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 Node* original_dest = dest;
388 bool dest_needs_zeroing = false;
389 bool acopy_to_uninitialized = false;
390
391 // See if this is the initialization of a newly-allocated array.
392 // If so, we will take responsibility here for initializing it to zero.
393 // (Note: Because tightly_coupled_allocation performs checks on the
394 // out-edges of the dest, we need to avoid making derived pointers
395 // from it until we have checked its uses.)
396 if (ReduceBulkZeroing
397 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
398 && basic_elem_type != T_CONFLICT // avoid corner case
399 && !src->eqv_uncast(dest)
400 && alloc != nullptr
401 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
402 assert(ac->is_alloc_tightly_coupled(), "sanity");
403 // acopy to uninitialized tightly coupled allocations
404 // needs zeroing outside the copy range
405 // and the acopy itself will be to uninitialized memory
406 acopy_to_uninitialized = true;
407 if (alloc->maybe_set_complete(&_igvn)) {
408 // "You break it, you buy it."
409 InitializeNode* init = alloc->initialization();
410 assert(init->is_complete(), "we just did this");
411 init->set_complete_with_arraycopy();
412 assert(dest->is_CheckCastPP(), "sanity");
413 assert(dest->in(0)->in(0) == init, "dest pinned");
414 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
415 // From this point on, every exit path is responsible for
416 // initializing any non-copied parts of the object to zero.
417 // Also, if this flag is set we make sure that arraycopy interacts properly
418 // with G1, eliding pre-barriers. See CR 6627983.
419 dest_needs_zeroing = true;
420 } else {
421 // dest_need_zeroing = false;
422 }
423 } else {
424 // No zeroing elimination needed here.
425 alloc = nullptr;
426 acopy_to_uninitialized = false;
427 //original_dest = dest;
428 //dest_needs_zeroing = false;
429 }
430
431 uint alias_idx = C->get_alias_index(adr_type);
432
433 // Results are placed here:
434 enum { fast_path = 1, // normal void-returning assembly stub
435 checked_path = 2, // special assembly stub with cleanup
436 slow_call_path = 3, // something went wrong; call the VM
437 zero_path = 4, // bypass when length of copy is zero
438 bcopy_path = 5, // copy primitive array by 64-bit blocks
439 PATH_LIMIT = 6
469 checked_i_o = *io;
470 checked_mem = mem->memory_at(alias_idx);
471 checked_value = cv;
472 *ctrl = top();
473 }
474
475 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
476 if (not_pos != nullptr) {
477 Node* local_ctrl = not_pos, *local_io = *io;
478 MergeMemNode* local_mem = MergeMemNode::make(mem);
479 transform_later(local_mem);
480
481 // (6) length must not be negative.
482 if (!length_never_negative) {
483 generate_negative_guard(&local_ctrl, copy_length, slow_region);
484 }
485
486 // copy_length is 0.
487 if (dest_needs_zeroing) {
488 assert(!local_ctrl->is_top(), "no ctrl?");
489 Node* dest_length = alloc->in(AllocateNode::ALength);
490 if (copy_length->eqv_uncast(dest_length)
491 || _igvn.find_int_con(dest_length, 1) <= 0) {
492 // There is no zeroing to do. No need for a secondary raw memory barrier.
493 } else {
494 // Clear the whole thing since there are no source elements to copy.
495 generate_clear_array(local_ctrl, local_mem,
496 adr_type, dest, basic_elem_type,
497 intcon(0), nullptr,
498 alloc->in(AllocateNode::AllocSize));
499 // Use a secondary InitializeNode as raw memory barrier.
500 // Currently it is needed only on this path since other
501 // paths have stub or runtime calls as raw memory barriers.
502 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
503 Compile::AliasIdxRaw,
504 top());
505 transform_later(mb);
506 mb->set_req(TypeFunc::Control,local_ctrl);
507 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
508 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
509 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
510
511 InitializeNode* init = mb->as_Initialize();
512 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
513 }
514 }
515
516 // Present the results of the fast call.
517 result_region->init_req(zero_path, local_ctrl);
518 result_i_o ->init_req(zero_path, local_io);
519 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
520 }
521
522 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
523 // We have to initialize the *uncopied* part of the array to zero.
524 // The copy destination is the slice dest[off..off+len]. The other slices
525 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
526 Node* dest_size = alloc->in(AllocateNode::AllocSize);
527 Node* dest_length = alloc->in(AllocateNode::ALength);
528 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
529
530 // If there is a head section that needs zeroing, do it now.
531 if (_igvn.find_int_con(dest_offset, -1) != 0) {
532 generate_clear_array(*ctrl, mem,
533 adr_type, dest, basic_elem_type,
534 intcon(0), dest_offset,
535 nullptr);
536 }
537
538 // Next, perform a dynamic check on the tail length.
539 // It is often zero, and we can win big if we prove this.
540 // There are two wins: Avoid generating the ClearArray
541 // with its attendant messy index arithmetic, and upgrade
542 // the copy to a more hardware-friendly word size of 64 bits.
543 Node* tail_ctl = nullptr;
544 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
545 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
546 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
547 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
548 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
549 }
550
551 // At this point, let's assume there is no tail.
552 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
553 // There is no tail. Try an upgrade to a 64-bit copy.
562 src, src_offset, dest, dest_offset,
563 dest_size, acopy_to_uninitialized);
564 if (didit) {
565 // Present the results of the block-copying fast call.
566 result_region->init_req(bcopy_path, local_ctrl);
567 result_i_o ->init_req(bcopy_path, local_io);
568 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
569 }
570 }
571 if (didit) {
572 *ctrl = top(); // no regular fast path
573 }
574 }
575
576 // Clear the tail, if any.
577 if (tail_ctl != nullptr) {
578 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
579 *ctrl = tail_ctl;
580 if (notail_ctl == nullptr) {
581 generate_clear_array(*ctrl, mem,
582 adr_type, dest, basic_elem_type,
583 dest_tail, nullptr,
584 dest_size);
585 } else {
586 // Make a local merge.
587 Node* done_ctl = transform_later(new RegionNode(3));
588 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
589 done_ctl->init_req(1, notail_ctl);
590 done_mem->init_req(1, mem->memory_at(alias_idx));
591 generate_clear_array(*ctrl, mem,
592 adr_type, dest, basic_elem_type,
593 dest_tail, nullptr,
594 dest_size);
595 done_ctl->init_req(2, *ctrl);
596 done_mem->init_req(2, mem->memory_at(alias_idx));
597 *ctrl = done_ctl;
598 mem->set_memory_at(alias_idx, done_mem);
599 }
600 }
601 }
602
603 BasicType copy_type = basic_elem_type;
604 assert(basic_elem_type != T_ARRAY, "caller must fix this");
605 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
606 // If src and dest have compatible element types, we can copy bits.
