1 /*
2 * Copyright (c) 2005, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.hpp"
26 #include "ci/ciInlineKlass.hpp"
27 #include "ci/ciInstanceKlass.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "gc/shared/collectedHeap.inline.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "libadt/vectset.hpp"
32 #include "memory/universe.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/inlinetypenode.hpp"
42 #include "opto/intrinsicnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/macro.hpp"
46 #include "opto/memnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/node.hpp"
49 #include "opto/opaquenode.hpp"
50 #include "opto/opcodes.hpp"
51 #include "opto/phaseX.hpp"
52 #include "opto/reachability.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/runtime.hpp"
55 #include "opto/subnode.hpp"
56 #include "opto/subtypenode.hpp"
57 #include "opto/type.hpp"
58 #include "prims/jvmtiExport.hpp"
59 #include "runtime/continuation.hpp"
60 #include "runtime/sharedRuntime.hpp"
61 #include "runtime/stubRoutines.hpp"
62 #include "utilities/globalDefinitions.hpp"
63 #include "utilities/macros.hpp"
64 #include "utilities/powerOfTwo.hpp"
65 #if INCLUDE_G1GC
66 #include "gc/g1/g1ThreadLocalData.hpp"
67 #endif // INCLUDE_G1GC
68
69
70 //
71 // Replace any references to "oldref" in inputs to "use" with "newref".
72 // Returns the number of replacements made.
73 //
74 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
75 int nreplacements = 0;
76 uint req = use->req();
77 for (uint j = 0; j < use->len(); j++) {
78 Node *uin = use->in(j);
79 if (uin == oldref) {
80 if (j < req)
81 use->set_req(j, newref);
82 else
83 use->set_prec(j, newref);
84 nreplacements++;
85 } else if (j >= req && uin == nullptr) {
86 break;
87 }
88 }
89 return nreplacements;
90 }
91
92
93 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
94 Node* cmp = word;
95 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
96 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
97 transform_later(iff);
98
99 // Fast path taken.
100 Node *fast_taken = transform_later(new IfFalseNode(iff));
101
102 // Fast path not-taken, i.e. slow path
103 Node *slow_taken = transform_later(new IfTrueNode(iff));
104
105 region->init_req(edge, fast_taken); // Capture fast-control
106 return slow_taken;
107 }
108
109 //--------------------copy_predefined_input_for_runtime_call--------------------
110 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
111 // Set fixed predefined input arguments
112 call->init_req( TypeFunc::Control, ctrl );
113 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
114 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
115 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
116 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
117 }
118
119 //------------------------------make_slow_call---------------------------------
120 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
121 address slow_call, const char* leaf_name, Node* slow_path,
122 Node* parm0, Node* parm1, Node* parm2) {
123
124 // Slow-path call
125 CallNode *call = leaf_name
126 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
127 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
128
129 // Slow path call has no side-effects, uses few values
130 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
131 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
132 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
133 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
134 call->copy_call_debug_info(&_igvn, oldcall);
135 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
136 _igvn.replace_node(oldcall, call);
137 transform_later(call);
138
139 return call;
140 }
141
142 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
143 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
144 bs->eliminate_gc_barrier(&_igvn, p2x);
145 #ifndef PRODUCT
146 if (PrintOptoStatistics) {
147 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
148 }
149 #endif
150 }
151
152 // Search for a memory operation for the specified memory slice.
153 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
154 Node *orig_mem = mem;
155 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
156 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
157 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
158 while (true) {
159 if (mem == alloc_mem || mem == start_mem ) {
160 return mem; // hit one of our sentinels
161 } else if (mem->is_MergeMem()) {
162 mem = mem->as_MergeMem()->memory_at(alias_idx);
163 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
164 Node *in = mem->in(0);
165 // we can safely skip over safepoints, calls, locks and membars because we
166 // already know that the object is safe to eliminate.
167 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
168 return in;
169 } else if (in->is_Call()) {
170 CallNode *call = in->as_Call();
171 if (call->may_modify(tinst, phase)) {
172 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
173 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
174 return in;
175 }
176 }
177 mem = in->in(TypeFunc::Memory);
178 } else if (in->is_MemBar()) {
179 ArrayCopyNode* ac = nullptr;
180 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
181 if (ac != nullptr) {
182 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
183 return ac;
184 }
185 }
186 mem = in->in(TypeFunc::Memory);
187 } else if (in->is_LoadFlat() || in->is_StoreFlat()) {
188 mem = in->in(TypeFunc::Memory);
189 } else {
190 #ifdef ASSERT
191 in->dump();
192 mem->dump();
193 assert(false, "unexpected projection");
194 #endif
195 }
196 } else if (mem->is_Store()) {
197 const TypePtr* atype = mem->as_Store()->adr_type();
198 int adr_idx = phase->C->get_alias_index(atype);
199 if (adr_idx == alias_idx) {
200 assert(atype->isa_oopptr(), "address type must be oopptr");
201 int adr_offset = atype->flat_offset();
202 uint adr_iid = atype->is_oopptr()->instance_id();
203 // Array elements references have the same alias_idx
204 // but different offset and different instance_id.
205 if (adr_offset == offset && adr_iid == alloc->_idx) {
206 return mem;
207 }
208 } else {
209 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
210 }
211 mem = mem->in(MemNode::Memory);
212 } else if (mem->is_ClearArray()) {
213 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
214 // Can not bypass initialization of the instance
215 // we are looking.
216 DEBUG_ONLY(intptr_t offset;)
217 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
218 InitializeNode* init = alloc->as_Allocate()->initialization();
219 // We are looking for stored value, return Initialize node
220 // or memory edge from Allocate node.
221 if (init != nullptr) {
222 return init;
223 } else {
224 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
225 }
226 }
227 // Otherwise skip it (the call updated 'mem' value).
228 } else if (mem->Opcode() == Op_SCMemProj) {
229 mem = mem->in(0);
230 Node* adr = nullptr;
231 if (mem->is_LoadStore()) {
232 adr = mem->in(MemNode::Address);
233 } else {
234 assert(mem->Opcode() == Op_EncodeISOArray ||
235 mem->Opcode() == Op_StrCompressedCopy, "sanity");
236 adr = mem->in(3); // Destination array
237 }
238 const TypePtr* atype = adr->bottom_type()->is_ptr();
239 int adr_idx = phase->C->get_alias_index(atype);
240 if (adr_idx == alias_idx) {
241 DEBUG_ONLY(mem->dump();)
242 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
243 return nullptr;
244 }
245 mem = mem->in(MemNode::Memory);
246 } else if (mem->Opcode() == Op_StrInflatedCopy) {
247 Node* adr = mem->in(3); // Destination array
248 const TypePtr* atype = adr->bottom_type()->is_ptr();
249 int adr_idx = phase->C->get_alias_index(atype);
250 if (adr_idx == alias_idx) {
251 DEBUG_ONLY(mem->dump();)
252 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
253 return nullptr;
254 }
255 mem = mem->in(MemNode::Memory);
256 } else {
257 return mem;
258 }
259 assert(mem != orig_mem, "dead memory loop");
260 }
261 }
262
263 // Determine if there is an interfering store between a rematerialization load and an arraycopy that is in the process
264 // of being elided. Starting from the given rematerialization load this method starts a BFS traversal upwards through
265 // the memory graph towards the provided ArrayCopyNode. For every node encountered on the traversal, check that it is
266 // independent from the provided rematerialization. Returns false if every node on the traversal is independent and
267 // true otherwise.
268 bool has_interfering_store(const ArrayCopyNode* ac, LoadNode* load, PhaseGVN* phase) {
269 assert(ac != nullptr && load != nullptr, "sanity");
270 AccessAnalyzer acc(phase, load);
271 ResourceMark rm;
272 Unique_Node_List to_visit;
273 to_visit.push(load->in(MemNode::Memory));
274
275 for (uint worklist_idx = 0; worklist_idx < to_visit.size(); worklist_idx++) {
276 Node* mem = to_visit.at(worklist_idx);
277
278 if (mem->is_Proj() && mem->in(0) == ac) {
279 // Reached the target, so visit what is left on the worklist.
280 continue;
281 }
282
283 if (mem->is_Phi()) {
284 assert(mem->bottom_type() == Type::MEMORY, "do not leave memory graph");
285 // Add all non-control inputs of phis to be visited.
286 for (uint phi_in = 1; phi_in < mem->len(); phi_in++) {
287 Node* input = mem->in(phi_in);
288 if (input != nullptr) {
289 to_visit.push(input);
290 }
291 }
292 continue;
293 }
294
295 AccessAnalyzer::AccessIndependence ind = acc.detect_access_independence(mem);
296 if (ind.independent) {
297 to_visit.push(ind.mem);
298 } else {
299 return true;
300 }
301 }
302 // Did not find modification of source element in memory graph.
303 return false;
304 }
305
306 // Generate loads from source of the arraycopy for fields of destination needed at a deoptimization point.
307 // Returns nullptr if the load cannot be created because the arraycopy is not suitable for elimination
308 // (e.g. copy inside the array with non-constant offsets) or the inputs do not match our assumptions (e.g.
309 // the arraycopy does not actually write something at the provided offset).
310 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type* ftype, AllocateNode* alloc) {
311 assert((ctl == ac->control() && mem == ac->memory()) != (mem != ac->memory() && ctl->is_Proj() && ctl->as_Proj()->is_uncommon_trap_proj()),
312 "Either the control and memory are the same as for the arraycopy or they are pinned in an uncommon trap.");
313 BasicType bt = ft;
314 const Type *type = ftype;
315 if (ft == T_NARROWOOP) {
316 bt = T_OBJECT;
317 type = ftype->make_oopptr();
318 }
319 Node* base = ac->in(ArrayCopyNode::Src);
320 Node* adr = nullptr;
321 const TypePtr* adr_type = nullptr;
322
323 if (ac->is_clonebasic()) {
324 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
325 adr = _igvn.transform(AddPNode::make_with_base(base, _igvn.MakeConX(offset)));
326 adr_type = _igvn.type(base)->is_ptr();
327 if (adr_type->isa_aryptr()) {
328 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset);
329 } else {
330 adr_type = adr_type->add_offset(offset);
331 }
332 } else {
333 if (!ac->modifies(offset, offset, &_igvn, true)) {
334 // If the arraycopy does not copy to this offset, we cannot generate a rematerialization load for it.
335 return nullptr;
336 }
337 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
338 uint shift = exact_log2(type2aelembytes(bt));
339 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
340 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
341 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
342 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
343
344 adr_type = _igvn.type(base)->is_aryptr();
345 if (((const TypeAryPtr*)adr_type)->is_flat()) {
346 shift = ((const TypeAryPtr*)adr_type)->flat_log_elem_size();
347 }
348 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
349 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
350 adr = _igvn.transform(AddPNode::make_with_base(base, base, _igvn.MakeConX(off)));
351 adr_type = _igvn.type(adr)->is_aryptr();
352 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
353 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
354 // Don't emit a new load from src if src == dst but try to get the value from memory instead
355 return value_from_mem(ac, ctl, ft, ftype, (const TypeAryPtr*)adr_type, alloc);
356 }
357 } else {
358 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
359 // Non constant offset in the array: we can't statically
360 // determine the value
361 return nullptr;
362 }
363 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
364 #ifdef _LP64
365 diff = _igvn.transform(new ConvI2LNode(diff));
366 #endif
367 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
368
369 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
370 adr = _igvn.transform(AddPNode::make_with_base(base, base, off));
371 // In the case of a flat inline type array, each field has its
372 // own slice so we need to extract the field being accessed from
373 // the address computation
374 adr_type = ((TypeAryPtr*)adr_type)->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
375 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
376 }
377 }
378 assert(adr != nullptr && adr_type != nullptr, "sanity");
379
380 // Create the rematerialization load ...
381 MergeMemNode* mergemem = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
382 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
383 Node* res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemem, adr, adr_type, type, bt);
384 assert(res != nullptr, "load should have been created");
385
386 // ... and ensure that pinning the rematerialization load inside the uncommon path is safe.
387 if (mem != ac->memory() && ctl->is_Proj() && ctl->as_Proj()->is_uncommon_trap_proj() && res->is_Load() &&
388 has_interfering_store(ac, res->as_Load(), &_igvn)) {
389 // Not safe: use control and memory from the arraycopy to ensure correct memory state.
390 _igvn.remove_dead_node(res, PhaseIterGVN::NodeOrigin::Graph); // Clean up the unusable rematerialization load.
391 return make_arraycopy_load(ac, offset, ac->control(), ac->memory(), ft, ftype, alloc);
392 }
393
394 if (ftype->isa_narrowoop()) {
395 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
396 res = _igvn.transform(new EncodePNode(res, ftype));
397 }
398 return res;
399 }
400
401 //
402 // Given a Memory Phi, compute a value Phi containing the values from stores
403 // on the input paths.
404 // Note: this function is recursive, its depth is limited by the "level" argument
405 // Returns the computed Phi, or null if it cannot compute it.
406 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
407 assert(mem->is_Phi(), "sanity");
408 int alias_idx = C->get_alias_index(adr_t);
409 int offset = adr_t->flat_offset();
410 int instance_id = adr_t->instance_id();
411
412 // Check if an appropriate value phi already exists.
413 Node* region = mem->in(0);
414 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
415 Node* phi = region->fast_out(k);
416 if (phi->is_Phi() && phi != mem &&
417 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
418 return phi;
419 }
420 }
421 // Check if an appropriate new value phi already exists.
422 Node* new_phi = value_phis->find(mem->_idx);
423 if (new_phi != nullptr)
424 return new_phi;
425
426 if (level <= 0) {
427 return nullptr; // Give up: phi tree too deep
428 }
429 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
430 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
431 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
432
433 uint length = mem->req();
434 GrowableArray <Node *> values(length, length, nullptr);
435
436 // create a new Phi for the value
437 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
438 transform_later(phi);
439 value_phis->push(phi, mem->_idx);
440
441 for (uint j = 1; j < length; j++) {
442 Node *in = mem->in(j);
443 if (in == nullptr || in->is_top()) {
444 values.at_put(j, in);
445 } else {
446 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
447 if (val == start_mem || val == alloc_mem) {
448 // hit a sentinel, return appropriate value
449 Node* init_value = value_from_alloc(ft, adr_t, alloc);
450 if (init_value == nullptr) {
451 return nullptr;
452 } else {
453 values.at_put(j, init_value);
454 continue;
455 }
456 }
457 if (val->is_Initialize()) {
458 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
459 }
460 if (val == nullptr) {
461 return nullptr; // can't find a value on this path
462 }
463 if (val == mem) {
464 values.at_put(j, mem);
465 } else if (val->is_Store()) {
466 Node* n = val->in(MemNode::ValueIn);
467 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
468 n = bs->step_over_gc_barrier(n);
469 if (is_subword_type(ft)) {
470 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
471 }
472 values.at_put(j, n);
473 } else if (val->is_Proj() && val->in(0) == alloc) {
474 Node* init_value = value_from_alloc(ft, adr_t, alloc);
475 if (init_value == nullptr) {
476 return nullptr;
477 } else {
478 values.at_put(j, init_value);
479 }
480 } else if (val->is_Phi()) {
481 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
482 if (val == nullptr) {
483 return nullptr;
484 }
485 values.at_put(j, val);
486 } else if (val->Opcode() == Op_SCMemProj) {
487 assert(val->in(0)->is_LoadStore() ||
488 val->in(0)->Opcode() == Op_EncodeISOArray ||
489 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
490 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
491 return nullptr;
492 } else if (val->is_ArrayCopy()) {
493 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
494 if (res == nullptr) {
495 return nullptr;
496 }
497 values.at_put(j, res);
498 } else if (val->is_top()) {
499 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
500 // IGVN will clean this up later.
