1 /*
2 * Copyright (c) 2016, 2023, 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 "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "gc/shared/c2/cardTableBarrierSetC2.hpp"
29 #include "gc/shared/gc_globals.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/graphKit.hpp"
32 #include "runtime/sharedRuntime.hpp"
33 #include "utilities/macros.hpp"
34 #include "utilities/powerOfTwo.hpp"
35
36 ArrayCopyNode::ArrayCopyNode(Compile* C, bool alloc_tightly_coupled, bool has_negative_length_guard)
37 : CallNode(arraycopy_type(), nullptr, TypePtr::BOTTOM),
38 _kind(None),
39 _alloc_tightly_coupled(alloc_tightly_coupled),
40 _has_negative_length_guard(has_negative_length_guard),
41 _arguments_validated(false),
42 _src_type(TypeOopPtr::BOTTOM),
43 _dest_type(TypeOopPtr::BOTTOM) {
44 init_class_id(Class_ArrayCopy);
45 init_flags(Flag_is_macro);
46 C->add_macro_node(this);
47 }
48
49 uint ArrayCopyNode::size_of() const { return sizeof(*this); }
50
51 ArrayCopyNode* ArrayCopyNode::make(GraphKit* kit, bool may_throw,
52 Node* src, Node* src_offset,
53 Node* dest, Node* dest_offset,
54 Node* length,
55 bool alloc_tightly_coupled,
56 bool has_negative_length_guard,
57 Node* src_klass, Node* dest_klass,
58 Node* src_length, Node* dest_length) {
59
60 ArrayCopyNode* ac = new ArrayCopyNode(kit->C, alloc_tightly_coupled, has_negative_length_guard);
61 kit->set_predefined_input_for_runtime_call(ac);
62
63 ac->init_req(ArrayCopyNode::Src, src);
64 ac->init_req(ArrayCopyNode::SrcPos, src_offset);
65 ac->init_req(ArrayCopyNode::Dest, dest);
66 ac->init_req(ArrayCopyNode::DestPos, dest_offset);
67 ac->init_req(ArrayCopyNode::Length, length);
68 ac->init_req(ArrayCopyNode::SrcLen, src_length);
69 ac->init_req(ArrayCopyNode::DestLen, dest_length);
70 ac->init_req(ArrayCopyNode::SrcKlass, src_klass);
71 ac->init_req(ArrayCopyNode::DestKlass, dest_klass);
72
73 if (may_throw) {
74 ac->set_req(TypeFunc::I_O , kit->i_o());
75 kit->add_safepoint_edges(ac, false);
76 }
77
78 return ac;
79 }
80
81 void ArrayCopyNode::connect_outputs(GraphKit* kit, bool deoptimize_on_exception) {
82 kit->set_all_memory_call(this, true);
83 kit->set_control(kit->gvn().transform(new ProjNode(this,TypeFunc::Control)));
84 kit->set_i_o(kit->gvn().transform(new ProjNode(this, TypeFunc::I_O)));
85 kit->make_slow_call_ex(this, kit->env()->Throwable_klass(), true, deoptimize_on_exception);
86 kit->set_all_memory_call(this);
87 }
88
89 #ifndef PRODUCT
90 const char* ArrayCopyNode::_kind_names[] = {"arraycopy", "arraycopy, validated arguments", "clone", "oop array clone", "CopyOf", "CopyOfRange"};
91
92 void ArrayCopyNode::dump_spec(outputStream *st) const {
93 CallNode::dump_spec(st);
94 st->print(" (%s%s)", _kind_names[_kind], _alloc_tightly_coupled ? ", tightly coupled allocation" : "");
95 }
96
97 void ArrayCopyNode::dump_compact_spec(outputStream* st) const {
98 st->print("%s%s", _kind_names[_kind], _alloc_tightly_coupled ? ",tight" : "");
99 }
100 #endif
101
102 intptr_t ArrayCopyNode::get_length_if_constant(PhaseGVN *phase) const {
103 // check that length is constant
104 Node* length = in(ArrayCopyNode::Length);
105 const Type* length_type = phase->type(length);
106
107 if (length_type == Type::TOP) {
108 return -1;
109 }
110
111 assert(is_clonebasic() || is_arraycopy() || is_copyof() || is_copyofrange(), "unexpected array copy type");
112
113 return is_clonebasic() ? length->find_intptr_t_con(-1) : length->find_int_con(-1);
114 }
115
116 int ArrayCopyNode::get_count(PhaseGVN *phase) const {
117 Node* src = in(ArrayCopyNode::Src);
118 const Type* src_type = phase->type(src);
119
120 if (is_clonebasic()) {
121 if (src_type->isa_instptr()) {
122 const TypeInstPtr* inst_src = src_type->is_instptr();
123 ciInstanceKlass* ik = inst_src->instance_klass();
124 // ciInstanceKlass::nof_nonstatic_fields() doesn't take injected
125 // fields into account. They are rare anyway so easier to simply
126 // skip instances with injected fields.