607 // Types S[] and D[] are compatible if D is a supertype of S.
608 //
609 // If they are not, we will use checked_oop_disjoint_arraycopy,
610 // which performs a fast optimistic per-oop check, and backs off
611 // further to JVM_ArrayCopy on the first per-oop check that fails.
612 // (Actually, we don't move raw bits only; the GC requires card marks.)
749 Node* length_minus = new SubINode(copy_length, slow_offset);
750 transform_later(length_minus);
751
752 // Tweak the node variables to adjust the code produced below:
753 src_offset = src_off_plus;
754 dest_offset = dest_off_plus;
755 copy_length = length_minus;
756 }
757 }
758 *ctrl = slow_control;
759 if (!(*ctrl)->is_top()) {
760 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
761 MergeMemNode* local_mem = MergeMemNode::make(mem);
762 transform_later(local_mem);
763
764 // Generate the slow path, if needed.
765 local_mem->set_memory_at(alias_idx, slow_mem);
766
767 if (dest_needs_zeroing) {
768 generate_clear_array(local_ctrl, local_mem,
769 adr_type, dest, basic_elem_type,
770 intcon(0), nullptr,
771 alloc->in(AllocateNode::AllocSize));
772 }
773
774 local_mem = generate_slow_arraycopy(ac,
775 &local_ctrl, local_mem, &local_io,
776 adr_type,
777 src, src_offset, dest, dest_offset,
778 copy_length, /*dest_uninitialized*/false);
779
780 result_region->init_req(slow_call_path, local_ctrl);
781 result_i_o ->init_req(slow_call_path, local_io);
782 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
783 } else {
784 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
785 // projections were not extracted
786 }
787
788 // Remove unused edges.
789 for (uint i = 1; i < result_region->req(); i++) {
818 // a subsequent store that would make this object accessible by
819 // other threads.
820 assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
821 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
822 } else {
823 int mem_bar_alias_idx = Compile::AliasIdxBot;
824 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
825 // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
826 // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
827 // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
828 // broken graph.
829 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
830 }
831 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
832 }
833
834 assert((*ctrl)->is_Proj(), "MemBar control projection");
835 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
836 (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
837
838 _igvn.replace_node(_callprojs.fallthrough_memproj, out_mem);
839 if (_callprojs.fallthrough_ioproj != nullptr) {
840 _igvn.replace_node(_callprojs.fallthrough_ioproj, *io);
841 }
842 _igvn.replace_node(_callprojs.fallthrough_catchproj, *ctrl);
843
844 #ifdef ASSERT
845 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
846 if (dest_t->is_known_instance()) {
847 ArrayCopyNode* ac = nullptr;
848 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
849 assert(ac == nullptr, "no arraycopy anymore");
850 }
851 #endif
852
853 return out_mem;
854 }
855
856 // Helper for initialization of arrays, creating a ClearArray.
857 // It writes zero bits in [start..end), within the body of an array object.
858 // The memory effects are all chained onto the 'adr_type' alias category.
859 //
860 // Since the object is otherwise uninitialized, we are free
861 // to put a little "slop" around the edges of the cleared area,
862 // as long as it does not go back into the array's header,
863 // or beyond the array end within the heap.
864 //
865 // The lower edge can be rounded down to the nearest jint and the
866 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
867 //
868 // Arguments:
869 // adr_type memory slice where writes are generated
870 // dest oop of the destination array
871 // basic_elem_type element type of the destination
872 // slice_idx array index of first element to store
873 // slice_len number of elements to store (or null)
874 // dest_size total size in bytes of the array object
875 //
876 // Exactly one of slice_len or dest_size must be non-null.
877 // If dest_size is non-null, zeroing extends to the end of the object.
878 // If slice_len is non-null, the slice_idx value must be a constant.
879 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
880 const TypePtr* adr_type,
881 Node* dest,
882 BasicType basic_elem_type,
883 Node* slice_idx,
884 Node* slice_len,
885 Node* dest_size) {
886 // one or the other but not both of slice_len and dest_size:
887 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
888 if (slice_len == nullptr) slice_len = top();
889 if (dest_size == nullptr) dest_size = top();
890
891 uint alias_idx = C->get_alias_index(adr_type);
892
893 // operate on this memory slice:
894 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
895
896 // scaling and rounding of indexes:
897 int scale = exact_log2(type2aelembytes(basic_elem_type));
898 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
899 int clear_low = (-1 << scale) & (BytesPerInt - 1);
900 int bump_bit = (-1 << scale) & BytesPerInt;
901
902 // determine constant starts and ends
903 const intptr_t BIG_NEG = -128;
904 assert(BIG_NEG + 2*abase < 0, "neg enough");
905 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
906 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
907 if (slice_len_con == 0) {
908 return; // nothing to do here
909 }
910 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
911 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
912 if (slice_idx_con >= 0 && slice_len_con >= 0) {
913 assert(end_con < 0, "not two cons");
914 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
915 BytesPerLong);
916 }
917
918 if (start_con >= 0 && end_con >= 0) {
919 // Constant start and end. Simple.
920 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
921 start_con, end_con, &_igvn);
922 } else if (start_con >= 0 && dest_size != top()) {
923 // Constant start, pre-rounded end after the tail of the array.
924 Node* end = dest_size;
925 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
926 start_con, end, &_igvn);
927 } else if (start_con >= 0 && slice_len != top()) {
928 // Constant start, non-constant end. End needs rounding up.
929 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
930 intptr_t end_base = abase + (slice_idx_con << scale);
931 int end_round = (-1 << scale) & (BytesPerLong - 1);
932 Node* end = ConvI2X(slice_len);
933 if (scale != 0)
934 end = transform_later(new LShiftXNode(end, intcon(scale) ));
935 end_base += end_round;
936 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
937 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
938 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
939 start_con, end, &_igvn);
940 } else if (start_con < 0 && dest_size != top()) {
941 // Non-constant start, pre-rounded end after the tail of the array.
942 // This is almost certainly a "round-to-end" operation.
