1 /*
2 * Copyright (c) 2012, 2024, 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 "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);
47 transform_later(*ctrl);
48 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
49 transform_later(mem_proj);
50 *mem = mem_proj;
51 }
52
53 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
54 uint shift = exact_log2(type2aelembytes(elembt));
55 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
56 Node* base = basic_plus_adr(ary, header);
57 #ifdef _LP64
58 // see comment in GraphKit::array_element_address
59 int index_max = max_jint - 1; // array size is max_jint, index is one less
60 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
61 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
62 #endif
63 Node* scale = new LShiftXNode(idx, intcon(shift));
64 transform_later(scale);
65 return basic_plus_adr(ary, base, scale);
66 }
67
68 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
69 return transform_later(new ConvI2LNode(offset));
70 }
71
72 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
73 const TypeFunc* call_type, address call_addr,
74 const char* call_name,
75 const TypePtr* adr_type,
76 Node* parm0, Node* parm1,
77 Node* parm2, Node* parm3,
78 Node* parm4, Node* parm5,
79 Node* parm6, Node* parm7) {
80 Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
81 call->init_req(TypeFunc::Control, ctrl);
82 call->init_req(TypeFunc::I_O , top());
83 call->init_req(TypeFunc::Memory , mem);
84 call->init_req(TypeFunc::ReturnAdr, top());
85 call->init_req(TypeFunc::FramePtr, top());
86
87 // Hook each parm in order. Stop looking at the first null.
88 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
89 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
90 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
91 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
92 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
93 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
94 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
95 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
96 /* close each nested if ===> */ } } } } } } } }
97 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
98
99 return call;
100 }
101
102
103 //------------------------------generate_guard---------------------------
104 // Helper function for generating guarded fast-slow graph structures.
105 // The given 'test', if true, guards a slow path. If the test fails
106 // then a fast path can be taken. (We generally hope it fails.)
107 // In all cases, GraphKit::control() is updated to the fast path.
108 // The returned value represents the control for the slow path.
109 // The return value is never 'top'; it is either a valid control
110 // or null if it is obvious that the slow path can never be taken.
111 // Also, if region and the slow control are not null, the slow edge
112 // is appended to the region.
113 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
114 if ((*ctrl)->is_top()) {
115 // Already short circuited.
116 return nullptr;
117 }
118 // Build an if node and its projections.
119 // If test is true we take the slow path, which we assume is uncommon.
120 if (_igvn.type(test) == TypeInt::ZERO) {
121 // The slow branch is never taken. No need to build this guard.
122 return nullptr;
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 != nullptr) {
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
171 last = new AddINode(last, offset);
172 transform_later(last);
173 }
174 Node* cmp_lt = new CmpUNode(array_length, last);
175 transform_later(cmp_lt);
176 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
177 transform_later(bol_lt);
178 generate_guard(ctrl, bol_lt, region, PROB_MIN);
179 }
180
181 //
182 // Partial in-lining handling for smaller conjoint/disjoint array copies having
183 // length(in bytes) less than ArrayOperationPartialInlineSize.
184 // if (length <= ArrayOperationPartialInlineSize) {
185 // partial_inlining_block:
186 // mask = Mask_Gen
187 // vload = LoadVectorMasked src , mask
188 // StoreVectorMasked dst, mask, vload
189 // } else {
190 // stub_block:
191 // callstub array_copy
192 // }
193 // exit_block:
194 // Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
195 // mem = MergeMem (Phi)
196 // control = stub_block
197 //
198 // Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
199 // edges. Remaining edges for exit_block coming from stub_block are connected by the caller
200 // post stub nodes creation.
201 //
202
203 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
204 RegionNode** exit_block, Node** result_memory, Node* length,
205 Node* src_start, Node* dst_start, BasicType type) {
206 const TypePtr *src_adr_type = _igvn.type(src_start)->isa_ptr();
207 Node* inline_block = nullptr;
208 Node* stub_block = nullptr;
209
210 int const_len = -1;
211 const TypeInt* lty = nullptr;
212 uint shift = exact_log2(type2aelembytes(type));
213 if (length->Opcode() == Op_ConvI2L) {
214 lty = _igvn.type(length->in(1))->isa_int();
215 } else {
216 lty = _igvn.type(length)->isa_int();
217 }
218 if (lty && lty->is_con()) {
219 const_len = lty->get_con() << shift;
220 }
221
222 // Return if copy length is greater than partial inline size limit or
223 // target does not supports masked load/stores.
224 int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, const_len);
225 if ( const_len > ArrayOperationPartialInlineSize ||
226 !Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type) ||
227 !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
228 !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
229 return;
230 }
231
232 int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
233 Node* casted_length = new CastLLNode(*ctrl, length, TypeLong::make(0, inline_limit, Type::WidenMin));
234 transform_later(casted_length);
235 Node* copy_bytes = new LShiftXNode(length, intcon(shift));
236 transform_later(copy_bytes);
237
238 Node* cmp_le = new CmpULNode(copy_bytes, longcon(ArrayOperationPartialInlineSize));
239 transform_later(cmp_le);
240 Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
241 transform_later(bol_le);
242 inline_block = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
243 stub_block = *ctrl;
244
245 Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
246 transform_later(mask_gen);
247
248 unsigned vec_size = lane_count * type2aelembytes(type);
249 if (C->max_vector_size() < vec_size) {
250 C->set_max_vector_size(vec_size);
251 }
252
253 const TypeVect * vt = TypeVect::make(type, lane_count);
254 Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
255 Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
256 src_adr_type, vt, mask_gen);
257 transform_later(masked_load);
258
259 mm = (*mem)->memory_at(C->get_alias_index(adr_type));
260 Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
261 masked_load, adr_type, mask_gen);
262 transform_later(masked_store);
263
264 // Convergence region for inline_block and stub_block.
