-String
3281 // I 'forget' that the Object is really now known to be a String. I have to
3282 // do this because we don't have true union types for interfaces - if I store
3283 // a Baz into an array-of-Interface and then tell the optimizer it's an
3284 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3285 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3286 // replace_in_map( obj, res );
3287
3288 // Return final merged results
3289 set_control( _gvn.transform(region) );
3290 record_for_igvn(region);
3291
3292 return record_profiled_receiver_for_speculation(res);
3293 }
3294
3295 //------------------------------next_monitor-----------------------------------
3296 // What number should be given to the next monitor?
3297 int GraphKit::next_monitor() {
3298 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3299 int next = current + C->sync_stack_slots();
3300 // Keep the toplevel high water mark current:
3301 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3302 return current;
3303 }
3304
3305 //------------------------------insert_mem_bar---------------------------------
3306 // Memory barrier to avoid floating things around
3307 // The membar serves as a pinch point between both control and all memory slices.
3308 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3309 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3310 mb->init_req(TypeFunc::Control, control());
3311 mb->init_req(TypeFunc::Memory, reset_memory());
3312 Node* membar = _gvn.transform(mb);
3340 }
3341 Node* membar = _gvn.transform(mb);
3342 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3343 if (alias_idx == Compile::AliasIdxBot) {
3344 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3345 } else {
3346 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3347 }
3348 return membar;
3349 }
3350
3351 //------------------------------shared_lock------------------------------------
3352 // Emit locking code.
3353 FastLockNode* GraphKit::shared_lock(Node* obj) {
3354 // bci is either a monitorenter bc or InvocationEntryBci
3355 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3356 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3357
3358 if( !GenerateSynchronizationCode )
3359 return NULL; // Not locking things?
3360 if (stopped()) // Dead monitor?
3361 return NULL;
3362
3363 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3364
3365 // Box the stack location
3366 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3367 Node* mem = reset_memory();
3368
3369 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3370 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3371 // Create the counters for this fast lock.
3372 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3373 }
3374
3375 // Create the rtm counters for this fast lock if needed.
3376 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3377
3378 // Add monitor to debug info for the slow path. If we block inside the
3379 // slow path and de-opt, we need the monitor hanging around
3412 }
3413 #endif
3414
3415 return flock;
3416 }
3417
3418
3419 //------------------------------shared_unlock----------------------------------
3420 // Emit unlocking code.
3421 void GraphKit::shared_unlock(Node* box, Node* obj) {
3422 // bci is either a monitorenter bc or InvocationEntryBci
3423 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3424 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3425
3426 if( !GenerateSynchronizationCode )
3427 return;
3428 if (stopped()) { // Dead monitor?
3429 map()->pop_monitor(); // Kill monitor from debug info
3430 return;
3431 }
3432
3433 // Memory barrier to avoid floating things down past the locked region
3434 insert_mem_bar(Op_MemBarReleaseLock);
3435
3436 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3437 UnlockNode *unlock = new UnlockNode(C, tf);
3438 #ifdef ASSERT
3439 unlock->set_dbg_jvms(sync_jvms());
3440 #endif
3441 uint raw_idx = Compile::AliasIdxRaw;
3442 unlock->init_req( TypeFunc::Control, control() );
3443 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3444 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3445 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3446 unlock->init_req( TypeFunc::ReturnAdr, top() );
3447
3448 unlock->init_req(TypeFunc::Parms + 0, obj);
3449 unlock->init_req(TypeFunc::Parms + 1, box);
3450 unlock = _gvn.transform(unlock)->as_Unlock();
3451
3452 Node* mem = reset_memory();
3453
3454 // unlock has no side-effects, sets few values
3455 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3456
3457 // Kill monitor from debug info
3458 map()->pop_monitor( );
3459 }
3460
3461 //-------------------------------get_layout_helper-----------------------------
3462 // If the given klass is a constant or known to be an array,
3463 // fetch the constant layout helper value into constant_value
3464 // and return (Node*)NULL. Otherwise, load the non-constant
3465 // layout helper value, and return the node which represents it.
3466 // This two-faced routine is useful because allocation sites
3467 // almost always feature constant types.
3468 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3469 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3470 if (!StressReflectiveCode && inst_klass != NULL) {
3471 ciKlass* klass = inst_klass->klass();
3472 bool xklass = inst_klass->klass_is_exact();
3473 if (xklass || klass->is_array_klass()) {
3474 jint lhelper = klass->layout_helper();
3475 if (lhelper != Klass::_lh_neutral_value) {
3476 constant_value = lhelper;
3477 return (Node*) NULL;
3478 }
3479 }
3480 }
3481 constant_value = Klass::_lh_neutral_value; // put in a known value
3482 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3483 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3484 }
3485
3486 // We just put in an allocate/initialize with a big raw-memory effect.
3487 // Hook selected additional alias categories on the initialization.
3488 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3489 MergeMemNode* init_in_merge,
3490 Node* init_out_raw) {
3491 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3492 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3493
3515
3516 // a normal slow-call doesn't change i_o, but an allocation does
3517 // we create a separate i_o projection for the normal control path
3518 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3519 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3520
3521 // put in an initialization barrier
3522 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3523 rawoop)->as_Initialize();
3524 assert(alloc->initialization() == init, "2-way macro link must work");
3525 assert(init ->allocation() == alloc, "2-way macro link must work");
3526 {
3527 // Extract memory strands which may participate in the new object's
3528 // initialization, and source them from the new InitializeNode.
3529 // This will allow us to observe initializations when they occur,
3530 // and link them properly (as a group) to the InitializeNode.
3531 assert(init->in(InitializeNode::Memory) == malloc, "");
3532 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3533 init->set_req(InitializeNode::Memory, minit_in);
3534 record_for_igvn(minit_in); // fold it up later, if possible
3535 Node* minit_out = memory(rawidx);
3536 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3537 // Add an edge in the MergeMem for the header fields so an access
3538 // to one of those has correct memory state
3539 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3540 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3541 if (oop_type->isa_aryptr()) {
3542 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3543 int elemidx = C->get_alias_index(telemref);
3544 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3545 } else if (oop_type->isa_instptr()) {
3546 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3547 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3548 ciField* field = ik->nonstatic_field_at(i);
3549 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3550 continue; // do not bother to track really large numbers of fields
3551 // Find (or create) the alias category for this field:
3552 int fieldidx = C->alias_type(field)->index();
3553 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3554 }
3555 }
3556 }
3557
3558 // Cast raw oop to the real thing...
3559 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3560 javaoop = _gvn.transform(javaoop);
3561 C->set_recent_alloc(control(), javaoop);
3562 assert(just_allocated_object(control()) == javaoop, "just allocated");
3563
3564 #ifdef ASSERT
3565 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3576 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3577 }
3578 }
3579 #endif //ASSERT
3580
3581 return javaoop;
3582 }
3583
3584 //---------------------------new_instance--------------------------------------
3585 // This routine takes a klass_node which may be constant (for a static type)
3586 // or may be non-constant (for reflective code). It will work equally well
3587 // for either, and the graph will fold nicely if the optimizer later reduces
3588 // the type to a constant.
3589 // The optional arguments are for specialized use by intrinsics:
3590 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3591 // - If 'return_size_val', report the the total object size to the caller.
3592 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3593 Node* GraphKit::new_instance(Node* klass_node,
3594 Node* extra_slow_test,
3595 Node* *return_size_val,
3596 bool deoptimize_on_exception) {
3597 // Compute size in doublewords
3598 // The size is always an integral number of doublewords, represented
3599 // as a positive bytewise size stored in the klass's layout_helper.
3600 // The layout_helper also encodes (in a low bit) the need for a slow path.
3601 jint layout_con = Klass::_lh_neutral_value;
3602 Node* layout_val = get_layout_helper(klass_node, layout_con);
3603 int layout_is_con = (layout_val == NULL);
3604
3605 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3606 // Generate the initial go-slow test. It's either ALWAYS (return a
3607 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3608 // case) a computed value derived from the layout_helper.
3609 Node* initial_slow_test = NULL;
3610 if (layout_is_con) {
3611 assert(!StressReflectiveCode, "stress mode does not use these paths");
3612 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3613 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3614 } else { // reflective case
3615 // This reflective path is used by Unsafe.allocateInstance.
3616 // (It may be stress-tested by specifying StressReflectiveCode.)
3617 // Basically, we want to get into the VM is there's an illegal argument.
3618 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3619 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3620 if (extra_slow_test != intcon(0)) {
3621 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3622 }
3623 // (Macro-expander will further convert this to a Bool, if necessary.)
3634
3635 // Clear the low bits to extract layout_helper_size_in_bytes:
3636 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3637 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3638 size = _gvn.transform( new AndXNode(size, mask) );
3639 }
3640 if (return_size_val != NULL) {
3641 (*return_size_val) = size;
3642 }
3643
3644 // This is a precise notnull oop of the klass.
3645 // (Actually, it need not be precise if this is a reflective allocation.)
3646 // It's what we cast the result to.
3647 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3648 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3649 const TypeOopPtr* oop_type = tklass->as_instance_type();
3650
3651 // Now generate allocation code
3652
3653 // The entire memory state is needed for slow path of the allocation
3654 // since GC and deoptimization can happened.
3655 Node *mem = reset_memory();
3656 set_all_memory(mem); // Create new memory state
3657
3658 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3659 control(), mem, i_o(),
3660 size, klass_node,
3661 initial_slow_test);
3662
3663 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3664 }
3665
3666 //-------------------------------new_array-------------------------------------
3667 // helper for both newarray and anewarray
3668 // The 'length' parameter is (obviously) the length of the array.
3669 // See comments on new_instance for the meaning of the other arguments.
3670 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3671 Node* length, // number of array elements
3672 int nargs, // number of arguments to push back for uncommon trap
3673 Node* *return_size_val,
3674 bool deoptimize_on_exception) {
3675 jint layout_con = Klass::_lh_neutral_value;
3676 Node* layout_val = get_layout_helper(klass_node, layout_con);
3677 int layout_is_con = (layout_val == NULL);
3678
3679 if (!layout_is_con && !StressReflectiveCode &&
3680 !too_many_traps(Deoptimization::Reason_class_check)) {
3681 // This is a reflective array creation site.
3682 // Optimistically assume that it is a subtype of Object[],
3683 // so that we can fold up all the address arithmetic.
3684 layout_con = Klass::array_layout_helper(T_OBJECT);
3685 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3686 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3687 { BuildCutout unless(this, bol_lh, PROB_MAX);
3688 inc_sp(nargs);
3689 uncommon_trap(Deoptimization::Reason_class_check,
3690 Deoptimization::Action_maybe_recompile);
3691 }
3692 layout_val = NULL;
3693 layout_is_con = true;
3694 }
3695
3696 // Generate the initial go-slow test. Make sure we do not overflow
3697 // if length is huge (near 2Gig) or negative! We do not need
3698 // exact double-words here, just a close approximation of needed
3699 // double-words. We can't add any offset or rounding bits, lest we
3700 // take a size -1 of bytes and make it positive. Use an unsigned
3701 // compare, so negative sizes look hugely positive.
3702 int fast_size_limit = FastAllocateSizeLimit;
3703 if (layout_is_con) {
3704 assert(!StressReflectiveCode, "stress mode does not use these paths");
3705 // Increase the size limit if we have exact knowledge of array type.
3706 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3707 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3708 }
3709
3710 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3711 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3712
3713 // --- Size Computation ---
3714 // array_size = round_to_heap(array_header + (length << elem_shift));
3715 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3716 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3717 // The rounding mask is strength-reduced, if possible.