943 Node* start = slice_idx;
944 start = ConvI2X(start);
945 if (scale != 0)
946 start = transform_later(new LShiftXNode( start, intcon(scale) ));
947 start = transform_later(new AddXNode(start, MakeConX(abase)) );
948 if ((bump_bit | clear_low) != 0) {
949 int to_clear = (bump_bit | clear_low);
950 // Align up mod 8, then store a jint zero unconditionally
951 // just before the mod-8 boundary.
952 if (((abase + bump_bit) & ~to_clear) - bump_bit
953 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
954 bump_bit = 0;
955 assert((abase & to_clear) == 0, "array base must be long-aligned");
956 } else {
957 // Bump 'start' up to (or past) the next jint boundary:
958 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
959 assert((abase & clear_low) == 0, "array base must be int-aligned");
960 }
961 // Round bumped 'start' down to jlong boundary in body of array.
962 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
963 if (bump_bit != 0) {
964 // Store a zero to the immediately preceding jint:
965 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
966 Node* p1 = basic_plus_adr(dest, x1);
967 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
968 mem = transform_later(mem);
969 }
970 }
971 Node* end = dest_size; // pre-rounded
972 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
973 start, end, &_igvn);
974 } else {
975 // Non-constant start, unrounded non-constant end.
976 // (Nobody zeroes a random midsection of an array using this routine.)
977 ShouldNotReachHere(); // fix caller
978 }
979
980 // Done.
981 merge_mem->set_memory_at(alias_idx, mem);
982 }
983
984 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
985 const TypePtr* adr_type,
986 BasicType basic_elem_type,
987 AllocateNode* alloc,
988 Node* src, Node* src_offset,
989 Node* dest, Node* dest_offset,
990 Node* dest_size, bool dest_uninitialized) {
991 // See if there is an advantage from block transfer.
992 int scale = exact_log2(type2aelembytes(basic_elem_type));
1068 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1069 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1070 "slow_arraycopy", TypePtr::BOTTOM);
1071
1072 call->init_req(TypeFunc::Control, *ctrl);
1073 call->init_req(TypeFunc::I_O , *io);
1074 call->init_req(TypeFunc::Memory , mem);
1075 call->init_req(TypeFunc::ReturnAdr, top());
1076 call->init_req(TypeFunc::FramePtr, top());
1077 call->init_req(TypeFunc::Parms+0, src);
1078 call->init_req(TypeFunc::Parms+1, src_offset);
1079 call->init_req(TypeFunc::Parms+2, dest);
1080 call->init_req(TypeFunc::Parms+3, dest_offset);
1081 call->init_req(TypeFunc::Parms+4, copy_length);
1082 call->copy_call_debug_info(&_igvn, ac);
1083
1084 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1085 _igvn.replace_node(ac, call);
1086 transform_later(call);
1087
1088 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1089 *ctrl = _callprojs.fallthrough_catchproj->clone();
1090 transform_later(*ctrl);
1091
1092 Node* m = _callprojs.fallthrough_memproj->clone();
1093 transform_later(m);
1094
1095 uint alias_idx = C->get_alias_index(adr_type);
1096 MergeMemNode* out_mem;
1097 if (alias_idx != Compile::AliasIdxBot) {
1098 out_mem = MergeMemNode::make(mem);
1099 out_mem->set_memory_at(alias_idx, m);
1100 } else {
1101 out_mem = MergeMemNode::make(m);
1102 }
1103 transform_later(out_mem);
1104
1105 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1106 // could be null. Skip clone and update null fallthrough_ioproj.
1107 if (_callprojs.fallthrough_ioproj != nullptr) {
1108 *io = _callprojs.fallthrough_ioproj->clone();
1109 transform_later(*io);
1110 } else {
1111 *io = nullptr;
1112 }
1113
1114 return out_mem;
1115 }
1116
1117 // Helper function; generates code for cases requiring runtime checks.
1118 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1119 const TypePtr* adr_type,
1120 Node* dest_elem_klass,
1121 Node* src, Node* src_offset,
1122 Node* dest, Node* dest_offset,
1123 Node* copy_length, bool dest_uninitialized) {
1124 if ((*ctrl)->is_top()) return nullptr;
1125
1126 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1127 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1128 return nullptr;
1220
1221 // Connecting remaining edges for exit_block coming from stub_block.
1222 if (exit_block) {
1223 exit_block->init_req(2, *ctrl);
1224
1225 // Memory edge corresponding to stub_region.
1226 result_memory->init_req(2, *mem);
1227
1228 uint alias_idx = C->get_alias_index(adr_type);
1229 if (alias_idx != Compile::AliasIdxBot) {
1230 *mem = MergeMemNode::make(*mem);
1231 (*mem)->set_memory_at(alias_idx, result_memory);
1232 } else {
1233 *mem = MergeMemNode::make(result_memory);
1234 }
1235 transform_later(*mem);
1236 *ctrl = exit_block;
1237 }
1238 }
1239
1240 #undef XTOP
1241
1242 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1243 Node* ctrl = ac->in(TypeFunc::Control);
1244 Node* io = ac->in(TypeFunc::I_O);
1245 Node* src = ac->in(ArrayCopyNode::Src);
1246 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1247 Node* dest = ac->in(ArrayCopyNode::Dest);
1248 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1249 Node* length = ac->in(ArrayCopyNode::Length);
1250 MergeMemNode* merge_mem = nullptr;
1251
1252 if (ac->is_clonebasic()) {
1253 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1254 bs->clone_at_expansion(this, ac);
1255 return;
1256 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1257 Node* mem = ac->in(TypeFunc::Memory);
1258 merge_mem = MergeMemNode::make(mem);
1259 transform_later(merge_mem);
1260
1261 AllocateArrayNode* alloc = nullptr;
1262 if (ac->is_alloc_tightly_coupled()) {
1263 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1264 assert(alloc != nullptr, "expect alloc");
1265 }
1266
1267 const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1268 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1269 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1270 }
1271 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1272 adr_type, T_OBJECT,
1273 src, src_offset, dest, dest_offset, length,
1274 true, ac->has_negative_length_guard());
1275
1276 return;
1277 }
1278
1279 AllocateArrayNode* alloc = nullptr;
1280 if (ac->is_alloc_tightly_coupled()) {
1281 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1282 assert(alloc != nullptr, "expect alloc");
1283 }
1284
1285 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1286
1287 // Compile time checks. If any of these checks cannot be verified at compile time,
1288 // we do not make a fast path for this call. Instead, we let the call remain as it
1289 // is. The checks we choose to mandate at compile time are:
1290 //
1291 // (1) src and dest are arrays.
