6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/tlab_globals.hpp"
28 #include "opto/arraycopynode.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/convertnode.hpp"
31 #include "opto/vectornode.hpp"
32 #include "opto/graphKit.hpp"
33 #include "opto/macro.hpp"
34 #include "opto/runtime.hpp"
35 #include "opto/castnode.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, Node* precedent) {
41 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, 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);
122 }
123
124 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
125 transform_later(iff);
126
127 Node* if_slow = new IfTrueNode(iff);
128 transform_later(if_slow);
129
130 if (region != NULL) {
131 region->add_req(if_slow);
132 }
133
134 Node* if_fast = new IfFalseNode(iff);
135 transform_later(if_fast);
136
137 *ctrl = if_fast;
138
139 return if_slow;
140 }
141
142 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
143 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
144 }
145
146 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
147 if ((*ctrl)->is_top())
148 return; // already stopped
149 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
150 return; // index is already adequately typed
151 Node* cmp_lt = new CmpINode(index, intcon(0));
152 transform_later(cmp_lt);
153 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
154 transform_later(bol_lt);
155 generate_guard(ctrl, bol_lt, region, PROB_MIN);
156 }
157
158 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
159 if ((*ctrl)->is_top())
160 return; // already stopped
161 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
162 if (zero_offset && subseq_length->eqv_uncast(array_length))
163 return; // common case of whole-array copy
164 Node* last = subseq_length;
165 if (!zero_offset) { // last += offset
266
267 *ctrl = stub_block;
268 }
269
270
271 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
272 if ((*ctrl)->is_top()) return NULL;
273
274 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
275 return NULL; // index is already adequately typed
276 Node* cmp_le = new CmpINode(index, intcon(0));
277 transform_later(cmp_le);
278 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
279 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
280 transform_later(bol_le);
281 Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
282
283 return is_notp;
284 }
285
286 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
287 transform_later(call);
288
289 *ctrl = new ProjNode(call,TypeFunc::Control);
290 transform_later(*ctrl);
291 Node* newmem = new ProjNode(call, TypeFunc::Memory);
292 transform_later(newmem);
293
294 uint alias_idx = C->get_alias_index(adr_type);
295 if (alias_idx != Compile::AliasIdxBot) {
296 *mem = MergeMemNode::make(*mem);
297 (*mem)->set_memory_at(alias_idx, newmem);
298 } else {
299 *mem = MergeMemNode::make(newmem);
300 }
301 transform_later(*mem);
302 }
303
304 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
305 Node* src_offset,
360 // }
361 // }
362 // // adjust params for remaining work:
363 // if (slowval != -1) {
364 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
365 // }
366 // slow_region:
367 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
368 // return // via slow_call_path
369 //
370 // This routine is used from several intrinsics: System.arraycopy,
371 // Object.clone (the array subcase), and Arrays.copyOf[Range].
372 //
373 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
374 Node** ctrl, MergeMemNode* mem, Node** io,
375 const TypePtr* adr_type,
376 BasicType basic_elem_type,
377 Node* src, Node* src_offset,
378 Node* dest, Node* dest_offset,
379 Node* copy_length,
380 bool disjoint_bases,
381 bool length_never_negative,
382 RegionNode* slow_region) {
383 if (slow_region == NULL) {
384 slow_region = new RegionNode(1);
385 transform_later(slow_region);
386 }
387
388 Node* original_dest = dest;
389 bool dest_needs_zeroing = false;
390 bool acopy_to_uninitialized = false;
391
392 // See if this is the initialization of a newly-allocated array.
393 // If so, we will take responsibility here for initializing it to zero.
394 // (Note: Because tightly_coupled_allocation performs checks on the
395 // out-edges of the dest, we need to avoid making derived pointers
396 // from it until we have checked its uses.)
397 if (ReduceBulkZeroing
398 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
399 && basic_elem_type != T_CONFLICT // avoid corner case
400 && !src->eqv_uncast(dest)
401 && alloc != NULL
402 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
403 assert(ac->is_alloc_tightly_coupled(), "sanity");
404 // acopy to uninitialized tightly coupled allocations
405 // needs zeroing outside the copy range
406 // and the acopy itself will be to uninitialized memory
407 acopy_to_uninitialized = true;
408 if (alloc->maybe_set_complete(&_igvn)) {
409 // "You break it, you buy it."
410 InitializeNode* init = alloc->initialization();
411 assert(init->is_complete(), "we just did this");
412 init->set_complete_with_arraycopy();
413 assert(dest->is_CheckCastPP(), "sanity");
414 assert(dest->in(0)->in(0) == init, "dest pinned");
415 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
416 // From this point on, every exit path is responsible for
417 // initializing any non-copied parts of the object to zero.
418 // Also, if this flag is set we make sure that arraycopy interacts properly
419 // with G1, eliding pre-barriers. See CR 6627983.
420 dest_needs_zeroing = true;
421 } else {
422 // dest_need_zeroing = false;
423 }
424 } else {
425 // No zeroing elimination needed here.
426 alloc = NULL;
427 acopy_to_uninitialized = false;
428 //original_dest = dest;
429 //dest_needs_zeroing = false;
430 }
431
432 uint alias_idx = C->get_alias_index(adr_type);
433
434 // Results are placed here:
435 enum { fast_path = 1, // normal void-returning assembly stub
436 checked_path = 2, // special assembly stub with cleanup
437 slow_call_path = 3, // something went wrong; call the VM
438 zero_path = 4, // bypass when length of copy is zero
439 bcopy_path = 5, // copy primitive array by 64-bit blocks
440 PATH_LIMIT = 6
470 checked_i_o = *io;
471 checked_mem = mem->memory_at(alias_idx);
472 checked_value = cv;
473 *ctrl = top();
474 }
475
476 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
477 if (not_pos != NULL) {
478 Node* local_ctrl = not_pos, *local_io = *io;
479 MergeMemNode* local_mem = MergeMemNode::make(mem);
480 transform_later(local_mem);
481
482 // (6) length must not be negative.
483 if (!length_never_negative) {
484 generate_negative_guard(&local_ctrl, copy_length, slow_region);
485 }
486
487 // copy_length is 0.
488 if (dest_needs_zeroing) {
489 assert(!local_ctrl->is_top(), "no ctrl?");
490 Node* dest_length = alloc->in(AllocateNode::ALength);
491 if (copy_length->eqv_uncast(dest_length)
492 || _igvn.find_int_con(dest_length, 1) <= 0) {
493 // There is no zeroing to do. No need for a secondary raw memory barrier.
494 } else {
495 // Clear the whole thing since there are no source elements to copy.
496 generate_clear_array(local_ctrl, local_mem,
497 adr_type, dest, basic_elem_type,
498 intcon(0), NULL,
499 alloc->in(AllocateNode::AllocSize));
500 // Use a secondary InitializeNode as raw memory barrier.
501 // Currently it is needed only on this path since other
502 // paths have stub or runtime calls as raw memory barriers.
503 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
504 Compile::AliasIdxRaw,
505 top());
506 transform_later(mb);
507 mb->set_req(TypeFunc::Control,local_ctrl);
508 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
509 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
510 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
511
512 InitializeNode* init = mb->as_Initialize();
513 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
514 }
515 }
516
517 // Present the results of the fast call.
518 result_region->init_req(zero_path, local_ctrl);
519 result_i_o ->init_req(zero_path, local_io);
520 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
521 }
522
523 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
524 // We have to initialize the *uncopied* part of the array to zero.