3718 int round_mask = MinObjAlignmentInBytes - 1;
3719 Node* header_size = NULL;
3720 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3721 // (T_BYTE has the weakest alignment and size restrictions...)
3722 if (layout_is_con) {
3723 int hsize = Klass::layout_helper_header_size(layout_con);
3724 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3725 BasicType etype = Klass::layout_helper_element_type(layout_con);
3726 if ((round_mask & ~right_n_bits(eshift)) == 0)
3727 round_mask = 0; // strength-reduce it if it goes away completely
3728 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3729 assert(header_size_min <= hsize, "generic minimum is smallest");
3730 header_size_min = hsize;
3731 header_size = intcon(hsize + round_mask);
3732 } else {
3733 Node* hss = intcon(Klass::_lh_header_size_shift);
3734 Node* hsm = intcon(Klass::_lh_header_size_mask);
3735 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3736 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3737 Node* mask = intcon(round_mask);
3738 header_size = _gvn.transform( new AddINode(hsize, mask) );
3739 }
3740
3741 Node* elem_shift = NULL;
3742 if (layout_is_con) {
3743 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3744 if (eshift != 0)
3745 elem_shift = intcon(eshift);
3746 } else {
3747 // There is no need to mask or shift this value.
3748 // The semantics of LShiftINode include an implicit mask to 0x1F.
3792 // places, one where the length is sharply limited, and the other
3793 // after a successful allocation.
3794 Node* abody = lengthx;
3795 if (elem_shift != NULL)
3796 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3797 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3798 if (round_mask != 0) {
3799 Node* mask = MakeConX(~round_mask);
3800 size = _gvn.transform( new AndXNode(size, mask) );
3801 }
3802 // else if round_mask == 0, the size computation is self-rounding
3803
3804 if (return_size_val != NULL) {
3805 // This is the size
3806 (*return_size_val) = size;
3807 }
3808
3809 // Now generate allocation code
3810
3811 // The entire memory state is needed for slow path of the allocation
3812 // since GC and deoptimization can happened.
3813 Node *mem = reset_memory();
3814 set_all_memory(mem); // Create new memory state
3815
3816 if (initial_slow_test->is_Bool()) {
3817 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3818 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3819 }
3820
3821 // Create the AllocateArrayNode and its result projections
3822 AllocateArrayNode* alloc
3823 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3824 control(), mem, i_o(),
3825 size, klass_node,
3826 initial_slow_test,
3827 length);
3828
3829 // Cast to correct type. Note that the klass_node may be constant or not,
3830 // and in the latter case the actual array type will be inexact also.
3831 // (This happens via a non-constant argument to inline_native_newArray.)
3832 // In any case, the value of klass_node provides the desired array type.
3833 const TypeInt* length_type = _gvn.find_int_type(length);
3834 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3835 if (ary_type->isa_aryptr() && length_type != NULL) {
3836 // Try to get a better type than POS for the size
3837 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3838 }
3839
3840 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3841
3842 // Cast length on remaining path to be as narrow as possible
3843 if (map()->find_edge(length) >= 0) {
3844 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3845 if (ccast != length) {
3846 _gvn.set_type_bottom(ccast);
3847 record_for_igvn(ccast);
3848 replace_in_map(length, ccast);
3849 }
3850 }
3851
3852 return javaoop;
3853 }
3854
3972 set_all_memory(ideal.merged_memory());
3973 set_i_o(ideal.i_o());
3974 set_control(ideal.ctrl());
3975 }
3976
3977 void GraphKit::final_sync(IdealKit& ideal) {
3978 // Final sync IdealKit and graphKit.
3979 sync_kit(ideal);
3980 }
3981
3982 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
3983 Node* len = load_array_length(load_String_value(str, set_ctrl));
3984 Node* coder = load_String_coder(str, set_ctrl);
3985 // Divide length by 2 if coder is UTF16
3986 return _gvn.transform(new RShiftINode(len, coder));
3987 }
3988
3989 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
3990 int value_offset = java_lang_String::value_offset();
3991 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3992 false, NULL, 0);
3993 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3994 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3995 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
3996 ciTypeArrayKlass::make(T_BYTE), true, 0);
3997 Node* p = basic_plus_adr(str, str, value_offset);
3998 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
3999 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4000 return load;
4001 }
4002
4003 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4004 if (!CompactStrings) {
4005 return intcon(java_lang_String::CODER_UTF16);
4006 }
4007 int coder_offset = java_lang_String::coder_offset();
4008 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4009 false, NULL, 0);
4010 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4011
4012 Node* p = basic_plus_adr(str, str, coder_offset);
4013 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4014 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4015 return load;
4016 }
4017
4018 void GraphKit::store_String_value(Node* str, Node* value) {
4019 int value_offset = java_lang_String::value_offset();
4020 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4021 false, NULL, 0);
4022 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4023
4024 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4025 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4026 }
4027
4028 void GraphKit::store_String_coder(Node* str, Node* value) {
4029 int coder_offset = java_lang_String::coder_offset();
4030 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4031 false, NULL, 0);
4032 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4033
4034 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4035 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4036 }
4037
4038 // Capture src and dst memory state with a MergeMemNode
4039 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4040 if (src_type == dst_type) {
4041 // Types are equal, we don't need a MergeMemNode
4042 return memory(src_type);
4043 }
4044 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4045 record_for_igvn(merge); // fold it up later, if possible
4046 int src_idx = C->get_alias_index(src_type);
4047 int dst_idx = C->get_alias_index(dst_type);
4048 merge->set_memory_at(src_idx, memory(src_idx));
4049 merge->set_memory_at(dst_idx, memory(dst_idx));
4050 return merge;
4051 }
4122 i_char->init_req(2, AddI(i_char, intcon(2)));
4123
4124 set_control(IfFalse(iff));
4125 set_memory(st, TypeAryPtr::BYTES);
4126 }
4127
4128 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4129 if (!field->is_constant()) {
4130 return NULL; // Field not marked as constant.
4131 }
4132 ciInstance* holder = NULL;
4133 if (!field->is_static()) {
4134 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4135 if (const_oop != NULL && const_oop->is_instance()) {
4136 holder = const_oop->as_instance();
4137 }
4138 }
4139 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4140 /*is_unsigned_load=*/false);
4141 if (con_type != NULL) {
4142 return makecon(con_type);
4143 }
4144 return NULL;
4145 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/machnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subtypenode.hpp"
50 #include "runtime/deoptimization.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/bitMap.inline.hpp"
53 #include "utilities/powerOfTwo.hpp"
54
55 //----------------------------GraphKit-----------------------------------------
56 // Main utility constructor.
57 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
58 : Phase(Phase::Parser),
59 _env(C->env()),
60 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
61 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
62 {
63 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
64 _exceptions = jvms->map()->next_exception();
65 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
66 set_jvms(jvms);
67 #ifdef ASSERT
68 if (_gvn.is_IterGVN() != NULL) {
69 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
70 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
71 _worklist_size = _gvn.C->for_igvn()->size();
72 }
73 #endif
74 }
75
76 // Private constructor for parser.
77 GraphKit::GraphKit()
78 : Phase(Phase::Parser),
79 _env(C->env()),
80 _gvn(*C->initial_gvn()),
81 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
82 {
83 _exceptions = NULL;
84 set_map(NULL);
85 debug_only(_sp = -99);
86 debug_only(set_bci(-99));
87 }
88
89
90
91 //---------------------------clean_stack---------------------------------------
92 // Clear away rubbish from the stack area of the JVM state.
93 // This destroys any arguments that may be waiting on the stack.
822 tty->print_cr("Zombie local %d: ", local);
823 jvms->dump();
824 }
825 return false;
826 }
827 }
828 }
829 return true;
830 }
831
832 #endif //ASSERT
833
834 // Helper function for enforcing certain bytecodes to reexecute if
835 // deoptimization happens
836 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
837 ciMethod* cur_method = jvms->method();
838 int cur_bci = jvms->bci();
839 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
840 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
841 return Interpreter::bytecode_should_reexecute(code) ||
842 (is_anewarray && (code == Bytecodes::_multianewarray));
843 // Reexecute _multianewarray bytecode which was replaced with
844 // sequence of [a]newarray. See Parse::do_multianewarray().
845 //
846 // Note: interpreter should not have it set since this optimization
847 // is limited by dimensions and guarded by flag so in some cases
848 // multianewarray() runtime calls will be generated and
849 // the bytecode should not be reexecutes (stack will not be reset).
850 } else {
851 return false;
852 }
853 }
854
855 // Helper function for adding JVMState and debug information to node
856 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
857 // Add the safepoint edges to the call (or other safepoint).
858
859 // Make sure dead locals are set to top. This
860 // should help register allocation time and cut down on the size
861 // of the deoptimization information.
862 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
863
864 // Walk the inline list to fill in the correct set of JVMState's
865 // Also fill in the associated edges for each JVMState.
866
867 // If the bytecode needs to be reexecuted we need to put
868 // the arguments back on the stack.
869 const bool should_reexecute = jvms()->should_reexecute();
870 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
871
872 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
1076 ciSignature* declared_signature = NULL;
1077 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1078 assert(declared_signature != NULL, "cannot be null");
1079 inputs = declared_signature->arg_size_for_bc(code);
1080 int size = declared_signature->return_type()->size();
1081 depth = size - inputs;
1082 }
1083 break;
1084
1085 case Bytecodes::_multianewarray:
1086 {
1087 ciBytecodeStream iter(method());
1088 iter.reset_to_bci(bci());
1089 iter.next();
1090 inputs = iter.get_dimensions();
1091 assert(rsize == 1, "");
1092 depth = rsize - inputs;
1093 }
1094 break;
1095
1096 case Bytecodes::_withfield: {
1097 bool ignored_will_link;
1098 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1099 int size = field->type()->size();
1100 inputs = size+1;
1101 depth = rsize - inputs;
1102 break;
1103 }
1104
1105 case Bytecodes::_ireturn:
1106 case Bytecodes::_lreturn:
1107 case Bytecodes::_freturn:
1108 case Bytecodes::_dreturn:
1109 case Bytecodes::_areturn:
1110 assert(rsize == -depth, "");
1111 inputs = rsize;
1112 break;
1113
1114 case Bytecodes::_jsr:
1115 case Bytecodes::_jsr_w:
1116 inputs = 0;
1117 depth = 1; // S.B. depth=1, not zero
1118 break;
1119
1120 default:
1121 // bytecode produces a typed result
1122 inputs = rsize - depth;
1123 assert(inputs >= 0, "");
1124 break;
1167 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1168 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1169 return _gvn.transform( new AndLNode(conv, mask) );
1170 }
1171
1172 Node* GraphKit::ConvL2I(Node* offset) {
1173 // short-circuit a common case
1174 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1175 if (offset_con != (jlong)Type::OffsetBot) {
1176 return intcon((int) offset_con);
1177 }
1178 return _gvn.transform( new ConvL2INode(offset));
1179 }
1180
1181 //-------------------------load_object_klass-----------------------------------
1182 Node* GraphKit::load_object_klass(Node* obj) {
1183 // Special-case a fresh allocation to avoid building nodes:
1184 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1185 if (akls != NULL) return akls;
1186 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1187 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT));
1188 }
1189
1190 //-------------------------load_array_length-----------------------------------
1191 Node* GraphKit::load_array_length(Node* array) {
1192 // Special-case a fresh allocation to avoid building nodes:
1193 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1194 Node *alen;
1195 if (alloc == NULL) {
1196 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1197 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1198 } else {
1199 alen = alloc->Ideal_length();
1200 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1201 if (ccast != alen) {
1202 alen = _gvn.transform(ccast);
1203 }
1204 }
1205 return alen;
1206 }
1207
1210 // the incoming address with NULL casted away. You are allowed to use the
1211 // not-null value only if you are control dependent on the test.
