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, bool is_forced_failure) {
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
700 tty->print("=== At SafePoint node %d ", sfpt->_idx);
701 if (is_forced_failure) {
702 tty->print_raw("forcibly abort elimination");
703 } else {
704 tty->print("can't find value of field [%s] of array element [%d]", flattened_field->name()->as_utf8(), elem_idx);
705 }
706 tty->print(", which prevents elimination of: ");
707 alloc->dump();
708 }
709 #endif // PRODUCT
710 };
711
712 // Create a new InlineTypeNode and retrieve the field values from memory
713 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, null_free);
714 transform_later(vt);
715 if (null_free) {
716 vt->set_null_marker(_igvn);
717 } else {
718 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
719 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
720 Node* nm_value = value_from_mem(sfpt, sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
721 bool force_scalarization_failure = StressEliminateAllocations &&
722 (C->random() % StressEliminateAllocationsMean == 0);
723 if (nm_value != nullptr && !force_scalarization_failure) {
724 vt->set_null_marker(_igvn, nm_value);
725 } else {
726 report_failure(nm_offset_in_element, nm_value != nullptr);
727 return nullptr;
728 }
729 }
730
731 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
732 ciField* field = vt->field(i);
733 ciType* field_type = field->type();
734 int field_offset_in_element = offset_in_element + field->offset_in_bytes() - vk->payload_offset();
735 Node* field_value = nullptr;
736 assert(!field->is_flat() || field->type()->is_inlinetype(), "must be an inline type");
737 if (field->is_flat()) {
738 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);
739 } else {
740 const Type* ft = Type::get_const_type(field_type);
741 BasicType bt = type2field[field_type->basic_type()];
742 if (UseCompressedOops && !is_java_primitive(bt)) {
743 ft = ft->make_narrowoop();
744 bt = T_NARROWOOP;
745 }
746 // Each inline type field has its own memory slice
747 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
748 field_value = value_from_mem(sfpt, sfpt->control(), bt, ft, field_adr_type, alloc);
749 bool force_scalarization_failure = StressEliminateAllocations &&
750 (C->random() % StressEliminateAllocationsMean == 0);
751 if (field_value == nullptr || force_scalarization_failure) {
752 report_failure(field_offset_in_element, field_value != nullptr);
753 return nullptr;
754 } else if (ft->isa_narrowoop()) {
755 assert(UseCompressedOops, "unexpected narrow oop");
756 if (field_value->is_EncodeP()) {
757 field_value = field_value->in(1);
758 } else if (!field_value->is_InlineType()) {
759 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
760 }
761 }
762 }
763 if (field_value != nullptr) {
764 vt->set_field_value(i, field_value);
765 } else {
766 return nullptr;
767 }
768 }
769 return vt;
770 }
771
772 // Check the possibility of scalar replacement.
773 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode* alloc, Unique_Node_List* safepoints) {
774 // Scan the uses of the allocation to check for anything that would
775 // prevent us from eliminating it.
776 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
777 DEBUG_ONLY( Node* disq_node = nullptr; )
778 bool can_eliminate = true;
779 bool reduce_merge_precheck = (safepoints == nullptr);
780
781 Unique_Node_List worklist;
782 Node* res = alloc->result_cast();
783 const TypeOopPtr* res_type = nullptr;
784 if (res == nullptr) {
785 // All users were eliminated.
786 } else if (!res->is_CheckCastPP()) {
787 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
788 can_eliminate = false;
789 } else {
790 worklist.push(res);
791 res_type = igvn->type(res)->isa_oopptr();
792 if (res_type == nullptr) {
793 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
794 can_eliminate = false;
795 } else if (!res_type->klass_is_exact()) {
796 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
797 can_eliminate = false;
798 } else if (res_type->isa_aryptr()) {
799 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
800 if (length < 0) {
801 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
802 can_eliminate = false;
803 }
804 }
805 }
806
807 while (can_eliminate && worklist.size() > 0) {
808 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
809 res = worklist.pop();
810 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
811 Node* use = res->fast_out(j);
812
813 if (use->is_AddP()) {
814 const TypePtr* addp_type = igvn->type(use)->is_ptr();
815 int offset = addp_type->offset();
816
817 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
818 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
819 can_eliminate = false;
820 break;
821 }
822 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
823 k < kmax && can_eliminate; k++) {
824 Node* n = use->fast_out(k);
825 if ((n->is_Mem() && n->as_Mem()->is_mismatched_access()) || n->is_LoadFlat() || n->is_StoreFlat()) {
826 DEBUG_ONLY(disq_node = n);
827 NOT_PRODUCT(fail_eliminate = "Mismatched access");
828 can_eliminate = false;
829 }
830 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
831 DEBUG_ONLY(disq_node = n;)
832 if (n->is_Load() || n->is_LoadStore()) {
833 NOT_PRODUCT(fail_eliminate = "Field load";)
834 } else {
835 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
836 }
837 can_eliminate = false;
838 }
839 }
840 } else if (use->is_ArrayCopy() &&
841 (use->as_ArrayCopy()->is_clonebasic() ||
842 use->as_ArrayCopy()->is_arraycopy_validated() ||
843 use->as_ArrayCopy()->is_copyof_validated() ||
844 use->as_ArrayCopy()->is_copyofrange_validated()) &&
845 use->in(ArrayCopyNode::Dest) == res) {
846 // ok to eliminate
847 } else if (use->is_ReachabilityFence() && OptimizeReachabilityFences) {
848 // ok to eliminate
849 } else if (use->is_SafePoint()) {
850 SafePointNode* sfpt = use->as_SafePoint();
851 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
852 // Object is passed as argument.
853 DEBUG_ONLY(disq_node = use;)
854 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
855 can_eliminate = false;
856 }
857 Node* sfptMem = sfpt->memory();
858 if (sfptMem == nullptr || sfptMem->is_top()) {
859 DEBUG_ONLY(disq_node = use;)
860 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
861 can_eliminate = false;
862 } else if (!reduce_merge_precheck) {
863 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
864 safepoints->push(sfpt);
865 }
866 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
867 // Look at uses
868 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
869 Node* u = use->fast_out(k);
870 if (u->is_InlineType()) {
871 // Use in flat field can be eliminated
872 InlineTypeNode* vt = u->as_InlineType();
873 for (uint i = 0; i < vt->field_count(); ++i) {
874 if (vt->field_value(i) == use && !vt->field(i)->is_flat()) {
875 can_eliminate = false; // Use in non-flat field
876 break;
877 }
878 }
879 } else {
880 // Add other uses to the worklist to process individually
881 worklist.push(use);
882 }
883 }
884 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
885 // Store to mark word of inline type larval buffer
886 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
887 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
888 // Inline type buffer allocations are followed by a membar
889 } else if (reduce_merge_precheck &&
890 (use->is_Phi() || use->is_EncodeP() ||
891 use->Opcode() == Op_MemBarRelease ||
892 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
893 // Nothing to do
894 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
895 if (use->is_Phi()) {
896 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
897 NOT_PRODUCT(fail_eliminate = "Object is return value";)
898 } else {
899 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
900 }
901 DEBUG_ONLY(disq_node = use;)
902 } else {
903 if (use->Opcode() == Op_Return) {
904 NOT_PRODUCT(fail_eliminate = "Object is return value";)
905 } else {
906 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
907 }
908 DEBUG_ONLY(disq_node = use;)
909 }
910 can_eliminate = false;
911 } else {
912 assert(use->Opcode() == Op_CastP2X, "should be");
913 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
914 }
915 }
916 }
917
918 #ifndef PRODUCT
919 if (PrintEliminateAllocations && safepoints != nullptr) {
920 if (can_eliminate) {
921 tty->print("Scalar ");
922 if (res == nullptr)
923 alloc->dump();
924 else
925 res->dump();
926 } else {
927 tty->print("NotScalar (%s)", fail_eliminate);
928 if (res == nullptr)
929 alloc->dump();
930 else
931 res->dump();
932 #ifdef ASSERT
933 if (disq_node != nullptr) {
934 tty->print(" >>>> ");
935 disq_node->dump();
936 }
937 #endif /*ASSERT*/
938 }
939 }
940
941 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
942 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
943 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
944 }
945 #endif
946 return can_eliminate;
947 }
948
949 void PhaseMacroExpand::undo_previous_scalarizations(Unique_Node_List& safepoints_done, AllocateNode* alloc) {
950 Node* res = alloc->result_cast();
951 int nfields = 0;
952 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
953
954 if (res != nullptr) {
955 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
956
957 if (res_type->isa_instptr()) {
958 // find the fields of the class which will be needed for safepoint debug information
959 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
960 nfields = iklass->nof_nonstatic_fields();
961 } else {
962 // find the array's elements which will be needed for safepoint debug information
963 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
964 assert(nfields >= 0, "must be an array klass.");
965 }
966 }
967
968 // rollback processed safepoints
969 while (safepoints_done.size() > 0) {
970 SafePointNode* sfpt_done = safepoints_done.pop()->as_SafePoint();
971
972 SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(igvn());
973
974 sfpt_done->remove_non_debug_edges(non_debug_edges_worklist);
975
976 // remove any extra entries we added to the safepoint
977 assert(sfpt_done->jvms()->endoff() == sfpt_done->req(), "no extra edges past debug info allowed");
978 uint last = sfpt_done->req() - 1;
979 for (int k = 0; k < nfields; k++) {
980 sfpt_done->del_req(last--);
981 }
982 JVMState *jvms = sfpt_done->jvms();
983 jvms->set_endoff(sfpt_done->req());
984 // Now make a pass over the debug information replacing any references
985 // to SafePointScalarObjectNode with the allocated object.
986 int start = jvms->debug_start();
987 int end = jvms->debug_end();
988 for (int i = start; i < end; i++) {
989 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
990 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
991 if (scobj->first_index(jvms) == sfpt_done->req() &&
992 scobj->n_fields() == (uint)nfields) {
993 assert(scobj->alloc() == alloc, "sanity");
994 sfpt_done->set_req(i, res);
995 }
996 }
997 }
998
999 sfpt_done->restore_non_debug_edges(non_debug_edges_worklist);
1000
1001 _igvn._worklist.push(sfpt_done);
1002 }
1003 }
1004
1005 #ifdef ASSERT
1006 // Verify if a value can be written into a field.
1007 void verify_type_compatability(const Type* value_type, const Type* field_type) {
1008 BasicType value_bt = value_type->basic_type();
1009 BasicType field_bt = field_type->basic_type();
1010
1011 // Primitive types must match.
1012 if (is_java_primitive(value_bt) && value_bt == field_bt) { return; }
1013
1014 // I have been struggling to make a similar assert for non-primitive
1015 // types. I we can add one in the future. For now, I just let them
1016 // pass without checks.
1017 // In particular, I was struggling with a value that came from a call,
1018 // and had only a non-null check CastPP. There was also a checkcast
1019 // in the graph to verify the interface, but the corresponding
1020 // CheckCastPP result was not updated in the stack slot, and so
1021 // we ended up using the CastPP. That means that the field knows
1022 // that it should get an oop from an interface, but the value lost
1023 // that information, and so it is not a subtype.
1024 // There may be other issues, feel free to investigate further!
1025 if (!is_java_primitive(value_bt)) { return; }
1026
1027 tty->print_cr("value not compatible for field: %s vs %s",
1028 type2name(value_bt),
1029 type2name(field_bt));
1030 tty->print("value_type: ");
1031 value_type->dump();
1032 tty->cr();
1033 tty->print("field_type: ");
1034 field_type->dump();
1035 tty->cr();
1036 assert(false, "value_type does not fit field_type");
1037 }
1038 #endif
1039
1040 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1041 if (UseCompressedOops && field_type->isa_narrowoop()) {
1042 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
1043 // to be able scalar replace the allocation.
1044 if (field_val->is_EncodeP()) {
1045 field_val = field_val->in(1);
1046 } else if (!field_val->is_InlineType()) {
1047 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
1048 }
1049 }
1050
1051 // Keep track of inline types to scalarize them later
1052 if (field_val->is_InlineType()) {
1053 value_worklist->push(field_val);
1054 } else if (field_val->is_Phi()) {
1055 PhiNode* phi = field_val->as_Phi();
1056 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
1057 // scalarize all its fields in safepoints.
