1 /*
2 * Copyright (c) 2005, 2025, 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 "compiler/compileLog.hpp"
26 #include "gc/shared/collectedHeap.inline.hpp"
27 #include "gc/shared/tlab_globals.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/universe.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/node.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/phaseX.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subnode.hpp"
50 #include "opto/subtypenode.hpp"
51 #include "opto/type.hpp"
52 #include "prims/jvmtiExport.hpp"
53 #include "runtime/continuation.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #if INCLUDE_G1GC
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_G1GC
60
61
62 //
63 // Replace any references to "oldref" in inputs to "use" with "newref".
64 // Returns the number of replacements made.
65 //
66 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
67 int nreplacements = 0;
68 uint req = use->req();
69 for (uint j = 0; j < use->len(); j++) {
70 Node *uin = use->in(j);
71 if (uin == oldref) {
72 if (j < req)
73 use->set_req(j, newref);
74 else
75 use->set_prec(j, newref);
76 nreplacements++;
77 } else if (j >= req && uin == nullptr) {
78 break;
79 }
80 }
81 return nreplacements;
82 }
83
84 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
85 assert(old != nullptr, "sanity");
86 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
87 Node* use = old->fast_out(i);
88 _igvn.rehash_node_delayed(use);
89 imax -= replace_input(use, old, target);
90 // back up iterator
91 --i;
92 }
93 assert(old->outcnt() == 0, "all uses must be deleted");
94 }
95
96 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
97 Node* cmp = word;
98 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
99 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
100 transform_later(iff);
101
102 // Fast path taken.
103 Node *fast_taken = transform_later(new IfFalseNode(iff));
104
105 // Fast path not-taken, i.e. slow path
106 Node *slow_taken = transform_later(new IfTrueNode(iff));
107
108 region->init_req(edge, fast_taken); // Capture fast-control
109 return slow_taken;
110 }
111
112 //--------------------copy_predefined_input_for_runtime_call--------------------
113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
114 // Set fixed predefined input arguments
115 call->init_req( TypeFunc::Control, ctrl );
116 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
117 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
118 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
119 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
120 }
121
122 //------------------------------make_slow_call---------------------------------
123 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
124 address slow_call, const char* leaf_name, Node* slow_path,
125 Node* parm0, Node* parm1, Node* parm2) {
126
127 // Slow-path call
128 CallNode *call = leaf_name
129 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
130 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
131
132 // Slow path call has no side-effects, uses few values
133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
134 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
135 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
136 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
137 call->copy_call_debug_info(&_igvn, oldcall);
138 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
139 _igvn.replace_node(oldcall, call);
140 transform_later(call);
141
142 return call;
143 }
144
145 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
146 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
147 bs->eliminate_gc_barrier(this, p2x);
148 #ifndef PRODUCT
149 if (PrintOptoStatistics) {
150 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
151 }
152 #endif
153 }
154
155 // Search for a memory operation for the specified memory slice.
156 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
157 Node *orig_mem = mem;
158 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
159 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
160 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
161 while (true) {
162 if (mem == alloc_mem || mem == start_mem ) {
163 return mem; // hit one of our sentinels
164 } else if (mem->is_MergeMem()) {
165 mem = mem->as_MergeMem()->memory_at(alias_idx);
166 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
167 Node *in = mem->in(0);
168 // we can safely skip over safepoints, calls, locks and membars because we
169 // already know that the object is safe to eliminate.
170 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
171 return in;
172 } else if (in->is_Call()) {
173 CallNode *call = in->as_Call();
174 if (call->may_modify(tinst, phase)) {
175 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
176 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
177 return in;
178 }
179 }
180 mem = in->in(TypeFunc::Memory);
181 } else if (in->is_MemBar()) {
182 ArrayCopyNode* ac = nullptr;
183 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
184 if (ac != nullptr) {
185 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
186 return ac;
187 }
188 }
189 mem = in->in(TypeFunc::Memory);
190 } else {
191 #ifdef ASSERT
192 in->dump();
193 mem->dump();
194 assert(false, "unexpected projection");
195 #endif
196 }
197 } else if (mem->is_Store()) {
198 const TypePtr* atype = mem->as_Store()->adr_type();
199 int adr_idx = phase->C->get_alias_index(atype);
200 if (adr_idx == alias_idx) {
201 assert(atype->isa_oopptr(), "address type must be oopptr");
202 int adr_offset = atype->offset();
203 uint adr_iid = atype->is_oopptr()->instance_id();
204 // Array elements references have the same alias_idx
205 // but different offset and different instance_id.
206 if (adr_offset == offset && adr_iid == alloc->_idx) {
207 return mem;
208 }
209 } else {
210 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
211 }
212 mem = mem->in(MemNode::Memory);
213 } else if (mem->is_ClearArray()) {
214 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
215 // Can not bypass initialization of the instance
216 // we are looking.
217 DEBUG_ONLY(intptr_t offset;)
218 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
219 InitializeNode* init = alloc->as_Allocate()->initialization();
220 // We are looking for stored value, return Initialize node
221 // or memory edge from Allocate node.
222 if (init != nullptr) {
223 return init;
224 } else {
225 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
226 }
227 }
228 // Otherwise skip it (the call updated 'mem' value).
229 } else if (mem->Opcode() == Op_SCMemProj) {
230 mem = mem->in(0);
231 Node* adr = nullptr;
232 if (mem->is_LoadStore()) {
233 adr = mem->in(MemNode::Address);
234 } else {
235 assert(mem->Opcode() == Op_EncodeISOArray ||
236 mem->Opcode() == Op_StrCompressedCopy, "sanity");
237 adr = mem->in(3); // Destination array
238 }
239 const TypePtr* atype = adr->bottom_type()->is_ptr();
240 int adr_idx = phase->C->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 DEBUG_ONLY(mem->dump();)
243 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
244 return nullptr;
245 }
246 mem = mem->in(MemNode::Memory);
247 } else if (mem->Opcode() == Op_StrInflatedCopy) {
248 Node* adr = mem->in(3); // Destination array
249 const TypePtr* atype = adr->bottom_type()->is_ptr();
250 int adr_idx = phase->C->get_alias_index(atype);
251 if (adr_idx == alias_idx) {
252 DEBUG_ONLY(mem->dump();)
253 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
254 return nullptr;
255 }
256 mem = mem->in(MemNode::Memory);
257 } else {
258 return mem;
259 }
260 assert(mem != orig_mem, "dead memory loop");
261 }
262 }
263
264 // Generate loads from source of the arraycopy for fields of
265 // destination needed at a deoptimization point
266 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
267 BasicType bt = ft;
268 const Type *type = ftype;
269 if (ft == T_NARROWOOP) {
270 bt = T_OBJECT;
271 type = ftype->make_oopptr();
272 }
273 Node* res = nullptr;
274 if (ac->is_clonebasic()) {
275 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
276 Node* base = ac->in(ArrayCopyNode::Src);
277 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
278 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
279 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
280 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
281 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
282 } else {
283 if (ac->modifies(offset, offset, &_igvn, true)) {
284 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
285 uint shift = exact_log2(type2aelembytes(bt));
286 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
287 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
288 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
289 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
290
291 Node* adr = nullptr;
292 const TypePtr* adr_type = nullptr;
293 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
294 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
295 Node* base = ac->in(ArrayCopyNode::Src);
296 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
297 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
298 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
299 // Don't emit a new load from src if src == dst but try to get the value from memory instead
300 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
301 }
302 } else {
303 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
304 #ifdef _LP64
305 diff = _igvn.transform(new ConvI2LNode(diff));
306 #endif
307 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
308
309 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
310 Node* base = ac->in(ArrayCopyNode::Src);
311 adr = _igvn.transform(new AddPNode(base, base, off));
312 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
313 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
314 // Non constant offset in the array: we can't statically
315 // determine the value
316 return nullptr;
317 }
318 }
319 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
320 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
321 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
322 }
323 }
324 if (res != nullptr) {
325 if (ftype->isa_narrowoop()) {
326 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
327 res = _igvn.transform(new EncodePNode(res, ftype));
328 }
329 return res;
330 }
331 return nullptr;
332 }
333
334 //
335 // Given a Memory Phi, compute a value Phi containing the values from stores
336 // on the input paths.
337 // Note: this function is recursive, its depth is limited by the "level" argument
338 // Returns the computed Phi, or null if it cannot compute it.
339 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) {
340 assert(mem->is_Phi(), "sanity");
341 int alias_idx = C->get_alias_index(adr_t);
342 int offset = adr_t->offset();
343 int instance_id = adr_t->instance_id();
344
345 // Check if an appropriate value phi already exists.
346 Node* region = mem->in(0);
347 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
348 Node* phi = region->fast_out(k);
349 if (phi->is_Phi() && phi != mem &&
350 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
351 return phi;
352 }
353 }
354 // Check if an appropriate new value phi already exists.
355 Node* new_phi = value_phis->find(mem->_idx);
356 if (new_phi != nullptr)
357 return new_phi;
358
359 if (level <= 0) {
360 return nullptr; // Give up: phi tree too deep
361 }
362 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
363 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
364 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
365
366 uint length = mem->req();
367 GrowableArray <Node *> values(length, length, nullptr);
368
369 // create a new Phi for the value
370 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
371 transform_later(phi);
372 value_phis->push(phi, mem->_idx);
373
374 for (uint j = 1; j < length; j++) {
375 Node *in = mem->in(j);
376 if (in == nullptr || in->is_top()) {
377 values.at_put(j, in);
378 } else {
379 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
380 if (val == start_mem || val == alloc_mem) {
381 // hit a sentinel, return appropriate 0 value
382 values.at_put(j, _igvn.zerocon(ft));
383 continue;
384 }
385 if (val->is_Initialize()) {
386 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
387 }
388 if (val == nullptr) {
389 return nullptr; // can't find a value on this path
390 }
391 if (val == mem) {
392 values.at_put(j, mem);
393 } else if (val->is_Store()) {
394 Node* n = val->in(MemNode::ValueIn);
395 if (is_subword_type(ft)) {
396 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
397 }
398 values.at_put(j, n);
399 } else if(val->is_Proj() && val->in(0) == alloc) {
400 values.at_put(j, _igvn.zerocon(ft));
401 } else if (val->is_Phi()) {
402 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
403 if (val == nullptr) {
404 return nullptr;
405 }
406 values.at_put(j, val);
407 } else if (val->Opcode() == Op_SCMemProj) {
408 assert(val->in(0)->is_LoadStore() ||
409 val->in(0)->Opcode() == Op_EncodeISOArray ||
410 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
411 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
412 return nullptr;
413 } else if (val->is_ArrayCopy()) {
414 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
415 if (res == nullptr) {
416 return nullptr;
417 }
418 values.at_put(j, res);
419 } else if (val->is_top()) {
420 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
421 // IGVN will clean this up later.