127 if ((!inst_src->klass_is_exact() && (ik->is_interface() || ik->has_subklass())) || ik->has_injected_fields()) {
128 return -1;
129 }
130 int nb_fields = ik->nof_nonstatic_fields();
131 return nb_fields;
132 } else {
133 const TypeAryPtr* ary_src = src_type->isa_aryptr();
134 assert (ary_src != nullptr, "not an array or instance?");
135 // clone passes a length as a rounded number of longs. If we're
136 // cloning an array we'll do it element by element. If the
137 // length of the input array is constant, ArrayCopyNode::Length
138 // must be too. Note that the opposite does not need to hold,
139 // because different input array lengths (e.g. int arrays with
140 // 3 or 4 elements) might lead to the same length input
141 // (e.g. 2 double-words).
142 assert(!ary_src->size()->is_con() || (get_length_if_constant(phase) >= 0) ||
143 phase->is_IterGVN() || phase->C->inlining_incrementally() || StressReflectiveCode, "inconsistent");
144 if (ary_src->size()->is_con()) {
145 return ary_src->size()->get_con();
146 }
147 return -1;
148 }
149 }
150
151 return get_length_if_constant(phase);
152 }
153
154 Node* ArrayCopyNode::load(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, const Type *type, BasicType bt) {
155 // 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
156 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk floating
157 // above runtime checks that guarantee it is within bounds.
158 DecoratorSet decorators = C2_READ_ACCESS | C2_CONTROL_DEPENDENT_LOAD | IN_HEAP | C2_ARRAY_COPY | C2_UNKNOWN_CONTROL_LOAD;
159 C2AccessValuePtr addr(adr, adr_type);
160 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
161 Node* res = bs->load_at(access, type);
162 ctl = access.ctl();
163 return res;
164 }
165
166 void ArrayCopyNode::store(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, Node* val, const Type *type, BasicType bt) {
167 DecoratorSet decorators = C2_WRITE_ACCESS | IN_HEAP | C2_ARRAY_COPY;
168 if (is_alloc_tightly_coupled()) {
169 decorators |= C2_TIGHTLY_COUPLED_ALLOC;
170 }
171 C2AccessValuePtr addr(adr, adr_type);
172 C2AccessValue value(val, type);
173 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
174 bs->store_at(access, value);
175 ctl = access.ctl();
176 }
177
178
179 Node* ArrayCopyNode::try_clone_instance(PhaseGVN *phase, bool can_reshape, int count) {
180 if (!is_clonebasic()) {
181 return nullptr;
182 }
183
184 Node* base_src = in(ArrayCopyNode::Src);
185 Node* base_dest = in(ArrayCopyNode::Dest);
186 Node* ctl = in(TypeFunc::Control);
187 Node* in_mem = in(TypeFunc::Memory);
188
189 const Type* src_type = phase->type(base_src);
190 const TypeInstPtr* inst_src = src_type->isa_instptr();
191 if (inst_src == nullptr) {
192 return nullptr;
193 }
194
195 MergeMemNode* mem = phase->transform(MergeMemNode::make(in_mem))->as_MergeMem();
196 if (can_reshape) {
197 phase->is_IterGVN()->_worklist.push(mem);
198 }
199
200
201 ciInstanceKlass* ik = inst_src->instance_klass();
202
203 if (!inst_src->klass_is_exact()) {
204 assert(!ik->is_interface(), "inconsistent klass hierarchy");
205 if (ik->has_subklass()) {
206 // Concurrent class loading.