525 // The copy destination is the slice dest[off..off+len]. The other slices
526 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
527 Node* dest_size = alloc->in(AllocateNode::AllocSize);
528 Node* dest_length = alloc->in(AllocateNode::ALength);
529 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
530
531 // If there is a head section that needs zeroing, do it now.
532 if (_igvn.find_int_con(dest_offset, -1) != 0) {
533 generate_clear_array(*ctrl, mem,
534 adr_type, dest, basic_elem_type,
535 intcon(0), dest_offset,
536 NULL);
537 }
538
539 // Next, perform a dynamic check on the tail length.
540 // It is often zero, and we can win big if we prove this.
541 // There are two wins: Avoid generating the ClearArray
542 // with its attendant messy index arithmetic, and upgrade
543 // the copy to a more hardware-friendly word size of 64 bits.
544 Node* tail_ctl = NULL;
545 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
546 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
547 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
548 tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
549 assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
550 }
551
552 // At this point, let's assume there is no tail.
553 if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
554 // There is no tail. Try an upgrade to a 64-bit copy.
563 src, src_offset, dest, dest_offset,
564 dest_size, acopy_to_uninitialized);
565 if (didit) {
566 // Present the results of the block-copying fast call.
567 result_region->init_req(bcopy_path, local_ctrl);
568 result_i_o ->init_req(bcopy_path, local_io);
569 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
570 }
571 }
572 if (didit) {
573 *ctrl = top(); // no regular fast path
574 }
575 }
576
577 // Clear the tail, if any.
578 if (tail_ctl != NULL) {
579 Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
580 *ctrl = tail_ctl;
581 if (notail_ctl == NULL) {
582 generate_clear_array(*ctrl, mem,
583 adr_type, dest, basic_elem_type,
584 dest_tail, NULL,
585 dest_size);
586 } else {
587 // Make a local merge.
588 Node* done_ctl = transform_later(new RegionNode(3));
589 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
590 done_ctl->init_req(1, notail_ctl);
591 done_mem->init_req(1, mem->memory_at(alias_idx));
592 generate_clear_array(*ctrl, mem,
593 adr_type, dest, basic_elem_type,
594 dest_tail, NULL,
595 dest_size);
596 done_ctl->init_req(2, *ctrl);
597 done_mem->init_req(2, mem->memory_at(alias_idx));
598 *ctrl = done_ctl;
599 mem->set_memory_at(alias_idx, done_mem);
600 }
601 }
602 }
603
604 BasicType copy_type = basic_elem_type;
605 assert(basic_elem_type != T_ARRAY, "caller must fix this");
606 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
607 // If src and dest have compatible element types, we can copy bits.
608 // Types S[] and D[] are compatible if D is a supertype of S.
609 //
610 // If they are not, we will use checked_oop_disjoint_arraycopy,
611 // which performs a fast optimistic per-oop check, and backs off
612 // further to JVM_ArrayCopy on the first per-oop check that fails.
613 // (Actually, we don't move raw bits only; the GC requires card marks.)
751 Node* length_minus = new SubINode(copy_length, slow_offset);
752 transform_later(length_minus);
753
754 // Tweak the node variables to adjust the code produced below:
755 src_offset = src_off_plus;
756 dest_offset = dest_off_plus;
757 copy_length = length_minus;
758 }
759 }
760 *ctrl = slow_control;
761 if (!(*ctrl)->is_top()) {
762 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
763 MergeMemNode* local_mem = MergeMemNode::make(mem);
764 transform_later(local_mem);
765
766 // Generate the slow path, if needed.
767 local_mem->set_memory_at(alias_idx, slow_mem);
768
769 if (dest_needs_zeroing) {
770 generate_clear_array(local_ctrl, local_mem,
771 adr_type, dest, basic_elem_type,
772 intcon(0), NULL,
773 alloc->in(AllocateNode::AllocSize));
774 }
775
776 local_mem = generate_slow_arraycopy(ac,
777 &local_ctrl, local_mem, &local_io,
778 adr_type,
779 src, src_offset, dest, dest_offset,
780 copy_length, /*dest_uninitialized*/false);
781
782 result_region->init_req(slow_call_path, local_ctrl);
783 result_i_o ->init_req(slow_call_path, local_io);
784 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
785 } else {
786 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
787 // projections were not extracted
788 }
789
790 // Remove unused edges.
791 for (uint i = 1; i < result_region->req(); i++) {
813 // The next memory barrier is added to avoid it. If the arraycopy can be
814 // optimized away (which it can, sometimes) then we can manually remove
815 // the membar also.
816 //
817 // Do not let reads from the cloned object float above the arraycopy.
818 if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
819 // Do not let stores that initialize this object be reordered with
820 // a subsequent store that would make this object accessible by
821 // other threads.
822 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
823 } else {
824 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
825 }
826
827 if (is_partial_array_copy) {
828 assert((*ctrl)->is_Proj(), "MemBar control projection");
829 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
830 (*ctrl)->in(0)->isa_MemBar()->set_trailing_partial_array_copy();
831 }
832
833 _igvn.replace_node(_callprojs.fallthrough_memproj, out_mem);
834 if (_callprojs.fallthrough_ioproj != NULL) {
835 _igvn.replace_node(_callprojs.fallthrough_ioproj, *io);
836 }
837 _igvn.replace_node(_callprojs.fallthrough_catchproj, *ctrl);
838
839 #ifdef ASSERT
840 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
841 if (dest_t->is_known_instance() && !is_partial_array_copy) {
842 ArrayCopyNode* ac = NULL;
843 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
844 assert(ac == NULL, "no arraycopy anymore");
845 }
846 #endif
847
848 return out_mem;
849 }
850
851 // Helper for initialization of arrays, creating a ClearArray.
852 // It writes zero bits in [start..end), within the body of an array object.
853 // The memory effects are all chained onto the 'adr_type' alias category.
854 //
855 // Since the object is otherwise uninitialized, we are free
856 // to put a little "slop" around the edges of the cleared area,
857 // as long as it does not go back into the array's header,
858 // or beyond the array end within the heap.
859 //
860 // The lower edge can be rounded down to the nearest jint and the
861 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
862 //
863 // Arguments:
864 // adr_type memory slice where writes are generated
865 // dest oop of the destination array
866 // basic_elem_type element type of the destination
867 // slice_idx array index of first element to store
868 // slice_len number of elements to store (or NULL)
869 // dest_size total size in bytes of the array object
870 //
871 // Exactly one of slice_len or dest_size must be non-NULL.
872 // If dest_size is non-NULL, zeroing extends to the end of the object.
873 // If slice_len is non-NULL, the slice_idx value must be a constant.
874 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
875 const TypePtr* adr_type,
876 Node* dest,
877 BasicType basic_elem_type,
878 Node* slice_idx,
879 Node* slice_len,
880 Node* dest_size) {
881 // one or the other but not both of slice_len and dest_size:
882 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
883 if (slice_len == NULL) slice_len = top();
884 if (dest_size == NULL) dest_size = top();
885
886 uint alias_idx = C->get_alias_index(adr_type);
887
888 // operate on this memory slice:
889 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
890
891 // scaling and rounding of indexes:
892 int scale = exact_log2(type2aelembytes(basic_elem_type));
893 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
894 int clear_low = (-1 << scale) & (BytesPerInt - 1);
895 int bump_bit = (-1 << scale) & BytesPerInt;
896
897 // determine constant starts and ends
898 const intptr_t BIG_NEG = -128;
899 assert(BIG_NEG + 2*abase < 0, "neg enough");
900 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
901 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
902 if (slice_len_con == 0) {
903 return; // nothing to do here
904 }
905 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
906 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
907 if (slice_idx_con >= 0 && slice_len_con >= 0) {
908 assert(end_con < 0, "not two cons");
909 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
910 BytesPerLong);
911 }
912
913 if (start_con >= 0 && end_con >= 0) {
914 // Constant start and end. Simple.