422 values.at_put(j, val);
423 } else {
424 DEBUG_ONLY( val->dump(); )
425 assert(false, "unknown node on this path");
426 return nullptr; // unknown node on this path
427 }
428 }
429 }
430 // Set Phi's inputs
431 for (uint j = 1; j < length; j++) {
432 if (values.at(j) == mem) {
433 phi->init_req(j, phi);
434 } else {
435 phi->init_req(j, values.at(j));
436 }
437 }
438 return phi;
439 }
440
441 // Search the last value stored into the object's field.
442 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
443 assert(adr_t->is_known_instance_field(), "instance required");
444 int instance_id = adr_t->instance_id();
445 assert((uint)instance_id == alloc->_idx, "wrong allocation");
446
447 int alias_idx = C->get_alias_index(adr_t);
448 int offset = adr_t->offset();
449 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
450 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
451 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
452 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
453 VectorSet visited;
454
455 bool done = sfpt_mem == alloc_mem;
456 Node *mem = sfpt_mem;
457 while (!done) {
458 if (visited.test_set(mem->_idx)) {
459 return nullptr; // found a loop, give up
460 }
461 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
462 if (mem == start_mem || mem == alloc_mem) {
463 done = true; // hit a sentinel, return appropriate 0 value
464 } else if (mem->is_Initialize()) {
465 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
466 if (mem == nullptr) {
467 done = true; // Something go wrong.
468 } else if (mem->is_Store()) {
469 const TypePtr* atype = mem->as_Store()->adr_type();
470 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
471 done = true;
472 }
473 } else if (mem->is_Store()) {
474 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
475 assert(atype != nullptr, "address type must be oopptr");
476 assert(C->get_alias_index(atype) == alias_idx &&
477 atype->is_known_instance_field() && atype->offset() == offset &&
478 atype->instance_id() == instance_id, "store is correct memory slice");
479 done = true;
480 } else if (mem->is_Phi()) {
481 // try to find a phi's unique input
482 Node *unique_input = nullptr;
483 Node *top = C->top();
484 for (uint i = 1; i < mem->req(); i++) {
485 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
486 if (n == nullptr || n == top || n == mem) {
487 continue;
488 } else if (unique_input == nullptr) {
489 unique_input = n;
490 } else if (unique_input != n) {
491 unique_input = top;
492 break;
493 }
494 }
495 if (unique_input != nullptr && unique_input != top) {
496 mem = unique_input;
497 } else {
498 done = true;
499 }
500 } else if (mem->is_ArrayCopy()) {
501 done = true;
502 } else {
503 DEBUG_ONLY( mem->dump(); )
504 assert(false, "unexpected node");
505 }
506 }
507 if (mem != nullptr) {
508 if (mem == start_mem || mem == alloc_mem) {
509 // hit a sentinel, return appropriate 0 value
510 return _igvn.zerocon(ft);
511 } else if (mem->is_Store()) {
512 return mem->in(MemNode::ValueIn);
513 } else if (mem->is_Phi()) {
514 // attempt to produce a Phi reflecting the values on the input paths of the Phi
515 Node_Stack value_phis(8);
516 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
517 if (phi != nullptr) {
518 return phi;
519 } else {
520 // Kill all new Phis
521 while(value_phis.is_nonempty()) {
522 Node* n = value_phis.node();
523 _igvn.replace_node(n, C->top());
524 value_phis.pop();
525 }
526 }
527 } else if (mem->is_ArrayCopy()) {
528 Node* ctl = mem->in(0);
529 Node* m = mem->in(TypeFunc::Memory);
530 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
531 // pin the loads in the uncommon trap path
532 ctl = sfpt_ctl;
533 m = sfpt_mem;
534 }
535 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
536 }
537 }
538 // Something go wrong.
539 return nullptr;
540 }
541
542 // Check the possibility of scalar replacement.
543 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
544 // Scan the uses of the allocation to check for anything that would
545 // prevent us from eliminating it.
546 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
547 DEBUG_ONLY( Node* disq_node = nullptr; )
548 bool can_eliminate = true;
549 bool reduce_merge_precheck = (safepoints == nullptr);
550
551 Node* res = alloc->result_cast();
552 const TypeOopPtr* res_type = nullptr;
553 if (res == nullptr) {
554 // All users were eliminated.
555 } else if (!res->is_CheckCastPP()) {
556 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
557 can_eliminate = false;
558 } else {
559 res_type = igvn->type(res)->isa_oopptr();
560 if (res_type == nullptr) {
561 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
562 can_eliminate = false;
563 } else if (!res_type->klass_is_exact()) {
564 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
565 can_eliminate = false;
566 } else if (res_type->isa_aryptr()) {
567 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
568 if (length < 0) {
569 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
570 can_eliminate = false;
571 }
572 }
573 }
574
575 if (can_eliminate && res != nullptr) {
576 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
577 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
578 j < jmax && can_eliminate; j++) {
579 Node* use = res->fast_out(j);
580
581 if (use->is_AddP()) {
582 const TypePtr* addp_type = igvn->type(use)->is_ptr();
583 int offset = addp_type->offset();
584
585 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
586 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
587 can_eliminate = false;
588 break;
589 }
590 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
591 k < kmax && can_eliminate; k++) {
592 Node* n = use->fast_out(k);
593 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !reduce_merge_precheck) {
594 DEBUG_ONLY(disq_node = n;)
595 if (n->is_Load() || n->is_LoadStore()) {
596 NOT_PRODUCT(fail_eliminate = "Field load";)
597 } else {
598 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
599 }
600 can_eliminate = false;
601 }
602 }
603 } else if (use->is_ArrayCopy() &&
604 (use->as_ArrayCopy()->is_clonebasic() ||
605 use->as_ArrayCopy()->is_arraycopy_validated() ||
606 use->as_ArrayCopy()->is_copyof_validated() ||
607 use->as_ArrayCopy()->is_copyofrange_validated()) &&
608 use->in(ArrayCopyNode::Dest) == res) {
609 // ok to eliminate
610 } else if (use->is_SafePoint()) {
611 SafePointNode* sfpt = use->as_SafePoint();
612 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
613 // Object is passed as argument.
614 DEBUG_ONLY(disq_node = use;)
615 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
616 can_eliminate = false;
617 }
618 Node* sfptMem = sfpt->memory();
619 if (sfptMem == nullptr || sfptMem->is_top()) {
620 DEBUG_ONLY(disq_node = use;)
621 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
622 can_eliminate = false;
623 } else if (!reduce_merge_precheck) {
624 safepoints->append_if_missing(sfpt);
625 }
626 } else if (reduce_merge_precheck &&
627 (use->is_Phi() || use->is_EncodeP() ||
628 use->Opcode() == Op_MemBarRelease ||
629 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
630 // Nothing to do
631 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
632 if (use->is_Phi()) {
633 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
634 NOT_PRODUCT(fail_eliminate = "Object is return value";)
635 } else {
636 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
637 }
638 DEBUG_ONLY(disq_node = use;)
639 } else {
640 if (use->Opcode() == Op_Return) {
641 NOT_PRODUCT(fail_eliminate = "Object is return value";)
642 } else {
643 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
644 }
645 DEBUG_ONLY(disq_node = use;)
646 }
647 can_eliminate = false;
648 }
649 }
650 }
651
652 #ifndef PRODUCT
653 if (PrintEliminateAllocations && safepoints != nullptr) {
654 if (can_eliminate) {
655 tty->print("Scalar ");
656 if (res == nullptr)
657 alloc->dump();
658 else
659 res->dump();
660 } else if (alloc->_is_scalar_replaceable) {
661 tty->print("NotScalar (%s)", fail_eliminate);
662 if (res == nullptr)
663 alloc->dump();
664 else
665 res->dump();
666 #ifdef ASSERT
667 if (disq_node != nullptr) {
668 tty->print(" >>>> ");
669 disq_node->dump();
670 }
671 #endif /*ASSERT*/
672 }
673 }
674
675 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
676 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
677 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
678 }
679 #endif
680 return can_eliminate;
681 }
682
683 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
684 Node* res = alloc->result_cast();
685 int nfields = 0;
686 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
687
688 if (res != nullptr) {
689 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
690
691 if (res_type->isa_instptr()) {
692 // find the fields of the class which will be needed for safepoint debug information
693 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
694 nfields = iklass->nof_nonstatic_fields();
695 } else {
696 // find the array's elements which will be needed for safepoint debug information
697 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
698 assert(nfields >= 0, "must be an array klass.");
699 }
700 }
701
702 // rollback processed safepoints
703 while (safepoints_done.length() > 0) {
704 SafePointNode* sfpt_done = safepoints_done.pop();
705 // remove any extra entries we added to the safepoint
706 uint last = sfpt_done->req() - 1;
707 for (int k = 0; k < nfields; k++) {
708 sfpt_done->del_req(last--);
709 }
710 JVMState *jvms = sfpt_done->jvms();
711 jvms->set_endoff(sfpt_done->req());
712 // Now make a pass over the debug information replacing any references
713 // to SafePointScalarObjectNode with the allocated object.
714 int start = jvms->debug_start();
715 int end = jvms->debug_end();
716 for (int i = start; i < end; i++) {
717 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
718 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
719 if (scobj->first_index(jvms) == sfpt_done->req() &&
720 scobj->n_fields() == (uint)nfields) {
721 assert(scobj->alloc() == alloc, "sanity");
722 sfpt_done->set_req(i, res);
723 }
724 }
725 }
726 _igvn._worklist.push(sfpt_done);
727 }
728 }
729
730 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {
731 // Fields of scalar objs are referenced only at the end
732 // of regular debuginfo at the last (youngest) JVMS.
733 // Record relative start index.
734 ciInstanceKlass* iklass = nullptr;
735 BasicType basic_elem_type = T_ILLEGAL;
736 const Type* field_type = nullptr;
737 const TypeOopPtr* res_type = nullptr;
738 int nfields = 0;
739 int array_base = 0;
740 int element_size = 0;
741 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
742 Node* res = alloc->result_cast();
743
744 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
745 assert(sfpt->jvms() != nullptr, "missed JVMS");
746
747 if (res != nullptr) { // Could be null when there are no users
748 res_type = _igvn.type(res)->isa_oopptr();
749
750 if (res_type->isa_instptr()) {
751 // find the fields of the class which will be needed for safepoint debug information
752 iklass = res_type->is_instptr()->instance_klass();
753 nfields = iklass->nof_nonstatic_fields();
754 } else {
755 // find the array's elements which will be needed for safepoint debug information
756 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
757 assert(nfields >= 0, "must be an array klass.");
758 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
759 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
760 element_size = type2aelembytes(basic_elem_type);
761 field_type = res_type->is_aryptr()->elem();
762 }
763 }
764
765 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
766 sobj->init_req(0, C->root());
767 transform_later(sobj);
768
769 // Scan object's fields adding an input to the safepoint for each field.