207 // Fail fast and return NodeSentinel to indicate that the transform failed.
208 return NodeSentinel;
209 } else {
210 phase->C->dependencies()->assert_leaf_type(ik);
211 }
212 }
213
214 assert(ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem, "too many fields");
215
216 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
217 for (int i = 0; i < count; i++) {
218 ciField* field = ik->nonstatic_field_at(i);
219 const TypePtr* adr_type = phase->C->alias_type(field)->adr_type();
220 Node* off = phase->MakeConX(field->offset_in_bytes());
221 Node* next_src = phase->transform(new AddPNode(base_src,base_src,off));
222 Node* next_dest = phase->transform(new AddPNode(base_dest,base_dest,off));
223 BasicType bt = field->layout_type();
224
225 const Type *type;
226 if (bt == T_OBJECT) {
227 if (!field->type()->is_loaded()) {
228 type = TypeInstPtr::BOTTOM;
229 } else {
230 ciType* field_klass = field->type();
231 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
232 }
233 } else {
234 type = Type::get_const_basic_type(bt);
235 }
236
237 Node* v = load(bs, phase, ctl, mem, next_src, adr_type, type, bt);
238 store(bs, phase, ctl, mem, next_dest, adr_type, v, type, bt);
239 }
240
241 if (!finish_transform(phase, can_reshape, ctl, mem)) {
242 // Return NodeSentinel to indicate that the transform failed
243 return NodeSentinel;
244 }
245
246 return mem;
247 }
248
249 bool ArrayCopyNode::prepare_array_copy(PhaseGVN *phase, bool can_reshape,
250 Node*& adr_src,
251 Node*& base_src,
252 Node*& adr_dest,
253 Node*& base_dest,
254 BasicType& copy_type,
255 const Type*& value_type,
256 bool& disjoint_bases) {
257 base_src = in(ArrayCopyNode::Src);
258 base_dest = in(ArrayCopyNode::Dest);
259 const Type* src_type = phase->type(base_src);
260 const TypeAryPtr* ary_src = src_type->isa_aryptr();
261
262 Node* src_offset = in(ArrayCopyNode::SrcPos);
263 Node* dest_offset = in(ArrayCopyNode::DestPos);
264
265 if (is_arraycopy() || is_copyofrange() || is_copyof()) {
266 const Type* dest_type = phase->type(base_dest);
267 const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
268
269 // newly allocated object is guaranteed to not overlap with source object
270 disjoint_bases = is_alloc_tightly_coupled();
271 if (ary_src == nullptr || ary_src->elem() == Type::BOTTOM ||
272 ary_dest == nullptr || ary_dest->elem() == Type::BOTTOM) {
273 // We don't know if arguments are arrays
274 return false;
275 }
276
277 BasicType src_elem = ary_src->elem()->array_element_basic_type();
278 BasicType dest_elem = ary_dest->elem()->array_element_basic_type();
279 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
280 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
281
282 if (src_elem != dest_elem || dest_elem == T_VOID) {
283 // We don't know if arguments are arrays of the same type
284 return false;
285 }
286
287 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
288 if (bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), dest_elem, false, false, BarrierSetC2::Optimization)) {
289 // It's an object array copy but we can't emit the card marking
290 // that is needed
291 return false;
292 }
293
294 value_type = ary_src->elem();
295
296 uint shift = exact_log2(type2aelembytes(dest_elem));
297 uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
298
299 src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
300 if (src_offset->is_top()) {
301 // Offset is out of bounds (the ArrayCopyNode will be removed)
302 return false;
303 }
304 dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
305 if (dest_offset->is_top()) {
306 // Offset is out of bounds (the ArrayCopyNode will be removed)
307 if (can_reshape) {
308 // record src_offset, so it can be deleted later (if it is dead)
