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, int alias_idx, Node* precedent) {
41 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
42 mb->init_req(TypeFunc::Control, *ctrl);
43 mb->init_req(TypeFunc::Memory, *mem);
44 transform_later(mb);
45 *ctrl = new ProjNode(mb,TypeFunc::Control);
46 transform_later(*ctrl);
47 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
48 transform_later(mem_proj);
49 if (alias_idx == Compile::AliasIdxBot) {
50 *mem = mem_proj;
51 } else {
52 MergeMemNode* mm = (*mem)->clone()->as_MergeMem();
53 mm->set_memory_at(alias_idx, mem_proj);
54 transform_later(mm);
55 *mem = mm;
56 }
57 }
58
59 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
60 uint shift = exact_log2(type2aelembytes(elembt));
61 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
62 Node* base = basic_plus_adr(ary, header);
63 #ifdef _LP64
64 // see comment in GraphKit::array_element_address
65 int index_max = max_jint - 1; // array size is max_jint, index is one less
66 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
67 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
68 #endif
69 Node* scale = new LShiftXNode(idx, intcon(shift));
70 transform_later(scale);
71 return basic_plus_adr(ary, base, scale);
72 }
73
74 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
75 return transform_later(new ConvI2LNode(offset));
76 }
77
78 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
79 const TypeFunc* call_type, address call_addr,
80 const char* call_name,
81 const TypePtr* adr_type,
82 Node* parm0, Node* parm1,
83 Node* parm2, Node* parm3,
84 Node* parm4, Node* parm5,
85 Node* parm6, Node* parm7) {
86 Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
87 call->init_req(TypeFunc::Control, ctrl);
88 call->init_req(TypeFunc::I_O , top());
89 call->init_req(TypeFunc::Memory , mem);
90 call->init_req(TypeFunc::ReturnAdr, top());
91 call->init_req(TypeFunc::FramePtr, top());
92
93 // Hook each parm in order. Stop looking at the first null.
94 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
95 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
96 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
97 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
98 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
99 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
100 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
101 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
102 /* close each nested if ===> */ } } } } } } } }
103 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
104
105 return call;
106 }
107
108
109 //------------------------------generate_guard---------------------------
110 // Helper function for generating guarded fast-slow graph structures.
111 // The given 'test', if true, guards a slow path. If the test fails
112 // then a fast path can be taken. (We generally hope it fails.)
113 // In all cases, GraphKit::control() is updated to the fast path.
114 // The returned value represents the control for the slow path.
115 // The return value is never 'top'; it is either a valid control
116 // or null if it is obvious that the slow path can never be taken.
117 // Also, if region and the slow control are not null, the slow edge
118 // is appended to the region.
119 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
120 if ((*ctrl)->is_top()) {
121 // Already short circuited.
122 return nullptr;
123 }
124 // Build an if node and its projections.
125 // If test is true we take the slow path, which we assume is uncommon.
126 if (_igvn.type(test) == TypeInt::ZERO) {
127 // The slow branch is never taken. No need to build this guard.
128 return nullptr;
129 }
130
131 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
132 transform_later(iff);
133
134 Node* if_slow = new IfTrueNode(iff);
135 transform_later(if_slow);
136
137 if (region != nullptr) {
138 region->add_req(if_slow);
139 }
140
141 Node* if_fast = new IfFalseNode(iff);
142 transform_later(if_fast);
143
144 *ctrl = if_fast;
145
146 return if_slow;
147 }
148
149 Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
150 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
151 }
152
153 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
154 return generate_guard(ctrl, test, region, PROB_FAIR);
155 }
156
157 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
158 if ((*ctrl)->is_top())
159 return; // already stopped
160 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
161 return; // index is already adequately typed
162 Node* cmp_lt = new CmpINode(index, intcon(0));
163 transform_later(cmp_lt);
164 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
165 transform_later(bol_lt);
166 generate_guard(ctrl, bol_lt, region, PROB_MIN);
167 }
168
169 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
170 if ((*ctrl)->is_top())
171 return; // already stopped
172 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
173 if (zero_offset && subseq_length->eqv_uncast(array_length))
174 return; // common case of whole-array copy
175 Node* last = subseq_length;
176 if (!zero_offset) { // last += offset
177 last = new AddINode(last, offset);
178 transform_later(last);
179 }
180 Node* cmp_lt = new CmpUNode(array_length, last);
181 transform_later(cmp_lt);
182 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
183 transform_later(bol_lt);
184 generate_guard(ctrl, bol_lt, region, PROB_MIN);
185 }
186
187 //
188 // Partial in-lining handling for smaller conjoint/disjoint array copies having
189 // length(in bytes) less than ArrayOperationPartialInlineSize.
