16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/assembler.inline.hpp"
26 #include "code/aotCodeCache.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compilerDirectives.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "gc/shared/barrierSet.hpp"
35 #include "gc/shared/c2/barrierSetC2.hpp"
36 #include "memory/allocation.hpp"
37 #include "opto/ad.hpp"
38 #include "opto/block.hpp"
39 #include "opto/c2_MacroAssembler.hpp"
40 #include "opto/c2compiler.hpp"
41 #include "opto/callnode.hpp"
42 #include "opto/cfgnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/machnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/optoreg.hpp"
47 #include "opto/output.hpp"
48 #include "opto/regalloc.hpp"
49 #include "opto/type.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "utilities/macros.hpp"
52 #include "utilities/powerOfTwo.hpp"
53 #include "utilities/xmlstream.hpp"
54
55 #ifndef PRODUCT
212 _first_block_size(0),
213 _handler_table(),
214 _inc_table(),
215 _stub_list(),
216 _oop_map_set(nullptr),
217 _scratch_buffer_blob(nullptr),
218 _scratch_locs_memory(nullptr),
219 _scratch_const_size(-1),
220 _in_scratch_emit_size(false),
221 _frame_slots(0),
222 _code_offsets(),
223 _node_bundling_limit(0),
224 _node_bundling_base(nullptr),
225 _orig_pc_slot(0),
226 _orig_pc_slot_offset_in_bytes(0),
227 _buf_sizes(),
228 _block(nullptr),
229 _index(0) {
230 C->set_output(this);
231 if (C->stub_name() == nullptr) {
232 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
233 }
234 }
235
236 PhaseOutput::~PhaseOutput() {
237 C->set_output(nullptr);
238 if (_scratch_buffer_blob != nullptr) {
239 BufferBlob::free(_scratch_buffer_blob);
240 }
241 }
242
243 void PhaseOutput::perform_mach_node_analysis() {
244 // Late barrier analysis must be done after schedule and bundle
245 // Otherwise liveness based spilling will fail
246 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
247 bs->late_barrier_analysis();
248
249 pd_perform_mach_node_analysis();
250
251 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
252 }
253
254 // Convert Nodes to instruction bits and pass off to the VM
255 void PhaseOutput::Output() {
256 // RootNode goes
257 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
258
259 // The number of new nodes (mostly MachNop) is proportional to
260 // the number of java calls and inner loops which are aligned.
261 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
262 C->inner_loops()*(OptoLoopAlignment-1)),
263 "out of nodes before code generation" ) ) {
264 return;
265 }
266 // Make sure I can find the Start Node
267 Block *entry = C->cfg()->get_block(1);
268 Block *broot = C->cfg()->get_root_block();
269
270 const StartNode *start = entry->head()->as_Start();
271
272 // Replace StartNode with prolog
273 MachPrologNode *prolog = new MachPrologNode();
274 entry->map_node(prolog, 0);
275 C->cfg()->map_node_to_block(prolog, entry);
276 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
277
278 // Virtual methods need an unverified entry point
279
280 if( C->is_osr_compilation() ) {
281 if( PoisonOSREntry ) {
282 // TODO: Should use a ShouldNotReachHereNode...
283 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
284 }
285 } else {
286 if( C->method() && !C->method()->flags().is_static() ) {
287 // Insert unvalidated entry point
288 C->cfg()->insert( broot, 0, new MachUEPNode() );
289 }
290
291 }
292
293 // Break before main entry point
294 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
295 (OptoBreakpoint && C->is_method_compilation()) ||
296 (OptoBreakpointOSR && C->is_osr_compilation()) ||
297 (OptoBreakpointC2R && !C->method()) ) {
298 // checking for C->method() means that OptoBreakpoint does not apply to
299 // runtime stubs or frame converters
300 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
301 }
302
303 // Insert epilogs before every return
304 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
305 Block* block = C->cfg()->get_block(i);
306 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
307 Node* m = block->end();
308 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
309 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
310 block->add_inst(epilog);
311 C->cfg()->map_node_to_block(epilog, block);
312 }
313 }
314 }
315
316 // Keeper of sizing aspects
317 _buf_sizes = BufferSizingData();
318
319 // Initialize code buffer
320 estimate_buffer_size(_buf_sizes._const);
321 if (C->failing()) return;
322
323 // Pre-compute the length of blocks and replace
324 // long branches with short if machine supports it.
325 // Must be done before ScheduleAndBundle due to SPARC delay slots
326 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
327 blk_starts[0] = 0;
328 shorten_branches(blk_starts);
329
330 ScheduleAndBundle();
331 if (C->failing()) {
332 return;
333 }
334
335 perform_mach_node_analysis();
336
337 // Complete sizing of codebuffer
338 CodeBuffer* cb = init_buffer();
339 if (cb == nullptr || C->failing()) {
340 return;
341 }
342
343 BuildOopMaps();
344
345 if (C->failing()) {
346 return;
347 }
348
349 C2_MacroAssembler masm(cb);
350 fill_buffer(&masm, blk_starts);
351 }
352
353 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
354 // Determine if we need to generate a stack overflow check.
355 // Do it if the method is not a stub function and
356 // has java calls or has frame size > vm_page_size/8.
357 // The debug VM checks that deoptimization doesn't trigger an
358 // unexpected stack overflow (compiled method stack banging should
359 // guarantee it doesn't happen) so we always need the stack bang in
360 // a debug VM.
361 return (C->stub_function() == nullptr &&
362 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
363 DEBUG_ONLY(|| true)));
364 }
365
366 bool PhaseOutput::need_register_stack_bang() const {
367 // Determine if we need to generate a register stack overflow check.
368 // This is only used on architectures which have split register
369 // and memory stacks.
370 // Bang if the method is not a stub function and has java calls
471 // Sum all instruction sizes to compute block size
472 uint last_inst = block->number_of_nodes();
473 uint blk_size = 0;
474 for (uint j = 0; j < last_inst; j++) {
475 _index = j;
476 Node* nj = block->get_node(_index);
477 // Handle machine instruction nodes
478 if (nj->is_Mach()) {
479 MachNode* mach = nj->as_Mach();
480 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
481 reloc_size += mach->reloc();
482 if (mach->is_MachCall()) {
483 // add size information for trampoline stub
484 // class CallStubImpl is platform-specific and defined in the *.ad files.