1292 const Type* src_type = src->Value(&_igvn);
1293 const Type* dest_type = dest->Value(&_igvn);
1294 const TypeAryPtr* top_src = src_type->isa_aryptr();
1295 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1296
1297 BasicType src_elem = T_CONFLICT;
1298 BasicType dest_elem = T_CONFLICT;
1299
1300 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1301 src_elem = top_src->elem()->array_element_basic_type();
1302 }
1303 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1304 dest_elem = top_dest->elem()->array_element_basic_type();
1305 }
1306 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1307 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1308
1309 if (ac->is_arraycopy_validated() &&
1310 dest_elem != T_CONFLICT &&
1311 src_elem == T_CONFLICT) {
1312 src_elem = dest_elem;
1313 }
1314
1315 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1316 // Conservatively insert a memory barrier on all memory slices.
1317 // Do not let writes into the source float below the arraycopy.
1318 {
1319 Node* mem = ac->in(TypeFunc::Memory);
1320 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1321
1322 merge_mem = MergeMemNode::make(mem);
1323 transform_later(merge_mem);
1324 }
1325
1326 // Call StubRoutines::generic_arraycopy stub.
1327 Node* mem = generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1328 TypeRawPtr::BOTTOM, T_CONFLICT,
1329 src, src_offset, dest, dest_offset, length,
1330 // If a negative length guard was generated for the ArrayCopyNode,
1331 // the length of the array can never be negative.
1332 false, ac->has_negative_length_guard());
1333 return;
1334 }
1335
1336 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1337
1338 // (2) src and dest arrays must have elements of the same BasicType
1339 // Figure out the size and type of the elements we will be copying.
1340 if (src_elem != dest_elem || dest_elem == T_VOID) {
1341 // The component types are not the same or are not recognized. Punt.
1342 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1343 {
1344 Node* mem = ac->in(TypeFunc::Memory);
1345 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1346 }
1347
1348 _igvn.replace_node(_callprojs.fallthrough_memproj, merge_mem);
1349 if (_callprojs.fallthrough_ioproj != nullptr) {
1350 _igvn.replace_node(_callprojs.fallthrough_ioproj, io);
1351 }
1352 _igvn.replace_node(_callprojs.fallthrough_catchproj, ctrl);
1353 return;
1354 }
1355
1356 //---------------------------------------------------------------------------
1357 // We will make a fast path for this call to arraycopy.
1358
1359 // We have the following tests left to perform:
1360 //
1361 // (3) src and dest must not be null.
1362 // (4) src_offset must not be negative.
1363 // (5) dest_offset must not be negative.
1364 // (6) length must not be negative.
1365 // (7) src_offset + length must not exceed length of src.
1366 // (8) dest_offset + length must not exceed length of dest.
1367 // (9) each element of an oop array must be assignable
1368
1369 {
1370 Node* mem = ac->in(TypeFunc::Memory);
1371 merge_mem = MergeMemNode::make(mem);
1372 transform_later(merge_mem);
1373 }
1374
1375 RegionNode* slow_region = new RegionNode(1);
1376 transform_later(slow_region);
1377
1378 if (!ac->is_arraycopy_validated()) {
1379 // (3) operands must not be null
1380 // We currently perform our null checks with the null_check routine.
1381 // This means that the null exceptions will be reported in the caller
1382 // rather than (correctly) reported inside of the native arraycopy call.
1383 // This should be corrected, given time. We do our null check with the
1384 // stack pointer restored.
1385 // null checks done library_call.cpp
1386
1387 // (4) src_offset must not be negative.
1388 generate_negative_guard(&ctrl, src_offset, slow_region);
1389
1390 // (5) dest_offset must not be negative.
1391 generate_negative_guard(&ctrl, dest_offset, slow_region);
1392
1393 // (6) length must not be negative (moved to generate_arraycopy()).
1394 // generate_negative_guard(length, slow_region);
1395
1396 // (7) src_offset + length must not exceed length of src.
1397 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1398 assert(alen != nullptr, "need src len");
1399 generate_limit_guard(&ctrl,
1400 src_offset, length,
1401 alen,
1402 slow_region);
1403
1404 // (8) dest_offset + length must not exceed length of dest.
1405 alen = ac->in(ArrayCopyNode::DestLen);
1406 assert(alen != nullptr, "need dest len");
1407 generate_limit_guard(&ctrl,
1408 dest_offset, length,
1409 alen,
1410 slow_region);
1411
1412 // (9) each element of an oop array must be assignable
1413 // The generate_arraycopy subroutine checks this.
1414 }
1415 // This is where the memory effects are placed:
1416 const TypePtr* adr_type = nullptr;
1417 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1418 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1419 } else {
1420 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1421 }
1422
1423 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1424 adr_type, dest_elem,
1425 src, src_offset, dest, dest_offset, length,
1426 // If a negative length guard was generated for the ArrayCopyNode,
1427 // the length of the array can never be negative.
1428 false, ac->has_negative_length_guard(), slow_region);
1429 }
|
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(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(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 Node* original_dest = dest;
442 bool dest_needs_zeroing = false;
443 bool acopy_to_uninitialized = false;
444 Node* init_value = nullptr;
445 Node* raw_init_value = nullptr;
446
447 // See if this is the initialization of a newly-allocated array.
448 // If so, we will take responsibility here for initializing it to zero.
449 // (Note: Because tightly_coupled_allocation performs checks on the
450 // out-edges of the dest, we need to avoid making derived pointers
451 // from it until we have checked its uses.)
452 if (ReduceBulkZeroing
453 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
454 && basic_elem_type != T_CONFLICT // avoid corner case
455 && !src->eqv_uncast(dest)
456 && alloc != nullptr
457 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
458 assert(ac->is_alloc_tightly_coupled(), "sanity");
459 // acopy to uninitialized tightly coupled allocations
460 // needs zeroing outside the copy range
461 // and the acopy itself will be to uninitialized memory
462 acopy_to_uninitialized = true;
463 if (alloc->maybe_set_complete(&_igvn)) {
464 // "You break it, you buy it."
465 InitializeNode* init = alloc->initialization();
466 assert(init->is_complete(), "we just did this");
467 init->set_complete_with_arraycopy();
468 assert(dest->is_CheckCastPP(), "sanity");
469 assert(dest->in(0)->in(0) == init, "dest pinned");
470 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
471 // From this point on, every exit path is responsible for
472 // initializing any non-copied parts of the object to zero.