770 for (int j = 0; j < nfields; j++) {
771 intptr_t offset;
772 ciField* field = nullptr;
773 if (iklass != nullptr) {
774 field = iklass->nonstatic_field_at(j);
775 offset = field->offset_in_bytes();
776 ciType* elem_type = field->type();
777 basic_elem_type = field->layout_type();
778
779 // The next code is taken from Parse::do_get_xxx().
780 if (is_reference_type(basic_elem_type)) {
781 if (!elem_type->is_loaded()) {
782 field_type = TypeInstPtr::BOTTOM;
783 } else if (field != nullptr && field->is_static_constant()) {
784 ciObject* con = field->constant_value().as_object();
785 // Do not "join" in the previous type; it doesn't add value,
786 // and may yield a vacuous result if the field is of interface type.
787 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
788 assert(field_type != nullptr, "field singleton type must be consistent");
789 } else {
790 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
791 }
792 if (UseCompressedOops) {
793 field_type = field_type->make_narrowoop();
794 basic_elem_type = T_NARROWOOP;
795 }
796 } else {
797 field_type = Type::get_const_basic_type(basic_elem_type);
798 }
799 } else {
800 offset = array_base + j * (intptr_t)element_size;
801 }
802
803 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
804
805 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
806
807 // We weren't able to find a value for this field,
808 // give up on eliminating this allocation.
809 if (field_val == nullptr) {
810 uint last = sfpt->req() - 1;
811 for (int k = 0; k < j; k++) {
812 sfpt->del_req(last--);
813 }
814 _igvn._worklist.push(sfpt);
815
816 #ifndef PRODUCT
817 if (PrintEliminateAllocations) {
818 if (field != nullptr) {
819 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
820 field->print();
821 int field_idx = C->get_alias_index(field_addr_type);
822 tty->print(" (alias_idx=%d)", field_idx);
823 } else { // Array's element
824 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
825 }
826 tty->print(", which prevents elimination of: ");
827 if (res == nullptr)
828 alloc->dump();
829 else
830 res->dump();
831 }
832 #endif
833
834 return nullptr;
835 }
836
837 if (UseCompressedOops && field_type->isa_narrowoop()) {
838 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
839 // to be able scalar replace the allocation.
840 if (field_val->is_EncodeP()) {
841 field_val = field_val->in(1);
842 } else {
843 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
844 }
845 }
846 sfpt->add_req(field_val);
847 }
848
849 sfpt->jvms()->set_endoff(sfpt->req());
850
851 return sobj;
852 }
853
854 // Do scalar replacement.
855 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
856 GrowableArray <SafePointNode *> safepoints_done;
857 Node* res = alloc->result_cast();
858 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
859
860 // Process the safepoint uses
861 while (safepoints.length() > 0) {
862 SafePointNode* sfpt = safepoints.pop();
863 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
864
865 if (sobj == nullptr) {
866 undo_previous_scalarizations(safepoints_done, alloc);
867 return false;
868 }
869
870 // Now make a pass over the debug information replacing any references
871 // to the allocated object with "sobj"
872 JVMState *jvms = sfpt->jvms();
873 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
874 _igvn._worklist.push(sfpt);
875
876 // keep it for rollback
877 safepoints_done.append_if_missing(sfpt);
878 }
879
880 return true;
881 }
882
883 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
884 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
885 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
886 if (ctl_proj != nullptr) {
887 igvn.replace_node(ctl_proj, n->in(0));
888 }
889 if (mem_proj != nullptr) {
890 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
891 }
892 }
893
894 // Process users of eliminated allocation.
895 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
896 Node* res = alloc->result_cast();
897 if (res != nullptr) {
898 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
899 Node *use = res->last_out(j);
900 uint oc1 = res->outcnt();
901
902 if (use->is_AddP()) {
903 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
904 Node *n = use->last_out(k);
905 uint oc2 = use->outcnt();
906 if (n->is_Store()) {
907 #ifdef ASSERT
908 // Verify that there is no dependent MemBarVolatile nodes,
909 // they should be removed during IGVN, see MemBarNode::Ideal().
910 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
911 p < pmax; p++) {
912 Node* mb = n->fast_out(p);
913 assert(mb->is_Initialize() || !mb->is_MemBar() ||
914 mb->req() <= MemBarNode::Precedent ||
915 mb->in(MemBarNode::Precedent) != n,
916 "MemBarVolatile should be eliminated for non-escaping object");
917 }
918 #endif
919 _igvn.replace_node(n, n->in(MemNode::Memory));
920 } else {
921 eliminate_gc_barrier(n);
922 }
923 k -= (oc2 - use->outcnt());
924 }
925 _igvn.remove_dead_node(use);
926 } else if (use->is_ArrayCopy()) {
927 // Disconnect ArrayCopy node
928 ArrayCopyNode* ac = use->as_ArrayCopy();
929 if (ac->is_clonebasic()) {
930 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
931 disconnect_projections(ac, _igvn);
932 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
933 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
934 disconnect_projections(membar_before->as_MemBar(), _igvn);
935 if (membar_after->is_MemBar()) {
936 disconnect_projections(membar_after->as_MemBar(), _igvn);
937 }
938 } else {
939 assert(ac->is_arraycopy_validated() ||
940 ac->is_copyof_validated() ||
941 ac->is_copyofrange_validated(), "unsupported");
942 CallProjections callprojs;
943 ac->extract_projections(&callprojs, true);
944
945 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
946 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
947 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
948
949 // Set control to top. IGVN will remove the remaining projections
950 ac->set_req(0, top());
951 ac->replace_edge(res, top(), &_igvn);
952
953 // Disconnect src right away: it can help find new
954 // opportunities for allocation elimination
955 Node* src = ac->in(ArrayCopyNode::Src);
956 ac->replace_edge(src, top(), &_igvn);
957 // src can be top at this point if src and dest of the
958 // arraycopy were the same
959 if (src->outcnt() == 0 && !src->is_top()) {
960 _igvn.remove_dead_node(src);
961 }
962 }
963 _igvn._worklist.push(ac);
964 } else {
965 eliminate_gc_barrier(use);
966 }
967 j -= (oc1 - res->outcnt());
968 }
969 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
970 _igvn.remove_dead_node(res);
971 }
972
973 //
974 // Process other users of allocation's projections
975 //
976 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
977 // First disconnect stores captured by Initialize node.
978 // If Initialize node is eliminated first in the following code,
979 // it will kill such stores and DUIterator_Last will assert.
980 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
981 Node* use = _callprojs.resproj->fast_out(j);
982 if (use->is_AddP()) {
983 // raw memory addresses used only by the initialization
984 _igvn.replace_node(use, C->top());
985 --j; --jmax;
986 }
987 }
988 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
989 Node* use = _callprojs.resproj->last_out(j);
990 uint oc1 = _callprojs.resproj->outcnt();
991 if (use->is_Initialize()) {
992 // Eliminate Initialize node.
993 InitializeNode *init = use->as_Initialize();
994 assert(init->outcnt() <= 2, "only a control and memory projection expected");
995 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
996 if (ctrl_proj != nullptr) {
997 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
998 #ifdef ASSERT
999 // If the InitializeNode has no memory out, it will die, and tmp will become null
1000 Node* tmp = init->in(TypeFunc::Control);
1001 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1002 #endif
1003 }
1004 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1005 if (mem_proj != nullptr) {
1006 Node *mem = init->in(TypeFunc::Memory);
1007 #ifdef ASSERT
1008 if (mem->is_MergeMem()) {
1009 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
1010 } else {
1011 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1012 }
1013 #endif
1014 _igvn.replace_node(mem_proj, mem);
1015 }
1016 } else {
1017 assert(false, "only Initialize or AddP expected");
1018 }
1019 j -= (oc1 - _callprojs.resproj->outcnt());
1020 }
1021 }
1022 if (_callprojs.fallthrough_catchproj != nullptr) {
1023 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1024 }
1025 if (_callprojs.fallthrough_memproj != nullptr) {
1026 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1027 }
1028 if (_callprojs.catchall_memproj != nullptr) {
1029 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1030 }
1031 if (_callprojs.fallthrough_ioproj != nullptr) {
1032 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1033 }
1034 if (_callprojs.catchall_ioproj != nullptr) {
1035 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1036 }
1037 if (_callprojs.catchall_catchproj != nullptr) {
1038 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1039 }
1040 }
1041
1042 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1043 // If reallocation fails during deoptimization we'll pop all
1044 // interpreter frames for this compiled frame and that won't play
1045 // nice with JVMTI popframe.
1046 // We avoid this issue by eager reallocation when the popframe request
1047 // is received.
1048 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1049 return false;
1050 }
1051 Node* klass = alloc->in(AllocateNode::KlassNode);
1052 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1053 Node* res = alloc->result_cast();
1054 // Eliminate boxing allocations which are not used
1055 // regardless scalar replaceable status.
1056 bool boxing_alloc = C->eliminate_boxing() &&
1057 tklass->isa_instklassptr() &&
1058 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1059 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1060 return false;
1061 }
1062
1063 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1064
1065 GrowableArray <SafePointNode *> safepoints;
1066 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1067 return false;
1068 }
1069
1070 if (!alloc->_is_scalar_replaceable) {
1071 assert(res == nullptr, "sanity");
1072 // We can only eliminate allocation if all debug info references
1073 // are already replaced with SafePointScalarObject because
1074 // we can't search for a fields value without instance_id.
1075 if (safepoints.length() > 0) {
1076 return false;
1077 }
1078 }
1079
1080 if (!scalar_replacement(alloc, safepoints)) {
1081 return false;
1082 }
1083
1084 CompileLog* log = C->log();
1085 if (log != nullptr) {
1086 log->head("eliminate_allocation type='%d'",
1087 log->identify(tklass->exact_klass()));
1088 JVMState* p = alloc->jvms();
1089 while (p != nullptr) {
1090 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1091 p = p->caller();
1092 }
1093 log->tail("eliminate_allocation");
1094 }
1095
1096 process_users_of_allocation(alloc);
1097
1098 #ifndef PRODUCT
1099 if (PrintEliminateAllocations) {
1100 if (alloc->is_AllocateArray())
1101 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1102 else
1103 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1104 }
1105 #endif
1106
1107 return true;
1108 }
1109
1110 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1111 // EA should remove all uses of non-escaping boxing node.
