1 /*
2 * Copyright (c) 2016, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "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 "runtime/sharedRuntime.hpp"
32 #include "utilities/macros.hpp"
33 #include "utilities/powerOfTwo.hpp"
34
35 const TypeFunc* ArrayCopyNode::_arraycopy_type_Type = nullptr;
36
37 ArrayCopyNode::ArrayCopyNode(Compile* C, bool alloc_tightly_coupled, bool has_negative_length_guard)
38 : CallNode(arraycopy_type(), nullptr, TypePtr::BOTTOM),
39 _kind(None),
40 _alloc_tightly_coupled(alloc_tightly_coupled),
41 _has_negative_length_guard(has_negative_length_guard),
42 _arguments_validated(false),
43 _src_type(TypeOopPtr::BOTTOM),
44 _dest_type(TypeOopPtr::BOTTOM) {
45 init_class_id(Class_ArrayCopy);
46 init_flags(Flag_is_macro);
47 C->add_macro_node(this);
48 }
49
50 uint ArrayCopyNode::size_of() const { return sizeof(*this); }
51
52 ArrayCopyNode* ArrayCopyNode::make(GraphKit* kit, bool may_throw,
53 Node* src, Node* src_offset,
54 Node* dest, Node* dest_offset,
55 Node* length,
56 bool alloc_tightly_coupled,
57 bool has_negative_length_guard,
58 Node* src_klass, Node* dest_klass,
59 Node* src_length, Node* dest_length) {
60
61 ArrayCopyNode* ac = new ArrayCopyNode(kit->C, alloc_tightly_coupled, has_negative_length_guard);
62 kit->set_predefined_input_for_runtime_call(ac);
63
64 ac->init_req(ArrayCopyNode::Src, src);
65 ac->init_req(ArrayCopyNode::SrcPos, src_offset);
66 ac->init_req(ArrayCopyNode::Dest, dest);
67 ac->init_req(ArrayCopyNode::DestPos, dest_offset);
68 ac->init_req(ArrayCopyNode::Length, length);
69 ac->init_req(ArrayCopyNode::SrcLen, src_length);
70 ac->init_req(ArrayCopyNode::DestLen, dest_length);
71 ac->init_req(ArrayCopyNode::SrcKlass, src_klass);
72 ac->init_req(ArrayCopyNode::DestKlass, dest_klass);
73
74 if (may_throw) {
75 ac->set_req(TypeFunc::I_O , kit->i_o());
76 kit->add_safepoint_edges(ac, false);
77 }
78
79 return ac;
80 }
81
82 void ArrayCopyNode::connect_outputs(GraphKit* kit, bool deoptimize_on_exception) {
83 kit->set_all_memory_call(this, true);
84 kit->set_control(kit->gvn().transform(new ProjNode(this,TypeFunc::Control)));
85 kit->set_i_o(kit->gvn().transform(new ProjNode(this, TypeFunc::I_O)));
86 kit->make_slow_call_ex(this, kit->env()->Throwable_klass(), true, deoptimize_on_exception);
87 kit->set_all_memory_call(this);
88 }
89
90 #ifndef PRODUCT
91 const char* ArrayCopyNode::_kind_names[] = {"arraycopy", "arraycopy, validated arguments", "clone", "oop array clone", "CopyOf", "CopyOfRange"};
92
93 void ArrayCopyNode::dump_spec(outputStream *st) const {
94 CallNode::dump_spec(st);
95 st->print(" (%s%s)", _kind_names[_kind], _alloc_tightly_coupled ? ", tightly coupled allocation" : "");
96 }
97
98 void ArrayCopyNode::dump_compact_spec(outputStream* st) const {
99 st->print("%s%s", _kind_names[_kind], _alloc_tightly_coupled ? ",tight" : "");
100 }
101 #endif
102
103 intptr_t ArrayCopyNode::get_length_if_constant(PhaseGVN *phase) const {
104 // check that length is constant
105 Node* length = in(ArrayCopyNode::Length);
106 const Type* length_type = phase->type(length);
107
108 if (length_type == Type::TOP) {
109 return -1;
110 }
111
112 assert(is_clonebasic() || is_arraycopy() || is_copyof() || is_copyofrange(), "unexpected array copy type");
113
114 return is_clonebasic() ? length->find_intptr_t_con(-1) : length->find_int_con(-1);
115 }
116
117 int ArrayCopyNode::get_count(PhaseGVN *phase) const {
118 Node* src = in(ArrayCopyNode::Src);
119 const Type* src_type = phase->type(src);
120
121 if (is_clonebasic()) {
122 if (src_type->isa_instptr()) {
123 const TypeInstPtr* inst_src = src_type->is_instptr();
124 ciInstanceKlass* ik = inst_src->instance_klass();
125 // ciInstanceKlass::nof_nonstatic_fields() doesn't take injected
126 // fields into account. They are rare anyway so easier to simply
127 // skip instances with injected fields.