1212 #ifndef PRODUCT
1213 extern int explicit_null_checks_inserted,
1214 explicit_null_checks_elided;
1215 #endif
1216 Node* GraphKit::null_check_common(Node* value, BasicType type,
1217 // optional arguments for variations:
1218 bool assert_null,
1219 Node* *null_control,
1220 bool speculative) {
1221 assert(!assert_null || null_control == NULL, "not both at once");
1222 if (stopped()) return top();
1223 NOT_PRODUCT(explicit_null_checks_inserted++);
1224
1225 // Construct NULL check
1226 Node *chk = NULL;
1227 switch(type) {
1228 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1229 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1230 case T_INLINE_TYPE : // fall through
1231 case T_ARRAY : // fall through
1232 type = T_OBJECT; // simplify further tests
1233 case T_OBJECT : {
1234 const Type *t = _gvn.type( value );
1235
1236 const TypeOopPtr* tp = t->isa_oopptr();
1237 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1238 // Only for do_null_check, not any of its siblings:
1239 && !assert_null && null_control == NULL) {
1240 // Usually, any field access or invocation on an unloaded oop type
1241 // will simply fail to link, since the statically linked class is
1242 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1243 // the static class is loaded but the sharper oop type is not.
1244 // Rather than checking for this obscure case in lots of places,
1245 // we simply observe that a null check on an unloaded class
1246 // will always be followed by a nonsense operation, so we
1247 // can just issue the uncommon trap here.
1248 // Our access to the unloaded class will only be correct
1249 // after it has been loaded and initialized, which requires
1250 // a trip through the interpreter.
1382 }
1383
1384 if (assert_null) {
1385 // Cast obj to null on this path.
1386 replace_in_map(value, zerocon(type));
1387 return zerocon(type);
1388 }
1389
1390 // Cast obj to not-null on this path, if there is no null_control.
1391 // (If there is a null_control, a non-null value may come back to haunt us.)
1392 if (type == T_OBJECT) {
1393 Node* cast = cast_not_null(value, false);
1394 if (null_control == NULL || (*null_control) == top())
1395 replace_in_map(value, cast);
1396 value = cast;
1397 }
1398
1399 return value;
1400 }
1401
1402 Node* GraphKit::null2default(Node* value, ciInlineKlass* vk) {
1403 assert(!vk->is_scalarizable(), "Should only be used for non scalarizable inline klasses");
1404 Node* null_ctl = top();
1405 value = null_check_oop(value, &null_ctl);
1406 if (!null_ctl->is_top()) {
1407 // Return default value if oop is null
1408 Node* region = new RegionNode(3);
1409 region->init_req(1, control());
1410 region->init_req(2, null_ctl);
1411 value = PhiNode::make(region, value, TypeInstPtr::make(TypePtr::BotPTR, vk));
1412 value->set_req(2, InlineTypeNode::default_oop(gvn(), vk));
1413 set_control(gvn().transform(region));
1414 value = gvn().transform(value);
1415 }
1416 return value;
1417 }
1418
1419 //------------------------------cast_not_null----------------------------------
1420 // Cast obj to not-null on this path
1421 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1422 if (obj->is_InlineType()) {
1423 return obj;
1424 }
1425 const Type *t = _gvn.type(obj);
1426 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1427 // Object is already not-null?
1428 if( t == t_not_null ) return obj;
1429
1430 Node *cast = new CastPPNode(obj,t_not_null);
1431 cast->init_req(0, control());
1432 cast = _gvn.transform( cast );
1433
1434 if (t->is_inlinetypeptr() && t->inline_klass()->is_scalarizable()) {
1435 // Scalarize inline type now that we know it's non-null
1436 cast = InlineTypeNode::make_from_oop(this, cast, t->inline_klass())->as_ptr(&gvn());
1437 }
1438
1439 // Scan for instances of 'obj' in the current JVM mapping.
1440 // These instances are known to be not-null after the test.
1441 if (do_replace_in_map)
1442 replace_in_map(obj, cast);
1443
1444 return cast; // Return casted value
1445 }
1446
1447 // Sometimes in intrinsics, we implicitly know an object is not null
1448 // (there's no actual null check) so we can cast it to not null. In
1449 // the course of optimizations, the input to the cast can become null.
1450 // In that case that data path will die and we need the control path
1451 // to become dead as well to keep the graph consistent. So we have to
1452 // add a check for null for which one branch can't be taken. It uses
1453 // an Opaque4 node that will cause the check to be removed after loop
1454 // opts so the test goes away and the compiled code doesn't execute a
1455 // useless check.
1456 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1457 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1458 return value;
1542 MemNode::MemOrd mo,
1543 LoadNode::ControlDependency control_dependency,
1544 bool require_atomic_access,
1545 bool unaligned,
1546 bool mismatched,
1547 bool unsafe,
1548 uint8_t barrier_data) {
1549 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1550 const TypePtr* adr_type = NULL; // debug-mode-only argument
1551 debug_only(adr_type = C->get_adr_type(adr_idx));
1552 Node* mem = memory(adr_idx);
1553 Node* ld;
1554 if (require_atomic_access && bt == T_LONG) {
1555 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1556 } else if (require_atomic_access && bt == T_DOUBLE) {
1557 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1558 } else {
1559 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1560 }
1561 ld = _gvn.transform(ld);
1562
1563 if (((bt == T_OBJECT || bt == T_INLINE_TYPE) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1564 // Improve graph before escape analysis and boxing elimination.
1565 record_for_igvn(ld);
1566 }
1567 return ld;
1568 }
1569
1570 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1571 int adr_idx,
1572 MemNode::MemOrd mo,
1573 bool require_atomic_access,
1574 bool unaligned,
1575 bool mismatched,
1576 bool unsafe) {
1577 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1578 const TypePtr* adr_type = NULL;
1579 debug_only(adr_type = C->get_adr_type(adr_idx));
1580 Node *mem = memory(adr_idx);
1581 Node* st;
1582 if (require_atomic_access && bt == T_LONG) {
1583 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1594 }
1595 if (unsafe) {
1596 st->as_Store()->set_unsafe_access();
1597 }
1598 st = _gvn.transform(st);
1599 set_memory(st, adr_idx);
1600 // Back-to-back stores can only remove intermediate store with DU info
1601 // so push on worklist for optimizer.
1602 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1603 record_for_igvn(st);
1604
1605 return st;
1606 }
1607
1608 Node* GraphKit::access_store_at(Node* obj,
1609 Node* adr,
1610 const TypePtr* adr_type,
1611 Node* val,
1612 const Type* val_type,
1613 BasicType bt,
1614 DecoratorSet decorators,
1615 bool safe_for_replace) {
1616 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1617 // could be delayed during Parse (for example, in adjust_map_after_if()).
1618 // Execute transformation here to avoid barrier generation in such case.
1619 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1620 val = _gvn.makecon(TypePtr::NULL_PTR);
1621 }
1622
1623 if (stopped()) {
1624 return top(); // Dead path ?
1625 }
1626
1627 assert(val != NULL, "not dead path");
1628 if (val->is_InlineType()) {
1629 // Store to non-flattened field. Buffer the inline type and make sure
1630 // the store is re-executed if the allocation triggers deoptimization.
1631 PreserveReexecuteState preexecs(this);
1632 jvms()->set_should_reexecute(true);
1633 val = val->as_InlineType()->buffer(this, safe_for_replace);
1634 }
1635
1636 C2AccessValuePtr addr(adr, adr_type);
1637 C2AccessValue value(val, val_type);
1638 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1639 if (access.is_raw()) {
1640 return _barrier_set->BarrierSetC2::store_at(access, value);
1641 } else {
1642 return _barrier_set->store_at(access, value);
1643 }
1644 }
1645
1646 Node* GraphKit::access_load_at(Node* obj, // containing obj
1647 Node* adr, // actual adress to store val at
1648 const TypePtr* adr_type,
1649 const Type* val_type,
1650 BasicType bt,
1651 DecoratorSet decorators,
1652 Node* ctl) {
1653 if (stopped()) {
1654 return top(); // Dead path ?
1655 }
1656
1657 C2AccessValuePtr addr(adr, adr_type);
1658 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1659 if (access.is_raw()) {
1660 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1661 } else {
1662 return _barrier_set->load_at(access, val_type);
1663 }
1664 }
1665
1666 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1667 const Type* val_type,
1668 BasicType bt,
1669 DecoratorSet decorators) {
1670 if (stopped()) {
1671 return top(); // Dead path ?
1672 }
1673
1674 C2AccessValuePtr addr(adr, NULL);
1675 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1676 if (access.is_raw()) {
1677 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1678 } else {
1736 }
1737 }
1738
1739 Node* GraphKit::access_atomic_add_at(Node* obj,
1740 Node* adr,
1741 const TypePtr* adr_type,
1742 int alias_idx,
1743 Node* new_val,
1744 const Type* value_type,
1745 BasicType bt,
1746 DecoratorSet decorators) {
1747 C2AccessValuePtr addr(adr, adr_type);
1748 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1749 if (access.is_raw()) {
1750 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1751 } else {
1752 return _barrier_set->atomic_add_at(access, new_val, value_type);
1753 }
1754 }
1755
1756 void GraphKit::access_clone(Node* src_base, Node* dst_base, Node* countx, bool is_array) {
1757 return _barrier_set->clone(this, src_base, dst_base, countx, is_array);
1758 }
1759
1760 //-------------------------array_element_address-------------------------
1761 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1762 const TypeInt* sizetype, Node* ctrl) {
1763 uint shift = exact_log2(type2aelembytes(elembt));
1764 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1765 if (arytype_klass != NULL && arytype_klass->is_flat_array_klass()) {
1766 ciFlatArrayKlass* vak = arytype_klass->as_flat_array_klass();
1767 shift = vak->log2_element_size();
1768 }
1769 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1770
1771 // short-circuit a common case (saves lots of confusing waste motion)
1772 jint idx_con = find_int_con(idx, -1);
1773 if (idx_con >= 0) {
1774 intptr_t offset = header + ((intptr_t)idx_con << shift);
1775 return basic_plus_adr(ary, offset);
1776 }
1777
1778 // must be correct type for alignment purposes
1779 Node* base = basic_plus_adr(ary, header);
1780 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1781 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1782 return basic_plus_adr(ary, base, scale);
1783 }
1784
1785 //-------------------------load_array_element-------------------------
1786 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1787 const Type* elemtype = arytype->elem();
1788 BasicType elembt = elemtype->array_element_basic_type();
1789 assert(elembt != T_INLINE_TYPE, "inline types are not supported by this method");
1790 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1791 if (elembt == T_NARROWOOP) {
1792 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1793 }
1794 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1795 return ld;
1796 }
1797
1798 //-------------------------set_arguments_for_java_call-------------------------
1799 // Arguments (pre-popped from the stack) are taken from the JVMS.
1800 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1801 PreserveReexecuteState preexecs(this);
1802 if (EnableValhalla) {
1803 // Make sure the call is re-executed, if buffering of inline type arguments triggers deoptimization
1804 jvms()->set_should_reexecute(true);
1805 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1806 inc_sp(arg_size);
1807 }
1808 // Add the call arguments
1809 const TypeTuple* domain = call->tf()->domain_sig();
1810 ExtendedSignature sig_cc = ExtendedSignature(call->method()->get_sig_cc(), SigEntryFilter());
1811 uint nargs = domain->cnt();
1812 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1813 Node* arg = argument(i-TypeFunc::Parms);
1814 const Type* t = domain->field_at(i);
1815 if (call->method()->has_scalarized_args() && t->is_inlinetypeptr() && !t->maybe_null()) {
1816 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1817 InlineTypeNode* vt = arg->as_InlineType();
1818 vt->pass_fields(this, call, sig_cc, idx);
1819 // If an inline type argument is passed as fields, attach the Method* to the call site
1820 // to be able to access the extended signature later via attached_method_before_pc().
