1 /*
2 * Copyright (c) 2005, 2024, 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 "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/rootnode.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "opto/type.hpp"
53 #include "prims/jvmtiExport.hpp"
54 #include "runtime/continuation.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #if INCLUDE_G1GC
59 #include "gc/g1/g1ThreadLocalData.hpp"
60 #endif // INCLUDE_G1GC
61
62
63 //
64 // Replace any references to "oldref" in inputs to "use" with "newref".
65 // Returns the number of replacements made.
66 //
67 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
68 int nreplacements = 0;
69 uint req = use->req();
70 for (uint j = 0; j < use->len(); j++) {
71 Node *uin = use->in(j);
72 if (uin == oldref) {
73 if (j < req)
74 use->set_req(j, newref);
75 else
76 use->set_prec(j, newref);
77 nreplacements++;
78 } else if (j >= req && uin == nullptr) {
79 break;
80 }
81 }
82 return nreplacements;
83 }
84
85 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
86 assert(old != nullptr, "sanity");
87 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
88 Node* use = old->fast_out(i);
89 _igvn.rehash_node_delayed(use);
90 imax -= replace_input(use, old, target);
91 // back up iterator
92 --i;
93 }
94 assert(old->outcnt() == 0, "all uses must be deleted");
95 }
96
97 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
98 Node* cmp;
99 if (mask != 0) {
100 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
101 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
102 } else {
103 cmp = word;
104 }
105 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
106 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
107 transform_later(iff);
108
109 // Fast path taken.
110 Node *fast_taken = transform_later(new IfFalseNode(iff));
111
112 // Fast path not-taken, i.e. slow path
113 Node *slow_taken = transform_later(new IfTrueNode(iff));
114
115 if (return_fast_path) {
116 region->init_req(edge, slow_taken); // Capture slow-control
117 return fast_taken;
118 } else {
119 region->init_req(edge, fast_taken); // Capture fast-control
120 return slow_taken;
121 }
122 }
123
124 //--------------------copy_predefined_input_for_runtime_call--------------------
125 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
126 // Set fixed predefined input arguments
127 call->init_req( TypeFunc::Control, ctrl );
128 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
129 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
130 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
131 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
132 }
133
134 //------------------------------make_slow_call---------------------------------
135 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
136 address slow_call, const char* leaf_name, Node* slow_path,
137 Node* parm0, Node* parm1, Node* parm2) {
138
139 // Slow-path call
140 CallNode *call = leaf_name
141 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
142 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
143
144 // Slow path call has no side-effects, uses few values
145 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
146 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
147 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
148 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
149 call->copy_call_debug_info(&_igvn, oldcall);
150 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
151 _igvn.replace_node(oldcall, call);
152 transform_later(call);
153
154 return call;
155 }
156
157 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
158 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
159 bs->eliminate_gc_barrier(this, p2x);
160 #ifndef PRODUCT
161 if (PrintOptoStatistics) {
162 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
163 }
164 #endif
165 }
166
167 // Search for a memory operation for the specified memory slice.
168 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
169 Node *orig_mem = mem;
170 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
171 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
172 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
173 while (true) {
174 if (mem == alloc_mem || mem == start_mem ) {
175 return mem; // hit one of our sentinels
176 } else if (mem->is_MergeMem()) {
177 mem = mem->as_MergeMem()->memory_at(alias_idx);
178 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
179 Node *in = mem->in(0);
180 // we can safely skip over safepoints, calls, locks and membars because we
181 // already know that the object is safe to eliminate.
182 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
183 return in;
184 } else if (in->is_Call()) {
185 CallNode *call = in->as_Call();
186 if (call->may_modify(tinst, phase)) {
187 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
188 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
189 return in;
190 }
191 }
192 mem = in->in(TypeFunc::Memory);
193 } else if (in->is_MemBar()) {
194 ArrayCopyNode* ac = nullptr;
195 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
196 if (ac != nullptr) {
197 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
198 return ac;
199 }
200 }
201 mem = in->in(TypeFunc::Memory);
202 } else {
203 #ifdef ASSERT
204 in->dump();
205 mem->dump();
206 assert(false, "unexpected projection");
207 #endif
208 }
209 } else if (mem->is_Store()) {
210 const TypePtr* atype = mem->as_Store()->adr_type();
211 int adr_idx = phase->C->get_alias_index(atype);
212 if (adr_idx == alias_idx) {
213 assert(atype->isa_oopptr(), "address type must be oopptr");
214 int adr_offset = atype->offset();
215 uint adr_iid = atype->is_oopptr()->instance_id();
216 // Array elements references have the same alias_idx
217 // but different offset and different instance_id.
218 if (adr_offset == offset && adr_iid == alloc->_idx) {
219 return mem;
220 }
221 } else {
222 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
223 }
224 mem = mem->in(MemNode::Memory);
225 } else if (mem->is_ClearArray()) {
226 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
227 // Can not bypass initialization of the instance
228 // we are looking.
229 debug_only(intptr_t offset;)
230 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
231 InitializeNode* init = alloc->as_Allocate()->initialization();
232 // We are looking for stored value, return Initialize node
233 // or memory edge from Allocate node.
234 if (init != nullptr) {
235 return init;
236 } else {
237 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
238 }
239 }
240 // Otherwise skip it (the call updated 'mem' value).
241 } else if (mem->Opcode() == Op_SCMemProj) {
242 mem = mem->in(0);
243 Node* adr = nullptr;
244 if (mem->is_LoadStore()) {
245 adr = mem->in(MemNode::Address);
246 } else {
247 assert(mem->Opcode() == Op_EncodeISOArray ||
248 mem->Opcode() == Op_StrCompressedCopy, "sanity");
249 adr = mem->in(3); // Destination array
250 }
251 const TypePtr* atype = adr->bottom_type()->is_ptr();
252 int adr_idx = phase->C->get_alias_index(atype);
253 if (adr_idx == alias_idx) {
254 DEBUG_ONLY(mem->dump();)
255 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
256 return nullptr;
257 }
258 mem = mem->in(MemNode::Memory);
259 } else if (mem->Opcode() == Op_StrInflatedCopy) {
260 Node* adr = mem->in(3); // Destination array
261 const TypePtr* atype = adr->bottom_type()->is_ptr();
262 int adr_idx = phase->C->get_alias_index(atype);
263 if (adr_idx == alias_idx) {
264 DEBUG_ONLY(mem->dump();)
265 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
266 return nullptr;
267 }
268 mem = mem->in(MemNode::Memory);
269 } else {
270 return mem;
271 }
272 assert(mem != orig_mem, "dead memory loop");
273 }
274 }
275
276 // Generate loads from source of the arraycopy for fields of
277 // destination needed at a deoptimization point
278 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
279 BasicType bt = ft;
280 const Type *type = ftype;
281 if (ft == T_NARROWOOP) {
282 bt = T_OBJECT;
283 type = ftype->make_oopptr();
284 }
285 Node* res = nullptr;
286 if (ac->is_clonebasic()) {
287 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
288 Node* base = ac->in(ArrayCopyNode::Src);
289 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
290 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
291 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
292 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
293 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
294 } else {
295 if (ac->modifies(offset, offset, &_igvn, true)) {
296 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
297 uint shift = exact_log2(type2aelembytes(bt));
298 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
299 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
300 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
301 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
302
303 Node* adr = nullptr;
304 const TypePtr* adr_type = nullptr;
305 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
306 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
307 Node* base = ac->in(ArrayCopyNode::Src);
308 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
309 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
310 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
311 // Don't emit a new load from src if src == dst but try to get the value from memory instead
312 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
313 }
314 } else {
315 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
316 #ifdef _LP64
317 diff = _igvn.transform(new ConvI2LNode(diff));
318 #endif
319 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
320
321 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
322 Node* base = ac->in(ArrayCopyNode::Src);
323 adr = _igvn.transform(new AddPNode(base, base, off));
324 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
325 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
326 // Non constant offset in the array: we can't statically
327 // determine the value
328 return nullptr;
329 }
330 }
331 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
332 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
333 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
334 }
335 }
336 if (res != nullptr) {
337 if (ftype->isa_narrowoop()) {
338 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
339 res = _igvn.transform(new EncodePNode(res, ftype));
340 }
341 return res;
342 }
343 return nullptr;
344 }
345
346 //
347 // Given a Memory Phi, compute a value Phi containing the values from stores
348 // on the input paths.
349 // Note: this function is recursive, its depth is limited by the "level" argument
350 // Returns the computed Phi, or null if it cannot compute it.
351 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) {
352 assert(mem->is_Phi(), "sanity");
353 int alias_idx = C->get_alias_index(adr_t);
354 int offset = adr_t->offset();
355 int instance_id = adr_t->instance_id();
356
357 // Check if an appropriate value phi already exists.
358 Node* region = mem->in(0);
359 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
360 Node* phi = region->fast_out(k);
361 if (phi->is_Phi() && phi != mem &&
362 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
363 return phi;
364 }
365 }
366 // Check if an appropriate new value phi already exists.
367 Node* new_phi = value_phis->find(mem->_idx);
368 if (new_phi != nullptr)
369 return new_phi;
370
371 if (level <= 0) {
372 return nullptr; // Give up: phi tree too deep
373 }
374 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
375 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
376 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
377
378 uint length = mem->req();
379 GrowableArray <Node *> values(length, length, nullptr);
380
381 // create a new Phi for the value
382 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
383 transform_later(phi);
384 value_phis->push(phi, mem->_idx);
385
386 for (uint j = 1; j < length; j++) {
387 Node *in = mem->in(j);
388 if (in == nullptr || in->is_top()) {
389 values.at_put(j, in);
390 } else {
391 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
392 if (val == start_mem || val == alloc_mem) {
393 // hit a sentinel, return appropriate 0 value
394 values.at_put(j, _igvn.zerocon(ft));
395 continue;
396 }
397 if (val->is_Initialize()) {
398 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
399 }
400 if (val == nullptr) {
401 return nullptr; // can't find a value on this path
402 }
403 if (val == mem) {
404 values.at_put(j, mem);
405 } else if (val->is_Store()) {
406 Node* n = val->in(MemNode::ValueIn);
407 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
408 n = bs->step_over_gc_barrier(n);
409 if (is_subword_type(ft)) {
410 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
411 }
412 values.at_put(j, n);
413 } else if(val->is_Proj() && val->in(0) == alloc) {
414 values.at_put(j, _igvn.zerocon(ft));
415 } else if (val->is_Phi()) {
416 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
417 if (val == nullptr) {
418 return nullptr;
419 }
420 values.at_put(j, val);
421 } else if (val->Opcode() == Op_SCMemProj) {
422 assert(val->in(0)->is_LoadStore() ||
423 val->in(0)->Opcode() == Op_EncodeISOArray ||
424 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
425 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
426 return nullptr;
427 } else if (val->is_ArrayCopy()) {
428 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
429 if (res == nullptr) {
430 return nullptr;
431 }
432 values.at_put(j, res);
433 } else {
434 DEBUG_ONLY( val->dump(); )
435 assert(false, "unknown node on this path");
436 return nullptr; // unknown node on this path
437 }
438 }
439 }
440 // Set Phi's inputs
441 for (uint j = 1; j < length; j++) {
442 if (values.at(j) == mem) {
443 phi->init_req(j, phi);
444 } else {
445 phi->init_req(j, values.at(j));
446 }
447 }
448 return phi;
449 }
450
451 // Search the last value stored into the object's field.