915 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
916 start_con, end_con, &_igvn);
917 } else if (start_con >= 0 && dest_size != top()) {
918 // Constant start, pre-rounded end after the tail of the array.
919 Node* end = dest_size;
920 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
921 start_con, end, &_igvn);
922 } else if (start_con >= 0 && slice_len != top()) {
923 // Constant start, non-constant end. End needs rounding up.
924 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
925 intptr_t end_base = abase + (slice_idx_con << scale);
926 int end_round = (-1 << scale) & (BytesPerLong - 1);
927 Node* end = ConvI2X(slice_len);
928 if (scale != 0)
929 end = transform_later(new LShiftXNode(end, intcon(scale) ));
930 end_base += end_round;
931 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
932 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
933 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
934 start_con, end, &_igvn);
935 } else if (start_con < 0 && dest_size != top()) {
936 // Non-constant start, pre-rounded end after the tail of the array.
937 // This is almost certainly a "round-to-end" operation.
938 Node* start = slice_idx;
939 start = ConvI2X(start);
940 if (scale != 0)
941 start = transform_later(new LShiftXNode( start, intcon(scale) ));
942 start = transform_later(new AddXNode(start, MakeConX(abase)) );
943 if ((bump_bit | clear_low) != 0) {
944 int to_clear = (bump_bit | clear_low);
945 // Align up mod 8, then store a jint zero unconditionally
946 // just before the mod-8 boundary.
947 if (((abase + bump_bit) & ~to_clear) - bump_bit
948 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
949 bump_bit = 0;
950 assert((abase & to_clear) == 0, "array base must be long-aligned");
951 } else {
952 // Bump 'start' up to (or past) the next jint boundary:
953 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
954 assert((abase & clear_low) == 0, "array base must be int-aligned");
955 }
956 // Round bumped 'start' down to jlong boundary in body of array.
957 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
958 if (bump_bit != 0) {
959 // Store a zero to the immediately preceding jint:
960 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
961 Node* p1 = basic_plus_adr(dest, x1);
962 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
963 mem = transform_later(mem);
964 }
965 }
966 Node* end = dest_size; // pre-rounded
967 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
968 start, end, &_igvn);
969 } else {
970 // Non-constant start, unrounded non-constant end.
971 // (Nobody zeroes a random midsection of an array using this routine.)
972 ShouldNotReachHere(); // fix caller
973 }
974
975 // Done.
976 merge_mem->set_memory_at(alias_idx, mem);
977 }
978
979 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
980 const TypePtr* adr_type,
981 BasicType basic_elem_type,
982 AllocateNode* alloc,
983 Node* src, Node* src_offset,
984 Node* dest, Node* dest_offset,
985 Node* dest_size, bool dest_uninitialized) {
986 // See if there is an advantage from block transfer.
987 int scale = exact_log2(type2aelembytes(basic_elem_type));
1063 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1064 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1065 "slow_arraycopy", TypePtr::BOTTOM);
1066
1067 call->init_req(TypeFunc::Control, *ctrl);
1068 call->init_req(TypeFunc::I_O , *io);
1069 call->init_req(TypeFunc::Memory , mem);
1070 call->init_req(TypeFunc::ReturnAdr, top());
1071 call->init_req(TypeFunc::FramePtr, top());
1072 call->init_req(TypeFunc::Parms+0, src);
1073 call->init_req(TypeFunc::Parms+1, src_offset);
1074 call->init_req(TypeFunc::Parms+2, dest);
1075 call->init_req(TypeFunc::Parms+3, dest_offset);
1076 call->init_req(TypeFunc::Parms+4, copy_length);
1077 call->copy_call_debug_info(&_igvn, ac);
1078
1079 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1080 _igvn.replace_node(ac, call);
1081 transform_later(call);
1082
1083 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1084 *ctrl = _callprojs.fallthrough_catchproj->clone();
1085 transform_later(*ctrl);
1086
1087 Node* m = _callprojs.fallthrough_memproj->clone();
1088 transform_later(m);
1089
1090 uint alias_idx = C->get_alias_index(adr_type);
1091 MergeMemNode* out_mem;
1092 if (alias_idx != Compile::AliasIdxBot) {
1093 out_mem = MergeMemNode::make(mem);
1094 out_mem->set_memory_at(alias_idx, m);
1095 } else {
1096 out_mem = MergeMemNode::make(m);
1097 }
1098 transform_later(out_mem);
1099
1100 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1101 // could be NULL. Skip clone and update NULL fallthrough_ioproj.
1102 if (_callprojs.fallthrough_ioproj != NULL) {
1103 *io = _callprojs.fallthrough_ioproj->clone();
1104 transform_later(*io);
1105 } else {
1106 *io = NULL;
1107 }
1108
1109 return out_mem;
1110 }
1111
1112 // Helper function; generates code for cases requiring runtime checks.
1113 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1114 const TypePtr* adr_type,
1115 Node* dest_elem_klass,
1116 Node* src, Node* src_offset,
1117 Node* dest, Node* dest_offset,
1118 Node* copy_length, bool dest_uninitialized) {
1119 if ((*ctrl)->is_top()) return NULL;
1120
1121 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1122 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1123 return NULL;
1215 if (exit_block) {
1216 exit_block->init_req(2, *ctrl);
1217
1218 // Memory edge corresponding to stub_region.
1219 result_memory->init_req(2, *mem);
1220
1221 uint alias_idx = C->get_alias_index(adr_type);
1222 if (alias_idx != Compile::AliasIdxBot) {
1223 *mem = MergeMemNode::make(*mem);
1224 (*mem)->set_memory_at(alias_idx, result_memory);
1225 } else {
1226 *mem = MergeMemNode::make(result_memory);
1227 }
1228 transform_later(*mem);
1229 *ctrl = exit_block;
1230 return true;
1231 }
1232 return false;
1233 }
1234
1235 #undef XTOP
1236
1237 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1238 Node* ctrl = ac->in(TypeFunc::Control);
1239 Node* io = ac->in(TypeFunc::I_O);
1240 Node* src = ac->in(ArrayCopyNode::Src);
1241 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1242 Node* dest = ac->in(ArrayCopyNode::Dest);
1243 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1244 Node* length = ac->in(ArrayCopyNode::Length);
1245 MergeMemNode* merge_mem = NULL;
1246
1247 if (ac->is_clonebasic()) {
1248 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1249 bs->clone_at_expansion(this, ac);
1250 return;
1251 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1252 Node* mem = ac->in(TypeFunc::Memory);
1253 merge_mem = MergeMemNode::make(mem);
1254 transform_later(merge_mem);
1255
1256 AllocateArrayNode* alloc = NULL;
1257 if (ac->is_alloc_tightly_coupled()) {
1258 alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1259 assert(alloc != NULL, "expect alloc");
1260 }
1261
1262 const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1263 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1264 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1265 }
1266 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1267 adr_type, T_OBJECT,
1268 src, src_offset, dest, dest_offset, length,
1269 true, !ac->is_copyofrange());
1270
1271 return;
1272 }
1273
1274 AllocateArrayNode* alloc = NULL;
1275 if (ac->is_alloc_tightly_coupled()) {
1276 alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1277 assert(alloc != NULL, "expect alloc");
1278 }
1279
1280 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1281
1282 // Compile time checks. If any of these checks cannot be verified at compile time,
1283 // we do not make a fast path for this call. Instead, we let the call remain as it
1284 // is. The checks we choose to mandate at compile time are:
1285 //
1286 // (1) src and dest are arrays.