501 values.at_put(j, val);
502 } else {
503 DEBUG_ONLY( val->dump(); )
504 assert(false, "unknown node on this path");
505 return nullptr; // unknown node on this path
506 }
507 }
508 }
509 // Set Phi's inputs
510 for (uint j = 1; j < length; j++) {
511 if (values.at(j) == mem) {
512 phi->init_req(j, phi);
513 } else {
514 phi->init_req(j, values.at(j));
515 }
516 }
517 return phi;
518 }
519
520 // Extract the initial value of a field in an allocation
521 Node* PhaseMacroExpand::value_from_alloc(BasicType ft, const TypeOopPtr* adr_t, AllocateNode* alloc) {
522 Node* init_value = alloc->in(AllocateNode::InitValue);
523 if (init_value == nullptr) {
524 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "conflicting InitValue and RawInitValue");
525 return _igvn.zerocon(ft);
526 }
527
528 const TypeAryPtr* ary_t = adr_t->isa_aryptr();
529 assert(ary_t != nullptr, "must be a pointer into an array");
530
531 // If this is not a flat array, then it must be an oop array with elements being init_value
532 if (ary_t->is_not_flat()) {
533 #ifdef ASSERT
534 BasicType init_bt = init_value->bottom_type()->basic_type();
535 assert(ft == init_bt ||
536 (!is_java_primitive(ft) && !is_java_primitive(init_bt) && type2aelembytes(ft, true) == type2aelembytes(init_bt, true)) ||
537 (is_subword_type(ft) && init_bt == T_INT),
538 "invalid init_value of type %s for field of type %s", type2name(init_bt), type2name(ft));
539 #endif // ASSERT
540 return init_value;
541 }
542
543 assert(ary_t->klass_is_exact() && ary_t->is_flat(), "must be an exact flat array");
544 assert(ary_t->field_offset().get() != Type::OffsetBot, "unknown offset");
545 if (init_value->is_EncodeP()) {
546 init_value = init_value->in(1);
547 }
548 // Cannot look through init_value if it is an oop
549 if (!init_value->is_InlineType()) {
550 return nullptr;
551 }
552
553 ciInlineKlass* vk = init_value->bottom_type()->inline_klass();
554 if (ary_t->field_offset().get() == vk->null_marker_offset_in_payload()) {
555 init_value = init_value->as_InlineType()->get_null_marker();
556 } else {
557 init_value = init_value->as_InlineType()->field_value_by_offset(ary_t->field_offset().get() + vk->payload_offset(), true);
558 }
559
560 if (ft == T_NARROWOOP) {
561 assert(init_value->bottom_type()->isa_ptr(), "must be a pointer");
562 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_narrowoop()));
563 }
564
565 #ifdef ASSERT
566 BasicType init_bt = init_value->bottom_type()->basic_type();
567 assert(ft == init_bt ||
568 (!is_java_primitive(ft) && !is_java_primitive(init_bt) && type2aelembytes(ft, true) == type2aelembytes(init_bt, true)) ||
569 (is_subword_type(ft) && init_bt == T_INT),
570 "invalid init_value of type %s for field of type %s", type2name(init_bt), type2name(ft));
571 #endif // ASSERT
572
573 return init_value;
574 }
575
576 // Search the last value stored into the object's field.
577 Node* PhaseMacroExpand::value_from_mem(Node* origin, Node* ctl, BasicType ft, const Type* ftype, const TypeOopPtr* adr_t, AllocateNode* alloc) {
578 assert(adr_t->is_known_instance_field(), "instance required");
579 int instance_id = adr_t->instance_id();
580 assert((uint)instance_id == alloc->_idx, "wrong allocation");
581
582 int alias_idx = C->get_alias_index(adr_t);
583 int offset = adr_t->flat_offset();
584 Node* orig_mem = origin->in(TypeFunc::Memory);
585 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
586 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
587 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
588 VectorSet visited;
589
590 bool done = orig_mem == alloc_mem;
591 Node *mem = orig_mem;
592 while (!done) {
593 if (visited.test_set(mem->_idx)) {
594 return nullptr; // found a loop, give up
595 }
596 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
597 if (mem == start_mem || mem == alloc_mem) {
598 done = true; // hit a sentinel, return appropriate 0 value
599 } else if (mem->is_Initialize()) {
600 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
601 if (mem == nullptr) {
602 done = true; // Something went wrong.
603 } else if (mem->is_Store()) {
604 const TypePtr* atype = mem->as_Store()->adr_type();
605 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
606 done = true;
607 }
608 } else if (mem->is_Store()) {
609 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
610 assert(atype != nullptr, "address type must be oopptr");
611 assert(C->get_alias_index(atype) == alias_idx &&
612 atype->is_known_instance_field() && atype->flat_offset() == offset &&
613 atype->instance_id() == instance_id, "store is correct memory slice");
614 done = true;
615 } else if (mem->is_Phi()) {
616 // try to find a phi's unique input
617 Node *unique_input = nullptr;
618 Node *top = C->top();
619 for (uint i = 1; i < mem->req(); i++) {
620 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
621 if (n == nullptr || n == top || n == mem) {
622 continue;
623 } else if (unique_input == nullptr) {
624 unique_input = n;
625 } else if (unique_input != n) {
626 unique_input = top;
627 break;
628 }
629 }
630 if (unique_input != nullptr && unique_input != top) {
631 mem = unique_input;
632 } else {
633 done = true;
634 }
635 } else if (mem->is_ArrayCopy()) {
636 done = true;
637 } else if (mem->is_top()) {
638 // The slice is on a dead path. Returning nullptr would lead to elimination
639 // bailout, but we want to prevent that. Just forwarding the top is also legal,
640 // and IGVN can just clean things up, and remove whatever receives top.
641 return mem;
642 } else {
643 DEBUG_ONLY( mem->dump(); )
644 assert(false, "unexpected node");
645 }
646 }
647 if (mem != nullptr) {
648 if (mem == start_mem || mem == alloc_mem) {
649 // hit a sentinel, return appropriate value
650 return value_from_alloc(ft, adr_t, alloc);
651 } else if (mem->is_Store()) {
652 Node* n = mem->in(MemNode::ValueIn);
653 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
654 n = bs->step_over_gc_barrier(n);
655 return n;
656 } else if (mem->is_Phi()) {
657 // attempt to produce a Phi reflecting the values on the input paths of the Phi
658 Node_Stack value_phis(8);
659 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
660 if (phi != nullptr) {
661 return phi;
662 } else {
663 // Kill all new Phis
664 while(value_phis.is_nonempty()) {
665 Node* n = value_phis.node();
666 _igvn.replace_node(n, C->top());
667 value_phis.pop();
668 }
669 }
670 } else if (mem->is_ArrayCopy()) {
671 // Rematerialize the scalar-replaced array. If possible, pin the loads to the uncommon path of the uncommon trap.
672 // Check for each element of the source array, whether it was modified. If not, pin both memory and control to
673 // the uncommon path. Otherwise, use the control and memory state of the arraycopy. Control and memory state must
674 // come from the same source to prevent anti-dependence problems in the backend.
675 ArrayCopyNode* ac = mem->as_ArrayCopy();
676 Node* ac_ctl = ac->control();
677 Node* ac_mem = ac->memory();
678 if (ctl->is_Proj() && ctl->as_Proj()->is_uncommon_trap_proj()) {
679 // pin the loads in the uncommon trap path
680 ac_ctl = ctl;
681 ac_mem = orig_mem;
682 }
683 return make_arraycopy_load(ac, offset, ac_ctl, ac_mem, ft, ftype, alloc);
684 }
685 }
686 // Something went wrong.
687 return nullptr;
688 }
689
690 // Search the last value stored into the inline type's fields (for flat arrays).
691 Node* PhaseMacroExpand::inline_type_from_mem(ciInlineKlass* vk, const TypeAryPtr* elem_adr_type, int elem_idx, int offset_in_element, bool null_free, AllocateNode* alloc, SafePointNode* sfpt) {
692 auto report_failure = [&](int field_offset_in_element) {
693 #ifndef PRODUCT
694 if (PrintEliminateAllocations) {
695 ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
696 int offset = field_offset_in_element + elem_klass->payload_offset();
697 ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
698 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
699 tty->print("=== At SafePoint node %d can't find value of field [%s] of array element [%d]", sfpt->_idx, flattened_field->name()->as_utf8(), elem_idx);
700 tty->print(", which prevents elimination of: ");
701 alloc->dump();
702 }
703 #endif // PRODUCT
704 };
705
706 // Create a new InlineTypeNode and retrieve the field values from memory
707 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, null_free);
708 transform_later(vt);
709 if (null_free) {
710 vt->set_null_marker(_igvn);
711 } else {
712 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
713 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
714 Node* nm_value = value_from_mem(sfpt, sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
715 if (nm_value != nullptr) {
716 vt->set_null_marker(_igvn, nm_value);
717 } else {
718 report_failure(nm_offset_in_element);
719 return nullptr;
720 }
721 }
722
723 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
724 ciField* field = vt->field(i);
725 ciType* field_type = field->type();
726 int field_offset_in_element = offset_in_element + field->offset_in_bytes() - vk->payload_offset();
727 Node* field_value = nullptr;
728 assert(!field->is_flat() || field->type()->is_inlinetype(), "must be an inline type");
729 if (field->is_flat()) {
730 field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, field->is_null_free(), alloc, sfpt);
731 } else {
732 const Type* ft = Type::get_const_type(field_type);
733 BasicType bt = type2field[field_type->basic_type()];
734 if (UseCompressedOops && !is_java_primitive(bt)) {
735 ft = ft->make_narrowoop();
736 bt = T_NARROWOOP;
737 }
738 // Each inline type field has its own memory slice
739 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
740 field_value = value_from_mem(sfpt, sfpt->control(), bt, ft, field_adr_type, alloc);
741 if (field_value == nullptr) {
742 report_failure(field_offset_in_element);
743 } else if (ft->isa_narrowoop()) {
744 assert(UseCompressedOops, "unexpected narrow oop");
745 if (field_value->is_EncodeP()) {
746 field_value = field_value->in(1);
747 } else if (!field_value->is_InlineType()) {
748 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
749 }
750 }
751 }
752 if (field_value != nullptr) {
753 vt->set_field_value(i, field_value);
754 } else {
755 return nullptr;
756 }
757 }
758 return vt;
759 }
760
761 // Check the possibility of scalar replacement.
762 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode* alloc, Unique_Node_List* safepoints) {
763 // Scan the uses of the allocation to check for anything that would
764 // prevent us from eliminating it.
765 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
766 DEBUG_ONLY( Node* disq_node = nullptr; )
767 bool can_eliminate = true;
768 bool reduce_merge_precheck = (safepoints == nullptr);
769
770 Unique_Node_List worklist;
771 Node* res = alloc->result_cast();
772 const TypeOopPtr* res_type = nullptr;
773 if (res == nullptr) {
774 // All users were eliminated.
775 } else if (!res->is_CheckCastPP()) {
776 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
777 can_eliminate = false;
778 } else {
779 worklist.push(res);
780 res_type = igvn->type(res)->isa_oopptr();
781 if (res_type == nullptr) {
782 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
783 can_eliminate = false;
784 } else if (!res_type->klass_is_exact()) {
785 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
786 can_eliminate = false;
787 } else if (res_type->isa_aryptr()) {
788 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
789 if (length < 0) {
790 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
791 can_eliminate = false;
792 }
793 }
794 }
795
796 while (can_eliminate && worklist.size() > 0) {
797 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
798 res = worklist.pop();
799 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
800 Node* use = res->fast_out(j);
801
802 if (use->is_AddP()) {
803 const TypePtr* addp_type = igvn->type(use)->is_ptr();
804 int offset = addp_type->offset();
805
806 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
807 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
808 can_eliminate = false;
809 break;
810 }
811 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
812 k < kmax && can_eliminate; k++) {
813 Node* n = use->fast_out(k);
814 if ((n->is_Mem() && n->as_Mem()->is_mismatched_access()) || n->is_LoadFlat() || n->is_StoreFlat()) {
815 DEBUG_ONLY(disq_node = n);
816 NOT_PRODUCT(fail_eliminate = "Mismatched access");
817 can_eliminate = false;
818 }
819 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
820 DEBUG_ONLY(disq_node = n;)
821 if (n->is_Load() || n->is_LoadStore()) {
822 NOT_PRODUCT(fail_eliminate = "Field load";)
823 } else {
824 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
825 }
826 can_eliminate = false;
827 }
828 }
829 } else if (use->is_ArrayCopy() &&
830 (use->as_ArrayCopy()->is_clonebasic() ||
831 use->as_ArrayCopy()->is_arraycopy_validated() ||
832 use->as_ArrayCopy()->is_copyof_validated() ||
833 use->as_ArrayCopy()->is_copyofrange_validated()) &&
834 use->in(ArrayCopyNode::Dest) == res) {
835 // ok to eliminate
836 } else if (use->is_ReachabilityFence() && OptimizeReachabilityFences) {
837 // ok to eliminate
838 } else if (use->is_SafePoint()) {
839 SafePointNode* sfpt = use->as_SafePoint();
840 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
841 // Object is passed as argument.
