1 /*
2 * Copyright (c) 2012, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.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);
46 transform_later(*ctrl);
47 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
48 transform_later(mem_proj);
49 *mem = mem_proj;
50 }
51
52 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
53 uint shift = exact_log2(type2aelembytes(elembt));
54 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
55 Node* base = basic_plus_adr(ary, header);
56 #ifdef _LP64
57 // see comment in GraphKit::array_element_address
58 int index_max = max_jint - 1; // array size is max_jint, index is one less
59 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
60 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
61 #endif
62 Node* scale = new LShiftXNode(idx, intcon(shift));
63 transform_later(scale);
64 return basic_plus_adr(ary, base, scale);
65 }
66
67 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
68 return transform_later(new ConvI2LNode(offset));
69 }
70
71 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
72 const TypeFunc* call_type, address call_addr,
73 const char* call_name,
74 const TypePtr* adr_type,
75 Node* parm0, Node* parm1,
76 Node* parm2, Node* parm3,
77 Node* parm4, Node* parm5,
78 Node* parm6, Node* parm7) {
79 Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
80 call->init_req(TypeFunc::Control, ctrl);
81 call->init_req(TypeFunc::I_O , top());
82 call->init_req(TypeFunc::Memory , mem);
83 call->init_req(TypeFunc::ReturnAdr, top());
84 call->init_req(TypeFunc::FramePtr, top());
85
86 // Hook each parm in order. Stop looking at the first null.
87 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
88 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
89 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
90 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
91 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
92 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
93 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
94 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
95 /* close each nested if ===> */ } } } } } } } }
96 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
97
98 return call;
99 }
100
101
102 //------------------------------generate_guard---------------------------
103 // Helper function for generating guarded fast-slow graph structures.
104 // The given 'test', if true, guards a slow path. If the test fails
105 // then a fast path can be taken. (We generally hope it fails.)
106 // In all cases, GraphKit::control() is updated to the fast path.
107 // The returned value represents the control for the slow path.
108 // The return value is never 'top'; it is either a valid control
109 // or null if it is obvious that the slow path can never be taken.
110 // Also, if region and the slow control are not null, the slow edge
111 // is appended to the region.
112 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
113 if ((*ctrl)->is_top()) {
114 // Already short circuited.
115 return nullptr;
116 }
117 // Build an if node and its projections.
118 // If test is true we take the slow path, which we assume is uncommon.
119 if (_igvn.type(test) == TypeInt::ZERO) {
120 // The slow branch is never taken. No need to build this guard.
121 return nullptr;
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 != nullptr) {
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 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 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
147 return generate_guard(ctrl, test, region, PROB_FAIR);
148 }
149
150 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
151 if ((*ctrl)->is_top())
152 return; // already stopped
153 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
154 return; // index is already adequately typed
155 Node* cmp_lt = new CmpINode(index, intcon(0));
156 transform_later(cmp_lt);
157 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
158 transform_later(bol_lt);
159 generate_guard(ctrl, bol_lt, region, PROB_MIN);
160 }
161
162 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
163 if ((*ctrl)->is_top())
164 return; // already stopped
165 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
166 if (zero_offset && subseq_length->eqv_uncast(array_length))
167 return; // common case of whole-array copy
168 Node* last = subseq_length;
169 if (!zero_offset) { // last += offset
170 last = new AddINode(last, offset);
171 transform_later(last);
172 }
173 Node* cmp_lt = new CmpUNode(array_length, last);
174 transform_later(cmp_lt);
175 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
176 transform_later(bol_lt);
177 generate_guard(ctrl, bol_lt, region, PROB_MIN);
178 }
179
180 //
181 // Partial in-lining handling for smaller conjoint/disjoint array copies having
182 // length(in bytes) less than ArrayOperationPartialInlineSize.
183 // if (length <= ArrayOperationPartialInlineSize) {
184 // partial_inlining_block:
185 // mask = Mask_Gen
186 // vload = LoadVectorMasked src , mask
187 // StoreVectorMasked dst, mask, vload
188 // } else {
189 // stub_block:
190 // callstub array_copy
191 // }
192 // exit_block:
193 // Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
194 // mem = MergeMem (Phi)
195 // control = stub_block
196 //
197 // Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
198 // edges. Remaining edges for exit_block coming from stub_block are connected by the caller
199 // post stub nodes creation.
200 //
201
202 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
203 RegionNode** exit_block, Node** result_memory, Node* length,
204 Node* src_start, Node* dst_start, BasicType type) {
205 const TypePtr *src_adr_type = _igvn.type(src_start)->isa_ptr();
206 Node* inline_block = nullptr;
207 Node* stub_block = nullptr;
208
209 int const_len = -1;
210 const TypeInt* lty = nullptr;
211 uint shift = exact_log2(type2aelembytes(type));
212 if (length->Opcode() == Op_ConvI2L) {
213 lty = _igvn.type(length->in(1))->isa_int();
214 } else {
215 lty = _igvn.type(length)->isa_int();
216 }
217 if (lty && lty->is_con()) {
218 const_len = lty->get_con() << shift;
219 }
220
221 // Return if copy length is greater than partial inline size limit or
222 // target does not supports masked load/stores.
223 int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, const_len);
224 if ( const_len > ArrayOperationPartialInlineSize ||
225 !Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type) ||
226 !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
227 !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
228 return;
229 }
230
231 int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
232 Node* casted_length = new CastLLNode(*ctrl, length, TypeLong::make(0, inline_limit, Type::WidenMin));
233 transform_later(casted_length);
234 Node* copy_bytes = new LShiftXNode(length, intcon(shift));
235 transform_later(copy_bytes);
236
237 Node* cmp_le = new CmpULNode(copy_bytes, longcon(ArrayOperationPartialInlineSize));
238 transform_later(cmp_le);
239 Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
240 transform_later(bol_le);
241 inline_block = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
242 stub_block = *ctrl;
243
244 Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
245 transform_later(mask_gen);
246
247 unsigned vec_size = lane_count * type2aelembytes(type);
248 if (C->max_vector_size() < vec_size) {
249 C->set_max_vector_size(vec_size);
250 }
251
252 const TypeVect * vt = TypeVect::make(type, lane_count);
253 Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
254 Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
255 src_adr_type, vt, mask_gen);
256 transform_later(masked_load);
257
258 mm = (*mem)->memory_at(C->get_alias_index(adr_type));
259 Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
260 masked_load, adr_type, mask_gen);
261 transform_later(masked_store);
262
263 // Convergence region for inline_block and stub_block.