1112 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1113 return false;
1114 }
1115
1116 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1117
1118 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1119
1120 const TypeTuple* r = boxing->tf()->range();
1121 assert(r->cnt() > TypeFunc::Parms, "sanity");
1122 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1123 assert(t != nullptr, "sanity");
1124
1125 CompileLog* log = C->log();
1126 if (log != nullptr) {
1127 log->head("eliminate_boxing type='%d'",
1128 log->identify(t->instance_klass()));
1129 JVMState* p = boxing->jvms();
1130 while (p != nullptr) {
1131 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1132 p = p->caller();
1133 }
1134 log->tail("eliminate_boxing");
1135 }
1136
1137 process_users_of_allocation(boxing);
1138
1139 #ifndef PRODUCT
1140 if (PrintEliminateAllocations) {
1141 tty->print("++++ Eliminated: %d ", boxing->_idx);
1142 boxing->method()->print_short_name(tty);
1143 tty->cr();
1144 }
1145 #endif
1146
1147 return true;
1148 }
1149
1150
1151 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1152 Node* adr = basic_plus_adr(base, offset);
1153 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1154 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1155 transform_later(value);
1156 return value;
1157 }
1158
1159
1160 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1161 Node* adr = basic_plus_adr(base, offset);
1162 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1163 transform_later(mem);
1164 return mem;
1165 }
1166
1167 //=============================================================================
1168 //
1169 // A L L O C A T I O N
1170 //
1171 // Allocation attempts to be fast in the case of frequent small objects.
1172 // It breaks down like this:
1173 //
1174 // 1) Size in doublewords is computed. This is a constant for objects and
1175 // variable for most arrays. Doubleword units are used to avoid size
1176 // overflow of huge doubleword arrays. We need doublewords in the end for
1177 // rounding.
1178 //
1179 // 2) Size is checked for being 'too large'. Too-large allocations will go
1180 // the slow path into the VM. The slow path can throw any required
1181 // exceptions, and does all the special checks for very large arrays. The
1182 // size test can constant-fold away for objects. For objects with
1183 // finalizers it constant-folds the otherway: you always go slow with
1184 // finalizers.
1185 //
1186 // 3) If NOT using TLABs, this is the contended loop-back point.
1187 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1188 //
1189 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1190 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1191 // "size*8" we always enter the VM, where "largish" is a constant picked small
1192 // enough that there's always space between the eden max and 4Gig (old space is
1193 // there so it's quite large) and large enough that the cost of entering the VM
1194 // is dwarfed by the cost to initialize the space.
1195 //
1196 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1197 // down. If contended, repeat at step 3. If using TLABs normal-store
1198 // adjusted heap top back down; there is no contention.
1199 //
1200 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1201 // fields.
1202 //
1203 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1204 // oop flavor.
1205 //
1206 //=============================================================================
1207 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1208 // Allocations bigger than this always go the slow route.
1209 // This value must be small enough that allocation attempts that need to
1210 // trigger exceptions go the slow route. Also, it must be small enough so
1211 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1212 //=============================================================================j//
1213 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1214 // The allocator will coalesce int->oop copies away. See comment in
1215 // coalesce.cpp about how this works. It depends critically on the exact
1216 // code shape produced here, so if you are changing this code shape
1217 // make sure the GC info for the heap-top is correct in and around the
1218 // slow-path call.
1219 //
1220
1221 void PhaseMacroExpand::expand_allocate_common(
1222 AllocateNode* alloc, // allocation node to be expanded
1223 Node* length, // array length for an array allocation
1224 const TypeFunc* slow_call_type, // Type of slow call
1225 address slow_call_address, // Address of slow call
1226 Node* valid_length_test // whether length is valid or not
1227 )
1228 {
1229 Node* ctrl = alloc->in(TypeFunc::Control);
1230 Node* mem = alloc->in(TypeFunc::Memory);
1231 Node* i_o = alloc->in(TypeFunc::I_O);
1232 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1233 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1234 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1235 assert(ctrl != nullptr, "must have control");
1236
1237 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1238 // they will not be used if "always_slow" is set
1239 enum { slow_result_path = 1, fast_result_path = 2 };
1240 Node *result_region = nullptr;
1241 Node *result_phi_rawmem = nullptr;
1242 Node *result_phi_rawoop = nullptr;
1243 Node *result_phi_i_o = nullptr;
1244
1245 // The initial slow comparison is a size check, the comparison
1246 // we want to do is a BoolTest::gt
1247 bool expand_fast_path = true;
1248 int tv = _igvn.find_int_con(initial_slow_test, -1);
1249 if (tv >= 0) {
1250 // InitialTest has constant result
1251 // 0 - can fit in TLAB
1252 // 1 - always too big or negative
1253 assert(tv <= 1, "0 or 1 if a constant");
1254 expand_fast_path = (tv == 0);
1255 initial_slow_test = nullptr;
1256 } else {
1257 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1258 }
1259
1260 if (!UseTLAB) {
1261 // Force slow-path allocation
1262 expand_fast_path = false;
1263 initial_slow_test = nullptr;
1264 }
1265
1266 bool allocation_has_use = (alloc->result_cast() != nullptr);
1267 if (!allocation_has_use) {
1268 InitializeNode* init = alloc->initialization();
1269 if (init != nullptr) {
1270 init->remove(&_igvn);
1271 }
1272 if (expand_fast_path && (initial_slow_test == nullptr)) {
1273 // Remove allocation node and return.
1274 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1275 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1276 #ifndef PRODUCT
1277 if (PrintEliminateAllocations) {
1278 tty->print("NotUsed ");
1279 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1280 if (res != nullptr) {
1281 res->dump();
1282 } else {
1283 alloc->dump();
1284 }
1285 }
1286 #endif
1287 yank_alloc_node(alloc);
1288 return;
1289 }
1290 }
1291
1292 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1293 Node *slow_region = nullptr;
1294 Node *toobig_false = ctrl;
1295
1296 // generate the initial test if necessary
1297 if (initial_slow_test != nullptr ) {
1298 assert (expand_fast_path, "Only need test if there is a fast path");
1299 slow_region = new RegionNode(3);
1300
1301 // Now make the initial failure test. Usually a too-big test but
1302 // might be a TRUE for finalizers.
1303 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1304 transform_later(toobig_iff);
1305 // Plug the failing-too-big test into the slow-path region
1306 Node *toobig_true = new IfTrueNode( toobig_iff );
1307 transform_later(toobig_true);
1308 slow_region ->init_req( too_big_or_final_path, toobig_true );
1309 toobig_false = new IfFalseNode( toobig_iff );
1310 transform_later(toobig_false);
1311 } else {
1312 // No initial test, just fall into next case
1313 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1314 toobig_false = ctrl;
1315 DEBUG_ONLY(slow_region = NodeSentinel);
1316 }
1317
1318 // If we are here there are several possibilities
1319 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1320 // no_initial_check means a constant allocation.
1321 // - If check always evaluates to false -> expand_fast_path is false (see above)
1322 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1323 // if !allocation_has_use the fast path is empty
1324 // if !allocation_has_use && no_initial_check
1325 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1326 // removed by yank_alloc_node above.
1327
1328 Node *slow_mem = mem; // save the current memory state for slow path
1329 // generate the fast allocation code unless we know that the initial test will always go slow
1330 if (expand_fast_path) {
1331 // Fast path modifies only raw memory.
1332 if (mem->is_MergeMem()) {
1333 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1334 }
1335
1336 // allocate the Region and Phi nodes for the result
1337 result_region = new RegionNode(3);
1338 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1339 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1340
1341 // Grab regular I/O before optional prefetch may change it.
1342 // Slow-path does no I/O so just set it to the original I/O.
1343 result_phi_i_o->init_req(slow_result_path, i_o);
1344
1345 // Name successful fast-path variables
1346 Node* fast_oop_ctrl;
1347 Node* fast_oop_rawmem;
1348 if (allocation_has_use) {
1349 Node* needgc_ctrl = nullptr;
1350 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1351
1352 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1353 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1354 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1355 fast_oop_ctrl, fast_oop_rawmem,
1356 prefetch_lines);
1357
1358 if (initial_slow_test != nullptr) {
1359 // This completes all paths into the slow merge point
1360 slow_region->init_req(need_gc_path, needgc_ctrl);
1361 transform_later(slow_region);
1362 } else {
1363 // No initial slow path needed!
1364 // Just fall from the need-GC path straight into the VM call.
1365 slow_region = needgc_ctrl;
1366 }
1367
1368 InitializeNode* init = alloc->initialization();
1369 fast_oop_rawmem = initialize_object(alloc,
1370 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1371 klass_node, length, size_in_bytes);
1372 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1373 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1374
1375 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1376 } else {
1377 assert (initial_slow_test != nullptr, "sanity");
1378 fast_oop_ctrl = toobig_false;
1379 fast_oop_rawmem = mem;
1380 transform_later(slow_region);
1381 }
1382
1383 // Plug in the successful fast-path into the result merge point
1384 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1385 result_phi_i_o ->init_req(fast_result_path, i_o);
1386 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1387 } else {
1388 slow_region = ctrl;
1389 result_phi_i_o = i_o; // Rename it to use in the following code.
1390 }
1391
1392 // Generate slow-path call
1393 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1394 OptoRuntime::stub_name(slow_call_address),
1395 TypePtr::BOTTOM);
1396 call->init_req(TypeFunc::Control, slow_region);
1397 call->init_req(TypeFunc::I_O, top()); // does no i/o
1398 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1399 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1400 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1401
1402 call->init_req(TypeFunc::Parms+0, klass_node);
1403 if (length != nullptr) {
1404 call->init_req(TypeFunc::Parms+1, length);
1405 }
1406
1407 // Copy debug information and adjust JVMState information, then replace
1408 // allocate node with the call
1409 call->copy_call_debug_info(&_igvn, alloc);
1410 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1411 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1412 // path dies).
1413 if (valid_length_test != nullptr) {
1414 call->add_req(valid_length_test);
1415 }
1416 if (expand_fast_path) {
1417 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1418 } else {
1419 // Hook i_o projection to avoid its elimination during allocation
1420 // replacement (when only a slow call is generated).
1421 call->set_req(TypeFunc::I_O, result_phi_i_o);
1422 }
1423 _igvn.replace_node(alloc, call);
1424 transform_later(call);
1425
1426 // Identify the output projections from the allocate node and
1427 // adjust any references to them.
