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,
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();
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) {
274 Node* src_offset = in(ArrayCopyNode::SrcPos);
275 Node* dest_offset = in(ArrayCopyNode::DestPos);
276
277 if (is_arraycopy() || is_copyofrange() || is_copyof()) {
278 const Type* dest_type = phase->type(base_dest);
279 const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
280
281 // newly allocated object is guaranteed to not overlap with source object
282 disjoint_bases = is_alloc_tightly_coupled();
283 if (ary_src == nullptr || ary_src->elem() == Type::BOTTOM ||
284 ary_dest == nullptr || ary_dest->elem() == Type::BOTTOM) {
285 // We don't know if arguments are arrays
286 return false;
287 }
288
289 BasicType src_elem = ary_src->elem()->array_element_basic_type();
290 BasicType dest_elem = ary_dest->elem()->array_element_basic_type();
291 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
292 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
293
294 if (src_elem != dest_elem || dest_elem == T_VOID) {
295 // We don't know if arguments are arrays of the same type
296 return false;
297 }
298
299 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
300 if (bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), dest_elem, false, false, BarrierSetC2::Optimization)) {
301 // It's an object array copy but we can't emit the card marking
302 // that is needed
303 return false;
304 }
305
306 value_type = ary_src->elem();
307
308 uint shift = exact_log2(type2aelembytes(dest_elem));
309 uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
310
311 src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
312 if (src_offset->is_top()) {
313 // Offset is out of bounds (the ArrayCopyNode will be removed)
314 return false;
315 }
316 dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
317 if (dest_offset->is_top()) {
318 // Offset is out of bounds (the ArrayCopyNode will be removed)
319 if (can_reshape) {
320 // record src_offset, so it can be deleted later (if it is dead)
321 phase->is_IterGVN()->_worklist.push(src_offset);
322 }
323 return false;
324 }
325
326 Node* hook = new Node(1);
327 hook->init_req(0, dest_offset);
328
329 Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
330
331 hook->destruct(phase);
332
333 Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
334
335 adr_src = phase->transform(AddPNode::make_with_base(base_src, src_scale));
336 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, dest_scale));
337
338 adr_src = phase->transform(AddPNode::make_with_base(base_src, adr_src, phase->MakeConX(header)));
339 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, adr_dest, phase->MakeConX(header)));
340
341 copy_type = dest_elem;
342 } else {
343 assert(ary_src != nullptr, "should be a clone");
344 assert(is_clonebasic(), "should be");
345
346 disjoint_bases = true;
347
348 BasicType elem = ary_src->isa_aryptr()->elem()->array_element_basic_type();
349 if (is_reference_type(elem, true)) {
350 elem = T_OBJECT;
351 }
352
353 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
354 if (bs->array_copy_requires_gc_barriers(true, elem, true, is_clone_inst(), BarrierSetC2::Optimization)) {
355 return false;
356 }
357
358 adr_src = phase->transform(AddPNode::make_with_base(base_src, src_offset));
359 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, dest_offset));
360
361 // The address is offsetted to an aligned address where a raw copy would start.
362 // If the clone copy is decomposed into load-stores - the address is adjusted to
363 // point at where the array starts.
364 const Type* toff = phase->type(src_offset);
365 int offset = toff->isa_long() ? (int) toff->is_long()->get_con() : (int) toff->is_int()->get_con();
366 int diff = arrayOopDesc::base_offset_in_bytes(elem) - offset;
367 assert(diff >= 0, "clone should not start after 1st array element");
368 if (diff > 0) {
369 adr_src = phase->transform(AddPNode::make_with_base(base_src, adr_src, phase->MakeConX(diff)));
370 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, adr_dest, phase->MakeConX(diff)));
371 }
372 copy_type = elem;
373 value_type = ary_src->elem();
374 }
375 return true;
376 }
377
378 const TypePtr* ArrayCopyNode::get_address_type(PhaseGVN* phase, const TypePtr* atp, Node* n) {
379 if (atp == TypeOopPtr::BOTTOM) {
380 atp = phase->type(n)->isa_ptr();
381 }
382 // adjust atp to be the correct array element address type
383 return atp->add_offset(Type::OffsetBot);
384 }
385
386 void ArrayCopyNode::array_copy_test_overlap(PhaseGVN *phase, bool can_reshape, bool disjoint_bases, int count, Node*& forward_ctl, Node*& backward_ctl) {
387 Node* ctl = in(TypeFunc::Control);
388 if (!disjoint_bases && count > 1) {
389 Node* src_offset = in(ArrayCopyNode::SrcPos);
390 Node* dest_offset = in(ArrayCopyNode::DestPos);
391 assert(src_offset != nullptr && dest_offset != nullptr, "should be");
392 Node* cmp = phase->transform(new CmpINode(src_offset, dest_offset));
393 Node *bol = phase->transform(new BoolNode(cmp, BoolTest::lt));
394 IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
395
396 phase->transform(iff);
397
398 forward_ctl = phase->transform(new IfFalseNode(iff));
399 backward_ctl = phase->transform(new IfTrueNode(iff));
400 } else {
401 forward_ctl = ctl;
402 }
403 }
404
405 Node* ArrayCopyNode::array_copy_forward(PhaseGVN *phase,
406 bool can_reshape,
407 Node*& forward_ctl,
408 Node* mem,
409 const TypePtr* atp_src,
410 const TypePtr* atp_dest,
411 Node* adr_src,
412 Node* base_src,
413 Node* adr_dest,
414 Node* base_dest,
415 BasicType copy_type,
416 const Type* value_type,
417 int count) {
418 if (!forward_ctl->is_top()) {
419 // copy forward
420 MergeMemNode* mm = MergeMemNode::make(mem);
421
422 if (count > 0) {
423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
424 Node* v = load(bs, phase, forward_ctl, mm, adr_src, atp_src, value_type, copy_type);
425 store(bs, phase, forward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
426 for (int i = 1; i < count; i++) {
427 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
428 Node* next_src = phase->transform(AddPNode::make_with_base(base_src,adr_src,off));
429 // We may have narrowed the type of next_src right before calling this method but because this runs with
430 // PhaseIterGVN::_delay_transform true, explicitly update the type of the AddP so it's consistent with its
431 // base and load() picks the right memory slice.