842 DEBUG_ONLY(disq_node = use;)
843 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
844 can_eliminate = false;
845 }
846 Node* sfptMem = sfpt->memory();
847 if (sfptMem == nullptr || sfptMem->is_top()) {
848 DEBUG_ONLY(disq_node = use;)
849 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
850 can_eliminate = false;
851 } else if (!reduce_merge_precheck) {
852 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
853 safepoints->push(sfpt);
854 }
855 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
856 // Look at uses
857 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
858 Node* u = use->fast_out(k);
859 if (u->is_InlineType()) {
860 // Use in flat field can be eliminated
861 InlineTypeNode* vt = u->as_InlineType();
862 for (uint i = 0; i < vt->field_count(); ++i) {
863 if (vt->field_value(i) == use && !vt->field(i)->is_flat()) {
864 can_eliminate = false; // Use in non-flat field
865 break;
866 }
867 }
868 } else {
869 // Add other uses to the worklist to process individually
870 worklist.push(use);
871 }
872 }
873 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
874 // Store to mark word of inline type larval buffer
875 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
876 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
877 // Inline type buffer allocations are followed by a membar
878 } else if (reduce_merge_precheck &&
879 (use->is_Phi() || use->is_EncodeP() ||
880 use->Opcode() == Op_MemBarRelease ||
881 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
882 // Nothing to do
883 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
884 if (use->is_Phi()) {
885 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
886 NOT_PRODUCT(fail_eliminate = "Object is return value";)
887 } else {
888 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
889 }
890 DEBUG_ONLY(disq_node = use;)
891 } else {
892 if (use->Opcode() == Op_Return) {
893 NOT_PRODUCT(fail_eliminate = "Object is return value";)
894 } else {
895 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
896 }
897 DEBUG_ONLY(disq_node = use;)
898 }
899 can_eliminate = false;
900 } else {
901 assert(use->Opcode() == Op_CastP2X, "should be");
902 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
903 }
904 }
905 }
906
907 #ifndef PRODUCT
908 if (PrintEliminateAllocations && safepoints != nullptr) {
909 if (can_eliminate) {
910 tty->print("Scalar ");
911 if (res == nullptr)
912 alloc->dump();
913 else
914 res->dump();
915 } else {
916 tty->print("NotScalar (%s)", fail_eliminate);
917 if (res == nullptr)
918 alloc->dump();
919 else
920 res->dump();
921 #ifdef ASSERT
922 if (disq_node != nullptr) {
923 tty->print(" >>>> ");
924 disq_node->dump();
925 }
926 #endif /*ASSERT*/
927 }
928 }
929
930 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
931 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
932 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
933 }
934 #endif
935 return can_eliminate;
936 }
937
938 void PhaseMacroExpand::undo_previous_scalarizations(Unique_Node_List& safepoints_done, AllocateNode* alloc) {
939 Node* res = alloc->result_cast();
940 int nfields = 0;
941 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
942
943 if (res != nullptr) {
944 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
945
946 if (res_type->isa_instptr()) {
947 // find the fields of the class which will be needed for safepoint debug information
948 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
949 nfields = iklass->nof_nonstatic_fields();
950 } else {
951 // find the array's elements which will be needed for safepoint debug information
952 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
953 assert(nfields >= 0, "must be an array klass.");
954 }
955 }
956
957 // rollback processed safepoints
958 while (safepoints_done.size() > 0) {
959 SafePointNode* sfpt_done = safepoints_done.pop()->as_SafePoint();
960
961 SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(igvn());
962
963 sfpt_done->remove_non_debug_edges(non_debug_edges_worklist);
964
965 // remove any extra entries we added to the safepoint
966 assert(sfpt_done->jvms()->endoff() == sfpt_done->req(), "no extra edges past debug info allowed");
967 uint last = sfpt_done->req() - 1;
968 for (int k = 0; k < nfields; k++) {
969 sfpt_done->del_req(last--);
970 }
971 JVMState *jvms = sfpt_done->jvms();
972 jvms->set_endoff(sfpt_done->req());
973 // Now make a pass over the debug information replacing any references
974 // to SafePointScalarObjectNode with the allocated object.
975 int start = jvms->debug_start();
976 int end = jvms->debug_end();
977 for (int i = start; i < end; i++) {
978 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
979 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
980 if (scobj->first_index(jvms) == sfpt_done->req() &&
981 scobj->n_fields() == (uint)nfields) {
982 assert(scobj->alloc() == alloc, "sanity");
983 sfpt_done->set_req(i, res);
984 }
985 }
986 }
987
988 sfpt_done->restore_non_debug_edges(non_debug_edges_worklist);
989
990 _igvn._worklist.push(sfpt_done);
991 }
992 }
993
994 #ifdef ASSERT
995 // Verify if a value can be written into a field.
996 void verify_type_compatability(const Type* value_type, const Type* field_type) {
997 BasicType value_bt = value_type->basic_type();
998 BasicType field_bt = field_type->basic_type();
999
1000 // Primitive types must match.
1001 if (is_java_primitive(value_bt) && value_bt == field_bt) { return; }
1002
1003 // I have been struggling to make a similar assert for non-primitive
1004 // types. I we can add one in the future. For now, I just let them
1005 // pass without checks.
1006 // In particular, I was struggling with a value that came from a call,
1007 // and had only a non-null check CastPP. There was also a checkcast
1008 // in the graph to verify the interface, but the corresponding
1009 // CheckCastPP result was not updated in the stack slot, and so
1010 // we ended up using the CastPP. That means that the field knows
1011 // that it should get an oop from an interface, but the value lost
1012 // that information, and so it is not a subtype.
1013 // There may be other issues, feel free to investigate further!
1014 if (!is_java_primitive(value_bt)) { return; }
1015
1016 tty->print_cr("value not compatible for field: %s vs %s",
1017 type2name(value_bt),
1018 type2name(field_bt));
1019 tty->print("value_type: ");
1020 value_type->dump();
1021 tty->cr();
1022 tty->print("field_type: ");
1023 field_type->dump();
1024 tty->cr();
1025 assert(false, "value_type does not fit field_type");
1026 }
1027 #endif
1028
1029 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1030 if (UseCompressedOops && field_type->isa_narrowoop()) {
1031 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
1032 // to be able scalar replace the allocation.
1033 if (field_val->is_EncodeP()) {
1034 field_val = field_val->in(1);
1035 } else if (!field_val->is_InlineType()) {
1036 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
1037 }
1038 }
1039
1040 // Keep track of inline types to scalarize them later
1041 if (field_val->is_InlineType()) {
1042 value_worklist->push(field_val);
1043 } else if (field_val->is_Phi()) {
1044 PhiNode* phi = field_val->as_Phi();
1045 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
1046 // scalarize all its fields in safepoints.
1047 field_val = phi->try_push_inline_types_down(&_igvn, true);
1048 if (field_val->is_InlineType()) {
1049 value_worklist->push(field_val);
1050 }
1051 }
1052 DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
1053 sfpt->add_req(field_val);
1054 }
1055
1056 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1057 const Type* elem_type = array_type->elem();
1058 BasicType basic_elem_type = elem_type->array_element_basic_type();
1059
1060 intptr_t elem_size;
1061 uint header_size;
1062 if (array_type->is_flat()) {
1063 elem_size = array_type->flat_elem_size();
1064 header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1065 } else {
1066 elem_size = type2aelembytes(basic_elem_type);
1067 header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1068 }
1069
1070 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
1071 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
1072 intptr_t elem_offset = header_size + elem_idx * elem_size;
1073 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
1074 Node* elem_val;
1075 if (array_type->is_flat()) {
1076 ciInlineKlass* elem_klass = elem_type->inline_klass();
1077 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
1078 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
1079 } else {
1080 elem_val = value_from_mem(sfpt, sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
1081 #ifndef PRODUCT
1082 if (PrintEliminateAllocations && elem_val == nullptr) {
1083 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, elem_idx);
1084 tty->print(", which prevents elimination of: ");
1085 alloc->dump();
1086 }
1087 #endif // PRODUCT
1088 }
1089 if (elem_val == nullptr) {
1090 return false;
1091 }
1092
1093 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
1094 }
1095
1096 return true;
1097 }
1098
1099 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
1100 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
1101 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
1102 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1103 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
1104 auto report_failure = [&](int offset) {
1105 #ifndef PRODUCT
1106 if (PrintEliminateAllocations) {
1107 ciInstanceKlass* base_klass = base_type->instance_klass();
1108 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
1109 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
1110 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
1111 flattened_field->print();
1112 int field_idx = C->alias_type(flattened_field)->index();
1113 tty->print(" (alias_idx=%d)", field_idx);
1114 tty->print(", which prevents elimination of: ");
1115 base->dump();
1116 }
1117 #endif // PRODUCT
1118 };
1119
1120 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
1121 ciField* field = iklass->declared_nonstatic_field_at(i);
1122 if (field->is_flat()) {
1123 ciInlineKlass* fvk = field->type()->as_inline_klass();
1124 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
1125 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
1126 if (!success) {
1127 return false;
1128 }
1129
1130 // The null marker of a field is added right after we scalarize that field
1131 if (!field->is_null_free()) {
1132 int nm_offset = offset_minus_header + field->null_marker_offset();
1133 Node* null_marker = value_from_mem(sfpt, sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
1134 if (null_marker == nullptr) {
1135 report_failure(nm_offset);
1136 return false;
1137 }
1138 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
1139 }
1140
1141 continue;
1142 }
1143
1144 int offset = offset_minus_header + field->offset_in_bytes();
1145 ciType* elem_type = field->type();
1146 BasicType basic_elem_type = field->layout_type();
1147
1148 const Type* field_type;
1149 if (is_reference_type(basic_elem_type)) {
1150 if (!elem_type->is_loaded()) {
1151 field_type = TypeInstPtr::BOTTOM;
1152 } else {
1153 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
1154 }
1155 if (UseCompressedOops) {
1156 field_type = field_type->make_narrowoop();
1157 basic_elem_type = T_NARROWOOP;
1158 }
1159 } else {
1160 field_type = Type::get_const_basic_type(basic_elem_type);
1161 }
1162
1163 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1164 Node* field_val = value_from_mem(sfpt, sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1165 if (field_val == nullptr) {
1166 report_failure(offset);
1167 return false;
1168 }
1169 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1170 }
1171
1172 return true;
1173 }
1174
1175 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1176 Unique_Node_List* value_worklist) {
1177 assert(sfpt->jvms()->endoff() == sfpt->req(), "no extra edges past debug info allowed");
1178
1179 // Fields of scalar objs are referenced only at the end
1180 // of regular debuginfo at the last (youngest) JVMS.
1181 // Record relative start index.
1182 ciInstanceKlass* iklass = nullptr;
1183 const TypeOopPtr* res_type = nullptr;
1184 int nfields = 0;
1185 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1186 Node* res = alloc->result_cast();
1187
1188 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1189 assert(sfpt->jvms() != nullptr, "missed JVMS");
1190 uint before_sfpt_req = sfpt->req();
1191
1192 if (res != nullptr) { // Could be null when there are no users
1193 res_type = _igvn.type(res)->isa_oopptr();
1194
1195 if (res_type->isa_instptr()) {
1196 // find the fields of the class which will be needed for safepoint debug information
1197 iklass = res_type->is_instptr()->instance_klass();
1198 nfields = iklass->nof_nonstatic_fields();
1199 } else {
1200 // find the array's elements which will be needed for safepoint debug information
1201 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1202 assert(nfields >= 0, "must be an array klass.");
1203 }
1204
1205 if (res->bottom_type()->is_inlinetypeptr()) {
1206 // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1207 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1208 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1209 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1210 // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1211 // later in PhaseOutput::filLocArray() for inline types.
1212 sfpt->add_req(_igvn.intcon(1));
1213 }
1214 }
1215
1216 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1217 sobj->init_req(0, C->root());
1218 transform_later(sobj);
1219
1220 if (res == nullptr) {
1221 sfpt->jvms()->set_endoff(sfpt->req());
1222 return sobj;
1223 }
1224
1225 bool success;
1226 if (iklass == nullptr) {
1227 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1228 } else {
1229 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1230 }
1231
1232 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1233 if (!success) {
1234 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1235 sfpt->del_req(i);
1236 }
1237 _igvn._worklist.push(sfpt);
1238 return nullptr;
1239 }
1240
1241 sfpt->jvms()->set_endoff(sfpt->req());
1242 return sobj;
1243 }
1244
1245 // Do scalar replacement.
1246 bool PhaseMacroExpand::scalar_replacement(AllocateNode* alloc, Unique_Node_List& safepoints) {
1247 Unique_Node_List safepoints_done;
1248 Node* res = alloc->result_cast();
1249 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1250 const TypeOopPtr* res_type = nullptr;
1251 if (res != nullptr) { // Could be null when there are no users
1252 res_type = _igvn.type(res)->isa_oopptr();
1253 }
1254
1255 // Process the safepoint uses
1256 Unique_Node_List value_worklist;
1257 while (safepoints.size() > 0) {
1258 SafePointNode* sfpt = safepoints.pop()->as_SafePoint();
1259
1260 SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(igvn());
1261
1262 // All sfpt inputs are implicitly included into debug info during the scalarization process below.
1263 // Keep non-debug inputs separately, so they stay non-debug.
1264 sfpt->remove_non_debug_edges(non_debug_edges_worklist);
1265
1266 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1267
1268 if (sobj == nullptr) {
1269 sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1270 undo_previous_scalarizations(safepoints_done, alloc);
1271 return false;
1272 }
1273
1274 // Now make a pass over the debug information replacing any references
1275 // to the allocated object with "sobj"
1276 JVMState *jvms = sfpt->jvms();
1277 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1278 non_debug_edges_worklist.remove_edge_if_present(res); // drop scalarized input from non-debug info
1279 sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1280 _igvn._worklist.push(sfpt);
1281
1282 // keep it for rollback
1283 safepoints_done.push(sfpt);
1284 }
1285 // Scalarize inline types that were added to the safepoint.
1286 // Don't allow linking a constant oop (if available) for flat array elements
1287 // because Deoptimization::reassign_flat_array_elements needs field values.
1288 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1289 for (uint i = 0; i < value_worklist.size(); ++i) {
1290 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1291 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1292 }
1293 return true;
1294 }
1295
1296 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1297 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1298 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1299 if (ctl_proj != nullptr) {
1300 igvn.replace_node(ctl_proj, n->in(0));
1301 }
1302 if (mem_proj != nullptr) {
1303 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1304 }
1305 }
1306
1307 // Process users of eliminated allocation.
1308 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1309 Unique_Node_List worklist;
1310 Node* res = alloc->result_cast();
1311 if (res != nullptr) {
1312 worklist.push(res);
1313 }
1314 while (worklist.size() > 0) {
1315 res = worklist.pop();
1316 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1317 Node *use = res->last_out(j);
1318 uint oc1 = res->outcnt();
1319
1320 if (use->is_AddP()) {
1321 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1322 Node *n = use->last_out(k);
1323 uint oc2 = use->outcnt();
1324 if (n->is_Store()) {
1325 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1326 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1327 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1328 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1329 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1330 mb->remove(&_igvn);
1331 }
1332 }
1333 _igvn.replace_node(n, n->in(MemNode::Memory));
1334 } else {
1335 eliminate_gc_barrier(n);
1336 }
1337 k -= (oc2 - use->outcnt());
1338 }
1339 _igvn.remove_dead_node(use, PhaseIterGVN::NodeOrigin::Graph);
1340 } else if (use->is_ArrayCopy()) {
1341 // Disconnect ArrayCopy node
1342 ArrayCopyNode* ac = use->as_ArrayCopy();
1343 if (ac->is_clonebasic()) {
1344 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1345 disconnect_projections(ac, _igvn);
1346 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1347 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1348 disconnect_projections(membar_before->as_MemBar(), _igvn);
1349 if (membar_after->is_MemBar()) {
1350 disconnect_projections(membar_after->as_MemBar(), _igvn);
1351 }
1352 } else {
1353 assert(ac->is_arraycopy_validated() ||
1354 ac->is_copyof_validated() ||
1355 ac->is_copyofrange_validated(), "unsupported");
1356 CallProjections* callprojs = ac->extract_projections(true);
1357
1358 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1359 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1360 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1361
1362 // Set control to top. IGVN will remove the remaining projections
1363 ac->set_req(0, top());
1364 ac->replace_edge(res, top(), &_igvn);
1365
1366 // Disconnect src right away: it can help find new
1367 // opportunities for allocation elimination
1368 Node* src = ac->in(ArrayCopyNode::Src);
1369 ac->replace_edge(src, top(), &_igvn);
1370 // src can be top at this point if src and dest of the
1371 // arraycopy were the same
1372 if (src->outcnt() == 0 && !src->is_top()) {
1373 _igvn.remove_dead_node(src, PhaseIterGVN::NodeOrigin::Graph);
1374 }
1375 }
1376 _igvn._worklist.push(ac);
1377 } else if (use->is_InlineType()) {
1378 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1379 // Cut off oop input and remove known instance id from type
1380 _igvn.rehash_node_delayed(use);
1381 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1382 use->as_InlineType()->set_is_buffered(_igvn, false);
1383 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1384 _igvn.set_type(use, toop);
1385 use->as_InlineType()->set_type(toop);
1386 // Process users
1387 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1388 Node* u = use->fast_out(k);
1389 if (!u->is_InlineType() && !u->is_StoreFlat()) {
1390 worklist.push(u);
1391 }
1392 }
1393 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1394 // Store to mark word of inline type larval buffer
1395 assert(inline_alloc, "Unexpected store to mark word");
1396 _igvn.replace_node(use, use->in(MemNode::Memory));
1397 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1398 // Inline type buffer allocations are followed by a membar
1399 assert(inline_alloc, "Unexpected MemBarRelease");
1400 use->as_MemBar()->remove(&_igvn);
1401 } else if (use->is_ReachabilityFence() && OptimizeReachabilityFences) {
1402 use->as_ReachabilityFence()->clear_referent(_igvn); // redundant fence; will be removed during IGVN
1403 } else {
1404 eliminate_gc_barrier(use);
1405 }
1406 j -= (oc1 - res->outcnt());
1407 }
1408 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1409 _igvn.remove_dead_node(res, PhaseIterGVN::NodeOrigin::Graph);
1410 }
1411
1412 //
1413 // Process other users of allocation's projections
1414 //
1415 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1416 // First disconnect stores captured by Initialize node.