1428 // The control and io projections look like:
1429 //
1430 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1431 // Allocate Catch
1432 // ^---Proj(io) <-------+ ^---CatchProj(io)
1433 //
1434 // We are interested in the CatchProj nodes.
1435 //
1436 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1437
1438 // An allocate node has separate memory projections for the uses on
1439 // the control and i_o paths. Replace the control memory projection with
1440 // result_phi_rawmem (unless we are only generating a slow call when
1441 // both memory projections are combined)
1442 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1443 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1444 }
1445 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1446 // catchall_memproj so we end up with a call that has only 1 memory projection.
1447 if (_callprojs.catchall_memproj != nullptr ) {
1448 if (_callprojs.fallthrough_memproj == nullptr) {
1449 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1450 transform_later(_callprojs.fallthrough_memproj);
1451 }
1452 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1453 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1454 }
1455
1456 // An allocate node has separate i_o projections for the uses on the control
1457 // and i_o paths. Always replace the control i_o projection with result i_o
1458 // otherwise incoming i_o become dead when only a slow call is generated
1459 // (it is different from memory projections where both projections are
1460 // combined in such case).
1461 if (_callprojs.fallthrough_ioproj != nullptr) {
1462 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1463 }
1464 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1465 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1466 if (_callprojs.catchall_ioproj != nullptr ) {
1467 if (_callprojs.fallthrough_ioproj == nullptr) {
1468 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1469 transform_later(_callprojs.fallthrough_ioproj);
1470 }
1471 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1472 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1473 }
1474
1475 // if we generated only a slow call, we are done
1476 if (!expand_fast_path) {
1477 // Now we can unhook i_o.
1478 if (result_phi_i_o->outcnt() > 1) {
1479 call->set_req(TypeFunc::I_O, top());
1480 } else {
1481 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1482 // Case of new array with negative size known during compilation.
1483 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1484 // following code since call to runtime will throw exception.
1485 // As result there will be no users of i_o after the call.
1486 // Leave i_o attached to this call to avoid problems in preceding graph.
1487 }
1488 return;
1489 }
1490
1491 if (_callprojs.fallthrough_catchproj != nullptr) {
1492 ctrl = _callprojs.fallthrough_catchproj->clone();
1493 transform_later(ctrl);
1494 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1495 } else {
1496 ctrl = top();
1497 }
1498 Node *slow_result;
1499 if (_callprojs.resproj == nullptr) {
1500 // no uses of the allocation result
1501 slow_result = top();
1502 } else {
1503 slow_result = _callprojs.resproj->clone();
1504 transform_later(slow_result);
1505 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1506 }
1507
1508 // Plug slow-path into result merge point
1509 result_region->init_req( slow_result_path, ctrl);
1510 transform_later(result_region);
1511 if (allocation_has_use) {
1512 result_phi_rawoop->init_req(slow_result_path, slow_result);
1513 transform_later(result_phi_rawoop);
1514 }
1515 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1516 transform_later(result_phi_rawmem);
1517 transform_later(result_phi_i_o);
1518 // This completes all paths into the result merge point
1519 }
1520
1521 // Remove alloc node that has no uses.
1522 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1523 Node* ctrl = alloc->in(TypeFunc::Control);
1524 Node* mem = alloc->in(TypeFunc::Memory);
1525 Node* i_o = alloc->in(TypeFunc::I_O);
1526
1527 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1528 if (_callprojs.resproj != nullptr) {
1529 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1530 Node* use = _callprojs.resproj->fast_out(i);
1531 use->isa_MemBar()->remove(&_igvn);
1532 --imax;
1533 --i; // back up iterator
1534 }
1535 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1536 _igvn.remove_dead_node(_callprojs.resproj);
1537 }
1538 if (_callprojs.fallthrough_catchproj != nullptr) {
1539 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1540 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1541 }
1542 if (_callprojs.catchall_catchproj != nullptr) {
1543 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1544 _callprojs.catchall_catchproj->set_req(0, top());
1545 }
1546 if (_callprojs.fallthrough_proj != nullptr) {
1547 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1548 _igvn.remove_dead_node(catchnode);
1549 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1550 }
1551 if (_callprojs.fallthrough_memproj != nullptr) {
1552 migrate_outs(_callprojs.fallthrough_memproj, mem);
1553 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1554 }
1555 if (_callprojs.fallthrough_ioproj != nullptr) {
1556 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1557 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1558 }
1559 if (_callprojs.catchall_memproj != nullptr) {
1560 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1561 _callprojs.catchall_memproj->set_req(0, top());
1562 }
1563 if (_callprojs.catchall_ioproj != nullptr) {
1564 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1565 _callprojs.catchall_ioproj->set_req(0, top());
1566 }
1567 #ifndef PRODUCT
1568 if (PrintEliminateAllocations) {
1569 if (alloc->is_AllocateArray()) {
1570 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1571 } else {
1572 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1573 }
1574 }
1575 #endif
1576 _igvn.remove_dead_node(alloc);
1577 }
1578
1579 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1580 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1581 // If initialization is performed by an array copy, any required
1582 // MemBarStoreStore was already added. If the object does not
1583 // escape no need for a MemBarStoreStore. If the object does not
1584 // escape in its initializer and memory barrier (MemBarStoreStore or
1585 // stronger) is already added at exit of initializer, also no need
1586 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1587 // so that stores that initialize this object can't be reordered
1588 // with a subsequent store that makes this object accessible by
1589 // other threads.
1590 // Other threads include java threads and JVM internal threads
1591 // (for example concurrent GC threads). Current concurrent GC
1592 // implementation: G1 will not scan newly created object,
1593 // so it's safe to skip storestore barrier when allocation does
1594 // not escape.
1595 if (!alloc->does_not_escape_thread() &&
1596 !alloc->is_allocation_MemBar_redundant() &&
1597 (init == nullptr || !init->is_complete_with_arraycopy())) {
1598 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1599 // No InitializeNode or no stores captured by zeroing
1600 // elimination. Simply add the MemBarStoreStore after object
1601 // initialization.
1602 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1603 transform_later(mb);
1604
1605 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1606 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1607 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1608 transform_later(fast_oop_ctrl);
1609 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1610 transform_later(fast_oop_rawmem);
1611 } else {
1612 // Add the MemBarStoreStore after the InitializeNode so that
1613 // all stores performing the initialization that were moved
1614 // before the InitializeNode happen before the storestore
1615 // barrier.
1616
1617 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1618 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1619
1620 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1621 transform_later(mb);
1622
1623 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1624 transform_later(ctrl);
1625 Node* mem = new ProjNode(init, TypeFunc::Memory);
1626 transform_later(mem);
1627
1628 // The MemBarStoreStore depends on control and memory coming
1629 // from the InitializeNode
1630 mb->init_req(TypeFunc::Memory, mem);
1631 mb->init_req(TypeFunc::Control, ctrl);
1632
1633 ctrl = new ProjNode(mb, TypeFunc::Control);
1634 transform_later(ctrl);
1635 mem = new ProjNode(mb, TypeFunc::Memory);
1636 transform_later(mem);
1637
1638 // All nodes that depended on the InitializeNode for control
1639 // and memory must now depend on the MemBarNode that itself
1640 // depends on the InitializeNode
1641 if (init_ctrl != nullptr) {
1642 _igvn.replace_node(init_ctrl, ctrl);
1643 }
1644 if (init_mem != nullptr) {
1645 _igvn.replace_node(init_mem, mem);
1646 }
1647 }
1648 }
1649 }
1650
1651 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1652 Node*& ctrl, Node*& rawmem) {
1653 if (C->env()->dtrace_alloc_probes()) {
1654 // Slow-path call
1655 int size = TypeFunc::Parms + 2;
1656 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1657 CAST_FROM_FN_PTR(address,
1658 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1659 "dtrace_object_alloc",
1660 TypeRawPtr::BOTTOM);
1661
1662 // Get base of thread-local storage area
1663 Node* thread = new ThreadLocalNode();
1664 transform_later(thread);
1665
1666 call->init_req(TypeFunc::Parms + 0, thread);
1667 call->init_req(TypeFunc::Parms + 1, oop);
1668 call->init_req(TypeFunc::Control, ctrl);
1669 call->init_req(TypeFunc::I_O , top()); // does no i/o
1670 call->init_req(TypeFunc::Memory , rawmem);
1671 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1672 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1673 transform_later(call);
1674 ctrl = new ProjNode(call, TypeFunc::Control);
1675 transform_later(ctrl);
1676 rawmem = new ProjNode(call, TypeFunc::Memory);
1677 transform_later(rawmem);
1678 }
1679 }
1680
1681 // Helper for PhaseMacroExpand::expand_allocate_common.
1682 // Initializes the newly-allocated storage.
1683 Node*
1684 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1685 Node* control, Node* rawmem, Node* object,
1686 Node* klass_node, Node* length,
1687 Node* size_in_bytes) {
1688 InitializeNode* init = alloc->initialization();
1689 // Store the klass & mark bits
1690 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1691 if (!mark_node->is_Con()) {
1692 transform_later(mark_node);
1693 }
1694 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1695
1696 if (!UseCompactObjectHeaders) {
1697 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1698 }
1699 int header_size = alloc->minimum_header_size(); // conservatively small
1700
1701 // Array length
1702 if (length != nullptr) { // Arrays need length field
1703 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1704 // conservatively small header size:
1705 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1706 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1707 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1708 if (is_reference_type(elem, true)) {
1709 elem = T_OBJECT;
1710 }
1711 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1712 }
1713 }
1714
1715 // Clear the object body, if necessary.
1716 if (init == nullptr) {
1717 // The init has somehow disappeared; be cautious and clear everything.
1718 //
1719 // This can happen if a node is allocated but an uncommon trap occurs
1720 // immediately. In this case, the Initialize gets associated with the
1721 // trap, and may be placed in a different (outer) loop, if the Allocate
1722 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1723 // there can be two Allocates to one Initialize. The answer in all these
1724 // edge cases is safety first. It is always safe to clear immediately
1725 // within an Allocate, and then (maybe or maybe not) clear some more later.
1726 if (!(UseTLAB && ZeroTLAB)) {
1727 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1728 header_size, size_in_bytes,
1729 &_igvn);
1730 }
1731 } else {
1732 if (!init->is_complete()) {
1733 // Try to win by zeroing only what the init does not store.
1734 // We can also try to do some peephole optimizations,
1735 // such as combining some adjacent subword stores.
1736 rawmem = init->complete_stores(control, rawmem, object,
1737 header_size, size_in_bytes, &_igvn);
1738 }
1739 // We have no more use for this link, since the AllocateNode goes away:
1740 init->set_req(InitializeNode::RawAddress, top());
1741 // (If we keep the link, it just confuses the register allocator,
1742 // who thinks he sees a real use of the address by the membar.)