473 // Also, if this flag is set we make sure that arraycopy interacts properly
474 // with G1, eliding pre-barriers. See CR 6627983.
475 dest_needs_zeroing = true;
476 init_value = alloc->in(AllocateNode::InitValue);
477 raw_init_value = alloc->in(AllocateNode::RawInitValue);
478 } else {
479 // dest_need_zeroing = false;
480 }
481 } else {
482 // No zeroing elimination needed here.
483 alloc = nullptr;
484 acopy_to_uninitialized = false;
485 //original_dest = dest;
486 //dest_needs_zeroing = false;
487 }
488
489 uint alias_idx = C->get_alias_index(adr_type);
490
491 // Results are placed here:
492 enum { fast_path = 1, // normal void-returning assembly stub
493 checked_path = 2, // special assembly stub with cleanup
494 slow_call_path = 3, // something went wrong; call the VM
495 zero_path = 4, // bypass when length of copy is zero
496 bcopy_path = 5, // copy primitive array by 64-bit blocks
497 PATH_LIMIT = 6
527 checked_i_o = *io;
528 checked_mem = mem->memory_at(alias_idx);
529 checked_value = cv;
530 *ctrl = top();
531 }
532
533 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
534 if (not_pos != nullptr) {
535 Node* local_ctrl = not_pos, *local_io = *io;
536 MergeMemNode* local_mem = MergeMemNode::make(mem);
537 transform_later(local_mem);
538
539 // (6) length must not be negative.
540 if (!length_never_negative) {
541 generate_negative_guard(&local_ctrl, copy_length, slow_region);
542 }
543
544 // copy_length is 0.
545 if (dest_needs_zeroing) {
546 assert(!local_ctrl->is_top(), "no ctrl?");
547 if (copy_length->eqv_uncast(dest_length)
548 || _igvn.find_int_con(dest_length, 1) <= 0) {
549 // There is no zeroing to do. No need for a secondary raw memory barrier.
550 } else {
551 // Clear the whole thing since there are no source elements to copy.
552 generate_clear_array(local_ctrl, local_mem,
553 adr_type, dest,
554 init_value, raw_init_value,
555 basic_elem_type,
556 intcon(0), nullptr,
557 alloc->in(AllocateNode::AllocSize));
558 // Use a secondary InitializeNode as raw memory barrier.
559 // Currently it is needed only on this path since other
560 // paths have stub or runtime calls as raw memory barriers.
561 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
562 Compile::AliasIdxRaw,
563 top());
564 transform_later(mb);
565 mb->set_req(TypeFunc::Control,local_ctrl);
566 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
567 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
568 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
569
570 InitializeNode* init = mb->as_Initialize();
571 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
572 }
573 }
574
575 // Present the results of the fast call.
576 result_region->init_req(zero_path, local_ctrl);
577 result_i_o ->init_req(zero_path, local_io);
578 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
579 }
580
581 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
582 // We have to initialize the *uncopied* part of the array to zero.
583 // The copy destination is the slice dest[off..off+len]. The other slices
584 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
585 Node* dest_size = alloc->in(AllocateNode::AllocSize);
586 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
587
588 // If there is a head section that needs zeroing, do it now.
589 if (_igvn.find_int_con(dest_offset, -1) != 0) {
590 generate_clear_array(*ctrl, mem,
591 adr_type, dest,
592 init_value, raw_init_value,
593 basic_elem_type,
594 intcon(0), dest_offset,
595 nullptr);
596 }
597
598 // Next, perform a dynamic check on the tail length.
599 // It is often zero, and we can win big if we prove this.
600 // There are two wins: Avoid generating the ClearArray
601 // with its attendant messy index arithmetic, and upgrade
602 // the copy to a more hardware-friendly word size of 64 bits.
603 Node* tail_ctl = nullptr;
604 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
605 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
606 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
607 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
608 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
609 }
610
611 // At this point, let's assume there is no tail.
612 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
613 // There is no tail. Try an upgrade to a 64-bit copy.
622 src, src_offset, dest, dest_offset,
623 dest_size, acopy_to_uninitialized);
624 if (didit) {
625 // Present the results of the block-copying fast call.
626 result_region->init_req(bcopy_path, local_ctrl);
627 result_i_o ->init_req(bcopy_path, local_io);
628 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
629 }
630 }
631 if (didit) {
632 *ctrl = top(); // no regular fast path
633 }
634 }
635
636 // Clear the tail, if any.
637 if (tail_ctl != nullptr) {
638 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
639 *ctrl = tail_ctl;
640 if (notail_ctl == nullptr) {
641 generate_clear_array(*ctrl, mem,
642 adr_type, dest,
643 init_value, raw_init_value,
644 basic_elem_type,
645 dest_tail, nullptr,
646 dest_size);
647 } else {
648 // Make a local merge.
649 Node* done_ctl = transform_later(new RegionNode(3));
650 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
651 done_ctl->init_req(1, notail_ctl);
652 done_mem->init_req(1, mem->memory_at(alias_idx));
653 generate_clear_array(*ctrl, mem,
654 adr_type, dest,
655 init_value, raw_init_value,
656 basic_elem_type,
657 dest_tail, nullptr,
658 dest_size);
659 done_ctl->init_req(2, *ctrl);
660 done_mem->init_req(2, mem->memory_at(alias_idx));
661 *ctrl = done_ctl;
662 mem->set_memory_at(alias_idx, done_mem);
663 }
664 }
665 }
666
667 BasicType copy_type = basic_elem_type;
668 assert(basic_elem_type != T_ARRAY, "caller must fix this");
669 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
670 // If src and dest have compatible element types, we can copy bits.
671 // Types S[] and D[] are compatible if D is a supertype of S.
672 //
673 // If they are not, we will use checked_oop_disjoint_arraycopy,
674 // which performs a fast optimistic per-oop check, and backs off
675 // further to JVM_ArrayCopy on the first per-oop check that fails.
676 // (Actually, we don't move raw bits only; the GC requires card marks.)
813 Node* length_minus = new SubINode(copy_length, slow_offset);
814 transform_later(length_minus);
815
816 // Tweak the node variables to adjust the code produced below:
817 src_offset = src_off_plus;
818 dest_offset = dest_off_plus;
819 copy_length = length_minus;
820 }
821 }
822 *ctrl = slow_control;
823 if (!(*ctrl)->is_top()) {
824 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
825 MergeMemNode* local_mem = MergeMemNode::make(mem);
826 transform_later(local_mem);
827
828 // Generate the slow path, if needed.