432 phase->set_type(next_src, next_src->Value(phase));
433 Node* next_dest = phase->transform(AddPNode::make_with_base(base_dest,adr_dest,off));
434 // Same as above
435 phase->set_type(next_dest, next_dest->Value(phase));
436 v = load(bs, phase, forward_ctl, mm, next_src, atp_src, value_type, copy_type);
437 store(bs, phase, forward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
438 }
439 } else if (can_reshape) {
440 PhaseIterGVN* igvn = phase->is_IterGVN();
441 igvn->_worklist.push(adr_src);
442 igvn->_worklist.push(adr_dest);
443 }
444 return mm;
445 }
446 return phase->C->top();
447 }
448
449 Node* ArrayCopyNode::array_copy_backward(PhaseGVN *phase,
450 bool can_reshape,
451 Node*& backward_ctl,
452 Node* mem,
453 const TypePtr* atp_src,
454 const TypePtr* atp_dest,
455 Node* adr_src,
456 Node* base_src,
457 Node* adr_dest,
458 Node* base_dest,
459 BasicType copy_type,
460 const Type* value_type,
461 int count) {
462 if (!backward_ctl->is_top()) {
463 // copy backward
464 MergeMemNode* mm = MergeMemNode::make(mem);
465
466 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
467 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");
468
469 if (count > 0) {
470 for (int i = count-1; i >= 1; i--) {
471 Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
472 Node* next_src = phase->transform(AddPNode::make_with_base(base_src,adr_src,off));
473 // We may have narrowed the type of next_src right before calling this method but because this runs with
474 // PhaseIterGVN::_delay_transform true, explicitly update the type of the AddP so it's consistent with its
475 // base and store() picks the right memory slice.
476 phase->set_type(next_src, next_src->Value(phase));
477 Node* next_dest = phase->transform(AddPNode::make_with_base(base_dest,adr_dest,off));
478 phase->set_type(next_dest, next_dest->Value(phase));
479 Node* v = load(bs, phase, backward_ctl, mm, next_src, atp_src, value_type, copy_type);
480 store(bs, phase, backward_ctl, mm, next_dest, atp_dest, v, value_type, copy_type);
481 }
482 Node* v = load(bs, phase, backward_ctl, mm, adr_src, atp_src, value_type, copy_type);
483 store(bs, phase, backward_ctl, mm, adr_dest, atp_dest, v, value_type, copy_type);
484 } else if (can_reshape) {
485 PhaseIterGVN* igvn = phase->is_IterGVN();
486 igvn->_worklist.push(adr_src);
487 igvn->_worklist.push(adr_dest);
488 }
489 return phase->transform(mm);
490 }
491 return phase->C->top();
492 }
493
494 bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
495 Node* ctl, Node *mem) {
496 if (can_reshape) {
497 PhaseIterGVN* igvn = phase->is_IterGVN();
498 igvn->set_delay_transform(false);
499 if (is_clonebasic()) {
500 Node* out_mem = proj_out(TypeFunc::Memory);
501
502 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
503 if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
504 out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
505 assert(bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, is_clone_inst(), BarrierSetC2::Optimization), "can only happen with card marking");
506 return false;
507 }
508
509 igvn->replace_node(out_mem->raw_out(0), mem);
510
511 Node* out_ctl = proj_out(TypeFunc::Control);
512 igvn->replace_node(out_ctl, ctl);
513 } else {
514 // replace fallthrough projections of the ArrayCopyNode by the
515 // new memory, control and the input IO.
516 CallProjections callprojs;
517 extract_projections(&callprojs, true, false);
518
519 if (callprojs.fallthrough_ioproj != nullptr) {
520 igvn->replace_node(callprojs.fallthrough_ioproj, in(TypeFunc::I_O));
521 }
522 if (callprojs.fallthrough_memproj != nullptr) {
523 igvn->replace_node(callprojs.fallthrough_memproj, mem);
524 }
525 if (callprojs.fallthrough_catchproj != nullptr) {
526 igvn->replace_node(callprojs.fallthrough_catchproj, ctl);
527 }
528
529 // The ArrayCopyNode is not disconnected. It still has the
530 // projections for the exception case. Replace current
531 // ArrayCopyNode with a dummy new one with a top() control so
532 // that this part of the graph stays consistent but is
533 // eventually removed.