264 *exit_block = new RegionNode(3);
265 transform_later(*exit_block);
266 (*exit_block)->init_req(1, inline_block);
267 *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
268 transform_later(*result_memory);
269 (*result_memory)->init_req(1, masked_store);
270
271 *ctrl = stub_block;
272 }
273
274
275 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
276 if ((*ctrl)->is_top()) return nullptr;
277
278 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
279 return nullptr; // index is already adequately typed
280 Node* cmp_le = new CmpINode(index, intcon(0));
281 transform_later(cmp_le);
282 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
283 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
284 transform_later(bol_le);
285 Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
286
287 return is_notp;
288 }
289
290 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
291 // Load markword and check if obj is locked
292 Node* mark = make_load(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
293 Node* locked_bit = MakeConX(markWord::unlocked_value);
294 locked_bit = transform_later(new AndXNode(locked_bit, mark));
295 Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
296 Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
297 IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
298 Node* locked_region = transform_later(new RegionNode(3));
299 Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
300
301 // Unlocked: Use bits from mark word
302 locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
303 mark_phi->init_req(1, mark);
304
305 // Locked: Load prototype header from klass
306 *ctrl = transform_later(new IfFalseNode(iff));
307 // Make loads control dependent to make sure they are only executed if array is locked
308 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
309 Node* klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
310 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
311 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));
312
313 locked_region->init_req(2, *ctrl);
314 mark_phi->init_req(2, proto);
315 *ctrl = locked_region;
316
317 // Now check if mark word bits are set
318 Node* mask = MakeConX(mask_val);
319 Node* masked = transform_later(new AndXNode(mark_phi, mask));
320 cmp = transform_later(new CmpXNode(masked, mask));
321 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
322 return generate_fair_guard(ctrl, bol, region);
323 }
324
325 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
326 return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
327 }
328
329 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
330 return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
331 }
332
333 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
334 transform_later(call);
335
336 *ctrl = new ProjNode(call,TypeFunc::Control);
337 transform_later(*ctrl);
338 Node* newmem = new ProjNode(call, TypeFunc::Memory);
339 transform_later(newmem);
340
341 uint alias_idx = C->get_alias_index(adr_type);
342 if (alias_idx != Compile::AliasIdxBot) {
343 *mem = MergeMemNode::make(*mem);
344 (*mem)->set_memory_at(alias_idx, newmem);
345 } else {
346 *mem = MergeMemNode::make(newmem);
347 }
348 transform_later(*mem);
349 }
350
351 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
352 Node* src_offset,
353 Node* dest_offset,
354 bool disjoint_bases,
355 const char* &name,
356 bool dest_uninitialized) {
357 const TypeInt* src_offset_inttype = _igvn.find_int_type(src_offset);
358 const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
359
360 bool aligned = false;
361 bool disjoint = disjoint_bases;
362
363 // if the offsets are the same, we can treat the memory regions as
364 // disjoint, because either the memory regions are in different arrays,
365 // or they are identical (which we can treat as disjoint.) We can also
366 // treat a copy with a destination index less that the source index
367 // as disjoint since a low->high copy will work correctly in this case.
368 if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
369 dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
370 // both indices are constants
371 int s_offs = src_offset_inttype->get_con();
372 int d_offs = dest_offset_inttype->get_con();
373 int element_size = type2aelembytes(t);
374 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
375 ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
376 if (s_offs >= d_offs) disjoint = true;
377 } else if (src_offset == dest_offset && src_offset != nullptr) {
378 // This can occur if the offsets are identical non-constants.
379 disjoint = true;
380 }
381
382 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
383 }
384
385 #define XTOP LP64_ONLY(COMMA top())
386
387 // Generate an optimized call to arraycopy.
388 // Caller must guard against non-arrays.
389 // Caller must determine a common array basic-type for both arrays.
390 // Caller must validate offsets against array bounds.
391 // The slow_region has already collected guard failure paths
392 // (such as out of bounds length or non-conformable array types).
393 // The generated code has this shape, in general:
394 //
395 // if (length == 0) return // via zero_path
396 // slowval = -1
397 // if (types unknown) {
398 // slowval = call generic copy loop
399 // if (slowval == 0) return // via checked_path
400 // } else if (indexes in bounds) {
401 // if ((is object array) && !(array type check)) {
402 // slowval = call checked copy loop
403 // if (slowval == 0) return // via checked_path
404 // } else {
405 // call bulk copy loop
406 // return // via fast_path
407 // }
408 // }
409 // // adjust params for remaining work:
410 // if (slowval != -1) {
411 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
412 // }
413 // slow_region:
414 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
415 // return // via slow_call_path
416 //
417 // This routine is used from several intrinsics: System.arraycopy,
418 // Object.clone (the array subcase), and Arrays.copyOf[Range].
419 //
420 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
421 Node** ctrl, MergeMemNode* mem, Node** io,
422 const TypePtr* adr_type,
423 BasicType basic_elem_type,
424 Node* src, Node* src_offset,
425 Node* dest, Node* dest_offset,
426 Node* copy_length,
427 Node* dest_length,
428 bool disjoint_bases,
429 bool length_never_negative,
430 RegionNode* slow_region) {
431 if (slow_region == nullptr) {
432 slow_region = new RegionNode(1);
433 transform_later(slow_region);
434 }
435
436 Node* original_dest = dest;
437 bool dest_needs_zeroing = false;
438 bool acopy_to_uninitialized = false;
439 Node* init_value = nullptr;
440 Node* raw_init_value = nullptr;
441
442 // See if this is the initialization of a newly-allocated array.
443 // If so, we will take responsibility here for initializing it to zero.
444 // (Note: Because tightly_coupled_allocation performs checks on the
445 // out-edges of the dest, we need to avoid making derived pointers
446 // from it until we have checked its uses.)
447 if (ReduceBulkZeroing
448 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
449 && basic_elem_type != T_CONFLICT // avoid corner case
450 && !src->eqv_uncast(dest)
451 && alloc != nullptr
452 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
453 assert(ac->is_alloc_tightly_coupled(), "sanity");
454 // acopy to uninitialized tightly coupled allocations
455 // needs zeroing outside the copy range
456 // and the acopy itself will be to uninitialized memory
457 acopy_to_uninitialized = true;
458 if (alloc->maybe_set_complete(&_igvn)) {
459 // "You break it, you buy it."
460 InitializeNode* init = alloc->initialization();
461 assert(init->is_complete(), "we just did this");
462 init->set_complete_with_arraycopy();
463 assert(dest->is_CheckCastPP(), "sanity");
464 assert(dest->in(0)->in(0) == init, "dest pinned");
465 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
466 // From this point on, every exit path is responsible for
467 // initializing any non-copied parts of the object to zero.
