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/tlab_globals.hpp"
27 #include "oops/objArrayKlass.hpp"
28 #include "opto/arraycopynode.hpp"
29 #include "opto/castnode.hpp"
30 #include "opto/convertnode.hpp"
31 #include "opto/graphKit.hpp"
32 #include "opto/macro.hpp"
33 #include "opto/runtime.hpp"
34 #include "opto/vectornode.hpp"
35 #include "runtime/stubRoutines.hpp"
36 #include "utilities/align.hpp"
37 #include "utilities/powerOfTwo.hpp"
38
39 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, int alias_idx, Node* precedent) {
40 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
41 mb->init_req(TypeFunc::Control, *ctrl);
42 mb->init_req(TypeFunc::Memory, *mem);
43 transform_later(mb);
44 *ctrl = new ProjNode(mb,TypeFunc::Control);
45 transform_later(*ctrl);
46 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
47 transform_later(mem_proj);
48 if (alias_idx == Compile::AliasIdxBot) {
49 *mem = mem_proj;
50 } else {
51 MergeMemNode* mm = (*mem)->clone()->as_MergeMem();
52 mm->set_memory_at(alias_idx, mem_proj);
53 transform_later(mm);
54 *mem = mm;
55 }
56 }
57
58 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
59 uint shift = exact_log2(type2aelembytes(elembt));
60 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
61 Node* base = basic_plus_adr(ary, header);
62 #ifdef _LP64
63 // see comment in GraphKit::array_element_address
64 int index_max = max_jint - 1; // array size is max_jint, index is one less
65 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
66 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
67 #endif
68 Node* scale = new LShiftXNode(idx, intcon(shift));
69 transform_later(scale);
70 return basic_plus_adr(ary, base, scale);
71 }
72
73 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
74 return transform_later(new ConvI2LNode(offset));
75 }
76
77 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
78 const TypeFunc* call_type, address call_addr,
79 const char* call_name,
128 }
129
130 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
131 transform_later(iff);
132
133 Node* if_slow = new IfTrueNode(iff);
134 transform_later(if_slow);
135
136 if (region != nullptr) {
137 region->add_req(if_slow);
138 }
139
140 Node* if_fast = new IfFalseNode(iff);
141 transform_later(if_fast);
142
143 *ctrl = if_fast;
144
145 return if_slow;
146 }
147
148 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
149 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
150 }
151
152 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
153 if ((*ctrl)->is_top())
154 return; // already stopped
155 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
156 return; // index is already adequately typed
157 Node* cmp_lt = new CmpINode(index, intcon(0));
158 transform_later(cmp_lt);
159 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
160 transform_later(bol_lt);
161 generate_guard(ctrl, bol_lt, region, PROB_MIN);
162 }
163
164 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
165 if ((*ctrl)->is_top())
166 return; // already stopped
167 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
168 if (zero_offset && subseq_length->eqv_uncast(array_length))
169 return; // common case of whole-array copy
170 Node* last = subseq_length;
171 if (!zero_offset) { // last += offset
265
266 *ctrl = stub_block;
267 }
268
269
270 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
271 if ((*ctrl)->is_top()) return nullptr;
272
273 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
274 return nullptr; // index is already adequately typed
275 Node* cmp_le = new CmpINode(index, intcon(0));
276 transform_later(cmp_le);
277 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
278 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
279 transform_later(bol_le);
280 Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
281
282 return is_notp;
283 }
284
285 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
286 transform_later(call);
287
288 *ctrl = new ProjNode(call,TypeFunc::Control);
289 transform_later(*ctrl);
290 Node* newmem = new ProjNode(call, TypeFunc::Memory);
291 transform_later(newmem);
292
293 uint alias_idx = C->get_alias_index(adr_type);
294 if (alias_idx != Compile::AliasIdxBot) {
295 *mem = MergeMemNode::make(*mem);
296 (*mem)->set_memory_at(alias_idx, newmem);
297 } else {
298 *mem = MergeMemNode::make(newmem);
299 }
300 transform_later(*mem);
301 }
302
303 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
304 Node* src_offset,
359 // }
360 // }
361 // // adjust params for remaining work:
362 // if (slowval != -1) {
363 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
364 // }
365 // slow_region:
366 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
367 // return // via slow_call_path
368 //
369 // This routine is used from several intrinsics: System.arraycopy,
370 // Object.clone (the array subcase), and Arrays.copyOf[Range].
371 //
372 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
373 Node** ctrl, MergeMemNode* mem, Node** io,
374 const TypePtr* adr_type,
375 BasicType basic_elem_type,
376 Node* src, Node* src_offset,
377 Node* dest, Node* dest_offset,
378 Node* copy_length,
379 bool disjoint_bases,
380 bool length_never_negative,
381 RegionNode* slow_region) {
382 if (slow_region == nullptr) {
383 slow_region = new RegionNode(1);
384 transform_later(slow_region);
385 }
386
387 bool dest_needs_zeroing = false;
388 bool acopy_to_uninitialized = false;
389
390 // See if this is the initialization of a newly-allocated array.
391 // If so, we will take responsibility here for initializing it to zero.
392 // (Note: Because tightly_coupled_allocation performs checks on the
393 // out-edges of the dest, we need to avoid making derived pointers
394 // from it until we have checked its uses.)
395 if (ReduceBulkZeroing
396 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
397 && basic_elem_type != T_CONFLICT // avoid corner case
398 && !src->eqv_uncast(dest)
399 && alloc != nullptr
400 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
401 assert(ac->is_alloc_tightly_coupled(), "sanity");
402 // acopy to uninitialized tightly coupled allocations
403 // needs zeroing outside the copy range
404 // and the acopy itself will be to uninitialized memory
405 acopy_to_uninitialized = true;
406 if (alloc->maybe_set_complete(&_igvn)) {
407 // "You break it, you buy it."
408 InitializeNode* init = alloc->initialization();
409 assert(init->is_complete(), "we just did this");
410 init->set_complete_with_arraycopy();
411 assert(dest->is_CheckCastPP(), "sanity");
412 assert(dest->in(0)->in(0) == init, "dest pinned");
413 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
414 // From this point on, every exit path is responsible for
415 // initializing any non-copied parts of the object to zero.
416 // Also, if this flag is set we make sure that arraycopy interacts properly
417 // with G1, eliding pre-barriers. See CR 6627983.
418 dest_needs_zeroing = true;
419 } else {
420 // dest_need_zeroing = false;
421 }
422 } else {
423 // No zeroing elimination needed here.
424 alloc = nullptr;
425 acopy_to_uninitialized = false;
426 //dest_needs_zeroing = false;
427 }
428
429 uint alias_idx = C->get_alias_index(adr_type);
430
431 // Results are placed here:
432 enum { fast_path = 1, // normal void-returning assembly stub
433 checked_path = 2, // special assembly stub with cleanup
434 slow_call_path = 3, // something went wrong; call the VM
435 zero_path = 4, // bypass when length of copy is zero
436 bcopy_path = 5, // copy primitive array by 64-bit blocks
437 PATH_LIMIT = 6
438 };
467 checked_i_o = *io;
468 checked_mem = mem->memory_at(alias_idx);
469 checked_value = cv;
470 *ctrl = top();
471 }
472
473 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
474 if (not_pos != nullptr) {
475 Node* local_ctrl = not_pos, *local_io = *io;
476 MergeMemNode* local_mem = MergeMemNode::make(mem);
477 transform_later(local_mem);
478
479 // (6) length must not be negative.
480 if (!length_never_negative) {
481 generate_negative_guard(&local_ctrl, copy_length, slow_region);
482 }
483
484 // copy_length is 0.
485 if (dest_needs_zeroing) {
486 assert(!local_ctrl->is_top(), "no ctrl?");
487 Node* dest_length = alloc->in(AllocateNode::ALength);
488 if (copy_length->eqv_uncast(dest_length)
489 || _igvn.find_int_con(dest_length, 1) <= 0) {
490 // There is no zeroing to do. No need for a secondary raw memory barrier.