534
535 set_req(0, phase->C->top());
536 remove_dead_region(phase, can_reshape);
537 }
538 } else {
539 if (in(TypeFunc::Control) != ctl) {
540 // we can't return new memory and control from Ideal at parse time
541 assert(!is_clonebasic() || UseShenandoahGC, "added control for clone?");
542 phase->record_for_igvn(this);
543 return false;
544 }
545 }
546 return true;
547 }
548
549
550 Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
551 if (remove_dead_region(phase, can_reshape)) return this;
552
553 if (StressArrayCopyMacroNode && !can_reshape) {
554 phase->record_for_igvn(this);
555 return nullptr;
556 }
557
558 // See if it's a small array copy and we can inline it as
559 // loads/stores
560 // Here we can only do:
561 // - arraycopy if all arguments were validated before and we don't
562 // need card marking
563 // - clone for which we don't need to do card marking
564
565 if (!is_clonebasic() && !is_arraycopy_validated() &&
566 !is_copyofrange_validated() && !is_copyof_validated()) {
567 return nullptr;
568 }
569
570 assert(in(TypeFunc::Control) != nullptr &&
571 in(TypeFunc::Memory) != nullptr &&
573 in(ArrayCopyNode::Dest) != nullptr &&
574 in(ArrayCopyNode::Length) != nullptr &&
575 in(ArrayCopyNode::SrcPos) != nullptr &&
576 in(ArrayCopyNode::DestPos) != nullptr, "broken inputs");
577
578 if (in(TypeFunc::Control)->is_top() ||
579 in(TypeFunc::Memory)->is_top() ||
580 phase->type(in(ArrayCopyNode::Src)) == Type::TOP ||
581 phase->type(in(ArrayCopyNode::Dest)) == Type::TOP ||
582 (in(ArrayCopyNode::SrcPos) != nullptr && in(ArrayCopyNode::SrcPos)->is_top()) ||
583 (in(ArrayCopyNode::DestPos) != nullptr && in(ArrayCopyNode::DestPos)->is_top())) {
584 return nullptr;
585 }
586
587 int count = get_count(phase);
588
589 if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
590 return nullptr;
591 }
592
593 Node* mem = try_clone_instance(phase, can_reshape, count);
594 if (mem != nullptr) {
595 return (mem == NodeSentinel) ? nullptr : mem;
596 }
597
598 Node* adr_src = nullptr;
599 Node* base_src = nullptr;
600 Node* adr_dest = nullptr;
601 Node* base_dest = nullptr;
602 BasicType copy_type = T_ILLEGAL;
603 const Type* value_type = nullptr;
604 bool disjoint_bases = false;
605
606 Node* src = in(ArrayCopyNode::Src);
607 Node* dest = in(ArrayCopyNode::Dest);
608 // EA may have moved an input to a new slice. EA stores the new address types in the ArrayCopy node itself
609 // (_src_type/_dest_type). phase->type(src) and _src_type or phase->type(dest) and _dest_type may be different
610 // when this transformation runs if igvn hasn't had a chance to propagate the new types yet. Make sure the new
611 // types are taken into account so new Load/Store nodes are created on the right slice.
612 const TypePtr* atp_src = get_address_type(phase, _src_type, src);
613 const TypePtr* atp_dest = get_address_type(phase, _dest_type, dest);
614 phase->set_type(src, phase->type(src)->join_speculative(atp_src));
615 phase->set_type(dest, phase->type(dest)->join_speculative(atp_dest));
616
617 if (!prepare_array_copy(phase, can_reshape,
618 adr_src, base_src, adr_dest, base_dest,
619 copy_type, value_type, disjoint_bases)) {
620 assert(adr_src == nullptr, "no node can be left behind");
621 assert(adr_dest == nullptr, "no node can be left behind");
622 return nullptr;
623 }
624
625 Node* in_mem = in(TypeFunc::Memory);
626
627 if (can_reshape) {
628 assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
629 phase->is_IterGVN()->set_delay_transform(true);
630 }
631
632 Node* backward_ctl = phase->C->top();
633 Node* forward_ctl = phase->C->top();
634 array_copy_test_overlap(phase, can_reshape, disjoint_bases, count, forward_ctl, backward_ctl);
635
636 Node* forward_mem = array_copy_forward(phase, can_reshape, forward_ctl,
637 in_mem,
638 atp_src, atp_dest,
639 adr_src, base_src, adr_dest, base_dest,
640 copy_type, value_type, count);
641
642 Node* backward_mem = array_copy_backward(phase, can_reshape, backward_ctl,
643 in_mem,
644 atp_src, atp_dest,
645 adr_src, base_src, adr_dest, base_dest,
646 copy_type, value_type, count);
647
648 Node* ctl = nullptr;
649 if (!forward_ctl->is_top() && !backward_ctl->is_top()) {
650 ctl = new RegionNode(3);
651 ctl->init_req(1, forward_ctl);
652 ctl->init_req(2, backward_ctl);
653 ctl = phase->transform(ctl);
654 MergeMemNode* forward_mm = forward_mem->as_MergeMem();
655 MergeMemNode* backward_mm = backward_mem->as_MergeMem();
656 for (MergeMemStream mms(forward_mm, backward_mm); mms.next_non_empty2(); ) {
657 if (mms.memory() != mms.memory2()) {
658 Node* phi = new PhiNode(ctl, Type::MEMORY, phase->C->get_adr_type(mms.alias_idx()));
659 phi->init_req(1, mms.memory());
660 phi->init_req(2, mms.memory2());
661 phi = phase->transform(phi);
662 mms.set_memory(phi);
663 }
664 }
665 mem = forward_mem;
666 } else if (!forward_ctl->is_top()) {
667 ctl = forward_ctl;
668 mem = forward_mem;
669 } else {
670 assert(!backward_ctl->is_top(), "no copy?");
671 ctl = backward_ctl;
672 mem = backward_mem;
673 }
674
675 if (can_reshape) {
676 assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
677 phase->is_IterGVN()->set_delay_transform(false);
678 }
679
680 if (!finish_transform(phase, can_reshape, ctl, mem)) {
681 if (can_reshape) {
682 // put in worklist, so that if it happens to be dead it is removed
683 phase->is_IterGVN()->_worklist.push(mem);
684 }
685 return nullptr;
686 }
687
688 return mem;
689 }
690
691 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
692 Node* dest = in(ArrayCopyNode::Dest);
693 if (dest->is_top()) {
694 return false;
695 }
696 const TypeOopPtr* dest_t = phase->type(dest)->is_oopptr();
697 assert(!dest_t->is_known_instance() || _dest_type->is_known_instance(), "result of EA not recorded");
698 assert(in(ArrayCopyNode::Src)->is_top() || !phase->type(in(ArrayCopyNode::Src))->is_oopptr()->is_known_instance() ||