1821 // For example, see CompiledMethod::preserve_callee_argument_oops().
1822 call->set_override_symbolic_info(true);
1823 continue;
1824 } else if (arg->is_InlineType()) {
1825 // Pass inline type argument via oop to callee
1826 arg = arg->as_InlineType()->buffer(this);
1827 if (!is_late_inline) {
1828 arg = arg->as_InlineTypePtr()->get_oop();
1829 }
1830 }
1831 call->init_req(idx++, arg);
1832 // Skip reserved arguments
1833 BasicType bt = t->basic_type();
1834 while (SigEntry::next_is_reserved(sig_cc, bt, true)) {
1835 call->init_req(idx++, top());
1836 if (type2size[bt] == 2) {
1837 call->init_req(idx++, top());
1838 }
1839 }
1840 }
1841 }
1842
1843 //---------------------------set_edges_for_java_call---------------------------
1844 // Connect a newly created call into the current JVMS.
1845 // A return value node (if any) is returned from set_edges_for_java_call.
1846 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1847
1848 // Add the predefined inputs:
1849 call->init_req( TypeFunc::Control, control() );
1850 call->init_req( TypeFunc::I_O , i_o() );
1851 call->init_req( TypeFunc::Memory , reset_memory() );
1852 call->init_req( TypeFunc::FramePtr, frameptr() );
1853 call->init_req( TypeFunc::ReturnAdr, top() );
1854
1855 add_safepoint_edges(call, must_throw);
1856
1857 Node* xcall = _gvn.transform(call);
1858
1859 if (xcall == top()) {
1860 set_control(top());
1861 return;
1862 }
1863 assert(xcall == call, "call identity is stable");
1864
1865 // Re-use the current map to produce the result.
1866
1867 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1868 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1869 set_all_memory_call(xcall, separate_io_proj);
1870
1871 //return xcall; // no need, caller already has it
1872 }
1873
1874 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1875 if (stopped()) return top(); // maybe the call folded up?
1876
1877 // Note: Since any out-of-line call can produce an exception,
1878 // we always insert an I_O projection from the call into the result.
1879
1880 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1881
1882 if (separate_io_proj) {
1883 // The caller requested separate projections be used by the fall
1884 // through and exceptional paths, so replace the projections for
1885 // the fall through path.
1886 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1887 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1888 }
1889
1890 // Capture the return value, if any.
1891 Node* ret;
1892 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1893 ret = top();
1894 } else if (call->tf()->returns_inline_type_as_fields()) {
1895 // Return of multiple values (inline type fields): we create a
1896 // InlineType node, each field is a projection from the call.
1897 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1898 const Array<SigEntry>* sig_array = vk->extended_sig();
1899 GrowableArray<SigEntry> sig = GrowableArray<SigEntry>(sig_array->length());
1900 sig.appendAll(sig_array);
1901 ExtendedSignature sig_cc = ExtendedSignature(&sig, SigEntryFilter());
1902 uint base_input = TypeFunc::Parms + 1;
1903 ret = InlineTypeNode::make_from_multi(this, call, sig_cc, vk, base_input, false);
1904 } else {
1905 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1906 }
1907
1908 return ret;
1909 }
1910
1911 //--------------------set_predefined_input_for_runtime_call--------------------
1912 // Reading and setting the memory state is way conservative here.
1913 // The real problem is that I am not doing real Type analysis on memory,
1914 // so I cannot distinguish card mark stores from other stores. Across a GC
1915 // point the Store Barrier and the card mark memory has to agree. I cannot
1916 // have a card mark store and its barrier split across the GC point from
1917 // either above or below. Here I get that to happen by reading ALL of memory.
1918 // A better answer would be to separate out card marks from other memory.
1919 // For now, return the input memory state, so that it can be reused
1920 // after the call, if this call has restricted memory effects.
1921 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1922 // Set fixed predefined input arguments
1923 Node* memory = reset_memory();
1924 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1925 call->init_req( TypeFunc::Control, control() );
1926 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1927 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1978 if (use->is_MergeMem()) {
1979 wl.push(use);
1980 }
1981 }
1982 }
1983
1984 // Replace the call with the current state of the kit.
1985 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1986 JVMState* ejvms = NULL;
1987 if (has_exceptions()) {
1988 ejvms = transfer_exceptions_into_jvms();
1989 }
1990
1991 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1992 ReplacedNodes replaced_nodes_exception;
1993 Node* ex_ctl = top();
1994
1995 SafePointNode* final_state = stop();
1996
1997 // Find all the needed outputs of this call
1998 CallProjections* callprojs = call->extract_projections(true);
1999
2000 Unique_Node_List wl;
2001 Node* init_mem = call->in(TypeFunc::Memory);
2002 Node* final_mem = final_state->in(TypeFunc::Memory);
2003 Node* final_ctl = final_state->in(TypeFunc::Control);
2004 Node* final_io = final_state->in(TypeFunc::I_O);
2005
2006 // Replace all the old call edges with the edges from the inlining result
2007 if (callprojs->fallthrough_catchproj != NULL) {
2008 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2009 }
2010 if (callprojs->fallthrough_memproj != NULL) {
2011 if (final_mem->is_MergeMem()) {
2012 // Parser's exits MergeMem was not transformed but may be optimized
2013 final_mem = _gvn.transform(final_mem);
2014 }
2015 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2016 add_mergemem_users_to_worklist(wl, final_mem);
2017 }
2018 if (callprojs->fallthrough_ioproj != NULL) {
2019 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2020 }
2021
2022 // Replace the result with the new result if it exists and is used
2023 if (callprojs->resproj[0] != NULL && result != NULL) {
2024 assert(callprojs->nb_resproj == 1, "unexpected number of results");
2025 C->gvn_replace_by(callprojs->resproj[0], result);
2026 }
2027
2028 if (ejvms == NULL) {
2029 // No exception edges to simply kill off those paths
2030 if (callprojs->catchall_catchproj != NULL) {
2031 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2032 }
2033 if (callprojs->catchall_memproj != NULL) {
2034 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2035 }
2036 if (callprojs->catchall_ioproj != NULL) {
2037 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2038 }
2039 // Replace the old exception object with top
2040 if (callprojs->exobj != NULL) {
2041 C->gvn_replace_by(callprojs->exobj, C->top());
2042 }
2043 } else {
2044 GraphKit ekit(ejvms);
2045
2046 // Load my combined exception state into the kit, with all phis transformed:
2047 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2048 replaced_nodes_exception = ex_map->replaced_nodes();
2049
2050 Node* ex_oop = ekit.use_exception_state(ex_map);
2051
2052 if (callprojs->catchall_catchproj != NULL) {
2053 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2054 ex_ctl = ekit.control();
2055 }
2056 if (callprojs->catchall_memproj != NULL) {
2057 Node* ex_mem = ekit.reset_memory();
2058 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2059 add_mergemem_users_to_worklist(wl, ex_mem);
2060 }
2061 if (callprojs->catchall_ioproj != NULL) {
2062 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2063 }
2064
2065 // Replace the old exception object with the newly created one
2066 if (callprojs->exobj != NULL) {
2067 C->gvn_replace_by(callprojs->exobj, ex_oop);
2068 }
2069 }
2070
2071 // Disconnect the call from the graph
2072 call->disconnect_inputs(NULL, C);
2073 C->gvn_replace_by(call, C->top());
2074
2075 // Clean up any MergeMems that feed other MergeMems since the
2076 // optimizer doesn't like that.
2077 while (wl.size() > 0) {
2078 _gvn.transform(wl.pop());
2079 }
2080
2081 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2082 replaced_nodes.apply(C, final_ctl);
2083 }
2084 if (!ex_ctl->is_top() && do_replaced_nodes) {
2085 replaced_nodes_exception.apply(C, ex_ctl);
2086 }
2087 }
2088
2089
2090 //------------------------------increment_counter------------------------------
2091 // for statistics: increment a VM counter by 1
2092
2093 void GraphKit::increment_counter(address counter_addr) {
2094 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2095 increment_counter(adr1);
2096 }
2097
2098 void GraphKit::increment_counter(Node* counter_addr) {
2099 int adr_type = Compile::AliasIdxRaw;
2100 Node* ctrl = control();
2101 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2282 speculative = speculative->with_inline_depth(jvms()->depth());
2283 } else if (current_type->would_improve_ptr(ptr_kind)) {
2284 // Profiling report that null was never seen so we can change the
2285 // speculative type to non null ptr.
2286 if (ptr_kind == ProfileAlwaysNull) {
2287 speculative = TypePtr::NULL_PTR;
2288 } else {
2289 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2290 const TypePtr* ptr = TypePtr::NOTNULL;
2291 if (speculative != NULL) {
2292 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2293 } else {
2294 speculative = ptr;
2295 }
2296 }
2297 }
2298
2299 if (speculative != current_type->speculative()) {
2300 // Build a type with a speculative type (what we think we know
2301 // about the type but will need a guard when we use it)
2302 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2303 // We're changing the type, we need a new CheckCast node to carry
2304 // the new type. The new type depends on the control: what
2305 // profiling tells us is only valid from here as far as we can
2306 // tell.
2307 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2308 cast = _gvn.transform(cast);
2309 replace_in_map(n, cast);
2310 n = cast;
2311 }
2312
2313 return n;
2314 }
2315
2316 /**
2317 * Record profiling data from receiver profiling at an invoke with the
2318 * type system so that it can propagate it (speculation)
2319 *
2320 * @param n receiver node
2321 *
2322 * @return node with improved type
2323 */
2324 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2325 if (!UseTypeSpeculation) {
2326 return n;
2327 }
2328 ciKlass* exact_kls = profile_has_unique_klass();
2329 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2330 if ((java_bc() == Bytecodes::_checkcast ||
2331 java_bc() == Bytecodes::_instanceof ||
2332 java_bc() == Bytecodes::_aastore) &&
2333 method()->method_data()->is_mature()) {
2334 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2335 if (data != NULL) {
2336 if (java_bc() == Bytecodes::_aastore) {
2337 ciKlass* array_type = NULL;
2338 ciKlass* element_type = NULL;
2339 ProfilePtrKind element_ptr = ProfileMaybeNull;
2340 bool flat_array = true;
2341 bool null_free_array = true;
2342 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2343 exact_kls = element_type;
2344 ptr_kind = element_ptr;
2345 } else {
2346 if (!data->as_BitData()->null_seen()) {
2347 ptr_kind = ProfileNeverNull;
2348 } else {
2349 assert(data->is_ReceiverTypeData(), "bad profile data type");
2350 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2351 uint i = 0;
2352 for (; i < call->row_limit(); i++) {
2353 ciKlass* receiver = call->receiver(i);
2354 if (receiver != NULL) {
2355 break;
2356 }
2357 }
2358 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2359 }
2360 }
2361 }
2362 }
2363 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2364 }
2365
2366 /**
2367 * Record profiling data from argument profiling at an invoke with the
2368 * type system so that it can propagate it (speculation)
2369 *
2370 * @param dest_method target method for the call
2371 * @param bc what invoke bytecode is this?