491 } else {
492 // Clear the whole thing since there are no source elements to copy.
493 generate_clear_array(local_ctrl, local_mem,
494 adr_type, dest, basic_elem_type,
495 intcon(0), nullptr,
496 alloc->in(AllocateNode::AllocSize));
497 // Use a secondary InitializeNode as raw memory barrier.
498 // Currently it is needed only on this path since other
499 // paths have stub or runtime calls as raw memory barriers.
500 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
501 Compile::AliasIdxRaw,
502 top());
503 transform_later(mb);
504 mb->set_req(TypeFunc::Control,local_ctrl);
505 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
506 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
507 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
508
509 InitializeNode* init = mb->as_Initialize();
510 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
511 }
512 }
513
514 // Present the results of the fast call.
515 result_region->init_req(zero_path, local_ctrl);
516 result_i_o ->init_req(zero_path, local_io);
517 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
518 }
519
520 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
521 // We have to initialize the *uncopied* part of the array to zero.
522 // The copy destination is the slice dest[off..off+len]. The other slices
523 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
524 Node* dest_size = alloc->in(AllocateNode::AllocSize);
525 Node* dest_length = alloc->in(AllocateNode::ALength);
526 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
527
528 // If there is a head section that needs zeroing, do it now.
529 if (_igvn.find_int_con(dest_offset, -1) != 0) {
530 generate_clear_array(*ctrl, mem,
531 adr_type, dest, basic_elem_type,
532 intcon(0), dest_offset,
533 nullptr);
534 }
535
536 // Next, perform a dynamic check on the tail length.
537 // It is often zero, and we can win big if we prove this.
538 // There are two wins: Avoid generating the ClearArray
539 // with its attendant messy index arithmetic, and upgrade
540 // the copy to a more hardware-friendly word size of 64 bits.
541 Node* tail_ctl = nullptr;
542 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
543 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
544 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
545 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
546 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
547 }
548
549 // At this point, let's assume there is no tail.
550 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
551 // There is no tail. Try an upgrade to a 64-bit copy.
559 basic_elem_type, src, src_offset,
560 dest, dest_offset, dest_size, acopy_to_uninitialized);
561 if (didit) {
562 // Present the results of the block-copying fast call.
563 result_region->init_req(bcopy_path, local_ctrl);
564 result_i_o ->init_req(bcopy_path, local_io);
565 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
566 }
567 }
568 if (didit) {
569 *ctrl = top(); // no regular fast path
570 }
571 }
572
573 // Clear the tail, if any.
574 if (tail_ctl != nullptr) {
575 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
576 *ctrl = tail_ctl;
577 if (notail_ctl == nullptr) {
578 generate_clear_array(*ctrl, mem,
579 adr_type, dest, basic_elem_type,
580 dest_tail, nullptr,
581 dest_size);
582 } else {
583 // Make a local merge.
584 Node* done_ctl = transform_later(new RegionNode(3));
585 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
586 done_ctl->init_req(1, notail_ctl);
587 done_mem->init_req(1, mem->memory_at(alias_idx));
588 generate_clear_array(*ctrl, mem,
589 adr_type, dest, basic_elem_type,
590 dest_tail, nullptr,
591 dest_size);
592 done_ctl->init_req(2, *ctrl);
593 done_mem->init_req(2, mem->memory_at(alias_idx));
594 *ctrl = done_ctl;
595 mem->set_memory_at(alias_idx, done_mem);
596 }
597 }
598 }
599
600 BasicType copy_type = basic_elem_type;
601 assert(basic_elem_type != T_ARRAY, "caller must fix this");
602 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
603 // If src and dest have compatible element types, we can copy bits.
604 // Types S[] and D[] are compatible if D is a supertype of S.
605 //
606 // If they are not, we will use checked_oop_disjoint_arraycopy,
607 // which performs a fast optimistic per-oop check, and backs off
608 // further to JVM_ArrayCopy on the first per-oop check that fails.
609 // (Actually, we don't move raw bits only; the GC requires card marks.)
746 Node* length_minus = new SubINode(copy_length, slow_offset);
747 transform_later(length_minus);
748
749 // Tweak the node variables to adjust the code produced below:
750 src_offset = src_off_plus;
751 dest_offset = dest_off_plus;
752 copy_length = length_minus;
753 }
754 }
755 *ctrl = slow_control;
756 if (!(*ctrl)->is_top()) {
757 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
758 MergeMemNode* local_mem = MergeMemNode::make(mem);
759 transform_later(local_mem);
760
761 // Generate the slow path, if needed.
762 local_mem->set_memory_at(alias_idx, slow_mem);
763
764 if (dest_needs_zeroing) {
765 generate_clear_array(local_ctrl, local_mem,
766 adr_type, dest, basic_elem_type,
767 intcon(0), nullptr,
768 alloc->in(AllocateNode::AllocSize));
769 }
770
771 local_mem = generate_slow_arraycopy(ac,
772 &local_ctrl, local_mem, &local_io,
773 adr_type,
774 src, src_offset, dest, dest_offset,
775 copy_length, /*dest_uninitialized*/false);
776
777 result_region->init_req(slow_call_path, local_ctrl);
778 result_i_o ->init_req(slow_call_path, local_io);
779 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
780 } else {
781 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
782 // projections were not extracted
783 }
784
785 // Remove unused edges.
786 for (uint i = 1; i < result_region->req(); i++) {
815 // a subsequent store that would make this object accessible by
816 // other threads.
817 assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
818 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
819 } else {
820 int mem_bar_alias_idx = Compile::AliasIdxBot;
821 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
822 // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
823 // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
824 // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
825 // broken graph.
826 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
827 }
828 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
829 }
830
831 assert((*ctrl)->is_Proj(), "MemBar control projection");
832 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
833 (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
834
835 _igvn.replace_node(_callprojs.fallthrough_memproj, out_mem);
836 if (_callprojs.fallthrough_ioproj != nullptr) {
837 _igvn.replace_node(_callprojs.fallthrough_ioproj, *io);
838 }
839 _igvn.replace_node(_callprojs.fallthrough_catchproj, *ctrl);
840
841 #ifdef ASSERT
842 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
843 if (dest_t->is_known_instance()) {
844 ArrayCopyNode* ac = nullptr;
845 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
846 assert(ac == nullptr, "no arraycopy anymore");
847 }
848 #endif
849
850 return out_mem;
851 }
852
853 // Helper for initialization of arrays, creating a ClearArray.
854 // It writes zero bits in [start..end), within the body of an array object.
855 // The memory effects are all chained onto the 'adr_type' alias category.
856 //
857 // Since the object is otherwise uninitialized, we are free
858 // to put a little "slop" around the edges of the cleared area,
859 // as long as it does not go back into the array's header,
860 // or beyond the array end within the heap.
861 //
862 // The lower edge can be rounded down to the nearest jint and the
863 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
864 //
865 // Arguments:
866 // adr_type memory slice where writes are generated
867 // dest oop of the destination array
868 // basic_elem_type element type of the destination
869 // slice_idx array index of first element to store
870 // slice_len number of elements to store (or null)
871 // dest_size total size in bytes of the array object
872 //
873 // Exactly one of slice_len or dest_size must be non-null.
874 // If dest_size is non-null, zeroing extends to the end of the object.
875 // If slice_len is non-null, the slice_idx value must be a constant.
