1 /*
2 * Copyright (c) 2016, 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 "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "gc/shared/c2/cardTableBarrierSetC2.hpp"
28 #include "gc/shared/gc_globals.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/graphKit.hpp"
31 #include "utilities/powerOfTwo.hpp"
32
33 const TypeFunc* ArrayCopyNode::_arraycopy_type_Type = nullptr;
34
35 ArrayCopyNode::ArrayCopyNode(Compile* C, bool alloc_tightly_coupled, bool has_negative_length_guard)
36 : CallNode(arraycopy_type(), nullptr, TypePtr::BOTTOM),
37 _kind(None),
38 _alloc_tightly_coupled(alloc_tightly_coupled),
39 _has_negative_length_guard(has_negative_length_guard),
40 _arguments_validated(false),
41 _src_type(TypeOopPtr::BOTTOM),
42 _dest_type(TypeOopPtr::BOTTOM) {
43 init_class_id(Class_ArrayCopy);
44 init_flags(Flag_is_macro);
45 C->add_macro_node(this);
46 }
47
48 uint ArrayCopyNode::size_of() const { return sizeof(*this); }
49
50 ArrayCopyNode* ArrayCopyNode::make(GraphKit* kit, bool may_throw,
51 Node* src, Node* src_offset,
52 Node* dest, Node* dest_offset,
53 Node* length,
54 bool alloc_tightly_coupled,
55 bool has_negative_length_guard,
56 Node* src_klass, Node* dest_klass,
57 Node* src_length, Node* dest_length) {
58
59 ArrayCopyNode* ac = new ArrayCopyNode(kit->C, alloc_tightly_coupled, has_negative_length_guard);
60 kit->set_predefined_input_for_runtime_call(ac);
61
62 ac->init_req(ArrayCopyNode::Src, src);
63 ac->init_req(ArrayCopyNode::SrcPos, src_offset);
64 ac->init_req(ArrayCopyNode::Dest, dest);
65 ac->init_req(ArrayCopyNode::DestPos, dest_offset);
66 ac->init_req(ArrayCopyNode::Length, length);
67 ac->init_req(ArrayCopyNode::SrcLen, src_length);
68 ac->init_req(ArrayCopyNode::DestLen, dest_length);
69 ac->init_req(ArrayCopyNode::SrcKlass, src_klass);
70 ac->init_req(ArrayCopyNode::DestKlass, dest_klass);
71
72 if (may_throw) {
73 ac->set_req(TypeFunc::I_O , kit->i_o());
74 kit->add_safepoint_edges(ac, false);
75 }
76
77 return ac;
78 }
79
80 void ArrayCopyNode::connect_outputs(GraphKit* kit, bool deoptimize_on_exception) {
81 kit->set_all_memory_call(this, true);
82 kit->set_control(kit->gvn().transform(new ProjNode(this,TypeFunc::Control)));
83 kit->set_i_o(kit->gvn().transform(new ProjNode(this, TypeFunc::I_O)));
84 kit->make_slow_call_ex(this, kit->env()->Throwable_klass(), true, deoptimize_on_exception);
85 kit->set_all_memory_call(this);
86 }
87
88 #ifndef PRODUCT
89 const char* ArrayCopyNode::_kind_names[] = {"arraycopy", "arraycopy, validated arguments", "clone", "oop array clone", "CopyOf", "CopyOfRange"};
90
91 void ArrayCopyNode::dump_spec(outputStream *st) const {
92 CallNode::dump_spec(st);
93 st->print(" (%s%s)", _kind_names[_kind], _alloc_tightly_coupled ? ", tightly coupled allocation" : "");
94 }
95
96 void ArrayCopyNode::dump_compact_spec(outputStream* st) const {
97 st->print("%s%s", _kind_names[_kind], _alloc_tightly_coupled ? ",tight" : "");
98 }
99 #endif
100
101 intptr_t ArrayCopyNode::get_length_if_constant(PhaseGVN *phase) const {
102 // check that length is constant
103 Node* length = in(ArrayCopyNode::Length);
104 const Type* length_type = phase->type(length);
105
106 if (length_type == Type::TOP) {
107 return -1;
108 }
109
110 assert(is_clonebasic() || is_arraycopy() || is_copyof() || is_copyofrange(), "unexpected array copy type");
111
112 return is_clonebasic() ? length->find_intptr_t_con(-1) : length->find_int_con(-1);
113 }
114
115 int ArrayCopyNode::get_count(PhaseGVN *phase) const {
116 Node* src = in(ArrayCopyNode::Src);
117 const Type* src_type = phase->type(src);
118
119 if (is_clonebasic()) {
120 if (src_type->isa_instptr()) {
121 const TypeInstPtr* inst_src = src_type->is_instptr();
122 ciInstanceKlass* ik = inst_src->instance_klass();
123 // ciInstanceKlass::nof_nonstatic_fields() doesn't take injected
124 // fields into account. They are rare anyway so easier to simply
125 // skip instances with injected fields.