1743 }
1744
1745 return rawmem;
1746 }
1747
1748 // Generate prefetch instructions for next allocations.
1749 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1750 Node*& contended_phi_rawmem,
1751 Node* old_eden_top, Node* new_eden_top,
1752 intx lines) {
1753 enum { fall_in_path = 1, pf_path = 2 };
1754 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1755 // Generate prefetch allocation with watermark check.
1756 // As an allocation hits the watermark, we will prefetch starting
1757 // at a "distance" away from watermark.
1758
1759 Node *pf_region = new RegionNode(3);
1760 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1761 TypeRawPtr::BOTTOM );
1762 // I/O is used for Prefetch
1763 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1764
1765 Node *thread = new ThreadLocalNode();
1766 transform_later(thread);
1767
1768 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1769 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1770 transform_later(eden_pf_adr);
1771
1772 Node *old_pf_wm = new LoadPNode(needgc_false,
1773 contended_phi_rawmem, eden_pf_adr,
1774 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1775 MemNode::unordered);
1776 transform_later(old_pf_wm);
1777
1778 // check against new_eden_top
1779 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1780 transform_later(need_pf_cmp);
1781 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1782 transform_later(need_pf_bol);
1783 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1784 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1785 transform_later(need_pf_iff);
1786
1787 // true node, add prefetchdistance
1788 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1789 transform_later(need_pf_true);
1790
1791 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1792 transform_later(need_pf_false);
1793
1794 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1795 _igvn.MakeConX(AllocatePrefetchDistance) );
1796 transform_later(new_pf_wmt );
1797 new_pf_wmt->set_req(0, need_pf_true);
1798
1799 Node *store_new_wmt = new StorePNode(need_pf_true,
1800 contended_phi_rawmem, eden_pf_adr,
1801 TypeRawPtr::BOTTOM, new_pf_wmt,
1802 MemNode::unordered);
1803 transform_later(store_new_wmt);
1804
1805 // adding prefetches
1806 pf_phi_abio->init_req( fall_in_path, i_o );
1807
1808 Node *prefetch_adr;
1809 Node *prefetch;
1810 uint step_size = AllocatePrefetchStepSize;
1811 uint distance = 0;
1812
1813 for ( intx i = 0; i < lines; i++ ) {
1814 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1815 _igvn.MakeConX(distance) );
1816 transform_later(prefetch_adr);
1817 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1818 transform_later(prefetch);
1819 distance += step_size;
1820 i_o = prefetch;
1821 }
1822 pf_phi_abio->set_req( pf_path, i_o );
1823
1824 pf_region->init_req( fall_in_path, need_pf_false );
1825 pf_region->init_req( pf_path, need_pf_true );
1826
1827 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1828 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1829
1830 transform_later(pf_region);
1831 transform_later(pf_phi_rawmem);
1832 transform_later(pf_phi_abio);
1833
1834 needgc_false = pf_region;
1835 contended_phi_rawmem = pf_phi_rawmem;
1836 i_o = pf_phi_abio;
1837 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1838 // Insert a prefetch instruction for each allocation.
1839 // This code is used to generate 1 prefetch instruction per cache line.
1840
1841 // Generate several prefetch instructions.
1842 uint step_size = AllocatePrefetchStepSize;
1843 uint distance = AllocatePrefetchDistance;
1844
1845 // Next cache address.
1846 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1847 _igvn.MakeConX(step_size + distance));
1848 transform_later(cache_adr);
1849 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1850 transform_later(cache_adr);
1851 // Address is aligned to execute prefetch to the beginning of cache line size
1852 // (it is important when BIS instruction is used on SPARC as prefetch).
1853 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1854 cache_adr = new AndXNode(cache_adr, mask);
1855 transform_later(cache_adr);
1856 cache_adr = new CastX2PNode(cache_adr);
1857 transform_later(cache_adr);
1858
1859 // Prefetch
1860 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1861 prefetch->set_req(0, needgc_false);
1862 transform_later(prefetch);
1863 contended_phi_rawmem = prefetch;
1864 Node *prefetch_adr;
1865 distance = step_size;
1866 for ( intx i = 1; i < lines; i++ ) {
1867 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1868 _igvn.MakeConX(distance) );
1869 transform_later(prefetch_adr);
1870 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1871 transform_later(prefetch);
1872 distance += step_size;
1873 contended_phi_rawmem = prefetch;
1874 }
1875 } else if( AllocatePrefetchStyle > 0 ) {
1876 // Insert a prefetch for each allocation only on the fast-path
1877 Node *prefetch_adr;
1878 Node *prefetch;
1879 // Generate several prefetch instructions.
1880 uint step_size = AllocatePrefetchStepSize;
1881 uint distance = AllocatePrefetchDistance;
1882 for ( intx i = 0; i < lines; i++ ) {
1883 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1884 _igvn.MakeConX(distance) );
1885 transform_later(prefetch_adr);
1886 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1887 // Do not let it float too high, since if eden_top == eden_end,
1888 // both might be null.
1889 if( i == 0 ) { // Set control for first prefetch, next follows it
1890 prefetch->init_req(0, needgc_false);
1891 }
1892 transform_later(prefetch);
1893 distance += step_size;
1894 i_o = prefetch;
1895 }
1896 }
1897 return i_o;
1898 }
1899
1900
1901 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1902 expand_allocate_common(alloc, nullptr,
1903 OptoRuntime::new_instance_Type(),
1904 OptoRuntime::new_instance_Java(), nullptr);
1905 }
1906
1907 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1908 Node* length = alloc->in(AllocateNode::ALength);
1909 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1910 InitializeNode* init = alloc->initialization();
1911 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1912 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1913 address slow_call_address; // Address of slow call
1914 if (init != nullptr && init->is_complete_with_arraycopy() &&
1915 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
1916 // Don't zero type array during slow allocation in VM since
1917 // it will be initialized later by arraycopy in compiled code.
1918 slow_call_address = OptoRuntime::new_array_nozero_Java();
1919 } else {
1920 slow_call_address = OptoRuntime::new_array_Java();
1921 }
1922 expand_allocate_common(alloc, length,
1923 OptoRuntime::new_array_Type(),
1924 slow_call_address, valid_length_test);
1925 }
1926
1927 //-------------------mark_eliminated_box----------------------------------
1928 //
1929 // During EA obj may point to several objects but after few ideal graph
1930 // transformations (CCP) it may point to only one non escaping object
1931 // (but still using phi), corresponding locks and unlocks will be marked
1932 // for elimination. Later obj could be replaced with a new node (new phi)
1933 // and which does not have escape information. And later after some graph
1934 // reshape other locks and unlocks (which were not marked for elimination
1935 // before) are connected to this new obj (phi) but they still will not be
1936 // marked for elimination since new obj has no escape information.
1937 // Mark all associated (same box and obj) lock and unlock nodes for
1938 // elimination if some of them marked already.
1939 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
1940 BoxLockNode* oldbox = box->as_BoxLock();
1941 if (oldbox->is_eliminated()) {
1942 return; // This BoxLock node was processed already.
1943 }
1944 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
1945 // New implementation (EliminateNestedLocks) has separate BoxLock
1946 // node for each locked region so mark all associated locks/unlocks as
1947 // eliminated even if different objects are referenced in one locked region
1948 // (for example, OSR compilation of nested loop inside locked scope).
1949 if (EliminateNestedLocks ||
1950 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
1951 // Box is used only in one lock region. Mark this box as eliminated.
1952 oldbox->set_local(); // This verifies correct state of BoxLock
1953 _igvn.hash_delete(oldbox);
1954 oldbox->set_eliminated(); // This changes box's hash value
1955 _igvn.hash_insert(oldbox);
1956
1957 for (uint i = 0; i < oldbox->outcnt(); i++) {
1958 Node* u = oldbox->raw_out(i);
1959 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1960 AbstractLockNode* alock = u->as_AbstractLock();
1961 // Check lock's box since box could be referenced by Lock's debug info.
1962 if (alock->box_node() == oldbox) {
1963 // Mark eliminated all related locks and unlocks.
1964 #ifdef ASSERT
1965 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1966 #endif
1967 alock->set_non_esc_obj();
1968 }
1969 }
1970 }
1971 return;
1972 }
1973
1974 // Create new "eliminated" BoxLock node and use it in monitor debug info
1975 // instead of oldbox for the same object.
1976 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1977
1978 // Note: BoxLock node is marked eliminated only here and it is used
1979 // to indicate that all associated lock and unlock nodes are marked
1980 // for elimination.
1981 newbox->set_local(); // This verifies correct state of BoxLock
1982 newbox->set_eliminated();
1983 transform_later(newbox);
1984
1985 // Replace old box node with new box for all users of the same object.
1986 for (uint i = 0; i < oldbox->outcnt();) {
1987 bool next_edge = true;
1988
1989 Node* u = oldbox->raw_out(i);
1990 if (u->is_AbstractLock()) {
1991 AbstractLockNode* alock = u->as_AbstractLock();
1992 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1993 // Replace Box and mark eliminated all related locks and unlocks.
1994 #ifdef ASSERT
1995 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1996 #endif
1997 alock->set_non_esc_obj();
1998 _igvn.rehash_node_delayed(alock);
1999 alock->set_box_node(newbox);
2000 next_edge = false;
2001 }
2002 }
2003 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2004 FastLockNode* flock = u->as_FastLock();
2005 assert(flock->box_node() == oldbox, "sanity");
2006 _igvn.rehash_node_delayed(flock);
2007 flock->set_box_node(newbox);
2008 next_edge = false;
2009 }
2010
2011 // Replace old box in monitor debug info.
2012 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2013 SafePointNode* sfn = u->as_SafePoint();
2014 JVMState* youngest_jvms = sfn->jvms();
2015 int max_depth = youngest_jvms->depth();
2016 for (int depth = 1; depth <= max_depth; depth++) {
2017 JVMState* jvms = youngest_jvms->of_depth(depth);
2018 int num_mon = jvms->nof_monitors();
2019 // Loop over monitors
2020 for (int idx = 0; idx < num_mon; idx++) {
2021 Node* obj_node = sfn->monitor_obj(jvms, idx);
2022 Node* box_node = sfn->monitor_box(jvms, idx);
2023 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2024 int j = jvms->monitor_box_offset(idx);
2025 _igvn.replace_input_of(u, j, newbox);
2026 next_edge = false;
2027 }
2028 }
2029 }
2030 }
2031 if (next_edge) i++;
2032 }
2033 }
2034
2035 //-----------------------mark_eliminated_locking_nodes-----------------------
2036 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2037 if (!alock->is_balanced()) {
2038 return; // Can't do any more elimination for this locking region
2039 }
2040 if (EliminateNestedLocks) {
2041 if (alock->is_nested()) {
2042 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2043 return;
2044 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2045 // Only Lock node has JVMState needed here.