128 if ((!inst_src->klass_is_exact() && (ik->is_interface() || ik->has_subklass())) || ik->has_injected_fields()) {
129 return -1;
130 }
131 int nb_fields = ik->nof_nonstatic_fields();
132 return nb_fields;
133 } else {
134 const TypeAryPtr* ary_src = src_type->isa_aryptr();
135 assert (ary_src != nullptr, "not an array or instance?");
136 // clone passes a length as a rounded number of longs. If we're
137 // cloning an array we'll do it element by element. If the
138 // length of the input array is constant, ArrayCopyNode::Length
139 // must be too. Note that the opposite does not need to hold,
140 // because different input array lengths (e.g. int arrays with
141 // 3 or 4 elements) might lead to the same length input
142 // (e.g. 2 double-words).
143 assert(!ary_src->size()->is_con() || (get_length_if_constant(phase) >= 0) ||
144 phase->is_IterGVN() || phase->C->inlining_incrementally() || StressReflectiveCode, "inconsistent");
145 if (ary_src->size()->is_con()) {
146 return ary_src->size()->get_con();
147 }
148 return -1;
149 }
150 }
151
152 return get_length_if_constant(phase);
153 }
154
155 Node* ArrayCopyNode::load(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, const Type *type, BasicType bt) {
156 // 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
157 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk floating
158 // above runtime checks that guarantee it is within bounds.
159 DecoratorSet decorators = C2_READ_ACCESS | C2_CONTROL_DEPENDENT_LOAD | IN_HEAP | C2_ARRAY_COPY | C2_UNKNOWN_CONTROL_LOAD;
160 C2AccessValuePtr addr(adr, adr_type);
161 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
162 Node* res = bs->load_at(access, type);
163 ctl = access.ctl();
164 return res;
165 }
166
167 void ArrayCopyNode::store(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, Node* val, const Type *type, BasicType bt) {
168 DecoratorSet decorators = C2_WRITE_ACCESS | IN_HEAP | C2_ARRAY_COPY;
169 if (is_alloc_tightly_coupled()) {
170 decorators |= C2_TIGHTLY_COUPLED_ALLOC;
171 }
172 C2AccessValuePtr addr(adr, adr_type);
173 C2AccessValue value(val, type);
174 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
175 bs->store_at(access, value);
176 ctl = access.ctl();
177 }
178
179
180 Node* ArrayCopyNode::try_clone_instance(PhaseGVN *phase, bool can_reshape, int count) {
181 if (!is_clonebasic()) {
182 return nullptr;
183 }
184
185 Node* base_src = in(ArrayCopyNode::Src);
186 Node* base_dest = in(ArrayCopyNode::Dest);
187 Node* ctl = in(TypeFunc::Control);
188 Node* in_mem = in(TypeFunc::Memory);
189
190 const Type* src_type = phase->type(base_src);
191 const TypeInstPtr* inst_src = src_type->isa_instptr();
192 if (inst_src == nullptr) {
193 return nullptr;
194 }
195
196 MergeMemNode* mem = phase->transform(MergeMemNode::make(in_mem))->as_MergeMem();
197 if (can_reshape) {
198 phase->is_IterGVN()->_worklist.push(mem);
199 }
200
201
202 ciInstanceKlass* ik = inst_src->instance_klass();
203
204 if (!inst_src->klass_is_exact()) {
205 assert(!ik->is_interface(), "inconsistent klass hierarchy");
206 if (ik->has_subklass()) {
207 // Concurrent class loading.