126 if ((!inst_src->klass_is_exact() && (ik->is_interface() || ik->has_subklass())) || ik->has_injected_fields()) {
127 return -1;
128 }
129 int nb_fields = ik->nof_nonstatic_fields();
130 return nb_fields;
131 } else {
132 const TypeAryPtr* ary_src = src_type->isa_aryptr();
133 assert (ary_src != nullptr, "not an array or instance?");
134 // clone passes a length as a rounded number of longs. If we're
135 // cloning an array we'll do it element by element. If the
136 // length of the input array is constant, ArrayCopyNode::Length
137 // must be too. Note that the opposite does not need to hold,
138 // because different input array lengths (e.g. int arrays with
139 // 3 or 4 elements) might lead to the same length input
140 // (e.g. 2 double-words).
141 assert(!ary_src->size()->is_con() || (get_length_if_constant(phase) >= 0) ||
142 phase->is_IterGVN() || phase->C->inlining_incrementally() || StressReflectiveCode, "inconsistent");
143 if (ary_src->size()->is_con()) {
144 return ary_src->size()->get_con();
145 }
146 return -1;
147 }
148 }
149
150 return get_length_if_constant(phase);
151 }
152
153 Node* ArrayCopyNode::load(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, const Type *type, BasicType bt) {
154 // Pin the load: if this is an array load, it's going to be dependent on a condition that's not a range check for that
155 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk floating
156 // above runtime checks that guarantee it is within bounds.
157 DecoratorSet decorators = C2_READ_ACCESS | C2_CONTROL_DEPENDENT_LOAD | IN_HEAP | C2_ARRAY_COPY | C2_UNKNOWN_CONTROL_LOAD;
158 C2AccessValuePtr addr(adr, adr_type);
159 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
160 Node* res = bs->load_at(access, type);
161 ctl = access.ctl();
162 return res;
163 }
164
165 void ArrayCopyNode::store(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, Node* val, const Type *type, BasicType bt) {
166 DecoratorSet decorators = C2_WRITE_ACCESS | IN_HEAP | C2_ARRAY_COPY;
167 if (is_alloc_tightly_coupled()) {
168 decorators |= C2_TIGHTLY_COUPLED_ALLOC;
169 }
170 C2AccessValuePtr addr(adr, adr_type);
171 C2AccessValue value(val, type);
172 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
173 bs->store_at(access, value);
174 ctl = access.ctl();
175 }
176
177
178 Node* ArrayCopyNode::try_clone_instance(PhaseGVN *phase, bool can_reshape, int count) {
179 if (!is_clonebasic()) {
180 return nullptr;
181 }
182
183 Node* base_src = in(ArrayCopyNode::Src);
184 Node* base_dest = in(ArrayCopyNode::Dest);
185 Node* ctl = in(TypeFunc::Control);
186 Node* in_mem = in(TypeFunc::Memory);
187
188 const Type* src_type = phase->type(base_src);
189 const TypeInstPtr* inst_src = src_type->isa_instptr();
190 if (inst_src == nullptr) {
191 return nullptr;
192 }
193
194 MergeMemNode* mem = phase->transform(MergeMemNode::make(in_mem))->as_MergeMem();
195 if (can_reshape) {
196 phase->is_IterGVN()->_worklist.push(mem);
197 }
198
199
200 ciInstanceKlass* ik = inst_src->instance_klass();
201
202 if (!inst_src->klass_is_exact()) {
203 assert(!ik->is_interface(), "inconsistent klass hierarchy");
204 if (ik->has_subklass()) {
205 // Concurrent class loading.
206 // Fail fast and return NodeSentinel to indicate that the transform failed.