485 stub_size += CallStubImpl::size_call_trampoline();
486 reloc_size += CallStubImpl::reloc_call_trampoline();
487
488 MachCallNode *mcall = mach->as_MachCall();
489 // This destination address is NOT PC-relative
490
491 mcall->method_set((intptr_t)mcall->entry_point());
492
493 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
494 stub_size += CompiledDirectCall::to_interp_stub_size();
495 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
496 }
497 } else if (mach->is_MachSafePoint()) {
498 // If call/safepoint are adjacent, account for possible
499 // nop to disambiguate the two safepoints.
500 // ScheduleAndBundle() can rearrange nodes in a block,
501 // check for all offsets inside this block.
502 if (last_call_adr >= blk_starts[i]) {
503 blk_size += nop_size;
504 }
505 }
506 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
507 // Nop is inserted between "avoid back to back" instructions.
508 // ScheduleAndBundle() can rearrange nodes in a block,
509 // check for all offsets inside this block.
510 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
511 blk_size += nop_size;
726 // New functionality:
727 // Assert if the local is not top. In product mode let the new node
728 // override the old entry.
729 assert(local == C->top(), "LocArray collision");
730 if (local == C->top()) {
731 return;
732 }
733 array->pop();
734 }
735 const Type *t = local->bottom_type();
736
737 // Is it a safepoint scalar object node?
738 if (local->is_SafePointScalarObject()) {
739 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
740
741 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
742 if (sv == nullptr) {
743 ciKlass* cik = t->is_oopptr()->exact_klass();
744 assert(cik->is_instance_klass() ||
745 cik->is_array_klass(), "Not supported allocation.");
746 sv = new ObjectValue(spobj->_idx,
747 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
748 set_sv_for_object_node(objs, sv);
749
750 uint first_ind = spobj->first_index(sfpt->jvms());
751 for (uint i = 0; i < spobj->n_fields(); i++) {
752 Node* fld_node = sfpt->in(first_ind+i);
753 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
754 }
755 }
756 array->append(sv);
757 return;
758 } else if (local->is_SafePointScalarMerge()) {
759 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
760 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
761
762 if (mv == nullptr) {
763 GrowableArray<ScopeValue*> deps;
764
765 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
766 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
767 assert(deps.length() == 1, "missing value");
768
769 int selector_idx = smerge->selector_idx(sfpt->jvms());
770 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
976 if (!n->is_SafePointScalarObject()) {
977 continue;
978 }
979
980 ObjectValue* other = sv_for_node_id(objs, n->_idx);
981 if (ov == other) {
982 return true;
983 }
984 }
985 return false;
986 }
987
988 //--------------------------Process_OopMap_Node--------------------------------
989 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
990 // Handle special safepoint nodes for synchronization
991 MachSafePointNode *sfn = mach->as_MachSafePoint();
992 MachCallNode *mcall;
993
994 int safepoint_pc_offset = current_offset;
995 bool return_oop = false;
996 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
997 bool arg_escape = false;
998
999 // Add the safepoint in the DebugInfoRecorder
1000 if( !mach->is_MachCall() ) {
1001 mcall = nullptr;
1002 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1003 } else {
1004 mcall = mach->as_MachCall();
1005
1006 if (mcall->is_MachCallJava()) {
1007 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1008 }
1009
1010 // Check if a call returns an object.
1011 if (mcall->returns_pointer()) {
1012 return_oop = true;
1013 }
1014 safepoint_pc_offset += mcall->ret_addr_offset();
1015 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1016 }
1017
1018 // Loop over the JVMState list to add scope information
1019 // Do not skip safepoints with a null method, they need monitor info
1020 JVMState* youngest_jvms = sfn->jvms();
1021 int max_depth = youngest_jvms->depth();
1022
1023 // Allocate the object pool for scalar-replaced objects -- the map from
1024 // small-integer keys (which can be recorded in the local and ostack
1025 // arrays) to descriptions of the object state.
1026 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1027
1028 // Visit scopes from oldest to youngest.
1029 for (int depth = 1; depth <= max_depth; depth++) {
1030 JVMState* jvms = youngest_jvms->of_depth(depth);
1031 int idx;
1032 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1033 // Safepoints that do not have method() set only provide oop-map and monitor info
1062 // Build the growable array of ScopeValues for exp stack
1063 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1064
1065 // Loop over monitors and insert into array
1066 for (idx = 0; idx < num_mon; idx++) {
1067 // Grab the node that defines this monitor
1068 Node* box_node = sfn->monitor_box(jvms, idx);
1069 Node* obj_node = sfn->monitor_obj(jvms, idx);
1070
1071 // Create ScopeValue for object
1072 ScopeValue *scval = nullptr;
1073
1074 if (obj_node->is_SafePointScalarObject()) {
1075 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1076 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1077 if (scval == nullptr) {
1078 const Type *t = spobj->bottom_type();
1079 ciKlass* cik = t->is_oopptr()->exact_klass();
1080 assert(cik->is_instance_klass() ||
1081 cik->is_array_klass(), "Not supported allocation.");
1082 ObjectValue* sv = new ObjectValue(spobj->_idx,
1083 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1084 PhaseOutput::set_sv_for_object_node(objs, sv);
1085
1086 uint first_ind = spobj->first_index(youngest_jvms);
1087 for (uint i = 0; i < spobj->n_fields(); i++) {
1088 Node* fld_node = sfn->in(first_ind+i);
1089 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1090 }
1091 scval = sv;
1092 }
1093 } else if (obj_node->is_SafePointScalarMerge()) {
1094 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1095 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1096
1097 if (mv == nullptr) {
1098 GrowableArray<ScopeValue*> deps;
1099
1100 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1101 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1102 assert(deps.length() == 1, "missing value");
1103
1170 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1171 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1172 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1173
1174 // Make method available for all Safepoints
1175 ciMethod* scope_method = method ? method : C->method();
1176 // Describe the scope here
1177 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1178 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1179 // Now we can describe the scope.