265 *exit_block = new RegionNode(3);
266 transform_later(*exit_block);
267 (*exit_block)->init_req(1, inline_block);
268 *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
269 transform_later(*result_memory);
270 (*result_memory)->init_req(1, masked_store);
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 nullptr;
278
279 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
280 return nullptr; // 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, nullptr, PROB_MIN);
287
288 return is_notp;
289 }
290
291 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
292 // Load markword and check if obj is locked
293 Node* mark = make_load(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
294 Node* locked_bit = MakeConX(markWord::unlocked_value);
295 locked_bit = transform_later(new AndXNode(locked_bit, mark));
296 Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
297 Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
298 IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
299 Node* locked_region = transform_later(new RegionNode(3));
300 Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
301
302 // Unlocked: Use bits from mark word
303 locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
304 mark_phi->init_req(1, mark);
305
306 // Locked: Load prototype header from klass
307 *ctrl = transform_later(new IfFalseNode(iff));
308 // Make loads control dependent to make sure they are only executed if array is locked
309 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
310 Node* klass = transform_later(LoadKlassNode::make(_igvn, *ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
311 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
312 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));
313
314 locked_region->init_req(2, *ctrl);
315 mark_phi->init_req(2, proto);
316 *ctrl = locked_region;
317
318 // Now check if mark word bits are set
319 Node* mask = MakeConX(mask_val);
320 Node* masked = transform_later(new AndXNode(mark_phi, mask));
321 cmp = transform_later(new CmpXNode(masked, mask));
322 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
323 return generate_fair_guard(ctrl, bol, region);
324 }
325
326 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
327 return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
328 }
329
330 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
331 return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
332 }
333
334 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
335 transform_later(call);
336
337 *ctrl = new ProjNode(call,TypeFunc::Control);
338 transform_later(*ctrl);
339 Node* newmem = new ProjNode(call, TypeFunc::Memory);
340 transform_later(newmem);
341
342 uint alias_idx = C->get_alias_index(adr_type);
343 if (alias_idx != Compile::AliasIdxBot) {
344 *mem = MergeMemNode::make(*mem);
345 (*mem)->set_memory_at(alias_idx, newmem);
346 } else {
347 *mem = MergeMemNode::make(newmem);
348 }
349 transform_later(*mem);
350 }
351
352 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
353 Node* src_offset,
354 Node* dest_offset,
355 bool disjoint_bases,
356 const char* &name,
357 bool dest_uninitialized) {
358 const TypeInt* src_offset_inttype = _igvn.find_int_type(src_offset);
359 const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
360
361 bool aligned = false;
362 bool disjoint = disjoint_bases;
363
364 // if the offsets are the same, we can treat the memory regions as
365 // disjoint, because either the memory regions are in different arrays,
366 // or they are identical (which we can treat as disjoint.) We can also
367 // treat a copy with a destination index less that the source index
368 // as disjoint since a low->high copy will work correctly in this case.
369 if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
370 dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
371 // both indices are constants
372 int s_offs = src_offset_inttype->get_con();
373 int d_offs = dest_offset_inttype->get_con();
374 int element_size = type2aelembytes(t);
375 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
376 ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
377 if (s_offs >= d_offs) disjoint = true;
378 } else if (src_offset == dest_offset && src_offset != nullptr) {
379 // This can occur if the offsets are identical non-constants.
380 disjoint = true;
381 }
382
383 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
384 }
385
386 #define XTOP LP64_ONLY(COMMA top())
387
388 // Generate an optimized call to arraycopy.
389 // Caller must guard against non-arrays.
390 // Caller must determine a common array basic-type for both arrays.
391 // Caller must validate offsets against array bounds.
392 // The slow_region has already collected guard failure paths
393 // (such as out of bounds length or non-conformable array types).
394 // The generated code has this shape, in general:
395 //
396 // if (length == 0) return // via zero_path
397 // slowval = -1
398 // if (types unknown) {
399 // slowval = call generic copy loop
400 // if (slowval == 0) return // via checked_path
401 // } else if (indexes in bounds) {
402 // if ((is object array) && !(array type check)) {
403 // slowval = call checked copy loop
404 // if (slowval == 0) return // via checked_path
405 // } else {
406 // call bulk copy loop
407 // return // via fast_path
408 // }
409 // }
410 // // adjust params for remaining work:
411 // if (slowval != -1) {
412 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
413 // }
414 // slow_region:
415 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
416 // return // via slow_call_path
417 //
418 // This routine is used from several intrinsics: System.arraycopy,
419 // Object.clone (the array subcase), and Arrays.copyOf[Range].
420 //
421 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
422 Node** ctrl, MergeMemNode* mem, Node** io,
423 const TypePtr* adr_type,
424 BasicType basic_elem_type,
425 Node* src, Node* src_offset,
426 Node* dest, Node* dest_offset,
427 Node* copy_length,
428 Node* dest_length,
429 bool disjoint_bases,
430 bool length_never_negative,
431 RegionNode* slow_region) {
432 if (slow_region == nullptr) {
433 slow_region = new RegionNode(1);
434 transform_later(slow_region);
435 }
436
437 Node* original_dest = dest;
438 bool dest_needs_zeroing = false;
439 bool acopy_to_uninitialized = false;
440 Node* default_value = nullptr;
441 Node* raw_default_value = nullptr;
442
443 // See if this is the initialization of a newly-allocated array.
444 // If so, we will take responsibility here for initializing it to zero.
445 // (Note: Because tightly_coupled_allocation performs checks on the
446 // out-edges of the dest, we need to avoid making derived pointers
447 // from it until we have checked its uses.)
448 if (ReduceBulkZeroing
449 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
450 && basic_elem_type != T_CONFLICT // avoid corner case
451 && !src->eqv_uncast(dest)
452 && alloc != nullptr
453 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
454 assert(ac->is_alloc_tightly_coupled(), "sanity");
455 // acopy to uninitialized tightly coupled allocations
456 // needs zeroing outside the copy range
457 // and the acopy itself will be to uninitialized memory
458 acopy_to_uninitialized = true;
459 if (alloc->maybe_set_complete(&_igvn)) {
460 // "You break it, you buy it."
461 InitializeNode* init = alloc->initialization();
462 assert(init->is_complete(), "we just did this");
463 init->set_complete_with_arraycopy();
464 assert(dest->is_CheckCastPP(), "sanity");
465 assert(dest->in(0)->in(0) == init, "dest pinned");
466 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
467 // From this point on, every exit path is responsible for
468 // initializing any non-copied parts of the object to zero.
469 // Also, if this flag is set we make sure that arraycopy interacts properly
470 // with G1, eliding pre-barriers. See CR 6627983.
471 dest_needs_zeroing = true;
472 default_value = alloc->in(AllocateNode::DefaultValue);
473 raw_default_value = alloc->in(AllocateNode::RawDefaultValue);
474 } else {
475 // dest_need_zeroing = false;
476 }
477 } else {
478 // No zeroing elimination needed here.