468 // Also, if this flag is set we make sure that arraycopy interacts properly
469 // with G1, eliding pre-barriers. See CR 6627983.
470 dest_needs_zeroing = true;
471 init_value = alloc->in(AllocateNode::InitValue);
472 raw_init_value = alloc->in(AllocateNode::RawInitValue);
473 } else {
474 // dest_need_zeroing = false;
475 }
476 } else {
477 // No zeroing elimination needed here.
478 alloc = nullptr;
479 acopy_to_uninitialized = false;
480 //original_dest = dest;
481 //dest_needs_zeroing = false;
482 }
483
484 uint alias_idx = C->get_alias_index(adr_type);
485
486 // Results are placed here:
487 enum { fast_path = 1, // normal void-returning assembly stub
488 checked_path = 2, // special assembly stub with cleanup
489 slow_call_path = 3, // something went wrong; call the VM
490 zero_path = 4, // bypass when length of copy is zero
491 bcopy_path = 5, // copy primitive array by 64-bit blocks
492 PATH_LIMIT = 6
493 };
494 RegionNode* result_region = new RegionNode(PATH_LIMIT);
495 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
496 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
497 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
498 transform_later(result_region);
499 transform_later(result_i_o);
500 transform_later(result_memory);
501
502 // The slow_control path:
503 Node* slow_control;
504 Node* slow_i_o = *io;
505 Node* slow_mem = mem->memory_at(alias_idx);
506 DEBUG_ONLY(slow_control = (Node*) badAddress);
507
508 // Checked control path:
509 Node* checked_control = top();
510 Node* checked_mem = nullptr;
511 Node* checked_i_o = nullptr;
512 Node* checked_value = nullptr;
513
514 if (basic_elem_type == T_CONFLICT) {
515 assert(!dest_needs_zeroing, "");
516 Node* cv = generate_generic_arraycopy(ctrl, &mem,
517 adr_type,
518 src, src_offset, dest, dest_offset,
519 copy_length, acopy_to_uninitialized);
520 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
521 checked_control = *ctrl;
522 checked_i_o = *io;
523 checked_mem = mem->memory_at(alias_idx);
524 checked_value = cv;
525 *ctrl = top();
526 }
527
528 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
529 if (not_pos != nullptr) {
530 Node* local_ctrl = not_pos, *local_io = *io;
531 MergeMemNode* local_mem = MergeMemNode::make(mem);
532 transform_later(local_mem);
533
534 // (6) length must not be negative.
535 if (!length_never_negative) {
536 generate_negative_guard(&local_ctrl, copy_length, slow_region);
537 }
538
539 // copy_length is 0.
540 if (dest_needs_zeroing) {
541 assert(!local_ctrl->is_top(), "no ctrl?");
542 if (copy_length->eqv_uncast(dest_length)
543 || _igvn.find_int_con(dest_length, 1) <= 0) {
544 // There is no zeroing to do. No need for a secondary raw memory barrier.
545 } else {
546 // Clear the whole thing since there are no source elements to copy.
547 generate_clear_array(local_ctrl, local_mem,
548 adr_type, dest,
549 init_value, raw_init_value,
550 basic_elem_type,
551 intcon(0), nullptr,
552 alloc->in(AllocateNode::AllocSize));
553 // Use a secondary InitializeNode as raw memory barrier.
554 // Currently it is needed only on this path since other
555 // paths have stub or runtime calls as raw memory barriers.
556 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
557 Compile::AliasIdxRaw,
558 top());
559 transform_later(mb);
560 mb->set_req(TypeFunc::Control,local_ctrl);
561 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
562 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
563 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
564
565 InitializeNode* init = mb->as_Initialize();
566 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
567 }
568 }
569
570 // Present the results of the fast call.
571 result_region->init_req(zero_path, local_ctrl);
572 result_i_o ->init_req(zero_path, local_io);
573 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
574 }
575
576 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
577 // We have to initialize the *uncopied* part of the array to zero.
578 // The copy destination is the slice dest[off..off+len]. The other slices
579 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
580 Node* dest_size = alloc->in(AllocateNode::AllocSize);
581 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
582
583 // If there is a head section that needs zeroing, do it now.
584 if (_igvn.find_int_con(dest_offset, -1) != 0) {
585 generate_clear_array(*ctrl, mem,
586 adr_type, dest,
587 init_value, raw_init_value,
588 basic_elem_type,
589 intcon(0), dest_offset,
590 nullptr);
591 }
592
593 // Next, perform a dynamic check on the tail length.
594 // It is often zero, and we can win big if we prove this.
595 // There are two wins: Avoid generating the ClearArray
596 // with its attendant messy index arithmetic, and upgrade
597 // the copy to a more hardware-friendly word size of 64 bits.
598 Node* tail_ctl = nullptr;
599 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
600 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
601 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
602 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
603 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
604 }
605
606 // At this point, let's assume there is no tail.
607 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
608 // There is no tail. Try an upgrade to a 64-bit copy.
609 bool didit = false;
610 {
611 Node* local_ctrl = *ctrl, *local_io = *io;
612 MergeMemNode* local_mem = MergeMemNode::make(mem);
613 transform_later(local_mem);
614
615 didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
616 adr_type, basic_elem_type, alloc,
617 src, src_offset, dest, dest_offset,
618 dest_size, acopy_to_uninitialized);
619 if (didit) {
620 // Present the results of the block-copying fast call.
621 result_region->init_req(bcopy_path, local_ctrl);
622 result_i_o ->init_req(bcopy_path, local_io);
623 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
624 }
625 }
626 if (didit) {
627 *ctrl = top(); // no regular fast path
628 }
629 }
630
631 // Clear the tail, if any.
632 if (tail_ctl != nullptr) {
633 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
634 *ctrl = tail_ctl;
635 if (notail_ctl == nullptr) {
636 generate_clear_array(*ctrl, mem,
637 adr_type, dest,
638 init_value, raw_init_value,
639 basic_elem_type,
640 dest_tail, nullptr,
641 dest_size);
642 } else {
643 // Make a local merge.
644 Node* done_ctl = transform_later(new RegionNode(3));
645 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
646 done_ctl->init_req(1, notail_ctl);
647 done_mem->init_req(1, mem->memory_at(alias_idx));
648 generate_clear_array(*ctrl, mem,
649 adr_type, dest,
650 init_value, raw_init_value,
651 basic_elem_type,
652 dest_tail, nullptr,
653 dest_size);
654 done_ctl->init_req(2, *ctrl);
655 done_mem->init_req(2, mem->memory_at(alias_idx));
656 *ctrl = done_ctl;
657 mem->set_memory_at(alias_idx, done_mem);
658 }
659 }
660 }
661
662 BasicType copy_type = basic_elem_type;
663 assert(basic_elem_type != T_ARRAY, "caller must fix this");
664 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
665 // If src and dest have compatible element types, we can copy bits.