452 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
453 assert(adr_t->is_known_instance_field(), "instance required");
454 int instance_id = adr_t->instance_id();
455 assert((uint)instance_id == alloc->_idx, "wrong allocation");
456
457 int alias_idx = C->get_alias_index(adr_t);
458 int offset = adr_t->offset();
459 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
460 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
461 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
462 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
463 VectorSet visited;
464
465 bool done = sfpt_mem == alloc_mem;
466 Node *mem = sfpt_mem;
467 while (!done) {
468 if (visited.test_set(mem->_idx)) {
469 return nullptr; // found a loop, give up
470 }
471 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
472 if (mem == start_mem || mem == alloc_mem) {
473 done = true; // hit a sentinel, return appropriate 0 value
474 } else if (mem->is_Initialize()) {
475 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
476 if (mem == nullptr) {
477 done = true; // Something go wrong.
478 } else if (mem->is_Store()) {
479 const TypePtr* atype = mem->as_Store()->adr_type();
480 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
481 done = true;
482 }
483 } else if (mem->is_Store()) {
484 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
485 assert(atype != nullptr, "address type must be oopptr");
486 assert(C->get_alias_index(atype) == alias_idx &&
487 atype->is_known_instance_field() && atype->offset() == offset &&
488 atype->instance_id() == instance_id, "store is correct memory slice");
489 done = true;
490 } else if (mem->is_Phi()) {
491 // try to find a phi's unique input
492 Node *unique_input = nullptr;
493 Node *top = C->top();
494 for (uint i = 1; i < mem->req(); i++) {
495 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
496 if (n == nullptr || n == top || n == mem) {
497 continue;
498 } else if (unique_input == nullptr) {
499 unique_input = n;
500 } else if (unique_input != n) {
501 unique_input = top;
502 break;
503 }
504 }
505 if (unique_input != nullptr && unique_input != top) {
506 mem = unique_input;
507 } else {
508 done = true;
509 }
510 } else if (mem->is_ArrayCopy()) {
511 done = true;
512 } else {
513 DEBUG_ONLY( mem->dump(); )
514 assert(false, "unexpected node");
515 }
516 }
517 if (mem != nullptr) {
518 if (mem == start_mem || mem == alloc_mem) {
519 // hit a sentinel, return appropriate 0 value
520 return _igvn.zerocon(ft);
521 } else if (mem->is_Store()) {
522 Node* n = mem->in(MemNode::ValueIn);
523 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
524 n = bs->step_over_gc_barrier(n);
525 return n;
526 } else if (mem->is_Phi()) {
527 // attempt to produce a Phi reflecting the values on the input paths of the Phi
528 Node_Stack value_phis(8);
529 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
530 if (phi != nullptr) {
531 return phi;
532 } else {
533 // Kill all new Phis
534 while(value_phis.is_nonempty()) {
535 Node* n = value_phis.node();
536 _igvn.replace_node(n, C->top());
537 value_phis.pop();
538 }
539 }
540 } else if (mem->is_ArrayCopy()) {
541 Node* ctl = mem->in(0);
542 Node* m = mem->in(TypeFunc::Memory);
543 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
544 // pin the loads in the uncommon trap path
545 ctl = sfpt_ctl;
546 m = sfpt_mem;
547 }
548 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
549 }
550 }
551 // Something go wrong.
552 return nullptr;
553 }
554
555 // Check the possibility of scalar replacement.
556 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
557 // Scan the uses of the allocation to check for anything that would
558 // prevent us from eliminating it.
559 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
560 DEBUG_ONLY( Node* disq_node = nullptr; )
561 bool can_eliminate = true;
562 bool reduce_merge_precheck = (safepoints == nullptr);
563
564 Node* res = alloc->result_cast();
565 const TypeOopPtr* res_type = nullptr;
566 if (res == nullptr) {
567 // All users were eliminated.
568 } else if (!res->is_CheckCastPP()) {
569 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
570 can_eliminate = false;
571 } else {
572 res_type = igvn->type(res)->isa_oopptr();
573 if (res_type == nullptr) {
574 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
575 can_eliminate = false;
576 } else if (!res_type->klass_is_exact()) {
577 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
578 can_eliminate = false;
579 } else if (res_type->isa_aryptr()) {
580 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
581 if (length < 0) {
582 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
583 can_eliminate = false;
584 }
585 }
586 }
587
588 if (can_eliminate && res != nullptr) {
589 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
590 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
591 j < jmax && can_eliminate; j++) {
592 Node* use = res->fast_out(j);
593
594 if (use->is_AddP()) {
595 const TypePtr* addp_type = igvn->type(use)->is_ptr();
596 int offset = addp_type->offset();
597
598 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
599 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
600 can_eliminate = false;
601 break;
602 }
603 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
604 k < kmax && can_eliminate; k++) {
605 Node* n = use->fast_out(k);
606 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
607 DEBUG_ONLY(disq_node = n;)
608 if (n->is_Load() || n->is_LoadStore()) {
609 NOT_PRODUCT(fail_eliminate = "Field load";)
610 } else {
611 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
612 }
613 can_eliminate = false;
614 }
615 }
616 } else if (use->is_ArrayCopy() &&
617 (use->as_ArrayCopy()->is_clonebasic() ||
618 use->as_ArrayCopy()->is_arraycopy_validated() ||
619 use->as_ArrayCopy()->is_copyof_validated() ||
620 use->as_ArrayCopy()->is_copyofrange_validated()) &&
621 use->in(ArrayCopyNode::Dest) == res) {
622 // ok to eliminate
623 } else if (use->is_SafePoint()) {
624 SafePointNode* sfpt = use->as_SafePoint();
625 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
626 // Object is passed as argument.
627 DEBUG_ONLY(disq_node = use;)
628 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
629 can_eliminate = false;
630 }
631 Node* sfptMem = sfpt->memory();
632 if (sfptMem == nullptr || sfptMem->is_top()) {
633 DEBUG_ONLY(disq_node = use;)
634 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
635 can_eliminate = false;
636 } else if (!reduce_merge_precheck) {
637 safepoints->append_if_missing(sfpt);
638 }
639 } else if (reduce_merge_precheck &&
640 (use->is_Phi() || use->is_EncodeP() ||
641 use->Opcode() == Op_MemBarRelease ||
642 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
643 // Nothing to do
644 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
645 if (use->is_Phi()) {
646 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
647 NOT_PRODUCT(fail_eliminate = "Object is return value";)
648 } else {
649 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
650 }
651 DEBUG_ONLY(disq_node = use;)
652 } else {
653 if (use->Opcode() == Op_Return) {
654 NOT_PRODUCT(fail_eliminate = "Object is return value";)
655 } else {
656 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
657 }
658 DEBUG_ONLY(disq_node = use;)
659 }
660 can_eliminate = false;
661 }
662 }
663 }
664
665 #ifndef PRODUCT
666 if (PrintEliminateAllocations && safepoints != nullptr) {
667 if (can_eliminate) {
668 tty->print("Scalar ");
669 if (res == nullptr)
670 alloc->dump();
671 else
672 res->dump();
673 } else if (alloc->_is_scalar_replaceable) {
674 tty->print("NotScalar (%s)", fail_eliminate);
675 if (res == nullptr)
676 alloc->dump();
677 else
678 res->dump();
679 #ifdef ASSERT
680 if (disq_node != nullptr) {
681 tty->print(" >>>> ");
682 disq_node->dump();
683 }
684 #endif /*ASSERT*/
685 }
686 }
687
688 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
689 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
690 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
691 }
692 #endif
693 return can_eliminate;
694 }
695
696 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
697 Node* res = alloc->result_cast();
698 int nfields = 0;
699 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
700
701 if (res != nullptr) {
702 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
703
704 if (res_type->isa_instptr()) {
705 // find the fields of the class which will be needed for safepoint debug information
706 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
707 nfields = iklass->nof_nonstatic_fields();
708 } else {
709 // find the array's elements which will be needed for safepoint debug information
710 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
711 assert(nfields >= 0, "must be an array klass.");
712 }
713 }
714
715 // rollback processed safepoints
716 while (safepoints_done.length() > 0) {
717 SafePointNode* sfpt_done = safepoints_done.pop();
718 // remove any extra entries we added to the safepoint
719 uint last = sfpt_done->req() - 1;
720 for (int k = 0; k < nfields; k++) {
721 sfpt_done->del_req(last--);
722 }
723 JVMState *jvms = sfpt_done->jvms();
724 jvms->set_endoff(sfpt_done->req());
725 // Now make a pass over the debug information replacing any references
726 // to SafePointScalarObjectNode with the allocated object.
727 int start = jvms->debug_start();
728 int end = jvms->debug_end();
729 for (int i = start; i < end; i++) {
730 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
731 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
732 if (scobj->first_index(jvms) == sfpt_done->req() &&
733 scobj->n_fields() == (uint)nfields) {
734 assert(scobj->alloc() == alloc, "sanity");
735 sfpt_done->set_req(i, res);
736 }
737 }
738 }
739 _igvn._worklist.push(sfpt_done);
740 }
741 }
742
743 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {
744 // Fields of scalar objs are referenced only at the end
745 // of regular debuginfo at the last (youngest) JVMS.
746 // Record relative start index.