1180 methodHandle null_mh;
1181 bool rethrow_exception = false;
1182 C->debug_info()->describe_scope(
1183 safepoint_pc_offset,
1184 null_mh,
1185 scope_method,
1186 jvms->bci(),
1187 jvms->should_reexecute(),
1188 rethrow_exception,
1189 return_oop,
1190 has_ea_local_in_scope,
1191 arg_escape,
1192 locvals,
1193 expvals,
1194 monvals
1195 );
1196 } // End jvms loop
1197
1198 // Mark the end of the scope set.
1199 C->debug_info()->end_safepoint(safepoint_pc_offset);
1200 }
1201
1202
1203
1204 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1205 class NonSafepointEmitter {
1206 Compile* C;
1207 JVMState* _pending_jvms;
1208 int _pending_offset;
1209
1523 MachNode *nop = new MachNopNode(nops_cnt);
1524 block->insert_node(nop, j++);
1525 last_inst++;
1526 C->cfg()->map_node_to_block(nop, block);
1527 // Ensure enough space.
1528 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1529 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1530 C->record_failure("CodeCache is full");
1531 return;
1532 }
1533 nop->emit(masm, C->regalloc());
1534 masm->code()->flush_bundle(true);
1535 current_offset = masm->offset();
1536 }
1537
1538 bool observe_safepoint = is_sfn;
1539 // Remember the start of the last call in a basic block
1540 if (is_mcall) {
1541 MachCallNode *mcall = mach->as_MachCall();
1542
1543 // This destination address is NOT PC-relative
1544 mcall->method_set((intptr_t)mcall->entry_point());
1545
1546 // Save the return address
1547 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1548
1549 observe_safepoint = mcall->guaranteed_safepoint();
1550 }
1551
1552 // sfn will be valid whenever mcall is valid now because of inheritance
1553 if (observe_safepoint) {
1554 // Handle special safepoint nodes for synchronization
1555 if (!is_mcall) {
1556 MachSafePointNode *sfn = mach->as_MachSafePoint();
1557 // !!!!! Stubs only need an oopmap right now, so bail out
1558 if (sfn->jvms()->method() == nullptr) {
1559 // Write the oopmap directly to the code blob??!!
1560 continue;
1561 }
1562 } // End synchronization
1563
1564 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1662 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1663 node_offsets[n->_idx] = masm->offset();
1664 }
1665 #endif
1666 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1667
1668 // "Normal" instruction case
1669 DEBUG_ONLY(uint instr_offset = masm->offset());
1670 n->emit(masm, C->regalloc());
1671 current_offset = masm->offset();
1672
1673 // Above we only verified that there is enough space in the instruction section.
1674 // However, the instruction may emit stubs that cause code buffer expansion.
1675 // Bail out here if expansion failed due to a lack of code cache space.
1676 if (C->failing()) {
1677 return;
1678 }
1679
1680 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1681 "ret_addr_offset() not within emitted code");
1682
1683 #ifdef ASSERT
1684 uint n_size = n->size(C->regalloc());
1685 if (n_size < (current_offset-instr_offset)) {
1686 MachNode* mach = n->as_Mach();
1687 n->dump();
1688 mach->dump_format(C->regalloc(), tty);
1689 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1690 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1691 tty->print_cr(" ------------------- ");
1692 BufferBlob* blob = this->scratch_buffer_blob();
1693 address blob_begin = blob->content_begin();
1694 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1695 assert(false, "wrong size of mach node");
1696 }
1697 #endif
1698 non_safepoints.observe_instruction(n, current_offset);
1699
1700 // mcall is last "call" that can be a safepoint
1701 // record it so we can see if a poll will directly follow it
1702 // in which case we'll need a pad to make the PcDesc sites unique
2917 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2918 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2919 }
2920 }
2921 // Do not allow defs of new derived values to float above GC
2922 // points unless the base is definitely available at the GC point.
2923
2924 Node *m = b->get_node(i);
2925
2926 // Add precedence edge from following safepoint to use of derived pointer
2927 if( last_safept_node != end_node &&
2928 m != last_safept_node) {
2929 for (uint k = 1; k < m->req(); k++) {
2930 const Type *t = m->in(k)->bottom_type();
2931 if( t->isa_oop_ptr() &&
2932 t->is_ptr()->offset() != 0 ) {
2933 last_safept_node->add_prec( m );
2934 break;
2935 }
2936 }
2937 }
2938
2939 if( n->jvms() ) { // Precedence edge from derived to safept
2940 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2941 if( b->get_node(last_safept) != last_safept_node ) {
2942 last_safept = b->find_node(last_safept_node);
2943 }
2944 for( uint j=last_safept; j > i; j-- ) {
2945 Node *mach = b->get_node(j);
2946 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2947 mach->add_prec( n );
2948 }
2949 last_safept = i;
2950 last_safept_node = m;
2951 }
2952 }
2953
2954 if (fat_proj_seen) {
2955 // Garbage collect pinch nodes that were not consumed.
2956 // They are usually created by a fat kill MachProj for a call.
3065 }
3066 #endif
3067
3068 //-----------------------init_scratch_buffer_blob------------------------------
3069 // Construct a temporary BufferBlob and cache it for this compile.
3070 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3071 // If there is already a scratch buffer blob allocated and the
3072 // constant section is big enough, use it. Otherwise free the
3073 // current and allocate a new one.
3074 BufferBlob* blob = scratch_buffer_blob();
3075 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3076 // Use the current blob.
3077 } else {
3078 if (blob != nullptr) {
3079 BufferBlob::free(blob);
3080 }
3081
3082 ResourceMark rm;
3083 _scratch_const_size = const_size;
3084 int size = C2Compiler::initial_code_buffer_size(const_size);
3085 blob = BufferBlob::create("Compile::scratch_buffer", size);
3086 // Record the buffer blob for next time.
3087 set_scratch_buffer_blob(blob);
3088 // Have we run out of code space?
3089 if (scratch_buffer_blob() == nullptr) {
3090 // Let CompilerBroker disable further compilations.