1287 const Type* src_type = src->Value(&_igvn);
1288 const Type* dest_type = dest->Value(&_igvn);
1289 const TypeAryPtr* top_src = src_type->isa_aryptr();
1290 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1291
1292 BasicType src_elem = T_CONFLICT;
1293 BasicType dest_elem = T_CONFLICT;
1294
1295 if (top_src != NULL && top_src->elem() != Type::BOTTOM) {
1296 src_elem = top_src->elem()->array_element_basic_type();
1297 }
1298 if (top_dest != NULL && top_dest->elem() != Type::BOTTOM) {
1299 dest_elem = top_dest->elem()->array_element_basic_type();
1300 }
1301 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1302 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1303
1304 if (ac->is_arraycopy_validated() &&
1305 dest_elem != T_CONFLICT &&
1306 src_elem == T_CONFLICT) {
1307 src_elem = dest_elem;
1308 }
1309
1310 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1311 // Conservatively insert a memory barrier on all memory slices.
1312 // Do not let writes into the source float below the arraycopy.
1313 {
1314 Node* mem = ac->in(TypeFunc::Memory);
1315 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1316
1317 merge_mem = MergeMemNode::make(mem);
1318 transform_later(merge_mem);
1319 }
1320
1321 // Call StubRoutines::generic_arraycopy stub.
1322 Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1323 TypeRawPtr::BOTTOM, T_CONFLICT,
1324 src, src_offset, dest, dest_offset, length,
1325 // If a negative length guard was generated for the ArrayCopyNode,
1326 // the length of the array can never be negative.
1327 false, ac->has_negative_length_guard());
1328 return;
1329 }
1330
1331 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1332
1333 // (2) src and dest arrays must have elements of the same BasicType
1334 // Figure out the size and type of the elements we will be copying.
1335 if (src_elem != dest_elem || dest_elem == T_VOID) {
1336 // The component types are not the same or are not recognized. Punt.
1337 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1338 {
1339 Node* mem = ac->in(TypeFunc::Memory);
1340 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1341 }
1342
1343 _igvn.replace_node(_callprojs.fallthrough_memproj, merge_mem);
1344 if (_callprojs.fallthrough_ioproj != NULL) {
1345 _igvn.replace_node(_callprojs.fallthrough_ioproj, io);
1346 }
1347 _igvn.replace_node(_callprojs.fallthrough_catchproj, ctrl);
1348 return;
1349 }
1350
1351 //---------------------------------------------------------------------------
1352 // We will make a fast path for this call to arraycopy.
1353
1354 // We have the following tests left to perform:
1355 //
1356 // (3) src and dest must not be null.
1357 // (4) src_offset must not be negative.
1358 // (5) dest_offset must not be negative.
1359 // (6) length must not be negative.
1360 // (7) src_offset + length must not exceed length of src.
1361 // (8) dest_offset + length must not exceed length of dest.
1362 // (9) each element of an oop array must be assignable
1363
1364 {
1365 Node* mem = ac->in(TypeFunc::Memory);
1366 merge_mem = MergeMemNode::make(mem);
1367 transform_later(merge_mem);
1368 }
1369
1370 RegionNode* slow_region = new RegionNode(1);
1371 transform_later(slow_region);
1372
1373 if (!ac->is_arraycopy_validated()) {
1374 // (3) operands must not be null
1375 // We currently perform our null checks with the null_check routine.
1376 // This means that the null exceptions will be reported in the caller
1377 // rather than (correctly) reported inside of the native arraycopy call.
1378 // This should be corrected, given time. We do our null check with the
1379 // stack pointer restored.
1380 // null checks done library_call.cpp
1381
1382 // (4) src_offset must not be negative.
1383 generate_negative_guard(&ctrl, src_offset, slow_region);
1384
1385 // (5) dest_offset must not be negative.
1386 generate_negative_guard(&ctrl, dest_offset, slow_region);
1387
1388 // (6) length must not be negative (moved to generate_arraycopy()).
1389 // generate_negative_guard(length, slow_region);
1390
1391 // (7) src_offset + length must not exceed length of src.
1392 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1393 assert(alen != NULL, "need src len");
1394 generate_limit_guard(&ctrl,
1395 src_offset, length,
1396 alen,
1397 slow_region);
1398
1399 // (8) dest_offset + length must not exceed length of dest.
1400 alen = ac->in(ArrayCopyNode::DestLen);
1401 assert(alen != NULL, "need dest len");
1402 generate_limit_guard(&ctrl,
1403 dest_offset, length,
1404 alen,
1405 slow_region);
1406
1407 // (9) each element of an oop array must be assignable
1408 // The generate_arraycopy subroutine checks this.
1409 }
1410 // This is where the memory effects are placed:
1411 const TypePtr* adr_type = NULL;
1412 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1413 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1414 } else {
1415 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1416 }
1417
1418 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1419 adr_type, dest_elem,
1420 src, src_offset, dest, dest_offset, length,
1421 // If a negative length guard was generated for the ArrayCopyNode,
1422 // the length of the array can never be negative.
1423 false, ac->has_negative_length_guard(), slow_region);
1424 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/convertnode.hpp"
32 #include "opto/vectornode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/macro.hpp"
35 #include "opto/runtime.hpp"
36 #include "opto/castnode.hpp"
37 #include "runtime/stubRoutines.hpp"
38 #include "utilities/align.hpp"
39 #include "utilities/powerOfTwo.hpp"
40
41 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
42 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
43 mb->init_req(TypeFunc::Control, *ctrl);
44 mb->init_req(TypeFunc::Memory, *mem);
45 transform_later(mb);
46 *ctrl = new ProjNode(mb,TypeFunc::Control);
123 }
124
125 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
126 transform_later(iff);
127
128 Node* if_slow = new IfTrueNode(iff);
129 transform_later(if_slow);
130
131 if (region != NULL) {
132 region->add_req(if_slow);
133 }
134
135 Node* if_fast = new IfFalseNode(iff);
136 transform_later(if_fast);
137
138 *ctrl = if_fast;
139
140 return if_slow;
141 }
142
143 Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
144 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
145 }
146
147 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
148 return generate_guard(ctrl, test, region, PROB_FAIR);
149 }
150
151 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
152 if ((*ctrl)->is_top())
153 return; // already stopped
154 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
155 return; // index is already adequately typed
156 Node* cmp_lt = new CmpINode(index, intcon(0));
157 transform_later(cmp_lt);
158 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
159 transform_later(bol_lt);
160 generate_guard(ctrl, bol_lt, region, PROB_MIN);
161 }
162
163 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
164 if ((*ctrl)->is_top())
165 return; // already stopped
166 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
167 if (zero_offset && subseq_length->eqv_uncast(array_length))
168 return; // common case of whole-array copy
169 Node* last = subseq_length;
170 if (!zero_offset) { // last += offset
271
272 *ctrl = stub_block;
273 }
274
275
276 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
277 if ((*ctrl)->is_top()) return NULL;
278
279 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
280 return NULL; // index is already adequately typed
281 Node* cmp_le = new CmpINode(index, intcon(0));
282 transform_later(cmp_le);
283 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
284 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
285 transform_later(bol_le);
286 Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
287
288 return is_notp;
289 }
290
291 Node* PhaseMacroExpand::array_lh_test(Node* array, jint mask) {
292 Node* klass_adr = basic_plus_adr(array, oopDesc::klass_offset_in_bytes());
293 Node* klass = transform_later(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
294 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
295 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, NULL, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
296 Node* masked = transform_later(new AndINode(lh_val, intcon(mask)));
297 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
298 return transform_later(new BoolNode(cmp, BoolTest::ne));
299 }
300
301 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, RegionNode* region) {
302 assert(UseFlatArray, "can never be flattened");
303 return generate_fair_guard(ctrl, array_lh_test(array, Klass::_lh_array_tag_flat_value_bit_inplace), region);
304 }
305
306 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, RegionNode* region) {
307 assert(EnableValhalla, "can never be null free");
308 return generate_fair_guard(ctrl, array_lh_test(array, Klass::_lh_null_free_array_bit_inplace), region);
309 }
310
311 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
312 transform_later(call);
313
314 *ctrl = new ProjNode(call,TypeFunc::Control);
315 transform_later(*ctrl);
316 Node* newmem = new ProjNode(call, TypeFunc::Memory);
317 transform_later(newmem);
318
319 uint alias_idx = C->get_alias_index(adr_type);
320 if (alias_idx != Compile::AliasIdxBot) {
321 *mem = MergeMemNode::make(*mem);
322 (*mem)->set_memory_at(alias_idx, newmem);
323 } else {
324 *mem = MergeMemNode::make(newmem);
325 }
326 transform_later(*mem);
327 }
328
329 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
330 Node* src_offset,
385 // }
386 // }
387 // // adjust params for remaining work:
388 // if (slowval != -1) {
389 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
390 // }
391 // slow_region:
392 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
393 // return // via slow_call_path
394 //
395 // This routine is used from several intrinsics: System.arraycopy,
396 // Object.clone (the array subcase), and Arrays.copyOf[Range].