1417 // If Initialize node is eliminated first in the following code,
1418 // it will kill such stores and DUIterator_Last will assert.
1419 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1420 Node* use = _callprojs->resproj[0]->fast_out(j);
1421 if (use->is_AddP()) {
1422 // raw memory addresses used only by the initialization
1423 _igvn.replace_node(use, C->top());
1424 --j; --jmax;
1425 }
1426 }
1427 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1428 Node* use = _callprojs->resproj[0]->last_out(j);
1429 uint oc1 = _callprojs->resproj[0]->outcnt();
1430 if (use->is_Initialize()) {
1431 // Eliminate Initialize node.
1432 InitializeNode *init = use->as_Initialize();
1433 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1434 if (ctrl_proj != nullptr) {
1435 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1436 #ifdef ASSERT
1437 // If the InitializeNode has no memory out, it will die, and tmp will become null
1438 Node* tmp = init->in(TypeFunc::Control);
1439 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1440 #endif
1441 }
1442 Node* mem = init->in(TypeFunc::Memory);
1443 #ifdef ASSERT
1444 if (init->number_of_projs(TypeFunc::Memory) > 0) {
1445 if (mem->is_MergeMem()) {
1446 assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs->fallthrough_memproj, "allocation memory projection");
1447 } else {
1448 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1449 }
1450 }
1451 #endif
1452 init->replace_mem_projs_by(mem, &_igvn);
1453 assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1454 } else if (use->Opcode() == Op_MemBarStoreStore) {
1455 // Inline type buffer allocations are followed by a membar
1456 assert(inline_alloc, "Unexpected MemBarStoreStore");
1457 use->as_MemBar()->remove(&_igvn);
1458 } else {
1459 assert(false, "only Initialize or AddP expected");
1460 }
1461 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1462 }
1463 }
1464 if (_callprojs->fallthrough_catchproj != nullptr) {
1465 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1466 }
1467 if (_callprojs->fallthrough_memproj != nullptr) {
1468 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1469 }
1470 if (_callprojs->catchall_memproj != nullptr) {
1471 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1472 }
1473 if (_callprojs->fallthrough_ioproj != nullptr) {
1474 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1475 }
1476 if (_callprojs->catchall_ioproj != nullptr) {
1477 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1478 }
1479 if (_callprojs->catchall_catchproj != nullptr) {
1480 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1481 }
1482 }
1483
1484 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1485 // If reallocation fails during deoptimization we'll pop all
1486 // interpreter frames for this compiled frame and that won't play
1487 // nice with JVMTI popframe.
1488 // We avoid this issue by eager reallocation when the popframe request
1489 // is received.
1490 if (!EliminateAllocations) {
1491 return false;
1492 }
1493 Node* klass = alloc->in(AllocateNode::KlassNode);
1494 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1495
1496 // Attempt to eliminate inline type buffer allocations
1497 // regardless of usage and escape/replaceable status.
1498 bool inline_alloc = tklass->isa_instklassptr() &&
1499 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1500 if (!alloc->_is_non_escaping && !inline_alloc) {
1501 return false;
1502 }
1503 // Eliminate boxing allocations which are not used
1504 // regardless scalar replaceable status.
1505 Node* res = alloc->result_cast();
1506 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1507 tklass->isa_instklassptr() &&
1508 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1509 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1510 return false;
1511 }
1512
1513 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1514
1515 Unique_Node_List safepoints;
1516 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1517 return false;
1518 }
1519
1520 if (!alloc->_is_scalar_replaceable) {
1521 assert(res == nullptr || inline_alloc, "sanity");
1522 // We can only eliminate allocation if all debug info references
1523 // are already replaced with SafePointScalarObject because
1524 // we can't search for a fields value without instance_id.
1525 if (safepoints.size() > 0) {
1526 return false;
1527 }
1528 }
1529
1530 if (!scalar_replacement(alloc, safepoints)) {
1531 return false;
1532 }
1533
1534 CompileLog* log = C->log();
1535 if (log != nullptr) {
1536 log->head("eliminate_allocation type='%d'",
1537 log->identify(tklass->exact_klass()));
1538 JVMState* p = alloc->jvms();
1539 while (p != nullptr) {
1540 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1541 p = p->caller();
1542 }
1543 log->tail("eliminate_allocation");
1544 }
1545
1546 process_users_of_allocation(alloc, inline_alloc);
1547
1548 #ifndef PRODUCT
1549 if (PrintEliminateAllocations) {
1550 if (alloc->is_AllocateArray())
1551 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1552 else
1553 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1554 }
1555 #endif
1556
1557 return true;
1558 }
1559
1560 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode* call) {
1561 // EA should remove all uses of non-escaping boxing node.
1562 if (!C->eliminate_boxing()) {
1563 return false;
1564 }
1565 for (uint i = TypeFunc::Parms; i < call->tf()->range_cc()->cnt(); ++i) {
1566 if (call->proj_out_or_null(i) != nullptr) {
1567 return false;
1568 }
1569 }
1570
1571 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1572
1573 process_users_of_allocation(call);
1574
1575 CompileLog* log = C->log();
1576 if (log != nullptr) {
1577 const TypeInstPtr* t = nullptr;
1578 if (call->is_boxing_method()) {
1579 const TypeTuple* range = call->tf()->range_sig();
1580 t = range->field_at(TypeFunc::Parms)->isa_instptr();
1581 } else {
1582 assert(call->is_unboxing_method(), "Unexpected call");
1583 const TypeTuple* domain = call->tf()->domain_sig();
1584 t = domain->field_at(TypeFunc::Parms)->isa_instptr();
1585 assert(!t->maybe_null(), "missing receiver null check?");
1586 }
1587 log->head("eliminate_boxing type='%d'",
1588 log->identify(t->instance_klass()));
1589 JVMState* p = call->jvms();
1590 while (p != nullptr) {
1591 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1592 p = p->caller();
1593 }
1594 log->tail("eliminate_boxing");
1595 }
1596
1597 #ifndef PRODUCT
1598 if (PrintEliminateAllocations) {
1599 tty->print("++++ Eliminated: %d ", call->_idx);
1600 call->method()->print_short_name(tty);
1601 tty->cr();
1602 }
1603 #endif
1604
1605 return true;
1606 }
1607
1608 Node* PhaseMacroExpand::make_load_raw(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1609 Node* adr = off_heap_plus_addr(base, offset);
1610 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1611 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1612 transform_later(value);
1613 return value;
1614 }
1615
1616
1617 Node* PhaseMacroExpand::make_store_raw(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1618 Node* adr = off_heap_plus_addr(base, offset);
1619 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1620 transform_later(mem);
1621 return mem;
1622 }
1623
1624 //=============================================================================
1625 //
1626 // A L L O C A T I O N
1627 //
1628 // Allocation attempts to be fast in the case of frequent small objects.
1629 // It breaks down like this:
1630 //
1631 // 1) Size in doublewords is computed. This is a constant for objects and
1632 // variable for most arrays. Doubleword units are used to avoid size
1633 // overflow of huge doubleword arrays. We need doublewords in the end for
1634 // rounding.
1635 //
1636 // 2) Size is checked for being 'too large'. Too-large allocations will go
1637 // the slow path into the VM. The slow path can throw any required
1638 // exceptions, and does all the special checks for very large arrays. The
1639 // size test can constant-fold away for objects. For objects with
1640 // finalizers it constant-folds the otherway: you always go slow with
1641 // finalizers.
1642 //
1643 // 3) If NOT using TLABs, this is the contended loop-back point.
1644 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1645 //
1646 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1647 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1648 // "size*8" we always enter the VM, where "largish" is a constant picked small
1649 // enough that there's always space between the eden max and 4Gig (old space is
1650 // there so it's quite large) and large enough that the cost of entering the VM
1651 // is dwarfed by the cost to initialize the space.
1652 //
1653 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1654 // down. If contended, repeat at step 3. If using TLABs normal-store
1655 // adjusted heap top back down; there is no contention.
1656 //
1657 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1658 // fields.
1659 //
1660 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1661 // oop flavor.
1662 //
1663 //=============================================================================
1664 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1665 // Allocations bigger than this always go the slow route.
1666 // This value must be small enough that allocation attempts that need to
1667 // trigger exceptions go the slow route. Also, it must be small enough so
1668 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1669 //=============================================================================j//
1670 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1671 // The allocator will coalesce int->oop copies away. See comment in
1672 // coalesce.cpp about how this works. It depends critically on the exact
1673 // code shape produced here, so if you are changing this code shape
1674 // make sure the GC info for the heap-top is correct in and around the
1675 // slow-path call.
1676 //
1677
1678 void PhaseMacroExpand::expand_allocate_common(
1679 AllocateNode* alloc, // allocation node to be expanded
1680 Node* length, // array length for an array allocation
1681 Node* init_val, // value to initialize the array with
1682 const TypeFunc* slow_call_type, // Type of slow call
1683 address slow_call_address, // Address of slow call
1684 Node* valid_length_test // whether length is valid or not
1685 )
1686 {
1687 Node* ctrl = alloc->in(TypeFunc::Control);
1688 Node* mem = alloc->in(TypeFunc::Memory);
1689 Node* i_o = alloc->in(TypeFunc::I_O);
1690 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1691 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1692 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1693 assert(ctrl != nullptr, "must have control");
1694
1695 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1696 // they will not be used if "always_slow" is set
1697 enum { slow_result_path = 1, fast_result_path = 2 };
1698 Node *result_region = nullptr;
1699 Node *result_phi_rawmem = nullptr;
1700 Node *result_phi_rawoop = nullptr;
1701 Node *result_phi_i_o = nullptr;
1702
1703 // The initial slow comparison is a size check, the comparison
1704 // we want to do is a BoolTest::gt
1705 bool expand_fast_path = true;
1706 int tv = _igvn.find_int_con(initial_slow_test, -1);
1707 if (tv >= 0) {
1708 // InitialTest has constant result
1709 // 0 - can fit in TLAB
1710 // 1 - always too big or negative
1711 assert(tv <= 1, "0 or 1 if a constant");
1712 expand_fast_path = (tv == 0);
1713 initial_slow_test = nullptr;
1714 } else {
1715 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1716 }
1717
1718 if (!UseTLAB) {
1719 // Force slow-path allocation
1720 expand_fast_path = false;
1721 initial_slow_test = nullptr;
1722 }
1723
1724 // ArrayCopyNode right after an allocation operates on the raw result projection for the Allocate node so it's not
1725 // safe to remove such an allocation even if it has no result cast.
1726 bool allocation_has_use = (alloc->result_cast() != nullptr) || (alloc->initialization() != nullptr && alloc->initialization()->is_complete_with_arraycopy());
1727 if (!allocation_has_use) {
1728 InitializeNode* init = alloc->initialization();
1729 if (init != nullptr) {
1730 init->remove(&_igvn);
1731 }
1732 if (expand_fast_path && (initial_slow_test == nullptr)) {
1733 // Remove allocation node and return.
1734 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1735 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1736 #ifndef PRODUCT
1737 if (PrintEliminateAllocations) {
1738 tty->print("NotUsed ");
1739 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1740 if (res != nullptr) {
1741 res->dump();
1742 } else {
1743 alloc->dump();
1744 }
1745 }
1746 #endif
1747 yank_alloc_node(alloc);
1748 return;
1749 }
1750 }
1751
1752 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1753 Node *slow_region = nullptr;
1754 Node *toobig_false = ctrl;
1755
1756 // generate the initial test if necessary
1757 if (initial_slow_test != nullptr ) {
1758 assert (expand_fast_path, "Only need test if there is a fast path");
1759 slow_region = new RegionNode(3);
1760
1761 // Now make the initial failure test. Usually a too-big test but
1762 // might be a TRUE for finalizers.
1763 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1764 transform_later(toobig_iff);
1765 // Plug the failing-too-big test into the slow-path region
1766 Node* toobig_true = new IfTrueNode(toobig_iff);
1767 transform_later(toobig_true);
1768 slow_region ->init_req( too_big_or_final_path, toobig_true );
1769 toobig_false = new IfFalseNode(toobig_iff);
1770 transform_later(toobig_false);
1771 } else {
1772 // No initial test, just fall into next case
1773 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1774 toobig_false = ctrl;
1775 DEBUG_ONLY(slow_region = NodeSentinel);
1776 }
1777
1778 // If we are here there are several possibilities
1779 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1780 // no_initial_check means a constant allocation.
1781 // - If check always evaluates to false -> expand_fast_path is false (see above)
1782 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1783 // if !allocation_has_use the fast path is empty
1784 // if !allocation_has_use && no_initial_check
1785 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1786 // removed by yank_alloc_node above.
1787
1788 Node *slow_mem = mem; // save the current memory state for slow path
1789 // generate the fast allocation code unless we know that the initial test will always go slow
1790 if (expand_fast_path) {
1791 // Fast path modifies only raw memory.
1792 if (mem->is_MergeMem()) {
1793 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1794 }
1795
1796 // allocate the Region and Phi nodes for the result
1797 result_region = new RegionNode(3);
1798 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1799 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1800
1801 // Grab regular I/O before optional prefetch may change it.
1802 // Slow-path does no I/O so just set it to the original I/O.
1803 result_phi_i_o->init_req(slow_result_path, i_o);
1804
1805 // Name successful fast-path variables
1806 Node* fast_oop_ctrl;
1807 Node* fast_oop_rawmem;
1808
1809 if (allocation_has_use) {
1810 Node* needgc_ctrl = nullptr;
1811 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1812
1813 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1814 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1815 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1816 fast_oop_ctrl, fast_oop_rawmem,
1817 prefetch_lines);
1818
1819 if (initial_slow_test != nullptr) {
1820 // This completes all paths into the slow merge point
1821 slow_region->init_req(need_gc_path, needgc_ctrl);
1822 transform_later(slow_region);
1823 } else {
1824 // No initial slow path needed!
1825 // Just fall from the need-GC path straight into the VM call.