207 return NodeSentinel;
208 } else {
209 phase->C->dependencies()->assert_leaf_type(ik);
210 }
211 }
212
213 const TypeInstPtr* dest_type = phase->type(base_dest)->is_instptr();
214 if (dest_type->instance_klass() != ik) {
215 // At parse time, the exact type of the object to clone was not known. That inexact type was captured by the CheckCastPP
216 // of the newly allocated cloned object (in dest). The exact type is now known (in src), but the type for the cloned object
217 // (dest) was not updated. When copying the fields below, Store nodes may write to offsets for fields that don't exist in
218 // the inexact class. The stores would then be assigned an incorrect slice.
219 return NodeSentinel;
220 }
221
222 assert(ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem, "too many fields");
223
224 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
225 for (int i = 0; i < count; i++) {
226 ciField* field = ik->nonstatic_field_at(i);
227 const TypePtr* adr_type = phase->C->alias_type(field)->adr_type();
228 Node* off = phase->MakeConX(field->offset_in_bytes());
229 Node* next_src = phase->transform(AddPNode::make_with_base(base_src, off));
230 Node* next_dest = phase->transform(AddPNode::make_with_base(base_dest, off));
231 assert(phase->C->get_alias_index(adr_type) == phase->C->get_alias_index(phase->type(next_src)->isa_ptr()),
232 "slice of address and input slice don't match");
233 assert(phase->C->get_alias_index(adr_type) == phase->C->get_alias_index(phase->type(next_dest)->isa_ptr()),
234 "slice of address and input slice don't match");
235 BasicType bt = field->layout_type();
236
237 const Type *type;
238 if (bt == T_OBJECT) {
239 if (!field->type()->is_loaded()) {
240 type = TypeInstPtr::BOTTOM;
241 } else {
242 ciType* field_klass = field->type();
243 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
244 }
245 } else {
246 type = Type::get_const_basic_type(bt);
247 }
248
249 Node* v = load(bs, phase, ctl, mem, next_src, adr_type, type, bt);
250 store(bs, phase, ctl, mem, next_dest, adr_type, v, type, bt);
251 }
252
253 if (!finish_transform(phase, can_reshape, ctl, mem)) {
254 // Return NodeSentinel to indicate that the transform failed
255 return NodeSentinel;
256 }
257
258 return mem;
259 }
260
261 // We may have narrowed the type of base because this runs with PhaseIterGVN::_delay_transform true, explicitly
262 // update the type of the AddP so it's consistent with its base and load() picks the right memory slice.
263 Node* ArrayCopyNode::make_and_transform_addp(PhaseGVN* phase, Node* base, Node* offset) {
264 return make_and_transform_addp(phase, base, base, offset);
265 }
266
267 Node* ArrayCopyNode::make_and_transform_addp(PhaseGVN* phase, Node* base, Node* ptr, Node* offset) {
268 assert(phase->is_IterGVN() == nullptr || phase->is_IterGVN()->delay_transform(), "helper method when delay transform is set");
269 Node* addp = phase->transform(AddPNode::make_with_base(base, ptr, offset));
270 phase->set_type(addp, addp->Value(phase));
271 return addp;
272 }
273
274 bool ArrayCopyNode::prepare_array_copy(PhaseGVN *phase, bool can_reshape,
275 Node*& adr_src,
276 Node*& base_src,
277 Node*& adr_dest,
278 Node*& base_dest,
279 BasicType& copy_type,
280 const Type*& value_type,
281 bool& disjoint_bases) {
282 base_src = in(ArrayCopyNode::Src);
283 base_dest = in(ArrayCopyNode::Dest);
284 const Type* src_type = phase->type(base_src);
285 const TypeAryPtr* ary_src = src_type->isa_aryptr();
286
287 Node* src_offset = in(ArrayCopyNode::SrcPos);
288 Node* dest_offset = in(ArrayCopyNode::DestPos);
289
290 if (is_arraycopy() || is_copyofrange() || is_copyof()) {
291 const Type* dest_type = phase->type(base_dest);
292 const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
293
294 // newly allocated object is guaranteed to not overlap with source object
295 disjoint_bases = is_alloc_tightly_coupled();
296 if (ary_src == nullptr || ary_src->elem() == Type::BOTTOM ||
297 ary_dest == nullptr || ary_dest->elem() == Type::BOTTOM) {
298 // We don't know if arguments are arrays
299 return false;
300 }
301
302 BasicType src_elem = ary_src->elem()->array_element_basic_type();