397 //
398 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
399 Node** ctrl, MergeMemNode* mem, Node** io,
400 const TypePtr* adr_type,
401 BasicType basic_elem_type,
402 Node* src, Node* src_offset,
403 Node* dest, Node* dest_offset,
404 Node* copy_length,
405 Node* dest_length,
406 bool disjoint_bases,
407 bool length_never_negative,
408 RegionNode* slow_region) {
409 if (slow_region == NULL) {
410 slow_region = new RegionNode(1);
411 transform_later(slow_region);
412 }
413
414 Node* original_dest = dest;
415 bool dest_needs_zeroing = false;
416 bool acopy_to_uninitialized = false;
417 Node* default_value = NULL;
418 Node* raw_default_value = NULL;
419
420 // See if this is the initialization of a newly-allocated array.
421 // If so, we will take responsibility here for initializing it to zero.
422 // (Note: Because tightly_coupled_allocation performs checks on the
423 // out-edges of the dest, we need to avoid making derived pointers
424 // from it until we have checked its uses.)
425 if (ReduceBulkZeroing
426 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
427 && basic_elem_type != T_CONFLICT // avoid corner case
428 && !src->eqv_uncast(dest)
429 && alloc != NULL
430 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
431 assert(ac->is_alloc_tightly_coupled(), "sanity");
432 // acopy to uninitialized tightly coupled allocations
433 // needs zeroing outside the copy range
434 // and the acopy itself will be to uninitialized memory
435 acopy_to_uninitialized = true;
436 if (alloc->maybe_set_complete(&_igvn)) {
437 // "You break it, you buy it."
438 InitializeNode* init = alloc->initialization();
439 assert(init->is_complete(), "we just did this");
440 init->set_complete_with_arraycopy();
441 assert(dest->is_CheckCastPP(), "sanity");
442 assert(dest->in(0)->in(0) == init, "dest pinned");
443 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
444 // From this point on, every exit path is responsible for
445 // initializing any non-copied parts of the object to zero.
446 // Also, if this flag is set we make sure that arraycopy interacts properly
447 // with G1, eliding pre-barriers. See CR 6627983.
448 dest_needs_zeroing = true;
449 default_value = alloc->in(AllocateNode::DefaultValue);
450 raw_default_value = alloc->in(AllocateNode::RawDefaultValue);
451 } else {
452 // dest_need_zeroing = false;
453 }
454 } else {
455 // No zeroing elimination needed here.
456 alloc = NULL;
457 acopy_to_uninitialized = false;
458 //original_dest = dest;
459 //dest_needs_zeroing = false;
460 }
461
462 uint alias_idx = C->get_alias_index(adr_type);
463
464 // Results are placed here:
465 enum { fast_path = 1, // normal void-returning assembly stub
466 checked_path = 2, // special assembly stub with cleanup
467 slow_call_path = 3, // something went wrong; call the VM
468 zero_path = 4, // bypass when length of copy is zero
469 bcopy_path = 5, // copy primitive array by 64-bit blocks
470 PATH_LIMIT = 6
500 checked_i_o = *io;
501 checked_mem = mem->memory_at(alias_idx);
502 checked_value = cv;
503 *ctrl = top();
504 }
505
506 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
507 if (not_pos != NULL) {
508 Node* local_ctrl = not_pos, *local_io = *io;
509 MergeMemNode* local_mem = MergeMemNode::make(mem);
510 transform_later(local_mem);
511
512 // (6) length must not be negative.
513 if (!length_never_negative) {
514 generate_negative_guard(&local_ctrl, copy_length, slow_region);
515 }
516
517 // copy_length is 0.
518 if (dest_needs_zeroing) {
519 assert(!local_ctrl->is_top(), "no ctrl?");
520 if (copy_length->eqv_uncast(dest_length)
521 || _igvn.find_int_con(dest_length, 1) <= 0) {
522 // There is no zeroing to do. No need for a secondary raw memory barrier.
523 } else {
524 // Clear the whole thing since there are no source elements to copy.
525 generate_clear_array(local_ctrl, local_mem,
526 adr_type, dest,
527 default_value, raw_default_value,
528 basic_elem_type,
529 intcon(0), NULL,
530 alloc->in(AllocateNode::AllocSize));
531 // Use a secondary InitializeNode as raw memory barrier.
532 // Currently it is needed only on this path since other
533 // paths have stub or runtime calls as raw memory barriers.
534 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
535 Compile::AliasIdxRaw,
536 top());
537 transform_later(mb);
538 mb->set_req(TypeFunc::Control,local_ctrl);
539 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
540 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
541 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
542
543 InitializeNode* init = mb->as_Initialize();
544 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
545 }
546 }
547
548 // Present the results of the fast call.
549 result_region->init_req(zero_path, local_ctrl);
550 result_i_o ->init_req(zero_path, local_io);
551 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
552 }
553
554 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
555 // We have to initialize the *uncopied* part of the array to zero.
556 // The copy destination is the slice dest[off..off+len]. The other slices
557 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
558 Node* dest_size = alloc->in(AllocateNode::AllocSize);
559 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
560
561 // If there is a head section that needs zeroing, do it now.
562 if (_igvn.find_int_con(dest_offset, -1) != 0) {
563 generate_clear_array(*ctrl, mem,
564 adr_type, dest,
565 default_value, raw_default_value,
566 basic_elem_type,
567 intcon(0), dest_offset,
568 NULL);
569 }
570
571 // Next, perform a dynamic check on the tail length.
572 // It is often zero, and we can win big if we prove this.
573 // There are two wins: Avoid generating the ClearArray
574 // with its attendant messy index arithmetic, and upgrade
575 // the copy to a more hardware-friendly word size of 64 bits.
576 Node* tail_ctl = NULL;
577 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
578 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
579 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
580 tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
581 assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
582 }
583
584 // At this point, let's assume there is no tail.
585 if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
586 // There is no tail. Try an upgrade to a 64-bit copy.
595 src, src_offset, dest, dest_offset,
596 dest_size, acopy_to_uninitialized);
597 if (didit) {
598 // Present the results of the block-copying fast call.