1826 slow_region = needgc_ctrl;
1827 }
1828
1829 InitializeNode* init = alloc->initialization();
1830 fast_oop_rawmem = initialize_object(alloc,
1831 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1832 klass_node, length, size_in_bytes);
1833 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1834 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1835
1836 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1837 } else {
1838 assert (initial_slow_test != nullptr, "sanity");
1839 fast_oop_ctrl = toobig_false;
1840 fast_oop_rawmem = mem;
1841 transform_later(slow_region);
1842 }
1843
1844 // Plug in the successful fast-path into the result merge point
1845 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1846 result_phi_i_o ->init_req(fast_result_path, i_o);
1847 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1848 } else {
1849 slow_region = ctrl;
1850 result_phi_i_o = i_o; // Rename it to use in the following code.
1851 }
1852
1853 // Generate slow-path call
1854 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1855 OptoRuntime::stub_name(slow_call_address),
1856 TypePtr::BOTTOM);
1857 call->init_req(TypeFunc::Control, slow_region);
1858 call->init_req(TypeFunc::I_O, top()); // does no i/o
1859 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1860 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1861 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1862
1863 call->init_req(TypeFunc::Parms+0, klass_node);
1864 if (length != nullptr) {
1865 call->init_req(TypeFunc::Parms+1, length);
1866 if (init_val != nullptr) {
1867 call->init_req(TypeFunc::Parms+2, init_val);
1868 }
1869 }
1870
1871 // Copy debug information and adjust JVMState information, then replace
1872 // allocate node with the call
1873 call->copy_call_debug_info(&_igvn, alloc);
1874 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1875 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1876 // path dies).
1877 if (valid_length_test != nullptr) {
1878 call->add_req(valid_length_test);
1879 }
1880 if (expand_fast_path) {
1881 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1882 } else {
1883 // Hook i_o projection to avoid its elimination during allocation
1884 // replacement (when only a slow call is generated).
1885 call->set_req(TypeFunc::I_O, result_phi_i_o);
1886 }
1887 _igvn.replace_node(alloc, call);
1888 transform_later(call);
1889
1890 // Identify the output projections from the allocate node and
1891 // adjust any references to them.
1892 // The control and io projections look like:
1893 //
1894 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1895 // Allocate Catch
1896 // ^---Proj(io) <-------+ ^---CatchProj(io)
1897 //
1898 // We are interested in the CatchProj nodes.
1899 //
1900 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1901
1902 // An allocate node has separate memory projections for the uses on
1903 // the control and i_o paths. Replace the control memory projection with
1904 // result_phi_rawmem (unless we are only generating a slow call when
1905 // both memory projections are combined)
1906 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1907 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1908 }
1909 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1910 // catchall_memproj so we end up with a call that has only 1 memory projection.
1911 if (_callprojs->catchall_memproj != nullptr) {
1912 if (_callprojs->fallthrough_memproj == nullptr) {
1913 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1914 transform_later(_callprojs->fallthrough_memproj);
1915 }
1916 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1917 _igvn.remove_dead_node(_callprojs->catchall_memproj, PhaseIterGVN::NodeOrigin::Graph);
1918 }
1919
1920 // An allocate node has separate i_o projections for the uses on the control
1921 // and i_o paths. Always replace the control i_o projection with result i_o
1922 // otherwise incoming i_o become dead when only a slow call is generated
1923 // (it is different from memory projections where both projections are
1924 // combined in such case).
1925 if (_callprojs->fallthrough_ioproj != nullptr) {
1926 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1927 }
1928 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1929 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1930 if (_callprojs->catchall_ioproj != nullptr) {
1931 if (_callprojs->fallthrough_ioproj == nullptr) {
1932 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1933 transform_later(_callprojs->fallthrough_ioproj);
1934 }
1935 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1936 _igvn.remove_dead_node(_callprojs->catchall_ioproj, PhaseIterGVN::NodeOrigin::Graph);
1937 }
1938
1939 // if we generated only a slow call, we are done
1940 if (!expand_fast_path) {
1941 // Now we can unhook i_o.
1942 if (result_phi_i_o->outcnt() > 1) {
1943 call->set_req(TypeFunc::I_O, top());
1944 } else {
1945 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1946 // Case of new array with negative size known during compilation.
1947 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1948 // following code since call to runtime will throw exception.
1949 // As result there will be no users of i_o after the call.
1950 // Leave i_o attached to this call to avoid problems in preceding graph.
1951 }
1952 return;
1953 }
1954
1955 if (_callprojs->fallthrough_catchproj != nullptr) {
1956 ctrl = _callprojs->fallthrough_catchproj->clone();
1957 transform_later(ctrl);
1958 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1959 } else {
1960 ctrl = top();
1961 }
1962 Node *slow_result;
1963 if (_callprojs->resproj[0] == nullptr) {
1964 // no uses of the allocation result
1965 slow_result = top();
1966 } else {
1967 slow_result = _callprojs->resproj[0]->clone();
1968 transform_later(slow_result);
1969 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1970 }
1971
1972 // Plug slow-path into result merge point
1973 result_region->init_req( slow_result_path, ctrl);
1974 transform_later(result_region);
1975 if (allocation_has_use) {
1976 result_phi_rawoop->init_req(slow_result_path, slow_result);
1977 transform_later(result_phi_rawoop);
1978 }
1979 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1980 transform_later(result_phi_rawmem);
1981 transform_later(result_phi_i_o);
1982 // This completes all paths into the result merge point
1983 }
1984
1985 // Remove alloc node that has no uses.
1986 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1987 Node* ctrl = alloc->in(TypeFunc::Control);
1988 Node* mem = alloc->in(TypeFunc::Memory);
1989 Node* i_o = alloc->in(TypeFunc::I_O);
1990
1991 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1992 if (_callprojs->resproj[0] != nullptr) {
1993 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1994 Node* use = _callprojs->resproj[0]->fast_out(i);
1995 use->isa_MemBar()->remove(&_igvn);
1996 --imax;
1997 --i; // back up iterator
1998 }
1999 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
2000 _igvn.remove_dead_node(_callprojs->resproj[0], PhaseIterGVN::NodeOrigin::Graph);
2001 }
2002 if (_callprojs->fallthrough_catchproj != nullptr) {
2003 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
2004 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj, PhaseIterGVN::NodeOrigin::Graph);
2005 }
2006 if (_callprojs->catchall_catchproj != nullptr) {
2007 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
2008 _callprojs->catchall_catchproj->set_req(0, top());
2009 }
2010 if (_callprojs->fallthrough_proj != nullptr) {
2011 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
2012 _igvn.remove_dead_node(catchnode, PhaseIterGVN::NodeOrigin::Graph);
2013 _igvn.remove_dead_node(_callprojs->fallthrough_proj, PhaseIterGVN::NodeOrigin::Graph);
2014 }
2015 if (_callprojs->fallthrough_memproj != nullptr) {
2016 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
2017 _igvn.remove_dead_node(_callprojs->fallthrough_memproj, PhaseIterGVN::NodeOrigin::Graph);
2018 }
2019 if (_callprojs->fallthrough_ioproj != nullptr) {
2020 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
2021 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj, PhaseIterGVN::NodeOrigin::Graph);
2022 }
2023 if (_callprojs->catchall_memproj != nullptr) {
2024 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
2025 _callprojs->catchall_memproj->set_req(0, top());
2026 }
2027 if (_callprojs->catchall_ioproj != nullptr) {
2028 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
2029 _callprojs->catchall_ioproj->set_req(0, top());
2030 }
2031 #ifndef PRODUCT
2032 if (PrintEliminateAllocations) {
2033 if (alloc->is_AllocateArray()) {
2034 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
2035 } else {
2036 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
2037 }
2038 }
2039 #endif
2040 _igvn.remove_dead_node(alloc, PhaseIterGVN::NodeOrigin::Graph);
2041 }
2042
2043 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
2044 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
2045 // If initialization is performed by an array copy, any required
2046 // MemBarStoreStore was already added. If the object does not
2047 // escape no need for a MemBarStoreStore. If the object does not
2048 // escape in its initializer and memory barrier (MemBarStoreStore or
2049 // stronger) is already added at exit of initializer, also no need
2050 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
2051 // so that stores that initialize this object can't be reordered
2052 // with a subsequent store that makes this object accessible by
2053 // other threads.
2054 // Other threads include java threads and JVM internal threads
2055 // (for example concurrent GC threads). Current concurrent GC
2056 // implementation: G1 will not scan newly created object,
2057 // so it's safe to skip storestore barrier when allocation does
2058 // not escape.
2059 if (!alloc->does_not_escape_thread() &&
2060 !alloc->is_allocation_MemBar_redundant() &&
2061 (init == nullptr || !init->is_complete_with_arraycopy())) {
2062 if (init == nullptr || init->req() < InitializeNode::RawStores) {
2063 // No InitializeNode or no stores captured by zeroing
2064 // elimination. Simply add the MemBarStoreStore after object
2065 // initialization.
2066 // What we want is to prevent the compiler and the CPU from re-ordering the stores that initialize this object
2067 // with subsequent stores to any slice. As a consequence, this MemBar should capture the entire memory state at
2068 // this point in the IR and produce a new memory state that should cover all slices. However, the Initialize node
2069 // only captures/produces a partial memory state making it complicated to insert such a MemBar. Because
2070 // re-ordering by the compiler can't happen by construction (a later Store that publishes the just allocated
2071 // object reference is indirectly control dependent on the Initialize node), preventing reordering by the CPU is
2072 // sufficient. For that a MemBar on the raw memory slice is good enough.
2073 // If init is null, this allocation does have an InitializeNode but this logic can't locate it (see comment in
2074 // PhaseMacroExpand::initialize_object()).
2075 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
2076 transform_later(mb);
2077
2078 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
2079 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
2080 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
2081 transform_later(fast_oop_ctrl);
2082 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
2083 transform_later(fast_oop_rawmem);
2084 } else {
2085 // Add the MemBarStoreStore after the InitializeNode so that
2086 // all stores performing the initialization that were moved
2087 // before the InitializeNode happen before the storestore
2088 // barrier.
2089
2090 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
2091
2092 // See comment above that explains why a raw memory MemBar is good enough.
2093 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
2094 transform_later(mb);
2095
2096 Node* ctrl = new ProjNode(init, TypeFunc::Control);
2097 transform_later(ctrl);
2098 Node* old_raw_mem_proj = nullptr;
2099 auto find_raw_mem = [&](ProjNode* proj) {
2100 if (C->get_alias_index(proj->adr_type()) == Compile::AliasIdxRaw) {
2101 assert(old_raw_mem_proj == nullptr, "only one expected");
2102 old_raw_mem_proj = proj;
2103 }
2104 };
2105 init->for_each_proj(find_raw_mem, TypeFunc::Memory);
2106 assert(old_raw_mem_proj != nullptr, "should have found raw mem Proj");
2107 Node* raw_mem_proj = new ProjNode(init, TypeFunc::Memory);
2108 transform_later(raw_mem_proj);
2109
2110 // The MemBarStoreStore depends on control and memory coming
2111 // from the InitializeNode
2112 mb->init_req(TypeFunc::Memory, raw_mem_proj);
2113 mb->init_req(TypeFunc::Control, ctrl);
2114
2115 ctrl = new ProjNode(mb, TypeFunc::Control);
2116 transform_later(ctrl);
2117 Node* mem = new ProjNode(mb, TypeFunc::Memory);
2118 transform_later(mem);
2119
2120 // All nodes that depended on the InitializeNode for control
2121 // and memory must now depend on the MemBarNode that itself
2122 // depends on the InitializeNode
2123 if (init_ctrl != nullptr) {
2124 _igvn.replace_node(init_ctrl, ctrl);
2125 }
2126 _igvn.replace_node(old_raw_mem_proj, mem);
2127 }
2128 }
2129 }
2130
2131 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
2132 Node*& ctrl, Node*& rawmem) {
2133 if (C->env()->dtrace_alloc_probes()) {
2134 // Slow-path call
2135 int size = TypeFunc::Parms + 2;
2136 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
2137 CAST_FROM_FN_PTR(address,
2138 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
2139 "dtrace_object_alloc",
2140 TypeRawPtr::BOTTOM);
2141
2142 // Get base of thread-local storage area
2143 Node* thread = new ThreadLocalNode();
2144 transform_later(thread);
2145
2146 call->init_req(TypeFunc::Parms + 0, thread);
2147 call->init_req(TypeFunc::Parms + 1, oop);
2148 call->init_req(TypeFunc::Control, ctrl);
2149 call->init_req(TypeFunc::I_O , top()); // does no i/o
2150 call->init_req(TypeFunc::Memory , rawmem);
2151 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
2152 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
2153 transform_later(call);
2154 ctrl = new ProjNode(call, TypeFunc::Control);
2155 transform_later(ctrl);
2156 rawmem = new ProjNode(call, TypeFunc::Memory);
2157 transform_later(rawmem);
2158 }
2159 }
2160
2161 // Helper for PhaseMacroExpand::expand_allocate_common.
2162 // Initializes the newly-allocated storage.
2163 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
2164 Node* control, Node* rawmem, Node* object,
2165 Node* klass_node, Node* length,
2166 Node* size_in_bytes) {
2167 InitializeNode* init = alloc->initialization();
2168 // Store the klass & mark bits
2169 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
2170 if (!mark_node->is_Con()) {
2171 transform_later(mark_node);
2172 }
2173 rawmem = make_store_raw(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
2174
2175 if (!UseCompactObjectHeaders) {
2176 rawmem = make_store_raw(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2177 }
2178 int header_size = alloc->minimum_header_size(); // conservatively small
2179
2180 // Array length
2181 if (length != nullptr) { // Arrays need length field
2182 rawmem = make_store_raw(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
2183 // conservatively small header size:
2184 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2185 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
2186 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
2187 if (is_reference_type(elem, true)) {
2188 elem = T_OBJECT;
2189 }
2190 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2191 }
2192 }
2193
2194 // Clear the object body, if necessary.
2195 if (init == nullptr) {
2196 // The init has somehow disappeared; be cautious and clear everything.
2197 //
2198 // This can happen if a node is allocated but an uncommon trap occurs
2199 // immediately. In this case, the Initialize gets associated with the
2200 // trap, and may be placed in a different (outer) loop, if the Allocate
2201 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2202 // there can be two Allocates to one Initialize. The answer in all these
2203 // edge cases is safety first. It is always safe to clear immediately
2204 // within an Allocate, and then (maybe or maybe not) clear some more later.
2205 if (!(UseTLAB && ZeroTLAB)) {
2206 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2207 alloc->in(AllocateNode::InitValue),
2208 alloc->in(AllocateNode::RawInitValue),
2209 header_size, size_in_bytes,
2210 true,
2211 &_igvn);
2212 }
2213 } else {
2214 if (!init->is_complete()) {
2215 // Try to win by zeroing only what the init does not store.
2216 // We can also try to do some peephole optimizations,
2217 // such as combining some adjacent subword stores.
2218 rawmem = init->complete_stores(control, rawmem, object,
2219 header_size, size_in_bytes, &_igvn);
2220 }
2221 // We have no more use for this link, since the AllocateNode goes away:
2222 init->set_req(InitializeNode::RawAddress, top());
2223 // (If we keep the link, it just confuses the register allocator,
2224 // who thinks he sees a real use of the address by the membar.)
2225 }
2226
2227 return rawmem;
2228 }
2229
2230 // Generate prefetch instructions for next allocations.