876 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
877 const TypePtr* adr_type,
878 Node* dest,
879 BasicType basic_elem_type,
880 Node* slice_idx,
881 Node* slice_len,
882 Node* dest_size) {
883 // one or the other but not both of slice_len and dest_size:
884 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
885 if (slice_len == nullptr) slice_len = top();
886 if (dest_size == nullptr) dest_size = top();
887
888 uint alias_idx = C->get_alias_index(adr_type);
889
890 // operate on this memory slice:
891 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
892
893 // scaling and rounding of indexes:
894 int scale = exact_log2(type2aelembytes(basic_elem_type));
895 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
896 int clear_low = (-1 << scale) & (BytesPerInt - 1);
897 int bump_bit = (-1 << scale) & BytesPerInt;
898
899 // determine constant starts and ends
900 const intptr_t BIG_NEG = -128;
901 assert(BIG_NEG + 2*abase < 0, "neg enough");
902 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
903 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
904 if (slice_len_con == 0) {
905 return; // nothing to do here
906 }
907 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
908 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
909 if (slice_idx_con >= 0 && slice_len_con >= 0) {
910 assert(end_con < 0, "not two cons");
911 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
912 BytesPerLong);
913 }
914
915 if (start_con >= 0 && end_con >= 0) {
916 // Constant start and end. Simple.
917 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
918 start_con, end_con, false, &_igvn);
919 } else if (start_con >= 0 && dest_size != top()) {
920 // Constant start, pre-rounded end after the tail of the array.
921 Node* end = dest_size;
922 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
923 start_con, end, false, &_igvn);
924 } else if (start_con >= 0 && slice_len != top()) {
925 // Constant start, non-constant end. End needs rounding up.
926 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
927 intptr_t end_base = abase + (slice_idx_con << scale);
928 int end_round = (-1 << scale) & (BytesPerLong - 1);
929 Node* end = ConvI2X(slice_len);
930 if (scale != 0)
931 end = transform_later(new LShiftXNode(end, intcon(scale) ));
932 end_base += end_round;
933 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
934 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
935 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
936 start_con, end, false, &_igvn);
937 } else if (start_con < 0 && dest_size != top()) {
938 // Non-constant start, pre-rounded end after the tail of the array.
939 // This is almost certainly a "round-to-end" operation.
940 Node* start = slice_idx;
941 start = ConvI2X(start);
942 if (scale != 0)
943 start = transform_later(new LShiftXNode( start, intcon(scale) ));
944 start = transform_later(new AddXNode(start, MakeConX(abase)) );
945 if ((bump_bit | clear_low) != 0) {
946 int to_clear = (bump_bit | clear_low);
947 // Align up mod 8, then store a jint zero unconditionally
948 // just before the mod-8 boundary.
949 if (((abase + bump_bit) & ~to_clear) - bump_bit
950 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
951 bump_bit = 0;
952 assert((abase & to_clear) == 0, "array base must be long-aligned");
953 } else {
954 // Bump 'start' up to (or past) the next jint boundary:
955 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
956 assert((abase & clear_low) == 0, "array base must be int-aligned");
957 }
958 // Round bumped 'start' down to jlong boundary in body of array.
959 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
960 if (bump_bit != 0) {
961 // Store a zero to the immediately preceding jint:
962 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
963 Node* p1 = basic_plus_adr(dest, x1);
964 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
965 mem = transform_later(mem);
966 }
967 }
968 Node* end = dest_size; // pre-rounded
969 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
970 start, end, false, &_igvn);
971 } else {
972 // Non-constant start, unrounded non-constant end.
973 // (Nobody zeroes a random midsection of an array using this routine.)
974 ShouldNotReachHere(); // fix caller
975 }
976
977 // Done.
978 merge_mem->set_memory_at(alias_idx, mem);
979 }
980
981 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem,
982 const TypePtr* adr_type,
983 BasicType basic_elem_type,
984 Node* src, Node* src_offset,
985 Node* dest, Node* dest_offset,
986 Node* dest_size, bool dest_uninitialized) {
987 // See if there is an advantage from block transfer.
988 int scale = exact_log2(type2aelembytes(basic_elem_type));
989 if (scale >= LogBytesPerLong)
1064 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1065 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1066 "slow_arraycopy", TypePtr::BOTTOM);
1067
1068 call->init_req(TypeFunc::Control, *ctrl);
1069 call->init_req(TypeFunc::I_O , *io);
1070 call->init_req(TypeFunc::Memory , mem);
1071 call->init_req(TypeFunc::ReturnAdr, top());
1072 call->init_req(TypeFunc::FramePtr, top());
1073 call->init_req(TypeFunc::Parms+0, src);
1074 call->init_req(TypeFunc::Parms+1, src_offset);
1075 call->init_req(TypeFunc::Parms+2, dest);
1076 call->init_req(TypeFunc::Parms+3, dest_offset);
1077 call->init_req(TypeFunc::Parms+4, copy_length);
1078 call->copy_call_debug_info(&_igvn, ac);
1079
1080 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1081 _igvn.replace_node(ac, call);
1082 transform_later(call);
1083
1084 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1085 *ctrl = _callprojs.fallthrough_catchproj->clone();
1086 transform_later(*ctrl);
1087
1088 Node* m = _callprojs.fallthrough_memproj->clone();
1089 transform_later(m);
1090
1091 uint alias_idx = C->get_alias_index(adr_type);
1092 MergeMemNode* out_mem;
1093 if (alias_idx != Compile::AliasIdxBot) {
1094 out_mem = MergeMemNode::make(mem);
1095 out_mem->set_memory_at(alias_idx, m);
1096 } else {
1097 out_mem = MergeMemNode::make(m);
1098 }
1099 transform_later(out_mem);
1100
1101 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1102 // could be null. Skip clone and update null fallthrough_ioproj.
1103 if (_callprojs.fallthrough_ioproj != nullptr) {
1104 *io = _callprojs.fallthrough_ioproj->clone();
1105 transform_later(*io);
1106 } else {
1107 *io = nullptr;
1108 }
1109
1110 return out_mem;
1111 }
1112
1113 // Helper function; generates code for cases requiring runtime checks.
1114 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1115 const TypePtr* adr_type,
1116 Node* dest_elem_klass,
1117 Node* src, Node* src_offset,
1118 Node* dest, Node* dest_offset,
1119 Node* copy_length, bool dest_uninitialized) {
1120 if ((*ctrl)->is_top()) return nullptr;
1121
1122 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1123 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1124 return nullptr;
1216
1217 // Connecting remaining edges for exit_block coming from stub_block.
1218 if (exit_block) {
1219 exit_block->init_req(2, *ctrl);
1220
1221 // Memory edge corresponding to stub_region.
1222 result_memory->init_req(2, *mem);
1223
1224 uint alias_idx = C->get_alias_index(adr_type);
1225 if (alias_idx != Compile::AliasIdxBot) {
1226 *mem = MergeMemNode::make(*mem);
1227 (*mem)->set_memory_at(alias_idx, result_memory);
1228 } else {
1229 *mem = MergeMemNode::make(result_memory);
1230 }
1231 transform_later(*mem);
1232 *ctrl = exit_block;
1233 }
1234 }
1235
1236 #undef XTOP
1237
1238 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1239 Node* ctrl = ac->in(TypeFunc::Control);
1240 Node* io = ac->in(TypeFunc::I_O);
1241 Node* src = ac->in(ArrayCopyNode::Src);
1242 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1243 Node* dest = ac->in(ArrayCopyNode::Dest);
1244 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1245 Node* length = ac->in(ArrayCopyNode::Length);
1246 MergeMemNode* merge_mem = nullptr;
1247
1248 if (ac->is_clonebasic()) {
1249 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1250 bs->clone_at_expansion(this, ac);
1251 return;
1252 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1253 Node* mem = ac->in(TypeFunc::Memory);
1254 merge_mem = MergeMemNode::make(mem);
1255 transform_later(merge_mem);
1256
1257 AllocateArrayNode* alloc = nullptr;
1258 if (ac->is_alloc_tightly_coupled()) {
1259 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1260 assert(alloc != nullptr, "expect alloc");
1261 }
1262
1263 const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1264 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1265 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1266 }
1267 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1268 adr_type, T_OBJECT,
1269 src, src_offset, dest, dest_offset, length,
1270 true, ac->has_negative_length_guard());
1271
1272 return;
1273 }
1274
1275 AllocateArrayNode* alloc = nullptr;
1276 if (ac->is_alloc_tightly_coupled()) {
1277 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1278 assert(alloc != nullptr, "expect alloc");
1279 }
1280
1281 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1282
1283 // Compile time checks. If any of these checks cannot be verified at compile time,
1284 // we do not make a fast path for this call. Instead, we let the call remain as it
1285 // is. The checks we choose to mandate at compile time are:
1286 //
1287 // (1) src and dest are arrays.