208 // Fail fast and return NodeSentinel to indicate that the transform failed.
209 return NodeSentinel;
210 } else {
211 phase->C->dependencies()->assert_leaf_type(ik);
212 }
213 }
214
215 const TypeInstPtr* dest_type = phase->type(base_dest)->is_instptr();
216 if (dest_type->instance_klass() != ik) {
217 // At parse time, the exact type of the object to clone was not known. That inexact type was captured by the CheckCastPP
218 // of the newly allocated cloned object (in dest). The exact type is now known (in src), but the type for the cloned object
219 // (dest) was not updated. When copying the fields below, Store nodes may write to offsets for fields that don't exist in
220 // the inexact class. The stores would then be assigned an incorrect slice.
221 return NodeSentinel;
222 }
223
224 assert(ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem, "too many fields");
225
226 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
227 for (int i = 0; i < count; i++) {
228 ciField* field = ik->nonstatic_field_at(i);
229 const TypePtr* adr_type = phase->C->alias_type(field)->adr_type();
230 Node* off = phase->MakeConX(field->offset_in_bytes());
231 Node* next_src = phase->transform(new AddPNode(base_src,base_src,off));
232 Node* next_dest = phase->transform(new AddPNode(base_dest,base_dest,off));
233 assert(phase->C->get_alias_index(adr_type) == phase->C->get_alias_index(phase->type(next_src)->isa_ptr()),
234 "slice of address and input slice don't match");
235 assert(phase->C->get_alias_index(adr_type) == phase->C->get_alias_index(phase->type(next_dest)->isa_ptr()),
236 "slice of address and input slice don't match");
237 BasicType bt = field->layout_type();
238
239 const Type *type;
240 if (bt == T_OBJECT) {
241 if (!field->type()->is_loaded()) {
242 type = TypeInstPtr::BOTTOM;
243 } else {
244 ciType* field_klass = field->type();
245 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
246 }
247 } else {
248 type = Type::get_const_basic_type(bt);
249 }
250
251 Node* v = load(bs, phase, ctl, mem, next_src, adr_type, type, bt);
252 store(bs, phase, ctl, mem, next_dest, adr_type, v, type, bt);
253 }
254
255 if (!finish_transform(phase, can_reshape, ctl, mem)) {
256 // Return NodeSentinel to indicate that the transform failed
257 return NodeSentinel;
258 }
259
260 return mem;
261 }
262
263 bool ArrayCopyNode::prepare_array_copy(PhaseGVN *phase, bool can_reshape,
264 Node*& adr_src,
265 Node*& base_src,
266 Node*& adr_dest,
267 Node*& base_dest,
268 BasicType& copy_type,
269 const Type*& value_type,
270 bool& disjoint_bases) {
271 base_src = in(ArrayCopyNode::Src);
272 base_dest = in(ArrayCopyNode::Dest);
273 const Type* src_type = phase->type(base_src);
274 const TypeAryPtr* ary_src = src_type->isa_aryptr();
275
276 Node* src_offset = in(ArrayCopyNode::SrcPos);
277 Node* dest_offset = in(ArrayCopyNode::DestPos);
278
279 if (is_arraycopy() || is_copyofrange() || is_copyof()) {
280 const Type* dest_type = phase->type(base_dest);
281 const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
282
283 // newly allocated object is guaranteed to not overlap with source object
284 disjoint_bases = is_alloc_tightly_coupled();
285 if (ary_src == nullptr || ary_src->elem() == Type::BOTTOM ||
286 ary_dest == nullptr || ary_dest->elem() == Type::BOTTOM) {
287 // We don't know if arguments are arrays
288 return false;
289 }
290
291 BasicType src_elem = ary_src->elem()->array_element_basic_type();
292 BasicType dest_elem = ary_dest->elem()->array_element_basic_type();
293 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
294 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
295
296 if (src_elem != dest_elem || dest_elem == T_VOID) {
297 // We don't know if arguments are arrays of the same type
298 return false;
299 }
300
301 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