2231 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
2232 Node*& contended_phi_rawmem,
2233 Node* old_eden_top, Node* new_eden_top,
2234 intx lines) {
2235 enum { fall_in_path = 1, pf_path = 2 };
2236 if (UseTLAB && AllocatePrefetchStyle == 2) {
2237 // Generate prefetch allocation with watermark check.
2238 // As an allocation hits the watermark, we will prefetch starting
2239 // at a "distance" away from watermark.
2240
2241 Node* pf_region = new RegionNode(3);
2242 Node* pf_phi_rawmem = new PhiNode(pf_region, Type::MEMORY,
2243 TypeRawPtr::BOTTOM);
2244 // I/O is used for Prefetch
2245 Node* pf_phi_abio = new PhiNode(pf_region, Type::ABIO);
2246
2247 Node* thread = new ThreadLocalNode();
2248 transform_later(thread);
2249
2250 Node* eden_pf_adr = AddPNode::make_off_heap(thread,
2251 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())));
2252 transform_later(eden_pf_adr);
2253
2254 Node* old_pf_wm = new LoadPNode(needgc_false,
2255 contended_phi_rawmem, eden_pf_adr,
2256 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2257 MemNode::unordered);
2258 transform_later(old_pf_wm);
2259
2260 // check against new_eden_top
2261 Node* need_pf_cmp = new CmpPNode(new_eden_top, old_pf_wm);
2262 transform_later(need_pf_cmp);
2263 Node* need_pf_bol = new BoolNode(need_pf_cmp, BoolTest::ge);
2264 transform_later(need_pf_bol);
2265 IfNode* need_pf_iff = new IfNode(needgc_false, need_pf_bol,
2266 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2267 transform_later(need_pf_iff);
2268
2269 // true node, add prefetchdistance
2270 Node* need_pf_true = new IfTrueNode(need_pf_iff);
2271 transform_later(need_pf_true);
2272
2273 Node* need_pf_false = new IfFalseNode(need_pf_iff);
2274 transform_later(need_pf_false);
2275
2276 Node* new_pf_wmt = AddPNode::make_off_heap(old_pf_wm,
2277 _igvn.MakeConX(AllocatePrefetchDistance));
2278 transform_later(new_pf_wmt);
2279 new_pf_wmt->set_req(0, need_pf_true);
2280
2281 Node* store_new_wmt = new StorePNode(need_pf_true,
2282 contended_phi_rawmem, eden_pf_adr,
2283 TypeRawPtr::BOTTOM, new_pf_wmt,
2284 MemNode::unordered);
2285 transform_later(store_new_wmt);
2286
2287 // adding prefetches
2288 pf_phi_abio->init_req(fall_in_path, i_o);
2289
2290 Node* prefetch_adr;
2291 Node* prefetch;
2292 uint step_size = AllocatePrefetchStepSize;
2293 uint distance = 0;
2294
2295 for (intx i = 0; i < lines; i++) {
2296 prefetch_adr = AddPNode::make_off_heap(new_pf_wmt,
2297 _igvn.MakeConX(distance));
2298 transform_later(prefetch_adr);
2299 prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2300 transform_later(prefetch);
2301 distance += step_size;
2302 i_o = prefetch;
2303 }
2304 pf_phi_abio->set_req(pf_path, i_o);
2305
2306 pf_region->init_req(fall_in_path, need_pf_false);
2307 pf_region->init_req(pf_path, need_pf_true);
2308
2309 pf_phi_rawmem->init_req(fall_in_path, contended_phi_rawmem);
2310 pf_phi_rawmem->init_req(pf_path, store_new_wmt);
2311
2312 transform_later(pf_region);
2313 transform_later(pf_phi_rawmem);
2314 transform_later(pf_phi_abio);
2315
2316 needgc_false = pf_region;
2317 contended_phi_rawmem = pf_phi_rawmem;
2318 i_o = pf_phi_abio;
2319 } else if (UseTLAB && AllocatePrefetchStyle == 3) {
2320 // Insert a prefetch instruction for each allocation.
2321 // This code is used to generate 1 prefetch instruction per cache line.
2322
2323 // Generate several prefetch instructions.
2324 uint step_size = AllocatePrefetchStepSize;
2325 uint distance = AllocatePrefetchDistance;
2326
2327 // Next cache address.
2328 Node* cache_adr = AddPNode::make_off_heap(old_eden_top,
2329 _igvn.MakeConX(step_size + distance));
2330 transform_later(cache_adr);
2331 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2332 transform_later(cache_adr);
2333 // Address is aligned to execute prefetch to the beginning of cache line size.
2334 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2335 cache_adr = new AndXNode(cache_adr, mask);
2336 transform_later(cache_adr);
2337 cache_adr = new CastX2PNode(cache_adr);
2338 transform_later(cache_adr);
2339
2340 // Prefetch
2341 Node* prefetch = new PrefetchAllocationNode(contended_phi_rawmem, cache_adr);
2342 prefetch->set_req(0, needgc_false);
2343 transform_later(prefetch);
2344 contended_phi_rawmem = prefetch;
2345 Node* prefetch_adr;
2346 distance = step_size;
2347 for (intx i = 1; i < lines; i++) {
2348 prefetch_adr = AddPNode::make_off_heap(cache_adr,
2349 _igvn.MakeConX(distance));
2350 transform_later(prefetch_adr);
2351 prefetch = new PrefetchAllocationNode(contended_phi_rawmem, prefetch_adr);
2352 transform_later(prefetch);
2353 distance += step_size;
2354 contended_phi_rawmem = prefetch;
2355 }
2356 } else if (AllocatePrefetchStyle > 0) {
2357 // Insert a prefetch for each allocation only on the fast-path
2358 Node* prefetch_adr;
2359 Node* prefetch;
2360 // Generate several prefetch instructions.
2361 uint step_size = AllocatePrefetchStepSize;
2362 uint distance = AllocatePrefetchDistance;
2363 for (intx i = 0; i < lines; i++) {
2364 prefetch_adr = AddPNode::make_off_heap(new_eden_top,
2365 _igvn.MakeConX(distance));
2366 transform_later(prefetch_adr);
2367 prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2368 // Do not let it float too high, since if eden_top == eden_end,
2369 // both might be null.
2370 if (i == 0) { // Set control for first prefetch, next follows it
2371 prefetch->init_req(0, needgc_false);
2372 }
2373 transform_later(prefetch);
2374 distance += step_size;
2375 i_o = prefetch;
2376 }
2377 }
2378 return i_o;
2379 }
2380
2381
2382 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2383 expand_allocate_common(alloc, nullptr, nullptr,
2384 OptoRuntime::new_instance_Type(),
2385 OptoRuntime::new_instance_Java(), nullptr);
2386 }
2387
2388 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2389 Node* length = alloc->in(AllocateNode::ALength);
2390 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2391 InitializeNode* init = alloc->initialization();
2392 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2393 Node* init_value = alloc->in(AllocateNode::InitValue);
2394 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2395 assert(!ary_klass_t || !ary_klass_t->klass_is_exact() || !ary_klass_t->exact_klass()->is_obj_array_klass() ||
2396 ary_klass_t->is_refined_type(), "Must be a refined array klass");
2397 const TypeFunc* slow_call_type;
2398 address slow_call_address; // Address of slow call
2399 if (init != nullptr && init->is_complete_with_arraycopy() &&
2400 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2401 // Don't zero type array during slow allocation in VM since
2402 // it will be initialized later by arraycopy in compiled code.
2403 slow_call_address = OptoRuntime::new_array_nozero_Java();
2404 slow_call_type = OptoRuntime::new_array_nozero_Type();
2405 } else {
2406 slow_call_address = OptoRuntime::new_array_Java();
2407 slow_call_type = OptoRuntime::new_array_Type();
2408
2409 if (init_value == nullptr) {
2410 init_value = _igvn.zerocon(T_OBJECT);
2411 } else if (UseCompressedOops) {
2412 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2413 }
2414 }
2415 expand_allocate_common(alloc, length, init_value,
2416 slow_call_type,
2417 slow_call_address, valid_length_test);
2418 }
2419
2420 //-------------------mark_eliminated_box----------------------------------
2421 //
2422 // During EA obj may point to several objects but after few ideal graph
2423 // transformations (CCP) it may point to only one non escaping object
2424 // (but still using phi), corresponding locks and unlocks will be marked
2425 // for elimination. Later obj could be replaced with a new node (new phi)
2426 // and which does not have escape information. And later after some graph
2427 // reshape other locks and unlocks (which were not marked for elimination
2428 // before) are connected to this new obj (phi) but they still will not be
2429 // marked for elimination since new obj has no escape information.
2430 // Mark all associated (same box and obj) lock and unlock nodes for
2431 // elimination if some of them marked already.
2432 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2433 BoxLockNode* oldbox = box->as_BoxLock();
2434 if (oldbox->is_eliminated()) {
2435 return; // This BoxLock node was processed already.
2436 }
2437 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2438 // New implementation (EliminateNestedLocks) has separate BoxLock
2439 // node for each locked region so mark all associated locks/unlocks as
2440 // eliminated even if different objects are referenced in one locked region
2441 // (for example, OSR compilation of nested loop inside locked scope).
2442 if (EliminateNestedLocks ||
2443 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2444 // Box is used only in one lock region. Mark this box as eliminated.
2445 oldbox->set_local(); // This verifies correct state of BoxLock
2446 _igvn.hash_delete(oldbox);
2447 oldbox->set_eliminated(); // This changes box's hash value
2448 _igvn.hash_insert(oldbox);
2449
2450 for (uint i = 0; i < oldbox->outcnt(); i++) {
2451 Node* u = oldbox->raw_out(i);
2452 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2453 AbstractLockNode* alock = u->as_AbstractLock();
2454 // Check lock's box since box could be referenced by Lock's debug info.
2455 if (alock->box_node() == oldbox) {
2456 // Mark eliminated all related locks and unlocks.
2457 #ifdef ASSERT
2458 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2459 #endif
2460 alock->set_non_esc_obj();
2461 }
2462 }
2463 }
2464 return;
2465 }
2466
2467 // Create new "eliminated" BoxLock node and use it in monitor debug info
2468 // instead of oldbox for the same object.
2469 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2470
2471 // Note: BoxLock node is marked eliminated only here and it is used
2472 // to indicate that all associated lock and unlock nodes are marked
2473 // for elimination.
2474 newbox->set_local(); // This verifies correct state of BoxLock
2475 newbox->set_eliminated();
2476 transform_later(newbox);
2477
2478 // Replace old box node with new box for all users of the same object.
2479 for (uint i = 0; i < oldbox->outcnt();) {
2480 bool next_edge = true;
2481
2482 Node* u = oldbox->raw_out(i);
2483 if (u->is_AbstractLock()) {
2484 AbstractLockNode* alock = u->as_AbstractLock();
2485 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2486 // Replace Box and mark eliminated all related locks and unlocks.
2487 #ifdef ASSERT
2488 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2489 #endif
2490 alock->set_non_esc_obj();
2491 _igvn.rehash_node_delayed(alock);
2492 alock->set_box_node(newbox);
2493 next_edge = false;
2494 }
2495 }
2496 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2497 FastLockNode* flock = u->as_FastLock();
2498 assert(flock->box_node() == oldbox, "sanity");
2499 _igvn.rehash_node_delayed(flock);
2500 flock->set_box_node(newbox);
2501 next_edge = false;
2502 }
2503
2504 // Replace old box in monitor debug info.
2505 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2506 SafePointNode* sfn = u->as_SafePoint();
2507 JVMState* youngest_jvms = sfn->jvms();
2508 int max_depth = youngest_jvms->depth();
2509 for (int depth = 1; depth <= max_depth; depth++) {
2510 JVMState* jvms = youngest_jvms->of_depth(depth);
2511 int num_mon = jvms->nof_monitors();
2512 // Loop over monitors
2513 for (int idx = 0; idx < num_mon; idx++) {
2514 Node* obj_node = sfn->monitor_obj(jvms, idx);
2515 Node* box_node = sfn->monitor_box(jvms, idx);
2516 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2517 int j = jvms->monitor_box_offset(idx);
2518 _igvn.replace_input_of(u, j, newbox);
2519 next_edge = false;
2520 }
2521 }
2522 }
2523 }
2524 if (next_edge) i++;
2525 }
2526 }
2527
2528 //-----------------------mark_eliminated_locking_nodes-----------------------
2529 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2530 if (!alock->is_balanced()) {
2531 return; // Can't do any more elimination for this locking region
2532 }
2533 if (EliminateNestedLocks) {
2534 if (alock->is_nested()) {
2535 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2536 return;
2537 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2538 // Only Lock node has JVMState needed here.
2539 // Not that preceding claim is documented anywhere else.
2540 if (alock->jvms() != nullptr) {
2541 if (alock->as_Lock()->is_nested_lock_region()) {
2542 // Mark eliminated related nested locks and unlocks.
2543 Node* obj = alock->obj_node();
2544 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2545 assert(!box_node->is_eliminated(), "should not be marked yet");
2546 // Note: BoxLock node is marked eliminated only here
2547 // and it is used to indicate that all associated lock
2548 // and unlock nodes are marked for elimination.
2549 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2550 for (uint i = 0; i < box_node->outcnt(); i++) {
2551 Node* u = box_node->raw_out(i);
2552 if (u->is_AbstractLock()) {
2553 alock = u->as_AbstractLock();
2554 if (alock->box_node() == box_node) {
2555 // Verify that this Box is referenced only by related locks.
2556 assert(alock->obj_node()->eqv_uncast(obj), "");
2557 // Mark all related locks and unlocks.
2558 #ifdef ASSERT
2559 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2560 #endif
2561 alock->set_nested();
2562 }
2563 }
2564 }
2565 } else {
2566 #ifdef ASSERT
2567 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2568 if (C->log() != nullptr)
2569 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2570 #endif
2571 }
2572 }
2573 return;
2574 }
2575 // Process locks for non escaping object
2576 assert(alock->is_non_esc_obj(), "");
2577 } // EliminateNestedLocks
2578
2579 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2580 // Look for all locks of this object and mark them and
2581 // corresponding BoxLock nodes as eliminated.
2582 Node* obj = alock->obj_node();
2583 for (uint j = 0; j < obj->outcnt(); j++) {
2584 Node* o = obj->raw_out(j);
2585 if (o->is_AbstractLock() &&
2586 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2587 alock = o->as_AbstractLock();
2588 Node* box = alock->box_node();
2589 // Replace old box node with new eliminated box for all users
2590 // of the same object and mark related locks as eliminated.
2591 mark_eliminated_box(box, obj);
2592 }
2593 }
2594 }
2595 }
2596
2597 // we have determined that this lock/unlock can be eliminated, we simply
2598 // eliminate the node without expanding it.
2599 //
2600 // Note: The membar's associated with the lock/unlock are currently not
2601 // eliminated. This should be investigated as a future enhancement.
2602 //
2603 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2604
2605 if (!alock->is_eliminated()) {
2606 return false;
2607 }
2608 #ifdef ASSERT
2609 if (!alock->is_coarsened()) {
2610 // Check that new "eliminated" BoxLock node is created.
2611 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2612 assert(oldbox->is_eliminated(), "should be done already");
2613 }
2614 #endif
2615
2616 alock->log_lock_optimization(C, "eliminate_lock");
2617
2618 #ifndef PRODUCT
2619 if (PrintEliminateLocks) {
2620 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2621 }
2622 #endif
2623
2624 Node* mem = alock->in(TypeFunc::Memory);
2625 Node* ctrl = alock->in(TypeFunc::Control);
2626 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2627
2628 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2629 // There are 2 projections from the lock. The lock node will
2630 // be deleted when its last use is subsumed below.