747 ciInstanceKlass* iklass = nullptr;
748 BasicType basic_elem_type = T_ILLEGAL;
749 const Type* field_type = nullptr;
750 const TypeOopPtr* res_type = nullptr;
751 int nfields = 0;
752 int array_base = 0;
753 int element_size = 0;
754 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
755 Node* res = alloc->result_cast();
756
757 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
758 assert(sfpt->jvms() != nullptr, "missed JVMS");
759
760 if (res != nullptr) { // Could be null when there are no users
761 res_type = _igvn.type(res)->isa_oopptr();
762
763 if (res_type->isa_instptr()) {
764 // find the fields of the class which will be needed for safepoint debug information
765 iklass = res_type->is_instptr()->instance_klass();
766 nfields = iklass->nof_nonstatic_fields();
767 } else {
768 // find the array's elements which will be needed for safepoint debug information
769 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
770 assert(nfields >= 0, "must be an array klass.");
771 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
772 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
773 element_size = type2aelembytes(basic_elem_type);
774 field_type = res_type->is_aryptr()->elem();
775 }
776 }
777
778 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
779 sobj->init_req(0, C->root());
780 transform_later(sobj);
781
782 // Scan object's fields adding an input to the safepoint for each field.
783 for (int j = 0; j < nfields; j++) {
784 intptr_t offset;
785 ciField* field = nullptr;
786 if (iklass != nullptr) {
787 field = iklass->nonstatic_field_at(j);
788 offset = field->offset_in_bytes();
789 ciType* elem_type = field->type();
790 basic_elem_type = field->layout_type();
791
792 // The next code is taken from Parse::do_get_xxx().
793 if (is_reference_type(basic_elem_type)) {
794 if (!elem_type->is_loaded()) {
795 field_type = TypeInstPtr::BOTTOM;
796 } else if (field != nullptr && field->is_static_constant()) {
797 ciObject* con = field->constant_value().as_object();
798 // Do not "join" in the previous type; it doesn't add value,
799 // and may yield a vacuous result if the field is of interface type.
800 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
801 assert(field_type != nullptr, "field singleton type must be consistent");
802 } else {
803 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
804 }
805 if (UseCompressedOops) {
806 field_type = field_type->make_narrowoop();
807 basic_elem_type = T_NARROWOOP;
808 }
809 } else {
810 field_type = Type::get_const_basic_type(basic_elem_type);
811 }
812 } else {
813 offset = array_base + j * (intptr_t)element_size;
814 }
815
816 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
817
818 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
819
820 // We weren't able to find a value for this field,
821 // give up on eliminating this allocation.
822 if (field_val == nullptr) {
823 uint last = sfpt->req() - 1;
824 for (int k = 0; k < j; k++) {
825 sfpt->del_req(last--);
826 }
827 _igvn._worklist.push(sfpt);
828
829 #ifndef PRODUCT
830 if (PrintEliminateAllocations) {
831 if (field != nullptr) {
832 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
833 field->print();
834 int field_idx = C->get_alias_index(field_addr_type);
835 tty->print(" (alias_idx=%d)", field_idx);
836 } else { // Array's element
837 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
838 }
839 tty->print(", which prevents elimination of: ");
840 if (res == nullptr)
841 alloc->dump();
842 else
843 res->dump();
844 }
845 #endif
846
847 return nullptr;
848 }
849
850 if (UseCompressedOops && field_type->isa_narrowoop()) {
851 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
852 // to be able scalar replace the allocation.
853 if (field_val->is_EncodeP()) {
854 field_val = field_val->in(1);
855 } else {
856 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
857 }
858 }
859 sfpt->add_req(field_val);
860 }
861
862 sfpt->jvms()->set_endoff(sfpt->req());
863
864 return sobj;
865 }
866
867 // Do scalar replacement.
868 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
869 GrowableArray <SafePointNode *> safepoints_done;
870 Node* res = alloc->result_cast();
871 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
872
873 // Process the safepoint uses
874 while (safepoints.length() > 0) {
875 SafePointNode* sfpt = safepoints.pop();
876 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
877
878 if (sobj == nullptr) {
879 undo_previous_scalarizations(safepoints_done, alloc);
880 return false;
881 }
882
883 // Now make a pass over the debug information replacing any references
884 // to the allocated object with "sobj"
885 JVMState *jvms = sfpt->jvms();
886 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
887 _igvn._worklist.push(sfpt);
888
889 // keep it for rollback
890 safepoints_done.append_if_missing(sfpt);
891 }
892
893 return true;
894 }
895
896 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
897 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
898 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
899 if (ctl_proj != nullptr) {
900 igvn.replace_node(ctl_proj, n->in(0));
901 }
902 if (mem_proj != nullptr) {
903 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
904 }
905 }
906
907 // Process users of eliminated allocation.
908 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
909 Node* res = alloc->result_cast();
910 if (res != nullptr) {
911 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
912 Node *use = res->last_out(j);
913 uint oc1 = res->outcnt();
914
915 if (use->is_AddP()) {
916 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
917 Node *n = use->last_out(k);
918 uint oc2 = use->outcnt();
919 if (n->is_Store()) {
920 #ifdef ASSERT
921 // Verify that there is no dependent MemBarVolatile nodes,
922 // they should be removed during IGVN, see MemBarNode::Ideal().
923 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
924 p < pmax; p++) {
925 Node* mb = n->fast_out(p);
926 assert(mb->is_Initialize() || !mb->is_MemBar() ||
927 mb->req() <= MemBarNode::Precedent ||
928 mb->in(MemBarNode::Precedent) != n,
929 "MemBarVolatile should be eliminated for non-escaping object");
930 }
931 #endif
932 _igvn.replace_node(n, n->in(MemNode::Memory));
933 } else {
934 eliminate_gc_barrier(n);
935 }
936 k -= (oc2 - use->outcnt());
937 }
938 _igvn.remove_dead_node(use);
939 } else if (use->is_ArrayCopy()) {
940 // Disconnect ArrayCopy node
941 ArrayCopyNode* ac = use->as_ArrayCopy();
942 if (ac->is_clonebasic()) {
943 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
944 disconnect_projections(ac, _igvn);
945 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
946 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
947 disconnect_projections(membar_before->as_MemBar(), _igvn);
948 if (membar_after->is_MemBar()) {
949 disconnect_projections(membar_after->as_MemBar(), _igvn);
950 }
951 } else {
952 assert(ac->is_arraycopy_validated() ||
953 ac->is_copyof_validated() ||
954 ac->is_copyofrange_validated(), "unsupported");
955 CallProjections callprojs;
956 ac->extract_projections(&callprojs, true);
957
958 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
959 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
960 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
961
962 // Set control to top. IGVN will remove the remaining projections
963 ac->set_req(0, top());
964 ac->replace_edge(res, top(), &_igvn);
965
966 // Disconnect src right away: it can help find new
967 // opportunities for allocation elimination
968 Node* src = ac->in(ArrayCopyNode::Src);
969 ac->replace_edge(src, top(), &_igvn);
970 // src can be top at this point if src and dest of the
971 // arraycopy were the same
972 if (src->outcnt() == 0 && !src->is_top()) {
973 _igvn.remove_dead_node(src);
974 }
975 }
976 _igvn._worklist.push(ac);
977 } else {
978 eliminate_gc_barrier(use);
979 }
980 j -= (oc1 - res->outcnt());
981 }
982 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
983 _igvn.remove_dead_node(res);
984 }
985
986 //
987 // Process other users of allocation's projections
988 //
989 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
990 // First disconnect stores captured by Initialize node.
991 // If Initialize node is eliminated first in the following code,
992 // it will kill such stores and DUIterator_Last will assert.
993 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
994 Node* use = _callprojs.resproj->fast_out(j);
995 if (use->is_AddP()) {
996 // raw memory addresses used only by the initialization
997 _igvn.replace_node(use, C->top());
998 --j; --jmax;
999 }
1000 }
1001 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
1002 Node* use = _callprojs.resproj->last_out(j);
1003 uint oc1 = _callprojs.resproj->outcnt();
1004 if (use->is_Initialize()) {
1005 // Eliminate Initialize node.
1006 InitializeNode *init = use->as_Initialize();
1007 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1008 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1009 if (ctrl_proj != nullptr) {
1010 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1011 #ifdef ASSERT
1012 // If the InitializeNode has no memory out, it will die, and tmp will become null
1013 Node* tmp = init->in(TypeFunc::Control);
1014 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1015 #endif
1016 }
1017 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1018 if (mem_proj != nullptr) {
1019 Node *mem = init->in(TypeFunc::Memory);
1020 #ifdef ASSERT
1021 if (mem->is_MergeMem()) {
1022 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
1023 } else {
1024 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1025 }
1026 #endif
1027 _igvn.replace_node(mem_proj, mem);
1028 }
1029 } else {
1030 assert(false, "only Initialize or AddP expected");
1031 }
1032 j -= (oc1 - _callprojs.resproj->outcnt());
1033 }
1034 }
1035 if (_callprojs.fallthrough_catchproj != nullptr) {
1036 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1037 }
1038 if (_callprojs.fallthrough_memproj != nullptr) {
1039 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1040 }
1041 if (_callprojs.catchall_memproj != nullptr) {
1042 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1043 }
1044 if (_callprojs.fallthrough_ioproj != nullptr) {
1045 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1046 }
1047 if (_callprojs.catchall_ioproj != nullptr) {
1048 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1049 }
1050 if (_callprojs.catchall_catchproj != nullptr) {
1051 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1052 }
1053 }
1054
1055 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1056 // If reallocation fails during deoptimization we'll pop all
1057 // interpreter frames for this compiled frame and that won't play
1058 // nice with JVMTI popframe.
1059 // We avoid this issue by eager reallocation when the popframe request
1060 // is received.
1061 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1062 return false;
1063 }
1064 Node* klass = alloc->in(AllocateNode::KlassNode);
1065 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1066 Node* res = alloc->result_cast();
1067 // Eliminate boxing allocations which are not used
1068 // regardless scalar replaceable status.
1069 bool boxing_alloc = C->eliminate_boxing() &&
1070 tklass->isa_instklassptr() &&
1071 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1072 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1073 return false;
1074 }
1075
1076 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1077
1078 GrowableArray <SafePointNode *> safepoints;
1079 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1080 return false;
1081 }
1082
1083 if (!alloc->_is_scalar_replaceable) {
1084 assert(res == nullptr, "sanity");
1085 // We can only eliminate allocation if all debug info references
1086 // are already replaced with SafePointScalarObject because
1087 // we can't search for a fields value without instance_id.