829 local_mem->set_memory_at(alias_idx, slow_mem);
830
831 if (dest_needs_zeroing) {
832 generate_clear_array(local_ctrl, local_mem,
833 adr_type, dest,
834 init_value, raw_init_value,
835 basic_elem_type,
836 intcon(0), nullptr,
837 alloc->in(AllocateNode::AllocSize));
838 }
839
840 local_mem = generate_slow_arraycopy(ac,
841 &local_ctrl, local_mem, &local_io,
842 adr_type,
843 src, src_offset, dest, dest_offset,
844 copy_length, /*dest_uninitialized*/false);
845
846 result_region->init_req(slow_call_path, local_ctrl);
847 result_i_o ->init_req(slow_call_path, local_io);
848 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
849 } else {
850 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
851 // projections were not extracted
852 }
853
854 // Remove unused edges.
855 for (uint i = 1; i < result_region->req(); i++) {
884 // a subsequent store that would make this object accessible by
885 // other threads.
886 assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
887 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
888 } else {
889 int mem_bar_alias_idx = Compile::AliasIdxBot;
890 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
891 // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
892 // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
893 // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
894 // broken graph.
895 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
896 }
897 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
898 }
899
900 assert((*ctrl)->is_Proj(), "MemBar control projection");
901 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
902 (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
903
904 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
905 if (_callprojs->fallthrough_ioproj != nullptr) {
906 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
907 }
908 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
909
910 #ifdef ASSERT
911 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
912 if (dest_t->is_known_instance()) {
913 ArrayCopyNode* ac = nullptr;
914 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
915 assert(ac == nullptr, "no arraycopy anymore");
916 }
917 #endif
918
919 return out_mem;
920 }
921
922 // Helper for initialization of arrays, creating a ClearArray.
923 // It writes zero bits in [start..end), within the body of an array object.
924 // The memory effects are all chained onto the 'adr_type' alias category.
925 //
926 // Since the object is otherwise uninitialized, we are free
927 // to put a little "slop" around the edges of the cleared area,
928 // as long as it does not go back into the array's header,
929 // or beyond the array end within the heap.
930 //
931 // The lower edge can be rounded down to the nearest jint and the
932 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
933 //
934 // Arguments:
935 // adr_type memory slice where writes are generated
936 // dest oop of the destination array
937 // basic_elem_type element type of the destination
938 // slice_idx array index of first element to store
939 // slice_len number of elements to store (or null)
940 // dest_size total size in bytes of the array object
941 //
942 // Exactly one of slice_len or dest_size must be non-null.
943 // If dest_size is non-null, zeroing extends to the end of the object.
944 // If slice_len is non-null, the slice_idx value must be a constant.
945 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
946 const TypePtr* adr_type,
947 Node* dest,
948 Node* val,
949 Node* raw_val,
950 BasicType basic_elem_type,
951 Node* slice_idx,
952 Node* slice_len,
953 Node* dest_size) {
954 // one or the other but not both of slice_len and dest_size:
955 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
956 if (slice_len == nullptr) slice_len = top();
957 if (dest_size == nullptr) dest_size = top();
958
959 uint alias_idx = C->get_alias_index(adr_type);
960
961 // operate on this memory slice:
962 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
963
964 // scaling and rounding of indexes:
965 int scale = exact_log2(type2aelembytes(basic_elem_type));
966 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
967 int clear_low = (-1 << scale) & (BytesPerInt - 1);
968 int bump_bit = (-1 << scale) & BytesPerInt;
969
970 // determine constant starts and ends
971 const intptr_t BIG_NEG = -128;
972 assert(BIG_NEG + 2*abase < 0, "neg enough");
973 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
974 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
975 if (slice_len_con == 0) {
976 return; // nothing to do here
977 }
978 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
979 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
980 if (slice_idx_con >= 0 && slice_len_con >= 0) {
981 assert(end_con < 0, "not two cons");
982 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
983 BytesPerLong);
984 }
985
986 if (start_con >= 0 && end_con >= 0) {
987 // Constant start and end. Simple.
988 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
989 start_con, end_con, &_igvn);
990 } else if (start_con >= 0 && dest_size != top()) {
991 // Constant start, pre-rounded end after the tail of the array.
992 Node* end = dest_size;
993 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
994 start_con, end, &_igvn);
995 } else if (start_con >= 0 && slice_len != top()) {
996 // Constant start, non-constant end. End needs rounding up.
997 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
998 intptr_t end_base = abase + (slice_idx_con << scale);
999 int end_round = (-1 << scale) & (BytesPerLong - 1);
1000 Node* end = ConvI2X(slice_len);
1001 if (scale != 0)
1002 end = transform_later(new LShiftXNode(end, intcon(scale) ));
1003 end_base += end_round;
1004 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
1005 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
1006 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1007 start_con, end, &_igvn);
1008 } else if (start_con < 0 && dest_size != top()) {
1009 // Non-constant start, pre-rounded end after the tail of the array.
1010 // This is almost certainly a "round-to-end" operation.
1011 Node* start = slice_idx;
1012 start = ConvI2X(start);
1013 if (scale != 0)
1014 start = transform_later(new LShiftXNode( start, intcon(scale) ));
1015 start = transform_later(new AddXNode(start, MakeConX(abase)) );
1016 if ((bump_bit | clear_low) != 0) {
1017 int to_clear = (bump_bit | clear_low);
1018 // Align up mod 8, then store a jint zero unconditionally
1019 // just before the mod-8 boundary.
1020 if (((abase + bump_bit) & ~to_clear) - bump_bit
1021 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1022 bump_bit = 0;
1023 assert((abase & to_clear) == 0, "array base must be long-aligned");
1024 } else {
1025 // Bump 'start' up to (or past) the next jint boundary:
1026 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1027 assert((abase & clear_low) == 0, "array base must be int-aligned");
1028 }
1029 // Round bumped 'start' down to jlong boundary in body of array.
1030 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1031 if (bump_bit != 0) {
1032 // Store a zero to the immediately preceding jint:
1033 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1034 Node* p1 = basic_plus_adr(dest, x1);
1035 if (val == nullptr) {
1036 assert(raw_val == nullptr, "val may not be null");
1037 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1038 } else {
1039 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1040 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1041 }
1042 mem = transform_later(mem);
1043 }
1044 }
1045 Node* end = dest_size; // pre-rounded
1046 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1047 start, end, &_igvn);
1048 } else {
1049 // Non-constant start, unrounded non-constant end.
1050 // (Nobody zeroes a random midsection of an array using this routine.)