1058 field_val = phi->try_push_inline_types_down(&_igvn, true);
1059 if (field_val->is_InlineType()) {
1060 value_worklist->push(field_val);
1061 }
1062 }
1063 DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
1064 sfpt->add_req(field_val);
1065 }
1066
1067 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1068 const Type* elem_type = array_type->elem();
1069 BasicType basic_elem_type = elem_type->array_element_basic_type();
1070
1071 intptr_t elem_size;
1072 uint header_size;
1073 if (array_type->is_flat()) {
1074 elem_size = array_type->flat_elem_size();
1075 header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1076 } else {
1077 elem_size = type2aelembytes(basic_elem_type);
1078 header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1079 }
1080
1081 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
1082 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
1083 intptr_t elem_offset = header_size + elem_idx * elem_size;
1084 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
1085 Node* elem_val;
1086 if (array_type->is_flat()) {
1087 ciInlineKlass* elem_klass = elem_type->inline_klass();
1088 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
1089 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
1090 } else {
1091 elem_val = value_from_mem(sfpt, sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
1092 }
1093 bool force_scalarization_failure = StressEliminateAllocations &&
1094 (C->random() % StressEliminateAllocationsMean == 0);
1095 if (elem_val == nullptr || force_scalarization_failure) {
1096 #ifndef PRODUCT
1097 if (PrintEliminateAllocations) {
1098 tty->print("=== At SafePoint node %d ", sfpt->_idx);
1099 if (elem_val == nullptr) {
1100 tty->print("can't find value of array element [%d]", elem_idx);
1101 } else {
1102 assert(force_scalarization_failure, "sanity");
1103 tty->print_raw("forcibly abort elimination");
1104 }
1105 tty->print(", which prevents elimination of: ");
1106 alloc->dump();
1107 }
1108 #endif // PRODUCT
1109 return false;
1110 }
1111
1112 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
1113 }
1114
1115 return true;
1116 }
1117
1118 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
1119 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
1120 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
1121 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1122 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
1123 auto report_failure = [&](int offset, bool is_forced_failure) {
1124 #ifndef PRODUCT
1125 if (PrintEliminateAllocations) {
1126 ciInstanceKlass* base_klass = base_type->instance_klass();
1127 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
1128 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
1129 tty->print("=== At SafePoint node %d ", sfpt->_idx);
1130 if (is_forced_failure) {
1131 tty->print_raw("forcibly abort elimination");
1132 } else {
1133 tty->print_raw("can't find value of field: ");
1134 flattened_field->print();
1135 int field_idx = C->alias_type(flattened_field)->index();
1136 tty->print(" (alias_idx=%d)", field_idx);
1137 }
1138 tty->print(", which prevents elimination of: ");
1139 base->dump();
1140 }
1141 #endif // PRODUCT
1142 };
1143
1144 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
1145 ciField* field = iklass->declared_nonstatic_field_at(i);
1146 if (field->is_flat()) {
1147 ciInlineKlass* fvk = field->type()->as_inline_klass();
1148 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
1149 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
1150 if (!success) {
1151 return false;
1152 }
1153
1154 // The null marker of a field is added right after we scalarize that field
1155 if (!field->is_null_free()) {
1156 int nm_offset = offset_minus_header + field->null_marker_offset();
1157 Node* null_marker = value_from_mem(sfpt, sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
1158 bool force_scalarization_failure = StressEliminateAllocations &&
1159 (C->random() % StressEliminateAllocationsMean == 0);
1160 if (null_marker == nullptr || force_scalarization_failure) {
1161 report_failure(nm_offset, null_marker != nullptr);
1162 return false;
1163 }
1164 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
1165 }
1166
1167 continue;
1168 }
1169
1170 int offset = offset_minus_header + field->offset_in_bytes();
1171 ciType* elem_type = field->type();
1172 BasicType basic_elem_type = field->layout_type();
1173
1174 const Type* field_type;
1175 if (is_reference_type(basic_elem_type)) {
1176 if (!elem_type->is_loaded()) {
1177 field_type = TypeInstPtr::BOTTOM;
1178 } else {
1179 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
1180 }
1181 if (UseCompressedOops) {
1182 field_type = field_type->make_narrowoop();
1183 basic_elem_type = T_NARROWOOP;
1184 }
1185 } else {
1186 field_type = Type::get_const_basic_type(basic_elem_type);
1187 }
1188
1189 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1190 Node* field_val = value_from_mem(sfpt, sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1191 bool force_scalarization_failure = StressEliminateAllocations &&
1192 (C->random() % StressEliminateAllocationsMean == 0);
1193 if (field_val == nullptr || force_scalarization_failure) {
1194 report_failure(offset, field_val != nullptr);
1195 return false;
1196 }
1197 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1198 }
1199
1200 return true;
1201 }
1202
1203 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1204 Unique_Node_List* value_worklist) {
1205 assert(sfpt->jvms()->endoff() == sfpt->req(), "no extra edges past debug info allowed");
1206
1207 // Fields of scalar objs are referenced only at the end
1208 // of regular debuginfo at the last (youngest) JVMS.
1209 // Record relative start index.
1210 ciInstanceKlass* iklass = nullptr;
1211 const TypeOopPtr* res_type = nullptr;
1212 int nfields = 0;
1213 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1214 Node* res = alloc->result_cast();
1215
1216 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1217 assert(sfpt->jvms() != nullptr, "missed JVMS");
1218 uint before_sfpt_req = sfpt->req();
1219
1220 if (res != nullptr) { // Could be null when there are no users
1221 res_type = _igvn.type(res)->isa_oopptr();
1222
1223 if (res_type->isa_instptr()) {
1224 // find the fields of the class which will be needed for safepoint debug information
1225 iklass = res_type->is_instptr()->instance_klass();
1226 nfields = iklass->nof_nonstatic_fields();
1227 } else {
1228 // find the array's elements which will be needed for safepoint debug information
1229 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1230 assert(nfields >= 0, "must be an array klass.");
1231 }
1232
1233 if (res->bottom_type()->is_inlinetypeptr()) {
1234 // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1235 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1236 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1237 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1238 // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1239 // later in PhaseOutput::filLocArray() for inline types.
1240 sfpt->add_req(_igvn.intcon(1));
1241 }
1242 }
1243
1244 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1245 sobj->init_req(0, C->root());
1246 transform_later(sobj);
1247
1248 if (res == nullptr) {
1249 sfpt->jvms()->set_endoff(sfpt->req());
1250 return sobj;
1251 }
1252
1253 bool success;
1254 if (iklass == nullptr) {
1255 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1256 } else {
1257 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1258 }
1259
1260 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1261 if (!success) {
1262 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1263 sfpt->del_req(i);
1264 }
1265 _igvn._worklist.push(sfpt);
1266 return nullptr;
1267 }
1268
1269 sfpt->jvms()->set_endoff(sfpt->req());
1270 return sobj;
1271 }
1272
1273 // Do scalar replacement.
1274 bool PhaseMacroExpand::scalar_replacement(AllocateNode* alloc, Unique_Node_List& safepoints) {
1275 Unique_Node_List safepoints_done;
1276 Node* res = alloc->result_cast();
1277 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1278 const TypeOopPtr* res_type = nullptr;
1279 if (res != nullptr) { // Could be null when there are no users
1280 res_type = _igvn.type(res)->isa_oopptr();
1281 }
1282
1283 // Process the safepoint uses
1284 Unique_Node_List value_worklist;
1285 while (safepoints.size() > 0) {
1286 SafePointNode* sfpt = safepoints.pop()->as_SafePoint();
1287
1288 SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(igvn());
1289
1290 // All sfpt inputs are implicitly included into debug info during the scalarization process below.
1291 // Keep non-debug inputs separately, so they stay non-debug.
1292 sfpt->remove_non_debug_edges(non_debug_edges_worklist);
1293
1294 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1295
1296 if (sobj == nullptr) {
1297 sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1298 undo_previous_scalarizations(safepoints_done, alloc);
1299 return false;
1300 }
1301
1302 // Now make a pass over the debug information replacing any references
1303 // to the allocated object with "sobj"
1304 JVMState *jvms = sfpt->jvms();
1305 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1306 non_debug_edges_worklist.remove_edge_if_present(res); // drop scalarized input from non-debug info
1307 sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1308 _igvn._worklist.push(sfpt);
1309
1310 // keep it for rollback
1311 safepoints_done.push(sfpt);
1312 }
1313 // Scalarize inline types that were added to the safepoint.
1314 // Don't allow linking a constant oop (if available) for flat array elements
1315 // because Deoptimization::reassign_flat_array_elements needs field values.
1316 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1317 for (uint i = 0; i < value_worklist.size(); ++i) {
1318 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1319 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1320 }
1321 return true;
1322 }
1323
1324 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1325 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1326 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1327 if (ctl_proj != nullptr) {
1328 igvn.replace_node(ctl_proj, n->in(0));
1329 }
1330 if (mem_proj != nullptr) {
1331 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1332 }
1333 }
1334
1335 // Process users of eliminated allocation.
1336 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1337 Unique_Node_List worklist;
1338 Node* res = alloc->result_cast();
1339 if (res != nullptr) {
1340 worklist.push(res);
1341 }
1342 while (worklist.size() > 0) {
1343 res = worklist.pop();
1344 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1345 Node *use = res->last_out(j);
1346 uint oc1 = res->outcnt();
1347
1348 if (use->is_AddP()) {
1349 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1350 Node *n = use->last_out(k);
1351 uint oc2 = use->outcnt();
1352 if (n->is_Store()) {
1353 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1354 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1355 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1356 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1357 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1358 mb->remove(&_igvn);
1359 }
1360 }
1361 _igvn.replace_node(n, n->in(MemNode::Memory));
1362 } else {
1363 eliminate_gc_barrier(n);
1364 }
1365 k -= (oc2 - use->outcnt());
1366 }
1367 _igvn.remove_dead_node(use, PhaseIterGVN::NodeOrigin::Graph);
1368 } else if (use->is_ArrayCopy()) {
1369 // Disconnect ArrayCopy node
1370 ArrayCopyNode* ac = use->as_ArrayCopy();
1371 if (ac->is_clonebasic()) {
1372 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1373 disconnect_projections(ac, _igvn);
1374 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1375 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1376 disconnect_projections(membar_before->as_MemBar(), _igvn);
1377 if (membar_after->is_MemBar()) {
1378 disconnect_projections(membar_after->as_MemBar(), _igvn);
1379 }
1380 } else {
1381 assert(ac->is_arraycopy_validated() ||
1382 ac->is_copyof_validated() ||
1383 ac->is_copyofrange_validated(), "unsupported");
1384 CallProjections* callprojs = ac->extract_projections(true);
1385
1386 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1387 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1388 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1389
1390 // Set control to top. IGVN will remove the remaining projections
1391 ac->set_req(0, top());
1392 ac->replace_edge(res, top(), &_igvn);
1393
1394 // Disconnect src right away: it can help find new
1395 // opportunities for allocation elimination
1396 Node* src = ac->in(ArrayCopyNode::Src);
1397 ac->replace_edge(src, top(), &_igvn);
1398 // src can be top at this point if src and dest of the
1399 // arraycopy were the same
1400 if (src->outcnt() == 0 && !src->is_top()) {
1401 _igvn.remove_dead_node(src, PhaseIterGVN::NodeOrigin::Graph);
1402 }
1403 }
1404 _igvn._worklist.push(ac);
1405 } else if (use->is_InlineType()) {
1406 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1407 // Cut off oop input and remove known instance id from type
1408 _igvn.rehash_node_delayed(use);
1409 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1410 use->as_InlineType()->set_is_buffered(_igvn, false);
1411 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1412 _igvn.set_type(use, toop);
1413 use->as_InlineType()->set_type(toop);
1414 // Process users
1415 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1416 Node* u = use->fast_out(k);
1417 if (!u->is_InlineType() && !u->is_StoreFlat()) {
1418 worklist.push(u);
1419 }
1420 }
1421 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1422 // Store to mark word of inline type larval buffer
1423 assert(inline_alloc, "Unexpected store to mark word");
1424 _igvn.replace_node(use, use->in(MemNode::Memory));
1425 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1426 // Inline type buffer allocations are followed by a membar
1427 assert(inline_alloc, "Unexpected MemBarRelease");
1428 use->as_MemBar()->remove(&_igvn);
1429 } else if (use->is_ReachabilityFence() && OptimizeReachabilityFences) {
1430 use->as_ReachabilityFence()->clear_referent(_igvn); // redundant fence; will be removed during IGVN
1431 } else {
1432 eliminate_gc_barrier(use);
1433 }
1434 j -= (oc1 - res->outcnt());
1435 }
1436 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1437 _igvn.remove_dead_node(res, PhaseIterGVN::NodeOrigin::Graph);
1438 }
1439
1440 //
1441 // Process other users of allocation's projections
1442 //
1443 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1444 // First disconnect stores captured by Initialize node.
1445 // If Initialize node is eliminated first in the following code,
1446 // it will kill such stores and DUIterator_Last will assert.