303 BasicType dest_elem = ary_dest->elem()->array_element_basic_type();
304 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
305 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
306
307 if (src_elem != dest_elem || dest_elem == T_VOID) {
308 // We don't know if arguments are arrays of the same type
309 return false;
310 }
311
312 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
313 if (bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), dest_elem, false, false, BarrierSetC2::Optimization)) {
314 // It's an object array copy but we can't emit the card marking
315 // that is needed
316 return false;
317 }
318
319 value_type = ary_src->elem();
320
321 uint shift = exact_log2(type2aelembytes(dest_elem));
322 uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
323
324 src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
325 if (src_offset->is_top()) {
326 // Offset is out of bounds (the ArrayCopyNode will be removed)
327 return false;
328 }
329 dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
330 if (dest_offset->is_top()) {
331 // Offset is out of bounds (the ArrayCopyNode will be removed)
332 if (can_reshape) {
333 // record src_offset, so it can be deleted later (if it is dead)
334 phase->is_IterGVN()->_worklist.push(src_offset);
335 }
336 return false;
337 }
338
339 Node* hook = new Node(1);
340 hook->init_req(0, dest_offset);
341
342 Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
343
344 hook->destruct(phase);
345
346 Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
347
348 adr_src = make_and_transform_addp(phase, base_src, src_scale);
349 adr_dest = make_and_transform_addp(phase, base_dest, dest_scale);
350
351 adr_src = make_and_transform_addp(phase, base_src, adr_src, phase->MakeConX(header));
352 adr_dest = make_and_transform_addp(phase, base_dest, adr_dest, phase->MakeConX(header));
353 copy_type = dest_elem;
354 } else {
355 assert(ary_src != nullptr, "should be a clone");
356 assert(is_clonebasic(), "should be");
357
358 disjoint_bases = true;
359
360 BasicType elem = ary_src->isa_aryptr()->elem()->array_element_basic_type();
361 if (is_reference_type(elem, true)) {
362 elem = T_OBJECT;
363 }
364
365 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
366 if (bs->array_copy_requires_gc_barriers(true, elem, true, is_clone_inst(), BarrierSetC2::Optimization)) {
367 return false;
368 }
369
370 adr_src = make_and_transform_addp(phase, base_src, src_offset);
371 adr_dest = make_and_transform_addp(phase, base_dest, dest_offset);
372
373 // The address is offsetted to an aligned address where a raw copy would start.
374 // If the clone copy is decomposed into load-stores - the address is adjusted to
375 // point at where the array starts.
376 const Type* toff = phase->type(src_offset);
377 int offset = toff->isa_long() ? (int) toff->is_long()->get_con() : (int) toff->is_int()->get_con();
378 int diff = arrayOopDesc::base_offset_in_bytes(elem) - offset;
379 assert(diff >= 0, "clone should not start after 1st array element");
380 if (diff > 0) {
381 adr_src = make_and_transform_addp(phase, base_src, adr_src, phase->MakeConX(diff));
382 adr_dest = make_and_transform_addp(phase, base_dest, adr_dest, phase->MakeConX(diff));
383 }
384 copy_type = elem;
385 value_type = ary_src->elem();
386 }
387 return true;
388 }
389
390 const TypePtr* ArrayCopyNode::get_address_type(PhaseGVN* phase, const TypePtr* atp, Node* n) {
391 if (atp == TypeOopPtr::BOTTOM) {
392 atp = phase->type(n)->isa_ptr();
393 }
394 // adjust atp to be the correct array element address type
395 return atp->add_offset(Type::OffsetBot);
396 }
397
398 const TypePtr* ArrayCopyNode::get_src_adr_type(PhaseGVN* phase) const {
399 return get_address_type(phase, _src_type, in(Src));
400 }
401
402 void ArrayCopyNode::array_copy_test_overlap(PhaseGVN *phase, bool can_reshape, bool disjoint_bases, int count, Node*& forward_ctl, Node*& backward_ctl) {
403 Node* ctl = in(TypeFunc::Control);
404 if (!disjoint_bases && count > 1) {
405 Node* src_offset = in(ArrayCopyNode::SrcPos);
406 Node* dest_offset = in(ArrayCopyNode::DestPos);
407 assert(src_offset != nullptr && dest_offset != nullptr, "should be");
408 Node* cmp = phase->transform(new CmpINode(src_offset, dest_offset));