3091 C->record_failure("Not enough space for scratch buffer in CodeCache");
3092 return;
3093 }
3094 }
3095
3096 // Initialize the relocation buffers
3097 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3098 set_scratch_locs_memory(locs_buf);
3099 }
3100
3101
3102 //-----------------------scratch_emit_size-------------------------------------
3103 // Helper function that computes size by emitting code
3104 uint PhaseOutput::scratch_emit_size(const Node* n) {
3135 buf.insts()->set_scratch_emit();
3136 buf.stubs()->set_scratch_emit();
3137
3138 // Do the emission.
3139
3140 Label fakeL; // Fake label for branch instructions.
3141 Label* saveL = nullptr;
3142 uint save_bnum = 0;
3143 bool is_branch = n->is_MachBranch();
3144 C2_MacroAssembler masm(&buf);
3145 masm.bind(fakeL);
3146 if (is_branch) {
3147 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3148 n->as_MachBranch()->label_set(&fakeL, 0);
3149 }
3150 n->emit(&masm, C->regalloc());
3151
3152 // Emitting into the scratch buffer should not fail
3153 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3154
3155 if (is_branch) // Restore label.
3156 n->as_MachBranch()->label_set(saveL, save_bnum);
3157
3158 // End scratch_emit_size section.
3159 set_in_scratch_emit_size(false);
3160
3161 return buf.insts_size();
3162 }
3163
3164 void PhaseOutput::install() {
3165 if (!C->should_install_code()) {
3166 return;
3167 } else if (C->stub_function() != nullptr) {
3168 install_stub(C->stub_name());
3169 } else {
3170 install_code(C->method(),
3171 C->entry_bci(),
3172 CompileBroker::compiler2(),
3173 C->has_unsafe_access(),
3174 SharedRuntime::is_wide_vector(C->max_vector_size()));
3175 }
3176 }
3177
3178 void PhaseOutput::install_code(ciMethod* target,
3179 int entry_bci,
3180 AbstractCompiler* compiler,
3181 bool has_unsafe_access,
3182 bool has_wide_vectors) {
3183 // Check if we want to skip execution of all compiled code.
3184 {
3185 #ifndef PRODUCT
3186 if (OptoNoExecute) {
3187 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3188 return;
3189 }
3190 #endif
3191 Compile::TracePhase tp(_t_registerMethod);
3192
3193 if (C->is_osr_compilation()) {
3194 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3195 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3196 } else {
3197 if (!target->is_static()) {
3198 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3199 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3200 // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3201 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3202 }
3203 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3204 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3205 }
3206
3207 C->env()->register_method(target,
3208 entry_bci,
3209 &_code_offsets,
3210 _orig_pc_slot_offset_in_bytes,
3211 code_buffer(),
3212 frame_size_in_words(),
3213 oop_map_set(),
3214 &_handler_table,
3215 inc_table(),
3216 compiler,
3217 has_unsafe_access,
3218 SharedRuntime::is_wide_vector(C->max_vector_size()),
3219 C->has_monitors(),
3220 C->has_scoped_access(),
3221 0);
3222
3223 if (C->log() != nullptr) { // Print code cache state into compiler log
3224 C->log()->code_cache_state();
3225 }
3226 }
3227 }
3228 void PhaseOutput::install_stub(const char* stub_name) {
3229 // Entry point will be accessed using stub_entry_point();
3230 if (code_buffer() == nullptr) {
3231 Matcher::soft_match_failure();
3232 } else {
3233 if (PrintAssembly && (WizardMode || Verbose))
3234 tty->print_cr("### Stub::%s", stub_name);
3235
3236 if (!C->failing()) {
3237 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3238
3239 // Make the NMethod
3240 // For now we mark the frame as never safe for profile stackwalking
3241 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
|
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/assembler.inline.hpp"
26 #include "code/aotCodeCache.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compilerDirectives.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "gc/shared/barrierSet.hpp"
35 #include "gc/shared/c2/barrierSetC2.hpp"
36 #include "gc/shared/gc_globals.hpp"
37 #include "memory/allocation.hpp"
38 #include "opto/ad.hpp"
39 #include "opto/block.hpp"
40 #include "opto/c2_MacroAssembler.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callnode.hpp"
43 #include "opto/cfgnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/node.hpp"
47 #include "opto/optoreg.hpp"
48 #include "opto/output.hpp"
49 #include "opto/regalloc.hpp"
50 #include "opto/type.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/macros.hpp"
53 #include "utilities/powerOfTwo.hpp"
54 #include "utilities/xmlstream.hpp"
55
56 #ifndef PRODUCT
213 _first_block_size(0),
214 _handler_table(),
215 _inc_table(),
216 _stub_list(),
217 _oop_map_set(nullptr),
218 _scratch_buffer_blob(nullptr),
219 _scratch_locs_memory(nullptr),
220 _scratch_const_size(-1),
221 _in_scratch_emit_size(false),
222 _frame_slots(0),
223 _code_offsets(),
224 _node_bundling_limit(0),
225 _node_bundling_base(nullptr),
226 _orig_pc_slot(0),
227 _orig_pc_slot_offset_in_bytes(0),
228 _buf_sizes(),
229 _block(nullptr),
230 _index(0) {
231 C->set_output(this);
232 if (C->stub_name() == nullptr) {
233 int fixed_slots = C->fixed_slots();
234 if (C->needs_stack_repair()) {
235 fixed_slots -= 2;
236 }
237 // TODO 8284443 Only reserve extra slot if needed
238 if (InlineTypeReturnedAsFields) {
239 fixed_slots -= 2;
240 }
241 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
242 }
243 }
244
245 PhaseOutput::~PhaseOutput() {
246 C->set_output(nullptr);
247 if (_scratch_buffer_blob != nullptr) {
248 BufferBlob::free(_scratch_buffer_blob);
249 }
250 }
251
252 void PhaseOutput::perform_mach_node_analysis() {
253 // Late barrier analysis must be done after schedule and bundle
254 // Otherwise liveness based spilling will fail
255 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
256 bs->late_barrier_analysis();
257
258 pd_perform_mach_node_analysis();
259
260 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
261 }
262
263 // Convert Nodes to instruction bits and pass off to the VM
264 void PhaseOutput::Output() {
265 // RootNode goes
266 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
267
268 // The number of new nodes (mostly MachNop) is proportional to
269 // the number of java calls and inner loops which are aligned.