1447 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1448 Node* use = _callprojs->resproj[0]->fast_out(j);
1449 if (use->is_AddP()) {
1450 // raw memory addresses used only by the initialization
1451 _igvn.replace_node(use, C->top());
1452 --j; --jmax;
1453 }
1454 }
1455 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1456 Node* use = _callprojs->resproj[0]->last_out(j);
1457 uint oc1 = _callprojs->resproj[0]->outcnt();
1458 if (use->is_Initialize()) {
1459 // Eliminate Initialize node.
1460 InitializeNode *init = use->as_Initialize();
1461 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1462 if (ctrl_proj != nullptr) {
1463 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1464 #ifdef ASSERT
1465 // If the InitializeNode has no memory out, it will die, and tmp will become null
1466 Node* tmp = init->in(TypeFunc::Control);
1467 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1468 #endif
1469 }
1470 Node* mem = init->in(TypeFunc::Memory);
1471 #ifdef ASSERT
1472 if (init->number_of_projs(TypeFunc::Memory) > 0) {
1473 if (mem->is_MergeMem()) {
1474 assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs->fallthrough_memproj, "allocation memory projection");
1475 } else {
1476 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1477 }
1478 }
1479 #endif
1480 init->replace_mem_projs_by(mem, &_igvn);
1481 assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1482 } else if (use->Opcode() == Op_MemBarStoreStore) {
1483 // Inline type buffer allocations are followed by a membar
1484 assert(inline_alloc, "Unexpected MemBarStoreStore");
1485 use->as_MemBar()->remove(&_igvn);
1486 } else {
1487 assert(false, "only Initialize or AddP expected");
1488 }
1489 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1490 }
1491 }
1492 if (_callprojs->fallthrough_catchproj != nullptr) {
1493 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1494 }
1495 if (_callprojs->fallthrough_memproj != nullptr) {
1496 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1497 }
1498 if (_callprojs->catchall_memproj != nullptr) {
1499 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1500 }
1501 if (_callprojs->fallthrough_ioproj != nullptr) {
1502 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1503 }
1504 if (_callprojs->catchall_ioproj != nullptr) {
1505 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1506 }
1507 if (_callprojs->catchall_catchproj != nullptr) {
1508 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1509 }
1510 }
1511
1512 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1513 // If reallocation fails during deoptimization we'll pop all
1514 // interpreter frames for this compiled frame and that won't play
1515 // nice with JVMTI popframe.
1516 // We avoid this issue by eager reallocation when the popframe request
1517 // is received.
1518 if (!EliminateAllocations) {
1519 return false;
1520 }
1521 Node* klass = alloc->in(AllocateNode::KlassNode);
1522 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1523
1524 // Attempt to eliminate inline type buffer allocations
1525 // regardless of usage and escape/replaceable status.
1526 bool inline_alloc = tklass->isa_instklassptr() &&
1527 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1528 if (!alloc->_is_non_escaping && !inline_alloc) {
1529 return false;
1530 }
1531 // Eliminate boxing allocations which are not used
1532 // regardless scalar replaceable status.
1533 Node* res = alloc->result_cast();
1534 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1535 tklass->isa_instklassptr() &&
1536 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1537 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1538 return false;
1539 }
1540
1541 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1542
1543 Unique_Node_List safepoints;
1544 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1545 return false;
1546 }
1547
1548 if (!alloc->_is_scalar_replaceable) {
1549 assert(res == nullptr || inline_alloc, "sanity");
1550 // We can only eliminate allocation if all debug info references
1551 // are already replaced with SafePointScalarObject because
1552 // we can't search for a fields value without instance_id.
1553 if (safepoints.size() > 0) {
1554 return false;
1555 }
1556 }
1557
1558 if (!scalar_replacement(alloc, safepoints)) {
1559 return false;
1560 }
1561
1562 CompileLog* log = C->log();
1563 if (log != nullptr) {
1564 log->head("eliminate_allocation type='%d'",
1565 log->identify(tklass->exact_klass()));
1566 JVMState* p = alloc->jvms();
1567 while (p != nullptr) {
1568 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1569 p = p->caller();
1570 }
1571 log->tail("eliminate_allocation");
1572 }
1573
1574 process_users_of_allocation(alloc, inline_alloc);
1575
1576 #ifndef PRODUCT
1577 if (PrintEliminateAllocations) {
1578 if (alloc->is_AllocateArray())
1579 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1580 else
1581 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1582 }
1583 #endif
1584
1585 return true;
1586 }
1587
1588 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode* call) {
1589 // EA should remove all uses of non-escaping boxing node.
1590 if (!C->eliminate_boxing()) {
1591 return false;
1592 }
1593 for (uint i = TypeFunc::Parms; i < call->tf()->range_cc()->cnt(); ++i) {
1594 if (call->proj_out_or_null(i) != nullptr) {
1595 return false;
1596 }
1597 }
1598
1599 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1600
1601 process_users_of_allocation(call);
1602
1603 CompileLog* log = C->log();
1604 if (log != nullptr) {
1605 const TypeInstPtr* t = nullptr;
1606 if (call->is_boxing_method()) {
1607 const TypeTuple* range = call->tf()->range_sig();
1608 t = range->field_at(TypeFunc::Parms)->isa_instptr();
1609 } else {
1610 assert(call->is_unboxing_method(), "Unexpected call");
1611 const TypeTuple* domain = call->tf()->domain_sig();
1612 t = domain->field_at(TypeFunc::Parms)->isa_instptr();
1613 assert(!t->maybe_null(), "missing receiver null check?");
1614 }
1615 log->head("eliminate_boxing type='%d'",
1616 log->identify(t->instance_klass()));
1617 JVMState* p = call->jvms();
1618 while (p != nullptr) {
1619 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1620 p = p->caller();
1621 }
1622 log->tail("eliminate_boxing");
1623 }
1624
1625 #ifndef PRODUCT
1626 if (PrintEliminateAllocations) {
1627 tty->print("++++ Eliminated: %d ", call->_idx);
1628 call->method()->print_short_name(tty);
1629 tty->cr();
1630 }
1631 #endif
1632
1633 return true;
1634 }
1635
1636 Node* PhaseMacroExpand::make_load_raw(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1637 Node* adr = off_heap_plus_addr(base, offset);
1638 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1639 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1640 transform_later(value);
1641 return value;
1642 }
1643
1644
1645 Node* PhaseMacroExpand::make_store_raw(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1646 Node* adr = off_heap_plus_addr(base, offset);
1647 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1648 transform_later(mem);
1649 return mem;
1650 }
1651
1652 //=============================================================================
1653 //
1654 // A L L O C A T I O N
1655 //
1656 // Allocation attempts to be fast in the case of frequent small objects.
1657 // It breaks down like this:
1658 //
1659 // 1) Size in doublewords is computed. This is a constant for objects and
1660 // variable for most arrays. Doubleword units are used to avoid size
1661 // overflow of huge doubleword arrays. We need doublewords in the end for
1662 // rounding.
1663 //
1664 // 2) Size is checked for being 'too large'. Too-large allocations will go
1665 // the slow path into the VM. The slow path can throw any required
1666 // exceptions, and does all the special checks for very large arrays. The
1667 // size test can constant-fold away for objects. For objects with
1668 // finalizers it constant-folds the otherway: you always go slow with
1669 // finalizers.
1670 //
1671 // 3) If NOT using TLABs, this is the contended loop-back point.
1672 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1673 //
1674 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1675 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1676 // "size*8" we always enter the VM, where "largish" is a constant picked small
1677 // enough that there's always space between the eden max and 4Gig (old space is
1678 // there so it's quite large) and large enough that the cost of entering the VM
1679 // is dwarfed by the cost to initialize the space.
1680 //
1681 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1682 // down. If contended, repeat at step 3. If using TLABs normal-store
1683 // adjusted heap top back down; there is no contention.
1684 //
1685 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1686 // fields.
1687 //
1688 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1689 // oop flavor.
1690 //
1691 //=============================================================================
1692 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1693 // Allocations bigger than this always go the slow route.
1694 // This value must be small enough that allocation attempts that need to
1695 // trigger exceptions go the slow route. Also, it must be small enough so
1696 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1697 //=============================================================================j//
1698 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1699 // The allocator will coalesce int->oop copies away. See comment in
1700 // coalesce.cpp about how this works. It depends critically on the exact
1701 // code shape produced here, so if you are changing this code shape
1702 // make sure the GC info for the heap-top is correct in and around the
1703 // slow-path call.
1704 //
1705
1706 void PhaseMacroExpand::expand_allocate_common(
1707 AllocateNode* alloc, // allocation node to be expanded
1708 Node* length, // array length for an array allocation
1709 Node* init_val, // value to initialize the array with
1710 const TypeFunc* slow_call_type, // Type of slow call
1711 address slow_call_address, // Address of slow call
1712 Node* valid_length_test // whether length is valid or not
1713 )
1714 {
1715 Node* ctrl = alloc->in(TypeFunc::Control);
1716 Node* mem = alloc->in(TypeFunc::Memory);
1717 Node* i_o = alloc->in(TypeFunc::I_O);
1718 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1719 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1720 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1721 assert(ctrl != nullptr, "must have control");
1722
1723 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1724 // they will not be used if "always_slow" is set
1725 enum { slow_result_path = 1, fast_result_path = 2 };
1726 Node *result_region = nullptr;
1727 Node *result_phi_rawmem = nullptr;
1728 Node *result_phi_rawoop = nullptr;
1729 Node *result_phi_i_o = nullptr;
1730
1731 // The initial slow comparison is a size check, the comparison
1732 // we want to do is a BoolTest::gt
1733 bool expand_fast_path = true;
1734 int tv = _igvn.find_int_con(initial_slow_test, -1);
1735 if (tv >= 0) {
1736 // InitialTest has constant result
1737 // 0 - can fit in TLAB
1738 // 1 - always too big or negative
1739 assert(tv <= 1, "0 or 1 if a constant");
1740 expand_fast_path = (tv == 0);
1741 initial_slow_test = nullptr;
1742 } else {
1743 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1744 }
1745
1746 if (!UseTLAB) {
1747 // Force slow-path allocation
1748 expand_fast_path = false;
1749 initial_slow_test = nullptr;
1750 }
1751
1752 // ArrayCopyNode right after an allocation operates on the raw result projection for the Allocate node so it's not
1753 // safe to remove such an allocation even if it has no result cast.
1754 bool allocation_has_use = (alloc->result_cast() != nullptr) || (alloc->initialization() != nullptr && alloc->initialization()->is_complete_with_arraycopy());
1755 if (!allocation_has_use) {
1756 InitializeNode* init = alloc->initialization();
1757 if (init != nullptr) {
1758 init->remove(&_igvn);
1759 }
1760 if (expand_fast_path && (initial_slow_test == nullptr)) {
1761 // Remove allocation node and return.
1762 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1763 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1764 #ifndef PRODUCT
1765 if (PrintEliminateAllocations) {
1766 tty->print("NotUsed ");
1767 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1768 if (res != nullptr) {
1769 res->dump();
1770 } else {
1771 alloc->dump();
1772 }
1773 }
1774 #endif
1775 yank_alloc_node(alloc);
1776 return;
1777 }
1778 }
1779
1780 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1781 Node *slow_region = nullptr;
1782 Node *toobig_false = ctrl;
1783
1784 // generate the initial test if necessary
1785 if (initial_slow_test != nullptr ) {
1786 assert (expand_fast_path, "Only need test if there is a fast path");
1787 slow_region = new RegionNode(3);
1788
1789 // Now make the initial failure test. Usually a too-big test but
1790 // might be a TRUE for finalizers.
1791 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1792 transform_later(toobig_iff);
1793 // Plug the failing-too-big test into the slow-path region
1794 Node* toobig_true = new IfTrueNode(toobig_iff);
1795 transform_later(toobig_true);
1796 slow_region ->init_req( too_big_or_final_path, toobig_true );
1797 toobig_false = new IfFalseNode(toobig_iff);
1798 transform_later(toobig_false);
1799 } else {
1800 // No initial test, just fall into next case
1801 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1802 toobig_false = ctrl;
1803 DEBUG_ONLY(slow_region = NodeSentinel);
1804 }
1805
1806 // If we are here there are several possibilities
1807 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1808 // no_initial_check means a constant allocation.
1809 // - If check always evaluates to false -> expand_fast_path is false (see above)
1810 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1811 // if !allocation_has_use the fast path is empty
1812 // if !allocation_has_use && no_initial_check
1813 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1814 // removed by yank_alloc_node above.
1815
1816 Node *slow_mem = mem; // save the current memory state for slow path
1817 // generate the fast allocation code unless we know that the initial test will always go slow
1818 if (expand_fast_path) {
1819 // Fast path modifies only raw memory.