699 _src_type->is_known_instance(), "result of EA not recorded");
700
701 if (_dest_type != TypeOopPtr::BOTTOM || t_oop->is_known_instance()) {
758 // if must_modify is true, return true if the copy is guaranteed to
759 // write between offset_lo and offset_hi
760 bool ArrayCopyNode::modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseValues* phase, bool must_modify) const {
761 assert(_kind == ArrayCopy || _kind == CopyOf || _kind == CopyOfRange, "only for real array copies");
762
763 Node* dest = in(Dest);
764 Node* dest_pos = in(DestPos);
765 Node* len = in(Length);
766
767 const TypeInt *dest_pos_t = phase->type(dest_pos)->isa_int();
768 const TypeInt *len_t = phase->type(len)->isa_int();
769 const TypeAryPtr* ary_t = phase->type(dest)->isa_aryptr();
770
771 if (dest_pos_t == nullptr || len_t == nullptr || ary_t == nullptr) {
772 return !must_modify;
773 }
774
775 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
776 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
777
778 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
779 uint elemsize = type2aelembytes(ary_elem);
780
781 jlong dest_pos_plus_len_lo = (((jlong)dest_pos_t->_lo) + len_t->_lo) * elemsize + header;
782 jlong dest_pos_plus_len_hi = (((jlong)dest_pos_t->_hi) + len_t->_hi) * elemsize + header;
783 jlong dest_pos_lo = ((jlong)dest_pos_t->_lo) * elemsize + header;
784 jlong dest_pos_hi = ((jlong)dest_pos_t->_hi) * elemsize + header;
785
786 if (must_modify) {
787 if (offset_lo >= dest_pos_hi && offset_hi < dest_pos_plus_len_lo) {
788 return true;
789 }
790 } else {
791 if (offset_hi >= dest_pos_lo && offset_lo < dest_pos_plus_len_hi) {
792 return true;
793 }
794 }
795 return false;
796 }
797
798 // As an optimization, choose the optimal vector size for bounded copy length
799 int ArrayCopyNode::get_partial_inline_vector_lane_count(BasicType type, jlong max_len) {
800 assert(max_len > 0, JLONG_FORMAT, max_len);
801 // We only care whether max_size_in_bytes is not larger than 32, we also want to avoid
802 // multiplication overflow, so clamp max_len to [0, 64]
803 int max_size_in_bytes = MIN2<jlong>(max_len, 64) * type2aelembytes(type);
804 if (ArrayOperationPartialInlineSize > 16 && max_size_in_bytes <= 16) {
|
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.hpp"
26 #include "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 "opto/inlinetypenode.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,
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 if (is_clonebasic()) {
119 Node* src = in(ArrayCopyNode::Src);
120 const Type* src_type = phase->type(src);
121
122 if (src_type == Type::TOP) {
123 return -1;
124 }
125
126 if (src_type->isa_instptr()) {
127 const TypeInstPtr* inst_src = src_type->is_instptr();
128 ciInstanceKlass* ik = inst_src->instance_klass();
129 // ciInstanceKlass::nof_nonstatic_fields() doesn't take injected
130 // fields into account. They are rare anyway so easier to simply
131 // skip instances with injected fields.
132 if ((!inst_src->klass_is_exact() && (ik->is_interface() || ik->has_subklass())) || ik->has_injected_fields()) {
133 return -1;
134 }
135 int nb_fields = ik->nof_nonstatic_fields();
136 return nb_fields;
137 } else {
138 const TypeAryPtr* ary_src = src_type->isa_aryptr();
139 assert (ary_src != nullptr, "not an array or instance?");
140 // clone passes a length as a rounded number of longs. If we're
141 // cloning an array we'll do it element by element. If the
142 // length of the input array is constant, ArrayCopyNode::Length
143 // must be too. Note that the opposite does not need to hold,
144 // because different input array lengths (e.g. int arrays with
145 // 3 or 4 elements) might lead to the same length input
146 // (e.g. 2 double-words).
147 assert(!ary_src->size()->is_con() || (get_length_if_constant(phase) >= 0) ||
148 (UseArrayFlattening && ary_src->elem()->make_oopptr() != nullptr && ary_src->elem()->make_oopptr()->can_be_inline_type()) ||
149 phase->is_IterGVN() || phase->C->inlining_incrementally() || StressReflectiveCode, "inconsistent");
150 if (ary_src->size()->is_con()) {
151 return ary_src->size()->get_con();
152 }
153 return -1;
154 }
155 }
156
157 return get_length_if_constant(phase);
158 }
159
160 Node* ArrayCopyNode::load(BarrierSetC2* bs, PhaseGVN *phase, Node*& ctl, MergeMemNode* mem, Node* adr, const TypePtr* adr_type, const Type *type, BasicType bt) {
161 // 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
162 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk floating
163 // above runtime checks that guarantee it is within bounds.
164 DecoratorSet decorators = C2_READ_ACCESS | C2_CONTROL_DEPENDENT_LOAD | IN_HEAP | C2_ARRAY_COPY | C2_UNKNOWN_CONTROL_LOAD;
165 C2AccessValuePtr addr(adr, adr_type);
166 C2OptAccess access(*phase, ctl, mem, decorators, bt, adr->in(AddPNode::Base), addr);
167 Node* res = bs->load_at(access, type);
168 ctl = access.ctl();
182 }
183
184
185 Node* ArrayCopyNode::try_clone_instance(PhaseGVN *phase, bool can_reshape, int count) {
186 if (!is_clonebasic()) {
187 return nullptr;
188 }
189
190 Node* base_src = in(ArrayCopyNode::Src);
191 Node* base_dest = in(ArrayCopyNode::Dest);
192 Node* ctl = in(TypeFunc::Control);
193 Node* in_mem = in(TypeFunc::Memory);
194
195 const Type* src_type = phase->type(base_src);
196 const TypeInstPtr* inst_src = src_type->isa_instptr();
197 if (inst_src == nullptr) {
198 return nullptr;
199 }
200
201 MergeMemNode* mem = phase->transform(MergeMemNode::make(in_mem))->as_MergeMem();
202 phase->record_for_igvn(mem);
203 if (can_reshape) {
204 phase->is_IterGVN()->_worklist.push(mem);
205 }
206
207
208 ciInstanceKlass* ik = inst_src->instance_klass();
209
210 if (!inst_src->klass_is_exact()) {
211 assert(!ik->is_interface(), "inconsistent klass hierarchy");
212 if (ik->has_subklass()) {
213 // Concurrent class loading.
214 // Fail fast and return NodeSentinel to indicate that the transform failed.