409 Node *bol = phase->transform(new BoolNode(cmp, BoolTest::lt));
410 IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
411
412 phase->transform(iff);
413
414 forward_ctl = phase->transform(new IfFalseNode(iff));
415 backward_ctl = phase->transform(new IfTrueNode(iff));
416 } else {
417 forward_ctl = ctl;
418 }
419 }
420
421 Node* ArrayCopyNode::array_copy_forward(PhaseGVN *phase,
422 bool can_reshape,
423 Node*& forward_ctl,
424 Node* mem,
425 const TypePtr* atp_src,
426 const TypePtr* atp_dest,
427 Node* adr_src,
428 Node* base_src,
429 Node* adr_dest,
430 Node* base_dest,
431 BasicType copy_type,
432 const Type* value_type,
433 int count) {
434 if (!forward_ctl->is_top()) {
435 // copy forward
436 MergeMemNode* mm = MergeMemNode::make(mem);
437
438 if (count > 0) {
439 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
440 Node* v = load(bs, phase, forward_ctl, mm, adr_src, atp_src, value_type, copy_type);
441 store(bs, phase, forward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
442 for (int i = 1; i < count; i++) {
443 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
444 Node* next_src = make_and_transform_addp(phase, base_src,adr_src,off);
445 Node* next_dest = make_and_transform_addp(phase, base_dest,adr_dest,off);
446 // Same as above
447 phase->set_type(next_dest, next_dest->Value(phase));
448 v = load(bs, phase, forward_ctl, mm, next_src, atp_src, value_type, copy_type);
449 store(bs, phase, forward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
450 }
451 } else if (can_reshape) {
452 PhaseIterGVN* igvn = phase->is_IterGVN();
453 igvn->_worklist.push(adr_src);
454 igvn->_worklist.push(adr_dest);
455 }
456 return mm;
457 }
458 return phase->C->top();
459 }
460
461 Node* ArrayCopyNode::array_copy_backward(PhaseGVN *phase,
462 bool can_reshape,
463 Node*& backward_ctl,
464 Node* mem,
465 const TypePtr* atp_src,
466 const TypePtr* atp_dest,
467 Node* adr_src,
468 Node* base_src,
469 Node* adr_dest,
470 Node* base_dest,
471 BasicType copy_type,
472 const Type* value_type,
473 int count) {
474 if (!backward_ctl->is_top()) {
475 // copy backward
476 MergeMemNode* mm = MergeMemNode::make(mem);
477
478 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
479 assert(copy_type != T_OBJECT || !bs->array_copy_requires_gc_barriers(false, T_OBJECT, false, false, BarrierSetC2::Optimization), "only tightly coupled allocations for object arrays");
480
481 if (count > 0) {
482 for (int i = count-1; i >= 1; i--) {
483 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
484 Node* next_src = make_and_transform_addp(phase, base_src,adr_src,off);
485 Node* next_dest = make_and_transform_addp(phase, base_dest,adr_dest,off);
486 Node* v = load(bs, phase, backward_ctl, mm, next_src, atp_src, value_type, copy_type);
487 store(bs, phase, backward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
488 }
489 Node* v = load(bs, phase, backward_ctl, mm, adr_src, atp_src, value_type, copy_type);
490 store(bs, phase, backward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
491 } else if (can_reshape) {
492 PhaseIterGVN* igvn = phase->is_IterGVN();
493 igvn->_worklist.push(adr_src);
494 igvn->_worklist.push(adr_dest);
495 }
496 return phase->transform(mm);
497 }
498 return phase->C->top();
499 }
500
501 bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
502 Node* ctl, Node *mem) {
503 if (can_reshape) {
504 PhaseIterGVN* igvn = phase->is_IterGVN();
505 igvn->set_delay_transform(false);
506 if (is_clonebasic()) {
507 Node* out_mem = proj_out(TypeFunc::Memory);
508
509 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
510 if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
511 out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
512 assert(bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, is_clone_inst(), BarrierSetC2::Optimization), "can only happen with card marking");
513 return false;
514 }
515
516 igvn->replace_node(out_mem->raw_out(0), mem);
517
518 Node* out_ctl = proj_out(TypeFunc::Control);
519 igvn->replace_node(out_ctl, ctl);
520 } else {
521 // replace fallthrough projections of the ArrayCopyNode by the
522 // new memory, control and the input IO.