1088 if (safepoints.length() > 0) {
1089 return false;
1090 }
1091 }
1092
1093 if (!scalar_replacement(alloc, safepoints)) {
1094 return false;
1095 }
1096
1097 CompileLog* log = C->log();
1098 if (log != nullptr) {
1099 log->head("eliminate_allocation type='%d'",
1100 log->identify(tklass->exact_klass()));
1101 JVMState* p = alloc->jvms();
1102 while (p != nullptr) {
1103 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1104 p = p->caller();
1105 }
1106 log->tail("eliminate_allocation");
1107 }
1108
1109 process_users_of_allocation(alloc);
1110
1111 #ifndef PRODUCT
1112 if (PrintEliminateAllocations) {
1113 if (alloc->is_AllocateArray())
1114 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1115 else
1116 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1117 }
1118 #endif
1119
1120 return true;
1121 }
1122
1123 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1124 // EA should remove all uses of non-escaping boxing node.
1125 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1126 return false;
1127 }
1128
1129 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1130
1131 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1132
1133 const TypeTuple* r = boxing->tf()->range();
1134 assert(r->cnt() > TypeFunc::Parms, "sanity");
1135 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1136 assert(t != nullptr, "sanity");
1137
1138 CompileLog* log = C->log();
1139 if (log != nullptr) {
1140 log->head("eliminate_boxing type='%d'",
1141 log->identify(t->instance_klass()));
1142 JVMState* p = boxing->jvms();
1143 while (p != nullptr) {
1144 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1145 p = p->caller();
1146 }
1147 log->tail("eliminate_boxing");
1148 }
1149
1150 process_users_of_allocation(boxing);
1151
1152 #ifndef PRODUCT
1153 if (PrintEliminateAllocations) {
1154 tty->print("++++ Eliminated: %d ", boxing->_idx);
1155 boxing->method()->print_short_name(tty);
1156 tty->cr();
1157 }
1158 #endif
1159
1160 return true;
1161 }
1162
1163
1164 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1165 Node* adr = basic_plus_adr(base, offset);
1166 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1167 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1168 transform_later(value);
1169 return value;
1170 }
1171
1172
1173 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1174 Node* adr = basic_plus_adr(base, offset);
1175 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1176 transform_later(mem);
1177 return mem;
1178 }
1179
1180 //=============================================================================
1181 //
1182 // A L L O C A T I O N
1183 //
1184 // Allocation attempts to be fast in the case of frequent small objects.
1185 // It breaks down like this:
1186 //
1187 // 1) Size in doublewords is computed. This is a constant for objects and
1188 // variable for most arrays. Doubleword units are used to avoid size
1189 // overflow of huge doubleword arrays. We need doublewords in the end for
1190 // rounding.
1191 //
1192 // 2) Size is checked for being 'too large'. Too-large allocations will go
1193 // the slow path into the VM. The slow path can throw any required
1194 // exceptions, and does all the special checks for very large arrays. The
1195 // size test can constant-fold away for objects. For objects with
1196 // finalizers it constant-folds the otherway: you always go slow with
1197 // finalizers.
1198 //
1199 // 3) If NOT using TLABs, this is the contended loop-back point.
1200 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1201 //
1202 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1203 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1204 // "size*8" we always enter the VM, where "largish" is a constant picked small
1205 // enough that there's always space between the eden max and 4Gig (old space is
1206 // there so it's quite large) and large enough that the cost of entering the VM
1207 // is dwarfed by the cost to initialize the space.
1208 //
1209 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1210 // down. If contended, repeat at step 3. If using TLABs normal-store
1211 // adjusted heap top back down; there is no contention.
1212 //
1213 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1214 // fields.
1215 //
1216 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1217 // oop flavor.
1218 //
1219 //=============================================================================
1220 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1221 // Allocations bigger than this always go the slow route.
1222 // This value must be small enough that allocation attempts that need to
1223 // trigger exceptions go the slow route. Also, it must be small enough so
1224 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1225 //=============================================================================j//
1226 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1227 // The allocator will coalesce int->oop copies away. See comment in
1228 // coalesce.cpp about how this works. It depends critically on the exact
1229 // code shape produced here, so if you are changing this code shape
1230 // make sure the GC info for the heap-top is correct in and around the
1231 // slow-path call.
1232 //
1233
1234 void PhaseMacroExpand::expand_allocate_common(
1235 AllocateNode* alloc, // allocation node to be expanded
1236 Node* length, // array length for an array allocation
1237 const TypeFunc* slow_call_type, // Type of slow call
1238 address slow_call_address, // Address of slow call
1239 Node* valid_length_test // whether length is valid or not
1240 )
1241 {
1242 Node* ctrl = alloc->in(TypeFunc::Control);
1243 Node* mem = alloc->in(TypeFunc::Memory);
1244 Node* i_o = alloc->in(TypeFunc::I_O);
1245 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1246 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1247 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1248 assert(ctrl != nullptr, "must have control");
1249
1250 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1251 // they will not be used if "always_slow" is set
1252 enum { slow_result_path = 1, fast_result_path = 2 };
1253 Node *result_region = nullptr;
1254 Node *result_phi_rawmem = nullptr;
1255 Node *result_phi_rawoop = nullptr;
1256 Node *result_phi_i_o = nullptr;
1257
1258 // The initial slow comparison is a size check, the comparison
1259 // we want to do is a BoolTest::gt
1260 bool expand_fast_path = true;
1261 int tv = _igvn.find_int_con(initial_slow_test, -1);
1262 if (tv >= 0) {
1263 // InitialTest has constant result
1264 // 0 - can fit in TLAB
1265 // 1 - always too big or negative
1266 assert(tv <= 1, "0 or 1 if a constant");
1267 expand_fast_path = (tv == 0);
1268 initial_slow_test = nullptr;
1269 } else {
1270 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1271 }
1272
1273 if (!UseTLAB) {
1274 // Force slow-path allocation
1275 expand_fast_path = false;
1276 initial_slow_test = nullptr;
1277 }
1278
1279 bool allocation_has_use = (alloc->result_cast() != nullptr);
1280 if (!allocation_has_use) {
1281 InitializeNode* init = alloc->initialization();
1282 if (init != nullptr) {
1283 init->remove(&_igvn);
1284 }
1285 if (expand_fast_path && (initial_slow_test == nullptr)) {
1286 // Remove allocation node and return.
1287 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1288 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1289 #ifndef PRODUCT
1290 if (PrintEliminateAllocations) {
1291 tty->print("NotUsed ");
1292 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1293 if (res != nullptr) {
1294 res->dump();
1295 } else {
1296 alloc->dump();
1297 }
1298 }
1299 #endif
1300 yank_alloc_node(alloc);
1301 return;
1302 }
1303 }
1304
1305 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1306 Node *slow_region = nullptr;
1307 Node *toobig_false = ctrl;
1308
1309 // generate the initial test if necessary
1310 if (initial_slow_test != nullptr ) {
1311 assert (expand_fast_path, "Only need test if there is a fast path");
1312 slow_region = new RegionNode(3);
1313
1314 // Now make the initial failure test. Usually a too-big test but
1315 // might be a TRUE for finalizers.
1316 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1317 transform_later(toobig_iff);
1318 // Plug the failing-too-big test into the slow-path region
1319 Node *toobig_true = new IfTrueNode( toobig_iff );
1320 transform_later(toobig_true);
1321 slow_region ->init_req( too_big_or_final_path, toobig_true );
1322 toobig_false = new IfFalseNode( toobig_iff );
1323 transform_later(toobig_false);
1324 } else {
1325 // No initial test, just fall into next case
1326 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1327 toobig_false = ctrl;
1328 debug_only(slow_region = NodeSentinel);
1329 }
1330
1331 // If we are here there are several possibilities
1332 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1333 // no_initial_check means a constant allocation.
1334 // - If check always evaluates to false -> expand_fast_path is false (see above)
1335 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1336 // if !allocation_has_use the fast path is empty
1337 // if !allocation_has_use && no_initial_check
1338 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1339 // removed by yank_alloc_node above.
1340
1341 Node *slow_mem = mem; // save the current memory state for slow path
1342 // generate the fast allocation code unless we know that the initial test will always go slow
1343 if (expand_fast_path) {
1344 // Fast path modifies only raw memory.
1345 if (mem->is_MergeMem()) {
1346 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1347 }
1348
1349 // allocate the Region and Phi nodes for the result
1350 result_region = new RegionNode(3);
1351 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1352 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1353
1354 // Grab regular I/O before optional prefetch may change it.
1355 // Slow-path does no I/O so just set it to the original I/O.
1356 result_phi_i_o->init_req(slow_result_path, i_o);
1357
1358 // Name successful fast-path variables
1359 Node* fast_oop_ctrl;
1360 Node* fast_oop_rawmem;
1361 if (allocation_has_use) {
1362 Node* needgc_ctrl = nullptr;
1363 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1364
1365 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1366 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1367 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1368 fast_oop_ctrl, fast_oop_rawmem,
1369 prefetch_lines);
1370
1371 if (initial_slow_test != nullptr) {
1372 // This completes all paths into the slow merge point
1373 slow_region->init_req(need_gc_path, needgc_ctrl);
1374 transform_later(slow_region);
1375 } else {
1376 // No initial slow path needed!
1377 // Just fall from the need-GC path straight into the VM call.
1378 slow_region = needgc_ctrl;
1379 }
1380
1381 InitializeNode* init = alloc->initialization();
1382 fast_oop_rawmem = initialize_object(alloc,
1383 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1384 klass_node, length, size_in_bytes);
1385 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1386 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1387
1388 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1389 } else {
1390 assert (initial_slow_test != nullptr, "sanity");
1391 fast_oop_ctrl = toobig_false;
1392 fast_oop_rawmem = mem;
1393 transform_later(slow_region);
1394 }
1395
1396 // Plug in the successful fast-path into the result merge point
1397 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1398 result_phi_i_o ->init_req(fast_result_path, i_o);
1399 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1400 } else {
1401 slow_region = ctrl;
1402 result_phi_i_o = i_o; // Rename it to use in the following code.
1403 }
1404
1405 // Generate slow-path call
1406 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1407 OptoRuntime::stub_name(slow_call_address),
1408 TypePtr::BOTTOM);
1409 call->init_req(TypeFunc::Control, slow_region);
1410 call->init_req(TypeFunc::I_O, top()); // does no i/o
1411 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1412 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1413 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1414
1415 call->init_req(TypeFunc::Parms+0, klass_node);
1416 if (length != nullptr) {
1417 call->init_req(TypeFunc::Parms+1, length);
1418 }
1419
1420 // Copy debug information and adjust JVMState information, then replace
1421 // allocate node with the call
1422 call->copy_call_debug_info(&_igvn, alloc);
1423 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1424 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1425 // path dies).
1426 if (valid_length_test != nullptr) {
1427 call->add_req(valid_length_test);
1428 }
1429 if (expand_fast_path) {
1430 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1431 } else {
1432 // Hook i_o projection to avoid its elimination during allocation
1433 // replacement (when only a slow call is generated).