2631 assert(alock->outcnt() == 2 &&
2632 _callprojs->fallthrough_proj != nullptr &&
2633 _callprojs->fallthrough_memproj != nullptr,
2634 "Unexpected projections from Lock/Unlock");
2635
2636 Node* fallthroughproj = _callprojs->fallthrough_proj;
2637 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2638
2639 // The memory projection from a lock/unlock is RawMem
2640 // The input to a Lock is merged memory, so extract its RawMem input
2641 // (unless the MergeMem has been optimized away.)
2642 if (alock->is_Lock()) {
2643 // Search for MemBarAcquireLock node and delete it also.
2644 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2645 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2646 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2647 Node* memproj = membar->proj_out(TypeFunc::Memory);
2648 _igvn.replace_node(ctrlproj, fallthroughproj);
2649 _igvn.replace_node(memproj, memproj_fallthrough);
2650
2651 // Delete FastLock node also if this Lock node is unique user
2652 // (a loop peeling may clone a Lock node).
2653 Node* flock = alock->as_Lock()->fastlock_node();
2654 if (flock->outcnt() == 1) {
2655 assert(flock->unique_out() == alock, "sanity");
2656 _igvn.replace_node(flock, top());
2657 }
2658 }
2659
2660 // Search for MemBarReleaseLock node and delete it also.
2661 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2662 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2663 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2664 mem->is_Proj() && membar == mem->in(0), "");
2665 _igvn.replace_node(fallthroughproj, ctrl);
2666 _igvn.replace_node(memproj_fallthrough, mem);
2667 fallthroughproj = ctrl;
2668 memproj_fallthrough = mem;
2669 ctrl = membar->in(TypeFunc::Control);
2670 mem = membar->in(TypeFunc::Memory);
2671 }
2672
2673 _igvn.replace_node(fallthroughproj, ctrl);
2674 _igvn.replace_node(memproj_fallthrough, mem);
2675 return true;
2676 }
2677
2678
2679 //------------------------------expand_lock_node----------------------
2680 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2681
2682 Node* ctrl = lock->in(TypeFunc::Control);
2683 Node* mem = lock->in(TypeFunc::Memory);
2684 Node* obj = lock->obj_node();
2685 Node* box = lock->box_node();
2686 Node* flock = lock->fastlock_node();
2687
2688 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2689
2690 // Make the merge point
2691 Node *region;
2692 Node *mem_phi;
2693 Node *slow_path;
2694
2695 region = new RegionNode(3);
2696 // create a Phi for the memory state
2697 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2698
2699 // Optimize test; set region slot 2
2700 slow_path = opt_bits_test(ctrl, region, 2, flock);
2701 mem_phi->init_req(2, mem);
2702
2703 // Make slow path call
2704 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2705 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2706 obj, box, nullptr);
2707
2708 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2709
2710 // Slow path can only throw asynchronous exceptions, which are always
2711 // de-opted. So the compiler thinks the slow-call can never throw an
2712 // exception. If it DOES throw an exception we would need the debug
2713 // info removed first (since if it throws there is no monitor).
2714 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2715 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2716
2717 // Capture slow path
2718 // disconnect fall-through projection from call and create a new one
2719 // hook up users of fall-through projection to region
2720 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2721 transform_later(slow_ctrl);
2722 _igvn.hash_delete(_callprojs->fallthrough_proj);
2723 _callprojs->fallthrough_proj->disconnect_inputs(C);
2724 region->init_req(1, slow_ctrl);
2725 // region inputs are now complete
2726 transform_later(region);
2727 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2728
2729 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2730
2731 mem_phi->init_req(1, memproj);
2732
2733 transform_later(mem_phi);
2734
2735 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2736 }
2737
2738 //------------------------------expand_unlock_node----------------------
2739 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2740
2741 Node* ctrl = unlock->in(TypeFunc::Control);
2742 Node* mem = unlock->in(TypeFunc::Memory);
2743 Node* obj = unlock->obj_node();
2744 Node* box = unlock->box_node();
2745
2746 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2747
2748 // No need for a null check on unlock
2749
2750 // Make the merge point
2751 Node* region = new RegionNode(3);
2752
2753 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2754 funlock = transform_later( funlock )->as_FastUnlock();
2755 // Optimize test; set region slot 2
2756 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2757 Node *thread = transform_later(new ThreadLocalNode());
2758
2759 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2760 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2761 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2762
2763 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2764 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2765 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2766
2767 // No exceptions for unlocking
2768 // Capture slow path
2769 // disconnect fall-through projection from call and create a new one
2770 // hook up users of fall-through projection to region
2771 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2772 transform_later(slow_ctrl);
2773 _igvn.hash_delete(_callprojs->fallthrough_proj);
2774 _callprojs->fallthrough_proj->disconnect_inputs(C);
2775 region->init_req(1, slow_ctrl);
2776 // region inputs are now complete
2777 transform_later(region);
2778 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2779
2780 if (_callprojs->fallthrough_memproj != nullptr) {
2781 // create a Phi for the memory state
2782 Node* mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2783 Node* memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2784 mem_phi->init_req(1, memproj);
2785 mem_phi->init_req(2, mem);
2786 transform_later(mem_phi);
2787 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2788 }
2789 }
2790
2791 // An inline type might be returned from the call but we don't know its
2792 // type. Either we get a buffered inline type (and nothing needs to be done)
2793 // or one of the values being returned is the klass of the inline type
2794 // and we need to allocate an inline type instance of that type and
2795 // initialize it with other values being returned. In that case, we
2796 // first try a fast path allocation and initialize the value with the
2797 // inline klass's pack handler or we fall back to a runtime call.
2798 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2799 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2800 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2801 if (ret == nullptr) {
2802 return;
2803 }
2804 const TypeFunc* tf = call->_tf;
2805 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2806 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2807 call->_tf = new_tf;
2808 // Make sure the change of type is applied before projections are processed by igvn
2809 _igvn.set_type(call, call->Value(&_igvn));
2810 _igvn.set_type(ret, ret->Value(&_igvn));
2811
2812 // Before any new projection is added:
2813 CallProjections* projs = call->extract_projections(true, true);
2814
2815 // Create temporary hook nodes that will be replaced below.
2816 // Add an input to prevent hook nodes from being dead.
2817 Node* ctl = new Node(call);
2818 Node* mem = new Node(ctl);
2819 Node* io = new Node(ctl);
2820 Node* ex_ctl = new Node(ctl);
2821 Node* ex_mem = new Node(ctl);
2822 Node* ex_io = new Node(ctl);
2823 Node* res = new Node(ctl);
2824
2825 // Allocate a new buffered inline type only if a new one is not returned
2826 Node* cast = transform_later(new CastP2XNode(ctl, res));
2827 Node* mask = MakeConX(0x1);
2828 Node* masked = transform_later(new AndXNode(cast, mask));
2829 Node* cmp = transform_later(new CmpXNode(masked, mask));
2830 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2831 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2832 transform_later(allocation_iff);
2833 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2834 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2835 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2836
2837 // Try to allocate a new buffered inline instance either from TLAB or eden space
2838 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2839 CallLeafNoFPNode* handler_call;
2840 const bool alloc_in_place = UseTLAB;
2841 if (alloc_in_place) {
2842 Node* fast_oop_ctrl = nullptr;
2843 Node* fast_oop_rawmem = nullptr;
2844 Node* mask2 = MakeConX(-2);
2845 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2846 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2847 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2848 Node* layout_val = make_load_raw(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2849 Node* size_in_bytes = ConvI2X(layout_val);
2850 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2851 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2852 fast_oop_ctrl, fast_oop_rawmem,
2853 AllocateInstancePrefetchLines);
2854 // Allocation succeed, initialize buffered inline instance header firstly,
2855 // and then initialize its fields with an inline class specific handler
2856 Node* mark_word_node;
2857 if (UseCompactObjectHeaders) {
2858 // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2859 mark_word_node = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2860 } else {
2861 // Otherwise, use the static prototype.
2862 mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2863 }
2864
2865 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2866 if (!UseCompactObjectHeaders) {
2867 // COH: Everything is encoded in the mark word, so nothing left to do.
2868 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2869 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2870 }
2871 Node* members = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InlineKlass::adr_members_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2872 Node* pack_handler = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, members, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2873 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2874 nullptr,
2875 "pack handler",
2876 TypeRawPtr::BOTTOM);
2877 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2878 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2879 handler_call->init_req(TypeFunc::I_O, top());
2880 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2881 handler_call->init_req(TypeFunc::ReturnAdr, top());
2882 handler_call->init_req(TypeFunc::Parms, pack_handler);
2883 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2884 } else {
2885 needgc_ctrl = allocation_ctl;
2886 }
2887
2888 // Allocation failed, fall back to a runtime call
2889 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2890 StubRoutines::store_inline_type_fields_to_buf(),
2891 "store_inline_type_fields",
2892 TypePtr::BOTTOM);
2893 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2894 slow_call->init_req(TypeFunc::Memory, mem);
2895 slow_call->init_req(TypeFunc::I_O, io);
2896 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2897 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2898 slow_call->init_req(TypeFunc::Parms, res);
2899
2900 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2901 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2902 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2903 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2904 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2905 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2906 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2907
2908 Node* ex_r = new RegionNode(3);
2909 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2910 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2911 ex_r->init_req(1, slow_excp);
2912 ex_mem_phi->init_req(1, slow_mem);
2913 ex_io_phi->init_req(1, slow_io);
2914 ex_r->init_req(2, ex_ctl);
2915 ex_mem_phi->init_req(2, ex_mem);
2916 ex_io_phi->init_req(2, ex_io);
2917 transform_later(ex_r);
2918 transform_later(ex_mem_phi);
2919 transform_later(ex_io_phi);
2920
2921 // We don't know how many values are returned. This assumes the
2922 // worst case, that all available registers are used.
2923 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2924 if (domain->field_at(i) == Type::HALF) {
2925 slow_call->init_req(i, top());
2926 if (alloc_in_place) {
2927 handler_call->init_req(i+1, top());
2928 }
2929 continue;
2930 }
2931 Node* proj = transform_later(new ProjNode(call, i));
2932 slow_call->init_req(i, proj);
2933 if (alloc_in_place) {
2934 handler_call->init_req(i+1, proj);
2935 }
2936 }
2937 // We can safepoint at that new call
2938 slow_call->copy_call_debug_info(&_igvn, call);
2939 transform_later(slow_call);
2940 if (alloc_in_place) {
2941 transform_later(handler_call);
2942 }
2943
2944 Node* fast_ctl = nullptr;
2945 Node* fast_res = nullptr;
2946 MergeMemNode* fast_mem = nullptr;
2947 if (alloc_in_place) {
2948 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2949 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2950 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2951 fast_mem = MergeMemNode::make(mem);
2952 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2953 transform_later(fast_mem);
2954 }
2955
2956 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2957 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2958 Node* io_phi = new PhiNode(r, Type::ABIO);
2959 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2960 r->init_req(1, no_allocation_ctl);
2961 mem_phi->init_req(1, mem);
2962 io_phi->init_req(1, io);
2963 res_phi->init_req(1, no_allocation_res);
2964 r->init_req(2, slow_norm);
2965 mem_phi->init_req(2, slow_mem);
2966 io_phi->init_req(2, slow_io);
2967 res_phi->init_req(2, slow_res);
2968 if (alloc_in_place) {
2969 r->init_req(3, fast_ctl);
2970 mem_phi->init_req(3, fast_mem);
2971 io_phi->init_req(3, io);
2972 res_phi->init_req(3, fast_res);
2973 }
2974 transform_later(r);
2975 transform_later(mem_phi);
2976 transform_later(io_phi);
2977 transform_later(res_phi);
2978
2979 // Do not let stores that initialize this buffer be reordered with a subsequent
2980 // store that would make this buffer accessible by other threads.
2981 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2982 transform_later(mb);
2983 mb->init_req(TypeFunc::Memory, mem_phi);
2984 mb->init_req(TypeFunc::Control, r);
2985 r = new ProjNode(mb, TypeFunc::Control);
2986 transform_later(r);
2987 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2988 transform_later(mem_phi);
2989
2990 assert(projs->nb_resproj == 1, "unexpected number of results");
2991 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2992 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2993 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2994 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2995 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2996 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2997 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2998 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2999 Node* cn = projs->fallthrough_catchproj->in(0);
3000 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
3001
3002 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
3003 _igvn.replace_node(mem, projs->fallthrough_memproj);
3004 _igvn.replace_node(io, projs->fallthrough_ioproj);
3005 _igvn.replace_node(res, projs->resproj[0]);
3006 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
3007 _igvn.replace_node(ex_mem, projs->catchall_memproj);
3008 _igvn.replace_node(ex_io, projs->catchall_ioproj);
3009 }
3010
3011 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
3012 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
3013 Node* bol = check->unique_out();
3014 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
3015 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
3016 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
3017
3018 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
3019 Node* iff = bol->last_out(i);
3020 assert(iff->is_If(), "where's the if?");
3021
3022 if (iff->in(0)->is_top()) {
3023 _igvn.replace_input_of(iff, 1, C->top());
3024 continue;
3025 }
3026
3027 IfTrueNode* iftrue = iff->as_If()->true_proj();
3028 IfFalseNode* iffalse = iff->as_If()->false_proj();
3029 Node* ctrl = iff->in(0);
3030
3031 Node* subklass = nullptr;
3032 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
3033 subklass = obj_or_subklass;
3034 } else {
3035 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
3036 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3037 }
3038
3039 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
3040
3041 _igvn.replace_input_of(iff, 0, C->top());
3042 _igvn.replace_node(iftrue, not_subtype_ctrl);
3043 _igvn.replace_node(iffalse, ctrl);
3044 }
3045 _igvn.replace_node(check, C->top());
3046 }
3047
3048 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
3049 //
3050 // long mark = array1.mark | array2.mark | ...;
3051 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
3052 // if (locked_bit == 0) {
3053 // // One array is locked, load prototype header from the klass
3054 // mark = array1.klass.proto | array2.klass.proto | ...
3055 // }
3056 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
3057 // ...