302 if (bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), dest_elem, false, false, BarrierSetC2::Optimization)) {
303 // It's an object array copy but we can't emit the card marking
304 // that is needed
305 return false;
306 }
307
308 value_type = ary_src->elem();
309
310 uint shift = exact_log2(type2aelembytes(dest_elem));
311 uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
312
313 src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
314 if (src_offset->is_top()) {
315 // Offset is out of bounds (the ArrayCopyNode will be removed)
316 return false;
317 }
318 dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
319 if (dest_offset->is_top()) {
320 // Offset is out of bounds (the ArrayCopyNode will be removed)
321 if (can_reshape) {
322 // record src_offset, so it can be deleted later (if it is dead)
323 phase->is_IterGVN()->_worklist.push(src_offset);
324 }
325 return false;
326 }
327
328 Node* hook = new Node(1);
329 hook->init_req(0, dest_offset);
330
331 Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
332
333 hook->destruct(phase);
334
335 Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
336
337 adr_src = phase->transform(new AddPNode(base_src, base_src, src_scale));
338 adr_dest = phase->transform(new AddPNode(base_dest, base_dest, dest_scale));
339
340 adr_src = phase->transform(new AddPNode(base_src, adr_src, phase->MakeConX(header)));
341 adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, phase->MakeConX(header)));
342
343 copy_type = dest_elem;
344 } else {
345 assert(ary_src != nullptr, "should be a clone");
346 assert(is_clonebasic(), "should be");
347
348 disjoint_bases = true;
349
350 BasicType elem = ary_src->isa_aryptr()->elem()->array_element_basic_type();
351 if (is_reference_type(elem, true)) {
352 elem = T_OBJECT;
353 }
354
355 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
356 if (bs->array_copy_requires_gc_barriers(true, elem, true, is_clone_inst(), BarrierSetC2::Optimization)) {
357 return false;
358 }
359
360 adr_src = phase->transform(new AddPNode(base_src, base_src, src_offset));
361 adr_dest = phase->transform(new AddPNode(base_dest, base_dest, dest_offset));
362
363 // The address is offsetted to an aligned address where a raw copy would start.
364 // If the clone copy is decomposed into load-stores - the address is adjusted to
365 // point at where the array starts.
366 const Type* toff = phase->type(src_offset);
367 int offset = toff->isa_long() ? (int) toff->is_long()->get_con() : (int) toff->is_int()->get_con();
368 int diff = arrayOopDesc::base_offset_in_bytes(elem) - offset;
369 assert(diff >= 0, "clone should not start after 1st array element");
370 if (diff > 0) {
371 adr_src = phase->transform(new AddPNode(base_src, adr_src, phase->MakeConX(diff)));
372 adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, phase->MakeConX(diff)));
373 }
374 copy_type = elem;
375 value_type = ary_src->elem();
376 }
377 return true;
378 }
379
380 const TypePtr* ArrayCopyNode::get_address_type(PhaseGVN* phase, const TypePtr* atp, Node* n) {
381 if (atp == TypeOopPtr::BOTTOM) {
382 atp = phase->type(n)->isa_ptr();
383 }
384 // adjust atp to be the correct array element address type
385 return atp->add_offset(Type::OffsetBot);
386 }
387
388 void ArrayCopyNode::array_copy_test_overlap(PhaseGVN *phase, bool can_reshape, bool disjoint_bases, int count, Node*& forward_ctl, Node*& backward_ctl) {
389 Node* ctl = in(TypeFunc::Control);
390 if (!disjoint_bases && count > 1) {
391 Node* src_offset = in(ArrayCopyNode::SrcPos);
392 Node* dest_offset = in(ArrayCopyNode::DestPos);
393 assert(src_offset != nullptr && dest_offset != nullptr, "should be");
394 Node* cmp = phase->transform(new CmpINode(src_offset, dest_offset));
395 Node *bol = phase->transform(new BoolNode(cmp, BoolTest::lt));