190 // if (length <= ArrayOperationPartialInlineSize) {
191 // partial_inlining_block:
192 // mask = Mask_Gen
193 // vload = LoadVectorMasked src , mask
194 // StoreVectorMasked dst, mask, vload
195 // } else {
196 // stub_block:
197 // callstub array_copy
198 // }
199 // exit_block:
200 // Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
201 // mem = MergeMem (Phi)
202 // control = stub_block
203 //
204 // Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
205 // edges. Remaining edges for exit_block coming from stub_block are connected by the caller
206 // post stub nodes creation.
207 //
208
209 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
210 RegionNode** exit_block, Node** result_memory, Node* length,
211 Node* src_start, Node* dst_start, BasicType type) {
212 const TypePtr *src_adr_type = _igvn.type(src_start)->isa_ptr();
213 Node* inline_block = nullptr;
214 Node* stub_block = nullptr;
215
216 int const_len = -1;
217 const TypeInt* lty = nullptr;
218 uint shift = exact_log2(type2aelembytes(type));
219 if (length->Opcode() == Op_ConvI2L) {
220 lty = _igvn.type(length->in(1))->isa_int();
221 } else {
222 lty = _igvn.type(length)->isa_int();
223 }
224 if (lty && lty->is_con()) {
225 const_len = lty->get_con() << shift;
226 }
227
228 // Return if copy length is greater than partial inline size limit or
229 // target does not supports masked load/stores.
230 int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, const_len);
231 if ( const_len > ArrayOperationPartialInlineSize ||
232 !Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type) ||
233 !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
234 !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
235 return;
236 }
237
238 int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
239 Node* casted_length = new CastLLNode(*ctrl, length, TypeLong::make(0, inline_limit, Type::WidenMin));
240 transform_later(casted_length);
241 Node* copy_bytes = new LShiftXNode(length, intcon(shift));
242 transform_later(copy_bytes);
243
244 Node* cmp_le = new CmpULNode(copy_bytes, longcon(ArrayOperationPartialInlineSize));
245 transform_later(cmp_le);
246 Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
247 transform_later(bol_le);
248 inline_block = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
249 stub_block = *ctrl;
250
251 Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
252 transform_later(mask_gen);
253
254 unsigned vec_size = lane_count * type2aelembytes(type);
255 if (C->max_vector_size() < vec_size) {
256 C->set_max_vector_size(vec_size);
257 }
258
259 const TypeVect * vt = TypeVect::make(type, lane_count);
260 Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
261 Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
262 src_adr_type, vt, mask_gen);
263 transform_later(masked_load);
264
265 mm = (*mem)->memory_at(C->get_alias_index(adr_type));
266 Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
267 masked_load, adr_type, mask_gen);
268 transform_later(masked_store);
269
270 // Convergence region for inline_block and stub_block.