270 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
271 C->inner_loops()*(OptoLoopAlignment-1)),
272 "out of nodes before code generation" ) ) {
273 return;
274 }
275 // Make sure I can find the Start Node
276 Block *entry = C->cfg()->get_block(1);
277 Block *broot = C->cfg()->get_root_block();
278
279 const StartNode *start = entry->head()->as_Start();
280
281 // Replace StartNode with prolog
282 Label verified_entry;
283 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
284 entry->map_node(prolog, 0);
285 C->cfg()->map_node_to_block(prolog, entry);
286 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
287
288 // Virtual methods need an unverified entry point
289 if (C->is_osr_compilation()) {
290 if (PoisonOSREntry) {
291 // TODO: Should use a ShouldNotReachHereNode...
292 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
293 }
294 } else {
295 if (C->method()) {
296 if (C->method()->has_scalarized_args()) {
297 // Add entry point to unpack all inline type arguments
298 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
299 if (!C->method()->is_static()) {
300 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
301 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
302 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
303 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
304 }
305 } else if (!C->method()->is_static()) {
306 // Insert unvalidated entry point
307 C->cfg()->insert(broot, 0, new MachUEPNode());
308 }
309 }
310 }
311
312 // Break before main entry point
313 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
314 (OptoBreakpoint && C->is_method_compilation()) ||
315 (OptoBreakpointOSR && C->is_osr_compilation()) ||
316 (OptoBreakpointC2R && !C->method()) ) {
317 // checking for C->method() means that OptoBreakpoint does not apply to
318 // runtime stubs or frame converters
319 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
320 }
321
322 // Insert epilogs before every return
323 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
324 Block* block = C->cfg()->get_block(i);
325 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
326 Node* m = block->end();
327 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
328 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
329 block->add_inst(epilog);
330 C->cfg()->map_node_to_block(epilog, block);
331 }
332 }
333 }
334
335 // Keeper of sizing aspects
336 _buf_sizes = BufferSizingData();
337
338 // Initialize code buffer
339 estimate_buffer_size(_buf_sizes._const);
340 if (C->failing()) return;
341
342 // Pre-compute the length of blocks and replace
343 // long branches with short if machine supports it.
344 // Must be done before ScheduleAndBundle due to SPARC delay slots
345 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
346 blk_starts[0] = 0;
347 shorten_branches(blk_starts);
348
349 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
350 // Compute the offsets of the entry points required by the inline type calling convention
351 if (!C->method()->is_static()) {
352 // We have entries at the beginning of the method, implemented by the first 4 nodes.
353 // Entry (unverified) @ offset 0
354 // Verified_Inline_Entry_RO
355 // Inline_Entry (unverified)
356 // Verified_Inline_Entry
357 uint offset = 0;
358 _code_offsets.set_value(CodeOffsets::Entry, offset);
359
360 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
361 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
362
363 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
364 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
365
366 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
367 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
368 } else {
369 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
370 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
371 }
372 }
373
374 ScheduleAndBundle();
375 if (C->failing()) {
376 return;
377 }
378
379 perform_mach_node_analysis();
380
381 // Complete sizing of codebuffer
382 CodeBuffer* cb = init_buffer();
383 if (cb == nullptr || C->failing()) {
384 return;
385 }
386
387 BuildOopMaps();
388
389 if (C->failing()) {
390 return;
391 }
392
393 C2_MacroAssembler masm(cb);
394 fill_buffer(&masm, blk_starts);
395 if (C->failing()) {
396 // If we bailed out during matching, not all nodes were visited and the
397 // label might be in inconsistent state (used but not bound). Reset it.
398 verified_entry.reset();
399 }
400 }
401
402 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
403 // Determine if we need to generate a stack overflow check.
404 // Do it if the method is not a stub function and
405 // has java calls or has frame size > vm_page_size/8.
406 // The debug VM checks that deoptimization doesn't trigger an
407 // unexpected stack overflow (compiled method stack banging should
408 // guarantee it doesn't happen) so we always need the stack bang in
409 // a debug VM.
410 return (C->stub_function() == nullptr &&
411 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
412 DEBUG_ONLY(|| true)));
413 }
414
415 bool PhaseOutput::need_register_stack_bang() const {
416 // Determine if we need to generate a register stack overflow check.
417 // This is only used on architectures which have split register
418 // and memory stacks.
419 // Bang if the method is not a stub function and has java calls
520 // Sum all instruction sizes to compute block size
521 uint last_inst = block->number_of_nodes();
522 uint blk_size = 0;
523 for (uint j = 0; j < last_inst; j++) {
524 _index = j;
525 Node* nj = block->get_node(_index);
526 // Handle machine instruction nodes
527 if (nj->is_Mach()) {
528 MachNode* mach = nj->as_Mach();
529 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
530 reloc_size += mach->reloc();
531 if (mach->is_MachCall()) {
532 // add size information for trampoline stub
533 // class CallStubImpl is platform-specific and defined in the *.ad files.
534 stub_size += CallStubImpl::size_call_trampoline();
535 reloc_size += CallStubImpl::reloc_call_trampoline();
536
537 MachCallNode *mcall = mach->as_MachCall();
538 // This destination address is NOT PC-relative
539
540 if (mcall->entry_point() != nullptr) {
541 mcall->method_set((intptr_t)mcall->entry_point());
542 }
543
544 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
545 stub_size += CompiledDirectCall::to_interp_stub_size();
546 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
547 }
548 } else if (mach->is_MachSafePoint()) {
549 // If call/safepoint are adjacent, account for possible
550 // nop to disambiguate the two safepoints.
551 // ScheduleAndBundle() can rearrange nodes in a block,
552 // check for all offsets inside this block.
553 if (last_call_adr >= blk_starts[i]) {
554 blk_size += nop_size;
555 }
556 }
557 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
558 // Nop is inserted between "avoid back to back" instructions.