215 return NodeSentinel;
216 } else {
217 phase->C->dependencies()->assert_leaf_type(ik);
218 }
219 }
220
221 const TypeInstPtr* dest_type = phase->type(base_dest)->is_instptr();
222 if (dest_type->instance_klass() != ik) {
282 Node* src_offset = in(ArrayCopyNode::SrcPos);
283 Node* dest_offset = in(ArrayCopyNode::DestPos);
284
285 if (is_arraycopy() || is_copyofrange() || is_copyof()) {
286 const Type* dest_type = phase->type(base_dest);
287 const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
288
289 // newly allocated object is guaranteed to not overlap with source object
290 disjoint_bases = is_alloc_tightly_coupled();
291 if (ary_src == nullptr || ary_src->elem() == Type::BOTTOM ||
292 ary_dest == nullptr || ary_dest->elem() == Type::BOTTOM) {
293 // We don't know if arguments are arrays
294 return false;
295 }
296
297 BasicType src_elem = ary_src->elem()->array_element_basic_type();
298 BasicType dest_elem = ary_dest->elem()->array_element_basic_type();
299 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
300 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
301
302 // TODO 8350865 What about atomicity?
303 if (src_elem != dest_elem || ary_src->is_null_free() != ary_dest->is_null_free() || ary_src->is_flat() != ary_dest->is_flat() || dest_elem == T_VOID) {
304 // We don't know if arguments are arrays of the same type
305 return false;
306 }
307
308 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
309 if ((!ary_dest->is_flat() && bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), dest_elem, false, false, BarrierSetC2::Optimization)) ||
310 (ary_dest->is_flat() && ary_src->elem()->inline_klass()->contains_oops() &&
311 bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), T_OBJECT, false, false, BarrierSetC2::Optimization))) {
312 // It's an object array copy but we can't emit the card marking that is needed
313 return false;
314 }
315
316 value_type = ary_src->elem();
317
318 uint shift = exact_log2(type2aelembytes(dest_elem));
319 if (ary_dest->is_flat()) {
320 assert(ary_src->is_flat(), "src and dest must be flat");
321 shift = ary_src->flat_log_elem_size();
322 src_elem = T_FLAT_ELEMENT;
323 dest_elem = T_FLAT_ELEMENT;
324 }
325
326 const uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
327
328 src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
329 if (src_offset->is_top()) {
330 // Offset is out of bounds (the ArrayCopyNode will be removed)
331 return false;
332 }
333 dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
334 if (dest_offset->is_top()) {
335 // Offset is out of bounds (the ArrayCopyNode will be removed)
336 if (can_reshape) {
337 // record src_offset, so it can be deleted later (if it is dead)
338 phase->is_IterGVN()->_worklist.push(src_offset);
339 }
340 return false;
341 }
342
343 Node* hook = new Node(1);
344 hook->init_req(0, dest_offset);
345
346 Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
347
348 hook->destruct(phase);
349
350 Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
351
352 adr_src = phase->transform(AddPNode::make_with_base(base_src, src_scale));
353 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, dest_scale));
354
355 adr_src = phase->transform(AddPNode::make_with_base(base_src, adr_src, phase->MakeConX(header)));
356 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, adr_dest, phase->MakeConX(header)));
357
358 copy_type = dest_elem;
359 } else {
360 assert(ary_src != nullptr, "should be a clone");
361 assert(is_clonebasic(), "should be");
362
363 disjoint_bases = true;
364
365 if (ary_src->elem()->make_oopptr() != nullptr &&
366 ary_src->elem()->make_oopptr()->can_be_inline_type()) {
367 return false;
368 }
369
370 BasicType elem = ary_src->isa_aryptr()->elem()->array_element_basic_type();
371 if (is_reference_type(elem, true)) {
372 elem = T_OBJECT;
373 }
374
375 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
376 if ((!ary_src->is_flat() && bs->array_copy_requires_gc_barriers(true, elem, true, is_clone_inst(), BarrierSetC2::Optimization)) ||
377 (ary_src->is_flat() && ary_src->elem()->inline_klass()->contains_oops() &&
378 bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, is_clone_inst(), BarrierSetC2::Optimization))) {
379 // It's an object array copy but we can't emit the card marking that is needed
380 return false;
381 }
382
383 adr_src = phase->transform(AddPNode::make_with_base(base_src, src_offset));
384 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, dest_offset));
385
386 // The address is offsetted to an aligned address where a raw copy would start.
387 // If the clone copy is decomposed into load-stores - the address is adjusted to
388 // point at where the array starts.