479 alloc = nullptr;
480 acopy_to_uninitialized = false;
481 //original_dest = dest;
482 //dest_needs_zeroing = false;
483 }
484
485 uint alias_idx = C->get_alias_index(adr_type);
486
487 // Results are placed here:
488 enum { fast_path = 1, // normal void-returning assembly stub
489 checked_path = 2, // special assembly stub with cleanup
490 slow_call_path = 3, // something went wrong; call the VM
491 zero_path = 4, // bypass when length of copy is zero
492 bcopy_path = 5, // copy primitive array by 64-bit blocks
493 PATH_LIMIT = 6
494 };
495 RegionNode* result_region = new RegionNode(PATH_LIMIT);
496 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
497 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
498 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
499 transform_later(result_region);
500 transform_later(result_i_o);
501 transform_later(result_memory);
502
503 // The slow_control path:
504 Node* slow_control;
505 Node* slow_i_o = *io;
506 Node* slow_mem = mem->memory_at(alias_idx);
507 DEBUG_ONLY(slow_control = (Node*) badAddress);
508
509 // Checked control path:
510 Node* checked_control = top();
511 Node* checked_mem = nullptr;
512 Node* checked_i_o = nullptr;
513 Node* checked_value = nullptr;
514
515 if (basic_elem_type == T_CONFLICT) {
516 assert(!dest_needs_zeroing, "");
517 Node* cv = generate_generic_arraycopy(ctrl, &mem,
518 adr_type,
519 src, src_offset, dest, dest_offset,
520 copy_length, acopy_to_uninitialized);
521 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
522 checked_control = *ctrl;
523 checked_i_o = *io;
524 checked_mem = mem->memory_at(alias_idx);
525 checked_value = cv;
526 *ctrl = top();
527 }
528
529 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
530 if (not_pos != nullptr) {
531 Node* local_ctrl = not_pos, *local_io = *io;
532 MergeMemNode* local_mem = MergeMemNode::make(mem);
533 transform_later(local_mem);
534
535 // (6) length must not be negative.
536 if (!length_never_negative) {
537 generate_negative_guard(&local_ctrl, copy_length, slow_region);
538 }
539
540 // copy_length is 0.
541 if (dest_needs_zeroing) {
542 assert(!local_ctrl->is_top(), "no ctrl?");
543 if (copy_length->eqv_uncast(dest_length)
544 || _igvn.find_int_con(dest_length, 1) <= 0) {
545 // There is no zeroing to do. No need for a secondary raw memory barrier.
546 } else {
547 // Clear the whole thing since there are no source elements to copy.
548 generate_clear_array(local_ctrl, local_mem,
549 adr_type, dest,
550 default_value, raw_default_value,
551 basic_elem_type,
552 intcon(0), nullptr,
553 alloc->in(AllocateNode::AllocSize));
554 // Use a secondary InitializeNode as raw memory barrier.
555 // Currently it is needed only on this path since other
556 // paths have stub or runtime calls as raw memory barriers.
557 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
558 Compile::AliasIdxRaw,
559 top());
560 transform_later(mb);
561 mb->set_req(TypeFunc::Control,local_ctrl);
562 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
563 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
564 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
565
566 InitializeNode* init = mb->as_Initialize();
567 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
568 }
569 }
570
571 // Present the results of the fast call.
572 result_region->init_req(zero_path, local_ctrl);
573 result_i_o ->init_req(zero_path, local_io);
574 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
575 }
576
577 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
578 // We have to initialize the *uncopied* part of the array to zero.
579 // The copy destination is the slice dest[off..off+len]. The other slices
580 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
581 Node* dest_size = alloc->in(AllocateNode::AllocSize);
582 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
583
584 // If there is a head section that needs zeroing, do it now.
585 if (_igvn.find_int_con(dest_offset, -1) != 0) {
586 generate_clear_array(*ctrl, mem,
587 adr_type, dest,
588 default_value, raw_default_value,
589 basic_elem_type,
590 intcon(0), dest_offset,
591 nullptr);
592 }
593
594 // Next, perform a dynamic check on the tail length.
595 // It is often zero, and we can win big if we prove this.
596 // There are two wins: Avoid generating the ClearArray
597 // with its attendant messy index arithmetic, and upgrade
598 // the copy to a more hardware-friendly word size of 64 bits.
599 Node* tail_ctl = nullptr;
600 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
601 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
602 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
603 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
604 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
605 }
606
607 // At this point, let's assume there is no tail.
608 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
609 // There is no tail. Try an upgrade to a 64-bit copy.
610 bool didit = false;
611 {
612 Node* local_ctrl = *ctrl, *local_io = *io;
613 MergeMemNode* local_mem = MergeMemNode::make(mem);
614 transform_later(local_mem);
615
616 didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
617 adr_type, basic_elem_type, alloc,
618 src, src_offset, dest, dest_offset,
619 dest_size, acopy_to_uninitialized);
620 if (didit) {
621 // Present the results of the block-copying fast call.
622 result_region->init_req(bcopy_path, local_ctrl);
623 result_i_o ->init_req(bcopy_path, local_io);
624 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
625 }
626 }
627 if (didit) {
628 *ctrl = top(); // no regular fast path
629 }
630 }
631
632 // Clear the tail, if any.
633 if (tail_ctl != nullptr) {
634 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
635 *ctrl = tail_ctl;
636 if (notail_ctl == nullptr) {
637 generate_clear_array(*ctrl, mem,
638 adr_type, dest,
639 default_value, raw_default_value,
640 basic_elem_type,
641 dest_tail, nullptr,
642 dest_size);
643 } else {
644 // Make a local merge.
645 Node* done_ctl = transform_later(new RegionNode(3));
646 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
647 done_ctl->init_req(1, notail_ctl);
648 done_mem->init_req(1, mem->memory_at(alias_idx));
649 generate_clear_array(*ctrl, mem,
650 adr_type, dest,
651 default_value, raw_default_value,
652 basic_elem_type,
653 dest_tail, nullptr,
654 dest_size);
655 done_ctl->init_req(2, *ctrl);
656 done_mem->init_req(2, mem->memory_at(alias_idx));
657 *ctrl = done_ctl;
658 mem->set_memory_at(alias_idx, done_mem);
659 }
660 }
661 }
662
663 BasicType copy_type = basic_elem_type;
664 assert(basic_elem_type != T_ARRAY, "caller must fix this");
665 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
666 // If src and dest have compatible element types, we can copy bits.
667 // Types S[] and D[] are compatible if D is a supertype of S.