396 IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
397
398 phase->transform(iff);
399
400 forward_ctl = phase->transform(new IfFalseNode(iff));
401 backward_ctl = phase->transform(new IfTrueNode(iff));
402 } else {
403 forward_ctl = ctl;
404 }
405 }
406
407 Node* ArrayCopyNode::array_copy_forward(PhaseGVN *phase,
408 bool can_reshape,
409 Node*& forward_ctl,
410 Node* mem,
411 const TypePtr* atp_src,
412 const TypePtr* atp_dest,
413 Node* adr_src,
414 Node* base_src,
415 Node* adr_dest,
416 Node* base_dest,
417 BasicType copy_type,
418 const Type* value_type,
419 int count) {
420 if (!forward_ctl->is_top()) {
421 // copy forward
422 MergeMemNode* mm = MergeMemNode::make(mem);
423
424 if (count > 0) {
425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
426 Node* v = load(bs, phase, forward_ctl, mm, adr_src, atp_src, value_type, copy_type);
427 store(bs, phase, forward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
428 for (int i = 1; i < count; i++) {
429 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
430 Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
431 Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
432 v = load(bs, phase, forward_ctl, mm, next_src, atp_src, value_type, copy_type);
433 store(bs, phase, forward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
434 }
435 } else if (can_reshape) {
436 PhaseIterGVN* igvn = phase->is_IterGVN();
437 igvn->_worklist.push(adr_src);
438 igvn->_worklist.push(adr_dest);
439 }
440 return mm;
441 }
442 return phase->C->top();
443 }
444
445 Node* ArrayCopyNode::array_copy_backward(PhaseGVN *phase,
446 bool can_reshape,
447 Node*& backward_ctl,
448 Node* mem,
449 const TypePtr* atp_src,
450 const TypePtr* atp_dest,
451 Node* adr_src,
452 Node* base_src,
453 Node* adr_dest,
454 Node* base_dest,
455 BasicType copy_type,
456 const Type* value_type,
457 int count) {
458 if (!backward_ctl->is_top()) {
459 // copy backward
460 MergeMemNode* mm = MergeMemNode::make(mem);
461
462 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
463 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");
464
465 if (count > 0) {
466 for (int i = count-1; i >= 1; i--) {
467 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
468 Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
469 Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
470 Node* v = load(bs, phase, backward_ctl, mm, next_src, atp_src, value_type, copy_type);
471 store(bs, phase, backward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
472 }
473 Node* v = load(bs, phase, backward_ctl, mm, adr_src, atp_src, value_type, copy_type);
474 store(bs, phase, backward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
475 } else if (can_reshape) {
476 PhaseIterGVN* igvn = phase->is_IterGVN();
477 igvn->_worklist.push(adr_src);
478 igvn->_worklist.push(adr_dest);
479 }
480 return phase->transform(mm);
481 }
482 return phase->C->top();
483 }
484
485 bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
486 Node* ctl, Node *mem) {
487 if (can_reshape) {
488 PhaseIterGVN* igvn = phase->is_IterGVN();
489 igvn->set_delay_transform(false);
490 if (is_clonebasic()) {
491 Node* out_mem = proj_out(TypeFunc::Memory);
492
493 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
494 if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
495 out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
496 assert(bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, is_clone_inst(), BarrierSetC2::Optimization), "can only happen with card marking");
497 return false;
498 }
499
500 igvn->replace_node(out_mem->raw_out(0), mem);
501
502 Node* out_ctl = proj_out(TypeFunc::Control);
503 igvn->replace_node(out_ctl, ctl);
504 } else {
505 // replace fallthrough projections of the ArrayCopyNode by the
506 // new memory, control and the input IO.