1820 if (mem->is_MergeMem()) {
1821 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1822 }
1823
1824 // allocate the Region and Phi nodes for the result
1825 result_region = new RegionNode(3);
1826 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1827 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1828
1829 // Grab regular I/O before optional prefetch may change it.
1830 // Slow-path does no I/O so just set it to the original I/O.
1831 result_phi_i_o->init_req(slow_result_path, i_o);
1832
1833 // Name successful fast-path variables
1834 Node* fast_oop_ctrl;
1835 Node* fast_oop_rawmem;
1836
1837 if (allocation_has_use) {
1838 Node* needgc_ctrl = nullptr;
1839 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1840
1841 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1842 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1843 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1844 fast_oop_ctrl, fast_oop_rawmem,
1845 prefetch_lines);
1846
1847 if (initial_slow_test != nullptr) {
1848 // This completes all paths into the slow merge point
1849 slow_region->init_req(need_gc_path, needgc_ctrl);
1850 transform_later(slow_region);
1851 } else {
1852 // No initial slow path needed!
1853 // Just fall from the need-GC path straight into the VM call.
1854 slow_region = needgc_ctrl;
1855 }
1856
1857 InitializeNode* init = alloc->initialization();
1858 fast_oop_rawmem = initialize_object(alloc,
1859 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1860 klass_node, length, size_in_bytes);
1861 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1862 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1863
1864 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1865 } else {
1866 assert (initial_slow_test != nullptr, "sanity");
1867 fast_oop_ctrl = toobig_false;
1868 fast_oop_rawmem = mem;
1869 transform_later(slow_region);
1870 }
1871
1872 // Plug in the successful fast-path into the result merge point
1873 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1874 result_phi_i_o ->init_req(fast_result_path, i_o);
1875 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1876 } else {
1877 slow_region = ctrl;
1878 result_phi_i_o = i_o; // Rename it to use in the following code.
1879 }
1880
1881 // Generate slow-path call
1882 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1883 OptoRuntime::stub_name(slow_call_address),
1884 TypePtr::BOTTOM);
1885 call->init_req(TypeFunc::Control, slow_region);
1886 call->init_req(TypeFunc::I_O, top()); // does no i/o
1887 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1888 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1889 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1890
1891 call->init_req(TypeFunc::Parms+0, klass_node);
1892 if (length != nullptr) {
1893 call->init_req(TypeFunc::Parms+1, length);
1894 if (init_val != nullptr) {
1895 call->init_req(TypeFunc::Parms+2, init_val);
1896 }
1897 }
1898
1899 // Copy debug information and adjust JVMState information, then replace
1900 // allocate node with the call
1901 call->copy_call_debug_info(&_igvn, alloc);
1902 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1903 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1904 // path dies).
1905 if (valid_length_test != nullptr) {
1906 call->add_req(valid_length_test);
1907 }
1908 if (expand_fast_path) {
1909 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1910 } else {
1911 // Hook i_o projection to avoid its elimination during allocation
1912 // replacement (when only a slow call is generated).
1913 call->set_req(TypeFunc::I_O, result_phi_i_o);
1914 }
1915 _igvn.replace_node(alloc, call);
1916 transform_later(call);
1917
1918 // Identify the output projections from the allocate node and
1919 // adjust any references to them.
1920 // The control and io projections look like:
1921 //
1922 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1923 // Allocate Catch
1924 // ^---Proj(io) <-------+ ^---CatchProj(io)
1925 //
1926 // We are interested in the CatchProj nodes.
1927 //
1928 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1929
1930 // An allocate node has separate memory projections for the uses on
1931 // the control and i_o paths. Replace the control memory projection with
1932 // result_phi_rawmem (unless we are only generating a slow call when
1933 // both memory projections are combined)
1934 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1935 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1936 }
1937 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1938 // catchall_memproj so we end up with a call that has only 1 memory projection.
1939 if (_callprojs->catchall_memproj != nullptr) {
1940 if (_callprojs->fallthrough_memproj == nullptr) {
1941 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1942 transform_later(_callprojs->fallthrough_memproj);
1943 }
1944 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1945 _igvn.remove_dead_node(_callprojs->catchall_memproj, PhaseIterGVN::NodeOrigin::Graph);
1946 }
1947
1948 // An allocate node has separate i_o projections for the uses on the control
1949 // and i_o paths. Always replace the control i_o projection with result i_o
1950 // otherwise incoming i_o become dead when only a slow call is generated
1951 // (it is different from memory projections where both projections are
1952 // combined in such case).
1953 if (_callprojs->fallthrough_ioproj != nullptr) {
1954 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1955 }
1956 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1957 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1958 if (_callprojs->catchall_ioproj != nullptr) {
1959 if (_callprojs->fallthrough_ioproj == nullptr) {
1960 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1961 transform_later(_callprojs->fallthrough_ioproj);
1962 }
1963 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1964 _igvn.remove_dead_node(_callprojs->catchall_ioproj, PhaseIterGVN::NodeOrigin::Graph);
1965 }
1966
1967 // if we generated only a slow call, we are done
1968 if (!expand_fast_path) {
1969 // Now we can unhook i_o.
1970 if (result_phi_i_o->outcnt() > 1) {
1971 call->set_req(TypeFunc::I_O, top());
1972 } else {
1973 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1974 // Case of new array with negative size known during compilation.
1975 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1976 // following code since call to runtime will throw exception.
1977 // As result there will be no users of i_o after the call.
1978 // Leave i_o attached to this call to avoid problems in preceding graph.
1979 }
1980 return;
1981 }
1982
1983 if (_callprojs->fallthrough_catchproj != nullptr) {
1984 ctrl = _callprojs->fallthrough_catchproj->clone();
1985 transform_later(ctrl);
1986 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1987 } else {
1988 ctrl = top();
1989 }
1990 Node *slow_result;
1991 if (_callprojs->resproj[0] == nullptr) {
1992 // no uses of the allocation result
1993 slow_result = top();
1994 } else {
1995 slow_result = _callprojs->resproj[0]->clone();
1996 transform_later(slow_result);
1997 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1998 }
1999
2000 // Plug slow-path into result merge point
2001 result_region->init_req( slow_result_path, ctrl);
2002 transform_later(result_region);
2003 if (allocation_has_use) {
2004 result_phi_rawoop->init_req(slow_result_path, slow_result);
2005 transform_later(result_phi_rawoop);
2006 }
2007 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
2008 transform_later(result_phi_rawmem);
2009 transform_later(result_phi_i_o);
2010 // This completes all paths into the result merge point
2011 }
2012
2013 // Remove alloc node that has no uses.
2014 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
2015 Node* ctrl = alloc->in(TypeFunc::Control);
2016 Node* mem = alloc->in(TypeFunc::Memory);
2017 Node* i_o = alloc->in(TypeFunc::I_O);
2018
2019 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2020 if (_callprojs->resproj[0] != nullptr) {
2021 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
2022 Node* use = _callprojs->resproj[0]->fast_out(i);
2023 use->isa_MemBar()->remove(&_igvn);
2024 --imax;
2025 --i; // back up iterator
2026 }
2027 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
2028 _igvn.remove_dead_node(_callprojs->resproj[0], PhaseIterGVN::NodeOrigin::Graph);
2029 }
2030 if (_callprojs->fallthrough_catchproj != nullptr) {
2031 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
2032 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj, PhaseIterGVN::NodeOrigin::Graph);
2033 }
2034 if (_callprojs->catchall_catchproj != nullptr) {
2035 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
2036 _callprojs->catchall_catchproj->set_req(0, top());
2037 }
2038 if (_callprojs->fallthrough_proj != nullptr) {
2039 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
2040 _igvn.remove_dead_node(catchnode, PhaseIterGVN::NodeOrigin::Graph);
2041 _igvn.remove_dead_node(_callprojs->fallthrough_proj, PhaseIterGVN::NodeOrigin::Graph);
2042 }
2043 if (_callprojs->fallthrough_memproj != nullptr) {
2044 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
2045 _igvn.remove_dead_node(_callprojs->fallthrough_memproj, PhaseIterGVN::NodeOrigin::Graph);
2046 }
2047 if (_callprojs->fallthrough_ioproj != nullptr) {
2048 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
2049 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj, PhaseIterGVN::NodeOrigin::Graph);
2050 }
2051 if (_callprojs->catchall_memproj != nullptr) {
2052 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
2053 _callprojs->catchall_memproj->set_req(0, top());
2054 }
2055 if (_callprojs->catchall_ioproj != nullptr) {
2056 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
2057 _callprojs->catchall_ioproj->set_req(0, top());
2058 }
2059 #ifndef PRODUCT
2060 if (PrintEliminateAllocations) {
2061 if (alloc->is_AllocateArray()) {
2062 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
2063 } else {
2064 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
2065 }
2066 }
2067 #endif
2068 _igvn.remove_dead_node(alloc, PhaseIterGVN::NodeOrigin::Graph);
2069 }
2070
2071 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
2072 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
2073 // If initialization is performed by an array copy, any required
2074 // MemBarStoreStore was already added. If the object does not
2075 // escape no need for a MemBarStoreStore. If the object does not
2076 // escape in its initializer and memory barrier (MemBarStoreStore or
2077 // stronger) is already added at exit of initializer, also no need
2078 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
2079 // so that stores that initialize this object can't be reordered
2080 // with a subsequent store that makes this object accessible by
2081 // other threads.
2082 // Other threads include java threads and JVM internal threads
2083 // (for example concurrent GC threads). Current concurrent GC
2084 // implementation: G1 will not scan newly created object,
2085 // so it's safe to skip storestore barrier when allocation does
2086 // not escape.
2087 if (!alloc->does_not_escape_thread() &&
2088 !alloc->is_allocation_MemBar_redundant() &&
2089 (init == nullptr || !init->is_complete_with_arraycopy())) {
2090 if (init == nullptr || init->req() < InitializeNode::RawStores) {
2091 // No InitializeNode or no stores captured by zeroing
2092 // elimination. Simply add the MemBarStoreStore after object
2093 // initialization.
2094 // What we want is to prevent the compiler and the CPU from re-ordering the stores that initialize this object
2095 // with subsequent stores to any slice. As a consequence, this MemBar should capture the entire memory state at
2096 // this point in the IR and produce a new memory state that should cover all slices. However, the Initialize node
2097 // only captures/produces a partial memory state making it complicated to insert such a MemBar. Because
2098 // re-ordering by the compiler can't happen by construction (a later Store that publishes the just allocated
2099 // object reference is indirectly control dependent on the Initialize node), preventing reordering by the CPU is
2100 // sufficient. For that a MemBar on the raw memory slice is good enough.
2101 // If init is null, this allocation does have an InitializeNode but this logic can't locate it (see comment in
2102 // PhaseMacroExpand::initialize_object()).
2103 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
2104 transform_later(mb);
2105
2106 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
2107 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
2108 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
2109 transform_later(fast_oop_ctrl);
2110 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
2111 transform_later(fast_oop_rawmem);
2112 } else {
2113 // Add the MemBarStoreStore after the InitializeNode so that
2114 // all stores performing the initialization that were moved
2115 // before the InitializeNode happen before the storestore
2116 // barrier.
2117
2118 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
2119
2120 // See comment above that explains why a raw memory MemBar is good enough.