523 CallProjections callprojs;
524 extract_projections(&callprojs, true, false);
525
526 if (callprojs.fallthrough_ioproj != nullptr) {
527 igvn->replace_node(callprojs.fallthrough_ioproj, in(TypeFunc::I_O));
528 }
529 if (callprojs.fallthrough_memproj != nullptr) {
530 igvn->replace_node(callprojs.fallthrough_memproj, mem);
531 }
532 if (callprojs.fallthrough_catchproj != nullptr) {
533 igvn->replace_node(callprojs.fallthrough_catchproj, ctl);
534 }
535
536 // The ArrayCopyNode is not disconnected. It still has the
537 // projections for the exception case. Replace current
538 // ArrayCopyNode with a dummy new one with a top() control so
539 // that this part of the graph stays consistent but is
540 // eventually removed.
541
542 set_req(0, phase->C->top());
543 remove_dead_region(phase, can_reshape);
544 }
545 } else {
546 if (in(TypeFunc::Control) != ctl) {
547 // we can't return new memory and control from Ideal at parse time
548 assert(!is_clonebasic() || UseShenandoahGC, "added control for clone?");
549 phase->record_for_igvn(this);
550 return false;
551 }
552 }
553 return true;
554 }
555
556
557 Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
558 if (remove_dead_region(phase, can_reshape)) return this;
559
560 if (StressArrayCopyMacroNode && !can_reshape) {
561 phase->record_for_igvn(this);
562 return nullptr;
563 }
564
565 // See if it's a small array copy and we can inline it as
566 // loads/stores
567 // Here we can only do:
568 // - arraycopy if all arguments were validated before and we don't
569 // need card marking
570 // - clone for which we don't need to do card marking
571
572 if (!is_clonebasic() && !is_arraycopy_validated() &&
573 !is_copyofrange_validated() && !is_copyof_validated()) {
574 return nullptr;
575 }
576
577 assert(in(TypeFunc::Control) != nullptr &&
578 in(TypeFunc::Memory) != nullptr &&
579 in(ArrayCopyNode::Src) != nullptr &&
580 in(ArrayCopyNode::Dest) != nullptr &&
581 in(ArrayCopyNode::Length) != nullptr &&
582 in(ArrayCopyNode::SrcPos) != nullptr &&
583 in(ArrayCopyNode::DestPos) != nullptr, "broken inputs");
584
585 if (in(TypeFunc::Control)->is_top() ||
586 in(TypeFunc::Memory)->is_top() ||
587 phase->type(in(ArrayCopyNode::Src)) == Type::TOP ||
588 phase->type(in(ArrayCopyNode::Dest)) == Type::TOP ||
589 (in(ArrayCopyNode::SrcPos) != nullptr && in(ArrayCopyNode::SrcPos)->is_top()) ||
590 (in(ArrayCopyNode::DestPos) != nullptr && in(ArrayCopyNode::DestPos)->is_top())) {
591 return nullptr;
592 }
593
594 int count = get_count(phase);
595
596 if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
597 return nullptr;
598 }
599
600 Node* mem = try_clone_instance(phase, can_reshape, count);
601 if (mem != nullptr) {
602 return (mem == NodeSentinel) ? nullptr : mem;
603 }
604
605 Node* adr_src = nullptr;
606 Node* base_src = nullptr;
607 Node* adr_dest = nullptr;
608 Node* base_dest = nullptr;
609 BasicType copy_type = T_ILLEGAL;
610 const Type* value_type = nullptr;
611 bool disjoint_bases = false;
612
613 Node* src = in(ArrayCopyNode::Src);
614 Node* dest = in(ArrayCopyNode::Dest);
615 // EA may have moved an input to a new slice. EA stores the new address types in the ArrayCopy node itself
616 // (_src_type/_dest_type). phase->type(src) and _src_type or phase->type(dest) and _dest_type may be different
617 // when this transformation runs if igvn hasn't had a chance to propagate the new types yet. Make sure the new
618 // types are taken into account so new Load/Store nodes are created on the right slice.
619 const TypePtr* atp_src = get_address_type(phase, _src_type, src);
620 const TypePtr* atp_dest = get_address_type(phase, _dest_type, dest);
621 phase->set_type(src, phase->type(src)->join_speculative(atp_src));
622 phase->set_type(dest, phase->type(dest)->join_speculative(atp_dest));
623
624 // Control flow is going to be created, it's easier to do with _delay_transform set to true.
625
626 // prepare_array_copy() doesn't build control flow, but it creates AddP nodes. The src/dest type possibly gets
627 // narrowed above. If a newly created AddP node is commoned with a pre-existing one, then the type narrowing is lost.
