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