2121 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
2122 transform_later(mb);
2123
2124 Node* ctrl = new ProjNode(init, TypeFunc::Control);
2125 transform_later(ctrl);
2126 Node* old_raw_mem_proj = nullptr;
2127 auto find_raw_mem = [&](ProjNode* proj) {
2128 if (C->get_alias_index(proj->adr_type()) == Compile::AliasIdxRaw) {
2129 assert(old_raw_mem_proj == nullptr, "only one expected");
2130 old_raw_mem_proj = proj;
2131 }
2132 };
2133 init->for_each_proj(find_raw_mem, TypeFunc::Memory);
2134 assert(old_raw_mem_proj != nullptr, "should have found raw mem Proj");
2135 Node* raw_mem_proj = new ProjNode(init, TypeFunc::Memory);
2136 transform_later(raw_mem_proj);
2137
2138 // The MemBarStoreStore depends on control and memory coming
2139 // from the InitializeNode
2140 mb->init_req(TypeFunc::Memory, raw_mem_proj);
2141 mb->init_req(TypeFunc::Control, ctrl);
2142
2143 ctrl = new ProjNode(mb, TypeFunc::Control);
2144 transform_later(ctrl);
2145 Node* mem = new ProjNode(mb, TypeFunc::Memory);
2146 transform_later(mem);
2147
2148 // All nodes that depended on the InitializeNode for control
2149 // and memory must now depend on the MemBarNode that itself
2150 // depends on the InitializeNode
2151 if (init_ctrl != nullptr) {
2152 _igvn.replace_node(init_ctrl, ctrl);
2153 }
2154 _igvn.replace_node(old_raw_mem_proj, mem);
2155 }
2156 }
2157 }
2158
2159 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
2160 Node*& ctrl, Node*& rawmem) {
2161 if (C->env()->dtrace_alloc_probes()) {
2162 // Slow-path call
2163 int size = TypeFunc::Parms + 2;
2164 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
2165 CAST_FROM_FN_PTR(address,
2166 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
2167 "dtrace_object_alloc",
2168 TypeRawPtr::BOTTOM);
2169
2170 // Get base of thread-local storage area
2171 Node* thread = new ThreadLocalNode();
2172 transform_later(thread);
2173
2174 call->init_req(TypeFunc::Parms + 0, thread);
2175 call->init_req(TypeFunc::Parms + 1, oop);
2176 call->init_req(TypeFunc::Control, ctrl);
2177 call->init_req(TypeFunc::I_O , top()); // does no i/o
2178 call->init_req(TypeFunc::Memory , rawmem);
2179 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
2180 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
2181 transform_later(call);
2182 ctrl = new ProjNode(call, TypeFunc::Control);
2183 transform_later(ctrl);
2184 rawmem = new ProjNode(call, TypeFunc::Memory);
2185 transform_later(rawmem);
2186 }
2187 }
2188
2189 // Helper for PhaseMacroExpand::expand_allocate_common.
2190 // Initializes the newly-allocated storage.
2191 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
2192 Node* control, Node* rawmem, Node* object,
2193 Node* klass_node, Node* length,
2194 Node* size_in_bytes) {
2195 InitializeNode* init = alloc->initialization();
2196 // Store the klass & mark bits
2197 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
2198 if (!mark_node->is_Con()) {
2199 transform_later(mark_node);
2200 }
2201 rawmem = make_store_raw(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
2202
2203 if (!UseCompactObjectHeaders) {
2204 rawmem = make_store_raw(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2205 }
2206 int header_size = alloc->minimum_header_size(); // conservatively small
2207
2208 // Array length
2209 if (length != nullptr) { // Arrays need length field
2210 rawmem = make_store_raw(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
2211 // conservatively small header size:
2212 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2213 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
2214 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_exact_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
2215 if (is_reference_type(elem, true)) {
2216 elem = T_OBJECT;
2217 }
2218 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2219 }
2220 }
2221
2222 // Clear the object body, if necessary.
2223 if (init == nullptr) {
2224 // The init has somehow disappeared; be cautious and clear everything.
2225 //
2226 // This can happen if a node is allocated but an uncommon trap occurs
2227 // immediately. In this case, the Initialize gets associated with the
2228 // trap, and may be placed in a different (outer) loop, if the Allocate
2229 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2230 // there can be two Allocates to one Initialize. The answer in all these
2231 // edge cases is safety first. It is always safe to clear immediately
2232 // within an Allocate, and then (maybe or maybe not) clear some more later.
2233 if (!(UseTLAB && ZeroTLAB)) {
2234 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2235 alloc->in(AllocateNode::InitValue),
2236 alloc->in(AllocateNode::RawInitValue),
2237 header_size, size_in_bytes,
2238 true,
2239 &_igvn);
2240 }
2241 } else {
2242 if (!init->is_complete()) {
2243 // Try to win by zeroing only what the init does not store.
2244 // We can also try to do some peephole optimizations,
2245 // such as combining some adjacent subword stores.
2246 rawmem = init->complete_stores(control, rawmem, object,
2247 header_size, size_in_bytes, &_igvn);
2248 }
2249 // We have no more use for this link, since the AllocateNode goes away:
2250 init->set_req(InitializeNode::RawAddress, top());
2251 // (If we keep the link, it just confuses the register allocator,
2252 // who thinks he sees a real use of the address by the membar.)
2253 }
2254
2255 return rawmem;
2256 }
2257
2258 // Generate prefetch instructions for next allocations.
2259 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
2260 Node*& contended_phi_rawmem,
2261 Node* old_eden_top, Node* new_eden_top,
2262 intx lines) {
2263 enum { fall_in_path = 1, pf_path = 2 };
2264 if (UseTLAB && AllocatePrefetchStyle == 2) {
2265 // Generate prefetch allocation with watermark check.
2266 // As an allocation hits the watermark, we will prefetch starting
2267 // at a "distance" away from watermark.
2268
2269 Node* pf_region = new RegionNode(3);
2270 Node* pf_phi_rawmem = new PhiNode(pf_region, Type::MEMORY,
2271 TypeRawPtr::BOTTOM);
2272 // I/O is used for Prefetch
2273 Node* pf_phi_abio = new PhiNode(pf_region, Type::ABIO);
2274
2275 Node* thread = new ThreadLocalNode();
2276 transform_later(thread);
2277
2278 Node* eden_pf_adr = AddPNode::make_off_heap(thread,
2279 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())));
2280 transform_later(eden_pf_adr);
2281
2282 Node* old_pf_wm = new LoadPNode(needgc_false,
2283 contended_phi_rawmem, eden_pf_adr,
2284 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2285 MemNode::unordered);
2286 transform_later(old_pf_wm);
2287
2288 // check against new_eden_top
2289 Node* need_pf_cmp = new CmpPNode(new_eden_top, old_pf_wm);
2290 transform_later(need_pf_cmp);
2291 Node* need_pf_bol = new BoolNode(need_pf_cmp, BoolTest::ge);
2292 transform_later(need_pf_bol);
2293 IfNode* need_pf_iff = new IfNode(needgc_false, need_pf_bol,
2294 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2295 transform_later(need_pf_iff);
2296
2297 // true node, add prefetchdistance
2298 Node* need_pf_true = new IfTrueNode(need_pf_iff);
2299 transform_later(need_pf_true);
2300
2301 Node* need_pf_false = new IfFalseNode(need_pf_iff);
2302 transform_later(need_pf_false);
2303
2304 Node* new_pf_wmt = AddPNode::make_off_heap(old_pf_wm,
2305 _igvn.MakeConX(AllocatePrefetchDistance));
2306 transform_later(new_pf_wmt);
2307 new_pf_wmt->set_req(0, need_pf_true);
2308
2309 Node* store_new_wmt = new StorePNode(need_pf_true,
2310 contended_phi_rawmem, eden_pf_adr,
2311 TypeRawPtr::BOTTOM, new_pf_wmt,
2312 MemNode::unordered);
2313 transform_later(store_new_wmt);
2314
2315 // adding prefetches
2316 pf_phi_abio->init_req(fall_in_path, i_o);
2317
2318 Node* prefetch_adr;
2319 Node* prefetch;
2320 uint step_size = AllocatePrefetchStepSize;
2321 uint distance = 0;
2322
2323 for (intx i = 0; i < lines; i++) {
2324 prefetch_adr = AddPNode::make_off_heap(new_pf_wmt,
2325 _igvn.MakeConX(distance));
2326 transform_later(prefetch_adr);
2327 prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2328 transform_later(prefetch);
2329 distance += step_size;
2330 i_o = prefetch;
2331 }
2332 pf_phi_abio->set_req(pf_path, i_o);
2333
2334 pf_region->init_req(fall_in_path, need_pf_false);
2335 pf_region->init_req(pf_path, need_pf_true);
2336
2337 pf_phi_rawmem->init_req(fall_in_path, contended_phi_rawmem);
2338 pf_phi_rawmem->init_req(pf_path, store_new_wmt);
2339
2340 transform_later(pf_region);
2341 transform_later(pf_phi_rawmem);
2342 transform_later(pf_phi_abio);
2343
2344 needgc_false = pf_region;
2345 contended_phi_rawmem = pf_phi_rawmem;
2346 i_o = pf_phi_abio;
2347 } else if (UseTLAB && AllocatePrefetchStyle == 3) {
2348 // Insert a prefetch instruction for each allocation.
2349 // This code is used to generate 1 prefetch instruction per cache line.
2350
2351 // Generate several prefetch instructions.
2352 uint step_size = AllocatePrefetchStepSize;
2353 uint distance = AllocatePrefetchDistance;
2354
2355 // Next cache address.
2356 Node* cache_adr = AddPNode::make_off_heap(old_eden_top,
2357 _igvn.MakeConX(step_size + distance));
2358 transform_later(cache_adr);
2359 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2360 transform_later(cache_adr);
2361 // Address is aligned to execute prefetch to the beginning of cache line size.
2362 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2363 cache_adr = new AndXNode(cache_adr, mask);
2364 transform_later(cache_adr);
2365 cache_adr = new CastX2PNode(cache_adr);
2366 transform_later(cache_adr);
2367
2368 // Prefetch
2369 Node* prefetch = new PrefetchAllocationNode(contended_phi_rawmem, cache_adr);
2370 prefetch->set_req(0, needgc_false);
2371 transform_later(prefetch);
2372 contended_phi_rawmem = prefetch;
2373 Node* prefetch_adr;
2374 distance = step_size;
2375 for (intx i = 1; i < lines; i++) {
2376 prefetch_adr = AddPNode::make_off_heap(cache_adr,
2377 _igvn.MakeConX(distance));
2378 transform_later(prefetch_adr);
2379 prefetch = new PrefetchAllocationNode(contended_phi_rawmem, prefetch_adr);
2380 transform_later(prefetch);
2381 distance += step_size;
2382 contended_phi_rawmem = prefetch;
2383 }
2384 } else if (AllocatePrefetchStyle > 0) {
2385 // Insert a prefetch for each allocation only on the fast-path
2386 Node* prefetch_adr;
2387 Node* prefetch;
2388 // Generate several prefetch instructions.
2389 uint step_size = AllocatePrefetchStepSize;
2390 uint distance = AllocatePrefetchDistance;
2391 for (intx i = 0; i < lines; i++) {
2392 prefetch_adr = AddPNode::make_off_heap(new_eden_top,
2393 _igvn.MakeConX(distance));
2394 transform_later(prefetch_adr);
2395 prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2396 // Do not let it float too high, since if eden_top == eden_end,
2397 // both might be null.
2398 if (i == 0) { // Set control for first prefetch, next follows it
2399 prefetch->init_req(0, needgc_false);
2400 }
2401 transform_later(prefetch);
2402 distance += step_size;
2403 i_o = prefetch;
2404 }
2405 }
2406 return i_o;
2407 }
2408
2409
2410 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2411 expand_allocate_common(alloc, nullptr, nullptr,
2412 OptoRuntime::new_instance_Type(),
2413 OptoRuntime::new_instance_Java(), nullptr);
2414 }
2415
2416 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2417 Node* length = alloc->in(AllocateNode::ALength);
2418 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2419 InitializeNode* init = alloc->initialization();
2420 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2421 Node* init_value = alloc->in(AllocateNode::InitValue);
2422 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2423 assert(!ary_klass_t || !ary_klass_t->klass_is_exact() || !ary_klass_t->exact_klass()->is_obj_array_klass() ||
2424 ary_klass_t->is_refined_type(), "Must be a refined array klass");
2425 const TypeFunc* slow_call_type;
2426 address slow_call_address; // Address of slow call
2427 if (init != nullptr && init->is_complete_with_arraycopy() &&
2428 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2429 // Don't zero type array during slow allocation in VM since
2430 // it will be initialized later by arraycopy in compiled code.
2431 slow_call_address = OptoRuntime::new_array_nozero_Java();
2432 slow_call_type = OptoRuntime::new_array_nozero_Type();
2433 } else {
2434 slow_call_address = OptoRuntime::new_array_Java();
2435 slow_call_type = OptoRuntime::new_array_Type();
2436
2437 if (init_value == nullptr) {
2438 init_value = _igvn.zerocon(T_OBJECT);
2439 } else if (UseCompressedOops) {
2440 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2441 }
2442 }
2443 expand_allocate_common(alloc, length, init_value,
2444 slow_call_type,
2445 slow_call_address, valid_length_test);
2446 }
2447
2448 //-------------------mark_eliminated_box----------------------------------
2449 //
2450 // During EA obj may point to several objects but after few ideal graph
2451 // transformations (CCP) it may point to only one non escaping object
2452 // (but still using phi), corresponding locks and unlocks will be marked
2453 // for elimination. Later obj could be replaced with a new node (new phi)
2454 // and which does not have escape information. And later after some graph
2455 // reshape other locks and unlocks (which were not marked for elimination
2456 // before) are connected to this new obj (phi) but they still will not be
2457 // marked for elimination since new obj has no escape information.