1434 call->set_req(TypeFunc::I_O, result_phi_i_o);
1435 }
1436 _igvn.replace_node(alloc, call);
1437 transform_later(call);
1438
1439 // Identify the output projections from the allocate node and
1440 // adjust any references to them.
1441 // The control and io projections look like:
1442 //
1443 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1444 // Allocate Catch
1445 // ^---Proj(io) <-------+ ^---CatchProj(io)
1446 //
1447 // We are interested in the CatchProj nodes.
1448 //
1449 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1450
1451 // An allocate node has separate memory projections for the uses on
1452 // the control and i_o paths. Replace the control memory projection with
1453 // result_phi_rawmem (unless we are only generating a slow call when
1454 // both memory projections are combined)
1455 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1456 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1457 }
1458 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1459 // catchall_memproj so we end up with a call that has only 1 memory projection.
1460 if (_callprojs.catchall_memproj != nullptr ) {
1461 if (_callprojs.fallthrough_memproj == nullptr) {
1462 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1463 transform_later(_callprojs.fallthrough_memproj);
1464 }
1465 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1466 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1467 }
1468
1469 // An allocate node has separate i_o projections for the uses on the control
1470 // and i_o paths. Always replace the control i_o projection with result i_o
1471 // otherwise incoming i_o become dead when only a slow call is generated
1472 // (it is different from memory projections where both projections are
1473 // combined in such case).
1474 if (_callprojs.fallthrough_ioproj != nullptr) {
1475 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1476 }
1477 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1478 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1479 if (_callprojs.catchall_ioproj != nullptr ) {
1480 if (_callprojs.fallthrough_ioproj == nullptr) {
1481 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1482 transform_later(_callprojs.fallthrough_ioproj);
1483 }
1484 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1485 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1486 }
1487
1488 // if we generated only a slow call, we are done
1489 if (!expand_fast_path) {
1490 // Now we can unhook i_o.
1491 if (result_phi_i_o->outcnt() > 1) {
1492 call->set_req(TypeFunc::I_O, top());
1493 } else {
1494 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1495 // Case of new array with negative size known during compilation.
1496 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1497 // following code since call to runtime will throw exception.
1498 // As result there will be no users of i_o after the call.
1499 // Leave i_o attached to this call to avoid problems in preceding graph.
1500 }
1501 return;
1502 }
1503
1504 if (_callprojs.fallthrough_catchproj != nullptr) {
1505 ctrl = _callprojs.fallthrough_catchproj->clone();
1506 transform_later(ctrl);
1507 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1508 } else {
1509 ctrl = top();
1510 }
1511 Node *slow_result;
1512 if (_callprojs.resproj == nullptr) {
1513 // no uses of the allocation result
1514 slow_result = top();
1515 } else {
1516 slow_result = _callprojs.resproj->clone();
1517 transform_later(slow_result);
1518 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1519 }
1520
1521 // Plug slow-path into result merge point
1522 result_region->init_req( slow_result_path, ctrl);
1523 transform_later(result_region);
1524 if (allocation_has_use) {
1525 result_phi_rawoop->init_req(slow_result_path, slow_result);
1526 transform_later(result_phi_rawoop);
1527 }
1528 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1529 transform_later(result_phi_rawmem);
1530 transform_later(result_phi_i_o);
1531 // This completes all paths into the result merge point
1532 }
1533
1534 // Remove alloc node that has no uses.
1535 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1536 Node* ctrl = alloc->in(TypeFunc::Control);
1537 Node* mem = alloc->in(TypeFunc::Memory);
1538 Node* i_o = alloc->in(TypeFunc::I_O);
1539
1540 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1541 if (_callprojs.resproj != nullptr) {
1542 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1543 Node* use = _callprojs.resproj->fast_out(i);
1544 use->isa_MemBar()->remove(&_igvn);
1545 --imax;
1546 --i; // back up iterator
1547 }
1548 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1549 _igvn.remove_dead_node(_callprojs.resproj);
1550 }
1551 if (_callprojs.fallthrough_catchproj != nullptr) {
1552 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1553 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1554 }
1555 if (_callprojs.catchall_catchproj != nullptr) {
1556 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1557 _callprojs.catchall_catchproj->set_req(0, top());
1558 }
1559 if (_callprojs.fallthrough_proj != nullptr) {
1560 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1561 _igvn.remove_dead_node(catchnode);
1562 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1563 }
1564 if (_callprojs.fallthrough_memproj != nullptr) {
1565 migrate_outs(_callprojs.fallthrough_memproj, mem);
1566 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1567 }
1568 if (_callprojs.fallthrough_ioproj != nullptr) {
1569 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1570 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1571 }
1572 if (_callprojs.catchall_memproj != nullptr) {
1573 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1574 _callprojs.catchall_memproj->set_req(0, top());
1575 }
1576 if (_callprojs.catchall_ioproj != nullptr) {
1577 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1578 _callprojs.catchall_ioproj->set_req(0, top());
1579 }
1580 #ifndef PRODUCT
1581 if (PrintEliminateAllocations) {
1582 if (alloc->is_AllocateArray()) {
1583 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1584 } else {
1585 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1586 }
1587 }
1588 #endif
1589 _igvn.remove_dead_node(alloc);
1590 }
1591
1592 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1593 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1594 // If initialization is performed by an array copy, any required
1595 // MemBarStoreStore was already added. If the object does not
1596 // escape no need for a MemBarStoreStore. If the object does not
1597 // escape in its initializer and memory barrier (MemBarStoreStore or
1598 // stronger) is already added at exit of initializer, also no need
1599 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1600 // so that stores that initialize this object can't be reordered
1601 // with a subsequent store that makes this object accessible by
1602 // other threads.
1603 // Other threads include java threads and JVM internal threads
1604 // (for example concurrent GC threads). Current concurrent GC
1605 // implementation: G1 will not scan newly created object,
1606 // so it's safe to skip storestore barrier when allocation does
1607 // not escape.
1608 if (!alloc->does_not_escape_thread() &&
1609 !alloc->is_allocation_MemBar_redundant() &&
1610 (init == nullptr || !init->is_complete_with_arraycopy())) {
1611 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1612 // No InitializeNode or no stores captured by zeroing
1613 // elimination. Simply add the MemBarStoreStore after object
1614 // initialization.
1615 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1616 transform_later(mb);
1617
1618 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1619 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1620 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1621 transform_later(fast_oop_ctrl);
1622 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1623 transform_later(fast_oop_rawmem);
1624 } else {
1625 // Add the MemBarStoreStore after the InitializeNode so that
1626 // all stores performing the initialization that were moved
1627 // before the InitializeNode happen before the storestore
1628 // barrier.
1629
1630 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1631 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1632
1633 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1634 transform_later(mb);
1635
1636 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1637 transform_later(ctrl);
1638 Node* mem = new ProjNode(init, TypeFunc::Memory);
1639 transform_later(mem);
1640
1641 // The MemBarStoreStore depends on control and memory coming
1642 // from the InitializeNode
1643 mb->init_req(TypeFunc::Memory, mem);
1644 mb->init_req(TypeFunc::Control, ctrl);
1645
1646 ctrl = new ProjNode(mb, TypeFunc::Control);
1647 transform_later(ctrl);
1648 mem = new ProjNode(mb, TypeFunc::Memory);
1649 transform_later(mem);
1650
1651 // All nodes that depended on the InitializeNode for control
1652 // and memory must now depend on the MemBarNode that itself
1653 // depends on the InitializeNode
1654 if (init_ctrl != nullptr) {
1655 _igvn.replace_node(init_ctrl, ctrl);
1656 }
1657 if (init_mem != nullptr) {
1658 _igvn.replace_node(init_mem, mem);
1659 }
1660 }
1661 }
1662 }
1663
1664 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1665 Node*& ctrl, Node*& rawmem) {
1666 if (C->env()->dtrace_alloc_probes()) {
1667 // Slow-path call
1668 int size = TypeFunc::Parms + 2;
1669 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1670 CAST_FROM_FN_PTR(address,
1671 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1672 "dtrace_object_alloc",
1673 TypeRawPtr::BOTTOM);
1674
1675 // Get base of thread-local storage area
1676 Node* thread = new ThreadLocalNode();
1677 transform_later(thread);
1678
1679 call->init_req(TypeFunc::Parms + 0, thread);
1680 call->init_req(TypeFunc::Parms + 1, oop);
1681 call->init_req(TypeFunc::Control, ctrl);
1682 call->init_req(TypeFunc::I_O , top()); // does no i/o
1683 call->init_req(TypeFunc::Memory , rawmem);
1684 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1685 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1686 transform_later(call);
1687 ctrl = new ProjNode(call, TypeFunc::Control);
1688 transform_later(ctrl);
1689 rawmem = new ProjNode(call, TypeFunc::Memory);
1690 transform_later(rawmem);
1691 }
1692 }
1693
1694 // Helper for PhaseMacroExpand::expand_allocate_common.
1695 // Initializes the newly-allocated storage.
1696 Node*
1697 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1698 Node* control, Node* rawmem, Node* object,
1699 Node* klass_node, Node* length,
1700 Node* size_in_bytes) {
1701 InitializeNode* init = alloc->initialization();
1702 // Store the klass & mark bits
1703 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1704 if (!mark_node->is_Con()) {
1705 transform_later(mark_node);
1706 }
1707 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1708
1709 if (!UseCompactObjectHeaders) {
1710 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1711 }
1712 int header_size = alloc->minimum_header_size(); // conservatively small
1713
1714 // Array length
1715 if (length != nullptr) { // Arrays need length field
1716 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1717 // conservatively small header size:
1718 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1719 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1720 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1721 if (is_reference_type(elem, true)) {
1722 elem = T_OBJECT;
1723 }
1724 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1725 }
1726 }
1727
1728 // Clear the object body, if necessary.
1729 if (init == nullptr) {
1730 // The init has somehow disappeared; be cautious and clear everything.
1731 //
1732 // This can happen if a node is allocated but an uncommon trap occurs
1733 // immediately. In this case, the Initialize gets associated with the
1734 // trap, and may be placed in a different (outer) loop, if the Allocate
1735 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1736 // there can be two Allocates to one Initialize. The answer in all these
1737 // edge cases is safety first. It is always safe to clear immediately
1738 // within an Allocate, and then (maybe or maybe not) clear some more later.