309 phase->is_IterGVN()->_worklist.push(src_offset);
310 }
311 return false;
312 }
313
314 Node* hook = new Node(1);
315 hook->init_req(0, dest_offset);
316
317 Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
318
319 hook->destruct(phase);
320
321 Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
322
323 adr_src = phase->transform(new AddPNode(base_src, base_src, src_scale));
324 adr_dest = phase->transform(new AddPNode(base_dest, base_dest, dest_scale));
325
326 adr_src = phase->transform(new AddPNode(base_src, adr_src, phase->MakeConX(header)));
327 adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, phase->MakeConX(header)));
328
329 copy_type = dest_elem;
330 } else {
331 assert(ary_src != nullptr, "should be a clone");
332 assert(is_clonebasic(), "should be");
333
334 disjoint_bases = true;
335
336 BasicType elem = ary_src->isa_aryptr()->elem()->array_element_basic_type();
337 if (is_reference_type(elem, true)) {
338 elem = T_OBJECT;
339 }
340
341 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
342 if (bs->array_copy_requires_gc_barriers(true, elem, true, is_clone_inst(), BarrierSetC2::Optimization)) {
343 return false;
344 }
345
346 adr_src = phase->transform(new AddPNode(base_src, base_src, src_offset));
347 adr_dest = phase->transform(new AddPNode(base_dest, base_dest, dest_offset));
348
349 // The address is offsetted to an aligned address where a raw copy would start.
350 // If the clone copy is decomposed into load-stores - the address is adjusted to
351 // point at where the array starts.
352 const Type* toff = phase->type(src_offset);
353 int offset = toff->isa_long() ? (int) toff->is_long()->get_con() : (int) toff->is_int()->get_con();
354 int diff = arrayOopDesc::base_offset_in_bytes(elem) - offset;
355 assert(diff >= 0, "clone should not start after 1st array element");
356 if (diff > 0) {
357 adr_src = phase->transform(new AddPNode(base_src, adr_src, phase->MakeConX(diff)));
358 adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, phase->MakeConX(diff)));
359 }
360 copy_type = elem;
361 value_type = ary_src->elem();
362 }
363 return true;
364 }
365
366 const TypePtr* ArrayCopyNode::get_address_type(PhaseGVN* phase, const TypePtr* atp, Node* n) {
367 if (atp == TypeOopPtr::BOTTOM) {
368 atp = phase->type(n)->isa_ptr();
369 }
370 // adjust atp to be the correct array element address type
371 return atp->add_offset(Type::OffsetBot);
372 }
373
374 void ArrayCopyNode::array_copy_test_overlap(PhaseGVN *phase, bool can_reshape, bool disjoint_bases, int count, Node*& forward_ctl, Node*& backward_ctl) {
375 Node* ctl = in(TypeFunc::Control);
376 if (!disjoint_bases && count > 1) {
377 Node* src_offset = in(ArrayCopyNode::SrcPos);
378 Node* dest_offset = in(ArrayCopyNode::DestPos);
379 assert(src_offset != nullptr && dest_offset != nullptr, "should be");
380 Node* cmp = phase->transform(new CmpINode(src_offset, dest_offset));
381 Node *bol = phase->transform(new BoolNode(cmp, BoolTest::lt));
382 IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
383
384 phase->transform(iff);
385
386 forward_ctl = phase->transform(new IfFalseNode(iff));
387 backward_ctl = phase->transform(new IfTrueNode(iff));
388 } else {
389 forward_ctl = ctl;
390 }
391 }
392
393 Node* ArrayCopyNode::array_copy_forward(PhaseGVN *phase,
394 bool can_reshape,
395 Node*& forward_ctl,
396 Node* mem,
397 const TypePtr* atp_src,
398 const TypePtr* atp_dest,
399 Node* adr_src,
400 Node* base_src,
401 Node* adr_dest,
402 Node* base_dest,
403 BasicType copy_type,
404 const Type* value_type,
405 int count) {
406 if (!forward_ctl->is_top()) {
407 // copy forward
408 MergeMemNode* mm = MergeMemNode::make(mem);
409
410 if (count > 0) {
411 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