389 const Type* toff = phase->type(src_offset);
390 int offset = toff->isa_long() ? (int) toff->is_long()->get_con() : (int) toff->is_int()->get_con();
391 int diff = arrayOopDesc::base_offset_in_bytes(elem) - offset;
392 assert(diff >= 0, "clone should not start after 1st array element");
393 if (diff > 0) {
394 adr_src = phase->transform(AddPNode::make_with_base(base_src, adr_src, phase->MakeConX(diff)));
395 adr_dest = phase->transform(AddPNode::make_with_base(base_dest, adr_dest, phase->MakeConX(diff)));
396 }
397 copy_type = elem;
398 value_type = ary_src->elem();
399 }
400 return true;
401 }
402
403 const TypeAryPtr* ArrayCopyNode::get_address_type(PhaseGVN* phase, const TypePtr* atp, Node* n) {
404 if (atp == TypeOopPtr::BOTTOM) {
405 atp = phase->type(n)->isa_ptr();
406 }
407 // adjust atp to be the correct array element address type
408 return atp->add_offset(Type::OffsetBot)->is_aryptr();
409 }
410
411 void ArrayCopyNode::array_copy_test_overlap(GraphKit& kit, bool disjoint_bases, int count, Node*& backward_ctl) {
412 Node* ctl = kit.control();
413 if (!disjoint_bases && count > 1) {
414 PhaseGVN& gvn = kit.gvn();
415 Node* src_offset = in(ArrayCopyNode::SrcPos);
416 Node* dest_offset = in(ArrayCopyNode::DestPos);
417 assert(src_offset != nullptr && dest_offset != nullptr, "should be");
418 Node* cmp = gvn.transform(new CmpINode(src_offset, dest_offset));
419 Node *bol = gvn.transform(new BoolNode(cmp, BoolTest::lt));
420 IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
421
422 gvn.transform(iff);
423
424 kit.set_control(gvn.transform(new IfFalseNode(iff)));
425 backward_ctl = gvn.transform(new IfTrueNode(iff));
426 }
427 }
428
429 void ArrayCopyNode::copy(GraphKit& kit,
430 const TypeAryPtr* atp_src,
431 const TypeAryPtr* atp_dest,
432 int i,
433 Node* base_src,
434 Node* base_dest,
435 Node* adr_src,
436 Node* adr_dest,
437 BasicType copy_type,
438 const Type* value_type) {
439 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
440 Node* ctl = kit.control();
441 PhaseGVN& phase = kit.gvn();
442 if (atp_dest->is_flat()) {
443 ciInlineKlass* vk = atp_src->elem()->inline_klass();
444 for (int j = 0; j < vk->nof_nonstatic_fields(); j++) {
445 ciField* field = vk->nonstatic_field_at(j);
446 int off_in_vt = field->offset_in_bytes() - vk->payload_offset();
447 Node* off = kit.MakeConX(off_in_vt + i * atp_src->flat_elem_size());
448 ciType* ft = field->type();
449 BasicType bt = type2field[ft->basic_type()];
450 assert(!field->is_flat(), "flat field encountered");
451 const Type* rt = Type::get_const_type(ft);
452 const TypePtr* adr_type = atp_src->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
453 assert(!bs->array_copy_requires_gc_barriers(is_alloc_tightly_coupled(), bt, false, false, BarrierSetC2::Optimization), "GC barriers required");
454 Node* next_src = kit.gvn().transform(AddPNode::make_with_base(base_src, adr_src, off));
455 // We may have narrowed the type of next_src right before calling this method but because this runs with
456 // PhaseIterGVN::_delay_transform true, explicitly update the type of the AddP so it's consistent with its
457 // base and load() picks the right memory slice.
458 phase.set_type(next_src, next_src->Value(&phase));
459 Node* next_dest = kit.gvn().transform(AddPNode::make_with_base(base_dest, adr_dest, off));
460 // Same as above
461 phase.set_type(next_src, next_src->Value(&phase));
462 Node* v = load(bs, &kit.gvn(), ctl, kit.merged_memory(), next_src, adr_type, rt, bt);
463 store(bs, &kit.gvn(), ctl, kit.merged_memory(), next_dest, adr_type, v, rt, bt);
464 }
465 } else {
466 Node* off = kit.MakeConX(type2aelembytes(copy_type) * i);
467 Node* next_src = kit.gvn().transform(AddPNode::make_with_base(base_src, adr_src, off));
468 // We may have narrowed the type of next_src right before calling this method but because this runs with
469 // PhaseIterGVN::_delay_transform true, explicitly update the type of the AddP so it's consistent with its
470 // base and load() picks the right memory slice.
471 phase.set_type(next_src, next_src->Value(&phase));
472 Node* next_dest = kit.gvn().transform(AddPNode::make_with_base(base_dest, adr_dest, off));
473 phase.set_type(next_dest, next_dest->Value(&phase));
474 Node* v = load(bs, &kit.gvn(), ctl, kit.merged_memory(), next_src, atp_src, value_type, copy_type);
475 store(bs, &kit.gvn(), ctl, kit.merged_memory(), next_dest, atp_dest, v, value_type, copy_type);
476 }
477 kit.set_control(ctl);
478 }
479
480
481 void ArrayCopyNode::array_copy_forward(GraphKit& kit,
482 bool can_reshape,
483 const TypeAryPtr* atp_src,
484 const TypeAryPtr* atp_dest,
485 Node* adr_src,
486 Node* base_src,
487 Node* adr_dest,
488 Node* base_dest,
489 BasicType copy_type,
490 const Type* value_type,
491 int count) {
492 if (!kit.stopped()) {
493 // copy forward
494 if (count > 0) {
495 for (int i = 0; i < count; i++) {
496 copy(kit, atp_src, atp_dest, i, base_src, base_dest, adr_src, adr_dest, copy_type, value_type);
497 }
498 } else if (can_reshape) {
499 PhaseGVN& gvn = kit.gvn();
500 assert(gvn.is_IterGVN(), "");
501 gvn.record_for_igvn(adr_src);
502 gvn.record_for_igvn(adr_dest);
503 }
504 }
505 }
506
507 void ArrayCopyNode::array_copy_backward(GraphKit& kit,
508 bool can_reshape,
509 const TypeAryPtr* atp_src,
510 const TypeAryPtr* atp_dest,
511 Node* adr_src,
512 Node* base_src,
513 Node* adr_dest,
514 Node* base_dest,
515 BasicType copy_type,
516 const Type* value_type,
517 int count) {
518 if (!kit.stopped()) {
519 // copy backward
520 PhaseGVN& gvn = kit.gvn();
521
522 if (count > 0) {
523 for (int i = count-1; i >= 0; i--) {
524 copy(kit, atp_src, atp_dest, i, base_src, base_dest, adr_src, adr_dest, copy_type, value_type);
525 }
526 } else if(can_reshape) {
527 PhaseGVN& gvn = kit.gvn();
528 assert(gvn.is_IterGVN(), "");
529 gvn.record_for_igvn(adr_src);
530 gvn.record_for_igvn(adr_dest);
531 }
532 }
533 }
534
535 bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
536 Node* ctl, Node *mem) {
537 if (can_reshape) {
538 PhaseIterGVN* igvn = phase->is_IterGVN();
539 igvn->set_delay_transform(false);
540 if (is_clonebasic()) {
541 Node* out_mem = proj_out(TypeFunc::Memory);
542
543 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
544 if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
545 out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
546 assert(bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, is_clone_inst(), BarrierSetC2::Optimization), "can only happen with card marking");
547 return false;
548 }
549
550 igvn->replace_node(out_mem->raw_out(0), mem);
551
552 Node* out_ctl = proj_out(TypeFunc::Control);
553 igvn->replace_node(out_ctl, ctl);
554 } else {
555 // replace fallthrough projections of the ArrayCopyNode by the
556 // new memory, control and the input IO.