668 //
669 // If they are not, we will use checked_oop_disjoint_arraycopy,
670 // which performs a fast optimistic per-oop check, and backs off
671 // further to JVM_ArrayCopy on the first per-oop check that fails.
672 // (Actually, we don't move raw bits only; the GC requires card marks.)
673
674 // We don't need a subtype check for validated copies and Object[].clone()
675 bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
676 ac->is_copyofrange_validated() || ac->is_clone_oop_array();
677 if (!skip_subtype_check) {
678 // Get the klass* for both src and dest
679 Node* src_klass = ac->in(ArrayCopyNode::SrcKlass);
680 Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
681
682 assert(src_klass != nullptr && dest_klass != nullptr, "should have klasses");
683
684 // Generate the subtype check.
685 // This might fold up statically, or then again it might not.
686 //
687 // Non-static example: Copying List<String>.elements to a new String[].
688 // The backing store for a List<String> is always an Object[],
689 // but its elements are always type String, if the generic types
690 // are correct at the source level.
691 //
692 // Test S[] against D[], not S against D, because (probably)
693 // the secondary supertype cache is less busy for S[] than S.
694 // This usually only matters when D is an interface.
695 Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn, nullptr, -1);
696 // Plug failing path into checked_oop_disjoint_arraycopy
697 if (not_subtype_ctrl != top()) {
698 Node* local_ctrl = not_subtype_ctrl;
699 MergeMemNode* local_mem = MergeMemNode::make(mem);
700 transform_later(local_mem);
701
702 // (At this point we can assume disjoint_bases, since types differ.)
703 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
704 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
705 Node* n1 = LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
706 Node* dest_elem_klass = transform_later(n1);
707 Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
708 adr_type,
709 dest_elem_klass,
710 src, src_offset, dest, dest_offset,
711 ConvI2X(copy_length), acopy_to_uninitialized);
712 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
713 checked_control = local_ctrl;
714 checked_i_o = *io;
715 checked_mem = local_mem->memory_at(alias_idx);
716 checked_value = cv;
717 }
718 }
719 // At this point we know we do not need type checks on oop stores.
720
721 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
722 if (!bs->array_copy_requires_gc_barriers(alloc != nullptr, copy_type, false, false, BarrierSetC2::Expansion)) {
723 // If we do not need gc barriers, copy using the jint or jlong stub.
724 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
725 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
726 "sizes agree");
727 }
728 }
729
730 bool is_partial_array_copy = false;
731 if (!(*ctrl)->is_top()) {
732 // Generate the fast path, if possible.
733 Node* local_ctrl = *ctrl;
734 MergeMemNode* local_mem = MergeMemNode::make(mem);
735 transform_later(local_mem);
736 is_partial_array_copy = generate_unchecked_arraycopy(&local_ctrl, &local_mem,
737 adr_type, copy_type, disjoint_bases,
738 src, src_offset, dest, dest_offset,
739 ConvI2X(copy_length), acopy_to_uninitialized);
740
741 // Present the results of the fast call.
742 result_region->init_req(fast_path, local_ctrl);
743 result_i_o ->init_req(fast_path, *io);
744 result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
745 }
746
747 // Here are all the slow paths up to this point, in one bundle:
748 assert(slow_region != nullptr, "allocated on entry");
749 slow_control = slow_region;
750 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
751
752 *ctrl = checked_control;
753 if (!(*ctrl)->is_top()) {
754 // Clean up after the checked call.
755 // The returned value is either 0 or -1^K,
756 // where K = number of partially transferred array elements.
757 Node* cmp = new CmpINode(checked_value, intcon(0));
758 transform_later(cmp);
759 Node* bol = new BoolNode(cmp, BoolTest::eq);
760 transform_later(bol);
761 IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
762 transform_later(iff);
763
764 // If it is 0, we are done, so transfer to the end.
765 Node* checks_done = new IfTrueNode(iff);
766 transform_later(checks_done);
767 result_region->init_req(checked_path, checks_done);
768 result_i_o ->init_req(checked_path, checked_i_o);
769 result_memory->init_req(checked_path, checked_mem);
770
771 // If it is not zero, merge into the slow call.
772 *ctrl = new IfFalseNode(iff);
773 transform_later(*ctrl);
774 RegionNode* slow_reg2 = new RegionNode(3);
775 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
776 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
777 transform_later(slow_reg2);
778 transform_later(slow_i_o2);
779 transform_later(slow_mem2);
780 slow_reg2 ->init_req(1, slow_control);
781 slow_i_o2 ->init_req(1, slow_i_o);
782 slow_mem2 ->init_req(1, slow_mem);
783 slow_reg2 ->init_req(2, *ctrl);
784 slow_i_o2 ->init_req(2, checked_i_o);
785 slow_mem2 ->init_req(2, checked_mem);
786
787 slow_control = slow_reg2;
788 slow_i_o = slow_i_o2;
789 slow_mem = slow_mem2;
790
791 if (alloc != nullptr) {
792 // We'll restart from the very beginning, after zeroing the whole thing.
793 // This can cause double writes, but that's OK since dest is brand new.
794 // So we ignore the low 31 bits of the value returned from the stub.
795 } else {
796 // We must continue the copy exactly where it failed, or else
797 // another thread might see the wrong number of writes to dest.
798 Node* checked_offset = new XorINode(checked_value, intcon(-1));
799 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
800 transform_later(checked_offset);
801 transform_later(slow_offset);
802 slow_offset->init_req(1, intcon(0));
803 slow_offset->init_req(2, checked_offset);
804
805 // Adjust the arguments by the conditionally incoming offset.
806 Node* src_off_plus = new AddINode(src_offset, slow_offset);
807 transform_later(src_off_plus);
808 Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
809 transform_later(dest_off_plus);
810 Node* length_minus = new SubINode(copy_length, slow_offset);
811 transform_later(length_minus);
812
813 // Tweak the node variables to adjust the code produced below:
814 src_offset = src_off_plus;
815 dest_offset = dest_off_plus;
816 copy_length = length_minus;
817 }
818 }
819 *ctrl = slow_control;
820 if (!(*ctrl)->is_top()) {
821 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
822 MergeMemNode* local_mem = MergeMemNode::make(mem);
823 transform_later(local_mem);
824
825 // Generate the slow path, if needed.