2372 */
2373 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2374 if (!UseTypeSpeculation) {
2375 return;
2376 }
2377 const TypeFunc* tf = TypeFunc::make(dest_method);
2378 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2379 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2380 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2381 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2382 if (is_reference_type(targ->basic_type())) {
2383 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2384 ciKlass* better_type = NULL;
2385 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2386 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2387 }
2388 i++;
2389 }
2390 }
2391 }
2392
2393 /**
2394 * Record profiling data from parameter profiling at an invoke with
2395 * the type system so that it can propagate it (speculation)
2396 */
2397 void GraphKit::record_profiled_parameters_for_speculation() {
2398 if (!UseTypeSpeculation) {
2399 return;
2400 }
2401 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2432 if (Matcher::strict_fp_requires_explicit_rounding) {
2433 // If a strict caller invokes a non-strict callee, round a double result.
2434 // A non-strict method may return a double value which has an extended exponent,
2435 // but this must not be visible in a caller which is strict.
2436 BasicType result_type = dest_method->return_type()->basic_type();
2437 assert(method() != NULL, "must have caller context");
2438 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2439 // Destination method's return value is on top of stack
2440 // dstore_rounding() does gvn.transform
2441 Node *result = pop_pair();
2442 result = dstore_rounding(result);
2443 push_pair(result);
2444 }
2445 }
2446 }
2447
2448 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2449 if (Matcher::strict_fp_requires_explicit_rounding) {
2450 // (Note: TypeFunc::make has a cache that makes this fast.)
2451 const TypeFunc* tf = TypeFunc::make(dest_method);
2452 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2453 for (int j = 0; j < nargs; j++) {
2454 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2455 if (targ->basic_type() == T_DOUBLE) {
2456 // If any parameters are doubles, they must be rounded before
2457 // the call, dstore_rounding does gvn.transform
2458 Node *arg = argument(j);
2459 arg = dstore_rounding(arg);
2460 set_argument(j, arg);
2461 }
2462 }
2463 }
2464 }
2465
2466 // rounding for strict float precision conformance
2467 Node* GraphKit::precision_rounding(Node* n) {
2468 if (Matcher::strict_fp_requires_explicit_rounding) {
2469 #ifdef IA32
2470 if (_method->flags().is_strict() && UseSSE == 0) {
2471 return _gvn.transform(new RoundFloatNode(0, n));
2472 }
2473 #else
2474 Unimplemented();
2914
2915 // Now do a linear scan of the secondary super-klass array. Again, no real
2916 // performance impact (too rare) but it's gotta be done.
2917 // Since the code is rarely used, there is no penalty for moving it
2918 // out of line, and it can only improve I-cache density.
2919 // The decision to inline or out-of-line this final check is platform
2920 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2921 Node* psc = gvn.transform(
2922 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2923
2924 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2925 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2926 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2927
2928 // Return false path; set default control to true path.
2929 *ctrl = gvn.transform(r_ok_subtype);
2930 return gvn.transform(r_not_subtype);
2931 }
2932
2933 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2934 const Type* sub_t = _gvn.type(obj_or_subklass);
2935 if (sub_t->isa_inlinetype()) {
2936 obj_or_subklass = makecon(TypeKlassPtr::make(sub_t->inline_klass()));
2937 }
2938 if (ExpandSubTypeCheckAtParseTime) {
2939 MergeMemNode* mem = merged_memory();
2940 Node* ctrl = control();
2941 Node* subklass = obj_or_subklass;
2942 if (!sub_t->isa_klassptr()) {
2943 subklass = load_object_klass(obj_or_subklass);
2944 }
2945 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2946 set_control(ctrl);
2947 return n;
2948 }
2949
2950 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2951 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2952 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2953 set_control(_gvn.transform(new IfTrueNode(iff)));
2954 return _gvn.transform(new IfFalseNode(iff));
2955 }
2956
2957 // Profile-driven exact type check:
2958 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2959 float prob, Node* *casted_receiver) {
2960 Node* fail = top();
2961 const Type* rec_t = _gvn.type(receiver);
2962 if (false && rec_t->isa_inlinetype()) {
2963 if (klass->equals(rec_t->inline_klass())) {
2964 (*casted_receiver) = receiver; // Always passes
2965 } else {
2966 (*casted_receiver) = top(); // Always fails
2967 fail = control();
2968 set_control(top());
2969 }
2970 return fail;
2971 }
2972 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2973 Node* recv_klass = load_object_klass(receiver);
2974 fail = type_check(recv_klass, tklass, prob);
2975 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2976 assert(recv_xtype->klass_is_exact(), "");
2977
2978 // Subsume downstream occurrences of receiver with a cast to
2979 // recv_xtype, since now we know what the type will be.
2980 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2981 Node* res = _gvn.transform(cast);
2982 if (recv_xtype->is_inlinetypeptr() && recv_xtype->inline_klass()->is_scalarizable()) {
2983 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
2984 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
2985 }
2986
2987 (*casted_receiver) = res;
2988 // (User must make the replace_in_map call.)
2989
2990 return fail;
2991 }
2992
2993 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2994 float prob) {
2995 Node* want_klass = makecon(tklass);
2996 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2997 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2998 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2999 set_control( _gvn.transform( new IfTrueNode (iff)));
3000 Node* fail = _gvn.transform( new IfFalseNode(iff));
3001 return fail;
3002 }
3003
3004 //------------------------------subtype_check_receiver-------------------------
3005 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3006 Node** casted_receiver) {
3007 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3008 Node* want_klass = makecon(tklass);
3009
3010 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3011
3012 // Cast receiver after successful check
3013 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3014 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3015 (*casted_receiver) = _gvn.transform(cast);
3016
3017 return slow_ctl;
3018 }
3019
3020 //------------------------------seems_never_null-------------------------------
3021 // Use null_seen information if it is available from the profile.
3022 // If we see an unexpected null at a type check we record it and force a
3023 // recompile; the offending check will be recompiled to handle NULLs.
3024 // If we see several offending BCIs, then all checks in the
3025 // method will be recompiled.
3026 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3027 speculating = !_gvn.type(obj)->speculative_maybe_null();
3028 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3029 if (UncommonNullCast // Cutout for this technique
3030 && obj != null() // And not the -Xcomp stupid case?
3031 && !too_many_traps(reason)
3032 ) {
3033 if (speculating) {
3034 return true;
3035 }
3036 if (data == NULL)
3037 // Edge case: no mature data. Be optimistic here.
3038 return true;
3039 // If the profile has not seen a null, assume it won't happen.
3040 assert(java_bc() == Bytecodes::_checkcast ||
3041 java_bc() == Bytecodes::_instanceof ||
3042 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3043 if (java_bc() == Bytecodes::_aastore) {
3044 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull;
3045 }
3046 return !data->as_BitData()->null_seen();
3047 }
3048 speculating = false;
3049 return false;
3050 }
3051
3052 void GraphKit::guard_klass_being_initialized(Node* klass) {
3053 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3054 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3055 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3056 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3057 T_BYTE, MemNode::unordered);
3058 init_state = _gvn.transform(init_state);
3059
3060 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3061
3062 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3063 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3064
3065 { BuildCutout unless(this, tst, PROB_MAX);
3105
3106 //------------------------maybe_cast_profiled_receiver-------------------------
3107 // If the profile has seen exactly one type, narrow to exactly that type.
3108 // Subsequent type checks will always fold up.
3109 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3110 ciKlass* require_klass,
3111 ciKlass* spec_klass,
3112 bool safe_for_replace) {
3113 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3114
3115 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3116
3117 // Make sure we haven't already deoptimized from this tactic.
3118 if (too_many_traps_or_recompiles(reason))
3119 return NULL;
3120
3121 // (No, this isn't a call, but it's enough like a virtual call
3122 // to use the same ciMethod accessor to get the profile info...)
3123 // If we have a speculative type use it instead of profiling (which
3124 // may not help us)
3125 ciKlass* exact_kls = spec_klass;
3126 if (exact_kls == NULL) {
3127 if (java_bc() == Bytecodes::_aastore) {
3128 ciKlass* array_type = NULL;
3129 ciKlass* element_type = NULL;
3130 ProfilePtrKind element_ptr = ProfileMaybeNull;
3131 bool flat_array = true;
3132 bool null_free_array = true;
3133 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3134 exact_kls = element_type;
3135 } else {
3136 exact_kls = profile_has_unique_klass();
3137 }
3138 }
3139 if (exact_kls != NULL) {// no cast failures here
3140 if (require_klass == NULL ||
3141 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
3142 // If we narrow the type to match what the type profile sees or
3143 // the speculative type, we can then remove the rest of the
3144 // cast.
3145 // This is a win, even if the exact_kls is very specific,
3146 // because downstream operations, such as method calls,
3147 // will often benefit from the sharper type.
3148 Node* exact_obj = not_null_obj; // will get updated in place...
3149 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3150 &exact_obj);
3151 { PreserveJVMState pjvms(this);
3152 set_control(slow_ctl);
3153 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3154 }
3155 if (safe_for_replace) {
3156 replace_in_map(not_null_obj, exact_obj);
3157 }
3158 return exact_obj;
3223 // and the reflective instance-of call.
3224 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3225 kill_dead_locals(); // Benefit all the uncommon traps
3226 assert( !stopped(), "dead parse path should be checked in callers" );
3227 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3228 "must check for not-null not-dead klass in callers");
3229
3230 // Make the merge point
3231 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3232 RegionNode* region = new RegionNode(PATH_LIMIT);
3233 Node* phi = new PhiNode(region, TypeInt::BOOL);
3234 C->set_has_split_ifs(true); // Has chance for split-if optimization
3235
3236 ciProfileData* data = NULL;
3237 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3238 data = method()->method_data()->bci_to_data(bci());
3239 }
3240 bool speculative_not_null = false;
3241 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3242 && seems_never_null(obj, data, speculative_not_null));
3243 bool is_value = obj->is_InlineType();
3244
3245 // Null check; get casted pointer; set region slot 3
3246 Node* null_ctl = top();
3247 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3248
3249 // If not_null_obj is dead, only null-path is taken
3250 if (stopped()) { // Doing instance-of on a NULL?
3251 set_control(null_ctl);
3252 return intcon(0);
3253 }
3254 region->init_req(_null_path, null_ctl);
3255 phi ->init_req(_null_path, intcon(0)); // Set null path value
3256 if (null_ctl == top()) {
3257 // Do this eagerly, so that pattern matches like is_diamond_phi
3258 // will work even during parsing.
3259 assert(_null_path == PATH_LIMIT-1, "delete last");
3260 region->del_req(_null_path);
3261 phi ->del_req(_null_path);
3262 }
3263
3264 // Do we know the type check always succeed?
3265 if (!is_value) {
3266 bool known_statically = false;
3267 if (_gvn.type(superklass)->singleton()) {
3268 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3269 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3270 if (subk != NULL && subk->is_loaded()) {
3271 int static_res = C->static_subtype_check(superk, subk);
3272 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3273 }
3274 }
3275
3276 if (!known_statically) {
3277 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3278 // We may not have profiling here or it may not help us. If we
3279 // have a speculative type use it to perform an exact cast.
3280 ciKlass* spec_obj_type = obj_type->speculative_type();
3281 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3282 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3283 if (stopped()) { // Profile disagrees with this path.
3284 set_control(null_ctl); // Null is the only remaining possibility.
3285 return intcon(0);
3286 }
3287 if (cast_obj != NULL &&
3288 // A value that's sometimes null is not something we can optimize well
3289 !(cast_obj->is_InlineType() && null_ctl != top())) {
3290 not_null_obj = cast_obj;
3291 is_value = not_null_obj->is_InlineType();
3292 }
3293 }
3294 }
3295 }
3296
3297 // Generate the subtype check
3298 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3299
3300 // Plug in the success path to the general merge in slot 1.