666 // Types S[] and D[] are compatible if D is a supertype of S.
667 //
668 // If they are not, we will use checked_oop_disjoint_arraycopy,
669 // which performs a fast optimistic per-oop check, and backs off
670 // further to JVM_ArrayCopy on the first per-oop check that fails.
671 // (Actually, we don't move raw bits only; the GC requires card marks.)
672
673 // We don't need a subtype check for validated copies and Object[].clone()
674 bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
675 ac->is_copyofrange_validated() || ac->is_clone_oop_array();
676 if (!skip_subtype_check) {
677 // Get the klass* for both src and dest
678 Node* src_klass = ac->in(ArrayCopyNode::SrcKlass);
679 Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
680
681 assert(src_klass != nullptr && dest_klass != nullptr, "should have klasses");
682
683 // Generate the subtype check.
684 // This might fold up statically, or then again it might not.
685 //
686 // Non-static example: Copying List<String>.elements to a new String[].
687 // The backing store for a List<String> is always an Object[],
688 // but its elements are always type String, if the generic types
689 // are correct at the source level.
690 //
691 // Test S[] against D[], not S against D, because (probably)
692 // the secondary supertype cache is less busy for S[] than S.
693 // This usually only matters when D is an interface.
694 Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn, nullptr, -1);
695 // Plug failing path into checked_oop_disjoint_arraycopy
696 if (not_subtype_ctrl != top()) {
697 Node* local_ctrl = not_subtype_ctrl;
698 MergeMemNode* local_mem = MergeMemNode::make(mem);
699 transform_later(local_mem);
700
701 // (At this point we can assume disjoint_bases, since types differ.)
702 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
703 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
704 Node* n1 = LoadKlassNode::make(_igvn, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
705 Node* dest_elem_klass = transform_later(n1);
706 Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
707 adr_type,
708 dest_elem_klass,
709 src, src_offset, dest, dest_offset,
710 ConvI2X(copy_length), acopy_to_uninitialized);
711 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
712 checked_control = local_ctrl;
713 checked_i_o = *io;
714 checked_mem = local_mem->memory_at(alias_idx);
715 checked_value = cv;
716 }
717 }
718 // At this point we know we do not need type checks on oop stores.
719
720 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
721 if (!bs->array_copy_requires_gc_barriers(alloc != nullptr, copy_type, false, false, BarrierSetC2::Expansion)) {
722 // If we do not need gc barriers, copy using the jint or jlong stub.
723 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
724 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
725 "sizes agree");
726 }
727 }
728
729 bool is_partial_array_copy = false;
730 if (!(*ctrl)->is_top()) {
731 // Generate the fast path, if possible.
732 Node* local_ctrl = *ctrl;
733 MergeMemNode* local_mem = MergeMemNode::make(mem);
734 transform_later(local_mem);
735 is_partial_array_copy = generate_unchecked_arraycopy(&local_ctrl, &local_mem,
736 adr_type, copy_type, disjoint_bases,
737 src, src_offset, dest, dest_offset,
738 ConvI2X(copy_length), acopy_to_uninitialized);
739
740 // Present the results of the fast call.
741 result_region->init_req(fast_path, local_ctrl);
742 result_i_o ->init_req(fast_path, *io);
743 result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
744 }
745
746 // Here are all the slow paths up to this point, in one bundle:
747 assert(slow_region != nullptr, "allocated on entry");
748 slow_control = slow_region;
749 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
750
751 *ctrl = checked_control;
752 if (!(*ctrl)->is_top()) {
753 // Clean up after the checked call.
754 // The returned value is either 0 or -1^K,
755 // where K = number of partially transferred array elements.
756 Node* cmp = new CmpINode(checked_value, intcon(0));
757 transform_later(cmp);
758 Node* bol = new BoolNode(cmp, BoolTest::eq);
759 transform_later(bol);
760 IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
761 transform_later(iff);
762
763 // If it is 0, we are done, so transfer to the end.
764 Node* checks_done = new IfTrueNode(iff);
765 transform_later(checks_done);
766 result_region->init_req(checked_path, checks_done);
767 result_i_o ->init_req(checked_path, checked_i_o);
768 result_memory->init_req(checked_path, checked_mem);
769
770 // If it is not zero, merge into the slow call.
771 *ctrl = new IfFalseNode(iff);
772 transform_later(*ctrl);
773 RegionNode* slow_reg2 = new RegionNode(3);
774 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
775 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
776 transform_later(slow_reg2);
777 transform_later(slow_i_o2);
778 transform_later(slow_mem2);
779 slow_reg2 ->init_req(1, slow_control);
780 slow_i_o2 ->init_req(1, slow_i_o);
781 slow_mem2 ->init_req(1, slow_mem);
782 slow_reg2 ->init_req(2, *ctrl);
783 slow_i_o2 ->init_req(2, checked_i_o);
784 slow_mem2 ->init_req(2, checked_mem);
785
786 slow_control = slow_reg2;
787 slow_i_o = slow_i_o2;
788 slow_mem = slow_mem2;
789
790 if (alloc != nullptr) {
791 // We'll restart from the very beginning, after zeroing the whole thing.
792 // This can cause double writes, but that's OK since dest is brand new.
793 // So we ignore the low 31 bits of the value returned from the stub.
794 } else {
795 // We must continue the copy exactly where it failed, or else
796 // another thread might see the wrong number of writes to dest.
797 Node* checked_offset = new XorINode(checked_value, intcon(-1));
798 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
799 transform_later(checked_offset);
800 transform_later(slow_offset);
801 slow_offset->init_req(1, intcon(0));
802 slow_offset->init_req(2, checked_offset);
803
804 // Adjust the arguments by the conditionally incoming offset.
805 Node* src_off_plus = new AddINode(src_offset, slow_offset);
806 transform_later(src_off_plus);
807 Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
808 transform_later(dest_off_plus);
809 Node* length_minus = new SubINode(copy_length, slow_offset);
810 transform_later(length_minus);
811
812 // Tweak the node variables to adjust the code produced below:
813 src_offset = src_off_plus;
814 dest_offset = dest_off_plus;
815 copy_length = length_minus;
816 }
817 }
818 *ctrl = slow_control;
819 if (!(*ctrl)->is_top()) {
820 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
821 MergeMemNode* local_mem = MergeMemNode::make(mem);
822 transform_later(local_mem);
823
824 // Generate the slow path, if needed.