1288 const Type* src_type = src->Value(&_igvn);
1289 const Type* dest_type = dest->Value(&_igvn);
1290 const TypeAryPtr* top_src = src_type->isa_aryptr();
1291 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1292
1293 BasicType src_elem = T_CONFLICT;
1294 BasicType dest_elem = T_CONFLICT;
1295
1296 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1297 src_elem = top_src->elem()->array_element_basic_type();
1298 }
1299 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1300 dest_elem = top_dest->elem()->array_element_basic_type();
1301 }
1302 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1303 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1304
1305 if (ac->is_arraycopy_validated() &&
1306 dest_elem != T_CONFLICT &&
1307 src_elem == T_CONFLICT) {
1308 src_elem = dest_elem;
1309 }
1310
1311 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1312 // Conservatively insert a memory barrier on all memory slices.
1313 // Do not let writes into the source float below the arraycopy.
1314 {
1315 Node* mem = ac->in(TypeFunc::Memory);
1316 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1317
1318 merge_mem = MergeMemNode::make(mem);
1319 transform_later(merge_mem);
1320 }
1321
1322 // Call StubRoutines::generic_arraycopy stub.
1323 Node* mem = generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1324 TypeRawPtr::BOTTOM, T_CONFLICT,
1325 src, src_offset, dest, dest_offset, length,
1326 // If a negative length guard was generated for the ArrayCopyNode,
1327 // the length of the array can never be negative.
1328 false, ac->has_negative_length_guard());
1329 return;
1330 }
1331
1332 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1333
1334 // (2) src and dest arrays must have elements of the same BasicType
1335 // Figure out the size and type of the elements we will be copying.
1336 if (src_elem != dest_elem || dest_elem == T_VOID) {
1337 // The component types are not the same or are not recognized. Punt.
1338 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1339 {
1340 Node* mem = ac->in(TypeFunc::Memory);
1341 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1342 }
1343
1344 _igvn.replace_node(_callprojs.fallthrough_memproj, merge_mem);
1345 if (_callprojs.fallthrough_ioproj != nullptr) {
1346 _igvn.replace_node(_callprojs.fallthrough_ioproj, io);
1347 }
1348 _igvn.replace_node(_callprojs.fallthrough_catchproj, ctrl);
1349 return;
1350 }
1351
1352 //---------------------------------------------------------------------------
1353 // We will make a fast path for this call to arraycopy.
1354
1355 // We have the following tests left to perform:
1356 //
1357 // (3) src and dest must not be null.
1358 // (4) src_offset must not be negative.
1359 // (5) dest_offset must not be negative.
1360 // (6) length must not be negative.
1361 // (7) src_offset + length must not exceed length of src.
1362 // (8) dest_offset + length must not exceed length of dest.
1363 // (9) each element of an oop array must be assignable
1364
1365 {
1366 Node* mem = ac->in(TypeFunc::Memory);
1367 merge_mem = MergeMemNode::make(mem);
1368 transform_later(merge_mem);
1369 }
1370
1371 RegionNode* slow_region = new RegionNode(1);
1372 transform_later(slow_region);
1373
1374 if (!ac->is_arraycopy_validated()) {
1375 // (3) operands must not be null
1376 // We currently perform our null checks with the null_check routine.
1377 // This means that the null exceptions will be reported in the caller
1378 // rather than (correctly) reported inside of the native arraycopy call.
1379 // This should be corrected, given time. We do our null check with the
1380 // stack pointer restored.
1381 // null checks done library_call.cpp
1382
1383 // (4) src_offset must not be negative.
1384 generate_negative_guard(&ctrl, src_offset, slow_region);
1385
1386 // (5) dest_offset must not be negative.
1387 generate_negative_guard(&ctrl, dest_offset, slow_region);
1388
1389 // (6) length must not be negative (moved to generate_arraycopy()).
1390 // generate_negative_guard(length, slow_region);
1391
1392 // (7) src_offset + length must not exceed length of src.
1393 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1394 assert(alen != nullptr, "need src len");
1395 generate_limit_guard(&ctrl,
1396 src_offset, length,
1397 alen,
1398 slow_region);
1399
1400 // (8) dest_offset + length must not exceed length of dest.
1401 alen = ac->in(ArrayCopyNode::DestLen);
1402 assert(alen != nullptr, "need dest len");
1403 generate_limit_guard(&ctrl,
1404 dest_offset, length,
1405 alen,
1406 slow_region);
1407
1408 // (9) each element of an oop array must be assignable
1409 // The generate_arraycopy subroutine checks this.
1410 }
1411 // This is where the memory effects are placed:
1412 const TypePtr* adr_type = nullptr;
1413 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1414 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1415 } else {
1416 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1417 }
1418
1419 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1420 adr_type, dest_elem,
1421 src, src_offset, dest, dest_offset, length,
1422 // If a negative length guard was generated for the ArrayCopyNode,
1423 // the length of the array can never be negative.
1424 false, ac->has_negative_length_guard(), slow_region);
1425 }
|
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/tlab_globals.hpp"
28 #include "oops/objArrayKlass.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/convertnode.hpp"
32 #include "opto/graphKit.hpp"
33 #include "opto/macro.hpp"
34 #include "opto/runtime.hpp"
35 #include "opto/vectornode.hpp"
36 #include "runtime/stubRoutines.hpp"
37 #include "utilities/align.hpp"
38 #include "utilities/powerOfTwo.hpp"
39
40 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, int alias_idx, Node* precedent) {
41 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
42 mb->init_req(TypeFunc::Control, *ctrl);
43 mb->init_req(TypeFunc::Memory, *mem);
44 transform_later(mb);
45 *ctrl = new ProjNode(mb,TypeFunc::Control);
46 transform_later(*ctrl);
47 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
48 transform_later(mem_proj);
49 if (alias_idx == Compile::AliasIdxBot) {
50 *mem = mem_proj;
51 } else {
52 MergeMemNode* mm = (*mem)->clone()->as_MergeMem();
53 mm->set_memory_at(alias_idx, mem_proj);
54 transform_later(mm);
55 *mem = mm;
56 }
57 }
58
59 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
60 uint shift = exact_log2(type2aelembytes(elembt));
61 const TypeAryPtr* array_type = _igvn.type(ary)->isa_aryptr();
62 if (array_type != nullptr && array_type->is_aryptr()->is_flat()) {
63 // Use T_FLAT_ELEMENT to get proper alignment with COH when fetching the array element address.