599 result_region->init_req(bcopy_path, local_ctrl);
600 result_i_o ->init_req(bcopy_path, local_io);
601 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
602 }
603 }
604 if (didit) {
605 *ctrl = top(); // no regular fast path
606 }
607 }
608
609 // Clear the tail, if any.
610 if (tail_ctl != NULL) {
611 Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
612 *ctrl = tail_ctl;
613 if (notail_ctl == NULL) {
614 generate_clear_array(*ctrl, mem,
615 adr_type, dest,
616 default_value, raw_default_value,
617 basic_elem_type,
618 dest_tail, NULL,
619 dest_size);
620 } else {
621 // Make a local merge.
622 Node* done_ctl = transform_later(new RegionNode(3));
623 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
624 done_ctl->init_req(1, notail_ctl);
625 done_mem->init_req(1, mem->memory_at(alias_idx));
626 generate_clear_array(*ctrl, mem,
627 adr_type, dest,
628 default_value, raw_default_value,
629 basic_elem_type,
630 dest_tail, NULL,
631 dest_size);
632 done_ctl->init_req(2, *ctrl);
633 done_mem->init_req(2, mem->memory_at(alias_idx));
634 *ctrl = done_ctl;
635 mem->set_memory_at(alias_idx, done_mem);
636 }
637 }
638 }
639
640 BasicType copy_type = basic_elem_type;
641 assert(basic_elem_type != T_ARRAY, "caller must fix this");
642 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
643 // If src and dest have compatible element types, we can copy bits.
644 // Types S[] and D[] are compatible if D is a supertype of S.
645 //
646 // If they are not, we will use checked_oop_disjoint_arraycopy,
647 // which performs a fast optimistic per-oop check, and backs off
648 // further to JVM_ArrayCopy on the first per-oop check that fails.
649 // (Actually, we don't move raw bits only; the GC requires card marks.)
787 Node* length_minus = new SubINode(copy_length, slow_offset);
788 transform_later(length_minus);
789
790 // Tweak the node variables to adjust the code produced below:
791 src_offset = src_off_plus;
792 dest_offset = dest_off_plus;
793 copy_length = length_minus;
794 }
795 }
796 *ctrl = slow_control;
797 if (!(*ctrl)->is_top()) {
798 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
799 MergeMemNode* local_mem = MergeMemNode::make(mem);
800 transform_later(local_mem);
801
802 // Generate the slow path, if needed.
803 local_mem->set_memory_at(alias_idx, slow_mem);
804
805 if (dest_needs_zeroing) {
806 generate_clear_array(local_ctrl, local_mem,
807 adr_type, dest,
808 default_value, raw_default_value,
809 basic_elem_type,
810 intcon(0), NULL,
811 alloc->in(AllocateNode::AllocSize));
812 }
813
814 local_mem = generate_slow_arraycopy(ac,
815 &local_ctrl, local_mem, &local_io,
816 adr_type,
817 src, src_offset, dest, dest_offset,
818 copy_length, /*dest_uninitialized*/false);
819
820 result_region->init_req(slow_call_path, local_ctrl);
821 result_i_o ->init_req(slow_call_path, local_io);
822 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
823 } else {
824 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
825 // projections were not extracted
826 }
827
828 // Remove unused edges.
829 for (uint i = 1; i < result_region->req(); i++) {
851 // The next memory barrier is added to avoid it. If the arraycopy can be
852 // optimized away (which it can, sometimes) then we can manually remove
853 // the membar also.
854 //
855 // Do not let reads from the cloned object float above the arraycopy.
856 if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
857 // Do not let stores that initialize this object be reordered with
858 // a subsequent store that would make this object accessible by
859 // other threads.
860 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
861 } else {
862 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
863 }
864
865 if (is_partial_array_copy) {
866 assert((*ctrl)->is_Proj(), "MemBar control projection");
867 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
868 (*ctrl)->in(0)->isa_MemBar()->set_trailing_partial_array_copy();
869 }
870
871 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
872 if (_callprojs->fallthrough_ioproj != NULL) {
873 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
874 }
875 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
876
877 #ifdef ASSERT
878 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
879 if (dest_t->is_known_instance() && !is_partial_array_copy) {
880 ArrayCopyNode* ac = NULL;
881 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
882 assert(ac == NULL, "no arraycopy anymore");
883 }
884 #endif
885
886 return out_mem;
887 }
888
889 // Helper for initialization of arrays, creating a ClearArray.
890 // It writes zero bits in [start..end), within the body of an array object.
891 // The memory effects are all chained onto the 'adr_type' alias category.
892 //
893 // Since the object is otherwise uninitialized, we are free
894 // to put a little "slop" around the edges of the cleared area,
895 // as long as it does not go back into the array's header,
896 // or beyond the array end within the heap.
897 //
898 // The lower edge can be rounded down to the nearest jint and the
899 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
900 //
901 // Arguments:
902 // adr_type memory slice where writes are generated
903 // dest oop of the destination array
904 // basic_elem_type element type of the destination
905 // slice_idx array index of first element to store
906 // slice_len number of elements to store (or NULL)
907 // dest_size total size in bytes of the array object
908 //
909 // Exactly one of slice_len or dest_size must be non-NULL.
910 // If dest_size is non-NULL, zeroing extends to the end of the object.
911 // If slice_len is non-NULL, the slice_idx value must be a constant.
912 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
913 const TypePtr* adr_type,
914 Node* dest,
915 Node* val,
916 Node* raw_val,
917 BasicType basic_elem_type,
918 Node* slice_idx,
919 Node* slice_len,
920 Node* dest_size) {
921 // one or the other but not both of slice_len and dest_size:
922 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
923 if (slice_len == NULL) slice_len = top();
924 if (dest_size == NULL) dest_size = top();
925
926 uint alias_idx = C->get_alias_index(adr_type);
927
928 // operate on this memory slice:
929 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
930
931 // scaling and rounding of indexes:
932 assert(basic_elem_type != T_PRIMITIVE_OBJECT, "should have been converted to a basic type copy");
933 int scale = exact_log2(type2aelembytes(basic_elem_type));
934 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
935 int clear_low = (-1 << scale) & (BytesPerInt - 1);
936 int bump_bit = (-1 << scale) & BytesPerInt;
937
938 // determine constant starts and ends
939 const intptr_t BIG_NEG = -128;
940 assert(BIG_NEG + 2*abase < 0, "neg enough");
941 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
942 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
943 if (slice_len_con == 0) {
944 return; // nothing to do here
945 }
946 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
947 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
948 if (slice_idx_con >= 0 && slice_len_con >= 0) {
949 assert(end_con < 0, "not two cons");
950 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
951 BytesPerLong);
952 }
953
954 if (start_con >= 0 && end_con >= 0) {
955 // Constant start and end. Simple.
956 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
957 start_con, end_con, &_igvn);
958 } else if (start_con >= 0 && dest_size != top()) {
959 // Constant start, pre-rounded end after the tail of the array.
960 Node* end = dest_size;
961 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
962 start_con, end, &_igvn);
963 } else if (start_con >= 0 && slice_len != top()) {
964 // Constant start, non-constant end. End needs rounding up.
965 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
966 intptr_t end_base = abase + (slice_idx_con << scale);
967 int end_round = (-1 << scale) & (BytesPerLong - 1);
968 Node* end = ConvI2X(slice_len);
969 if (scale != 0)
970 end = transform_later(new LShiftXNode(end, intcon(scale) ));
971 end_base += end_round;
972 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
973 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
974 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
975 start_con, end, &_igvn);
976 } else if (start_con < 0 && dest_size != top()) {
977 // Non-constant start, pre-rounded end after the tail of the array.