3301 region->init_req(_obj_path, control());
3302 phi ->init_req(_obj_path, intcon(1));
3303
3304 // Plug in the failing path to the general merge in slot 2.
3305 region->init_req(_fail_path, not_subtype_ctrl);
3306 phi ->init_req(_fail_path, intcon(0));
3307
3308 // Return final merged results
3309 set_control( _gvn.transform(region) );
3310 record_for_igvn(region);
3311
3312 // If we know the type check always succeeds then we don't use the
3313 // profiling data at this bytecode. Don't lose it, feed it to the
3314 // type system as a speculative type.
3315 if (safe_for_replace && !is_value) {
3316 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3317 replace_in_map(obj, casted_obj);
3318 }
3319
3320 return _gvn.transform(phi);
3321 }
3322
3323 //-------------------------------gen_checkcast---------------------------------
3324 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3325 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3326 // uncommon-trap paths work. Adjust stack after this call.
3327 // If failure_control is supplied and not null, it is filled in with
3328 // the control edge for the cast failure. Otherwise, an appropriate
3329 // uncommon trap or exception is thrown.
3330 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control) {
3331 kill_dead_locals(); // Benefit all the uncommon traps
3332 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3333 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3334
3335 // Check if inline types are involved
3336 bool from_inline = obj->is_InlineType();
3337 bool to_inline = tk->klass()->is_inlinetype();
3338
3339 // Fast cutout: Check the case that the cast is vacuously true.
3340 // This detects the common cases where the test will short-circuit
3341 // away completely. We do this before we perform the null check,
3342 // because if the test is going to turn into zero code, we don't
3343 // want a residual null check left around. (Causes a slowdown,
3344 // for example, in some objArray manipulations, such as a[i]=a[j].)
3345 if (tk->singleton()) {
3346 ciKlass* klass = NULL;
3347 if (from_inline) {
3348 klass = _gvn.type(obj)->inline_klass();
3349 } else {
3350 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3351 if (objtp != NULL) {
3352 klass = objtp->klass();
3353 }
3354 }
3355 if (klass != NULL) {
3356 switch (C->static_subtype_check(tk->klass(), klass)) {
3357 case Compile::SSC_always_true:
3358 // If we know the type check always succeed then we don't use
3359 // the profiling data at this bytecode. Don't lose it, feed it
3360 // to the type system as a speculative type.
3361 if (!from_inline) {
3362 obj = record_profiled_receiver_for_speculation(obj);
3363 if (to_inline) {
3364 obj = null_check(obj);
3365 if (toop->inline_klass()->is_scalarizable()) {
3366 obj = InlineTypeNode::make_from_oop(this, obj, toop->inline_klass());
3367 }
3368 }
3369 }
3370 return obj;
3371 case Compile::SSC_always_false:
3372 if (from_inline || to_inline) {
3373 if (!from_inline) {
3374 null_check(obj);
3375 }
3376 // Inline type is never null. Always throw an exception.
3377 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3378 return top();
3379 } else {
3380 // It needs a null check because a null will *pass* the cast check.
3381 return null_assert(obj);
3382 }
3383 }
3384 }
3385 }
3386
3387 ciProfileData* data = NULL;
3388 bool safe_for_replace = false;
3389 if (failure_control == NULL) { // use MDO in regular case only
3390 assert(java_bc() == Bytecodes::_aastore ||
3391 java_bc() == Bytecodes::_checkcast,
3392 "interpreter profiles type checks only for these BCs");
3393 if (method()->method_data()->is_mature()) {
3394 data = method()->method_data()->bci_to_data(bci());
3395 }
3396 safe_for_replace = true;
3397 }
3398
3399 // Make the merge point
3400 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3401 RegionNode* region = new RegionNode(PATH_LIMIT);
3402 Node* phi = new PhiNode(region, toop);
3403 _gvn.set_type(region, Type::CONTROL);
3404 _gvn.set_type(phi, toop);
3405
3406 C->set_has_split_ifs(true); // Has chance for split-if optimization
3407
3408 // Use null-cast information if it is available
3409 bool speculative_not_null = false;
3410 bool never_see_null = ((failure_control == NULL) // regular case only
3411 && seems_never_null(obj, data, speculative_not_null));
3412
3413 // Null check; get casted pointer; set region slot 3
3414 Node* null_ctl = top();
3415 Node* not_null_obj = NULL;
3416 if (from_inline) {
3417 not_null_obj = obj;
3418 } else if (to_inline) {
3419 not_null_obj = null_check(obj);
3420 } else {
3421 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3422 }
3423
3424 // If not_null_obj is dead, only null-path is taken
3425 if (stopped()) { // Doing instance-of on a NULL?
3426 set_control(null_ctl);
3427 return null();
3428 }
3429 region->init_req(_null_path, null_ctl);
3430 phi ->init_req(_null_path, null()); // Set null path value
3431 if (null_ctl == top()) {
3432 // Do this eagerly, so that pattern matches like is_diamond_phi
3433 // will work even during parsing.
3434 assert(_null_path == PATH_LIMIT-1, "delete last");
3435 region->del_req(_null_path);
3436 phi ->del_req(_null_path);
3437 }
3438
3439 Node* cast_obj = NULL;
3440 if (!from_inline && tk->klass_is_exact()) {
3441 // The following optimization tries to statically cast the speculative type of the object
3442 // (for example obtained during profiling) to the type of the superklass and then do a
3443 // dynamic check that the type of the object is what we expect. To work correctly
3444 // for checkcast and aastore the type of superklass should be exact.
3445 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3446 // We may not have profiling here or it may not help us. If we have
3447 // a speculative type use it to perform an exact cast.
3448 ciKlass* spec_obj_type = obj_type->speculative_type();
3449 if (spec_obj_type != NULL || data != NULL) {
3450 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3451 if (cast_obj != NULL && cast_obj->is_InlineType()) {
3452 if (null_ctl != top()) {
3453 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3454 } else {
3455 return cast_obj;
3456 }
3457 }
3458 if (cast_obj != NULL) {
3459 if (failure_control != NULL) // failure is now impossible
3460 (*failure_control) = top();
3461 // adjust the type of the phi to the exact klass:
3462 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3463 }
3464 }
3465 }
3466
3467 if (cast_obj == NULL) {
3468 // Generate the subtype check
3469 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3470
3471 // Plug in success path into the merge
3472 cast_obj = from_inline ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3473 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3474 if (failure_control == NULL) {
3475 if (not_subtype_ctrl != top()) { // If failure is possible
3476 PreserveJVMState pjvms(this);
3477 set_control(not_subtype_ctrl);
3478 Node* obj_klass = NULL;
3479 if (from_inline) {
3480 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3481 } else {
3482 obj_klass = load_object_klass(not_null_obj);
3483 }
3484 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3485 }
3486 } else {
3487 (*failure_control) = not_subtype_ctrl;
3488 }
3489 }
3490
3491 region->init_req(_obj_path, control());
3492 phi ->init_req(_obj_path, cast_obj);
3493
3494 // A merge of NULL or Casted-NotNull obj
3495 Node* res = _gvn.transform(phi);
3496
3497 // Note I do NOT always 'replace_in_map(obj,result)' here.
3498 // if( tk->klass()->can_be_primary_super() )
3499 // This means that if I successfully store an Object into an array-of-String
3500 // I 'forget' that the Object is really now known to be a String. I have to
3501 // do this because we don't have true union types for interfaces - if I store
3502 // a Baz into an array-of-Interface and then tell the optimizer it's an
3503 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3504 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3505 // replace_in_map( obj, res );
3506
3507 // Return final merged results
3508 set_control( _gvn.transform(region) );
3509 record_for_igvn(region);
3510
3511 bool not_inline = !toop->can_be_inline_type();
3512 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array());
3513 if (EnableValhalla && not_flattened) {
3514 // Check if obj has been loaded from an array
3515 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3516 Node* array = NULL;
3517 if (obj->isa_Load()) {
3518 Node* address = obj->in(MemNode::Address);
3519 if (address->isa_AddP()) {
3520 array = address->as_AddP()->in(AddPNode::Base);
3521 }
3522 } else if (obj->is_Phi()) {
3523 Node* region = obj->in(0);
3524 // TODO make this more robust (see JDK-8231346)
3525 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) {
3526 IfNode* iff = region->in(2)->in(0)->isa_If();
3527 if (iff != NULL) {
3528 iff->is_non_flattened_array_check(&_gvn, &array);
3529 }
3530 }
3531 }
3532 if (array != NULL) {
3533 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3534 if (ary_t != NULL) {
3535 if (!ary_t->is_not_null_free() && not_inline) {
3536 // Casting array element to a non-inline-type, mark array as not null-free.
3537 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3538 replace_in_map(array, cast);
3539 } else if (!ary_t->is_not_flat()) {
3540 // Casting array element to a non-flattened type, mark array as not flat.
3541 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3542 replace_in_map(array, cast);
3543 }
3544 }
3545 }
3546 }
3547
3548 if (!from_inline) {
3549 res = record_profiled_receiver_for_speculation(res);
3550 if (to_inline && toop->inline_klass()->is_scalarizable()) {
3551 assert(!gvn().type(res)->maybe_null(), "Inline types are null-free");
3552 res = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3553 }
3554 }
3555 return res;
3556 }
3557
3558 // Check if 'obj' is an inline type by checking if it has the always_locked markWord pattern set.
3559 Node* GraphKit::is_inline_type(Node* obj) {
3560 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3561 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3562 Node* mask = _gvn.MakeConX(markWord::always_locked_pattern);
3563 Node* andx = _gvn.transform(new AndXNode(mark, mask));
3564 Node* cmp = _gvn.transform(new CmpXNode(andx, mask));
3565 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3566 }
3567
3568 // Check if 'ary' is a non-flattened array
3569 Node* GraphKit::is_non_flattened_array(Node* ary) {
3570 Node* kls = load_object_klass(ary);
3571 Node* tag = load_lh_array_tag(kls);
3572 Node* cmp = gen_lh_array_test(kls, Klass::_lh_array_tag_vt_value);
3573 return _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3574 }
3575
3576 // Check if 'ary' is a nullable array
3577 Node* GraphKit::is_nullable_array(Node* ary) {
3578 Node* kls = load_object_klass(ary);
3579 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3580 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3581 Node* null_free = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_null_free_shift)));
3582 null_free = _gvn.transform(new AndINode(null_free, intcon(Klass::_lh_null_free_mask)));
3583 Node* cmp = _gvn.transform(new CmpINode(null_free, intcon(0)));
3584 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3585 }
3586
3587 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3588 Node* GraphKit::gen_inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3589 const Type* val_t = _gvn.type(val);
3590 if (val->is_InlineType() || !TypePtr::NULL_PTR->higher_equal(val_t)) {
3591 return ary; // Never null
3592 }
3593 RegionNode* region = new RegionNode(3);
3594 Node* null_ctl = top();
3595 null_check_oop(val, &null_ctl);
3596 if (null_ctl != top()) {
3597 PreserveJVMState pjvms(this);
3598 set_control(null_ctl);
3599 {
3600 // Deoptimize if null-free array
3601 BuildCutout unless(this, is_nullable_array(ary), PROB_MAX);
3602 inc_sp(nargs);
3603 uncommon_trap(Deoptimization::Reason_null_check,
3604 Deoptimization::Action_none);
3605 }
3606 region->init_req(1, control());
3607 }
3608 region->init_req(2, control());
3609 set_control(_gvn.transform(region));
3610 record_for_igvn(region);
3611 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3612 if (val_t == TypePtr::NULL_PTR && !ary_t->is_not_null_free()) {
3613 // Since we were just successfully storing null, the array can't be null free.