825 local_mem->set_memory_at(alias_idx, slow_mem);
826
827 if (dest_needs_zeroing) {
828 generate_clear_array(local_ctrl, local_mem,
829 adr_type, dest,
830 init_value, raw_init_value,
831 basic_elem_type,
832 intcon(0), nullptr,
833 alloc->in(AllocateNode::AllocSize));
834 }
835
836 local_mem = generate_slow_arraycopy(ac,
837 &local_ctrl, local_mem, &local_io,
838 adr_type,
839 src, src_offset, dest, dest_offset,
840 copy_length, /*dest_uninitialized*/false);
841
842 result_region->init_req(slow_call_path, local_ctrl);
843 result_i_o ->init_req(slow_call_path, local_io);
844 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
845 } else {
846 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
847 // projections were not extracted
848 }
849
850 // Remove unused edges.
851 for (uint i = 1; i < result_region->req(); i++) {
852 if (result_region->in(i) == nullptr) {
853 result_region->init_req(i, top());
854 }
855 }
856
857 // Finished; return the combined state.
858 *ctrl = result_region;
859 *io = result_i_o;
860 mem->set_memory_at(alias_idx, result_memory);
861
862 // mem no longer guaranteed to stay a MergeMemNode
863 Node* out_mem = mem;
864 DEBUG_ONLY(mem = nullptr);
865
866 // The memory edges above are precise in order to model effects around
867 // array copies accurately to allow value numbering of field loads around
868 // arraycopy. Such field loads, both before and after, are common in Java
869 // collections and similar classes involving header/array data structures.
870 //
871 // But with low number of register or when some registers are used or killed
872 // by arraycopy calls it causes registers spilling on stack. See 6544710.
873 // The next memory barrier is added to avoid it. If the arraycopy can be
874 // optimized away (which it can, sometimes) then we can manually remove
875 // the membar also.
876 //
877 // Do not let reads from the cloned object float above the arraycopy.
878 if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
879 // Do not let stores that initialize this object be reordered with
880 // a subsequent store that would make this object accessible by
881 // other threads.
882 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
883 } else {
884 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
885 }
886
887 if (is_partial_array_copy) {
888 assert((*ctrl)->is_Proj(), "MemBar control projection");
889 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
890 (*ctrl)->in(0)->isa_MemBar()->set_trailing_partial_array_copy();
891 }
892
893 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
894 if (_callprojs->fallthrough_ioproj != nullptr) {
895 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
896 }
897 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
898
899 #ifdef ASSERT
900 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
901 if (dest_t->is_known_instance() && !is_partial_array_copy) {
902 ArrayCopyNode* ac = nullptr;
903 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
904 assert(ac == nullptr, "no arraycopy anymore");
905 }
906 #endif
907
908 return out_mem;
909 }
910
911 // Helper for initialization of arrays, creating a ClearArray.
912 // It writes zero bits in [start..end), within the body of an array object.
913 // The memory effects are all chained onto the 'adr_type' alias category.
914 //
915 // Since the object is otherwise uninitialized, we are free
916 // to put a little "slop" around the edges of the cleared area,
917 // as long as it does not go back into the array's header,
918 // or beyond the array end within the heap.
919 //
920 // The lower edge can be rounded down to the nearest jint and the
921 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
922 //
923 // Arguments:
924 // adr_type memory slice where writes are generated
925 // dest oop of the destination array
926 // basic_elem_type element type of the destination
927 // slice_idx array index of first element to store
928 // slice_len number of elements to store (or null)
929 // dest_size total size in bytes of the array object
930 //
931 // Exactly one of slice_len or dest_size must be non-null.
932 // If dest_size is non-null, zeroing extends to the end of the object.
933 // If slice_len is non-null, the slice_idx value must be a constant.
934 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
935 const TypePtr* adr_type,
936 Node* dest,
937 Node* val,
938 Node* raw_val,
939 BasicType basic_elem_type,
940 Node* slice_idx,
941 Node* slice_len,
942 Node* dest_size) {
943 // one or the other but not both of slice_len and dest_size:
944 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
945 if (slice_len == nullptr) slice_len = top();
946 if (dest_size == nullptr) dest_size = top();
947
948 uint alias_idx = C->get_alias_index(adr_type);
949
950 // operate on this memory slice:
951 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
952
953 // scaling and rounding of indexes:
954 int scale = exact_log2(type2aelembytes(basic_elem_type));
955 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
956 int clear_low = (-1 << scale) & (BytesPerInt - 1);
957 int bump_bit = (-1 << scale) & BytesPerInt;
958
959 // determine constant starts and ends
960 const intptr_t BIG_NEG = -128;
961 assert(BIG_NEG + 2*abase < 0, "neg enough");
962 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
963 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
964 if (slice_len_con == 0) {
965 return; // nothing to do here
966 }
967 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
968 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
969 if (slice_idx_con >= 0 && slice_len_con >= 0) {
970 assert(end_con < 0, "not two cons");
971 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
972 BytesPerLong);
973 }
974
975 if (start_con >= 0 && end_con >= 0) {
976 // Constant start and end. Simple.
977 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
978 start_con, end_con, &_igvn);
979 } else if (start_con >= 0 && dest_size != top()) {
980 // Constant start, pre-rounded end after the tail of the array.
981 Node* end = dest_size;
982 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
983 start_con, end, &_igvn);
984 } else if (start_con >= 0 && slice_len != top()) {
985 // Constant start, non-constant end. End needs rounding up.
986 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
987 intptr_t end_base = abase + (slice_idx_con << scale);
988 int end_round = (-1 << scale) & (BytesPerLong - 1);
989 Node* end = ConvI2X(slice_len);
990 if (scale != 0)
991 end = transform_later(new LShiftXNode(end, intcon(scale) ));
992 end_base += end_round;
993 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
994 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
995 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
996 start_con, end, &_igvn);
997 } else if (start_con < 0 && dest_size != top()) {
998 // Non-constant start, pre-rounded end after the tail of the array.
999 // This is almost certainly a "round-to-end" operation.
1000 Node* start = slice_idx;
1001 start = ConvI2X(start);
1002 if (scale != 0)
1003 start = transform_later(new LShiftXNode( start, intcon(scale) ));
1004 start = transform_later(new AddXNode(start, MakeConX(abase)) );
1005 if ((bump_bit | clear_low) != 0) {
1006 int to_clear = (bump_bit | clear_low);
1007 // Align up mod 8, then store a jint zero unconditionally
1008 // just before the mod-8 boundary.