1739 if (!(UseTLAB && ZeroTLAB)) {
1740 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1741 header_size, size_in_bytes,
1742 &_igvn);
1743 }
1744 } else {
1745 if (!init->is_complete()) {
1746 // Try to win by zeroing only what the init does not store.
1747 // We can also try to do some peephole optimizations,
1748 // such as combining some adjacent subword stores.
1749 rawmem = init->complete_stores(control, rawmem, object,
1750 header_size, size_in_bytes, &_igvn);
1751 }
1752 // We have no more use for this link, since the AllocateNode goes away:
1753 init->set_req(InitializeNode::RawAddress, top());
1754 // (If we keep the link, it just confuses the register allocator,
1755 // who thinks he sees a real use of the address by the membar.)
1756 }
1757
1758 return rawmem;
1759 }
1760
1761 // Generate prefetch instructions for next allocations.
1762 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1763 Node*& contended_phi_rawmem,
1764 Node* old_eden_top, Node* new_eden_top,
1765 intx lines) {
1766 enum { fall_in_path = 1, pf_path = 2 };
1767 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1768 // Generate prefetch allocation with watermark check.
1769 // As an allocation hits the watermark, we will prefetch starting
1770 // at a "distance" away from watermark.
1771
1772 Node *pf_region = new RegionNode(3);
1773 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1774 TypeRawPtr::BOTTOM );
1775 // I/O is used for Prefetch
1776 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1777
1778 Node *thread = new ThreadLocalNode();
1779 transform_later(thread);
1780
1781 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1782 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1783 transform_later(eden_pf_adr);
1784
1785 Node *old_pf_wm = new LoadPNode(needgc_false,
1786 contended_phi_rawmem, eden_pf_adr,
1787 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1788 MemNode::unordered);
1789 transform_later(old_pf_wm);
1790
1791 // check against new_eden_top
1792 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1793 transform_later(need_pf_cmp);
1794 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1795 transform_later(need_pf_bol);
1796 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1797 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1798 transform_later(need_pf_iff);
1799
1800 // true node, add prefetchdistance
1801 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1802 transform_later(need_pf_true);
1803
1804 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1805 transform_later(need_pf_false);
1806
1807 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1808 _igvn.MakeConX(AllocatePrefetchDistance) );
1809 transform_later(new_pf_wmt );
1810 new_pf_wmt->set_req(0, need_pf_true);
1811
1812 Node *store_new_wmt = new StorePNode(need_pf_true,
1813 contended_phi_rawmem, eden_pf_adr,
1814 TypeRawPtr::BOTTOM, new_pf_wmt,
1815 MemNode::unordered);
1816 transform_later(store_new_wmt);
1817
1818 // adding prefetches
1819 pf_phi_abio->init_req( fall_in_path, i_o );
1820
1821 Node *prefetch_adr;
1822 Node *prefetch;
1823 uint step_size = AllocatePrefetchStepSize;
1824 uint distance = 0;
1825
1826 for ( intx i = 0; i < lines; i++ ) {
1827 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1828 _igvn.MakeConX(distance) );
1829 transform_later(prefetch_adr);
1830 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1831 transform_later(prefetch);
1832 distance += step_size;
1833 i_o = prefetch;
1834 }
1835 pf_phi_abio->set_req( pf_path, i_o );
1836
1837 pf_region->init_req( fall_in_path, need_pf_false );
1838 pf_region->init_req( pf_path, need_pf_true );
1839
1840 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1841 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1842
1843 transform_later(pf_region);
1844 transform_later(pf_phi_rawmem);
1845 transform_later(pf_phi_abio);
1846
1847 needgc_false = pf_region;
1848 contended_phi_rawmem = pf_phi_rawmem;
1849 i_o = pf_phi_abio;
1850 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1851 // Insert a prefetch instruction for each allocation.
1852 // This code is used to generate 1 prefetch instruction per cache line.
1853
1854 // Generate several prefetch instructions.
1855 uint step_size = AllocatePrefetchStepSize;
1856 uint distance = AllocatePrefetchDistance;
1857
1858 // Next cache address.
1859 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1860 _igvn.MakeConX(step_size + distance));
1861 transform_later(cache_adr);
1862 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1863 transform_later(cache_adr);
1864 // Address is aligned to execute prefetch to the beginning of cache line size
1865 // (it is important when BIS instruction is used on SPARC as prefetch).
1866 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1867 cache_adr = new AndXNode(cache_adr, mask);
1868 transform_later(cache_adr);
1869 cache_adr = new CastX2PNode(cache_adr);
1870 transform_later(cache_adr);
1871
1872 // Prefetch
1873 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1874 prefetch->set_req(0, needgc_false);
1875 transform_later(prefetch);
1876 contended_phi_rawmem = prefetch;
1877 Node *prefetch_adr;
1878 distance = step_size;
1879 for ( intx i = 1; i < lines; i++ ) {
1880 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1881 _igvn.MakeConX(distance) );
1882 transform_later(prefetch_adr);
1883 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1884 transform_later(prefetch);
1885 distance += step_size;
1886 contended_phi_rawmem = prefetch;
1887 }
1888 } else if( AllocatePrefetchStyle > 0 ) {
1889 // Insert a prefetch for each allocation only on the fast-path
1890 Node *prefetch_adr;
1891 Node *prefetch;
1892 // Generate several prefetch instructions.
1893 uint step_size = AllocatePrefetchStepSize;
1894 uint distance = AllocatePrefetchDistance;
1895 for ( intx i = 0; i < lines; i++ ) {
1896 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1897 _igvn.MakeConX(distance) );
1898 transform_later(prefetch_adr);
1899 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1900 // Do not let it float too high, since if eden_top == eden_end,
1901 // both might be null.
1902 if( i == 0 ) { // Set control for first prefetch, next follows it
1903 prefetch->init_req(0, needgc_false);
1904 }
1905 transform_later(prefetch);
1906 distance += step_size;
1907 i_o = prefetch;
1908 }
1909 }
1910 return i_o;
1911 }
1912
1913
1914 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1915 expand_allocate_common(alloc, nullptr,
1916 OptoRuntime::new_instance_Type(),
1917 OptoRuntime::new_instance_Java(), nullptr);
1918 }
1919
1920 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1921 Node* length = alloc->in(AllocateNode::ALength);
1922 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1923 InitializeNode* init = alloc->initialization();
1924 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1925 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1926 address slow_call_address; // Address of slow call
1927 if (init != nullptr && init->is_complete_with_arraycopy() &&
1928 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
1929 // Don't zero type array during slow allocation in VM since
1930 // it will be initialized later by arraycopy in compiled code.
1931 slow_call_address = OptoRuntime::new_array_nozero_Java();
1932 } else {
1933 slow_call_address = OptoRuntime::new_array_Java();
1934 }
1935 expand_allocate_common(alloc, length,
1936 OptoRuntime::new_array_Type(),
1937 slow_call_address, valid_length_test);
1938 }
1939
1940 //-------------------mark_eliminated_box----------------------------------
1941 //
1942 // During EA obj may point to several objects but after few ideal graph
1943 // transformations (CCP) it may point to only one non escaping object
1944 // (but still using phi), corresponding locks and unlocks will be marked
1945 // for elimination. Later obj could be replaced with a new node (new phi)
1946 // and which does not have escape information. And later after some graph
1947 // reshape other locks and unlocks (which were not marked for elimination
1948 // before) are connected to this new obj (phi) but they still will not be
1949 // marked for elimination since new obj has no escape information.
1950 // Mark all associated (same box and obj) lock and unlock nodes for
1951 // elimination if some of them marked already.
1952 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
1953 BoxLockNode* oldbox = box->as_BoxLock();
1954 if (oldbox->is_eliminated()) {
1955 return; // This BoxLock node was processed already.
1956 }
1957 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
1958 // New implementation (EliminateNestedLocks) has separate BoxLock
1959 // node for each locked region so mark all associated locks/unlocks as
1960 // eliminated even if different objects are referenced in one locked region
1961 // (for example, OSR compilation of nested loop inside locked scope).
1962 if (EliminateNestedLocks ||
1963 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
1964 // Box is used only in one lock region. Mark this box as eliminated.
1965 oldbox->set_local(); // This verifies correct state of BoxLock
1966 _igvn.hash_delete(oldbox);
1967 oldbox->set_eliminated(); // This changes box's hash value
1968 _igvn.hash_insert(oldbox);
1969
1970 for (uint i = 0; i < oldbox->outcnt(); i++) {
1971 Node* u = oldbox->raw_out(i);
1972 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1973 AbstractLockNode* alock = u->as_AbstractLock();
1974 // Check lock's box since box could be referenced by Lock's debug info.
1975 if (alock->box_node() == oldbox) {
1976 // Mark eliminated all related locks and unlocks.
1977 #ifdef ASSERT
1978 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1979 #endif
1980 alock->set_non_esc_obj();
1981 }
1982 }
1983 }
1984 return;
1985 }
1986
1987 // Create new "eliminated" BoxLock node and use it in monitor debug info
1988 // instead of oldbox for the same object.
1989 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1990
1991 // Note: BoxLock node is marked eliminated only here and it is used
1992 // to indicate that all associated lock and unlock nodes are marked
1993 // for elimination.
1994 newbox->set_local(); // This verifies correct state of BoxLock
1995 newbox->set_eliminated();
1996 transform_later(newbox);
1997
1998 // Replace old box node with new box for all users of the same object.
1999 for (uint i = 0; i < oldbox->outcnt();) {
2000 bool next_edge = true;
2001
2002 Node* u = oldbox->raw_out(i);
2003 if (u->is_AbstractLock()) {
2004 AbstractLockNode* alock = u->as_AbstractLock();
2005 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2006 // Replace Box and mark eliminated all related locks and unlocks.
2007 #ifdef ASSERT
2008 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2009 #endif
2010 alock->set_non_esc_obj();
2011 _igvn.rehash_node_delayed(alock);
2012 alock->set_box_node(newbox);
2013 next_edge = false;
2014 }
2015 }
2016 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2017 FastLockNode* flock = u->as_FastLock();
2018 assert(flock->box_node() == oldbox, "sanity");
2019 _igvn.rehash_node_delayed(flock);
2020 flock->set_box_node(newbox);
2021 next_edge = false;
2022 }
2023
2024 // Replace old box in monitor debug info.