64 elembt = T_FLAT_ELEMENT;
65 }
66 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
67 Node* base = basic_plus_adr(ary, header);
68 #ifdef _LP64
69 // see comment in GraphKit::array_element_address
70 int index_max = max_jint - 1; // array size is max_jint, index is one less
71 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
72 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
73 #endif
74 Node* scale = new LShiftXNode(idx, intcon(shift));
75 transform_later(scale);
76 return basic_plus_adr(ary, base, scale);
77 }
78
79 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
80 return transform_later(new ConvI2LNode(offset));
81 }
82
83 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
84 const TypeFunc* call_type, address call_addr,
85 const char* call_name,
134 }
135
136 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
137 transform_later(iff);
138
139 Node* if_slow = new IfTrueNode(iff);
140 transform_later(if_slow);
141
142 if (region != nullptr) {
143 region->add_req(if_slow);
144 }
145
146 Node* if_fast = new IfFalseNode(iff);
147 transform_later(if_fast);
148
149 *ctrl = if_fast;
150
151 return if_slow;
152 }
153
154 Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
155 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
156 }
157
158 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
159 return generate_guard(ctrl, test, region, PROB_FAIR);
160 }
161
162 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
163 if ((*ctrl)->is_top())
164 return; // already stopped
165 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
166 return; // index is already adequately typed
167 Node* cmp_lt = new CmpINode(index, intcon(0));
168 transform_later(cmp_lt);
169 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
170 transform_later(bol_lt);
171 generate_guard(ctrl, bol_lt, region, PROB_MIN);
172 }
173
174 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
175 if ((*ctrl)->is_top())
176 return; // already stopped
177 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
178 if (zero_offset && subseq_length->eqv_uncast(array_length))
179 return; // common case of whole-array copy
180 Node* last = subseq_length;
181 if (!zero_offset) { // last += offset
275
276 *ctrl = stub_block;
277 }
278
279
280 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
281 if ((*ctrl)->is_top()) return nullptr;
282
283 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
284 return nullptr; // index is already adequately typed
285 Node* cmp_le = new CmpINode(index, intcon(0));
286 transform_later(cmp_le);
287 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
288 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
289 transform_later(bol_le);
290 Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
291
292 return is_notp;
293 }
294
295 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
296 // Load markword and check if obj is locked
297 Node* mark = make_load_raw(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
298 Node* locked_bit = MakeConX(markWord::unlocked_value);
299 locked_bit = transform_later(new AndXNode(locked_bit, mark));
300 Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
301 Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
302 IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
303 Node* locked_region = transform_later(new RegionNode(3));
304 Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
305
306 // Unlocked: Use bits from mark word
307 locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
308 mark_phi->init_req(1, mark);
309
310 // Locked: Load prototype header from klass
311 *ctrl = transform_later(new IfFalseNode(iff));
312 // Make loads control dependent to make sure they are only executed if array is locked
313 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
314 Node* klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
315 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
316 Node* proto = transform_later(LoadNode::make(_igvn, *ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
317
318 locked_region->init_req(2, *ctrl);
319 mark_phi->init_req(2, proto);
320 *ctrl = locked_region;
321
322 // Now check if mark word bits are set
323 Node* mask = MakeConX(mask_val);
324 Node* masked = transform_later(new AndXNode(mark_phi, mask));
325 cmp = transform_later(new CmpXNode(masked, mask));
326 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
327 return generate_fair_guard(ctrl, bol, region);
328 }
329
330 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
331 return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
332 }
333
334 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
335 return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
336 }
337
338 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
339 transform_later(call);
340
341 *ctrl = new ProjNode(call,TypeFunc::Control);
342 transform_later(*ctrl);
343 Node* newmem = new ProjNode(call, TypeFunc::Memory);
344 transform_later(newmem);
345
346 uint alias_idx = C->get_alias_index(adr_type);
347 if (alias_idx != Compile::AliasIdxBot) {
348 *mem = MergeMemNode::make(*mem);
349 (*mem)->set_memory_at(alias_idx, newmem);
350 } else {
351 *mem = MergeMemNode::make(newmem);
352 }
353 transform_later(*mem);
354 }
355
356 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
357 Node* src_offset,
412 // }
413 // }
414 // // adjust params for remaining work:
415 // if (slowval != -1) {
416 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
417 // }
418 // slow_region:
419 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
420 // return // via slow_call_path
421 //
422 // This routine is used from several intrinsics: System.arraycopy,
423 // Object.clone (the array subcase), and Arrays.copyOf[Range].
424 //
425 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
426 Node** ctrl, MergeMemNode* mem, Node** io,
427 const TypePtr* adr_type,
428 BasicType basic_elem_type,
429 Node* src, Node* src_offset,
430 Node* dest, Node* dest_offset,
431 Node* copy_length,
432 Node* dest_length,
433 bool disjoint_bases,
434 bool length_never_negative,
435 RegionNode* slow_region) {
436 if (slow_region == nullptr) {
437 slow_region = new RegionNode(1);
438 transform_later(slow_region);
439 }
440
441 bool dest_needs_zeroing = false;
442 bool acopy_to_uninitialized = false;
443 Node* init_value = nullptr;
444 Node* raw_init_value = nullptr;
445
446 // See if this is the initialization of a newly-allocated array.
447 // If so, we will take responsibility here for initializing it to zero.
448 // (Note: Because tightly_coupled_allocation performs checks on the
449 // out-edges of the dest, we need to avoid making derived pointers
450 // from it until we have checked its uses.)
451 if (ReduceBulkZeroing
452 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
453 && basic_elem_type != T_CONFLICT // avoid corner case
454 && !src->eqv_uncast(dest)
455 && alloc != nullptr
456 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
457 assert(ac->is_alloc_tightly_coupled(), "sanity");
458 // acopy to uninitialized tightly coupled allocations
459 // needs zeroing outside the copy range
460 // and the acopy itself will be to uninitialized memory
461 acopy_to_uninitialized = true;
462 if (alloc->maybe_set_complete(&_igvn)) {
463 // "You break it, you buy it."
464 InitializeNode* init = alloc->initialization();
465 assert(init->is_complete(), "we just did this");
466 init->set_complete_with_arraycopy();
467 assert(dest->is_CheckCastPP(), "sanity");
468 assert(dest->in(0)->in(0) == init, "dest pinned");
469 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
470 // From this point on, every exit path is responsible for
471 // initializing any non-copied parts of the object to zero.
472 // Also, if this flag is set we make sure that arraycopy interacts properly
473 // with G1, eliding pre-barriers. See CR 6627983.
474 dest_needs_zeroing = true;
475 init_value = alloc->in(AllocateNode::InitValue);
476 raw_init_value = alloc->in(AllocateNode::RawInitValue);
477 } else {
478 // dest_need_zeroing = false;
479 }
480 } else {
481 // No zeroing elimination needed here.
482 alloc = nullptr;
483 acopy_to_uninitialized = false;
484 //dest_needs_zeroing = false;
485 }
486
487 uint alias_idx = C->get_alias_index(adr_type);
488
489 // Results are placed here:
490 enum { fast_path = 1, // normal void-returning assembly stub
491 checked_path = 2, // special assembly stub with cleanup
492 slow_call_path = 3, // something went wrong; call the VM
493 zero_path = 4, // bypass when length of copy is zero
494 bcopy_path = 5, // copy primitive array by 64-bit blocks
495 PATH_LIMIT = 6
496 };
525 checked_i_o = *io;
526 checked_mem = mem->memory_at(alias_idx);
527 checked_value = cv;
528 *ctrl = top();
529 }
530
531 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
532 if (not_pos != nullptr) {
533 Node* local_ctrl = not_pos, *local_io = *io;
534 MergeMemNode* local_mem = MergeMemNode::make(mem);
535 transform_later(local_mem);
536
537 // (6) length must not be negative.
538 if (!length_never_negative) {
539 generate_negative_guard(&local_ctrl, copy_length, slow_region);
540 }
541
542 // copy_length is 0.
543 if (dest_needs_zeroing) {
544 assert(!local_ctrl->is_top(), "no ctrl?");
545 if (copy_length->eqv_uncast(dest_length)
546 || _igvn.find_int_con(dest_length, 1) <= 0) {
547 // There is no zeroing to do. No need for a secondary raw memory barrier.
548 } else {
549 // Clear the whole thing since there are no source elements to copy.
550 generate_clear_array(local_ctrl, local_mem,
551 adr_type, dest,
552 init_value, raw_init_value,
553 basic_elem_type,
554 intcon(0), nullptr,
555 alloc->in(AllocateNode::AllocSize));
556 // Use a secondary InitializeNode as raw memory barrier.
557 // Currently it is needed only on this path since other
558 // paths have stub or runtime calls as raw memory barriers.