2458 // Mark all associated (same box and obj) lock and unlock nodes for
2459 // elimination if some of them marked already.
2460 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2461 BoxLockNode* oldbox = box->as_BoxLock();
2462 if (oldbox->is_eliminated()) {
2463 return; // This BoxLock node was processed already.
2464 }
2465 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2466 // New implementation (EliminateNestedLocks) has separate BoxLock
2467 // node for each locked region so mark all associated locks/unlocks as
2468 // eliminated even if different objects are referenced in one locked region
2469 // (for example, OSR compilation of nested loop inside locked scope).
2470 if (EliminateNestedLocks ||
2471 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2472 // Box is used only in one lock region. Mark this box as eliminated.
2473 oldbox->set_local(); // This verifies correct state of BoxLock
2474 _igvn.hash_delete(oldbox);
2475 oldbox->set_eliminated(); // This changes box's hash value
2476 _igvn.hash_insert(oldbox);
2477
2478 for (uint i = 0; i < oldbox->outcnt(); i++) {
2479 Node* u = oldbox->raw_out(i);
2480 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2481 AbstractLockNode* alock = u->as_AbstractLock();
2482 // Check lock's box since box could be referenced by Lock's debug info.
2483 if (alock->box_node() == oldbox) {
2484 // Mark eliminated all related locks and unlocks.
2485 #ifdef ASSERT
2486 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2487 #endif
2488 alock->set_non_esc_obj();
2489 }
2490 }
2491 }
2492 return;
2493 }
2494
2495 // Create new "eliminated" BoxLock node and use it in monitor debug info
2496 // instead of oldbox for the same object.
2497 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2498
2499 // Note: BoxLock node is marked eliminated only here and it is used
2500 // to indicate that all associated lock and unlock nodes are marked
2501 // for elimination.
2502 newbox->set_local(); // This verifies correct state of BoxLock
2503 newbox->set_eliminated();
2504 transform_later(newbox);
2505
2506 // Replace old box node with new box for all users of the same object.
2507 for (uint i = 0; i < oldbox->outcnt();) {
2508 bool next_edge = true;
2509
2510 Node* u = oldbox->raw_out(i);
2511 if (u->is_AbstractLock()) {
2512 AbstractLockNode* alock = u->as_AbstractLock();
2513 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2514 // Replace Box and mark eliminated all related locks and unlocks.
2515 #ifdef ASSERT
2516 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2517 #endif
2518 alock->set_non_esc_obj();
2519 _igvn.rehash_node_delayed(alock);
2520 alock->set_box_node(newbox);
2521 next_edge = false;
2522 }
2523 }
2524 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2525 FastLockNode* flock = u->as_FastLock();
2526 assert(flock->box_node() == oldbox, "sanity");
2527 _igvn.rehash_node_delayed(flock);
2528 flock->set_box_node(newbox);
2529 next_edge = false;
2530 }
2531
2532 // Replace old box in monitor debug info.
2533 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2534 SafePointNode* sfn = u->as_SafePoint();
2535 JVMState* youngest_jvms = sfn->jvms();
2536 int max_depth = youngest_jvms->depth();
2537 for (int depth = 1; depth <= max_depth; depth++) {
2538 JVMState* jvms = youngest_jvms->of_depth(depth);
2539 int num_mon = jvms->nof_monitors();
2540 // Loop over monitors
2541 for (int idx = 0; idx < num_mon; idx++) {
2542 Node* obj_node = sfn->monitor_obj(jvms, idx);
2543 Node* box_node = sfn->monitor_box(jvms, idx);
2544 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2545 int j = jvms->monitor_box_offset(idx);
2546 _igvn.replace_input_of(u, j, newbox);
2547 next_edge = false;
2548 }
2549 }
2550 }
2551 }
2552 if (next_edge) i++;
2553 }
2554 }
2555
2556 //-----------------------mark_eliminated_locking_nodes-----------------------
2557 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2558 if (!alock->is_balanced()) {
2559 return; // Can't do any more elimination for this locking region
2560 }
2561 if (EliminateNestedLocks) {
2562 if (alock->is_nested()) {
2563 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2564 return;
2565 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2566 // Only Lock node has JVMState needed here.
2567 // Not that preceding claim is documented anywhere else.
2568 if (alock->jvms() != nullptr) {
2569 if (alock->as_Lock()->is_nested_lock_region()) {
2570 // Mark eliminated related nested locks and unlocks.
2571 Node* obj = alock->obj_node();
2572 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2573 assert(!box_node->is_eliminated(), "should not be marked yet");
2574 // Note: BoxLock node is marked eliminated only here
2575 // and it is used to indicate that all associated lock
2576 // and unlock nodes are marked for elimination.
2577 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2578 for (uint i = 0; i < box_node->outcnt(); i++) {
2579 Node* u = box_node->raw_out(i);
2580 if (u->is_AbstractLock()) {
2581 alock = u->as_AbstractLock();
2582 if (alock->box_node() == box_node) {
2583 // Verify that this Box is referenced only by related locks.
2584 assert(alock->obj_node()->eqv_uncast(obj), "");
2585 // Mark all related locks and unlocks.
2586 #ifdef ASSERT
2587 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2588 #endif
2589 alock->set_nested();
2590 }
2591 }
2592 }
2593 } else {
2594 #ifdef ASSERT
2595 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2596 if (C->log() != nullptr)
2597 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2598 #endif
2599 }
2600 }
2601 return;
2602 }
2603 // Process locks for non escaping object
2604 assert(alock->is_non_esc_obj(), "");
2605 } // EliminateNestedLocks
2606
2607 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2608 // Look for all locks of this object and mark them and
2609 // corresponding BoxLock nodes as eliminated.
2610 Node* obj = alock->obj_node();
2611 for (uint j = 0; j < obj->outcnt(); j++) {
2612 Node* o = obj->raw_out(j);
2613 if (o->is_AbstractLock() &&
2614 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2615 alock = o->as_AbstractLock();
2616 Node* box = alock->box_node();
2617 // Replace old box node with new eliminated box for all users
2618 // of the same object and mark related locks as eliminated.
2619 mark_eliminated_box(box, obj);
2620 }
2621 }
2622 }
2623 }
2624
2625 // we have determined that this lock/unlock can be eliminated, we simply
2626 // eliminate the node without expanding it.
2627 //
2628 // Note: The membar's associated with the lock/unlock are currently not
2629 // eliminated. This should be investigated as a future enhancement.
2630 //
2631 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2632
2633 if (!alock->is_eliminated()) {
2634 return false;
2635 }
2636 #ifdef ASSERT
2637 if (!alock->is_coarsened()) {
2638 // Check that new "eliminated" BoxLock node is created.
2639 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2640 assert(oldbox->is_eliminated(), "should be done already");
2641 }
2642 #endif
2643
2644 alock->log_lock_optimization(C, "eliminate_lock");
2645
2646 #ifndef PRODUCT
2647 if (PrintEliminateLocks) {
2648 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2649 }
2650 #endif
2651
2652 Node* mem = alock->in(TypeFunc::Memory);
2653 Node* ctrl = alock->in(TypeFunc::Control);
2654 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2655
2656 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2657 // There are 2 projections from the lock. The lock node will
2658 // be deleted when its last use is subsumed below.
2659 assert(alock->outcnt() == 2 &&
2660 _callprojs->fallthrough_proj != nullptr &&
2661 _callprojs->fallthrough_memproj != nullptr,
2662 "Unexpected projections from Lock/Unlock");
2663
2664 Node* fallthroughproj = _callprojs->fallthrough_proj;
2665 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2666
2667 // The memory projection from a lock/unlock is RawMem
2668 // The input to a Lock is merged memory, so extract its RawMem input
2669 // (unless the MergeMem has been optimized away.)
2670 if (alock->is_Lock()) {
2671 // Search for MemBarAcquireLock node and delete it also.
2672 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2673 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2674 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2675 Node* memproj = membar->proj_out(TypeFunc::Memory);
2676 _igvn.replace_node(ctrlproj, fallthroughproj);
2677 _igvn.replace_node(memproj, memproj_fallthrough);
2678
2679 // Delete FastLock node also if this Lock node is unique user
2680 // (a loop peeling may clone a Lock node).
2681 Node* flock = alock->as_Lock()->fastlock_node();
2682 if (flock->outcnt() == 1) {
2683 assert(flock->unique_out() == alock, "sanity");
2684 _igvn.replace_node(flock, top());
2685 }
2686 }
2687
2688 // Search for MemBarReleaseLock node and delete it also.
2689 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2690 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2691 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2692 mem->is_Proj() && membar == mem->in(0), "");
2693 _igvn.replace_node(fallthroughproj, ctrl);
2694 _igvn.replace_node(memproj_fallthrough, mem);
2695 fallthroughproj = ctrl;
2696 memproj_fallthrough = mem;
2697 ctrl = membar->in(TypeFunc::Control);
2698 mem = membar->in(TypeFunc::Memory);
2699 }
2700
2701 _igvn.replace_node(fallthroughproj, ctrl);
2702 _igvn.replace_node(memproj_fallthrough, mem);
2703 return true;
2704 }
2705
2706
2707 //------------------------------expand_lock_node----------------------
2708 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2709
2710 Node* ctrl = lock->in(TypeFunc::Control);
2711 Node* mem = lock->in(TypeFunc::Memory);
2712 Node* obj = lock->obj_node();
2713 Node* box = lock->box_node();
2714 Node* flock = lock->fastlock_node();
2715
2716 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2717
2718 // Make the merge point
2719 Node *region;
2720 Node *mem_phi;
2721 Node *slow_path;
2722
2723 region = new RegionNode(3);
2724 // create a Phi for the memory state
2725 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2726
2727 // Optimize test; set region slot 2
2728 slow_path = opt_bits_test(ctrl, region, 2, flock);
2729 mem_phi->init_req(2, mem);
2730
2731 // Make slow path call
2732 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2733 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2734 obj, box, nullptr);
2735
2736 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2737
2738 // Slow path can only throw asynchronous exceptions, which are always
2739 // de-opted. So the compiler thinks the slow-call can never throw an
2740 // exception. If it DOES throw an exception we would need the debug
2741 // info removed first (since if it throws there is no monitor).
2742 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2743 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2744
2745 // Capture slow path
2746 // disconnect fall-through projection from call and create a new one
2747 // hook up users of fall-through projection to region
2748 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2749 transform_later(slow_ctrl);
2750 _igvn.hash_delete(_callprojs->fallthrough_proj);
2751 _callprojs->fallthrough_proj->disconnect_inputs(C);
2752 region->init_req(1, slow_ctrl);
2753 // region inputs are now complete
2754 transform_later(region);
2755 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2756
2757 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2758
2759 mem_phi->init_req(1, memproj);
2760
2761 transform_later(mem_phi);
2762
2763 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2764 }
2765
2766 //------------------------------expand_unlock_node----------------------
2767 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2768
2769 Node* ctrl = unlock->in(TypeFunc::Control);
2770 Node* mem = unlock->in(TypeFunc::Memory);
2771 Node* obj = unlock->obj_node();
2772 Node* box = unlock->box_node();
2773
2774 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2775
2776 // No need for a null check on unlock
2777
2778 // Make the merge point
2779 Node* region = new RegionNode(3);
2780
2781 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2782 funlock = transform_later( funlock )->as_FastUnlock();
2783 // Optimize test; set region slot 2
2784 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2785 Node *thread = transform_later(new ThreadLocalNode());
2786
2787 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2788 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2789 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2790
2791 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2792 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2793 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2794
2795 // No exceptions for unlocking
2796 // Capture slow path
2797 // disconnect fall-through projection from call and create a new one
2798 // hook up users of fall-through projection to region
2799 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2800 transform_later(slow_ctrl);
2801 _igvn.hash_delete(_callprojs->fallthrough_proj);
2802 _callprojs->fallthrough_proj->disconnect_inputs(C);
2803 region->init_req(1, slow_ctrl);
2804 // region inputs are now complete
2805 transform_later(region);
2806 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2807
2808 if (_callprojs->fallthrough_memproj != nullptr) {
2809 // create a Phi for the memory state
2810 Node* mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2811 Node* memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2812 mem_phi->init_req(1, memproj);
2813 mem_phi->init_req(2, mem);
2814 transform_later(mem_phi);
2815 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2816 }
2817 }
2818
2819 // An inline type might be returned from the call but we don't know its
2820 // type. Either we get a buffered inline type (and nothing needs to be done)
2821 // or one of the values being returned is the klass of the inline type
2822 // and we need to allocate an inline type instance of that type and
2823 // initialize it with other values being returned. In that case, we
2824 // first try a fast path allocation and initialize the value with the
2825 // inline klass's pack handler or we fall back to a runtime call.