1009 if (((abase + bump_bit) & ~to_clear) - bump_bit
1010 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1011 bump_bit = 0;
1012 assert((abase & to_clear) == 0, "array base must be long-aligned");
1013 } else {
1014 // Bump 'start' up to (or past) the next jint boundary:
1015 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1016 assert((abase & clear_low) == 0, "array base must be int-aligned");
1017 }
1018 // Round bumped 'start' down to jlong boundary in body of array.
1019 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1020 if (bump_bit != 0) {
1021 // Store a zero to the immediately preceding jint:
1022 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1023 Node* p1 = basic_plus_adr(dest, x1);
1024 if (val == nullptr) {
1025 assert(raw_val == nullptr, "val may not be null");
1026 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1027 } else {
1028 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1029 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1030 }
1031 mem = transform_later(mem);
1032 }
1033 }
1034 Node* end = dest_size; // pre-rounded
1035 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1036 start, end, &_igvn);
1037 } else {
1038 // Non-constant start, unrounded non-constant end.
1039 // (Nobody zeroes a random midsection of an array using this routine.)
1040 ShouldNotReachHere(); // fix caller
1041 }
1042
1043 // Done.
1044 merge_mem->set_memory_at(alias_idx, mem);
1045 }
1046
1047 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1048 const TypePtr* adr_type,
1049 BasicType basic_elem_type,
1050 AllocateNode* alloc,
1051 Node* src, Node* src_offset,
1052 Node* dest, Node* dest_offset,
1053 Node* dest_size, bool dest_uninitialized) {
1054 // See if there is an advantage from block transfer.
1055 int scale = exact_log2(type2aelembytes(basic_elem_type));
1056 if (scale >= LogBytesPerLong)
1057 return false; // it is already a block transfer
1058
1059 // Look at the alignment of the starting offsets.
1060 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1061
1062 intptr_t src_off_con = (intptr_t) _igvn.find_int_con(src_offset, -1);
1063 intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1064 if (src_off_con < 0 || dest_off_con < 0) {
1065 // At present, we can only understand constants.
1066 return false;
1067 }
1068
1069 intptr_t src_off = abase + (src_off_con << scale);
1070 intptr_t dest_off = abase + (dest_off_con << scale);
1071
1072 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1073 // Non-aligned; too bad.
1074 // One more chance: Pick off an initial 32-bit word.
1075 // This is a common case, since abase can be odd mod 8.
1076 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1077 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1078 Node* sptr = basic_plus_adr(src, src_off);
1079 Node* dptr = basic_plus_adr(dest, dest_off);
1080 const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1081 assert(s_adr_type->isa_aryptr(), "impossible slice");
1082 uint s_alias_idx = C->get_alias_index(s_adr_type);
1083 uint d_alias_idx = C->get_alias_index(adr_type);
1084 bool is_mismatched = (basic_elem_type != T_INT);
1085 Node* sval = transform_later(
1086 LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1087 TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1088 false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1089 Node* st = transform_later(
1090 StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1091 sval, T_INT, MemNode::unordered));
1092 if (is_mismatched) {
1093 st->as_Store()->set_mismatched_access();
1094 }
1095 (*mem)->set_memory_at(d_alias_idx, st);
1096 src_off += BytesPerInt;
1097 dest_off += BytesPerInt;
1098 } else {
1099 return false;
1100 }
1101 }
1102 assert(src_off % BytesPerLong == 0, "");
1103 assert(dest_off % BytesPerLong == 0, "");
1104
1105 // Do this copy by giant steps.
1106 Node* sptr = basic_plus_adr(src, src_off);
1107 Node* dptr = basic_plus_adr(dest, dest_off);
1108 Node* countx = dest_size;
1109 countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1110 countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1111
1112 bool disjoint_bases = true; // since alloc isn't null
1113 generate_unchecked_arraycopy(ctrl, mem,
1114 adr_type, T_LONG, disjoint_bases,
1115 sptr, nullptr, dptr, nullptr, countx, dest_uninitialized);
1116
1117 return true;
1118 }
1119
1120 // Helper function; generates code for the slow case.
1121 // We make a call to a runtime method which emulates the native method,
1122 // but without the native wrapper overhead.
1123 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1124 Node** ctrl, Node* mem, Node** io,
1125 const TypePtr* adr_type,
1126 Node* src, Node* src_offset,
1127 Node* dest, Node* dest_offset,
1128 Node* copy_length, bool dest_uninitialized) {
1129 assert(!dest_uninitialized, "Invariant");
1130
1131 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1132 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1133 "slow_arraycopy", TypePtr::BOTTOM);
1134
1135 call->init_req(TypeFunc::Control, *ctrl);
1136 call->init_req(TypeFunc::I_O , *io);
1137 call->init_req(TypeFunc::Memory , mem);
1138 call->init_req(TypeFunc::ReturnAdr, top());
1139 call->init_req(TypeFunc::FramePtr, top());
1140 call->init_req(TypeFunc::Parms+0, src);
1141 call->init_req(TypeFunc::Parms+1, src_offset);
1142 call->init_req(TypeFunc::Parms+2, dest);
1143 call->init_req(TypeFunc::Parms+3, dest_offset);
1144 call->init_req(TypeFunc::Parms+4, copy_length);
1145 call->copy_call_debug_info(&_igvn, ac);
1146
1147 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1148 _igvn.replace_node(ac, call);
1149 transform_later(call);
1150
1151 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1152 *ctrl = _callprojs->fallthrough_catchproj->clone();
1153 transform_later(*ctrl);
1154
1155 Node* m = _callprojs->fallthrough_memproj->clone();
1156 transform_later(m);
1157
1158 uint alias_idx = C->get_alias_index(adr_type);
1159 MergeMemNode* out_mem;
1160 if (alias_idx != Compile::AliasIdxBot) {
1161 out_mem = MergeMemNode::make(mem);
1162 out_mem->set_memory_at(alias_idx, m);
1163 } else {
1164 out_mem = MergeMemNode::make(m);
1165 }
1166 transform_later(out_mem);
1167
1168 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1169 // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1170 if (_callprojs->fallthrough_ioproj != nullptr) {
1171 *io = _callprojs->fallthrough_ioproj->clone();
1172 transform_later(*io);
1173 } else {
1174 *io = nullptr;
1175 }
1176
1177 return out_mem;
1178 }
1179
1180 // Helper function; generates code for cases requiring runtime checks.