507 CallProjections callprojs;
508 extract_projections(&callprojs, true, false);
509
510 if (callprojs.fallthrough_ioproj != nullptr) {
511 igvn->replace_node(callprojs.fallthrough_ioproj, in(TypeFunc::I_O));
512 }
513 if (callprojs.fallthrough_memproj != nullptr) {
514 igvn->replace_node(callprojs.fallthrough_memproj, mem);
515 }
516 if (callprojs.fallthrough_catchproj != nullptr) {
517 igvn->replace_node(callprojs.fallthrough_catchproj, ctl);
518 }
519
520 // The ArrayCopyNode is not disconnected. It still has the
521 // projections for the exception case. Replace current
522 // ArrayCopyNode with a dummy new one with a top() control so
523 // that this part of the graph stays consistent but is
524 // eventually removed.
525
526 set_req(0, phase->C->top());
527 remove_dead_region(phase, can_reshape);
528 }
529 } else {
530 if (in(TypeFunc::Control) != ctl) {
531 // we can't return new memory and control from Ideal at parse time
532 assert(!is_clonebasic() || UseShenandoahGC, "added control for clone?");
533 phase->record_for_igvn(this);
534 return false;
535 }
536 }
537 return true;
538 }
539
540
541 Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
542 if (remove_dead_region(phase, can_reshape)) return this;
543
544 if (StressArrayCopyMacroNode && !can_reshape) {
545 phase->record_for_igvn(this);
546 return nullptr;
547 }
548
549 // See if it's a small array copy and we can inline it as
550 // loads/stores
551 // Here we can only do:
552 // - arraycopy if all arguments were validated before and we don't
553 // need card marking
554 // - clone for which we don't need to do card marking
555
556 if (!is_clonebasic() && !is_arraycopy_validated() &&
557 !is_copyofrange_validated() && !is_copyof_validated()) {
558 return nullptr;
559 }
560
561 assert(in(TypeFunc::Control) != nullptr &&
562 in(TypeFunc::Memory) != nullptr &&
563 in(ArrayCopyNode::Src) != nullptr &&
564 in(ArrayCopyNode::Dest) != nullptr &&
565 in(ArrayCopyNode::Length) != nullptr &&
566 in(ArrayCopyNode::SrcPos) != nullptr &&
567 in(ArrayCopyNode::DestPos) != nullptr, "broken inputs");
568
569 if (in(TypeFunc::Control)->is_top() ||
570 in(TypeFunc::Memory)->is_top() ||
571 phase->type(in(ArrayCopyNode::Src)) == Type::TOP ||
572 phase->type(in(ArrayCopyNode::Dest)) == Type::TOP ||
573 (in(ArrayCopyNode::SrcPos) != nullptr && in(ArrayCopyNode::SrcPos)->is_top()) ||
574 (in(ArrayCopyNode::DestPos) != nullptr && in(ArrayCopyNode::DestPos)->is_top())) {
575 return nullptr;
576 }
577
578 int count = get_count(phase);
579
580 if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
581 return nullptr;
582 }
583
584 Node* mem = try_clone_instance(phase, can_reshape, count);
585 if (mem != nullptr) {
586 return (mem == NodeSentinel) ? nullptr : mem;
587 }
588
589 Node* adr_src = nullptr;
590 Node* base_src = nullptr;
591 Node* adr_dest = nullptr;
592 Node* base_dest = nullptr;
593 BasicType copy_type = T_ILLEGAL;
594 const Type* value_type = nullptr;
595 bool disjoint_bases = false;
596
597 if (!prepare_array_copy(phase, can_reshape,
598 adr_src, base_src, adr_dest, base_dest,
599 copy_type, value_type, disjoint_bases)) {
600 assert(adr_src == nullptr, "no node can be left behind");
601 assert(adr_dest == nullptr, "no node can be left behind");
602 return nullptr;
603 }
604
605 Node* src = in(ArrayCopyNode::Src);
606 Node* dest = in(ArrayCopyNode::Dest);
607 const TypePtr* atp_src = get_address_type(phase, _src_type, src);