2826 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2827 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2828 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2829 if (ret == nullptr) {
2830 return;
2831 }
2832 const TypeFunc* tf = call->_tf;
2833 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2834 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2835 call->_tf = new_tf;
2836 // Make sure the change of type is applied before projections are processed by igvn
2837 _igvn.set_type(call, call->Value(&_igvn));
2838 _igvn.set_type(ret, ret->Value(&_igvn));
2839
2840 // Before any new projection is added:
2841 CallProjections* projs = call->extract_projections(true, true);
2842
2843 // Create temporary hook nodes that will be replaced below.
2844 // Add an input to prevent hook nodes from being dead.
2845 Node* ctl = new Node(call);
2846 Node* mem = new Node(ctl);
2847 Node* io = new Node(ctl);
2848 Node* ex_ctl = new Node(ctl);
2849 Node* ex_mem = new Node(ctl);
2850 Node* ex_io = new Node(ctl);
2851 Node* res = new Node(ctl);
2852
2853 // Allocate a new buffered inline type only if a new one is not returned
2854 Node* cast = transform_later(new CastP2XNode(ctl, res));
2855 Node* mask = MakeConX(0x1);
2856 Node* masked = transform_later(new AndXNode(cast, mask));
2857 Node* cmp = transform_later(new CmpXNode(masked, mask));
2858 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2859 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2860 transform_later(allocation_iff);
2861 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2862 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2863 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2864
2865 // Try to allocate a new buffered inline instance either from TLAB or eden space
2866 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2867 CallLeafNoFPNode* handler_call;
2868 const bool alloc_in_place = UseTLAB;
2869 if (alloc_in_place) {
2870 Node* fast_oop_ctrl = nullptr;
2871 Node* fast_oop_rawmem = nullptr;
2872 Node* mask2 = MakeConX(-2);
2873 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2874 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2875 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2876 Node* layout_val = make_load_raw(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2877 Node* size_in_bytes = ConvI2X(layout_val);
2878 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2879 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2880 fast_oop_ctrl, fast_oop_rawmem,
2881 AllocateInstancePrefetchLines);
2882 // Allocation succeed, initialize buffered inline instance header firstly,
2883 // and then initialize its fields with an inline class specific handler
2884 Node* mark_word_node;
2885 if (UseCompactObjectHeaders) {
2886 // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2887 mark_word_node = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2888 } else {
2889 // Otherwise, use the static prototype.
2890 mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2891 }
2892
2893 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2894 if (!UseCompactObjectHeaders) {
2895 // COH: Everything is encoded in the mark word, so nothing left to do.
2896 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2897 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2898 }
2899 Node* members = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InlineKlass::adr_members_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2900 Node* pack_handler = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, members, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2901 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2902 nullptr,
2903 "pack handler",
2904 TypeRawPtr::BOTTOM);
2905 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2906 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2907 handler_call->init_req(TypeFunc::I_O, top());
2908 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2909 handler_call->init_req(TypeFunc::ReturnAdr, top());
2910 handler_call->init_req(TypeFunc::Parms, pack_handler);
2911 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2912 } else {
2913 needgc_ctrl = allocation_ctl;
2914 }
2915
2916 // Allocation failed, fall back to a runtime call
2917 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2918 SharedRuntime::store_inline_type_fields_to_buf_entry(),
2919 "store_inline_type_fields",
2920 TypePtr::BOTTOM);
2921 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2922 slow_call->init_req(TypeFunc::Memory, mem);
2923 slow_call->init_req(TypeFunc::I_O, io);
2924 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2925 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2926 slow_call->init_req(TypeFunc::Parms, res);
2927
2928 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2929 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2930 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2931 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2932 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2933 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2934 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2935
2936 Node* ex_r = new RegionNode(3);
2937 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2938 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2939 ex_r->init_req(1, slow_excp);
2940 ex_mem_phi->init_req(1, slow_mem);
2941 ex_io_phi->init_req(1, slow_io);
2942 ex_r->init_req(2, ex_ctl);
2943 ex_mem_phi->init_req(2, ex_mem);
2944 ex_io_phi->init_req(2, ex_io);
2945 transform_later(ex_r);
2946 transform_later(ex_mem_phi);
2947 transform_later(ex_io_phi);
2948
2949 // We don't know how many values are returned. This assumes the
2950 // worst case, that all available registers are used.
2951 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2952 if (domain->field_at(i) == Type::HALF) {
2953 slow_call->init_req(i, top());
2954 if (alloc_in_place) {
2955 handler_call->init_req(i+1, top());
2956 }
2957 continue;
2958 }
2959 Node* proj = transform_later(new ProjNode(call, i));
2960 slow_call->init_req(i, proj);
2961 if (alloc_in_place) {
2962 handler_call->init_req(i+1, proj);
2963 }
2964 }
2965 // We can safepoint at that new call
2966 slow_call->copy_call_debug_info(&_igvn, call);
2967 transform_later(slow_call);
2968 if (alloc_in_place) {
2969 transform_later(handler_call);
2970 }
2971
2972 Node* fast_ctl = nullptr;
2973 Node* fast_res = nullptr;
2974 MergeMemNode* fast_mem = nullptr;
2975 if (alloc_in_place) {
2976 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2977 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2978 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2979 fast_mem = MergeMemNode::make(mem);
2980 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2981 transform_later(fast_mem);
2982 }
2983
2984 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2985 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2986 Node* io_phi = new PhiNode(r, Type::ABIO);
2987 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2988 r->init_req(1, no_allocation_ctl);
2989 mem_phi->init_req(1, mem);
2990 io_phi->init_req(1, io);
2991 res_phi->init_req(1, no_allocation_res);
2992 r->init_req(2, slow_norm);
2993 mem_phi->init_req(2, slow_mem);
2994 io_phi->init_req(2, slow_io);
2995 res_phi->init_req(2, slow_res);
2996 if (alloc_in_place) {
2997 r->init_req(3, fast_ctl);
2998 mem_phi->init_req(3, fast_mem);
2999 io_phi->init_req(3, io);
3000 res_phi->init_req(3, fast_res);
3001 }
3002 transform_later(r);
3003 transform_later(mem_phi);
3004 transform_later(io_phi);
3005 transform_later(res_phi);
3006
3007 // Do not let stores that initialize this buffer be reordered with a subsequent
3008 // store that would make this buffer accessible by other threads.
3009 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
3010 transform_later(mb);
3011 mb->init_req(TypeFunc::Memory, mem_phi);
3012 mb->init_req(TypeFunc::Control, r);
3013 r = new ProjNode(mb, TypeFunc::Control);
3014 transform_later(r);
3015 mem_phi = new ProjNode(mb, TypeFunc::Memory);
3016 transform_later(mem_phi);
3017
3018 assert(projs->nb_resproj == 1, "unexpected number of results");
3019 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
3020 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
3021 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
3022 _igvn.replace_in_uses(projs->resproj[0], res_phi);
3023 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
3024 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
3025 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
3026 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
3027 Node* cn = projs->fallthrough_catchproj->in(0);
3028 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
3029
3030 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
3031 _igvn.replace_node(mem, projs->fallthrough_memproj);
3032 _igvn.replace_node(io, projs->fallthrough_ioproj);
3033 _igvn.replace_node(res, projs->resproj[0]);
3034 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
3035 _igvn.replace_node(ex_mem, projs->catchall_memproj);
3036 _igvn.replace_node(ex_io, projs->catchall_ioproj);
3037 }
3038
3039 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
3040 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
3041 Node* bol = check->unique_out();
3042 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
3043 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
3044 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
3045
3046 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
3047 Node* iff = bol->last_out(i);
3048 assert(iff->is_If(), "where's the if?");
3049
3050 if (iff->in(0)->is_top()) {
3051 _igvn.replace_input_of(iff, 1, C->top());
3052 continue;
3053 }
3054
3055 IfTrueNode* iftrue = iff->as_If()->true_proj();
3056 IfFalseNode* iffalse = iff->as_If()->false_proj();
3057 Node* ctrl = iff->in(0);
3058
3059 Node* subklass = nullptr;
3060 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
3061 subklass = obj_or_subklass;
3062 } else {
3063 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
3064 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3065 }
3066
3067 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
3068
3069 _igvn.replace_input_of(iff, 0, C->top());
3070 _igvn.replace_node(iftrue, not_subtype_ctrl);
3071 _igvn.replace_node(iffalse, ctrl);
3072 }
3073 _igvn.replace_node(check, C->top());
3074 }
3075
3076 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
3077 //
3078 // long mark = array1.mark | array2.mark | ...;
3079 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
3080 // if (locked_bit == 0) {
3081 // // One array is locked, load prototype header from the klass
3082 // mark = array1.klass.proto | array2.klass.proto | ...
3083 // }
3084 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
3085 // ...
3086 // }
3087 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
3088 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
3089 if (array_inputs) {
3090 Node* mark = MakeConX(0);
3091 Node* locked_bit = MakeConX(markWord::unlocked_value);
3092 Node* mem = check->in(FlatArrayCheckNode::Memory);
3093 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3094 Node* ary = check->in(i);
3095 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
3096 assert(t != nullptr, "Mixing array and klass inputs");
3097 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3098 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
3099 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));
3100 mark = _igvn.transform(new OrXNode(mark, mark_load));
3101 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
3102 }
3103 assert(!mark->is_Con(), "Should have been optimized out");
3104 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3105 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
3106
3107 // BoolNode might be shared, replace each if user
3108 Node* old_bol = check->unique_out();
3109 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
3110 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3111 IfNode* old_iff = old_bol->last_out(i)->as_If();
3112 Node* ctrl = old_iff->in(0);
3113 RegionNode* region = new RegionNode(3);
3114 Node* mark_phi = new PhiNode(region, TypeX_X);
3115
3116 // Check if array is unlocked
3117 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
3118
3119 // Unlocked: Use bits from mark word
3120 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
3121 mark_phi->init_req(1, mark);
3122
3123 // Locked: Load prototype header from klass
3124 ctrl = _igvn.transform(new IfFalseNode(iff));
3125 Node* proto = MakeConX(0);
3126 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3127 Node* ary = check->in(i);
3128 // Make loads control dependent to make sure they are only executed if array is locked
3129 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3130 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3131 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3132 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));
3133 proto = _igvn.transform(new OrXNode(proto, proto_load));
3134 }
3135 region->init_req(2, ctrl);
3136 mark_phi->init_req(2, proto);
3137
3138 // Check if flat array bits are set
3139 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
3140 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
3141 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
3142 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
3143
3144 ctrl = _igvn.transform(region);
3145 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
3146 _igvn.replace_node(old_iff, iff);
3147 }
3148 _igvn.replace_node(check, C->top());
3149 } else {
3150 // Fall back to layout helper check
3151 Node* lhs = intcon(0);
3152 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3153 Node* array_or_klass = check->in(i);
3154 Node* klass = nullptr;
3155 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
3156 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3157 if (t->isa_oopptr() != nullptr) {
3158 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
3159 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3160 } else {
3161 assert(t->isa_klassptr(), "Unexpected input type");
3162 klass = array_or_klass;
3163 }
3164 Node* lh_addr = basic_plus_adr(top(), klass, in_bytes(Klass::layout_helper_offset()));
3165 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));
3166 lhs = _igvn.transform(new OrINode(lhs, lh_val));
3167 }
3168 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
3169 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
3170 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
3171 Node* m2b = transform_later(new Conv2BNode(masked));
3172 // The matcher expects the input to If/CMove nodes to be produced by a Bool(CmpI..)
3173 // pattern, but the input to other potential users (e.g. Phi) to be some
3174 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
3175 Node* old_bol = check->unique_out();
3176 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3177 Node* user = old_bol->last_out(i);
3178 for (uint j = 0; j < user->req(); j++) {
3179 Node* n = user->in(j);
3180 if (n == old_bol) {
3181 _igvn.replace_input_of(user, j, (user->is_If() || user->is_CMove()) ? bol : m2b);
3182 }
3183 }
3184 }
3185 _igvn.replace_node(check, C->top());
3186 }
3187 }
3188
3189 // Perform refining of strip mined loop nodes in the macro nodes list.
3190 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
3191 for (int i = C->macro_count(); i > 0; i--) {
3192 Node* n = C->macro_node(i - 1);
3193 if (n->is_OuterStripMinedLoop()) {
3194 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3195 }
3196 }
3197 }
3198
3199 //---------------------------eliminate_macro_nodes----------------------
3200 // Eliminate scalar replaced allocations and associated locks.