2046 // Not that preceding claim is documented anywhere else.
2047 if (alock->jvms() != nullptr) {
2048 if (alock->as_Lock()->is_nested_lock_region()) {
2049 // Mark eliminated related nested locks and unlocks.
2050 Node* obj = alock->obj_node();
2051 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2052 assert(!box_node->is_eliminated(), "should not be marked yet");
2053 // Note: BoxLock node is marked eliminated only here
2054 // and it is used to indicate that all associated lock
2055 // and unlock nodes are marked for elimination.
2056 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2057 for (uint i = 0; i < box_node->outcnt(); i++) {
2058 Node* u = box_node->raw_out(i);
2059 if (u->is_AbstractLock()) {
2060 alock = u->as_AbstractLock();
2061 if (alock->box_node() == box_node) {
2062 // Verify that this Box is referenced only by related locks.
2063 assert(alock->obj_node()->eqv_uncast(obj), "");
2064 // Mark all related locks and unlocks.
2065 #ifdef ASSERT
2066 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2067 #endif
2068 alock->set_nested();
2069 }
2070 }
2071 }
2072 } else {
2073 #ifdef ASSERT
2074 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2075 if (C->log() != nullptr)
2076 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2077 #endif
2078 }
2079 }
2080 return;
2081 }
2082 // Process locks for non escaping object
2083 assert(alock->is_non_esc_obj(), "");
2084 } // EliminateNestedLocks
2085
2086 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2087 // Look for all locks of this object and mark them and
2088 // corresponding BoxLock nodes as eliminated.
2089 Node* obj = alock->obj_node();
2090 for (uint j = 0; j < obj->outcnt(); j++) {
2091 Node* o = obj->raw_out(j);
2092 if (o->is_AbstractLock() &&
2093 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2094 alock = o->as_AbstractLock();
2095 Node* box = alock->box_node();
2096 // Replace old box node with new eliminated box for all users
2097 // of the same object and mark related locks as eliminated.
2098 mark_eliminated_box(box, obj);
2099 }
2100 }
2101 }
2102 }
2103
2104 // we have determined that this lock/unlock can be eliminated, we simply
2105 // eliminate the node without expanding it.
2106 //
2107 // Note: The membar's associated with the lock/unlock are currently not
2108 // eliminated. This should be investigated as a future enhancement.
2109 //
2110 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2111
2112 if (!alock->is_eliminated()) {
2113 return false;
2114 }
2115 #ifdef ASSERT
2116 if (!alock->is_coarsened()) {
2117 // Check that new "eliminated" BoxLock node is created.
2118 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2119 assert(oldbox->is_eliminated(), "should be done already");
2120 }
2121 #endif
2122
2123 alock->log_lock_optimization(C, "eliminate_lock");
2124
2125 #ifndef PRODUCT
2126 if (PrintEliminateLocks) {
2127 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2128 }
2129 #endif
2130
2131 Node* mem = alock->in(TypeFunc::Memory);
2132 Node* ctrl = alock->in(TypeFunc::Control);
2133 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2134
2135 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2136 // There are 2 projections from the lock. The lock node will
2137 // be deleted when its last use is subsumed below.
2138 assert(alock->outcnt() == 2 &&
2139 _callprojs.fallthrough_proj != nullptr &&
2140 _callprojs.fallthrough_memproj != nullptr,
2141 "Unexpected projections from Lock/Unlock");
2142
2143 Node* fallthroughproj = _callprojs.fallthrough_proj;
2144 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2145
2146 // The memory projection from a lock/unlock is RawMem
2147 // The input to a Lock is merged memory, so extract its RawMem input
2148 // (unless the MergeMem has been optimized away.)
2149 if (alock->is_Lock()) {
2150 // Search for MemBarAcquireLock node and delete it also.
2151 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2152 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2153 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2154 Node* memproj = membar->proj_out(TypeFunc::Memory);
2155 _igvn.replace_node(ctrlproj, fallthroughproj);
2156 _igvn.replace_node(memproj, memproj_fallthrough);
2157
2158 // Delete FastLock node also if this Lock node is unique user
2159 // (a loop peeling may clone a Lock node).
2160 Node* flock = alock->as_Lock()->fastlock_node();
2161 if (flock->outcnt() == 1) {
2162 assert(flock->unique_out() == alock, "sanity");
2163 _igvn.replace_node(flock, top());
2164 }
2165 }
2166
2167 // Search for MemBarReleaseLock node and delete it also.
2168 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2169 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2170 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2171 mem->is_Proj() && membar == mem->in(0), "");
2172 _igvn.replace_node(fallthroughproj, ctrl);
2173 _igvn.replace_node(memproj_fallthrough, mem);
2174 fallthroughproj = ctrl;
2175 memproj_fallthrough = mem;
2176 ctrl = membar->in(TypeFunc::Control);
2177 mem = membar->in(TypeFunc::Memory);
2178 }
2179
2180 _igvn.replace_node(fallthroughproj, ctrl);
2181 _igvn.replace_node(memproj_fallthrough, mem);
2182 return true;
2183 }
2184
2185
2186 //------------------------------expand_lock_node----------------------
2187 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2188
2189 Node* ctrl = lock->in(TypeFunc::Control);
2190 Node* mem = lock->in(TypeFunc::Memory);
2191 Node* obj = lock->obj_node();
2192 Node* box = lock->box_node();
2193 Node* flock = lock->fastlock_node();
2194
2195 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2196
2197 // Make the merge point
2198 Node *region;
2199 Node *mem_phi;
2200 Node *slow_path;
2201
2202 region = new RegionNode(3);
2203 // create a Phi for the memory state
2204 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2205
2206 // Optimize test; set region slot 2
2207 slow_path = opt_bits_test(ctrl, region, 2, flock);
2208 mem_phi->init_req(2, mem);
2209
2210 // Make slow path call
2211 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2212 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2213 obj, box, nullptr);
2214
2215 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2216
2217 // Slow path can only throw asynchronous exceptions, which are always
2218 // de-opted. So the compiler thinks the slow-call can never throw an
2219 // exception. If it DOES throw an exception we would need the debug
2220 // info removed first (since if it throws there is no monitor).
2221 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2222 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2223
2224 // Capture slow path
2225 // disconnect fall-through projection from call and create a new one
2226 // hook up users of fall-through projection to region
2227 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2228 transform_later(slow_ctrl);
2229 _igvn.hash_delete(_callprojs.fallthrough_proj);
2230 _callprojs.fallthrough_proj->disconnect_inputs(C);
2231 region->init_req(1, slow_ctrl);
2232 // region inputs are now complete
2233 transform_later(region);
2234 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2235
2236 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2237
2238 mem_phi->init_req(1, memproj);
2239
2240 transform_later(mem_phi);
2241
2242 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2243 }
2244
2245 //------------------------------expand_unlock_node----------------------
2246 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2247
2248 Node* ctrl = unlock->in(TypeFunc::Control);
2249 Node* mem = unlock->in(TypeFunc::Memory);
2250 Node* obj = unlock->obj_node();
2251 Node* box = unlock->box_node();
2252
2253 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2254
2255 // No need for a null check on unlock
2256
2257 // Make the merge point
2258 Node *region;
2259 Node *mem_phi;
2260
2261 region = new RegionNode(3);
2262 // create a Phi for the memory state
2263 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2264
2265 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2266 funlock = transform_later( funlock )->as_FastUnlock();
2267 // Optimize test; set region slot 2
2268 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2269 Node *thread = transform_later(new ThreadLocalNode());
2270
2271 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2272 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2273 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2274
2275 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2276 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2277 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2278
2279 // No exceptions for unlocking
2280 // Capture slow path
2281 // disconnect fall-through projection from call and create a new one
2282 // hook up users of fall-through projection to region
2283 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2284 transform_later(slow_ctrl);
2285 _igvn.hash_delete(_callprojs.fallthrough_proj);
2286 _callprojs.fallthrough_proj->disconnect_inputs(C);
2287 region->init_req(1, slow_ctrl);
2288 // region inputs are now complete
2289 transform_later(region);
2290 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2291
2292 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2293 mem_phi->init_req(1, memproj );
2294 mem_phi->init_req(2, mem);
2295 transform_later(mem_phi);
2296
2297 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2298 }
2299
2300 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2301 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2302 Node* bol = check->unique_out();
2303 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2304 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2305 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2306
2307 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2308 Node* iff = bol->last_out(i);
2309 assert(iff->is_If(), "where's the if?");
2310
2311 if (iff->in(0)->is_top()) {
2312 _igvn.replace_input_of(iff, 1, C->top());
2313 continue;
2314 }
2315
2316 Node* iftrue = iff->as_If()->proj_out(1);
2317 Node* iffalse = iff->as_If()->proj_out(0);
2318 Node* ctrl = iff->in(0);
2319
2320 Node* subklass = nullptr;
2321 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2322 subklass = obj_or_subklass;
2323 } else {
2324 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2325 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2326 }
2327
2328 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2329
2330 _igvn.replace_input_of(iff, 0, C->top());
2331 _igvn.replace_node(iftrue, not_subtype_ctrl);
2332 _igvn.replace_node(iffalse, ctrl);
2333 }
2334 _igvn.replace_node(check, C->top());
2335 }
2336
2337 // Perform refining of strip mined loop nodes in the macro nodes list.
2338 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
2339 for (int i = C->macro_count(); i > 0; i--) {
2340 Node* n = C->macro_node(i - 1);
2341 if (n->is_OuterStripMinedLoop()) {
2342 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2343 }
2344 }
2345 }
2346
2347 //---------------------------eliminate_macro_nodes----------------------
2348 // Eliminate scalar replaced allocations and associated locks.
2349 void PhaseMacroExpand::eliminate_macro_nodes() {
2350 if (C->macro_count() == 0)
2351 return;
2352
2353 if (StressMacroElimination) {
2354 C->shuffle_macro_nodes();
2355 }
2356 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2357
2358 // Before elimination may re-mark (change to Nested or NonEscObj)
2359 // all associated (same box and obj) lock and unlock nodes.
2360 int cnt = C->macro_count();
2361 for (int i=0; i < cnt; i++) {
2362 Node *n = C->macro_node(i);
2363 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2364 mark_eliminated_locking_nodes(n->as_AbstractLock());
2365 }
2366 }
2367 // Re-marking may break consistency of Coarsened locks.