1181 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1182 const TypePtr* adr_type,
1183 Node* dest_elem_klass,
1184 Node* src, Node* src_offset,
1185 Node* dest, Node* dest_offset,
1186 Node* copy_length, bool dest_uninitialized) {
1187 if ((*ctrl)->is_top()) return nullptr;
1188
1189 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1190 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1191 return nullptr;
1192 }
1193
1194 // Pick out the parameters required to perform a store-check
1195 // for the target array. This is an optimistic check. It will
1196 // look in each non-null element's class, at the desired klass's
1197 // super_check_offset, for the desired klass.
1198 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1199 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1200 Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1201 Node* check_offset = ConvI2X(transform_later(n3));
1202 Node* check_value = dest_elem_klass;
1203
1204 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1205 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1206
1207 const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1208 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1209 src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1210
1211 finish_arraycopy_call(call, ctrl, mem, adr_type);
1212
1213 Node* proj = new ProjNode(call, TypeFunc::Parms);
1214 transform_later(proj);
1215
1216 return proj;
1217 }
1218
1219 // Helper function; generates code for cases requiring runtime checks.
1220 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1221 const TypePtr* adr_type,
1222 Node* src, Node* src_offset,
1223 Node* dest, Node* dest_offset,
1224 Node* copy_length, bool dest_uninitialized) {
1225 if ((*ctrl)->is_top()) return nullptr;
1226 assert(!dest_uninitialized, "Invariant");
1227
1228 address copyfunc_addr = StubRoutines::generic_arraycopy();
1229 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1230 return nullptr;
1231 }
1232
1233 const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1234 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1235 src, src_offset, dest, dest_offset, copy_length);
1236
1237 finish_arraycopy_call(call, ctrl, mem, adr_type);
1238
1239 Node* proj = new ProjNode(call, TypeFunc::Parms);
1240 transform_later(proj);
1241
1242 return proj;
1243 }
1244
1245 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1246 bool PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1247 const TypePtr* adr_type,
1248 BasicType basic_elem_type,
1249 bool disjoint_bases,
1250 Node* src, Node* src_offset,
1251 Node* dest, Node* dest_offset,
1252 Node* copy_length, bool dest_uninitialized) {
1253 if ((*ctrl)->is_top()) return false;
1254
1255 Node* src_start = src;
1256 Node* dest_start = dest;
1257 if (src_offset != nullptr || dest_offset != nullptr) {
1258 src_start = array_element_address(src, src_offset, basic_elem_type);
1259 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1260 }
1261
1262 // Figure out which arraycopy runtime method to call.
1263 const char* copyfunc_name = "arraycopy";
1264 address copyfunc_addr =
1265 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1266 disjoint_bases, copyfunc_name, dest_uninitialized);
1267
1268 Node* result_memory = nullptr;
1269 RegionNode* exit_block = nullptr;
1270 if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1271 Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1272 generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1273 copy_length, src_start, dest_start, basic_elem_type);
1274 }
1275
1276 const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1277 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1278 src_start, dest_start, copy_length XTOP);
1279
1280 finish_arraycopy_call(call, ctrl, mem, adr_type);
1281
1282 // Connecting remaining edges for exit_block coming from stub_block.
1283 if (exit_block) {
1284 exit_block->init_req(2, *ctrl);
1285
1286 // Memory edge corresponding to stub_region.
1287 result_memory->init_req(2, *mem);
1288
1289 uint alias_idx = C->get_alias_index(adr_type);
1290 if (alias_idx != Compile::AliasIdxBot) {
1291 *mem = MergeMemNode::make(*mem);
1292 (*mem)->set_memory_at(alias_idx, result_memory);
1293 } else {
1294 *mem = MergeMemNode::make(result_memory);
1295 }
1296 transform_later(*mem);
1297 *ctrl = exit_block;
1298 return true;
1299 }
1300 return false;
1301 }
1302
1303 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1304 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1305 Node*& dest_length) {
1306 #ifdef ASSERT
1307 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1308 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1309 bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1310 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1311 #endif
1312 int elem_size = top_dest->flat_elem_size();
1313 if (elem_size >= 8) {
1314 if (elem_size > 8) {
1315 // treat as array of long but scale length, src offset and dest offset
1316 assert((elem_size % 8) == 0, "not a power of 2?");
1317 int factor = elem_size / 8;
1318 length = transform_later(new MulINode(length, intcon(factor)));
1319 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1320 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1321 if (dest_length != nullptr) {
1322 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1323 }
1324 elem_size = 8;
1325 }
1326 dest_elem = T_LONG;
1327 } else if (elem_size == 4) {
1328 dest_elem = T_INT;
1329 } else if (elem_size == 2) {
1330 dest_elem = T_CHAR;
1331 } else if (elem_size == 1) {
1332 dest_elem = T_BYTE;
1333 } else {
1334 ShouldNotReachHere();
1335 }
1336 return TypeRawPtr::BOTTOM;
1337 }
1338
1339 #undef XTOP
1340
1341 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1342 Node* ctrl = ac->in(TypeFunc::Control);
1343 Node* io = ac->in(TypeFunc::I_O);
1344 Node* src = ac->in(ArrayCopyNode::Src);
1345 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1346 Node* dest = ac->in(ArrayCopyNode::Dest);
1347 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1348 Node* length = ac->in(ArrayCopyNode::Length);
1349 MergeMemNode* merge_mem = nullptr;
1350
1351 if (ac->is_clonebasic()) {
1352 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1353 bs->clone_at_expansion(this, ac);
1354 return;
1355 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1356 const Type* src_type = _igvn.type(src);
1357 const Type* dest_type = _igvn.type(dest);
1358 const TypeAryPtr* top_src = src_type->isa_aryptr();
1359 // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1360 // Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1361 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1362 BasicType dest_elem = T_OBJECT;
1363 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1364 dest_elem = top_dest->elem()->array_element_basic_type();
1365 }
1366 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1367
1368 if (top_src != nullptr && top_src->is_flat()) {
1369 // If src is flat, dest is guaranteed to be flat as well
1370 top_dest = top_src;
1371 }
1372
1373 AllocateArrayNode* alloc = nullptr;
1374 Node* dest_length = nullptr;
1375 if (ac->is_alloc_tightly_coupled()) {
1376 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1377 assert(alloc != nullptr, "expect alloc");
1378 dest_length = alloc->in(AllocateNode::ALength);
1379 }
1380
1381 Node* mem = ac->in(TypeFunc::Memory);
1382 const TypePtr* adr_type = nullptr;
1383 if (top_dest != nullptr && top_dest->is_flat()) {
1384 assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1385 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1386 // on all slices to prevent writes into the source from floating below the arraycopy.