412 Node* v = load(bs, phase, forward_ctl, mm, adr_src, atp_src, value_type, copy_type);
413 store(bs, phase, forward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
414 for (int i = 1; i < count; i++) {
415 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
416 Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
417 Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
418 v = load(bs, phase, forward_ctl, mm, next_src, atp_src, value_type, copy_type);
419 store(bs, phase, forward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
420 }
421 } else if (can_reshape) {
422 PhaseIterGVN* igvn = phase->is_IterGVN();
423 igvn->_worklist.push(adr_src);
424 igvn->_worklist.push(adr_dest);
425 }
426 return mm;
427 }
428 return phase->C->top();
429 }
430
431 Node* ArrayCopyNode::array_copy_backward(PhaseGVN *phase,
432 bool can_reshape,
433 Node*& backward_ctl,
434 Node* mem,
435 const TypePtr* atp_src,
436 const TypePtr* atp_dest,
437 Node* adr_src,
438 Node* base_src,
439 Node* adr_dest,
440 Node* base_dest,
441 BasicType copy_type,
442 const Type* value_type,
443 int count) {
444 if (!backward_ctl->is_top()) {
445 // copy backward
446 MergeMemNode* mm = MergeMemNode::make(mem);
447
448 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
449 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");
450
451 if (count > 0) {
452 for (int i = count-1; i >= 1; i--) {
453 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
454 Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
455 Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
456 Node* v = load(bs, phase, backward_ctl, mm, next_src, atp_src, value_type, copy_type);
457 store(bs, phase, backward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
458 }
459 Node* v = load(bs, phase, backward_ctl, mm, adr_src, atp_src, value_type, copy_type);
460 store(bs, phase, backward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
461 } else if (can_reshape) {
462 PhaseIterGVN* igvn = phase->is_IterGVN();
463 igvn->_worklist.push(adr_src);
464 igvn->_worklist.push(adr_dest);
465 }
466 return phase->transform(mm);
467 }
468 return phase->C->top();
469 }
470
471 bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
472 Node* ctl, Node *mem) {
473 if (can_reshape) {
474 PhaseIterGVN* igvn = phase->is_IterGVN();
475 igvn->set_delay_transform(false);
476 if (is_clonebasic()) {
477 Node* out_mem = proj_out(TypeFunc::Memory);
478
479 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
480 if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
481 out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
482 assert(bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, is_clone_inst(), BarrierSetC2::Optimization), "can only happen with card marking");
483 return false;
484 }
485
486 igvn->replace_node(out_mem->raw_out(0), mem);
487
488 Node* out_ctl = proj_out(TypeFunc::Control);
489 igvn->replace_node(out_ctl, ctl);
490 } else {
491 // replace fallthrough projections of the ArrayCopyNode by the
492 // new memory, control and the input IO.
493 CallProjections callprojs;
494 extract_projections(&callprojs, true, false);
495
496 if (callprojs.fallthrough_ioproj != nullptr) {
497 igvn->replace_node(callprojs.fallthrough_ioproj, in(TypeFunc::I_O));
498 }
499 if (callprojs.fallthrough_memproj != nullptr) {
500 igvn->replace_node(callprojs.fallthrough_memproj, mem);
501 }
502 if (callprojs.fallthrough_catchproj != nullptr) {
503 igvn->replace_node(callprojs.fallthrough_catchproj, ctl);
504 }
505
506 // The ArrayCopyNode is not disconnected. It still has the
507 // projections for the exception case. Replace current
508 // ArrayCopyNode with a dummy new one with a top() control so
509 // that this part of the graph stays consistent but is
510 // eventually removed.