978 // This is almost certainly a "round-to-end" operation.
979 Node* start = slice_idx;
980 start = ConvI2X(start);
981 if (scale != 0)
982 start = transform_later(new LShiftXNode( start, intcon(scale) ));
983 start = transform_later(new AddXNode(start, MakeConX(abase)) );
984 if ((bump_bit | clear_low) != 0) {
985 int to_clear = (bump_bit | clear_low);
986 // Align up mod 8, then store a jint zero unconditionally
987 // just before the mod-8 boundary.
988 if (((abase + bump_bit) & ~to_clear) - bump_bit
989 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
990 bump_bit = 0;
991 assert((abase & to_clear) == 0, "array base must be long-aligned");
992 } else {
993 // Bump 'start' up to (or past) the next jint boundary:
994 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
995 assert((abase & clear_low) == 0, "array base must be int-aligned");
996 }
997 // Round bumped 'start' down to jlong boundary in body of array.
998 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
999 if (bump_bit != 0) {
1000 // Store a zero to the immediately preceding jint:
1001 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1002 Node* p1 = basic_plus_adr(dest, x1);
1003 if (val == NULL) {
1004 assert(raw_val == NULL, "val may not be null");
1005 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1006 } else {
1007 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1008 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1009 }
1010 mem = transform_later(mem);
1011 }
1012 }
1013 Node* end = dest_size; // pre-rounded
1014 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1015 start, end, &_igvn);
1016 } else {
1017 // Non-constant start, unrounded non-constant end.
1018 // (Nobody zeroes a random midsection of an array using this routine.)
1019 ShouldNotReachHere(); // fix caller
1020 }
1021
1022 // Done.
1023 merge_mem->set_memory_at(alias_idx, mem);
1024 }
1025
1026 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1027 const TypePtr* adr_type,
1028 BasicType basic_elem_type,
1029 AllocateNode* alloc,
1030 Node* src, Node* src_offset,
1031 Node* dest, Node* dest_offset,
1032 Node* dest_size, bool dest_uninitialized) {
1033 // See if there is an advantage from block transfer.
1034 int scale = exact_log2(type2aelembytes(basic_elem_type));
1110 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1111 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1112 "slow_arraycopy", TypePtr::BOTTOM);
1113
1114 call->init_req(TypeFunc::Control, *ctrl);
1115 call->init_req(TypeFunc::I_O , *io);
1116 call->init_req(TypeFunc::Memory , mem);
1117 call->init_req(TypeFunc::ReturnAdr, top());
1118 call->init_req(TypeFunc::FramePtr, top());
1119 call->init_req(TypeFunc::Parms+0, src);
1120 call->init_req(TypeFunc::Parms+1, src_offset);
1121 call->init_req(TypeFunc::Parms+2, dest);
1122 call->init_req(TypeFunc::Parms+3, dest_offset);
1123 call->init_req(TypeFunc::Parms+4, copy_length);
1124 call->copy_call_debug_info(&_igvn, ac);
1125
1126 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1127 _igvn.replace_node(ac, call);
1128 transform_later(call);
1129
1130 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1131 *ctrl = _callprojs->fallthrough_catchproj->clone();
1132 transform_later(*ctrl);
1133
1134 Node* m = _callprojs->fallthrough_memproj->clone();
1135 transform_later(m);
1136
1137 uint alias_idx = C->get_alias_index(adr_type);
1138 MergeMemNode* out_mem;
1139 if (alias_idx != Compile::AliasIdxBot) {
1140 out_mem = MergeMemNode::make(mem);
1141 out_mem->set_memory_at(alias_idx, m);
1142 } else {
1143 out_mem = MergeMemNode::make(m);
1144 }
1145 transform_later(out_mem);
1146
1147 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1148 // could be NULL. Skip clone and update NULL fallthrough_ioproj.
1149 if (_callprojs->fallthrough_ioproj != NULL) {
1150 *io = _callprojs->fallthrough_ioproj->clone();
1151 transform_later(*io);
1152 } else {
1153 *io = NULL;
1154 }
1155
1156 return out_mem;
1157 }
1158
1159 // Helper function; generates code for cases requiring runtime checks.
1160 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1161 const TypePtr* adr_type,
1162 Node* dest_elem_klass,
1163 Node* src, Node* src_offset,
1164 Node* dest, Node* dest_offset,
1165 Node* copy_length, bool dest_uninitialized) {
1166 if ((*ctrl)->is_top()) return NULL;
1167
1168 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1169 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1170 return NULL;
1262 if (exit_block) {
1263 exit_block->init_req(2, *ctrl);
1264
1265 // Memory edge corresponding to stub_region.
1266 result_memory->init_req(2, *mem);
1267
1268 uint alias_idx = C->get_alias_index(adr_type);
1269 if (alias_idx != Compile::AliasIdxBot) {
1270 *mem = MergeMemNode::make(*mem);
1271 (*mem)->set_memory_at(alias_idx, result_memory);
1272 } else {
1273 *mem = MergeMemNode::make(result_memory);
1274 }
1275 transform_later(*mem);
1276 *ctrl = exit_block;
1277 return true;
1278 }
1279 return false;
1280 }
1281
1282 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1283 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1284 Node*& dest_length) {
1285 #ifdef ASSERT
1286 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1287 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1288 bs->array_copy_requires_gc_barriers(dest_length != NULL, T_OBJECT, false, false, BarrierSetC2::Optimization);
1289 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1290 #endif
1291 int elem_size = top_dest->flat_elem_size();
1292 if (elem_size >= 8) {
1293 if (elem_size > 8) {
1294 // treat as array of long but scale length, src offset and dest offset
1295 assert((elem_size % 8) == 0, "not a power of 2?");
1296 int factor = elem_size / 8;
1297 length = transform_later(new MulINode(length, intcon(factor)));
1298 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1299 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1300 if (dest_length != NULL) {
1301 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1302 }
1303 elem_size = 8;
1304 }
1305 dest_elem = T_LONG;
1306 } else if (elem_size == 4) {
1307 dest_elem = T_INT;
1308 } else if (elem_size == 2) {
1309 dest_elem = T_CHAR;
1310 } else if (elem_size == 1) {
1311 dest_elem = T_BYTE;
1312 } else {
1313 ShouldNotReachHere();
1314 }
1315 return TypeRawPtr::BOTTOM;
1316 }
1317
1318 #undef XTOP
1319
1320 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1321 Node* ctrl = ac->in(TypeFunc::Control);
1322 Node* io = ac->in(TypeFunc::I_O);
1323 Node* src = ac->in(ArrayCopyNode::Src);
1324 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1325 Node* dest = ac->in(ArrayCopyNode::Dest);
1326 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1327 Node* length = ac->in(ArrayCopyNode::Length);
1328 MergeMemNode* merge_mem = NULL;
1329
1330 if (ac->is_clonebasic()) {
1331 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1332 bs->clone_at_expansion(this, ac);
1333 return;
1334 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1335 const Type* src_type = _igvn.type(src);
1336 const Type* dest_type = _igvn.type(dest);
1337 const TypeAryPtr* top_src = src_type->isa_aryptr();
1338 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1339 BasicType dest_elem = T_OBJECT;
1340 if (top_dest != NULL && top_dest->elem() != Type::BOTTOM) {
1341 dest_elem = top_dest->elem()->array_element_basic_type();
1342 }
1343 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1344
1345 if (top_src != NULL && top_src->is_flat()) {
1346 // If src is flat, dest is guaranteed to be flat as well
1347 top_dest = top_src;
1348 }
1349
1350 AllocateArrayNode* alloc = NULL;
1351 Node* dest_length = NULL;
1352 if (ac->is_alloc_tightly_coupled()) {
1353 alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1354 assert(alloc != NULL, "expect alloc");
1355 dest_length = alloc->in(AllocateNode::ALength);
1356 }
1357
1358 Node* mem = ac->in(TypeFunc::Memory);
1359 const TypePtr* adr_type = NULL;
1360 if (top_dest->is_flat()) {
1361 assert(dest_length != NULL || StressReflectiveCode, "must be tightly coupled");
1362 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1363 // on all slices to prevent writes into the source from floating below the arraycopy.