557 CallProjections* callprojs = extract_projections(true, false);
558
559 if (callprojs->fallthrough_ioproj != nullptr) {
560 igvn->replace_node(callprojs->fallthrough_ioproj, in(TypeFunc::I_O));
561 }
562 if (callprojs->fallthrough_memproj != nullptr) {
563 igvn->replace_node(callprojs->fallthrough_memproj, mem);
564 }
565 if (callprojs->fallthrough_catchproj != nullptr) {
566 igvn->replace_node(callprojs->fallthrough_catchproj, ctl);
567 }
568
569 // The ArrayCopyNode is not disconnected. It still has the
570 // projections for the exception case. Replace current
571 // ArrayCopyNode with a dummy new one with a top() control so
572 // that this part of the graph stays consistent but is
573 // eventually removed.
574
575 set_req(0, phase->C->top());
576 remove_dead_region(phase, can_reshape);
577 }
578 } else {
579 if (in(TypeFunc::Control) != ctl) {
580 // we can't return new memory and control from Ideal at parse time
581 assert(!is_clonebasic() || UseShenandoahGC, "added control for clone?");
582 phase->record_for_igvn(this);
583 return false;
584 }
585 }
586 return true;
587 }
588
589
590 Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
591 // Perform any generic optimizations first
592 Node* result = SafePointNode::Ideal(phase, can_reshape);
593 if (result != nullptr) {
594 return result;
595 }
596
597 if (StressArrayCopyMacroNode && !can_reshape) {
598 phase->record_for_igvn(this);
599 return nullptr;
600 }
601
602 // See if it's a small array copy and we can inline it as
603 // loads/stores
604 // Here we can only do:
605 // - arraycopy if all arguments were validated before and we don't
606 // need card marking
607 // - clone for which we don't need to do card marking
608
609 if (!is_clonebasic() && !is_arraycopy_validated() &&
610 !is_copyofrange_validated() && !is_copyof_validated()) {
611 return nullptr;
612 }
613
614 assert(in(TypeFunc::Control) != nullptr &&
615 in(TypeFunc::Memory) != nullptr &&
617 in(ArrayCopyNode::Dest) != nullptr &&
618 in(ArrayCopyNode::Length) != nullptr &&
619 in(ArrayCopyNode::SrcPos) != nullptr &&
620 in(ArrayCopyNode::DestPos) != nullptr, "broken inputs");
621
622 if (in(TypeFunc::Control)->is_top() ||
623 in(TypeFunc::Memory)->is_top() ||
624 phase->type(in(ArrayCopyNode::Src)) == Type::TOP ||
625 phase->type(in(ArrayCopyNode::Dest)) == Type::TOP ||
626 (in(ArrayCopyNode::SrcPos) != nullptr && in(ArrayCopyNode::SrcPos)->is_top()) ||
627 (in(ArrayCopyNode::DestPos) != nullptr && in(ArrayCopyNode::DestPos)->is_top())) {
628 return nullptr;
629 }
630
631 int count = get_count(phase);
632
633 if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
634 return nullptr;
635 }
636
637 Node* src = in(ArrayCopyNode::Src);
638 Node* dest = in(ArrayCopyNode::Dest);
639 const Type* src_type = phase->type(src);
640 const Type* dest_type = phase->type(dest);
641
642 if (src_type->isa_aryptr() && dest_type->isa_instptr()) {
643 // clone used for load of unknown inline type can't be optimized at
644 // this point
645 return nullptr;
646 }
647
648 Node* mem = try_clone_instance(phase, can_reshape, count);
649 if (mem != nullptr) {
650 return (mem == NodeSentinel) ? nullptr : mem;
651 }
652
653 Node* adr_src = nullptr;
654 Node* base_src = nullptr;
655 Node* adr_dest = nullptr;
656 Node* base_dest = nullptr;
657 BasicType copy_type = T_ILLEGAL;
658 const Type* value_type = nullptr;
659 bool disjoint_bases = false;
660
661 // EA may have moved an input to a new slice. EA stores the new address types in the ArrayCopy node itself
662 // (_src_type/_dest_type). phase->type(src) and _src_type or phase->type(dest) and _dest_type may be different
663 // when this transformation runs if igvn hasn't had a chance to propagate the new types yet. Make sure the new
664 // types are taken into account so new Load/Store nodes are created on the right slice.