826 local_mem->set_memory_at(alias_idx, slow_mem);
827
828 if (dest_needs_zeroing) {
829 generate_clear_array(local_ctrl, local_mem,
830 adr_type, dest,
831 default_value, raw_default_value,
832 basic_elem_type,
833 intcon(0), nullptr,
834 alloc->in(AllocateNode::AllocSize));
835 }
836
837 local_mem = generate_slow_arraycopy(ac,
838 &local_ctrl, local_mem, &local_io,
839 adr_type,
840 src, src_offset, dest, dest_offset,
841 copy_length, /*dest_uninitialized*/false);
842
843 result_region->init_req(slow_call_path, local_ctrl);
844 result_i_o ->init_req(slow_call_path, local_io);
845 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
846 } else {
847 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
848 // projections were not extracted
849 }
850
851 // Remove unused edges.
852 for (uint i = 1; i < result_region->req(); i++) {
853 if (result_region->in(i) == nullptr) {
854 result_region->init_req(i, top());
855 }
856 }
857
858 // Finished; return the combined state.
859 *ctrl = result_region;
860 *io = result_i_o;
861 mem->set_memory_at(alias_idx, result_memory);
862
863 // mem no longer guaranteed to stay a MergeMemNode
864 Node* out_mem = mem;
865 DEBUG_ONLY(mem = nullptr);
866
867 // The memory edges above are precise in order to model effects around
868 // array copies accurately to allow value numbering of field loads around
869 // arraycopy. Such field loads, both before and after, are common in Java
870 // collections and similar classes involving header/array data structures.
871 //
872 // But with low number of register or when some registers are used or killed
873 // by arraycopy calls it causes registers spilling on stack. See 6544710.
874 // The next memory barrier is added to avoid it. If the arraycopy can be
875 // optimized away (which it can, sometimes) then we can manually remove
876 // the membar also.
877 //
878 // Do not let reads from the cloned object float above the arraycopy.
879 if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
880 // Do not let stores that initialize this object be reordered with
881 // a subsequent store that would make this object accessible by
882 // other threads.
883 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
884 } else {
885 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
886 }
887
888 if (is_partial_array_copy) {
889 assert((*ctrl)->is_Proj(), "MemBar control projection");
890 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
891 (*ctrl)->in(0)->isa_MemBar()->set_trailing_partial_array_copy();
892 }
893
894 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
895 if (_callprojs->fallthrough_ioproj != nullptr) {
896 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
897 }
898 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
899
900 #ifdef ASSERT
901 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
902 if (dest_t->is_known_instance() && !is_partial_array_copy) {
903 ArrayCopyNode* ac = nullptr;
904 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
905 assert(ac == nullptr, "no arraycopy anymore");
906 }
907 #endif
908
909 return out_mem;
910 }
911
912 // Helper for initialization of arrays, creating a ClearArray.
913 // It writes zero bits in [start..end), within the body of an array object.
914 // The memory effects are all chained onto the 'adr_type' alias category.
915 //
916 // Since the object is otherwise uninitialized, we are free
917 // to put a little "slop" around the edges of the cleared area,
918 // as long as it does not go back into the array's header,
919 // or beyond the array end within the heap.
920 //
921 // The lower edge can be rounded down to the nearest jint and the
922 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
923 //
924 // Arguments:
925 // adr_type memory slice where writes are generated
926 // dest oop of the destination array
927 // basic_elem_type element type of the destination
928 // slice_idx array index of first element to store
929 // slice_len number of elements to store (or null)
930 // dest_size total size in bytes of the array object
931 //
932 // Exactly one of slice_len or dest_size must be non-null.
933 // If dest_size is non-null, zeroing extends to the end of the object.
934 // If slice_len is non-null, the slice_idx value must be a constant.
935 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
936 const TypePtr* adr_type,
937 Node* dest,
938 Node* val,
939 Node* raw_val,
940 BasicType basic_elem_type,
941 Node* slice_idx,
942 Node* slice_len,
943 Node* dest_size) {
944 // one or the other but not both of slice_len and dest_size:
945 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
946 if (slice_len == nullptr) slice_len = top();
947 if (dest_size == nullptr) dest_size = top();
948
949 uint alias_idx = C->get_alias_index(adr_type);
950
951 // operate on this memory slice:
952 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
953
954 // scaling and rounding of indexes:
955 int scale = exact_log2(type2aelembytes(basic_elem_type));
956 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
957 int clear_low = (-1 << scale) & (BytesPerInt - 1);
958 int bump_bit = (-1 << scale) & BytesPerInt;
959
960 // determine constant starts and ends
961 const intptr_t BIG_NEG = -128;
962 assert(BIG_NEG + 2*abase < 0, "neg enough");
963 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
964 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
965 if (slice_len_con == 0) {
966 return; // nothing to do here
967 }
968 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
969 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
970 if (slice_idx_con >= 0 && slice_len_con >= 0) {
971 assert(end_con < 0, "not two cons");
972 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
973 BytesPerLong);
974 }
975
976 if (start_con >= 0 && end_con >= 0) {
977 // Constant start and end. Simple.
978 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
979 start_con, end_con, &_igvn);
980 } else if (start_con >= 0 && dest_size != top()) {
981 // Constant start, pre-rounded end after the tail of the array.
982 Node* end = dest_size;
983 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
984 start_con, end, &_igvn);
985 } else if (start_con >= 0 && slice_len != top()) {
986 // Constant start, non-constant end. End needs rounding up.
987 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
988 intptr_t end_base = abase + (slice_idx_con << scale);
989 int end_round = (-1 << scale) & (BytesPerLong - 1);
990 Node* end = ConvI2X(slice_len);
991 if (scale != 0)
992 end = transform_later(new LShiftXNode(end, intcon(scale) ));
993 end_base += end_round;
994 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
995 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
996 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
997 start_con, end, &_igvn);
998 } else if (start_con < 0 && dest_size != top()) {
999 // Non-constant start, pre-rounded end after the tail of the array.
1000 // This is almost certainly a "round-to-end" operation.
1001 Node* start = slice_idx;
1002 start = ConvI2X(start);
1003 if (scale != 0)
1004 start = transform_later(new LShiftXNode( start, intcon(scale) ));
1005 start = transform_later(new AddXNode(start, MakeConX(abase)) );
1006 if ((bump_bit | clear_low) != 0) {
1007 int to_clear = (bump_bit | clear_low);
1008 // Align up mod 8, then store a jint zero unconditionally
1009 // just before the mod-8 boundary.
1010 if (((abase + bump_bit) & ~to_clear) - bump_bit
1011 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1012 bump_bit = 0;
1013 assert((abase & to_clear) == 0, "array base must be long-aligned");
1014 } else {
1015 // Bump 'start' up to (or past) the next jint boundary:
1016 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1017 assert((abase & clear_low) == 0, "array base must be int-aligned");
1018 }
1019 // Round bumped 'start' down to jlong boundary in body of array.
