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