559 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
560 Compile::AliasIdxRaw,
561 top());
562 transform_later(mb);
563 mb->set_req(TypeFunc::Control,local_ctrl);
564 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
565 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
566 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
567
568 InitializeNode* init = mb->as_Initialize();
569 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
570 }
571 }
572
573 // Present the results of the fast call.
574 result_region->init_req(zero_path, local_ctrl);
575 result_i_o ->init_req(zero_path, local_io);
576 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
577 }
578
579 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
580 // We have to initialize the *uncopied* part of the array to zero.
581 // The copy destination is the slice dest[off..off+len]. The other slices
582 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
583 Node* dest_size = alloc->in(AllocateNode::AllocSize);
584 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
585
586 // If there is a head section that needs zeroing, do it now.
587 if (_igvn.find_int_con(dest_offset, -1) != 0) {
588 generate_clear_array(*ctrl, mem,
589 adr_type, dest,
590 init_value, raw_init_value,
591 basic_elem_type,
592 intcon(0), dest_offset,
593 nullptr);
594 }
595
596 // Next, perform a dynamic check on the tail length.
597 // It is often zero, and we can win big if we prove this.
598 // There are two wins: Avoid generating the ClearArray
599 // with its attendant messy index arithmetic, and upgrade
600 // the copy to a more hardware-friendly word size of 64 bits.
601 Node* tail_ctl = nullptr;
602 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
603 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
604 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
605 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
606 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
607 }
608
609 // At this point, let's assume there is no tail.
610 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
611 // There is no tail. Try an upgrade to a 64-bit copy.
619 basic_elem_type, src, src_offset,
620 dest, dest_offset, dest_size, acopy_to_uninitialized);
621 if (didit) {
622 // Present the results of the block-copying fast call.
623 result_region->init_req(bcopy_path, local_ctrl);
624 result_i_o ->init_req(bcopy_path, local_io);
625 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
626 }
627 }
628 if (didit) {
629 *ctrl = top(); // no regular fast path
630 }
631 }
632
633 // Clear the tail, if any.
634 if (tail_ctl != nullptr) {
635 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
636 *ctrl = tail_ctl;
637 if (notail_ctl == nullptr) {
638 generate_clear_array(*ctrl, mem,
639 adr_type, dest,
640 init_value, raw_init_value,
641 basic_elem_type,
642 dest_tail, nullptr,
643 dest_size);
644 } else {
645 // Make a local merge.
646 Node* done_ctl = transform_later(new RegionNode(3));
647 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
648 done_ctl->init_req(1, notail_ctl);
649 done_mem->init_req(1, mem->memory_at(alias_idx));
650 generate_clear_array(*ctrl, mem,
651 adr_type, dest,
652 init_value, raw_init_value,
653 basic_elem_type,
654 dest_tail, nullptr,
655 dest_size);
656 done_ctl->init_req(2, *ctrl);
657 done_mem->init_req(2, mem->memory_at(alias_idx));
658 *ctrl = done_ctl;
659 mem->set_memory_at(alias_idx, done_mem);
660 }
661 }
662 }
663
664 BasicType copy_type = basic_elem_type;
665 assert(basic_elem_type != T_ARRAY, "caller must fix this");
666 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
667 // If src and dest have compatible element types, we can copy bits.
668 // Types S[] and D[] are compatible if D is a supertype of S.
669 //
670 // If they are not, we will use checked_oop_disjoint_arraycopy,
671 // which performs a fast optimistic per-oop check, and backs off
672 // further to JVM_ArrayCopy on the first per-oop check that fails.
673 // (Actually, we don't move raw bits only; the GC requires card marks.)
810 Node* length_minus = new SubINode(copy_length, slow_offset);
811 transform_later(length_minus);
812
813 // Tweak the node variables to adjust the code produced below:
814 src_offset = src_off_plus;
815 dest_offset = dest_off_plus;
816 copy_length = length_minus;
817 }
818 }
819 *ctrl = slow_control;
820 if (!(*ctrl)->is_top()) {
821 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
822 MergeMemNode* local_mem = MergeMemNode::make(mem);
823 transform_later(local_mem);
824
825 // Generate the slow path, if needed.
826 local_mem->set_memory_at(alias_idx, slow_mem);
827
828 if (dest_needs_zeroing) {
829 generate_clear_array(local_ctrl, local_mem,
830 adr_type, dest,
831 init_value, raw_init_value,
832 basic_elem_type,
833 intcon(0), nullptr,
834 alloc->in(AllocateNode::AllocSize));
835 }
836
837 local_mem = generate_slow_arraycopy(ac,
838 &local_ctrl, local_mem, &local_io,
839 adr_type,
840 src, src_offset, dest, dest_offset,
841 copy_length, /*dest_uninitialized*/false);
842
843 result_region->init_req(slow_call_path, local_ctrl);
844 result_i_o ->init_req(slow_call_path, local_io);
845 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
846 } else {
847 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
848 // projections were not extracted
849 }
850
851 // Remove unused edges.
852 for (uint i = 1; i < result_region->req(); i++) {
881 // a subsequent store that would make this object accessible by
882 // other threads.
883 assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
884 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
885 } else {
886 int mem_bar_alias_idx = Compile::AliasIdxBot;
887 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
888 // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
889 // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
890 // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
891 // broken graph.
892 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
893 }
894 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
895 }
896
897 assert((*ctrl)->is_Proj(), "MemBar control projection");
898 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
899 (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
900
901 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
902 if (_callprojs->fallthrough_ioproj != nullptr) {
903 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
904 }
905 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
906
907 #ifdef ASSERT
908 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
909 if (dest_t->is_known_instance()) {
910 ArrayCopyNode* ac = nullptr;
911 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
912 assert(ac == nullptr, "no arraycopy anymore");
913 }
914 #endif
915
916 return out_mem;
917 }
918
919 // Helper for initialization of arrays, creating a ClearArray.
920 // It writes zero bits in [start..end), within the body of an array object.
921 // The memory effects are all chained onto the 'adr_type' alias category.
922 //
923 // Since the object is otherwise uninitialized, we are free
924 // to put a little "slop" around the edges of the cleared area,
925 // as long as it does not go back into the array's header,
926 // or beyond the array end within the heap.
927 //
928 // The lower edge can be rounded down to the nearest jint and the
929 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
930 //
931 // Arguments:
932 // adr_type memory slice where writes are generated
933 // dest oop of the destination array
934 // basic_elem_type element type of the destination
935 // slice_idx array index of first element to store
936 // slice_len number of elements to store (or null)
937 // dest_size total size in bytes of the array object
938 //
939 // Exactly one of slice_len or dest_size must be non-null.
940 // If dest_size is non-null, zeroing extends to the end of the object.
941 // If slice_len is non-null, the slice_idx value must be a constant.
942 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
943 const TypePtr* adr_type,
944 Node* dest,
945 Node* val,
946 Node* raw_val,
947 BasicType basic_elem_type,
948 Node* slice_idx,
949 Node* slice_len,
950 Node* dest_size) {
951 // one or the other but not both of slice_len and dest_size:
952 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
953 if (slice_len == nullptr) slice_len = top();
954 if (dest_size == nullptr) dest_size = top();
955
956 uint alias_idx = C->get_alias_index(adr_type);
957
958 // operate on this memory slice:
959 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
960
961 // scaling and rounding of indexes:
962 int scale = exact_log2(type2aelembytes(basic_elem_type));
963 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
964 int clear_low = (-1 << scale) & (BytesPerInt - 1);
965 int bump_bit = (-1 << scale) & BytesPerInt;
966
967 // determine constant starts and ends
968 const intptr_t BIG_NEG = -128;
969 assert(BIG_NEG + 2*abase < 0, "neg enough");
970 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
971 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
972 if (slice_len_con == 0) {
973 return; // nothing to do here
974 }
975 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
976 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
977 if (slice_idx_con >= 0 && slice_len_con >= 0) {
978 assert(end_con < 0, "not two cons");
979 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
980 BytesPerLong);
981 }
982
983 if (start_con >= 0 && end_con >= 0) {
984 // Constant start and end. Simple.