2025 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2026 SafePointNode* sfn = u->as_SafePoint();
2027 JVMState* youngest_jvms = sfn->jvms();
2028 int max_depth = youngest_jvms->depth();
2029 for (int depth = 1; depth <= max_depth; depth++) {
2030 JVMState* jvms = youngest_jvms->of_depth(depth);
2031 int num_mon = jvms->nof_monitors();
2032 // Loop over monitors
2033 for (int idx = 0; idx < num_mon; idx++) {
2034 Node* obj_node = sfn->monitor_obj(jvms, idx);
2035 Node* box_node = sfn->monitor_box(jvms, idx);
2036 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2037 int j = jvms->monitor_box_offset(idx);
2038 _igvn.replace_input_of(u, j, newbox);
2039 next_edge = false;
2040 }
2041 }
2042 }
2043 }
2044 if (next_edge) i++;
2045 }
2046 }
2047
2048 //-----------------------mark_eliminated_locking_nodes-----------------------
2049 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2050 if (alock->box_node()->as_BoxLock()->is_unbalanced()) {
2051 return; // Can't do any more elimination for this locking region
2052 }
2053 if (EliminateNestedLocks) {
2054 if (alock->is_nested()) {
2055 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2056 return;
2057 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2058 // Only Lock node has JVMState needed here.
2059 // Not that preceding claim is documented anywhere else.
2060 if (alock->jvms() != nullptr) {
2061 if (alock->as_Lock()->is_nested_lock_region()) {
2062 // Mark eliminated related nested locks and unlocks.
2063 Node* obj = alock->obj_node();
2064 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2065 assert(!box_node->is_eliminated(), "should not be marked yet");
2066 // Note: BoxLock node is marked eliminated only here
2067 // and it is used to indicate that all associated lock
2068 // and unlock nodes are marked for elimination.
2069 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2070 for (uint i = 0; i < box_node->outcnt(); i++) {
2071 Node* u = box_node->raw_out(i);
2072 if (u->is_AbstractLock()) {
2073 alock = u->as_AbstractLock();
2074 if (alock->box_node() == box_node) {
2075 // Verify that this Box is referenced only by related locks.
2076 assert(alock->obj_node()->eqv_uncast(obj), "");
2077 // Mark all related locks and unlocks.
2078 #ifdef ASSERT
2079 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2080 #endif
2081 alock->set_nested();
2082 }
2083 }
2084 }
2085 } else {
2086 #ifdef ASSERT
2087 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2088 if (C->log() != nullptr)
2089 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2090 #endif
2091 }
2092 }
2093 return;
2094 }
2095 // Process locks for non escaping object
2096 assert(alock->is_non_esc_obj(), "");
2097 } // EliminateNestedLocks
2098
2099 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2100 // Look for all locks of this object and mark them and
2101 // corresponding BoxLock nodes as eliminated.
2102 Node* obj = alock->obj_node();
2103 for (uint j = 0; j < obj->outcnt(); j++) {
2104 Node* o = obj->raw_out(j);
2105 if (o->is_AbstractLock() &&
2106 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2107 alock = o->as_AbstractLock();
2108 Node* box = alock->box_node();
2109 // Replace old box node with new eliminated box for all users
2110 // of the same object and mark related locks as eliminated.
2111 mark_eliminated_box(box, obj);
2112 }
2113 }
2114 }
2115 }
2116
2117 // we have determined that this lock/unlock can be eliminated, we simply
2118 // eliminate the node without expanding it.
2119 //
2120 // Note: The membar's associated with the lock/unlock are currently not
2121 // eliminated. This should be investigated as a future enhancement.
2122 //
2123 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2124
2125 if (!alock->is_eliminated()) {
2126 return false;
2127 }
2128 #ifdef ASSERT
2129 if (!alock->is_coarsened()) {
2130 // Check that new "eliminated" BoxLock node is created.
2131 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2132 assert(oldbox->is_eliminated(), "should be done already");
2133 }
2134 #endif
2135
2136 alock->log_lock_optimization(C, "eliminate_lock");
2137
2138 #ifndef PRODUCT
2139 if (PrintEliminateLocks) {
2140 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2141 }
2142 #endif
2143
2144 Node* mem = alock->in(TypeFunc::Memory);
2145 Node* ctrl = alock->in(TypeFunc::Control);
2146 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2147
2148 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2149 // There are 2 projections from the lock. The lock node will
2150 // be deleted when its last use is subsumed below.
2151 assert(alock->outcnt() == 2 &&
2152 _callprojs.fallthrough_proj != nullptr &&
2153 _callprojs.fallthrough_memproj != nullptr,
2154 "Unexpected projections from Lock/Unlock");
2155
2156 Node* fallthroughproj = _callprojs.fallthrough_proj;
2157 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2158
2159 // The memory projection from a lock/unlock is RawMem
2160 // The input to a Lock is merged memory, so extract its RawMem input
2161 // (unless the MergeMem has been optimized away.)
2162 if (alock->is_Lock()) {
2163 // Search for MemBarAcquireLock node and delete it also.
2164 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2165 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2166 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2167 Node* memproj = membar->proj_out(TypeFunc::Memory);
2168 _igvn.replace_node(ctrlproj, fallthroughproj);
2169 _igvn.replace_node(memproj, memproj_fallthrough);
2170
2171 // Delete FastLock node also if this Lock node is unique user
2172 // (a loop peeling may clone a Lock node).
2173 Node* flock = alock->as_Lock()->fastlock_node();
2174 if (flock->outcnt() == 1) {
2175 assert(flock->unique_out() == alock, "sanity");
2176 _igvn.replace_node(flock, top());
2177 }
2178 }
2179
2180 // Search for MemBarReleaseLock node and delete it also.
2181 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2182 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2183 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2184 mem->is_Proj() && membar == mem->in(0), "");
2185 _igvn.replace_node(fallthroughproj, ctrl);
2186 _igvn.replace_node(memproj_fallthrough, mem);
2187 fallthroughproj = ctrl;
2188 memproj_fallthrough = mem;
2189 ctrl = membar->in(TypeFunc::Control);
2190 mem = membar->in(TypeFunc::Memory);
2191 }
2192
2193 _igvn.replace_node(fallthroughproj, ctrl);
2194 _igvn.replace_node(memproj_fallthrough, mem);
2195 return true;
2196 }
2197
2198
2199 //------------------------------expand_lock_node----------------------
2200 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2201
2202 Node* ctrl = lock->in(TypeFunc::Control);
2203 Node* mem = lock->in(TypeFunc::Memory);
2204 Node* obj = lock->obj_node();
2205 Node* box = lock->box_node();
2206 Node* flock = lock->fastlock_node();
2207
2208 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2209
2210 // Make the merge point
2211 Node *region;
2212 Node *mem_phi;
2213 Node *slow_path;
2214
2215 region = new RegionNode(3);
2216 // create a Phi for the memory state
2217 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2218
2219 // Optimize test; set region slot 2
2220 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2221 mem_phi->init_req(2, mem);
2222
2223 // Make slow path call
2224 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2225 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2226 obj, box, nullptr);
2227
2228 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2229
2230 // Slow path can only throw asynchronous exceptions, which are always
2231 // de-opted. So the compiler thinks the slow-call can never throw an
2232 // exception. If it DOES throw an exception we would need the debug
2233 // info removed first (since if it throws there is no monitor).
2234 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2235 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2236
2237 // Capture slow path
2238 // disconnect fall-through projection from call and create a new one
2239 // hook up users of fall-through projection to region
2240 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2241 transform_later(slow_ctrl);
2242 _igvn.hash_delete(_callprojs.fallthrough_proj);
2243 _callprojs.fallthrough_proj->disconnect_inputs(C);
2244 region->init_req(1, slow_ctrl);
2245 // region inputs are now complete
2246 transform_later(region);
2247 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2248
2249 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2250
2251 mem_phi->init_req(1, memproj);
2252
2253 transform_later(mem_phi);
2254
2255 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2256 }
2257
2258 //------------------------------expand_unlock_node----------------------
2259 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2260
2261 Node* ctrl = unlock->in(TypeFunc::Control);
2262 Node* mem = unlock->in(TypeFunc::Memory);
2263 Node* obj = unlock->obj_node();
2264 Node* box = unlock->box_node();
2265
2266 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2267
2268 // No need for a null check on unlock
2269
2270 // Make the merge point
2271 Node *region;
2272 Node *mem_phi;
2273
2274 region = new RegionNode(3);
2275 // create a Phi for the memory state
2276 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2277
2278 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2279 funlock = transform_later( funlock )->as_FastUnlock();
2280 // Optimize test; set region slot 2
2281 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2282 Node *thread = transform_later(new ThreadLocalNode());
2283
2284 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2285 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2286 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2287
2288 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2289 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2290 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2291
2292 // No exceptions for unlocking
2293 // Capture slow path
2294 // disconnect fall-through projection from call and create a new one
2295 // hook up users of fall-through projection to region
2296 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2297 transform_later(slow_ctrl);
2298 _igvn.hash_delete(_callprojs.fallthrough_proj);
2299 _callprojs.fallthrough_proj->disconnect_inputs(C);
2300 region->init_req(1, slow_ctrl);
2301 // region inputs are now complete
2302 transform_later(region);
2303 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2304
2305 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2306 mem_phi->init_req(1, memproj );
2307 mem_phi->init_req(2, mem);
2308 transform_later(mem_phi);
2309
2310 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2311 }
2312
2313 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2314 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2315 Node* bol = check->unique_out();
2316 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2317 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2318 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2319
2320 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2321 Node* iff = bol->last_out(i);
2322 assert(iff->is_If(), "where's the if?");
2323
2324 if (iff->in(0)->is_top()) {
2325 _igvn.replace_input_of(iff, 1, C->top());
2326 continue;
2327 }
2328
2329 Node* iftrue = iff->as_If()->proj_out(1);
2330 Node* iffalse = iff->as_If()->proj_out(0);
2331 Node* ctrl = iff->in(0);
2332
2333 Node* subklass = nullptr;
2334 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2335 subklass = obj_or_subklass;
2336 } else {
2337 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2338 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2339 }
2340
2341 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2342
2343 _igvn.replace_input_of(iff, 0, C->top());
2344 _igvn.replace_node(iftrue, not_subtype_ctrl);
2345 _igvn.replace_node(iffalse, ctrl);
2346 }
2347 _igvn.replace_node(check, C->top());
2348 }
2349
2350 //---------------------------eliminate_macro_nodes----------------------
2351 // Eliminate scalar replaced allocations and associated locks.
2352 void PhaseMacroExpand::eliminate_macro_nodes() {
2353 if (C->macro_count() == 0)
2354 return;
2355 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2356
2357 // Before elimination may re-mark (change to Nested or NonEscObj)
2358 // all associated (same box and obj) lock and unlock nodes.