3058 // }
3059 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
3060 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
3061 if (array_inputs) {
3062 Node* mark = MakeConX(0);
3063 Node* locked_bit = MakeConX(markWord::unlocked_value);
3064 Node* mem = check->in(FlatArrayCheckNode::Memory);
3065 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3066 Node* ary = check->in(i);
3067 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
3068 assert(t != nullptr, "Mixing array and klass inputs");
3069 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3070 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
3071 Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3072 mark = _igvn.transform(new OrXNode(mark, mark_load));
3073 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
3074 }
3075 assert(!mark->is_Con(), "Should have been optimized out");
3076 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3077 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
3078
3079 // BoolNode might be shared, replace each if user
3080 Node* old_bol = check->unique_out();
3081 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
3082 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3083 IfNode* old_iff = old_bol->last_out(i)->as_If();
3084 Node* ctrl = old_iff->in(0);
3085 RegionNode* region = new RegionNode(3);
3086 Node* mark_phi = new PhiNode(region, TypeX_X);
3087
3088 // Check if array is unlocked
3089 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
3090
3091 // Unlocked: Use bits from mark word
3092 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
3093 mark_phi->init_req(1, mark);
3094
3095 // Locked: Load prototype header from klass
3096 ctrl = _igvn.transform(new IfFalseNode(iff));
3097 Node* proto = MakeConX(0);
3098 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3099 Node* ary = check->in(i);
3100 // Make loads control dependent to make sure they are only executed if array is locked
3101 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3102 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3103 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3104 Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3105 proto = _igvn.transform(new OrXNode(proto, proto_load));
3106 }
3107 region->init_req(2, ctrl);
3108 mark_phi->init_req(2, proto);
3109
3110 // Check if flat array bits are set
3111 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
3112 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
3113 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
3114 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
3115
3116 ctrl = _igvn.transform(region);
3117 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
3118 _igvn.replace_node(old_iff, iff);
3119 }
3120 _igvn.replace_node(check, C->top());
3121 } else {
3122 // Fall back to layout helper check
3123 Node* lhs = intcon(0);
3124 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3125 Node* array_or_klass = check->in(i);
3126 Node* klass = nullptr;
3127 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
3128 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3129 if (t->isa_oopptr() != nullptr) {
3130 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
3131 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3132 } else {
3133 assert(t->isa_klassptr(), "Unexpected input type");
3134 klass = array_or_klass;
3135 }
3136 Node* lh_addr = basic_plus_adr(top(), klass, in_bytes(Klass::layout_helper_offset()));
3137 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3138 lhs = _igvn.transform(new OrINode(lhs, lh_val));
3139 }
3140 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
3141 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
3142 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
3143 Node* m2b = transform_later(new Conv2BNode(masked));
3144 // The matcher expects the input to If/CMove nodes to be produced by a Bool(CmpI..)
3145 // pattern, but the input to other potential users (e.g. Phi) to be some
3146 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
3147 Node* old_bol = check->unique_out();
3148 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3149 Node* user = old_bol->last_out(i);
3150 for (uint j = 0; j < user->req(); j++) {
3151 Node* n = user->in(j);
3152 if (n == old_bol) {
3153 _igvn.replace_input_of(user, j, (user->is_If() || user->is_CMove()) ? bol : m2b);
3154 }
3155 }
3156 }
3157 _igvn.replace_node(check, C->top());
3158 }
3159 }
3160
3161 // Perform refining of strip mined loop nodes in the macro nodes list.
3162 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
3163 for (int i = C->macro_count(); i > 0; i--) {
3164 Node* n = C->macro_node(i - 1);
3165 if (n->is_OuterStripMinedLoop()) {
3166 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3167 }
3168 }
3169 }
3170
3171 //---------------------------eliminate_macro_nodes----------------------
3172 // Eliminate scalar replaced allocations and associated locks.
3173 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
3174 if (C->macro_count() == 0) {
3175 return;
3176 }
3177
3178 if (StressMacroElimination) {
3179 C->shuffle_macro_nodes();
3180 }
3181 NOT_PRODUCT(int membar_before = count_MemBar(C);)
3182
3183 int iteration = 0;
3184 while (C->macro_count() > 0) {
3185 if (iteration++ > 100) {
3186 assert(false, "Too slow convergence of macro elimination");
3187 break;
3188 }
3189
3190 // Postpone lock elimination to after EA when most allocations are eliminated
3191 // because they might block lock elimination if their escape state isn't
3192 // determined yet and we only got one chance at eliminating the lock.
3193 if (eliminate_locks) {
3194 // Before elimination may re-mark (change to Nested or NonEscObj)
3195 // all associated (same box and obj) lock and unlock nodes.
3196 int cnt = C->macro_count();
3197 for (int i=0; i < cnt; i++) {
3198 Node *n = C->macro_node(i);
3199 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3200 mark_eliminated_locking_nodes(n->as_AbstractLock());
3201 }
3202 }
3203 // Re-marking may break consistency of Coarsened locks.
3204 if (!C->coarsened_locks_consistent()) {
3205 return; // recompile without Coarsened locks if broken
3206 } else {
3207 // After coarsened locks are eliminated locking regions
3208 // become unbalanced. We should not execute any more
3209 // locks elimination optimizations on them.
3210 C->mark_unbalanced_boxes();
3211 }
3212 }
3213
3214 bool progress = false;
3215 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3216 Node* n = C->macro_node(i - 1);
3217 bool success = false;
3218 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3219 switch (n->class_id()) {
3220 case Node::Class_Allocate:
3221 case Node::Class_AllocateArray:
3222 success = eliminate_allocate_node(n->as_Allocate());
3223 #ifndef PRODUCT
3224 if (success && PrintOptoStatistics) {
3225 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3226 }
3227 #endif
3228 break;
3229 case Node::Class_CallStaticJava: {
3230 CallStaticJavaNode* call = n->as_CallStaticJava();
3231 if (!call->method()->is_method_handle_intrinsic()) {
3232 success = eliminate_boxing_node(n->as_CallStaticJava());
3233 }
3234 break;
3235 }
3236 case Node::Class_Lock:
3237 case Node::Class_Unlock:
3238 if (eliminate_locks) {
3239 success = eliminate_locking_node(n->as_AbstractLock());
3240 #ifndef PRODUCT
3241 if (success && PrintOptoStatistics) {
3242 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3243 }
3244 #endif
3245 }
3246 break;
3247 case Node::Class_ArrayCopy:
3248 break;
3249 case Node::Class_OuterStripMinedLoop:
3250 break;
3251 case Node::Class_SubTypeCheck:
3252 break;
3253 case Node::Class_Opaque1:
3254 break;
3255 case Node::Class_FlatArrayCheck:
3256 break;
3257 default:
3258 assert(n->Opcode() == Op_LoopLimit ||
3259 n->Opcode() == Op_ModD ||
3260 n->Opcode() == Op_ModF ||
3261 n->Opcode() == Op_PowD ||
3262 n->is_OpaqueConstantBool() ||
3263 n->is_OpaqueInitializedAssertionPredicate() ||
3264 n->Opcode() == Op_MaxL ||
3265 n->Opcode() == Op_MinL ||
3266 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3267 "unknown node type in macro list");
3268 }
3269 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3270 progress = progress || success;
3271 if (success) {
3272 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3273 }
3274 }
3275
3276 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3277 // transformation is pending). If an allocation is used only in safepoints, elimination of
3278 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3279 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3280 // iteration.
3281 _igvn.set_delay_transform(false);
3282 _igvn.optimize();
3283 if (C->failing()) {
3284 return;
3285 }
3286 _igvn.set_delay_transform(true);
3287
3288 if (!progress) {
3289 break;
3290 }
3291 }
3292 #ifndef PRODUCT
3293 if (PrintOptoStatistics) {
3294 int membar_after = count_MemBar(C);
3295 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3296 }
3297 #endif
3298 }
3299
3300 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3301 if (C->macro_count() == 0) {
3302 return;
3303 }
3304 refine_strip_mined_loop_macro_nodes();
3305 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3306 bool progress = true;
3307 while (progress) {
3308 progress = false;
3309 for (int i = C->macro_count(); i > 0; i--) {
3310 Node* n = C->macro_node(i-1);
3311 bool success = false;
3312 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3313 if (n->Opcode() == Op_LoopLimit) {
3314 // Remove it from macro list and put on IGVN worklist to optimize.
3315 C->remove_macro_node(n);
3316 _igvn._worklist.push(n);
3317 success = true;
3318 } else if (n->Opcode() == Op_CallStaticJava) {
3319 CallStaticJavaNode* call = n->as_CallStaticJava();
3320 if (!call->method()->is_method_handle_intrinsic()) {
3321 // Remove it from macro list and put on IGVN worklist to optimize.
3322 C->remove_macro_node(n);
3323 _igvn._worklist.push(n);
3324 success = true;
3325 }
3326 } else if (n->is_Opaque1()) {
3327 _igvn.replace_node(n, n->in(1));
3328 success = true;
3329 } else if (n->is_OpaqueConstantBool()) {
3330 // Tests with OpaqueConstantBool nodes are implicitly known. Replace the node with true/false. In debug builds,
3331 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3332 // we will execute a Halt node.
3333 #ifdef ASSERT
3334 _igvn.replace_node(n, n->in(1));
3335 #else
3336 _igvn.replace_node(n, _igvn.intcon(n->as_OpaqueConstantBool()->constant()));
3337 #endif
3338 success = true;
3339 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3340 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3341 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3342 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3343 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3344 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3345 #ifdef ASSERT
3346 _igvn.replace_node(n, n->in(1));
3347 #else
3348 _igvn.replace_node(n, _igvn.intcon(1));
3349 #endif // ASSERT
3350 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3351 C->remove_macro_node(n);
3352 success = true;
3353 } else if (n->Opcode() == Op_MaxL) {
3354 // Since MaxL and MinL are not implemented in the backend, we expand them to
3355 // a CMoveL construct now. At least until here, the type could be computed
3356 // precisely. CMoveL is not so smart, but we can give it at least the best
3357 // type we know abouot n now.
3358 Node* repl = MinMaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3359 _igvn.replace_node(n, repl);
3360 success = true;
3361 } else if (n->Opcode() == Op_MinL) {
3362 Node* repl = MinMaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3363 _igvn.replace_node(n, repl);
3364 success = true;
3365 }
3366 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3367 progress = progress || success;
3368 if (success) {
3369 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3370 }
3371 }
3372 }
3373 }
3374
3375 //------------------------------expand_macro_nodes----------------------
3376 // Returns true if a failure occurred.
3377 bool PhaseMacroExpand::expand_macro_nodes() {
3378 if (StressMacroExpansion) {
3379 C->shuffle_macro_nodes();
3380 }
3381
3382 // Clean up the graph so we're less likely to hit the maximum node
3383 // limit
3384 _igvn.set_delay_transform(false);
3385 _igvn.optimize();
3386 if (C->failing()) return true;
3387 _igvn.set_delay_transform(true);
3388
3389
3390 // Because we run IGVN after each expansion, some macro nodes may go
3391 // dead and be removed from the list as we iterate over it. Move
3392 // Allocate nodes (processed in a second pass) at the beginning of
3393 // the list and then iterate from the last element of the list until
3394 // an Allocate node is seen. This is robust to random deletion in
3395 // the list due to nodes going dead.
3396 C->sort_macro_nodes();
3397
3398 // expand arraycopy "macro" nodes first
3399 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3400 // before the allocate nodes are expanded.
3401 while (C->macro_count() > 0) {
3402 int macro_count = C->macro_count();
3403 Node * n = C->macro_node(macro_count-1);
3404 assert(n->is_macro(), "only macro nodes expected here");
3405 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3406 // node is unreachable, so don't try to expand it
3407 C->remove_macro_node(n);
3408 continue;
3409 }
3410 if (n->is_Allocate()) {
3411 break;
3412 }
3413 // Make sure expansion will not cause node limit to be exceeded.
3414 // Worst case is a macro node gets expanded into about 200 nodes.
3415 // Allow 50% more for optimization.
3416 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3417 return true;
3418 }
3419
3420 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3421 switch (n->class_id()) {
3422 case Node::Class_Lock:
3423 expand_lock_node(n->as_Lock());
3424 break;
3425 case Node::Class_Unlock:
3426 expand_unlock_node(n->as_Unlock());
3427 break;
3428 case Node::Class_ArrayCopy:
3429 expand_arraycopy_node(n->as_ArrayCopy());
3430 break;
3431 case Node::Class_SubTypeCheck:
3432 expand_subtypecheck_node(n->as_SubTypeCheck());
3433 break;
3434 case Node::Class_CallStaticJava:
3435 expand_mh_intrinsic_return(n->as_CallStaticJava());
3436 C->remove_macro_node(n);
3437 break;
3438 case Node::Class_FlatArrayCheck:
3439 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3440 break;
3441 default:
3442 switch (n->Opcode()) {
3443 case Op_ModD:
3444 case Op_ModF:
3445 case Op_PowD: {
3446 CallLeafPureNode* call_macro = n->as_CallLeafPure();
3447 CallLeafPureNode* call = call_macro->inline_call_leaf_pure_node();
3448 _igvn.replace_node(call_macro, call);
3449 transform_later(call);
3450 break;
3451 }
3452 default:
3453 assert(false, "unknown node type in macro list");
3454 }
3455 }
3456 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3457 if (C->failing()) return true;
3458 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3459
3460 // Clean up the graph so we're less likely to hit the maximum node
3461 // limit
3462 _igvn.set_delay_transform(false);
3463 _igvn.optimize();
3464 if (C->failing()) return true;
3465 _igvn.set_delay_transform(true);
3466 }
3467
3468 // All nodes except Allocate nodes are expanded now. There could be
3469 // new optimization opportunities (such as folding newly created
3470 // load from a just allocated object). Run IGVN.
3471
3472 // expand "macro" nodes
3473 // nodes are removed from the macro list as they are processed
3474 while (C->macro_count() > 0) {
3475 int macro_count = C->macro_count();
3476 Node * n = C->macro_node(macro_count-1);
3477 assert(n->is_macro(), "only macro nodes expected here");
3478 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3479 // node is unreachable, so don't try to expand it
3480 C->remove_macro_node(n);
3481 continue;
3482 }
3483 // Make sure expansion will not cause node limit to be exceeded.
3484 // Worst case is a macro node gets expanded into about 200 nodes.
3485 // Allow 50% more for optimization.
3486 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3487 return true;
3488 }
3489 switch (n->class_id()) {
3490 case Node::Class_Allocate:
3491 expand_allocate(n->as_Allocate());
3492 break;
3493 case Node::Class_AllocateArray:
3494 expand_allocate_array(n->as_AllocateArray());
3495 break;
3496 default:
3497 assert(false, "unknown node type in macro list");
3498 }
3499 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3500 if (C->failing()) return true;
3501 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3502
3503 // Clean up the graph so we're less likely to hit the maximum node
3504 // limit
3505 _igvn.set_delay_transform(false);
3506 _igvn.optimize();
3507 if (C->failing()) return true;
3508 _igvn.set_delay_transform(true);
3509 }
3510
3511 _igvn.set_delay_transform(false);
3512 return false;
3513 }
3514
3515 #ifndef PRODUCT
3516 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3517 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3518 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3519 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3520
3521 void PhaseMacroExpand::print_statistics() {
3522 tty->print("Objects scalar replaced = %d, ", AtomicAccess::load(&_objs_scalar_replaced_counter));
3523 tty->print("Monitor objects removed = %d, ", AtomicAccess::load(&_monitor_objects_removed_counter));
3524 tty->print("GC barriers removed = %d, ", AtomicAccess::load(&_GC_barriers_removed_counter));
3525 tty->print_cr("Memory barriers removed = %d", AtomicAccess::load(&_memory_barriers_removed_counter));
3526 }
3527
3528 int PhaseMacroExpand::count_MemBar(Compile *C) {
3529 if (!PrintOptoStatistics) {
3530 return 0;
3531 }
3532 Unique_Node_List ideal_nodes;
3533 int total = 0;
3534 ideal_nodes.map(C->live_nodes(), nullptr);
3535 ideal_nodes.push(C->root());
3536 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3537 Node* n = ideal_nodes.at(next);
3538 if (n->is_MemBar()) {
3539 total++;
3540 }
3541 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3542 Node* m = n->fast_out(i);
3543 ideal_nodes.push(m);
3544 }
3545 }
3546 return total;
3547 }
3548 #endif