985 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
986 start_con, end_con, false, &_igvn);
987 } else if (start_con >= 0 && dest_size != top()) {
988 // Constant start, pre-rounded end after the tail of the array.
989 Node* end = dest_size;
990 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
991 start_con, end, false, &_igvn);
992 } else if (start_con >= 0 && slice_len != top()) {
993 // Constant start, non-constant end. End needs rounding up.
994 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
995 intptr_t end_base = abase + (slice_idx_con << scale);
996 int end_round = (-1 << scale) & (BytesPerLong - 1);
997 Node* end = ConvI2X(slice_len);
998 if (scale != 0)
999 end = transform_later(new LShiftXNode(end, intcon(scale) ));
1000 end_base += end_round;
1001 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
1002 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
1003 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1004 start_con, end, false, &_igvn);
1005 } else if (start_con < 0 && dest_size != top()) {
1006 // Non-constant start, pre-rounded end after the tail of the array.
1007 // This is almost certainly a "round-to-end" operation.
1008 Node* start = slice_idx;
1009 start = ConvI2X(start);
1010 if (scale != 0)
1011 start = transform_later(new LShiftXNode( start, intcon(scale) ));
1012 start = transform_later(new AddXNode(start, MakeConX(abase)) );
1013 if ((bump_bit | clear_low) != 0) {
1014 int to_clear = (bump_bit | clear_low);
1015 // Align up mod 8, then store a jint zero unconditionally
1016 // just before the mod-8 boundary.
1017 if (((abase + bump_bit) & ~to_clear) - bump_bit
1018 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1019 bump_bit = 0;
1020 assert((abase & to_clear) == 0, "array base must be long-aligned");
1021 } else {
1022 // Bump 'start' up to (or past) the next jint boundary:
1023 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1024 assert((abase & clear_low) == 0, "array base must be int-aligned");
1025 }
1026 // Round bumped 'start' down to jlong boundary in body of array.
1027 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1028 if (bump_bit != 0) {
1029 // Store a zero to the immediately preceding jint:
1030 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1031 Node* p1 = basic_plus_adr(dest, x1);
1032 if (val == nullptr) {
1033 assert(raw_val == nullptr, "val may not be null");
1034 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1035 } else {
1036 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1037 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1038 }
1039 mem = transform_later(mem);
1040 }
1041 }
1042 Node* end = dest_size; // pre-rounded
1043 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1044 start, end, false, &_igvn);
1045 } else {
1046 // Non-constant start, unrounded non-constant end.
1047 // (Nobody zeroes a random midsection of an array using this routine.)
1048 ShouldNotReachHere(); // fix caller
1049 }
1050
1051 // Done.
1052 merge_mem->set_memory_at(alias_idx, mem);
1053 }
1054
1055 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem,
1056 const TypePtr* adr_type,
1057 BasicType basic_elem_type,
1058 Node* src, Node* src_offset,
1059 Node* dest, Node* dest_offset,
1060 Node* dest_size, bool dest_uninitialized) {
1061 // See if there is an advantage from block transfer.
1062 int scale = exact_log2(type2aelembytes(basic_elem_type));
1063 if (scale >= LogBytesPerLong)
1138 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1139 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1140 "slow_arraycopy", TypePtr::BOTTOM);
1141
1142 call->init_req(TypeFunc::Control, *ctrl);
1143 call->init_req(TypeFunc::I_O , *io);
1144 call->init_req(TypeFunc::Memory , mem);
1145 call->init_req(TypeFunc::ReturnAdr, top());
1146 call->init_req(TypeFunc::FramePtr, top());
1147 call->init_req(TypeFunc::Parms+0, src);
1148 call->init_req(TypeFunc::Parms+1, src_offset);
1149 call->init_req(TypeFunc::Parms+2, dest);
1150 call->init_req(TypeFunc::Parms+3, dest_offset);
1151 call->init_req(TypeFunc::Parms+4, copy_length);
1152 call->copy_call_debug_info(&_igvn, ac);
1153
1154 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1155 _igvn.replace_node(ac, call);
1156 transform_later(call);
1157
1158 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1159 *ctrl = _callprojs->fallthrough_catchproj->clone();
1160 transform_later(*ctrl);
1161
1162 Node* m = _callprojs->fallthrough_memproj->clone();
1163 transform_later(m);
1164
1165 uint alias_idx = C->get_alias_index(adr_type);
1166 MergeMemNode* out_mem;
1167 if (alias_idx != Compile::AliasIdxBot) {
1168 out_mem = MergeMemNode::make(mem);
1169 out_mem->set_memory_at(alias_idx, m);
1170 } else {
1171 out_mem = MergeMemNode::make(m);
1172 }
1173 transform_later(out_mem);
1174
1175 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1176 // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1177 if (_callprojs->fallthrough_ioproj != nullptr) {
1178 *io = _callprojs->fallthrough_ioproj->clone();
1179 transform_later(*io);
1180 } else {
1181 *io = nullptr;
1182 }
1183
1184 return out_mem;
1185 }
1186
1187 // Helper function; generates code for cases requiring runtime checks.
1188 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1189 const TypePtr* adr_type,
1190 Node* dest_elem_klass,
1191 Node* src, Node* src_offset,
1192 Node* dest, Node* dest_offset,
1193 Node* copy_length, bool dest_uninitialized) {
1194 if ((*ctrl)->is_top()) return nullptr;
1195
1196 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1197 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1198 return nullptr;
1290
1291 // Connecting remaining edges for exit_block coming from stub_block.
1292 if (exit_block) {
1293 exit_block->init_req(2, *ctrl);
1294
1295 // Memory edge corresponding to stub_region.
1296 result_memory->init_req(2, *mem);
1297
1298 uint alias_idx = C->get_alias_index(adr_type);
1299 if (alias_idx != Compile::AliasIdxBot) {
1300 *mem = MergeMemNode::make(*mem);
1301 (*mem)->set_memory_at(alias_idx, result_memory);
1302 } else {
1303 *mem = MergeMemNode::make(result_memory);
1304 }
1305 transform_later(*mem);
1306 *ctrl = exit_block;
1307 }
1308 }
1309
1310 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1311 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1312 Node*& dest_length) {
1313 #ifdef ASSERT
1314 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1315 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1316 bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1317 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1318 #endif
1319 int elem_size = top_dest->flat_elem_size();
1320 if (elem_size >= 8) {
1321 if (elem_size > 8) {
1322 // treat as array of long but scale length, src offset and dest offset
1323 assert((elem_size % 8) == 0, "not a power of 2?");
1324 int factor = elem_size / 8;
1325 length = transform_later(new MulINode(length, intcon(factor)));
1326 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1327 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1328 if (dest_length != nullptr) {
1329 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1330 }
1331 elem_size = 8;
1332 }
1333 dest_elem = T_LONG;
1334 } else if (elem_size == 4) {
1335 dest_elem = T_INT;
1336 } else if (elem_size == 2) {
1337 dest_elem = T_CHAR;
1338 } else if (elem_size == 1) {
1339 dest_elem = T_BYTE;
1340 } else {
1341 ShouldNotReachHere();
1342 }
1343 return TypeRawPtr::BOTTOM;
1344 }
1345
1346 #undef XTOP
1347
1348 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1349 Node* ctrl = ac->in(TypeFunc::Control);
1350 Node* io = ac->in(TypeFunc::I_O);
1351 Node* src = ac->in(ArrayCopyNode::Src);
1352 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1353 Node* dest = ac->in(ArrayCopyNode::Dest);
1354 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1355 Node* length = ac->in(ArrayCopyNode::Length);
1356 MergeMemNode* merge_mem = nullptr;
1357
1358 if (ac->is_clonebasic()) {
1359 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1360 bs->clone_at_expansion(this, ac);
1361 return;
1362 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1363 const Type* src_type = _igvn.type(src);
1364 const Type* dest_type = _igvn.type(dest);
1365 const TypeAryPtr* top_src = src_type->isa_aryptr();
1366 // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1367 // Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1368 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1369 BasicType dest_elem = T_OBJECT;
1370 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1371 dest_elem = top_dest->elem()->array_element_basic_type();
1372 }
1373 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1374
1375 if (top_src != nullptr && top_src->is_flat()) {
1376 // If src is flat, dest is guaranteed to be flat as well
1377 top_dest = top_src;
1378 }
1379
1380 AllocateArrayNode* alloc = nullptr;
1381 Node* dest_length = nullptr;
1382 if (ac->is_alloc_tightly_coupled()) {
1383 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1384 assert(alloc != nullptr, "expect alloc");
1385 dest_length = alloc->in(AllocateNode::ALength);
1386 }
1387
1388 Node* mem = ac->in(TypeFunc::Memory);
1389 const TypePtr* adr_type = nullptr;
1390 if (top_dest != nullptr && top_dest->is_flat()) {
1391 assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1392 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1393 // on all slices to prevent writes into the source from floating below the arraycopy.