1020 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1021 if (bump_bit != 0) {
1022 // Store a zero to the immediately preceding jint:
1023 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1024 Node* p1 = basic_plus_adr(dest, x1);
1025 if (val == nullptr) {
1026 assert(raw_val == nullptr, "val may not be null");
1027 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1028 } else {
1029 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1030 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1031 }
1032 mem = transform_later(mem);
1033 }
1034 }
1035 Node* end = dest_size; // pre-rounded
1036 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1037 start, end, &_igvn);
1038 } else {
1039 // Non-constant start, unrounded non-constant end.
1040 // (Nobody zeroes a random midsection of an array using this routine.)
1041 ShouldNotReachHere(); // fix caller
1042 }
1043
1044 // Done.
1045 merge_mem->set_memory_at(alias_idx, mem);
1046 }
1047
1048 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1049 const TypePtr* adr_type,
1050 BasicType basic_elem_type,
1051 AllocateNode* alloc,
1052 Node* src, Node* src_offset,
1053 Node* dest, Node* dest_offset,
1054 Node* dest_size, bool dest_uninitialized) {
1055 // See if there is an advantage from block transfer.
1056 int scale = exact_log2(type2aelembytes(basic_elem_type));
1057 if (scale >= LogBytesPerLong)
1058 return false; // it is already a block transfer
1059
1060 // Look at the alignment of the starting offsets.
1061 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1062
1063 intptr_t src_off_con = (intptr_t) _igvn.find_int_con(src_offset, -1);
1064 intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1065 if (src_off_con < 0 || dest_off_con < 0) {
1066 // At present, we can only understand constants.
1067 return false;
1068 }
1069
1070 intptr_t src_off = abase + (src_off_con << scale);
1071 intptr_t dest_off = abase + (dest_off_con << scale);
1072
1073 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1074 // Non-aligned; too bad.
1075 // One more chance: Pick off an initial 32-bit word.
1076 // This is a common case, since abase can be odd mod 8.
1077 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1078 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1079 Node* sptr = basic_plus_adr(src, src_off);
1080 Node* dptr = basic_plus_adr(dest, dest_off);
1081 const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1082 assert(s_adr_type->isa_aryptr(), "impossible slice");
1083 uint s_alias_idx = C->get_alias_index(s_adr_type);
1084 uint d_alias_idx = C->get_alias_index(adr_type);
1085 bool is_mismatched = (basic_elem_type != T_INT);
1086 Node* sval = transform_later(
1087 LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1088 TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1089 false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1090 Node* st = transform_later(
1091 StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1092 sval, T_INT, MemNode::unordered));
1093 if (is_mismatched) {
1094 st->as_Store()->set_mismatched_access();
1095 }
1096 (*mem)->set_memory_at(d_alias_idx, st);
1097 src_off += BytesPerInt;
1098 dest_off += BytesPerInt;
1099 } else {
1100 return false;
1101 }
1102 }
1103 assert(src_off % BytesPerLong == 0, "");
1104 assert(dest_off % BytesPerLong == 0, "");
1105
1106 // Do this copy by giant steps.
1107 Node* sptr = basic_plus_adr(src, src_off);
1108 Node* dptr = basic_plus_adr(dest, dest_off);
1109 Node* countx = dest_size;
1110 countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1111 countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1112
1113 bool disjoint_bases = true; // since alloc isn't null
1114 generate_unchecked_arraycopy(ctrl, mem,
1115 adr_type, T_LONG, disjoint_bases,
1116 sptr, nullptr, dptr, nullptr, countx, dest_uninitialized);
1117
1118 return true;
1119 }
1120
1121 // Helper function; generates code for the slow case.
1122 // We make a call to a runtime method which emulates the native method,
1123 // but without the native wrapper overhead.
1124 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1125 Node** ctrl, Node* mem, Node** io,
1126 const TypePtr* adr_type,
1127 Node* src, Node* src_offset,
1128 Node* dest, Node* dest_offset,
1129 Node* copy_length, bool dest_uninitialized) {
1130 assert(!dest_uninitialized, "Invariant");
1131
1132 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1133 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1134 "slow_arraycopy", TypePtr::BOTTOM);
1135
1136 call->init_req(TypeFunc::Control, *ctrl);
1137 call->init_req(TypeFunc::I_O , *io);
1138 call->init_req(TypeFunc::Memory , mem);
1139 call->init_req(TypeFunc::ReturnAdr, top());
1140 call->init_req(TypeFunc::FramePtr, top());
1141 call->init_req(TypeFunc::Parms+0, src);
1142 call->init_req(TypeFunc::Parms+1, src_offset);
1143 call->init_req(TypeFunc::Parms+2, dest);
1144 call->init_req(TypeFunc::Parms+3, dest_offset);
1145 call->init_req(TypeFunc::Parms+4, copy_length);
1146 call->copy_call_debug_info(&_igvn, ac);
1147
1148 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1149 _igvn.replace_node(ac, call);
1150 transform_later(call);
1151
1152 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1153 *ctrl = _callprojs->fallthrough_catchproj->clone();
1154 transform_later(*ctrl);
1155
1156 Node* m = _callprojs->fallthrough_memproj->clone();
1157 transform_later(m);
1158
1159 uint alias_idx = C->get_alias_index(adr_type);
1160 MergeMemNode* out_mem;
1161 if (alias_idx != Compile::AliasIdxBot) {
1162 out_mem = MergeMemNode::make(mem);
1163 out_mem->set_memory_at(alias_idx, m);
1164 } else {
1165 out_mem = MergeMemNode::make(m);
1166 }
1167 transform_later(out_mem);
1168
1169 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1170 // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1171 if (_callprojs->fallthrough_ioproj != nullptr) {
1172 *io = _callprojs->fallthrough_ioproj->clone();
1173 transform_later(*io);
1174 } else {
1175 *io = nullptr;
1176 }
1177
1178 return out_mem;
1179 }
1180
1181 // Helper function; generates code for cases requiring runtime checks.
1182 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1183 const TypePtr* adr_type,
1184 Node* dest_elem_klass,
1185 Node* src, Node* src_offset,
1186 Node* dest, Node* dest_offset,
1187 Node* copy_length, bool dest_uninitialized) {
1188 if ((*ctrl)->is_top()) return nullptr;
1189
1190 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1191 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1192 return nullptr;
1193 }
1194
1195 // Pick out the parameters required to perform a store-check
1196 // for the target array. This is an optimistic check. It will
1197 // look in each non-null element's class, at the desired klass's
1198 // super_check_offset, for the desired klass.