559 // ScheduleAndBundle() can rearrange nodes in a block,
560 // check for all offsets inside this block.
561 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
562 blk_size += nop_size;
777 // New functionality:
778 // Assert if the local is not top. In product mode let the new node
779 // override the old entry.
780 assert(local == C->top(), "LocArray collision");
781 if (local == C->top()) {
782 return;
783 }
784 array->pop();
785 }
786 const Type *t = local->bottom_type();
787
788 // Is it a safepoint scalar object node?
789 if (local->is_SafePointScalarObject()) {
790 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
791
792 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
793 if (sv == nullptr) {
794 ciKlass* cik = t->is_oopptr()->exact_klass();
795 assert(cik->is_instance_klass() ||
796 cik->is_array_klass(), "Not supported allocation.");
797 uint first_ind = spobj->first_index(sfpt->jvms());
798 // Nullable, scalarized inline types have a null_marker input
799 // that needs to be checked before using the field values.
800 ScopeValue* properties = nullptr;
801 if (cik->is_inlinetype()) {
802 Node* null_marker_node = sfpt->in(first_ind++);
803 assert(null_marker_node != nullptr, "null_marker node not found");
804 if (!null_marker_node->is_top()) {
805 const TypeInt* null_marker_type = null_marker_node->bottom_type()->is_int();
806 if (null_marker_node->is_Con()) {
807 properties = new ConstantIntValue(null_marker_type->get_con());
808 } else {
809 OptoReg::Name null_marker_reg = C->regalloc()->get_reg_first(null_marker_node);
810 properties = new_loc_value(C->regalloc(), null_marker_reg, Location::normal);
811 }
812 }
813 }
814 if (cik->is_array_klass() && !cik->is_type_array_klass()) {
815 jint props = ArrayKlass::ArrayProperties::DEFAULT;
816 if (cik->as_array_klass()->element_klass()->is_inlinetype()) {
817 if (cik->as_array_klass()->is_elem_null_free()) {
818 props |= ArrayKlass::ArrayProperties::NULL_RESTRICTED;
819 }
820 if (!cik->as_array_klass()->is_elem_atomic()) {
821 props |= ArrayKlass::ArrayProperties::NON_ATOMIC;
822 }
823 }
824 properties = new ConstantIntValue(props);
825 }
826 sv = new ObjectValue(spobj->_idx,
827 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, properties);
828 set_sv_for_object_node(objs, sv);
829
830 for (uint i = 0; i < spobj->n_fields(); i++) {
831 Node* fld_node = sfpt->in(first_ind+i);
832 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
833 }
834 }
835 array->append(sv);
836 return;
837 } else if (local->is_SafePointScalarMerge()) {
838 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
839 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
840
841 if (mv == nullptr) {
842 GrowableArray<ScopeValue*> deps;
843
844 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
845 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
846 assert(deps.length() == 1, "missing value");
847
848 int selector_idx = smerge->selector_idx(sfpt->jvms());
849 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
1055 if (!n->is_SafePointScalarObject()) {
1056 continue;
1057 }
1058
1059 ObjectValue* other = sv_for_node_id(objs, n->_idx);
1060 if (ov == other) {
1061 return true;
1062 }
1063 }
1064 return false;
1065 }
1066
1067 //--------------------------Process_OopMap_Node--------------------------------
1068 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1069 // Handle special safepoint nodes for synchronization
1070 MachSafePointNode *sfn = mach->as_MachSafePoint();
1071 MachCallNode *mcall;
1072
1073 int safepoint_pc_offset = current_offset;
1074 bool return_oop = false;
1075 bool return_scalarized = false;
1076 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1077 bool arg_escape = false;
1078
1079 // Add the safepoint in the DebugInfoRecorder
1080 if( !mach->is_MachCall() ) {
1081 mcall = nullptr;
1082 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1083 } else {
1084 mcall = mach->as_MachCall();
1085
1086 if (mcall->is_MachCallJava()) {
1087 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1088 }
1089
1090 // Check if a call returns an object.
1091 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1092 return_oop = true;
1093 }
1094 if (mcall->returns_scalarized()) {
1095 return_scalarized = true;
1096 }
1097 safepoint_pc_offset += mcall->ret_addr_offset();
1098 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1099 }
1100
1101 // Loop over the JVMState list to add scope information
1102 // Do not skip safepoints with a null method, they need monitor info
1103 JVMState* youngest_jvms = sfn->jvms();
1104 int max_depth = youngest_jvms->depth();
1105
1106 // Allocate the object pool for scalar-replaced objects -- the map from
1107 // small-integer keys (which can be recorded in the local and ostack
1108 // arrays) to descriptions of the object state.
1109 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1110
1111 // Visit scopes from oldest to youngest.
1112 for (int depth = 1; depth <= max_depth; depth++) {
1113 JVMState* jvms = youngest_jvms->of_depth(depth);
1114 int idx;
1115 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1116 // Safepoints that do not have method() set only provide oop-map and monitor info
1145 // Build the growable array of ScopeValues for exp stack
1146 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1147
1148 // Loop over monitors and insert into array
1149 for (idx = 0; idx < num_mon; idx++) {
1150 // Grab the node that defines this monitor
1151 Node* box_node = sfn->monitor_box(jvms, idx);
1152 Node* obj_node = sfn->monitor_obj(jvms, idx);
1153
1154 // Create ScopeValue for object
1155 ScopeValue *scval = nullptr;
1156
1157 if (obj_node->is_SafePointScalarObject()) {
1158 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1159 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1160 if (scval == nullptr) {
1161 const Type *t = spobj->bottom_type();
1162 ciKlass* cik = t->is_oopptr()->exact_klass();
1163 assert(cik->is_instance_klass() ||
1164 cik->is_array_klass(), "Not supported allocation.");
1165 assert(!cik->is_inlinetype(), "Synchronization on value object?");
1166 ScopeValue* properties = nullptr;
1167 if (cik->is_array_klass() && !cik->is_type_array_klass()) {
1168 jint props = ArrayKlass::ArrayProperties::DEFAULT;
1169 if (cik->as_array_klass()->element_klass()->is_inlinetype()) {
1170 if (cik->as_array_klass()->is_elem_null_free()) {
1171 props |= ArrayKlass::ArrayProperties::NULL_RESTRICTED;
1172 }
1173 if (!cik->as_array_klass()->is_elem_atomic()) {
1174 props |= ArrayKlass::ArrayProperties::NON_ATOMIC;
1175 }
1176 }
1177 properties = new ConstantIntValue(props);
1178 }
1179 ObjectValue* sv = new ObjectValue(spobj->_idx,
1180 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, properties);
1181 PhaseOutput::set_sv_for_object_node(objs, sv);
1182
1183 uint first_ind = spobj->first_index(youngest_jvms);
1184 for (uint i = 0; i < spobj->n_fields(); i++) {
1185 Node* fld_node = sfn->in(first_ind+i);
1186 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1187 }
1188 scval = sv;
1189 }
1190 } else if (obj_node->is_SafePointScalarMerge()) {
1191 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1192 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1193
1194 if (mv == nullptr) {
1195 GrowableArray<ScopeValue*> deps;
1196
1197 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1198 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1199 assert(deps.length() == 1, "missing value");
1200
1267 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1268 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1269 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1270
1271 // Make method available for all Safepoints
1272 ciMethod* scope_method = method ? method : C->method();
1273 // Describe the scope here
1274 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1275 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1276 // Now we can describe the scope.