1394 int mem_bar_alias_idx = Compile::AliasIdxBot;
1395 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1396 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1397 }
1398 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1399 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1400 } else {
1401 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1402 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1403 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1404 }
1405 if (ac->_src_type != ac->_dest_type) {
1406 adr_type = TypeRawPtr::BOTTOM;
1407 }
1408 }
1409 merge_mem = MergeMemNode::make(mem);
1410 transform_later(merge_mem);
1411
1412 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1413 adr_type, dest_elem,
1414 src, src_offset, dest, dest_offset, length,
1415 dest_length,
1416 true, ac->has_negative_length_guard());
1417
1418 return;
1419 }
1420
1421 AllocateArrayNode* alloc = nullptr;
1422 if (ac->is_alloc_tightly_coupled()) {
1423 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1424 assert(alloc != nullptr, "expect alloc");
1425 }
1426
1427 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1428
1429 // Compile time checks. If any of these checks cannot be verified at compile time,
1430 // we do not make a fast path for this call. Instead, we let the call remain as it
1431 // is. The checks we choose to mandate at compile time are:
1432 //
1433 // (1) src and dest are arrays.
1434 const Type* src_type = src->Value(&_igvn);
1435 const Type* dest_type = dest->Value(&_igvn);
1436 const TypeAryPtr* top_src = src_type->isa_aryptr();
1437 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1438
1439 BasicType src_elem = T_CONFLICT;
1440 BasicType dest_elem = T_CONFLICT;
1441
1442 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1443 src_elem = top_src->elem()->array_element_basic_type();
1444 }
1445 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1446 dest_elem = top_dest->elem()->array_element_basic_type();
1447 }
1448 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1449 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1450
1451 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1452 src_elem = dest_elem;
1453 }
1454
1455 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1456 // Conservatively insert a memory barrier on all memory slices.
1457 // Do not let writes into the source float below the arraycopy.
1458 {
1459 Node* mem = ac->in(TypeFunc::Memory);
1460 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1461
1462 merge_mem = MergeMemNode::make(mem);
1463 transform_later(merge_mem);
1464 }
1465
1466 // Call StubRoutines::generic_arraycopy stub.
1467 generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1468 TypeRawPtr::BOTTOM, T_CONFLICT,
1469 src, src_offset, dest, dest_offset, length,
1470 nullptr,
1471 // If a negative length guard was generated for the ArrayCopyNode,
1472 // the length of the array can never be negative.
1473 false, ac->has_negative_length_guard());
1474 return;
1475 }
1476
1477 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1478
1479 // (2) src and dest arrays must have elements of the same BasicType
1480 // Figure out the size and type of the elements we will be copying.
1481 //
1482 // We have no stub to copy flat inline type arrays with oop
1483 // fields if we need to emit write barriers.
1484 //
1485 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1486 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1487 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1488 bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1489 // The component types are not the same or are not recognized. Punt.
1490 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1491 {
1492 Node* mem = ac->in(TypeFunc::Memory);
1493 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1494 }
1495
1496 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1497 if (_callprojs->fallthrough_ioproj != nullptr) {
1498 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1499 }
1500 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1501 return;
1502 }
1503
1504 //---------------------------------------------------------------------------
1505 // We will make a fast path for this call to arraycopy.
1506
1507 // We have the following tests left to perform:
1508 //
1509 // (3) src and dest must not be null.
1510 // (4) src_offset must not be negative.
1511 // (5) dest_offset must not be negative.
1512 // (6) length must not be negative.
1513 // (7) src_offset + length must not exceed length of src.
1514 // (8) dest_offset + length must not exceed length of dest.
1515 // (9) each element of an oop array must be assignable
1516
1517 Node* mem = ac->in(TypeFunc::Memory);
1518 if (top_dest->is_flat()) {
1519 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1520 // on all slices to prevent writes into the source from floating below the arraycopy.
1521 int mem_bar_alias_idx = Compile::AliasIdxBot;
1522 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1523 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1524 }
1525 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1526 }
1527 merge_mem = MergeMemNode::make(mem);
1528 transform_later(merge_mem);
1529
1530 RegionNode* slow_region = new RegionNode(1);
1531 transform_later(slow_region);
1532
1533 if (!ac->is_arraycopy_validated()) {
1534 // (3) operands must not be null
1535 // We currently perform our null checks with the null_check routine.
1536 // This means that the null exceptions will be reported in the caller
1537 // rather than (correctly) reported inside of the native arraycopy call.
1538 // This should be corrected, given time. We do our null check with the
1539 // stack pointer restored.
1540 // null checks done library_call.cpp
1541
1542 // (4) src_offset must not be negative.
1543 generate_negative_guard(&ctrl, src_offset, slow_region);
1544
1545 // (5) dest_offset must not be negative.
1546 generate_negative_guard(&ctrl, dest_offset, slow_region);
1547
1548 // (6) length must not be negative (moved to generate_arraycopy()).
1549 // generate_negative_guard(length, slow_region);
1550
1551 // (7) src_offset + length must not exceed length of src.
1552 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1553 assert(alen != nullptr, "need src len");
1554 generate_limit_guard(&ctrl,
1555 src_offset, length,
1556 alen,
1557 slow_region);
1558
1559 // (8) dest_offset + length must not exceed length of dest.
1560 alen = ac->in(ArrayCopyNode::DestLen);
1561 assert(alen != nullptr, "need dest len");
1562 generate_limit_guard(&ctrl,
1563 dest_offset, length,
1564 alen,
1565 slow_region);
1566
1567 // (9) each element of an oop array must be assignable
1568 // The generate_arraycopy subroutine checks this.
1569
1570 // TODO 8350865 This is too strong
1571 // We need to be careful here because 'adjust_for_flat_array' will adjust offsets/length etc. which then does not work anymore for the slow call to SharedRuntime::slow_arraycopy_C.
1572 if (!(top_src->is_flat() && top_dest->is_flat() && top_src->is_null_free() == top_dest->is_null_free())) {
1573 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1574 generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1575 generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1576 }
1577 }
1578
1579 // This is where the memory effects are placed:
1580 const TypePtr* adr_type = nullptr;
1581 Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1582
1583 if (top_src->is_flat() && top_dest->is_flat() &&
1584 top_src->is_null_free() == top_dest->is_null_free()) {
1585 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1586 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1587 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1588 } else {
1589 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1590 }
1591
1592 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1593 adr_type, dest_elem,
1594 src, src_offset, dest, dest_offset, length,
1595 dest_length,
1596 // If a negative length guard was generated for the ArrayCopyNode,
1597 // the length of the array can never be negative.
1598 false, ac->has_negative_length_guard(),
1599 slow_region);
1600 }
|