3614 ary_t = ary_t->cast_to_not_null_free();
3615 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3616 if (safe_for_replace) {
3617 replace_in_map(ary, cast);
3618 }
3619 ary = cast;
3620 }
3621 return ary;
3622 }
3623
3624 Node* GraphKit::load_lh_array_tag(Node* kls) {
3625 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3626 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3627 return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3628 }
3629
3630 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) {
3631 Node* layout_val = load_lh_array_tag(kls);
3632 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value)));
3633 return cmp;
3634 }
3635
3636 //------------------------------next_monitor-----------------------------------
3637 // What number should be given to the next monitor?
3638 int GraphKit::next_monitor() {
3639 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3640 int next = current + C->sync_stack_slots();
3641 // Keep the toplevel high water mark current:
3642 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3643 return current;
3644 }
3645
3646 //------------------------------insert_mem_bar---------------------------------
3647 // Memory barrier to avoid floating things around
3648 // The membar serves as a pinch point between both control and all memory slices.
3649 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3650 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3651 mb->init_req(TypeFunc::Control, control());
3652 mb->init_req(TypeFunc::Memory, reset_memory());
3653 Node* membar = _gvn.transform(mb);
3681 }
3682 Node* membar = _gvn.transform(mb);
3683 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3684 if (alias_idx == Compile::AliasIdxBot) {
3685 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3686 } else {
3687 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3688 }
3689 return membar;
3690 }
3691
3692 //------------------------------shared_lock------------------------------------
3693 // Emit locking code.
3694 FastLockNode* GraphKit::shared_lock(Node* obj) {
3695 // bci is either a monitorenter bc or InvocationEntryBci
3696 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3697 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3698
3699 if( !GenerateSynchronizationCode )
3700 return NULL; // Not locking things?
3701
3702 if (stopped()) // Dead monitor?
3703 return NULL;
3704
3705 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3706
3707 // Box the stack location
3708 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3709 Node* mem = reset_memory();
3710
3711 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3712 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3713 // Create the counters for this fast lock.
3714 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3715 }
3716
3717 // Create the rtm counters for this fast lock if needed.
3718 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3719
3720 // Add monitor to debug info for the slow path. If we block inside the
3721 // slow path and de-opt, we need the monitor hanging around
3754 }
3755 #endif
3756
3757 return flock;
3758 }
3759
3760
3761 //------------------------------shared_unlock----------------------------------
3762 // Emit unlocking code.
3763 void GraphKit::shared_unlock(Node* box, Node* obj) {
3764 // bci is either a monitorenter bc or InvocationEntryBci
3765 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3766 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3767
3768 if( !GenerateSynchronizationCode )
3769 return;
3770 if (stopped()) { // Dead monitor?
3771 map()->pop_monitor(); // Kill monitor from debug info
3772 return;
3773 }
3774 assert(!obj->is_InlineTypeBase(), "should not unlock on inline type");
3775
3776 // Memory barrier to avoid floating things down past the locked region
3777 insert_mem_bar(Op_MemBarReleaseLock);
3778
3779 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3780 UnlockNode *unlock = new UnlockNode(C, tf);
3781 #ifdef ASSERT
3782 unlock->set_dbg_jvms(sync_jvms());
3783 #endif
3784 uint raw_idx = Compile::AliasIdxRaw;
3785 unlock->init_req( TypeFunc::Control, control() );
3786 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3787 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3788 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3789 unlock->init_req( TypeFunc::ReturnAdr, top() );
3790
3791 unlock->init_req(TypeFunc::Parms + 0, obj);
3792 unlock->init_req(TypeFunc::Parms + 1, box);
3793 unlock = _gvn.transform(unlock)->as_Unlock();
3794
3795 Node* mem = reset_memory();
3796
3797 // unlock has no side-effects, sets few values
3798 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3799
3800 // Kill monitor from debug info
3801 map()->pop_monitor( );
3802 }
3803
3804 //-------------------------------get_layout_helper-----------------------------
3805 // If the given klass is a constant or known to be an array,
3806 // fetch the constant layout helper value into constant_value
3807 // and return (Node*)NULL. Otherwise, load the non-constant
3808 // layout helper value, and return the node which represents it.
3809 // This two-faced routine is useful because allocation sites
3810 // almost always feature constant types.
3811 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3812 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3813 if (!StressReflectiveCode && inst_klass != NULL) {
3814 ciKlass* klass = inst_klass->klass();
3815 assert(klass != NULL, "klass should not be NULL");
3816 bool xklass = inst_klass->klass_is_exact();
3817 bool can_be_flattened = false;
3818 if (UseFlatArray && klass->is_obj_array_klass()) {
3819 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
3820 can_be_flattened = elem->can_be_inline_klass() && (!elem->is_inlinetype() || elem->flatten_array());
3821 }
3822 if (xklass || (klass->is_array_klass() && !can_be_flattened)) {
3823 jint lhelper = klass->layout_helper();
3824 if (lhelper != Klass::_lh_neutral_value) {
3825 constant_value = lhelper;
3826 return (Node*) NULL;
3827 }
3828 }
3829 }
3830 constant_value = Klass::_lh_neutral_value; // put in a known value
3831 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3832 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3833 }
3834
3835 // We just put in an allocate/initialize with a big raw-memory effect.
3836 // Hook selected additional alias categories on the initialization.
3837 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3838 MergeMemNode* init_in_merge,
3839 Node* init_out_raw) {
3840 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3841 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3842
3864
3865 // a normal slow-call doesn't change i_o, but an allocation does
3866 // we create a separate i_o projection for the normal control path
3867 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3868 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3869
3870 // put in an initialization barrier
3871 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3872 rawoop)->as_Initialize();
3873 assert(alloc->initialization() == init, "2-way macro link must work");
3874 assert(init ->allocation() == alloc, "2-way macro link must work");
3875 {
3876 // Extract memory strands which may participate in the new object's
3877 // initialization, and source them from the new InitializeNode.
3878 // This will allow us to observe initializations when they occur,
3879 // and link them properly (as a group) to the InitializeNode.
3880 assert(init->in(InitializeNode::Memory) == malloc, "");
3881 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3882 init->set_req(InitializeNode::Memory, minit_in);
3883 record_for_igvn(minit_in); // fold it up later, if possible
3884 _gvn.set_type(minit_in, Type::MEMORY);
3885 Node* minit_out = memory(rawidx);
3886 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3887 // Add an edge in the MergeMem for the header fields so an access
3888 // to one of those has correct memory state
3889 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3890 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3891 if (oop_type->isa_aryptr()) {
3892 const TypeAryPtr* arytype = oop_type->is_aryptr();
3893 if (arytype->klass()->is_flat_array_klass()) {
3894 // Initially all flattened array accesses share a single slice
3895 // but that changes after parsing. Prepare the memory graph so
3896 // it can optimize flattened array accesses properly once they
3897 // don't share a single slice.
3898 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
3899 C->set_flattened_accesses_share_alias(false);
3900 ciFlatArrayKlass* vak = arytype->klass()->as_flat_array_klass();
3901 ciInlineKlass* vk = vak->element_klass()->as_inline_klass();
3902 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3903 ciField* field = vk->nonstatic_field_at(i);
3904 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3905 continue; // do not bother to track really large numbers of fields
3906 int off_in_vt = field->offset() - vk->first_field_offset();
3907 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3908 int fieldidx = C->get_alias_index(adr_type, true);
3909 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3910 }
3911 C->set_flattened_accesses_share_alias(true);
3912 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
3913 } else {
3914 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3915 int elemidx = C->get_alias_index(telemref);
3916 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3917 }
3918 } else if (oop_type->isa_instptr()) {
3919 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3920 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3921 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3922 ciField* field = ik->nonstatic_field_at(i);
3923 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3924 continue; // do not bother to track really large numbers of fields
3925 // Find (or create) the alias category for this field:
3926 int fieldidx = C->alias_type(field)->index();
3927 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3928 }
3929 }
3930 }
3931
3932 // Cast raw oop to the real thing...
3933 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3934 javaoop = _gvn.transform(javaoop);
3935 C->set_recent_alloc(control(), javaoop);
3936 assert(just_allocated_object(control()) == javaoop, "just allocated");
3937
3938 #ifdef ASSERT
3939 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3950 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3951 }
3952 }
3953 #endif //ASSERT
3954
3955 return javaoop;
3956 }
3957
3958 //---------------------------new_instance--------------------------------------
3959 // This routine takes a klass_node which may be constant (for a static type)
3960 // or may be non-constant (for reflective code). It will work equally well
3961 // for either, and the graph will fold nicely if the optimizer later reduces
3962 // the type to a constant.
3963 // The optional arguments are for specialized use by intrinsics:
3964 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3965 // - If 'return_size_val', report the the total object size to the caller.
3966 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3967 Node* GraphKit::new_instance(Node* klass_node,
3968 Node* extra_slow_test,
3969 Node* *return_size_val,
3970 bool deoptimize_on_exception,
3971 InlineTypeBaseNode* inline_type_node) {
3972 // Compute size in doublewords
3973 // The size is always an integral number of doublewords, represented
3974 // as a positive bytewise size stored in the klass's layout_helper.
3975 // The layout_helper also encodes (in a low bit) the need for a slow path.
3976 jint layout_con = Klass::_lh_neutral_value;
3977 Node* layout_val = get_layout_helper(klass_node, layout_con);
3978 bool layout_is_con = (layout_val == NULL);
3979
3980 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3981 // Generate the initial go-slow test. It's either ALWAYS (return a
3982 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3983 // case) a computed value derived from the layout_helper.
3984 Node* initial_slow_test = NULL;
3985 if (layout_is_con) {
3986 assert(!StressReflectiveCode, "stress mode does not use these paths");
3987 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3988 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3989 } else { // reflective case
3990 // This reflective path is used by Unsafe.allocateInstance.
3991 // (It may be stress-tested by specifying StressReflectiveCode.)
3992 // Basically, we want to get into the VM is there's an illegal argument.
3993 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3994 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3995 if (extra_slow_test != intcon(0)) {
3996 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3997 }
3998 // (Macro-expander will further convert this to a Bool, if necessary.)
4009
4010 // Clear the low bits to extract layout_helper_size_in_bytes:
4011 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4012 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4013 size = _gvn.transform( new AndXNode(size, mask) );
4014 }
4015 if (return_size_val != NULL) {
4016 (*return_size_val) = size;
4017 }
4018
4019 // This is a precise notnull oop of the klass.
4020 // (Actually, it need not be precise if this is a reflective allocation.)
4021 // It's what we cast the result to.
4022 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4023 if (!tklass) tklass = TypeKlassPtr::OBJECT;
4024 const TypeOopPtr* oop_type = tklass->as_instance_type();
4025
4026 // Now generate allocation code
4027
4028 // The entire memory state is needed for slow path of the allocation
4029 // since GC and deoptimization can happen.
4030 Node *mem = reset_memory();
4031 set_all_memory(mem); // Create new memory state
4032
4033 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4034 control(), mem, i_o(),
4035 size, klass_node,
4036 initial_slow_test, inline_type_node);
4037
4038 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4039 }
4040
4041 // With compressed oops, the 64 bit init value for non flattened value
4042 // arrays is built from 2 32 bit compressed oops
4043 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) {
4044 Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value));
4045 Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32)));
4046 return kit.gvn().transform(new OrLNode(lower, upper));
4047 }
4048
4049 //-------------------------------new_array-------------------------------------
4050 // helper for newarray and anewarray
4051 // The 'length' parameter is (obviously) the length of the array.