1387 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1388 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1389 } else {
1390 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1391 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1392 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1393 }
1394 if (ac->_src_type != ac->_dest_type) {
1395 adr_type = TypeRawPtr::BOTTOM;
1396 }
1397 }
1398 merge_mem = MergeMemNode::make(mem);
1399 transform_later(merge_mem);
1400
1401 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1402 adr_type, dest_elem,
1403 src, src_offset, dest, dest_offset, length,
1404 dest_length,
1405 true, ac->has_negative_length_guard());
1406
1407 return;
1408 }
1409
1410 AllocateArrayNode* alloc = nullptr;
1411 if (ac->is_alloc_tightly_coupled()) {
1412 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1413 assert(alloc != nullptr, "expect alloc");
1414 }
1415
1416 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1417
1418 // Compile time checks. If any of these checks cannot be verified at compile time,
1419 // we do not make a fast path for this call. Instead, we let the call remain as it
1420 // is. The checks we choose to mandate at compile time are:
1421 //
1422 // (1) src and dest are arrays.
1423 const Type* src_type = src->Value(&_igvn);
1424 const Type* dest_type = dest->Value(&_igvn);
1425 const TypeAryPtr* top_src = src_type->isa_aryptr();
1426 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1427
1428 BasicType src_elem = T_CONFLICT;
1429 BasicType dest_elem = T_CONFLICT;
1430
1431 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1432 src_elem = top_src->elem()->array_element_basic_type();
1433 }
1434 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1435 dest_elem = top_dest->elem()->array_element_basic_type();
1436 }
1437 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1438 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1439
1440 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1441 src_elem = dest_elem;
1442 }
1443
1444 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1445 // Conservatively insert a memory barrier on all memory slices.
1446 // Do not let writes into the source float below the arraycopy.
1447 {
1448 Node* mem = ac->in(TypeFunc::Memory);
1449 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1450
1451 merge_mem = MergeMemNode::make(mem);
1452 transform_later(merge_mem);
1453 }
1454
1455 // Call StubRoutines::generic_arraycopy stub.
1456 generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1457 TypeRawPtr::BOTTOM, T_CONFLICT,
1458 src, src_offset, dest, dest_offset, length,
1459 nullptr,
1460 // If a negative length guard was generated for the ArrayCopyNode,
1461 // the length of the array can never be negative.
1462 false, ac->has_negative_length_guard());
1463 return;
1464 }
1465
1466 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1467
1468 // (2) src and dest arrays must have elements of the same BasicType
1469 // Figure out the size and type of the elements we will be copying.
1470 //
1471 // We have no stub to copy flat inline type arrays with oop
1472 // fields if we need to emit write barriers.
1473 //
1474 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1475 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1476 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1477 bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1478 // The component types are not the same or are not recognized. Punt.
1479 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1480 {
1481 Node* mem = ac->in(TypeFunc::Memory);
1482 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1483 }
1484
1485 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1486 if (_callprojs->fallthrough_ioproj != nullptr) {
1487 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1488 }
1489 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1490 return;
1491 }
1492
1493 //---------------------------------------------------------------------------
1494 // We will make a fast path for this call to arraycopy.
1495
1496 // We have the following tests left to perform:
1497 //
1498 // (3) src and dest must not be null.
1499 // (4) src_offset must not be negative.
1500 // (5) dest_offset must not be negative.
1501 // (6) length must not be negative.
1502 // (7) src_offset + length must not exceed length of src.
1503 // (8) dest_offset + length must not exceed length of dest.
1504 // (9) each element of an oop array must be assignable
1505
1506 Node* mem = ac->in(TypeFunc::Memory);
1507 if (top_dest->is_flat()) {
1508 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1509 // on all slices to prevent writes into the source from floating below the arraycopy.
1510 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1511 }
1512 merge_mem = MergeMemNode::make(mem);
1513 transform_later(merge_mem);
1514
1515 RegionNode* slow_region = new RegionNode(1);
1516 transform_later(slow_region);
1517
1518 if (!ac->is_arraycopy_validated()) {
1519 // (3) operands must not be null
1520 // We currently perform our null checks with the null_check routine.
1521 // This means that the null exceptions will be reported in the caller
1522 // rather than (correctly) reported inside of the native arraycopy call.
1523 // This should be corrected, given time. We do our null check with the
1524 // stack pointer restored.
1525 // null checks done library_call.cpp
1526
1527 // (4) src_offset must not be negative.
1528 generate_negative_guard(&ctrl, src_offset, slow_region);
1529
1530 // (5) dest_offset must not be negative.
1531 generate_negative_guard(&ctrl, dest_offset, slow_region);
1532
1533 // (6) length must not be negative (moved to generate_arraycopy()).
1534 // generate_negative_guard(length, slow_region);
1535
1536 // (7) src_offset + length must not exceed length of src.
1537 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1538 assert(alen != nullptr, "need src len");
1539 generate_limit_guard(&ctrl,
1540 src_offset, length,
1541 alen,
1542 slow_region);
1543
1544 // (8) dest_offset + length must not exceed length of dest.
1545 alen = ac->in(ArrayCopyNode::DestLen);
1546 assert(alen != nullptr, "need dest len");
1547 generate_limit_guard(&ctrl,
1548 dest_offset, length,
1549 alen,
1550 slow_region);
1551
1552 // (9) each element of an oop array must be assignable
1553 // The generate_arraycopy subroutine checks this.
1554
1555 // TODO 8350865 Fix below logic. Also handle atomicity.
1556 // 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.
1557 if (!(top_src->is_flat() && top_dest->is_flat())) {
1558 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1559 generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1560 }
1561
1562 // Handle inline type arrays
1563 if (!top_src->is_flat()) {
1564 if (UseArrayFlattening && !top_src->is_not_flat()) {
1565 // Src might be flat and dest might not be flat. Go to the slow path if src is flat.
1566 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1567 }
1568 if (EnableValhalla) {
1569 // No validation. The subtype check emitted at macro expansion time will not go to the slow
1570 // path but call checkcast_arraycopy which can not handle flat/null-free inline type arrays.
1571 generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1572 }
1573 } else {
1574 assert(top_dest->is_flat(), "dest array must be flat");
1575 }
1576 }
1577
1578 // This is where the memory effects are placed:
1579 const TypePtr* adr_type = nullptr;
1580 Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1581
1582 if (top_src->is_flat() && top_dest->is_flat()) {
1583 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1584 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1585 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1586 } else {
1587 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1588 }
1589
1590 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1591 adr_type, dest_elem,
1592 src, src_offset, dest, dest_offset, length,
1593 dest_length,
1594 // If a negative length guard was generated for the ArrayCopyNode,
1595 // the length of the array can never be negative.
1596 false, ac->has_negative_length_guard(),
1597 slow_region);
1598 }