608 const TypePtr* atp_dest = get_address_type(phase, _dest_type, dest);
609 Node* in_mem = in(TypeFunc::Memory);
610
611 if (can_reshape) {
612 assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
613 phase->is_IterGVN()->set_delay_transform(true);
614 }
615
616 Node* backward_ctl = phase->C->top();
617 Node* forward_ctl = phase->C->top();
618 array_copy_test_overlap(phase, can_reshape, disjoint_bases, count, forward_ctl, backward_ctl);
619
620 Node* forward_mem = array_copy_forward(phase, can_reshape, forward_ctl,
621 in_mem,
622 atp_src, atp_dest,
623 adr_src, base_src, adr_dest, base_dest,
624 copy_type, value_type, count);
625
626 Node* backward_mem = array_copy_backward(phase, can_reshape, backward_ctl,
627 in_mem,
628 atp_src, atp_dest,
629 adr_src, base_src, adr_dest, base_dest,
630 copy_type, value_type, count);
631
632 Node* ctl = nullptr;
633 if (!forward_ctl->is_top() && !backward_ctl->is_top()) {
634 ctl = new RegionNode(3);
635 ctl->init_req(1, forward_ctl);
636 ctl->init_req(2, backward_ctl);
637 ctl = phase->transform(ctl);
638 MergeMemNode* forward_mm = forward_mem->as_MergeMem();
639 MergeMemNode* backward_mm = backward_mem->as_MergeMem();
640 for (MergeMemStream mms(forward_mm, backward_mm); mms.next_non_empty2(); ) {
641 if (mms.memory() != mms.memory2()) {
642 Node* phi = new PhiNode(ctl, Type::MEMORY, phase->C->get_adr_type(mms.alias_idx()));
643 phi->init_req(1, mms.memory());
644 phi->init_req(2, mms.memory2());
645 phi = phase->transform(phi);
646 mms.set_memory(phi);
647 }
648 }
649 mem = forward_mem;
650 } else if (!forward_ctl->is_top()) {
651 ctl = forward_ctl;
652 mem = forward_mem;
653 } else {
654 assert(!backward_ctl->is_top(), "no copy?");
655 ctl = backward_ctl;
656 mem = backward_mem;
657 }
658
659 if (can_reshape) {
660 assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
661 phase->is_IterGVN()->set_delay_transform(false);
662 }
663
664 if (!finish_transform(phase, can_reshape, ctl, mem)) {
665 if (can_reshape) {
666 // put in worklist, so that if it happens to be dead it is removed
667 phase->is_IterGVN()->_worklist.push(mem);
668 }
669 return nullptr;
670 }
671
672 return mem;
673 }
674
675 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
676 Node* dest = in(ArrayCopyNode::Dest);
677 if (dest->is_top()) {
678 return false;
679 }
680 const TypeOopPtr* dest_t = phase->type(dest)->is_oopptr();
681 assert(!dest_t->is_known_instance() || _dest_type->is_known_instance(), "result of EA not recorded");
682 assert(in(ArrayCopyNode::Src)->is_top() || !phase->type(in(ArrayCopyNode::Src))->is_oopptr()->is_known_instance() ||
683 _src_type->is_known_instance(), "result of EA not recorded");
684
685 if (_dest_type != TypeOopPtr::BOTTOM || t_oop->is_known_instance()) {
686 assert(_dest_type == TypeOopPtr::BOTTOM || _dest_type->is_known_instance(), "result of EA is known instance");
687 return t_oop->instance_id() == _dest_type->instance_id();
688 }
689
690 return CallNode::may_modify_arraycopy_helper(dest_t, t_oop, phase);
691 }
692
693 bool ArrayCopyNode::may_modify_helper(const TypeOopPtr* t_oop, Node* n, PhaseValues* phase, ArrayCopyNode*& ac) {
694 if (n != nullptr &&
695 n->is_ArrayCopy() &&
696 n->as_ArrayCopy()->may_modify(t_oop, phase)) {
697 ac = n->as_ArrayCopy();
698 return true;
699 }
700 return false;
701 }
702
703 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, MemBarNode* mb, PhaseValues* phase, ArrayCopyNode*& ac) {