3201 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
3202 if (C->macro_count() == 0) {
3203 return;
3204 }
3205
3206 if (StressMacroElimination) {
3207 C->shuffle_macro_nodes();
3208 }
3209 NOT_PRODUCT(int membar_before = count_MemBar(C);)
3210
3211 int iteration = 0;
3212 while (C->macro_count() > 0) {
3213 if (iteration++ > 100) {
3214 assert(false, "Too slow convergence of macro elimination");
3215 break;
3216 }
3217
3218 // Postpone lock elimination to after EA when most allocations are eliminated
3219 // because they might block lock elimination if their escape state isn't
3220 // determined yet and we only got one chance at eliminating the lock.
3221 if (eliminate_locks) {
3222 // Before elimination may re-mark (change to Nested or NonEscObj)
3223 // all associated (same box and obj) lock and unlock nodes.
3224 int cnt = C->macro_count();
3225 for (int i=0; i < cnt; i++) {
3226 Node *n = C->macro_node(i);
3227 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3228 mark_eliminated_locking_nodes(n->as_AbstractLock());
3229 }
3230 }
3231 // Re-marking may break consistency of Coarsened locks.
3232 if (!C->coarsened_locks_consistent()) {
3233 return; // recompile without Coarsened locks if broken
3234 } else {
3235 // After coarsened locks are eliminated locking regions
3236 // become unbalanced. We should not execute any more
3237 // locks elimination optimizations on them.
3238 C->mark_unbalanced_boxes();
3239 }
3240 }
3241
3242 bool progress = false;
3243 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3244 Node* n = C->macro_node(i - 1);
3245 bool success = false;
3246 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3247 switch (n->class_id()) {
3248 case Node::Class_Allocate:
3249 case Node::Class_AllocateArray:
3250 success = eliminate_allocate_node(n->as_Allocate());
3251 #ifndef PRODUCT
3252 if (success && PrintOptoStatistics) {
3253 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3254 }
3255 #endif
3256 break;
3257 case Node::Class_CallStaticJava: {
3258 CallStaticJavaNode* call = n->as_CallStaticJava();
3259 if (!call->method()->is_method_handle_intrinsic()) {
3260 success = eliminate_boxing_node(n->as_CallStaticJava());
3261 }
3262 break;
3263 }
3264 case Node::Class_Lock:
3265 case Node::Class_Unlock:
3266 if (eliminate_locks) {
3267 success = eliminate_locking_node(n->as_AbstractLock());
3268 #ifndef PRODUCT
3269 if (success && PrintOptoStatistics) {
3270 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3271 }
3272 #endif
3273 }
3274 break;
3275 case Node::Class_ArrayCopy:
3276 break;
3277 case Node::Class_OuterStripMinedLoop:
3278 break;
3279 case Node::Class_SubTypeCheck:
3280 break;
3281 case Node::Class_Opaque1:
3282 break;
3283 case Node::Class_FlatArrayCheck:
3284 break;
3285 default:
3286 assert(n->Opcode() == Op_LoopLimit ||
3287 n->Opcode() == Op_ModD ||
3288 n->Opcode() == Op_ModF ||
3289 n->Opcode() == Op_PowD ||
3290 n->is_OpaqueConstantBool() ||
3291 n->is_OpaqueInitializedAssertionPredicate() ||
3292 n->Opcode() == Op_MaxL ||
3293 n->Opcode() == Op_MinL ||
3294 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3295 "unknown node type in macro list");
3296 }
3297 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3298 progress = progress || success;
3299 if (success) {
3300 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3301 }
3302 }
3303
3304 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3305 // transformation is pending). If an allocation is used only in safepoints, elimination of
3306 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3307 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3308 // iteration.
3309 _igvn.set_delay_transform(false);
3310 _igvn.optimize();
3311 if (C->failing()) {
3312 return;
3313 }
3314 _igvn.set_delay_transform(true);
3315
3316 if (!progress) {
3317 break;
3318 }
3319 }
3320 #ifndef PRODUCT
3321 if (PrintOptoStatistics) {
3322 int membar_after = count_MemBar(C);
3323 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3324 }
3325 #endif
3326 }
3327
3328 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3329 if (C->macro_count() == 0) {
3330 return;
3331 }
3332 refine_strip_mined_loop_macro_nodes();
3333 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3334 bool progress = true;
3335 while (progress) {
3336 progress = false;
3337 for (int i = C->macro_count(); i > 0; i--) {
3338 Node* n = C->macro_node(i-1);
3339 bool success = false;
3340 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3341 if (n->Opcode() == Op_LoopLimit) {
3342 // Remove it from macro list and put on IGVN worklist to optimize.
3343 C->remove_macro_node(n);
3344 _igvn._worklist.push(n);
3345 success = true;
3346 } else if (n->Opcode() == Op_CallStaticJava) {
3347 CallStaticJavaNode* call = n->as_CallStaticJava();
3348 if (!call->method()->is_method_handle_intrinsic()) {
3349 // Remove it from macro list and put on IGVN worklist to optimize.
3350 C->remove_macro_node(n);
3351 _igvn._worklist.push(n);
3352 success = true;
3353 }
3354 } else if (n->is_Opaque1()) {
3355 _igvn.replace_node(n, n->in(1));
3356 success = true;
3357 } else if (n->is_OpaqueConstantBool()) {
3358 // Tests with OpaqueConstantBool nodes are implicitly known. Replace the node with true/false. In debug builds,
3359 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3360 // we will execute a Halt node.
3361 #ifdef ASSERT
3362 _igvn.replace_node(n, n->in(1));
3363 #else
3364 _igvn.replace_node(n, _igvn.intcon(n->as_OpaqueConstantBool()->constant()));
3365 #endif
3366 success = true;
3367 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3368 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3369 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3370 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3371 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3372 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3373 #ifdef ASSERT
3374 _igvn.replace_node(n, n->in(1));
3375 #else
3376 _igvn.replace_node(n, _igvn.intcon(1));
3377 #endif // ASSERT
3378 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3379 C->remove_macro_node(n);
3380 success = true;
3381 } else if (n->Opcode() == Op_MaxL) {
3382 // Since MaxL and MinL are not implemented in the backend, we expand them to
3383 // a CMoveL construct now. At least until here, the type could be computed
3384 // precisely. CMoveL is not so smart, but we can give it at least the best
3385 // type we know abouot n now.
3386 Node* repl = MinMaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3387 _igvn.replace_node(n, repl);
3388 success = true;
3389 } else if (n->Opcode() == Op_MinL) {
3390 Node* repl = MinMaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3391 _igvn.replace_node(n, repl);
3392 success = true;
3393 }
3394 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3395 progress = progress || success;
3396 if (success) {
3397 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3398 }
3399 }
3400 }
3401 }
3402
3403 //------------------------------expand_macro_nodes----------------------
3404 // Returns true if a failure occurred.
3405 bool PhaseMacroExpand::expand_macro_nodes() {
3406 if (StressMacroExpansion) {
3407 C->shuffle_macro_nodes();
3408 }
3409
3410 // Clean up the graph so we're less likely to hit the maximum node
3411 // limit
3412 _igvn.set_delay_transform(false);
3413 _igvn.optimize();
3414 if (C->failing()) return true;
3415 _igvn.set_delay_transform(true);
3416
3417
3418 // Because we run IGVN after each expansion, some macro nodes may go
3419 // dead and be removed from the list as we iterate over it. Move
3420 // Allocate nodes (processed in a second pass) at the beginning of
3421 // the list and then iterate from the last element of the list until
3422 // an Allocate node is seen. This is robust to random deletion in
3423 // the list due to nodes going dead.
3424 C->sort_macro_nodes();
3425
3426 // expand arraycopy "macro" nodes first
3427 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3428 // before the allocate nodes are expanded.
3429 while (C->macro_count() > 0) {
3430 int macro_count = C->macro_count();
3431 Node * n = C->macro_node(macro_count-1);
3432 assert(n->is_macro(), "only macro nodes expected here");
3433 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3434 // node is unreachable, so don't try to expand it
3435 C->remove_macro_node(n);
3436 continue;
3437 }
3438 if (n->is_Allocate()) {
3439 break;
3440 }
3441 // Make sure expansion will not cause node limit to be exceeded.
3442 // Worst case is a macro node gets expanded into about 200 nodes.
3443 // Allow 50% more for optimization.
3444 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3445 return true;
3446 }
3447
3448 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3449 switch (n->class_id()) {
3450 case Node::Class_Lock:
3451 expand_lock_node(n->as_Lock());
3452 break;
3453 case Node::Class_Unlock:
3454 expand_unlock_node(n->as_Unlock());
3455 break;
3456 case Node::Class_ArrayCopy:
3457 expand_arraycopy_node(n->as_ArrayCopy());
3458 break;
3459 case Node::Class_SubTypeCheck:
3460 expand_subtypecheck_node(n->as_SubTypeCheck());
3461 break;
3462 case Node::Class_CallStaticJava:
3463 expand_mh_intrinsic_return(n->as_CallStaticJava());
3464 C->remove_macro_node(n);
3465 break;
3466 case Node::Class_FlatArrayCheck:
3467 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3468 break;
3469 default:
3470 switch (n->Opcode()) {
3471 case Op_ModD:
3472 case Op_ModF:
3473 case Op_PowD: {
3474 CallLeafPureNode* call_macro = n->as_CallLeafPure();
3475 CallLeafPureNode* call = call_macro->inline_call_leaf_pure_node();
3476 _igvn.replace_node(call_macro, call);
3477 transform_later(call);
3478 break;
3479 }
3480 default:
3481 assert(false, "unknown node type in macro list");
3482 }
3483 }
3484 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3485 if (C->failing()) return true;
3486 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3487
3488 // Clean up the graph so we're less likely to hit the maximum node
3489 // limit
3490 _igvn.set_delay_transform(false);
3491 _igvn.optimize();
3492 if (C->failing()) return true;
3493 _igvn.set_delay_transform(true);
3494 }
3495
3496 // All nodes except Allocate nodes are expanded now. There could be
3497 // new optimization opportunities (such as folding newly created
3498 // load from a just allocated object). Run IGVN.
3499
3500 // expand "macro" nodes
3501 // nodes are removed from the macro list as they are processed
3502 while (C->macro_count() > 0) {
3503 int macro_count = C->macro_count();
3504 Node * n = C->macro_node(macro_count-1);
3505 assert(n->is_macro(), "only macro nodes expected here");
3506 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3507 // node is unreachable, so don't try to expand it
3508 C->remove_macro_node(n);
3509 continue;
3510 }
3511 // Make sure expansion will not cause node limit to be exceeded.
3512 // Worst case is a macro node gets expanded into about 200 nodes.
3513 // Allow 50% more for optimization.
3514 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3515 return true;
3516 }
3517 switch (n->class_id()) {
3518 case Node::Class_Allocate:
3519 expand_allocate(n->as_Allocate());
3520 break;
3521 case Node::Class_AllocateArray:
3522 expand_allocate_array(n->as_AllocateArray());
3523 break;
3524 default:
3525 assert(false, "unknown node type in macro list");
3526 }
3527 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3528 if (C->failing()) return true;
3529 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3530
3531 // Clean up the graph so we're less likely to hit the maximum node
3532 // limit
3533 _igvn.set_delay_transform(false);
3534 _igvn.optimize();
3535 if (C->failing()) return true;
3536 _igvn.set_delay_transform(true);
3537 }
3538
3539 _igvn.set_delay_transform(false);
3540 return false;
3541 }
3542
3543 #ifndef PRODUCT
3544 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3545 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3546 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3547 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3548
3549 void PhaseMacroExpand::print_statistics() {
3550 tty->print("Objects scalar replaced = %d, ", AtomicAccess::load(&_objs_scalar_replaced_counter));
3551 tty->print("Monitor objects removed = %d, ", AtomicAccess::load(&_monitor_objects_removed_counter));
3552 tty->print("GC barriers removed = %d, ", AtomicAccess::load(&_GC_barriers_removed_counter));
3553 tty->print_cr("Memory barriers removed = %d", AtomicAccess::load(&_memory_barriers_removed_counter));
3554 }
3555
3556 int PhaseMacroExpand::count_MemBar(Compile *C) {
3557 if (!PrintOptoStatistics) {
3558 return 0;
3559 }
3560 Unique_Node_List ideal_nodes;
3561 int total = 0;
3562 ideal_nodes.map(C->live_nodes(), nullptr);
3563 ideal_nodes.push(C->root());
3564 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3565 Node* n = ideal_nodes.at(next);
3566 if (n->is_MemBar()) {
3567 total++;
3568 }
3569 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3570 Node* m = n->fast_out(i);
3571 ideal_nodes.push(m);
3572 }
3573 }
3574 return total;
3575 }
3576 #endif