2359 int cnt = C->macro_count();
2360 for (int i=0; i < cnt; i++) {
2361 Node *n = C->macro_node(i);
2362 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2363 mark_eliminated_locking_nodes(n->as_AbstractLock());
2364 }
2365 }
2366 // Re-marking may break consistency of Coarsened locks.
2367 if (!C->coarsened_locks_consistent()) {
2368 return; // recompile without Coarsened locks if broken
2369 } else {
2370 // After coarsened locks are eliminated locking regions
2371 // become unbalanced. We should not execute any more
2372 // locks elimination optimizations on them.
2373 C->mark_unbalanced_boxes();
2374 }
2375
2376 // First, attempt to eliminate locks
2377 bool progress = true;
2378 while (progress) {
2379 progress = false;
2380 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2381 Node* n = C->macro_node(i - 1);
2382 bool success = false;
2383 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2384 if (n->is_AbstractLock()) {
2385 success = eliminate_locking_node(n->as_AbstractLock());
2386 #ifndef PRODUCT
2387 if (success && PrintOptoStatistics) {
2388 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2389 }
2390 #endif
2391 }
2392 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2393 progress = progress || success;
2394 }
2395 }
2396 // Next, attempt to eliminate allocations
2397 _has_locks = false;
2398 progress = true;
2399 while (progress) {
2400 progress = false;
2401 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2402 Node* n = C->macro_node(i - 1);
2403 bool success = false;
2404 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2405 switch (n->class_id()) {
2406 case Node::Class_Allocate:
2407 case Node::Class_AllocateArray:
2408 success = eliminate_allocate_node(n->as_Allocate());
2409 #ifndef PRODUCT
2410 if (success && PrintOptoStatistics) {
2411 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2412 }
2413 #endif
2414 break;
2415 case Node::Class_CallStaticJava:
2416 success = eliminate_boxing_node(n->as_CallStaticJava());
2417 break;
2418 case Node::Class_Lock:
2419 case Node::Class_Unlock:
2420 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2421 _has_locks = true;
2422 break;
2423 case Node::Class_ArrayCopy:
2424 break;
2425 case Node::Class_OuterStripMinedLoop:
2426 break;
2427 case Node::Class_SubTypeCheck:
2428 break;
2429 case Node::Class_Opaque1:
2430 break;
2431 default:
2432 assert(n->Opcode() == Op_LoopLimit ||
2433 n->is_Opaque4() ||
2434 n->is_OpaqueInitializedAssertionPredicate() ||
2435 n->Opcode() == Op_MaxL ||
2436 n->Opcode() == Op_MinL ||
2437 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2438 "unknown node type in macro list");
2439 }
2440 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2441 progress = progress || success;
2442 }
2443 }
2444 #ifndef PRODUCT
2445 if (PrintOptoStatistics) {
2446 int membar_after = count_MemBar(C);
2447 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2448 }
2449 #endif
2450 }
2451
2452 //------------------------------expand_macro_nodes----------------------
2453 // Returns true if a failure occurred.
2454 bool PhaseMacroExpand::expand_macro_nodes() {
2455 // Do not allow new macro nodes once we started to expand
2456 C->reset_allow_macro_nodes();
2457 if (StressMacroExpansion) {
2458 C->shuffle_macro_nodes();
2459 }
2460 // Last attempt to eliminate macro nodes.
2461 eliminate_macro_nodes();
2462 if (C->failing()) return true;
2463
2464 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2465 bool progress = true;
2466 while (progress) {
2467 progress = false;
2468 for (int i = C->macro_count(); i > 0; i--) {
2469 Node* n = C->macro_node(i-1);
2470 bool success = false;
2471 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2472 if (n->Opcode() == Op_LoopLimit) {
2473 // Remove it from macro list and put on IGVN worklist to optimize.
2474 C->remove_macro_node(n);
2475 _igvn._worklist.push(n);
2476 success = true;
2477 } else if (n->Opcode() == Op_CallStaticJava) {
2478 // Remove it from macro list and put on IGVN worklist to optimize.
2479 C->remove_macro_node(n);
2480 _igvn._worklist.push(n);
2481 success = true;
2482 } else if (n->is_Opaque1()) {
2483 _igvn.replace_node(n, n->in(1));
2484 success = true;
2485 } else if (n->is_Opaque4()) {
2486 // With Opaque4 nodes, the expectation is that the test of input 1
2487 // is always equal to the constant value of input 2. So we can
2488 // remove the Opaque4 and replace it by input 2. In debug builds,
2489 // leave the non constant test in instead to sanity check that it
2490 // never fails (if it does, that subgraph was constructed so, at
2491 // runtime, a Halt node is executed).
2492 #ifdef ASSERT
2493 _igvn.replace_node(n, n->in(1));
2494 #else
2495 _igvn.replace_node(n, n->in(2));
2496 #endif
2497 success = true;
2498 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
2499 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
2500 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
2501 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
2502 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
2503 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
2504 #ifdef ASSERT
2505 _igvn.replace_node(n, n->in(1));
2506 #else
2507 _igvn.replace_node(n, _igvn.intcon(1));
2508 #endif // ASSERT
2509 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2510 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2511 C->remove_macro_node(n);
2512 success = true;
2513 } else if (n->Opcode() == Op_MaxL) {
2514 // Since MaxL and MinL are not implemented in the backend, we expand them to
2515 // a CMoveL construct now. At least until here, the type could be computed
2516 // precisely. CMoveL is not so smart, but we can give it at least the best
2517 // type we know abouot n now.
2518 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
2519 _igvn.replace_node(n, repl);
2520 success = true;
2521 } else if (n->Opcode() == Op_MinL) {
2522 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
2523 _igvn.replace_node(n, repl);
2524 success = true;
2525 }
2526 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2527 progress = progress || success;
2528 if (success) {
2529 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2530 }
2531 }
2532 }
2533
2534 // Clean up the graph so we're less likely to hit the maximum node
2535 // limit
2536 _igvn.set_delay_transform(false);
2537 _igvn.optimize();
2538 if (C->failing()) return true;
2539 _igvn.set_delay_transform(true);
2540
2541
2542 // Because we run IGVN after each expansion, some macro nodes may go
2543 // dead and be removed from the list as we iterate over it. Move
2544 // Allocate nodes (processed in a second pass) at the beginning of
2545 // the list and then iterate from the last element of the list until
2546 // an Allocate node is seen. This is robust to random deletion in
2547 // the list due to nodes going dead.
2548 C->sort_macro_nodes();
2549
2550 // expand arraycopy "macro" nodes first
2551 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2552 // before the allocate nodes are expanded.
2553 while (C->macro_count() > 0) {
2554 int macro_count = C->macro_count();
2555 Node * n = C->macro_node(macro_count-1);
2556 assert(n->is_macro(), "only macro nodes expected here");
2557 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
2558 // node is unreachable, so don't try to expand it
2559 C->remove_macro_node(n);
2560 continue;
2561 }
2562 if (n->is_Allocate()) {
2563 break;
2564 }
2565 // Make sure expansion will not cause node limit to be exceeded.
2566 // Worst case is a macro node gets expanded into about 200 nodes.
2567 // Allow 50% more for optimization.
2568 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2569 return true;
2570 }
2571
2572 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2573 switch (n->class_id()) {
2574 case Node::Class_Lock:
2575 expand_lock_node(n->as_Lock());
2576 break;
2577 case Node::Class_Unlock:
2578 expand_unlock_node(n->as_Unlock());
2579 break;
2580 case Node::Class_ArrayCopy:
2581 expand_arraycopy_node(n->as_ArrayCopy());
2582 break;
2583 case Node::Class_SubTypeCheck:
2584 expand_subtypecheck_node(n->as_SubTypeCheck());
2585 break;
2586 default:
2587 assert(false, "unknown node type in macro list");
2588 }
2589 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2590 if (C->failing()) return true;
2591 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2592
2593 // Clean up the graph so we're less likely to hit the maximum node
2594 // limit
2595 _igvn.set_delay_transform(false);
2596 _igvn.optimize();
2597 if (C->failing()) return true;
2598 _igvn.set_delay_transform(true);
2599 }
2600
2601 // All nodes except Allocate nodes are expanded now. There could be
2602 // new optimization opportunities (such as folding newly created
2603 // load from a just allocated object). Run IGVN.
2604
2605 // expand "macro" nodes
2606 // nodes are removed from the macro list as they are processed
2607 while (C->macro_count() > 0) {
2608 int macro_count = C->macro_count();
2609 Node * n = C->macro_node(macro_count-1);
2610 assert(n->is_macro(), "only macro nodes expected here");
2611 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
2612 // node is unreachable, so don't try to expand it
2613 C->remove_macro_node(n);
2614 continue;
2615 }
2616 // Make sure expansion will not cause node limit to be exceeded.
2617 // Worst case is a macro node gets expanded into about 200 nodes.
2618 // Allow 50% more for optimization.
2619 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2620 return true;
2621 }
2622 switch (n->class_id()) {
2623 case Node::Class_Allocate:
2624 expand_allocate(n->as_Allocate());
2625 break;
2626 case Node::Class_AllocateArray:
2627 expand_allocate_array(n->as_AllocateArray());
2628 break;
2629 default:
2630 assert(false, "unknown node type in macro list");
2631 }
2632 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2633 if (C->failing()) return true;
2634 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2635
2636 // Clean up the graph so we're less likely to hit the maximum node
2637 // limit
2638 _igvn.set_delay_transform(false);
2639 _igvn.optimize();
2640 if (C->failing()) return true;
2641 _igvn.set_delay_transform(true);
2642 }
2643
2644 _igvn.set_delay_transform(false);
2645 return false;
2646 }
2647
2648 #ifndef PRODUCT
2649 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
2650 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
2651 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
2652 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
2653
2654 void PhaseMacroExpand::print_statistics() {
2655 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
2656 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
2657 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
2658 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
2659 }
2660
2661 int PhaseMacroExpand::count_MemBar(Compile *C) {
2662 if (!PrintOptoStatistics) {
2663 return 0;
2664 }
2665 Unique_Node_List ideal_nodes;
2666 int total = 0;
2667 ideal_nodes.map(C->live_nodes(), nullptr);
2668 ideal_nodes.push(C->root());
2669 for (uint next = 0; next < ideal_nodes.size(); ++next) {
2670 Node* n = ideal_nodes.at(next);
2671 if (n->is_MemBar()) {
2672 total++;
2673 }
2674 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2675 Node* m = n->fast_out(i);
2676 ideal_nodes.push(m);
2677 }
2678 }
2679 return total;
2680 }
2681 #endif