1051 ShouldNotReachHere(); // fix caller
1052 }
1053
1054 // Done.
1055 merge_mem->set_memory_at(alias_idx, mem);
1056 }
1057
1058 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1059 const TypePtr* adr_type,
1060 BasicType basic_elem_type,
1061 AllocateNode* alloc,
1062 Node* src, Node* src_offset,
1063 Node* dest, Node* dest_offset,
1064 Node* dest_size, bool dest_uninitialized) {
1065 // See if there is an advantage from block transfer.
1066 int scale = exact_log2(type2aelembytes(basic_elem_type));
1142 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1143 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1144 "slow_arraycopy", TypePtr::BOTTOM);
1145
1146 call->init_req(TypeFunc::Control, *ctrl);
1147 call->init_req(TypeFunc::I_O , *io);
1148 call->init_req(TypeFunc::Memory , mem);
1149 call->init_req(TypeFunc::ReturnAdr, top());
1150 call->init_req(TypeFunc::FramePtr, top());
1151 call->init_req(TypeFunc::Parms+0, src);
1152 call->init_req(TypeFunc::Parms+1, src_offset);
1153 call->init_req(TypeFunc::Parms+2, dest);
1154 call->init_req(TypeFunc::Parms+3, dest_offset);
1155 call->init_req(TypeFunc::Parms+4, copy_length);
1156 call->copy_call_debug_info(&_igvn, ac);
1157
1158 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1159 _igvn.replace_node(ac, call);
1160 transform_later(call);
1161
1162 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1163 *ctrl = _callprojs->fallthrough_catchproj->clone();
1164 transform_later(*ctrl);
1165
1166 Node* m = _callprojs->fallthrough_memproj->clone();
1167 transform_later(m);
1168
1169 uint alias_idx = C->get_alias_index(adr_type);
1170 MergeMemNode* out_mem;
1171 if (alias_idx != Compile::AliasIdxBot) {
1172 out_mem = MergeMemNode::make(mem);
1173 out_mem->set_memory_at(alias_idx, m);
1174 } else {
1175 out_mem = MergeMemNode::make(m);
1176 }
1177 transform_later(out_mem);
1178
1179 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1180 // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1181 if (_callprojs->fallthrough_ioproj != nullptr) {
1182 *io = _callprojs->fallthrough_ioproj->clone();
1183 transform_later(*io);
1184 } else {
1185 *io = nullptr;
1186 }
1187
1188 return out_mem;
1189 }
1190
1191 // Helper function; generates code for cases requiring runtime checks.
1192 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1193 const TypePtr* adr_type,
1194 Node* dest_elem_klass,
1195 Node* src, Node* src_offset,
1196 Node* dest, Node* dest_offset,
1197 Node* copy_length, bool dest_uninitialized) {
1198 if ((*ctrl)->is_top()) return nullptr;
1199
1200 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1201 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1202 return nullptr;
1294
1295 // Connecting remaining edges for exit_block coming from stub_block.
1296 if (exit_block) {
1297 exit_block->init_req(2, *ctrl);
1298
1299 // Memory edge corresponding to stub_region.
1300 result_memory->init_req(2, *mem);
1301
1302 uint alias_idx = C->get_alias_index(adr_type);
1303 if (alias_idx != Compile::AliasIdxBot) {
1304 *mem = MergeMemNode::make(*mem);
1305 (*mem)->set_memory_at(alias_idx, result_memory);
1306 } else {
1307 *mem = MergeMemNode::make(result_memory);
1308 }
1309 transform_later(*mem);
1310 *ctrl = exit_block;
1311 }
1312 }
1313
1314 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1315 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1316 Node*& dest_length) {
1317 #ifdef ASSERT
1318 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1319 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1320 bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1321 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1322 #endif
1323 int elem_size = top_dest->flat_elem_size();
1324 if (elem_size >= 8) {
1325 if (elem_size > 8) {
1326 // treat as array of long but scale length, src offset and dest offset
1327 assert((elem_size % 8) == 0, "not a power of 2?");
1328 int factor = elem_size / 8;
1329 length = transform_later(new MulINode(length, intcon(factor)));
1330 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1331 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1332 if (dest_length != nullptr) {
1333 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1334 }
1335 elem_size = 8;
1336 }
1337 dest_elem = T_LONG;
1338 } else if (elem_size == 4) {
1339 dest_elem = T_INT;
1340 } else if (elem_size == 2) {
1341 dest_elem = T_CHAR;
1342 } else if (elem_size == 1) {
1343 dest_elem = T_BYTE;
1344 } else {
1345 ShouldNotReachHere();
1346 }
1347 return TypeRawPtr::BOTTOM;
1348 }
1349
1350 #undef XTOP
1351
1352 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1353 Node* ctrl = ac->in(TypeFunc::Control);
1354 Node* io = ac->in(TypeFunc::I_O);
1355 Node* src = ac->in(ArrayCopyNode::Src);
1356 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1357 Node* dest = ac->in(ArrayCopyNode::Dest);
1358 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1359 Node* length = ac->in(ArrayCopyNode::Length);
1360 MergeMemNode* merge_mem = nullptr;
1361
1362 if (ac->is_clonebasic()) {
1363 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1364 bs->clone_at_expansion(this, ac);
1365 return;
1366 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1367 const Type* src_type = _igvn.type(src);
1368 const Type* dest_type = _igvn.type(dest);
1369 const TypeAryPtr* top_src = src_type->isa_aryptr();
1370 // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1371 // Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1372 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1373 BasicType dest_elem = T_OBJECT;
1374 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1375 dest_elem = top_dest->elem()->array_element_basic_type();
1376 }
1377 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1378
1379 if (top_src != nullptr && top_src->is_flat()) {
1380 // If src is flat, dest is guaranteed to be flat as well
1381 top_dest = top_src;
1382 }
1383
1384 AllocateArrayNode* alloc = nullptr;
1385 Node* dest_length = nullptr;
1386 if (ac->is_alloc_tightly_coupled()) {
1387 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1388 assert(alloc != nullptr, "expect alloc");
1389 dest_length = alloc->in(AllocateNode::ALength);
1390 }
1391
1392 Node* mem = ac->in(TypeFunc::Memory);
1393 const TypePtr* adr_type = nullptr;
1394 if (top_dest != nullptr && top_dest->is_flat()) {
1395 assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1396 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1397 // on all slices to prevent writes into the source from floating below the arraycopy.