511
512 set_req(0, phase->C->top());
513 remove_dead_region(phase, can_reshape);
514 }
515 } else {
516 if (in(TypeFunc::Control) != ctl) {
517 // we can't return new memory and control from Ideal at parse time
518 assert(!is_clonebasic() || UseShenandoahGC, "added control for clone?");
519 phase->record_for_igvn(this);
520 return false;
521 }
522 }
523 return true;
524 }
525
526
527 Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
528 if (remove_dead_region(phase, can_reshape)) return this;
529
530 if (StressArrayCopyMacroNode && !can_reshape) {
531 phase->record_for_igvn(this);
532 return nullptr;
533 }
534
535 // See if it's a small array copy and we can inline it as
536 // loads/stores
537 // Here we can only do:
538 // - arraycopy if all arguments were validated before and we don't
539 // need card marking
540 // - clone for which we don't need to do card marking
541
542 if (!is_clonebasic() && !is_arraycopy_validated() &&
543 !is_copyofrange_validated() && !is_copyof_validated()) {
544 return nullptr;
545 }
546
547 assert(in(TypeFunc::Control) != nullptr &&
548 in(TypeFunc::Memory) != nullptr &&
549 in(ArrayCopyNode::Src) != nullptr &&
550 in(ArrayCopyNode::Dest) != nullptr &&
551 in(ArrayCopyNode::Length) != nullptr &&
552 in(ArrayCopyNode::SrcPos) != nullptr &&
553 in(ArrayCopyNode::DestPos) != nullptr, "broken inputs");
554
555 if (in(TypeFunc::Control)->is_top() ||
556 in(TypeFunc::Memory)->is_top() ||
557 phase->type(in(ArrayCopyNode::Src)) == Type::TOP ||
558 phase->type(in(ArrayCopyNode::Dest)) == Type::TOP ||
559 (in(ArrayCopyNode::SrcPos) != nullptr && in(ArrayCopyNode::SrcPos)->is_top()) ||
560 (in(ArrayCopyNode::DestPos) != nullptr && in(ArrayCopyNode::DestPos)->is_top())) {
561 return nullptr;
562 }
563
564 int count = get_count(phase);
565
566 if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
567 return nullptr;
568 }
569
570 Node* mem = try_clone_instance(phase, can_reshape, count);
571 if (mem != nullptr) {
572 return (mem == NodeSentinel) ? nullptr : mem;
573 }
574
575 Node* adr_src = nullptr;
576 Node* base_src = nullptr;
577 Node* adr_dest = nullptr;
578 Node* base_dest = nullptr;
579 BasicType copy_type = T_ILLEGAL;
580 const Type* value_type = nullptr;
581 bool disjoint_bases = false;
582
583 if (!prepare_array_copy(phase, can_reshape,
584 adr_src, base_src, adr_dest, base_dest,
585 copy_type, value_type, disjoint_bases)) {
586 assert(adr_src == nullptr, "no node can be left behind");
587 assert(adr_dest == nullptr, "no node can be left behind");
588 return nullptr;
589 }
590
591 Node* src = in(ArrayCopyNode::Src);
592 Node* dest = in(ArrayCopyNode::Dest);
593 const TypePtr* atp_src = get_address_type(phase, _src_type, src);
594 const TypePtr* atp_dest = get_address_type(phase, _dest_type, dest);
595 Node* in_mem = in(TypeFunc::Memory);
596
597 if (can_reshape) {
598 assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
599 phase->is_IterGVN()->set_delay_transform(true);
600 }
601
602 Node* backward_ctl = phase->C->top();
603 Node* forward_ctl = phase->C->top();
604 array_copy_test_overlap(phase, can_reshape, disjoint_bases, count, forward_ctl, backward_ctl);
605
606 Node* forward_mem = array_copy_forward(phase, can_reshape, forward_ctl,