1199 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1200 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1201 Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1202 Node* check_offset = ConvI2X(transform_later(n3));
1203 Node* check_value = dest_elem_klass;
1204
1205 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1206 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1207
1208 const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1209 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1210 src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1211
1212 finish_arraycopy_call(call, ctrl, mem, adr_type);
1213
1214 Node* proj = new ProjNode(call, TypeFunc::Parms);
1215 transform_later(proj);
1216
1217 return proj;
1218 }
1219
1220 // Helper function; generates code for cases requiring runtime checks.
1221 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1222 const TypePtr* adr_type,
1223 Node* src, Node* src_offset,
1224 Node* dest, Node* dest_offset,
1225 Node* copy_length, bool dest_uninitialized) {
1226 if ((*ctrl)->is_top()) return nullptr;
1227 assert(!dest_uninitialized, "Invariant");
1228
1229 address copyfunc_addr = StubRoutines::generic_arraycopy();
1230 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1231 return nullptr;
1232 }
1233
1234 const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1235 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1236 src, src_offset, dest, dest_offset, copy_length);
1237
1238 finish_arraycopy_call(call, ctrl, mem, adr_type);
1239
1240 Node* proj = new ProjNode(call, TypeFunc::Parms);
1241 transform_later(proj);
1242
1243 return proj;
1244 }
1245
1246 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1247 bool PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1248 const TypePtr* adr_type,
1249 BasicType basic_elem_type,
1250 bool disjoint_bases,
1251 Node* src, Node* src_offset,
1252 Node* dest, Node* dest_offset,
1253 Node* copy_length, bool dest_uninitialized) {
1254 if ((*ctrl)->is_top()) return false;
1255
1256 Node* src_start = src;
1257 Node* dest_start = dest;
1258 if (src_offset != nullptr || dest_offset != nullptr) {
1259 src_start = array_element_address(src, src_offset, basic_elem_type);
1260 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1261 }
1262
1263 // Figure out which arraycopy runtime method to call.
1264 const char* copyfunc_name = "arraycopy";
1265 address copyfunc_addr =
1266 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1267 disjoint_bases, copyfunc_name, dest_uninitialized);
1268
1269 Node* result_memory = nullptr;
1270 RegionNode* exit_block = nullptr;
1271 if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1272 Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1273 generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1274 copy_length, src_start, dest_start, basic_elem_type);
1275 }
1276
1277 const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1278 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1279 src_start, dest_start, copy_length XTOP);
1280
1281 finish_arraycopy_call(call, ctrl, mem, adr_type);
1282
1283 // Connecting remaining edges for exit_block coming from stub_block.
1284 if (exit_block) {
1285 exit_block->init_req(2, *ctrl);
1286
1287 // Memory edge corresponding to stub_region.
1288 result_memory->init_req(2, *mem);
1289
1290 uint alias_idx = C->get_alias_index(adr_type);
1291 if (alias_idx != Compile::AliasIdxBot) {
1292 *mem = MergeMemNode::make(*mem);
1293 (*mem)->set_memory_at(alias_idx, result_memory);
1294 } else {
1295 *mem = MergeMemNode::make(result_memory);
1296 }
1297 transform_later(*mem);
1298 *ctrl = exit_block;
1299 return true;
1300 }
1301 return false;
1302 }
1303
1304 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1305 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1306 Node*& dest_length) {
1307 #ifdef ASSERT
1308 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1309 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1310 bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1311 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1312 #endif
1313 int elem_size = top_dest->flat_elem_size();
1314 if (elem_size >= 8) {
1315 if (elem_size > 8) {
1316 // treat as array of long but scale length, src offset and dest offset
1317 assert((elem_size % 8) == 0, "not a power of 2?");
1318 int factor = elem_size / 8;
1319 length = transform_later(new MulINode(length, intcon(factor)));
1320 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1321 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1322 if (dest_length != nullptr) {
1323 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1324 }
1325 elem_size = 8;
1326 }
1327 dest_elem = T_LONG;
1328 } else if (elem_size == 4) {
1329 dest_elem = T_INT;
1330 } else if (elem_size == 2) {
1331 dest_elem = T_CHAR;
1332 } else if (elem_size == 1) {
1333 dest_elem = T_BYTE;
1334 } else {
1335 ShouldNotReachHere();
1336 }
1337 return TypeRawPtr::BOTTOM;
1338 }
1339
1340 #undef XTOP
1341
1342 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1343 Node* ctrl = ac->in(TypeFunc::Control);
1344 Node* io = ac->in(TypeFunc::I_O);
1345 Node* src = ac->in(ArrayCopyNode::Src);
1346 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1347 Node* dest = ac->in(ArrayCopyNode::Dest);
1348 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1349 Node* length = ac->in(ArrayCopyNode::Length);
1350 MergeMemNode* merge_mem = nullptr;
1351
1352 if (ac->is_clonebasic()) {
1353 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1354 bs->clone_at_expansion(this, ac);
1355 return;
1356 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1357 const Type* src_type = _igvn.type(src);
1358 const Type* dest_type = _igvn.type(dest);
1359 const TypeAryPtr* top_src = src_type->isa_aryptr();
1360 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1361 BasicType dest_elem = T_OBJECT;
1362 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1363 dest_elem = top_dest->elem()->array_element_basic_type();
1364 }
1365 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1366
1367 if (top_src != nullptr && top_src->is_flat()) {
1368 // If src is flat, dest is guaranteed to be flat as well
1369 top_dest = top_src;
1370 }
1371
1372 AllocateArrayNode* alloc = nullptr;
1373 Node* dest_length = nullptr;
1374 if (ac->is_alloc_tightly_coupled()) {
1375 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1376 assert(alloc != nullptr, "expect alloc");
1377 dest_length = alloc->in(AllocateNode::ALength);
1378 }
1379
1380 Node* mem = ac->in(TypeFunc::Memory);
1381 const TypePtr* adr_type = nullptr;
1382 if (top_dest->is_flat()) {
1383 assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1384 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1385 // on all slices to prevent writes into the source from floating below the arraycopy.