1398 int mem_bar_alias_idx = Compile::AliasIdxBot;
1399 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1400 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1401 }
1402 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1403 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1404 } else {
1405 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1406 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1407 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1408 }
1409 if (ac->_src_type != ac->_dest_type) {
1410 adr_type = TypeRawPtr::BOTTOM;
1411 }
1412 }
1413 merge_mem = MergeMemNode::make(mem);
1414 transform_later(merge_mem);
1415
1416 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1417 adr_type, dest_elem,
1418 src, src_offset, dest, dest_offset, length,
1419 dest_length,
1420 true, ac->has_negative_length_guard());
1421
1422 return;
1423 }
1424
1425 AllocateArrayNode* alloc = nullptr;
1426 if (ac->is_alloc_tightly_coupled()) {
1427 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1428 assert(alloc != nullptr, "expect alloc");
1429 }
1430
1431 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1432
1433 // Compile time checks. If any of these checks cannot be verified at compile time,
1434 // we do not make a fast path for this call. Instead, we let the call remain as it
1435 // is. The checks we choose to mandate at compile time are:
1436 //
1437 // (1) src and dest are arrays.
1438 const Type* src_type = src->Value(&_igvn);
1439 const Type* dest_type = dest->Value(&_igvn);
1440 const TypeAryPtr* top_src = src_type->isa_aryptr();
1441 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1442
1443 BasicType src_elem = T_CONFLICT;
1444 BasicType dest_elem = T_CONFLICT;
1445
1446 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1447 src_elem = top_src->elem()->array_element_basic_type();
1448 }
1449 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1450 dest_elem = top_dest->elem()->array_element_basic_type();
1451 }
1452 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1453 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1454
1455 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1456 src_elem = dest_elem;
1457 }
1458
1459 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1460 // Conservatively insert a memory barrier on all memory slices.
1461 // Do not let writes into the source float below the arraycopy.
1462 {
1463 Node* mem = ac->in(TypeFunc::Memory);
1464 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1465
1466 merge_mem = MergeMemNode::make(mem);
1467 transform_later(merge_mem);
1468 }
1469
1470 // Call StubRoutines::generic_arraycopy stub.
1471 generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1472 TypeRawPtr::BOTTOM, T_CONFLICT,
1473 src, src_offset, dest, dest_offset, length,
1474 nullptr,
1475 // If a negative length guard was generated for the ArrayCopyNode,
1476 // the length of the array can never be negative.
1477 false, ac->has_negative_length_guard());
1478 return;
1479 }
1480
1481 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1482
1483 // (2) src and dest arrays must have elements of the same BasicType
1484 // Figure out the size and type of the elements we will be copying.
1485 //
1486 // We have no stub to copy flat inline type arrays with oop
1487 // fields if we need to emit write barriers.
1488 //
1489 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1490 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1491 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1492 bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1493 // The component types are not the same or are not recognized. Punt.
1494 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1495 {
1496 Node* mem = ac->in(TypeFunc::Memory);
1497 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1498 }
1499
1500 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1501 if (_callprojs->fallthrough_ioproj != nullptr) {
1502 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1503 }
1504 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1505 return;
1506 }
1507
1508 //---------------------------------------------------------------------------
1509 // We will make a fast path for this call to arraycopy.
1510
1511 // We have the following tests left to perform:
1512 //
1513 // (3) src and dest must not be null.
1514 // (4) src_offset must not be negative.
1515 // (5) dest_offset must not be negative.
1516 // (6) length must not be negative.
1517 // (7) src_offset + length must not exceed length of src.
1518 // (8) dest_offset + length must not exceed length of dest.
1519 // (9) each element of an oop array must be assignable
1520
1521 Node* mem = ac->in(TypeFunc::Memory);
1522 if (top_dest->is_flat()) {
1523 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1524 // on all slices to prevent writes into the source from floating below the arraycopy.
1525 int mem_bar_alias_idx = Compile::AliasIdxBot;
1526 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1527 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1528 }
1529 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1530 }
1531 merge_mem = MergeMemNode::make(mem);
1532 transform_later(merge_mem);
1533
1534 RegionNode* slow_region = new RegionNode(1);
1535 transform_later(slow_region);
1536
1537 if (!ac->is_arraycopy_validated()) {
1538 // (3) operands must not be null
1539 // We currently perform our null checks with the null_check routine.
1540 // This means that the null exceptions will be reported in the caller
1541 // rather than (correctly) reported inside of the native arraycopy call.
1542 // This should be corrected, given time. We do our null check with the
1543 // stack pointer restored.
1544 // null checks done library_call.cpp
1545
1546 // (4) src_offset must not be negative.
1547 generate_negative_guard(&ctrl, src_offset, slow_region);
1548
1549 // (5) dest_offset must not be negative.
1550 generate_negative_guard(&ctrl, dest_offset, slow_region);
1551
1552 // (6) length must not be negative (moved to generate_arraycopy()).
1553 // generate_negative_guard(length, slow_region);
1554
1555 // (7) src_offset + length must not exceed length of src.
1556 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1557 assert(alen != nullptr, "need src len");
1558 generate_limit_guard(&ctrl,
1559 src_offset, length,
1560 alen,
1561 slow_region);
1562
1563 // (8) dest_offset + length must not exceed length of dest.
1564 alen = ac->in(ArrayCopyNode::DestLen);
1565 assert(alen != nullptr, "need dest len");
1566 generate_limit_guard(&ctrl,
1567 dest_offset, length,
1568 alen,
1569 slow_region);
1570
1571 // (9) each element of an oop array must be assignable
1572 // The generate_arraycopy subroutine checks this.
1573
1574 // TODO 8350865 This is too strong
1575 // 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.
1576 if (!(top_src->is_flat() && top_dest->is_flat() && top_src->is_null_free() == top_dest->is_null_free())) {
1577 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1578 generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1579 generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1580 }
1581 }
1582
1583 // This is where the memory effects are placed:
1584 const TypePtr* adr_type = nullptr;
1585 Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1586
1587 if (top_src->is_flat() && top_dest->is_flat() &&
1588 top_src->is_null_free() == top_dest->is_null_free()) {
1589 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1590 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1591 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1592 } else {
1593 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1594 }
1595
1596 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1597 adr_type, dest_elem,
1598 src, src_offset, dest, dest_offset, length,
1599 dest_length,
1600 // If a negative length guard was generated for the ArrayCopyNode,
1601 // the length of the array can never be negative.
1602 false, ac->has_negative_length_guard(),
1603 slow_region);
1604 }
|