665 const TypeAryPtr* atp_src = get_address_type(phase, _src_type, src);
666 const TypeAryPtr* atp_dest = get_address_type(phase, _dest_type, dest);
667 phase->set_type(src, phase->type(src)->join_speculative(atp_src));
668 phase->set_type(dest, phase->type(dest)->join_speculative(atp_dest));
669
670 if (!prepare_array_copy(phase, can_reshape,
671 adr_src, base_src, adr_dest, base_dest,
672 copy_type, value_type, disjoint_bases)) {
673 assert(adr_src == nullptr, "no node can be left behind");
674 assert(adr_dest == nullptr, "no node can be left behind");
675 return nullptr;
676 }
677
678 JVMState* new_jvms = nullptr;
679 SafePointNode* new_map = nullptr;
680 if (!is_clonebasic()) {
681 new_jvms = jvms()->clone_shallow(phase->C);
682 new_map = new SafePointNode(req(), new_jvms);
683 for (uint i = TypeFunc::FramePtr; i < req(); i++) {
684 new_map->init_req(i, in(i));
685 }
686 new_jvms->set_map(new_map);
687 } else {
688 new_jvms = new (phase->C) JVMState(0);
689 new_map = new SafePointNode(TypeFunc::Parms, new_jvms);
690 new_jvms->set_map(new_map);
691 }
692 new_map->set_control(in(TypeFunc::Control));
693 new_map->set_memory(MergeMemNode::make(in(TypeFunc::Memory)));
694 new_map->set_i_o(in(TypeFunc::I_O));
695 phase->record_for_igvn(new_map);
696
697 if (can_reshape) {
698 assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
699 phase->is_IterGVN()->set_delay_transform(true);
700 }
701
702 GraphKit kit(new_jvms, phase);
703
704 SafePointNode* backward_map = nullptr;
705 SafePointNode* forward_map = nullptr;
706 Node* backward_ctl = phase->C->top();
707
708 array_copy_test_overlap(kit, disjoint_bases, count, backward_ctl);
709
710 {
711 PreserveJVMState pjvms(&kit);
712
713 array_copy_forward(kit, can_reshape,
714 atp_src, atp_dest,
715 adr_src, base_src, adr_dest, base_dest,
716 copy_type, value_type, count);
717
718 forward_map = kit.stop();
719 }
720
721 kit.set_control(backward_ctl);
722 array_copy_backward(kit, can_reshape,
723 atp_src, atp_dest,
724 adr_src, base_src, adr_dest, base_dest,
725 copy_type, value_type, count);
726
727 backward_map = kit.stop();
728
729 if (!forward_map->control()->is_top() && !backward_map->control()->is_top()) {
730 assert(forward_map->i_o() == backward_map->i_o(), "need a phi on IO?");
731 Node* ctl = new RegionNode(3);
732 Node* mem = new PhiNode(ctl, Type::MEMORY, TypePtr::BOTTOM);
733 kit.set_map(forward_map);
734 ctl->init_req(1, kit.control());
735 mem->init_req(1, kit.reset_memory());
736 kit.set_map(backward_map);
737 ctl->init_req(2, kit.control());
738 mem->init_req(2, kit.reset_memory());
739 kit.set_control(phase->transform(ctl));
740 kit.set_all_memory(phase->transform(mem));
741 } else if (!forward_map->control()->is_top()) {
742 kit.set_map(forward_map);
743 } else {
744 assert(!backward_map->control()->is_top(), "no copy?");
745 kit.set_map(backward_map);
746 }
747
748 if (can_reshape) {
749 assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
750 phase->is_IterGVN()->set_delay_transform(false);
751 }
752
753 mem = kit.map()->memory();
754 if (!finish_transform(phase, can_reshape, kit.control(), mem)) {
755 if (!can_reshape) {
756 phase->record_for_igvn(this);
757 } else {
758 // put in worklist, so that if it happens to be dead it is removed
759 phase->is_IterGVN()->_worklist.push(mem);
760 }
761 return nullptr;
762 }
763
764 return mem;
765 }
766
767 bool ArrayCopyNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
768 Node* dest = in(ArrayCopyNode::Dest);
769 if (dest->is_top()) {
770 return false;
771 }
772 const TypeOopPtr* dest_t = phase->type(dest)->is_oopptr();
773 assert(!dest_t->is_known_instance() || _dest_type->is_known_instance(), "result of EA not recorded");
774 assert(in(ArrayCopyNode::Src)->is_top() || !phase->type(in(ArrayCopyNode::Src))->is_oopptr()->is_known_instance() ||
775 _src_type->is_known_instance(), "result of EA not recorded");
776
777 if (_dest_type != TypeOopPtr::BOTTOM || t_oop->is_known_instance()) {
834 // if must_modify is true, return true if the copy is guaranteed to
835 // write between offset_lo and offset_hi
836 bool ArrayCopyNode::modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseValues* phase, bool must_modify) const {
837 assert(_kind == ArrayCopy || _kind == CopyOf || _kind == CopyOfRange, "only for real array copies");
838
839 Node* dest = in(Dest);
840 Node* dest_pos = in(DestPos);
841 Node* len = in(Length);
842
843 const TypeInt *dest_pos_t = phase->type(dest_pos)->isa_int();
844 const TypeInt *len_t = phase->type(len)->isa_int();
845 const TypeAryPtr* ary_t = phase->type(dest)->isa_aryptr();
846
847 if (dest_pos_t == nullptr || len_t == nullptr || ary_t == nullptr) {
848 return !must_modify;
849 }
850
851 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
852 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
853
854 uint header;
855 uint elem_size;
856 if (ary_t->is_flat()) {
857 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
858 elem_size = ary_t->flat_elem_size();
859 } else {
860 header = arrayOopDesc::base_offset_in_bytes(ary_elem);
861 elem_size = type2aelembytes(ary_elem);
862 }
863
864 jlong dest_pos_plus_len_lo = (((jlong)dest_pos_t->_lo) + len_t->_lo) * elem_size + header;
865 jlong dest_pos_plus_len_hi = (((jlong)dest_pos_t->_hi) + len_t->_hi) * elem_size + header;
866 jlong dest_pos_lo = ((jlong)dest_pos_t->_lo) * elem_size + header;
867 jlong dest_pos_hi = ((jlong)dest_pos_t->_hi) * elem_size + header;
868
869 if (must_modify) {
870 if (offset_lo >= dest_pos_hi && offset_hi < dest_pos_plus_len_lo) {
871 return true;
872 }
873 } else {
874 if (offset_hi >= dest_pos_lo && offset_lo < dest_pos_plus_len_hi) {
875 return true;
876 }
877 }
878 return false;
879 }
880
881 // As an optimization, choose the optimal vector size for bounded copy length
882 int ArrayCopyNode::get_partial_inline_vector_lane_count(BasicType type, jlong max_len) {
883 assert(max_len > 0, JLONG_FORMAT, max_len);
884 // We only care whether max_size_in_bytes is not larger than 32, we also want to avoid
885 // multiplication overflow, so clamp max_len to [0, 64]
886 int max_size_in_bytes = MIN2<jlong>(max_len, 64) * type2aelembytes(type);
887 if (ArrayOperationPartialInlineSize > 16 && max_size_in_bytes <= 16) {
|