4052 // See comments on new_instance for the meaning of the other arguments.
4053 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4054 Node* length, // number of array elements
4055 int nargs, // number of arguments to push back for uncommon trap
4056 Node* *return_size_val,
4057 bool deoptimize_on_exception) {
4058 jint layout_con = Klass::_lh_neutral_value;
4059 Node* layout_val = get_layout_helper(klass_node, layout_con);
4060 bool layout_is_con = (layout_val == NULL);
4061
4062 if (!layout_is_con && !StressReflectiveCode &&
4063 !too_many_traps(Deoptimization::Reason_class_check)) {
4064 // This is a reflective array creation site.
4065 // Optimistically assume that it is a subtype of Object[],
4066 // so that we can fold up all the address arithmetic.
4067 layout_con = Klass::array_layout_helper(T_OBJECT);
4068 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4069 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4070 { BuildCutout unless(this, bol_lh, PROB_MAX);
4071 inc_sp(nargs);
4072 uncommon_trap(Deoptimization::Reason_class_check,
4073 Deoptimization::Action_maybe_recompile);
4074 }
4075 layout_val = NULL;
4076 layout_is_con = true;
4077 }
4078
4079 // Generate the initial go-slow test. Make sure we do not overflow
4080 // if length is huge (near 2Gig) or negative! We do not need
4081 // exact double-words here, just a close approximation of needed
4082 // double-words. We can't add any offset or rounding bits, lest we
4083 // take a size -1 of bytes and make it positive. Use an unsigned
4084 // compare, so negative sizes look hugely positive.
4085 int fast_size_limit = FastAllocateSizeLimit;
4086 if (layout_is_con) {
4087 assert(!StressReflectiveCode, "stress mode does not use these paths");
4088 // Increase the size limit if we have exact knowledge of array type.
4089 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4090 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4091 }
4092
4093 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4094 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4095
4096 // --- Size Computation ---
4097 // array_size = round_to_heap(array_header + (length << elem_shift));
4098 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4099 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4100 // The rounding mask is strength-reduced, if possible.
4101 int round_mask = MinObjAlignmentInBytes - 1;
4102 Node* header_size = NULL;
4103 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4104 // (T_BYTE has the weakest alignment and size restrictions...)
4105 if (layout_is_con) {
4106 int hsize = Klass::layout_helper_header_size(layout_con);
4107 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4108 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4109 if ((round_mask & ~right_n_bits(eshift)) == 0)
4110 round_mask = 0; // strength-reduce it if it goes away completely
4111 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4112 assert(header_size_min <= hsize, "generic minimum is smallest");
4113 header_size_min = hsize;
4114 header_size = intcon(hsize + round_mask);
4115 } else {
4116 Node* hss = intcon(Klass::_lh_header_size_shift);
4117 Node* hsm = intcon(Klass::_lh_header_size_mask);
4118 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4119 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4120 Node* mask = intcon(round_mask);
4121 header_size = _gvn.transform( new AddINode(hsize, mask) );
4122 }
4123
4124 Node* elem_shift = NULL;
4125 if (layout_is_con) {
4126 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4127 if (eshift != 0)
4128 elem_shift = intcon(eshift);
4129 } else {
4130 // There is no need to mask or shift this value.
4131 // The semantics of LShiftINode include an implicit mask to 0x1F.
4175 // places, one where the length is sharply limited, and the other
4176 // after a successful allocation.
4177 Node* abody = lengthx;
4178 if (elem_shift != NULL)
4179 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4180 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4181 if (round_mask != 0) {
4182 Node* mask = MakeConX(~round_mask);
4183 size = _gvn.transform( new AndXNode(size, mask) );
4184 }
4185 // else if round_mask == 0, the size computation is self-rounding
4186
4187 if (return_size_val != NULL) {
4188 // This is the size
4189 (*return_size_val) = size;
4190 }
4191
4192 // Now generate allocation code
4193
4194 // The entire memory state is needed for slow path of the allocation
4195 // since GC and deoptimization can happen.
4196 Node *mem = reset_memory();
4197 set_all_memory(mem); // Create new memory state
4198
4199 if (initial_slow_test->is_Bool()) {
4200 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4201 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4202 }
4203
4204 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4205 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4206 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4207
4208 // Inline type array variants:
4209 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue$ref")
4210 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue$val")
4211 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue$val")
4212 // Check if array is a null-free, non-flattened inline type array
4213 // that needs to be initialized with the default inline type.
4214 Node* default_value = NULL;
4215 Node* raw_default_value = NULL;
4216 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4217 // Array type is known
4218 ciKlass* elem_klass = ary_ptr->klass()->as_array_klass()->element_klass();
4219 if (elem_klass != NULL && elem_klass->is_inlinetype()) {
4220 ciInlineKlass* vk = elem_klass->as_inline_klass();
4221 if (!vk->flatten_array()) {
4222 default_value = InlineTypeNode::default_oop(gvn(), vk);
4223 if (UseCompressedOops) {
4224 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4225 raw_default_value = raw_default_for_coops(default_value, *this);
4226 } else {
4227 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4228 }
4229 }
4230 }
4231 } else if (ary_klass->klass()->can_be_inline_array_klass()) {
4232 // Array type is not known, add runtime checks
4233 assert(!ary_klass->klass_is_exact(), "unexpected exact type");
4234 Node* r = new RegionNode(4);
4235 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4236
4237 // Check if array is an object array
4238 Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_obj_value);
4239 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4240 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4241
4242 // Not an object array, initialize with all zero
4243 r->init_req(1, _gvn.transform(new IfFalseNode(iff)));
4244 default_value->init_req(1, null());
4245
4246 // Object array, check if null-free
4247 set_control(_gvn.transform(new IfTrueNode(iff)));
4248 Node* lhp = basic_plus_adr(klass_node, in_bytes(Klass::layout_helper_offset()));
4249 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
4250 Node* null_free = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_null_free_shift)));
4251 null_free = _gvn.transform(new AndINode(null_free, intcon(Klass::_lh_null_free_mask)));
4252 cmp = _gvn.transform(new CmpINode(null_free, intcon(0)));
4253 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
4254 iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4255
4256 // Not null-free, initialize with all zero
4257 r->init_req(2, _gvn.transform(new IfFalseNode(iff)));
4258 default_value->init_req(2, null());
4259
4260 // Null-free, non-flattened inline type array, initialize with the default value
4261 set_control(_gvn.transform(new IfTrueNode(iff)));
4262 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4263 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4264 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()));
4265 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4266 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset()));
4267 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4268 Node* elem_mirror = load_mirror_from_klass(eklass);
4269 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4270 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP);
4271 r->init_req(3, control());
4272 default_value->init_req(3, val);
4273
4274 set_control(_gvn.transform(r));
4275 default_value = _gvn.transform(default_value);
4276 if (UseCompressedOops) {
4277 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4278 raw_default_value = raw_default_for_coops(default_value, *this);
4279 } else {
4280 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4281 }
4282 }
4283
4284 // Create the AllocateArrayNode and its result projections
4285 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4286 control(), mem, i_o(),
4287 size, klass_node,
4288 initial_slow_test,
4289 length, default_value,
4290 raw_default_value);
4291
4292 // Cast to correct type. Note that the klass_node may be constant or not,
4293 // and in the latter case the actual array type will be inexact also.
4294 // (This happens via a non-constant argument to inline_native_newArray.)
4295 // In any case, the value of klass_node provides the desired array type.
4296 const TypeInt* length_type = _gvn.find_int_type(length);
4297 if (ary_type->isa_aryptr() && length_type != NULL) {
4298 // Try to get a better type than POS for the size
4299 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4300 }
4301
4302 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4303
4304 // Cast length on remaining path to be as narrow as possible
4305 if (map()->find_edge(length) >= 0) {
4306 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
4307 if (ccast != length) {
4308 _gvn.set_type_bottom(ccast);
4309 record_for_igvn(ccast);
4310 replace_in_map(length, ccast);
4311 }
4312 }
4313
4314 return javaoop;
4315 }
4316
4434 set_all_memory(ideal.merged_memory());
4435 set_i_o(ideal.i_o());
4436 set_control(ideal.ctrl());
4437 }
4438
4439 void GraphKit::final_sync(IdealKit& ideal) {
4440 // Final sync IdealKit and graphKit.
4441 sync_kit(ideal);
4442 }
4443
4444 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4445 Node* len = load_array_length(load_String_value(str, set_ctrl));
4446 Node* coder = load_String_coder(str, set_ctrl);
4447 // Divide length by 2 if coder is UTF16
4448 return _gvn.transform(new RShiftINode(len, coder));
4449 }
4450
4451 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4452 int value_offset = java_lang_String::value_offset();
4453 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4454 false, NULL, Type::Offset(0));
4455 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4456 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4457 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4458 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4459 Node* p = basic_plus_adr(str, str, value_offset);
4460 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4461 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4462 return load;
4463 }
4464
4465 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4466 if (!CompactStrings) {
4467 return intcon(java_lang_String::CODER_UTF16);
4468 }
4469 int coder_offset = java_lang_String::coder_offset();
4470 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4471 false, NULL, Type::Offset(0));
4472 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4473
4474 Node* p = basic_plus_adr(str, str, coder_offset);
4475 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4476 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4477 return load;
4478 }
4479
4480 void GraphKit::store_String_value(Node* str, Node* value) {
4481 int value_offset = java_lang_String::value_offset();
4482 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4483 false, NULL, Type::Offset(0));
4484 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4485
4486 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4487 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4488 }
4489
4490 void GraphKit::store_String_coder(Node* str, Node* value) {
4491 int coder_offset = java_lang_String::coder_offset();
4492 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4493 false, NULL, Type::Offset(0));
4494 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4495
4496 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4497 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4498 }
4499
4500 // Capture src and dst memory state with a MergeMemNode
4501 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4502 if (src_type == dst_type) {
4503 // Types are equal, we don't need a MergeMemNode
4504 return memory(src_type);
4505 }
4506 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4507 record_for_igvn(merge); // fold it up later, if possible
4508 int src_idx = C->get_alias_index(src_type);
4509 int dst_idx = C->get_alias_index(dst_type);
4510 merge->set_memory_at(src_idx, memory(src_idx));
4511 merge->set_memory_at(dst_idx, memory(dst_idx));
4512 return merge;
4513 }
4584 i_char->init_req(2, AddI(i_char, intcon(2)));
4585
4586 set_control(IfFalse(iff));
4587 set_memory(st, TypeAryPtr::BYTES);
4588 }
4589
4590 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4591 if (!field->is_constant()) {
4592 return NULL; // Field not marked as constant.
4593 }
4594 ciInstance* holder = NULL;
4595 if (!field->is_static()) {
4596 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4597 if (const_oop != NULL && const_oop->is_instance()) {
4598 holder = const_oop->as_instance();
4599 }
4600 }
4601 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4602 /*is_unsigned_load=*/false);
4603 if (con_type != NULL) {
4604 Node* con = makecon(con_type);
4605 assert(!field->type()->is_inlinetype() || (field->is_static() && !con_type->is_zero_type()), "sanity");
4606 // Check type of constant which might be more precise
4607 if (con_type->is_inlinetypeptr() && con_type->inline_klass()->is_scalarizable()) {
4608 // Load inline type from constant oop
4609 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4610 }
4611 return con;
4612 }
4613 return NULL;
4614 }
4615
4616 //---------------------------load_mirror_from_klass----------------------------
4617 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4618 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4619 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4620 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4621 // mirror = ((OopHandle)mirror)->resolve();
4622 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4623 }
|