1277 methodHandle null_mh;
1278 bool rethrow_exception = false;
1279 C->debug_info()->describe_scope(
1280 safepoint_pc_offset,
1281 null_mh,
1282 scope_method,
1283 jvms->bci(),
1284 jvms->should_reexecute(),
1285 rethrow_exception,
1286 return_oop,
1287 return_scalarized,
1288 has_ea_local_in_scope,
1289 arg_escape,
1290 locvals,
1291 expvals,
1292 monvals
1293 );
1294 } // End jvms loop
1295
1296 // Mark the end of the scope set.
1297 C->debug_info()->end_safepoint(safepoint_pc_offset);
1298 }
1299
1300
1301
1302 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1303 class NonSafepointEmitter {
1304 Compile* C;
1305 JVMState* _pending_jvms;
1306 int _pending_offset;
1307
1621 MachNode *nop = new MachNopNode(nops_cnt);
1622 block->insert_node(nop, j++);
1623 last_inst++;
1624 C->cfg()->map_node_to_block(nop, block);
1625 // Ensure enough space.
1626 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1627 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1628 C->record_failure("CodeCache is full");
1629 return;
1630 }
1631 nop->emit(masm, C->regalloc());
1632 masm->code()->flush_bundle(true);
1633 current_offset = masm->offset();
1634 }
1635
1636 bool observe_safepoint = is_sfn;
1637 // Remember the start of the last call in a basic block
1638 if (is_mcall) {
1639 MachCallNode *mcall = mach->as_MachCall();
1640
1641 if (mcall->entry_point() != nullptr) {
1642 // This destination address is NOT PC-relative
1643 mcall->method_set((intptr_t)mcall->entry_point());
1644 }
1645
1646 // Save the return address
1647 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1648
1649 observe_safepoint = mcall->guaranteed_safepoint();
1650 }
1651
1652 // sfn will be valid whenever mcall is valid now because of inheritance
1653 if (observe_safepoint) {
1654 // Handle special safepoint nodes for synchronization
1655 if (!is_mcall) {
1656 MachSafePointNode *sfn = mach->as_MachSafePoint();
1657 // !!!!! Stubs only need an oopmap right now, so bail out
1658 if (sfn->jvms()->method() == nullptr) {
1659 // Write the oopmap directly to the code blob??!!
1660 continue;
1661 }
1662 } // End synchronization
1663
1664 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1762 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1763 node_offsets[n->_idx] = masm->offset();
1764 }
1765 #endif
1766 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1767
1768 // "Normal" instruction case
1769 DEBUG_ONLY(uint instr_offset = masm->offset());
1770 n->emit(masm, C->regalloc());
1771 current_offset = masm->offset();
1772
1773 // Above we only verified that there is enough space in the instruction section.
1774 // However, the instruction may emit stubs that cause code buffer expansion.
1775 // Bail out here if expansion failed due to a lack of code cache space.
1776 if (C->failing()) {
1777 return;
1778 }
1779
1780 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1781 "ret_addr_offset() not within emitted code");
1782 #ifdef ASSERT
1783 uint n_size = n->size(C->regalloc());
1784 if (n_size < (current_offset-instr_offset)) {
1785 MachNode* mach = n->as_Mach();
1786 n->dump();
1787 mach->dump_format(C->regalloc(), tty);
1788 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1789 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1790 tty->print_cr(" ------------------- ");
1791 BufferBlob* blob = this->scratch_buffer_blob();
1792 address blob_begin = blob->content_begin();
1793 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1794 assert(false, "wrong size of mach node");
1795 }
1796 #endif
1797 non_safepoints.observe_instruction(n, current_offset);
1798
1799 // mcall is last "call" that can be a safepoint
1800 // record it so we can see if a poll will directly follow it
1801 // in which case we'll need a pad to make the PcDesc sites unique
3016 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3017 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3018 }
3019 }
3020 // Do not allow defs of new derived values to float above GC
3021 // points unless the base is definitely available at the GC point.
3022
3023 Node *m = b->get_node(i);
3024
3025 // Add precedence edge from following safepoint to use of derived pointer
3026 if( last_safept_node != end_node &&
3027 m != last_safept_node) {
3028 for (uint k = 1; k < m->req(); k++) {
3029 const Type *t = m->in(k)->bottom_type();
3030 if( t->isa_oop_ptr() &&
3031 t->is_ptr()->offset() != 0 ) {
3032 last_safept_node->add_prec( m );
3033 break;
3034 }
3035 }
3036
3037 // Do not allow a CheckCastPP node whose input is a raw pointer to
3038 // float past a safepoint. This can occur when a buffered inline
3039 // type is allocated in a loop and the CheckCastPP from that
3040 // allocation is reused outside the loop. If the use inside the
3041 // loop is scalarized the CheckCastPP will no longer be connected
3042 // to the loop safepoint. See JDK-8264340.