2368 if (!C->coarsened_locks_consistent()) {
2369 return; // recompile without Coarsened locks if broken
2370 } else {
2371 // After coarsened locks are eliminated locking regions
2372 // become unbalanced. We should not execute any more
2373 // locks elimination optimizations on them.
2374 C->mark_unbalanced_boxes();
2375 }
2376
2377 // First, attempt to eliminate locks
2378 bool progress = true;
2379 while (progress) {
2380 progress = false;
2381 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2382 Node* n = C->macro_node(i - 1);
2383 bool success = false;
2384 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2385 if (n->is_AbstractLock()) {
2386 success = eliminate_locking_node(n->as_AbstractLock());
2387 #ifndef PRODUCT
2388 if (success && PrintOptoStatistics) {
2389 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2390 }
2391 #endif
2392 }
2393 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2394 progress = progress || success;
2395 if (success) {
2396 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2397 }
2398 }
2399 }
2400 // Next, attempt to eliminate allocations
2401 progress = true;
2402 while (progress) {
2403 progress = false;
2404 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2405 Node* n = C->macro_node(i - 1);
2406 bool success = false;
2407 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2408 switch (n->class_id()) {
2409 case Node::Class_Allocate:
2410 case Node::Class_AllocateArray:
2411 success = eliminate_allocate_node(n->as_Allocate());
2412 #ifndef PRODUCT
2413 if (success && PrintOptoStatistics) {
2414 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2415 }
2416 #endif
2417 break;
2418 case Node::Class_CallStaticJava:
2419 success = eliminate_boxing_node(n->as_CallStaticJava());
2420 break;
2421 case Node::Class_Lock:
2422 case Node::Class_Unlock:
2423 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2424 break;
2425 case Node::Class_ArrayCopy:
2426 break;
2427 case Node::Class_OuterStripMinedLoop:
2428 break;
2429 case Node::Class_SubTypeCheck:
2430 break;
2431 case Node::Class_Opaque1:
2432 break;
2433 default:
2434 assert(n->Opcode() == Op_LoopLimit ||
2435 n->Opcode() == Op_ModD ||
2436 n->Opcode() == Op_ModF ||
2437 n->is_OpaqueNotNull() ||
2438 n->is_OpaqueInitializedAssertionPredicate() ||
2439 n->Opcode() == Op_MaxL ||
2440 n->Opcode() == Op_MinL ||
2441 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2442 "unknown node type in macro list");
2443 }
2444 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2445 progress = progress || success;
2446 if (success) {
2447 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2448 }
2449 }
2450 }
2451 #ifndef PRODUCT
2452 if (PrintOptoStatistics) {
2453 int membar_after = count_MemBar(C);
2454 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2455 }
2456 #endif
2457 }
2458
2459 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
2460 if (C->macro_count() == 0) {
2461 return;
2462 }
2463 refine_strip_mined_loop_macro_nodes();
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_OpaqueNotNull()) {
2486 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
2487 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
2488 // we will execute a Halt node.
2489 #ifdef ASSERT
2490 _igvn.replace_node(n, n->in(1));
2491 #else
2492 _igvn.replace_node(n, _igvn.intcon(1));
2493 #endif
2494 success = true;
2495 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
2496 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
2497 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
2498 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
2499 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
2500 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
2501 #ifdef ASSERT
2502 _igvn.replace_node(n, n->in(1));
2503 #else
2504 _igvn.replace_node(n, _igvn.intcon(1));
2505 #endif // ASSERT
2506 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2507 C->remove_macro_node(n);
2508 success = true;
2509 } else if (n->Opcode() == Op_MaxL) {
2510 // Since MaxL and MinL are not implemented in the backend, we expand them to
2511 // a CMoveL construct now. At least until here, the type could be computed
2512 // precisely. CMoveL is not so smart, but we can give it at least the best
2513 // type we know abouot n now.
2514 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
2515 _igvn.replace_node(n, repl);
2516 success = true;
2517 } else if (n->Opcode() == Op_MinL) {
2518 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
2519 _igvn.replace_node(n, repl);
2520 success = true;
2521 }
2522 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2523 progress = progress || success;
2524 if (success) {
2525 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2526 }
2527 }
2528 }
2529 }
2530
2531 //------------------------------expand_macro_nodes----------------------
2532 // Returns true if a failure occurred.
2533 bool PhaseMacroExpand::expand_macro_nodes() {
2534 if (StressMacroExpansion) {
2535 C->shuffle_macro_nodes();
2536 }
2537
2538 // Clean up the graph so we're less likely to hit the maximum node
2539 // limit
2540 _igvn.set_delay_transform(false);
2541 _igvn.optimize();
2542 if (C->failing()) return true;
2543 _igvn.set_delay_transform(true);
2544
2545
2546 // Because we run IGVN after each expansion, some macro nodes may go
2547 // dead and be removed from the list as we iterate over it. Move
2548 // Allocate nodes (processed in a second pass) at the beginning of
2549 // the list and then iterate from the last element of the list until
2550 // an Allocate node is seen. This is robust to random deletion in
2551 // the list due to nodes going dead.
2552 C->sort_macro_nodes();
2553
2554 // expand arraycopy "macro" nodes first
2555 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2556 // before the allocate nodes are expanded.
2557 while (C->macro_count() > 0) {
2558 int macro_count = C->macro_count();
2559 Node * n = C->macro_node(macro_count-1);
2560 assert(n->is_macro(), "only macro nodes expected here");
2561 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
2562 // node is unreachable, so don't try to expand it
2563 C->remove_macro_node(n);
2564 continue;
2565 }
2566 if (n->is_Allocate()) {
2567 break;
2568 }
2569 // Make sure expansion will not cause node limit to be exceeded.
2570 // Worst case is a macro node gets expanded into about 200 nodes.
2571 // Allow 50% more for optimization.
2572 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2573 return true;
2574 }
2575
2576 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2577 switch (n->class_id()) {
2578 case Node::Class_Lock:
2579 expand_lock_node(n->as_Lock());
2580 break;
2581 case Node::Class_Unlock:
2582 expand_unlock_node(n->as_Unlock());
2583 break;
2584 case Node::Class_ArrayCopy:
2585 expand_arraycopy_node(n->as_ArrayCopy());
2586 break;
2587 case Node::Class_SubTypeCheck:
2588 expand_subtypecheck_node(n->as_SubTypeCheck());
2589 break;
2590 default:
2591 switch (n->Opcode()) {
2592 case Op_ModD:
2593 case Op_ModF: {
2594 CallNode* mod_macro = n->as_Call();
2595 CallNode* call = new CallLeafPureNode(mod_macro->tf(), mod_macro->entry_point(),
2596 mod_macro->_name, TypeRawPtr::BOTTOM);
2597 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
2598 call->init_req(TypeFunc::I_O, C->top());
2599 call->init_req(TypeFunc::Memory, C->top());
2600 call->init_req(TypeFunc::ReturnAdr, C->top());
2601 call->init_req(TypeFunc::FramePtr, C->top());
2602 for (unsigned int i = 0; i < mod_macro->tf()->domain()->cnt() - TypeFunc::Parms; i++) {
2603 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
2604 }
2605 _igvn.replace_node(mod_macro, call);
2606 transform_later(call);
2607 break;
2608 }
2609 default:
2610 assert(false, "unknown node type in macro list");
2611 }
2612 }
2613 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2614 if (C->failing()) return true;
2615 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2616
2617 // Clean up the graph so we're less likely to hit the maximum node
2618 // limit
2619 _igvn.set_delay_transform(false);
2620 _igvn.optimize();
2621 if (C->failing()) return true;
2622 _igvn.set_delay_transform(true);
2623 }
2624
2625 // All nodes except Allocate nodes are expanded now. There could be
2626 // new optimization opportunities (such as folding newly created
2627 // load from a just allocated object). Run IGVN.
2628
2629 // expand "macro" nodes
2630 // nodes are removed from the macro list as they are processed
2631 while (C->macro_count() > 0) {
2632 int macro_count = C->macro_count();
2633 Node * n = C->macro_node(macro_count-1);
2634 assert(n->is_macro(), "only macro nodes expected here");
2635 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
2636 // node is unreachable, so don't try to expand it
2637 C->remove_macro_node(n);
2638 continue;
2639 }
2640 // Make sure expansion will not cause node limit to be exceeded.
2641 // Worst case is a macro node gets expanded into about 200 nodes.
2642 // Allow 50% more for optimization.
2643 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2644 return true;
2645 }
2646 switch (n->class_id()) {
2647 case Node::Class_Allocate:
2648 expand_allocate(n->as_Allocate());
2649 break;
2650 case Node::Class_AllocateArray:
2651 expand_allocate_array(n->as_AllocateArray());
2652 break;
2653 default:
2654 assert(false, "unknown node type in macro list");
2655 }
2656 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2657 if (C->failing()) return true;
2658 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2659
2660 // Clean up the graph so we're less likely to hit the maximum node
2661 // limit
2662 _igvn.set_delay_transform(false);
2663 _igvn.optimize();
2664 if (C->failing()) return true;
2665 _igvn.set_delay_transform(true);
2666 }
2667
2668 _igvn.set_delay_transform(false);
2669 return false;
2670 }
2671
2672 #ifndef PRODUCT
2673 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
2674 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
2675 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
2676 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
2677
2678 void PhaseMacroExpand::print_statistics() {
2679 tty->print("Objects scalar replaced = %d, ", AtomicAccess::load(&_objs_scalar_replaced_counter));
2680 tty->print("Monitor objects removed = %d, ", AtomicAccess::load(&_monitor_objects_removed_counter));
2681 tty->print("GC barriers removed = %d, ", AtomicAccess::load(&_GC_barriers_removed_counter));
2682 tty->print_cr("Memory barriers removed = %d", AtomicAccess::load(&_memory_barriers_removed_counter));
2683 }
2684
2685 int PhaseMacroExpand::count_MemBar(Compile *C) {
2686 if (!PrintOptoStatistics) {
2687 return 0;
2688 }
2689 Unique_Node_List ideal_nodes;
2690 int total = 0;
2691 ideal_nodes.map(C->live_nodes(), nullptr);
2692 ideal_nodes.push(C->root());
2693 for (uint next = 0; next < ideal_nodes.size(); ++next) {
2694 Node* n = ideal_nodes.at(next);
2695 if (n->is_MemBar()) {
2696 total++;
2697 }
2698 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2699 Node* m = n->fast_out(i);
2700 ideal_nodes.push(m);
2701 }
2702 }
2703 return total;
2704 }
2705 #endif