1386 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1387 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1388 } else {
1389 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1390 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1391 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1392 }
1393 if (ac->_src_type != ac->_dest_type) {
1394 adr_type = TypeRawPtr::BOTTOM;
1395 }
1396 }
1397 merge_mem = MergeMemNode::make(mem);
1398 transform_later(merge_mem);
1399
1400 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1401 adr_type, dest_elem,
1402 src, src_offset, dest, dest_offset, length,
1403 dest_length,
1404 true, ac->has_negative_length_guard());
1405
1406 return;
1407 }
1408
1409 AllocateArrayNode* alloc = nullptr;
1410 if (ac->is_alloc_tightly_coupled()) {
1411 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1412 assert(alloc != nullptr, "expect alloc");
1413 }
1414
1415 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1416
1417 // Compile time checks. If any of these checks cannot be verified at compile time,
1418 // we do not make a fast path for this call. Instead, we let the call remain as it
1419 // is. The checks we choose to mandate at compile time are:
1420 //
1421 // (1) src and dest are arrays.
1422 const Type* src_type = src->Value(&_igvn);
1423 const Type* dest_type = dest->Value(&_igvn);
1424 const TypeAryPtr* top_src = src_type->isa_aryptr();
1425 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1426
1427 BasicType src_elem = T_CONFLICT;
1428 BasicType dest_elem = T_CONFLICT;
1429
1430 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1431 src_elem = top_src->elem()->array_element_basic_type();
1432 }
1433 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1434 dest_elem = top_dest->elem()->array_element_basic_type();
1435 }
1436 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1437 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1438
1439 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1440 src_elem = dest_elem;
1441 }
1442
1443 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1444 // Conservatively insert a memory barrier on all memory slices.
1445 // Do not let writes into the source float below the arraycopy.
1446 {
1447 Node* mem = ac->in(TypeFunc::Memory);
1448 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1449
1450 merge_mem = MergeMemNode::make(mem);
1451 transform_later(merge_mem);
1452 }
1453
1454 // Call StubRoutines::generic_arraycopy stub.
1455 Node* mem = generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1456 TypeRawPtr::BOTTOM, T_CONFLICT,
1457 src, src_offset, dest, dest_offset, length,
1458 nullptr,
1459 // If a negative length guard was generated for the ArrayCopyNode,
1460 // the length of the array can never be negative.
1461 false, ac->has_negative_length_guard());
1462 return;
1463 }
1464
1465 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1466
1467 // (2) src and dest arrays must have elements of the same BasicType
1468 // Figure out the size and type of the elements we will be copying.
1469 //
1470 // We have no stub to copy flat inline type arrays with oop
1471 // fields if we need to emit write barriers.
1472 //
1473 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1474 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1475 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1476 bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1477 // The component types are not the same or are not recognized. Punt.
1478 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1479 {
1480 Node* mem = ac->in(TypeFunc::Memory);
1481 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1482 }
1483
1484 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1485 if (_callprojs->fallthrough_ioproj != nullptr) {
1486 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1487 }
1488 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1489 return;
1490 }
1491
1492 //---------------------------------------------------------------------------
1493 // We will make a fast path for this call to arraycopy.
1494
1495 // We have the following tests left to perform:
1496 //
1497 // (3) src and dest must not be null.
1498 // (4) src_offset must not be negative.
1499 // (5) dest_offset must not be negative.
1500 // (6) length must not be negative.
1501 // (7) src_offset + length must not exceed length of src.
1502 // (8) dest_offset + length must not exceed length of dest.
1503 // (9) each element of an oop array must be assignable
1504
1505 Node* mem = ac->in(TypeFunc::Memory);
1506 if (top_dest->is_flat()) {
1507 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1508 // on all slices to prevent writes into the source from floating below the arraycopy.
1509 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1510 }
1511 merge_mem = MergeMemNode::make(mem);
1512 transform_later(merge_mem);
1513
1514 RegionNode* slow_region = new RegionNode(1);
1515 transform_later(slow_region);
1516
1517 if (!ac->is_arraycopy_validated()) {
1518 // (3) operands must not be null
1519 // We currently perform our null checks with the null_check routine.
1520 // This means that the null exceptions will be reported in the caller
1521 // rather than (correctly) reported inside of the native arraycopy call.
1522 // This should be corrected, given time. We do our null check with the
1523 // stack pointer restored.
1524 // null checks done library_call.cpp
1525
1526 // (4) src_offset must not be negative.
1527 generate_negative_guard(&ctrl, src_offset, slow_region);
1528
1529 // (5) dest_offset must not be negative.
1530 generate_negative_guard(&ctrl, dest_offset, slow_region);
1531
1532 // (6) length must not be negative (moved to generate_arraycopy()).
1533 // generate_negative_guard(length, slow_region);
1534
1535 // (7) src_offset + length must not exceed length of src.
1536 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1537 assert(alen != nullptr, "need src len");
1538 generate_limit_guard(&ctrl,
1539 src_offset, length,
1540 alen,
1541 slow_region);
1542
1543 // (8) dest_offset + length must not exceed length of dest.
1544 alen = ac->in(ArrayCopyNode::DestLen);
1545 assert(alen != nullptr, "need dest len");
1546 generate_limit_guard(&ctrl,
1547 dest_offset, length,
1548 alen,
1549 slow_region);
1550
1551 // (9) each element of an oop array must be assignable
1552 // The generate_arraycopy subroutine checks this.
1553
1554 // Handle inline type arrays
1555 if (!top_src->is_flat()) {
1556 if (UseFlatArray && !top_src->is_not_flat()) {
1557 // Src might be flat and dest might not be flat. Go to the slow path if src is flat.
1558 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1559 }
1560 if (EnableValhalla) {
1561 // No validation. The subtype check emitted at macro expansion time will not go to the slow
1562 // path but call checkcast_arraycopy which can not handle flat/null-free inline type arrays.
1563 generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1564 }
1565 } else {
1566 assert(top_dest->is_flat(), "dest array must be flat");
1567 }
1568 }
1569
1570 // This is where the memory effects are placed:
1571 const TypePtr* adr_type = nullptr;
1572 Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1573
1574 if (top_dest->is_flat()) {
1575 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1576 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1577 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1578 } else {
1579 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1580 }
1581
1582 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1583 adr_type, dest_elem,
1584 src, src_offset, dest, dest_offset, length,
1585 dest_length,
1586 // If a negative length guard was generated for the ArrayCopyNode,
1587 // the length of the array can never be negative.
1588 false, ac->has_negative_length_guard(),
1589 slow_region);
1590 }