607 in_mem,
608 atp_src, atp_dest,
609 adr_src, base_src, adr_dest, base_dest,
610 copy_type, value_type, count);
611
612 Node* backward_mem = array_copy_backward(phase, can_reshape, backward_ctl,
613 in_mem,
614 atp_src, atp_dest,
615 adr_src, base_src, adr_dest, base_dest,
616 copy_type, value_type, count);
617
618 Node* ctl = nullptr;
619 if (!forward_ctl->is_top() && !backward_ctl->is_top()) {
620 ctl = new RegionNode(3);
621 ctl->init_req(1, forward_ctl);
622 ctl->init_req(2, backward_ctl);
623 ctl = phase->transform(ctl);
624 MergeMemNode* forward_mm = forward_mem->as_MergeMem();
625 MergeMemNode* backward_mm = backward_mem->as_MergeMem();
626 for (MergeMemStream mms(forward_mm, backward_mm); mms.next_non_empty2(); ) {
627 if (mms.memory() != mms.memory2()) {
628 Node* phi = new PhiNode(ctl, Type::MEMORY, phase->C->get_adr_type(mms.alias_idx()));
629 phi->init_req(1, mms.memory());
630 phi->init_req(2, mms.memory2());
631 phi = phase->transform(phi);
632 mms.set_memory(phi);
633 }
634 }
635 mem = forward_mem;
636 } else if (!forward_ctl->is_top()) {
637 ctl = forward_ctl;
638 mem = forward_mem;
639 } else {
640 assert(!backward_ctl->is_top(), "no copy?");
641 ctl = backward_ctl;
642 mem = backward_mem;
643 }
644
645 if (can_reshape) {
646 assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
647 phase->is_IterGVN()->set_delay_transform(false);
648 }
649
650 if (!finish_transform(phase, can_reshape, ctl, mem)) {
651 if (can_reshape) {
652 // put in worklist, so that if it happens to be dead it is removed
653 phase->is_IterGVN()->_worklist.push(mem);
654 }
655 return nullptr;
656 }
657
658 return mem;
659 }
660
661 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
662 Node* dest = in(ArrayCopyNode::Dest);
663 if (dest->is_top()) {
664 return false;
665 }
666 const TypeOopPtr* dest_t = phase->type(dest)->is_oopptr();
667 assert(!dest_t->is_known_instance() || _dest_type->is_known_instance(), "result of EA not recorded");
668 assert(in(ArrayCopyNode::Src)->is_top() || !phase->type(in(ArrayCopyNode::Src))->is_oopptr()->is_known_instance() ||
669 _src_type->is_known_instance(), "result of EA not recorded");
670
671 if (_dest_type != TypeOopPtr::BOTTOM || t_oop->is_known_instance()) {
672 assert(_dest_type == TypeOopPtr::BOTTOM || _dest_type->is_known_instance(), "result of EA is known instance");
673 return t_oop->instance_id() == _dest_type->instance_id();
674 }
675
676 return CallNode::may_modify_arraycopy_helper(dest_t, t_oop, phase);
677 }
678
679 bool ArrayCopyNode::may_modify_helper(const TypeOopPtr* t_oop, Node* n, PhaseValues* phase, CallNode*& call) {
680 if (n != nullptr &&
681 n->is_Call() &&
682 n->as_Call()->may_modify(t_oop, phase) &&
683 (n->as_Call()->is_ArrayCopy() || n->as_Call()->is_call_to_arraycopystub())) {
684 call = n->as_Call();
685 return true;
686 }
687 return false;
688 }
689
690 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, MemBarNode* mb, PhaseValues* phase, ArrayCopyNode*& ac) {
691 Node* c = mb->in(0);
692
693 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
694 // step over g1 gc barrier if we're at e.g. a clone with ReduceInitialCardMarks off
695 c = bs->step_over_gc_barrier(c);
696
697 CallNode* call = nullptr;