628 // Setting _delay_transform before prepare_array_copy() guarantees this doesn't happen.
629 if (can_reshape) {
630 assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
631 phase->is_IterGVN()->set_delay_transform(true);
632 }
633
634 if (!prepare_array_copy(phase, can_reshape,
635 adr_src, base_src, adr_dest, base_dest,
636 copy_type, value_type, disjoint_bases)) {
637 assert(adr_src == nullptr, "no node can be left behind");
638 assert(adr_dest == nullptr, "no node can be left behind");
639 if (can_reshape) {
640 assert(phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
641 phase->is_IterGVN()->set_delay_transform(false);
642 }
643
644 return nullptr;
645 }
646
647 Node* in_mem = in(TypeFunc::Memory);
648
649 Node* backward_ctl = phase->C->top();
650 Node* forward_ctl = phase->C->top();
651 array_copy_test_overlap(phase, can_reshape, disjoint_bases, count, forward_ctl, backward_ctl);
652
653 Node* forward_mem = array_copy_forward(phase, can_reshape, forward_ctl,
654 in_mem,
655 atp_src, atp_dest,
656 adr_src, base_src, adr_dest, base_dest,
657 copy_type, value_type, count);
658
659 Node* backward_mem = array_copy_backward(phase, can_reshape, backward_ctl,
660 in_mem,
661 atp_src, atp_dest,
662 adr_src, base_src, adr_dest, base_dest,
663 copy_type, value_type, count);
664
665 Node* ctl = nullptr;
666 if (!forward_ctl->is_top() && !backward_ctl->is_top()) {
667 ctl = new RegionNode(3);
668 ctl->init_req(1, forward_ctl);
669 ctl->init_req(2, backward_ctl);
670 ctl = phase->transform(ctl);
671 MergeMemNode* forward_mm = forward_mem->as_MergeMem();
672 MergeMemNode* backward_mm = backward_mem->as_MergeMem();
673 for (MergeMemStream mms(forward_mm, backward_mm); mms.next_non_empty2(); ) {
674 if (mms.memory() != mms.memory2()) {
675 Node* phi = new PhiNode(ctl, Type::MEMORY, phase->C->get_adr_type(mms.alias_idx()));
676 phi->init_req(1, mms.memory());
677 phi->init_req(2, mms.memory2());
678 phi = phase->transform(phi);
679 mms.set_memory(phi);
680 }
681 }
682 mem = forward_mem;
683 } else if (!forward_ctl->is_top()) {
684 ctl = forward_ctl;
685 mem = forward_mem;
686 } else {
687 assert(!backward_ctl->is_top(), "no copy?");
688 ctl = backward_ctl;
689 mem = backward_mem;
690 }
691
692 if (can_reshape) {
693 assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
694 phase->is_IterGVN()->set_delay_transform(false);
695 }
696
697 if (!finish_transform(phase, can_reshape, ctl, mem)) {
698 if (can_reshape) {
699 // put in worklist, so that if it happens to be dead it is removed
700 phase->is_IterGVN()->_worklist.push(mem);
701 }
702 return nullptr;
703 }
704
705 return mem;
706 }
707
708 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) const {
709 Node* dest = in(ArrayCopyNode::Dest);
710 if (dest->is_top()) {
711 return false;
712 }
713 const TypeOopPtr* dest_t = phase->type(dest)->is_oopptr();
714 assert(!dest_t->is_known_instance() || _dest_type->is_known_instance(), "result of EA not recorded");
715 assert(in(ArrayCopyNode::Src)->is_top() || !phase->type(in(ArrayCopyNode::Src))->is_oopptr()->is_known_instance() ||
716 _src_type->is_known_instance(), "result of EA not recorded");
717
718 if (_dest_type != TypeOopPtr::BOTTOM || t_oop->is_known_instance()) {
719 assert(_dest_type == TypeOopPtr::BOTTOM || _dest_type->is_known_instance(), "result of EA is known instance");
720 return t_oop->instance_id() == _dest_type->instance_id();
721 }
722
723 return CallNode::may_modify_arraycopy_helper(dest_t, t_oop, phase);
724 }
725
726 bool ArrayCopyNode::may_modify_helper(const TypeOopPtr* t_oop, Node* n, PhaseValues* phase, ArrayCopyNode*& ac) {
727 if (n != nullptr &&
728 n->is_ArrayCopy() &&
729 n->as_ArrayCopy()->may_modify(t_oop, phase)) {
730 ac = n->as_ArrayCopy();
731 return true;
732 }
733 return false;