3043 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3044 Node *def = m->in(1);
3045 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3046 last_safept_node->add_prec(m);
3047 }
3048 }
3049 }
3050
3051 if( n->jvms() ) { // Precedence edge from derived to safept
3052 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3053 if( b->get_node(last_safept) != last_safept_node ) {
3054 last_safept = b->find_node(last_safept_node);
3055 }
3056 for( uint j=last_safept; j > i; j-- ) {
3057 Node *mach = b->get_node(j);
3058 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3059 mach->add_prec( n );
3060 }
3061 last_safept = i;
3062 last_safept_node = m;
3063 }
3064 }
3065
3066 if (fat_proj_seen) {
3067 // Garbage collect pinch nodes that were not consumed.
3068 // They are usually created by a fat kill MachProj for a call.
3177 }
3178 #endif
3179
3180 //-----------------------init_scratch_buffer_blob------------------------------
3181 // Construct a temporary BufferBlob and cache it for this compile.
3182 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3183 // If there is already a scratch buffer blob allocated and the
3184 // constant section is big enough, use it. Otherwise free the
3185 // current and allocate a new one.
3186 BufferBlob* blob = scratch_buffer_blob();
3187 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3188 // Use the current blob.
3189 } else {
3190 if (blob != nullptr) {
3191 BufferBlob::free(blob);
3192 }
3193
3194 ResourceMark rm;
3195 _scratch_const_size = const_size;
3196 int size = C2Compiler::initial_code_buffer_size(const_size);
3197 if (C->has_scalarized_args()) {
3198 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3199 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3200 ciMethod* method = C->method();
3201 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3202 int arg_num = 0;
3203 if (!method->is_static()) {
3204 if (method->is_scalarized_arg(arg_num)) {
3205 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3206 }
3207 arg_num++;
3208 }
3209 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3210 if (method->is_scalarized_arg(arg_num)) {
3211 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3212 }
3213 arg_num++;
3214 }
3215 }
3216 blob = BufferBlob::create("Compile::scratch_buffer", size);
3217 // Record the buffer blob for next time.
3218 set_scratch_buffer_blob(blob);
3219 // Have we run out of code space?
3220 if (scratch_buffer_blob() == nullptr) {
3221 // Let CompilerBroker disable further compilations.
3222 C->record_failure("Not enough space for scratch buffer in CodeCache");
3223 return;
3224 }
3225 }
3226
3227 // Initialize the relocation buffers
3228 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3229 set_scratch_locs_memory(locs_buf);
3230 }
3231
3232
3233 //-----------------------scratch_emit_size-------------------------------------
3234 // Helper function that computes size by emitting code
3235 uint PhaseOutput::scratch_emit_size(const Node* n) {
3266 buf.insts()->set_scratch_emit();
3267 buf.stubs()->set_scratch_emit();
3268
3269 // Do the emission.
3270
3271 Label fakeL; // Fake label for branch instructions.
3272 Label* saveL = nullptr;
3273 uint save_bnum = 0;
3274 bool is_branch = n->is_MachBranch();
3275 C2_MacroAssembler masm(&buf);
3276 masm.bind(fakeL);
3277 if (is_branch) {
3278 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3279 n->as_MachBranch()->label_set(&fakeL, 0);
3280 }
3281 n->emit(&masm, C->regalloc());
3282
3283 // Emitting into the scratch buffer should not fail
3284 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3285
3286 // Restore label.
3287 if (is_branch) {
3288 n->as_MachBranch()->label_set(saveL, save_bnum);
3289 }
3290
3291 // End scratch_emit_size section.
3292 set_in_scratch_emit_size(false);
3293
3294 return buf.insts_size();
3295 }
3296
3297 void PhaseOutput::install() {
3298 if (!C->should_install_code()) {
3299 return;
3300 } else if (C->stub_function() != nullptr) {
3301 install_stub(C->stub_name());
3302 } else {
3303 install_code(C->method(),
3304 C->entry_bci(),
3305 CompileBroker::compiler2(),
3306 C->has_unsafe_access(),
3307 SharedRuntime::is_wide_vector(C->max_vector_size()));
3308 }
3309 }
3310
3311 void PhaseOutput::install_code(ciMethod* target,
3312 int entry_bci,
3313 AbstractCompiler* compiler,
3314 bool has_unsafe_access,
3315 bool has_wide_vectors) {
3316 // Check if we want to skip execution of all compiled code.
3317 {
3318 #ifndef PRODUCT
3319 if (OptoNoExecute) {
3320 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3321 return;
3322 }
3323 #endif
3324 Compile::TracePhase tp(_t_registerMethod);
3325
3326 if (C->is_osr_compilation()) {
3327 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3328 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3329 } else {
3330 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3331 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3332 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3333 }
3334 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3335 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3336 }
3337 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3338 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3339 }
3340 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3341 }
3342
3343 C->env()->register_method(target,
3344 entry_bci,
3345 &_code_offsets,
3346 _orig_pc_slot_offset_in_bytes,
3347 code_buffer(),
3348 frame_size_in_words(),
3349 _oop_map_set,
3350 &_handler_table,
3351 inc_table(),
3352 compiler,
3353 has_unsafe_access,
3354 SharedRuntime::is_wide_vector(C->max_vector_size()),
3355 C->has_monitors(),
3356 C->has_scoped_access(),
3357 0);
3358
3359 if (C->log() != nullptr) { // Print code cache state into compiler log
3360 C->log()->code_cache_state();
3361 }
3362 }
3363 }
3364 void PhaseOutput::install_stub(const char* stub_name) {
3365 // Entry point will be accessed using stub_entry_point();
3366 if (code_buffer() == nullptr) {
3367 Matcher::soft_match_failure();
3368 } else {
3369 if (PrintAssembly && (WizardMode || Verbose))
3370 tty->print_cr("### Stub::%s", stub_name);
3371
3372 if (!C->failing()) {
3373 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3374
3375 // Make the NMethod
3376 // For now we mark the frame as never safe for profile stackwalking
3377 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
|