698 guarantee(c != nullptr, "step_over_gc_barrier failed, there must be something to step to.");
699 if (c->is_Region()) {
700 for (uint i = 1; i < c->req(); i++) {
701 if (c->in(i) != nullptr) {
702 Node* n = c->in(i)->in(0);
703 if (may_modify_helper(t_oop, n, phase, call)) {
704 ac = call->isa_ArrayCopy();
705 assert(c == mb->in(0), "only for clone");
706 return true;
707 }
708 }
709 }
710 } else if (may_modify_helper(t_oop, c->in(0), phase, call)) {
711 ac = call->isa_ArrayCopy();
712 #ifdef ASSERT
713 bool use_ReduceInitialCardMarks = BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet) &&
714 static_cast<CardTableBarrierSetC2*>(bs)->use_ReduceInitialCardMarks();
715 assert(c == mb->in(0) || (ac != nullptr && ac->is_clonebasic() && !use_ReduceInitialCardMarks), "only for clone");
716 #endif
717 return true;
718 } else if (mb->trailing_partial_array_copy()) {
719 return true;
720 }
721
722 return false;
723 }
724
725 // Does this array copy modify offsets between offset_lo and offset_hi
726 // in the destination array
727 // if must_modify is false, return true if the copy could write
728 // between offset_lo and offset_hi
729 // if must_modify is true, return true if the copy is guaranteed to
730 // write between offset_lo and offset_hi
731 bool ArrayCopyNode::modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseValues* phase, bool must_modify) const {
732 assert(_kind == ArrayCopy || _kind == CopyOf || _kind == CopyOfRange, "only for real array copies");
733
734 Node* dest = in(Dest);
735 Node* dest_pos = in(DestPos);
736 Node* len = in(Length);
737
738 const TypeInt *dest_pos_t = phase->type(dest_pos)->isa_int();
739 const TypeInt *len_t = phase->type(len)->isa_int();
740 const TypeAryPtr* ary_t = phase->type(dest)->isa_aryptr();
741
742 if (dest_pos_t == nullptr || len_t == nullptr || ary_t == nullptr) {
743 return !must_modify;
744 }
745
746 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
747 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
748
749 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
750 uint elemsize = type2aelembytes(ary_elem);
751
752 jlong dest_pos_plus_len_lo = (((jlong)dest_pos_t->_lo) + len_t->_lo) * elemsize + header;
753 jlong dest_pos_plus_len_hi = (((jlong)dest_pos_t->_hi) + len_t->_hi) * elemsize + header;
754 jlong dest_pos_lo = ((jlong)dest_pos_t->_lo) * elemsize + header;
755 jlong dest_pos_hi = ((jlong)dest_pos_t->_hi) * elemsize + header;
756
757 if (must_modify) {
758 if (offset_lo >= dest_pos_hi && offset_hi < dest_pos_plus_len_lo) {
759 return true;
760 }
761 } else {
762 if (offset_hi >= dest_pos_lo && offset_lo < dest_pos_plus_len_hi) {
763 return true;
764 }
765 }
766 return false;
767 }
768
769 // As an optimization, choose optimum vector size for copy length known at compile time.
770 int ArrayCopyNode::get_partial_inline_vector_lane_count(BasicType type, int const_len) {
771 int lane_count = ArrayOperationPartialInlineSize/type2aelembytes(type);
772 if (const_len > 0) {
773 int size_in_bytes = const_len * type2aelembytes(type);
774 if (size_in_bytes <= 16)
775 lane_count = 16/type2aelembytes(type);
776 else if (size_in_bytes > 16 && size_in_bytes <= 32)
777 lane_count = 32/type2aelembytes(type);
778 }
779 return lane_count;
780 }