1364 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1365 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1366 } else {
1367 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1368 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1369 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1370 }
1371 if (ac->_src_type != ac->_dest_type) {
1372 adr_type = TypeRawPtr::BOTTOM;
1373 }
1374 }
1375 merge_mem = MergeMemNode::make(mem);
1376 transform_later(merge_mem);
1377
1378 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1379 adr_type, dest_elem,
1380 src, src_offset, dest, dest_offset, length,
1381 dest_length,
1382 true, !ac->is_copyofrange());
1383 return;
1384 }
1385
1386 AllocateArrayNode* alloc = NULL;
1387 if (ac->is_alloc_tightly_coupled()) {
1388 alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1389 assert(alloc != NULL, "expect alloc");
1390 }
1391
1392 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1393
1394 // Compile time checks. If any of these checks cannot be verified at compile time,
1395 // we do not make a fast path for this call. Instead, we let the call remain as it
1396 // is. The checks we choose to mandate at compile time are:
1397 //
1398 // (1) src and dest are arrays.
1399 const Type* src_type = src->Value(&_igvn);
1400 const Type* dest_type = dest->Value(&_igvn);
1401 const TypeAryPtr* top_src = src_type->isa_aryptr();
1402 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1403
1404 BasicType src_elem = T_CONFLICT;
1405 BasicType dest_elem = T_CONFLICT;
1406
1407 if (top_src != NULL && top_src->elem() != Type::BOTTOM) {
1408 src_elem = top_src->elem()->array_element_basic_type();
1409 }
1410 if (top_dest != NULL && top_dest->elem() != Type::BOTTOM) {
1411 dest_elem = top_dest->elem()->array_element_basic_type();
1412 }
1413 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1414 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1415
1416 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1417 src_elem = dest_elem;
1418 }
1419
1420 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1421 // Conservatively insert a memory barrier on all memory slices.
1422 // Do not let writes into the source float below the arraycopy.
1423 {
1424 Node* mem = ac->in(TypeFunc::Memory);
1425 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1426
1427 merge_mem = MergeMemNode::make(mem);
1428 transform_later(merge_mem);
1429 }
1430
1431 // Call StubRoutines::generic_arraycopy stub.
1432 Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1433 TypeRawPtr::BOTTOM, T_CONFLICT,
1434 src, src_offset, dest, dest_offset, length,
1435 NULL,
1436 // If a negative length guard was generated for the ArrayCopyNode,
1437 // the length of the array can never be negative.
1438 false, ac->has_negative_length_guard());
1439 return;
1440 }
1441
1442 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1443
1444 // (2) src and dest arrays must have elements of the same BasicType
1445 // Figure out the size and type of the elements we will be copying.
1446 //
1447 // We have no stub to copy flattened inline type arrays with oop
1448 // fields if we need to emit write barriers.
1449 //
1450 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1451 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1452 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1453 bs->array_copy_requires_gc_barriers(alloc != NULL, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1454 // The component types are not the same or are not recognized. Punt.
1455 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1456 {
1457 Node* mem = ac->in(TypeFunc::Memory);
1458 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1459 }
1460
1461 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1462 if (_callprojs->fallthrough_ioproj != NULL) {
1463 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1464 }
1465 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1466 return;
1467 }
1468
1469 //---------------------------------------------------------------------------
1470 // We will make a fast path for this call to arraycopy.
1471
1472 // We have the following tests left to perform:
1473 //
1474 // (3) src and dest must not be null.
1475 // (4) src_offset must not be negative.
1476 // (5) dest_offset must not be negative.
1477 // (6) length must not be negative.
1478 // (7) src_offset + length must not exceed length of src.
1479 // (8) dest_offset + length must not exceed length of dest.
1480 // (9) each element of an oop array must be assignable
1481
1482 Node* mem = ac->in(TypeFunc::Memory);
1483 if (top_dest->is_flat()) {
1484 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1485 // on all slices to prevent writes into the source from floating below the arraycopy.
1486 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1487 }
1488 merge_mem = MergeMemNode::make(mem);
1489 transform_later(merge_mem);
1490
1491 RegionNode* slow_region = new RegionNode(1);
1492 transform_later(slow_region);
1493
1494 if (!ac->is_arraycopy_validated()) {
1495 // (3) operands must not be null
1496 // We currently perform our null checks with the null_check routine.
1497 // This means that the null exceptions will be reported in the caller
1498 // rather than (correctly) reported inside of the native arraycopy call.
1499 // This should be corrected, given time. We do our null check with the
1500 // stack pointer restored.
1501 // null checks done library_call.cpp
1502
1503 // (4) src_offset must not be negative.
1504 generate_negative_guard(&ctrl, src_offset, slow_region);
1505
1506 // (5) dest_offset must not be negative.
1507 generate_negative_guard(&ctrl, dest_offset, slow_region);
1508
1509 // (6) length must not be negative (moved to generate_arraycopy()).
1510 // generate_negative_guard(length, slow_region);
1511
1512 // (7) src_offset + length must not exceed length of src.
1513 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1514 assert(alen != NULL, "need src len");
1515 generate_limit_guard(&ctrl,
1516 src_offset, length,
1517 alen,
1518 slow_region);
1519
1520 // (8) dest_offset + length must not exceed length of dest.
1521 alen = ac->in(ArrayCopyNode::DestLen);
1522 assert(alen != NULL, "need dest len");
1523 generate_limit_guard(&ctrl,
1524 dest_offset, length,
1525 alen,
1526 slow_region);
1527
1528 // (9) each element of an oop array must be assignable
1529 // The generate_arraycopy subroutine checks this.
1530
1531 // Handle inline type arrays
1532 if (!top_src->is_flat()) {
1533 if (UseFlatArray && !top_src->is_not_flat()) {
1534 // Src might be flat and dest might not be flat. Go to the slow path if src is flat.
1535 generate_flat_array_guard(&ctrl, src, slow_region);
1536 }
1537 if (EnableValhalla) {
1538 // No validation. The subtype check emitted at macro expansion time will not go to the slow
1539 // path but call checkcast_arraycopy which can not handle flat/null-free inline type arrays.
1540 generate_null_free_array_guard(&ctrl, dest, slow_region);
1541 }
1542 } else {
1543 assert(top_dest->is_flat(), "dest array must be flat");
1544 }
1545 }
1546
1547 // This is where the memory effects are placed:
1548 const TypePtr* adr_type = NULL;
1549 Node* dest_length = (alloc != NULL) ? alloc->in(AllocateNode::ALength) : NULL;
1550
1551 if (top_dest->is_flat()) {
1552 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1553 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1554 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1555 } else {
1556 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1557 }
1558
1559 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1560 adr_type, dest_elem,
1561 src, src_offset, dest, dest_offset, length,
1562 dest_length,
1563 // If a negative length guard was generated for the ArrayCopyNode,
1564 // the length of the array can never be negative.
1565 false, ac->has_negative_length_guard(),
1566 slow_region);
1567 }
|