704 if (mb->trailing_expanded_array_copy()) {
705 return true;
706 }
707
708 Node* c = mb->in(0);
709
710 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
711 // step over g1 gc barrier if we're at e.g. a clone with ReduceInitialCardMarks off
712 c = bs->step_over_gc_barrier(c);
713
714 CallNode* call = nullptr;
715 guarantee(c != nullptr, "step_over_gc_barrier failed, there must be something to step to.");
716 if (c->is_Region()) {
717 for (uint i = 1; i < c->req(); i++) {
718 if (c->in(i) != nullptr) {
719 Node* n = c->in(i)->in(0);
720 if (may_modify_helper(t_oop, n, phase, ac)) {
721 assert(c == mb->in(0), "only for clone");
722 return true;
723 }
724 }
725 }
726 } else if (may_modify_helper(t_oop, c->in(0), phase, ac)) {
727 #ifdef ASSERT
728 bool use_ReduceInitialCardMarks = BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet) &&
729 static_cast<CardTableBarrierSetC2*>(bs)->use_ReduceInitialCardMarks();
730 assert(c == mb->in(0) || (ac->is_clonebasic() && !use_ReduceInitialCardMarks), "only for clone");
731 #endif
732 return true;
733 }
734
735 return false;
736 }
737
738 // Does this array copy modify offsets between offset_lo and offset_hi
739 // in the destination array
740 // if must_modify is false, return true if the copy could write
741 // between offset_lo and offset_hi
742 // if must_modify is true, return true if the copy is guaranteed to
743 // write between offset_lo and offset_hi
744 bool ArrayCopyNode::modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseValues* phase, bool must_modify) const {
745 assert(_kind == ArrayCopy || _kind == CopyOf || _kind == CopyOfRange, "only for real array copies");
746
747 Node* dest = in(Dest);
748 Node* dest_pos = in(DestPos);
749 Node* len = in(Length);
750
751 const TypeInt *dest_pos_t = phase->type(dest_pos)->isa_int();
752 const TypeInt *len_t = phase->type(len)->isa_int();
753 const TypeAryPtr* ary_t = phase->type(dest)->isa_aryptr();
754
755 if (dest_pos_t == nullptr || len_t == nullptr || ary_t == nullptr) {
756 return !must_modify;
757 }
758
759 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
760 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
761
762 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
763 uint elemsize = type2aelembytes(ary_elem);
764
765 jlong dest_pos_plus_len_lo = (((jlong)dest_pos_t->_lo) + len_t->_lo) * elemsize + header;
766 jlong dest_pos_plus_len_hi = (((jlong)dest_pos_t->_hi) + len_t->_hi) * elemsize + header;
767 jlong dest_pos_lo = ((jlong)dest_pos_t->_lo) * elemsize + header;
768 jlong dest_pos_hi = ((jlong)dest_pos_t->_hi) * elemsize + header;
769
770 if (must_modify) {
771 if (offset_lo >= dest_pos_hi && offset_hi < dest_pos_plus_len_lo) {
772 return true;
773 }
774 } else {
775 if (offset_hi >= dest_pos_lo && offset_lo < dest_pos_plus_len_hi) {
776 return true;
777 }
778 }
779 return false;
780 }
781
782 // As an optimization, choose optimum vector size for copy length known at compile time.
783 int ArrayCopyNode::get_partial_inline_vector_lane_count(BasicType type, int const_len) {
784 int lane_count = ArrayOperationPartialInlineSize/type2aelembytes(type);
785 if (const_len > 0) {
786 int size_in_bytes = const_len * type2aelembytes(type);
787 if (size_in_bytes <= 16)
788 lane_count = 16/type2aelembytes(type);
789 else if (size_in_bytes > 16 && size_in_bytes <= 32)
790 lane_count = 32/type2aelembytes(type);
791 }
792 return lane_count;
793 }