734 }
735
736 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, MemBarNode* mb, PhaseValues* phase, ArrayCopyNode*& ac) {
737 if (mb->trailing_expanded_array_copy()) {
738 return true;
739 }
740
741 Node* c = mb->in(0);
742
743 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
744 // step over g1 gc barrier if we're at e.g. a clone with ReduceInitialCardMarks off
745 c = bs->step_over_gc_barrier(c);
746
747 CallNode* call = nullptr;
748 guarantee(c != nullptr, "step_over_gc_barrier failed, there must be something to step to.");
749 if (c->is_Region()) {
750 for (uint i = 1; i < c->req(); i++) {
751 if (c->in(i) != nullptr) {
752 Node* n = c->in(i)->in(0);
753 if (may_modify_helper(t_oop, n, phase, ac)) {
754 assert(c == mb->in(0), "only for clone");
755 return true;
756 }
757 }
758 }
759 } else if (may_modify_helper(t_oop, c->in(0), phase, ac)) {
760 #ifdef ASSERT
761 bool use_ReduceInitialCardMarks = BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet) &&
762 static_cast<CardTableBarrierSetC2*>(bs)->use_ReduceInitialCardMarks();
763 assert(c == mb->in(0) || (ac->is_clonebasic() && !use_ReduceInitialCardMarks), "only for clone");
764 #endif
765 return true;
766 }
767
768 return false;
769 }
770
771 // Does this array copy modify offsets between offset_lo and offset_hi
772 // in the destination array
773 // if must_modify is false, return true if the copy could write
774 // between offset_lo and offset_hi
775 // if must_modify is true, return true if the copy is guaranteed to
776 // write between offset_lo and offset_hi
777 bool ArrayCopyNode::modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseValues* phase, bool must_modify) const {
778 assert(_kind == ArrayCopy || _kind == CopyOf || _kind == CopyOfRange, "only for real array copies");
779
780 Node* dest = in(Dest);
781 Node* dest_pos = in(DestPos);
782 Node* len = in(Length);
783
784 const TypeInt *dest_pos_t = phase->type(dest_pos)->isa_int();
785 const TypeInt *len_t = phase->type(len)->isa_int();
786 const TypeAryPtr* ary_t = phase->type(dest)->isa_aryptr();
787
788 if (dest_pos_t == nullptr || len_t == nullptr || ary_t == nullptr) {
789 return !must_modify;
790 }
791
792 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
793 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
794
795 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
796 uint elemsize = type2aelembytes(ary_elem);
797
798 jlong dest_pos_plus_len_lo = (((jlong)dest_pos_t->_lo) + len_t->_lo) * elemsize + header;
799 jlong dest_pos_plus_len_hi = (((jlong)dest_pos_t->_hi) + len_t->_hi) * elemsize + header;
800 jlong dest_pos_lo = ((jlong)dest_pos_t->_lo) * elemsize + header;
801 jlong dest_pos_hi = ((jlong)dest_pos_t->_hi) * elemsize + header;
802
803 if (must_modify) {
804 if (offset_lo >= dest_pos_hi && offset_hi < dest_pos_plus_len_lo) {
805 return true;
806 }
807 } else {
808 if (offset_hi >= dest_pos_lo && offset_lo < dest_pos_plus_len_hi) {
809 return true;
810 }
811 }
812 return false;
813 }
814
815 // As an optimization, choose the optimal vector size for bounded copy length
816 int ArrayCopyNode::get_partial_inline_vector_lane_count(BasicType type, jlong max_len) {
817 assert(max_len > 0, JLONG_FORMAT, max_len);
818 // We only care whether max_size_in_bytes is not larger than 32, we also want to avoid
819 // multiplication overflow, so clamp max_len to [0, 64]
820 int max_size_in_bytes = MIN2<jlong>(max_len, 64) * type2aelembytes(type);
821 if (ArrayOperationPartialInlineSize > 16 && max_size_in_bytes <= 16) {
822 return 16 / type2aelembytes(type);
823 } else if (ArrayOperationPartialInlineSize > 32 && max_size_in_bytes <= 32) {
824 return 32 / type2aelembytes(type);
825 } else {
826 return ArrayOperationPartialInlineSize / type2aelembytes(type);
827 }
828 }