271 *exit_block = new RegionNode(3);
272 transform_later(*exit_block);
273 (*exit_block)->init_req(1, inline_block);
274 *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
275 transform_later(*result_memory);
276 (*result_memory)->init_req(1, masked_store);
277
278 *ctrl = stub_block;
279 }
280
281
282 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
283 if ((*ctrl)->is_top()) return nullptr;
284
285 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
286 return nullptr; // index is already adequately typed
287 Node* cmp_le = new CmpINode(index, intcon(0));
288 transform_later(cmp_le);
289 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
290 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
291 transform_later(bol_le);
292 Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
293
294 return is_notp;
295 }
296
297 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
298 // Load markword and check if obj is locked
299 Node* mark = make_load(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
300 Node* locked_bit = MakeConX(markWord::unlocked_value);
301 locked_bit = transform_later(new AndXNode(locked_bit, mark));
302 Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
303 Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
304 IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
305 Node* locked_region = transform_later(new RegionNode(3));
306 Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
307
308 // Unlocked: Use bits from mark word
309 locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
310 mark_phi->init_req(1, mark);
311
312 // Locked: Load prototype header from klass
313 *ctrl = transform_later(new IfFalseNode(iff));
314 // Make loads control dependent to make sure they are only executed if array is locked
315 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
316 Node* klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
317 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
318 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));
319
320 locked_region->init_req(2, *ctrl);
321 mark_phi->init_req(2, proto);
322 *ctrl = locked_region;
323
324 // Now check if mark word bits are set
325 Node* mask = MakeConX(mask_val);
326 Node* masked = transform_later(new AndXNode(mark_phi, mask));
327 cmp = transform_later(new CmpXNode(masked, mask));
328 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
329 return generate_fair_guard(ctrl, bol, region);
330 }
331
332 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
333 return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
334 }
335
336 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
337 return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
338 }
339
340 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
341 transform_later(call);
342
343 *ctrl = new ProjNode(call,TypeFunc::Control);
344 transform_later(*ctrl);
345 Node* newmem = new ProjNode(call, TypeFunc::Memory);
346 transform_later(newmem);
347
348 uint alias_idx = C->get_alias_index(adr_type);
349 if (alias_idx != Compile::AliasIdxBot) {
350 *mem = MergeMemNode::make(*mem);
351 (*mem)->set_memory_at(alias_idx, newmem);
352 } else {
353 *mem = MergeMemNode::make(newmem);
354 }
355 transform_later(*mem);
356 }
357
358 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
359 Node* src_offset,
360 Node* dest_offset,
361 bool disjoint_bases,
362 const char* &name,
363 bool dest_uninitialized) {
364 const TypeInt* src_offset_inttype = _igvn.find_int_type(src_offset);
365 const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
366
367 bool aligned = false;
368 bool disjoint = disjoint_bases;
369
370 // if the offsets are the same, we can treat the memory regions as
371 // disjoint, because either the memory regions are in different arrays,
372 // or they are identical (which we can treat as disjoint.) We can also
373 // treat a copy with a destination index less that the source index
374 // as disjoint since a low->high copy will work correctly in this case.
375 if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
376 dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
377 // both indices are constants
378 int s_offs = src_offset_inttype->get_con();
379 int d_offs = dest_offset_inttype->get_con();
380 int element_size = type2aelembytes(t);
381 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
382 ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
383 if (s_offs >= d_offs) disjoint = true;
384 } else if (src_offset == dest_offset && src_offset != nullptr) {
385 // This can occur if the offsets are identical non-constants.
386 disjoint = true;
387 }
388
389 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
390 }
391
392 #define XTOP LP64_ONLY(COMMA top())
393
394 // Generate an optimized call to arraycopy.
395 // Caller must guard against non-arrays.
396 // Caller must determine a common array basic-type for both arrays.
397 // Caller must validate offsets against array bounds.
398 // The slow_region has already collected guard failure paths
399 // (such as out of bounds length or non-conformable array types).
400 // The generated code has this shape, in general:
401 //
402 // if (length == 0) return // via zero_path
403 // slowval = -1
404 // if (types unknown) {
405 // slowval = call generic copy loop
406 // if (slowval == 0) return // via checked_path
407 // } else if (indexes in bounds) {
408 // if ((is object array) && !(array type check)) {
409 // slowval = call checked copy loop
410 // if (slowval == 0) return // via checked_path
411 // } else {
412 // call bulk copy loop
413 // return // via fast_path
414 // }
415 // }
416 // // adjust params for remaining work:
417 // if (slowval != -1) {
418 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
419 // }
420 // slow_region:
421 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
422 // return // via slow_call_path
423 //
424 // This routine is used from several intrinsics: System.arraycopy,
425 // Object.clone (the array subcase), and Arrays.copyOf[Range].
426 //
427 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
428 Node** ctrl, MergeMemNode* mem, Node** io,
429 const TypePtr* adr_type,
430 BasicType basic_elem_type,
431 Node* src, Node* src_offset,
432 Node* dest, Node* dest_offset,
433 Node* copy_length,
434 Node* dest_length,
435 bool disjoint_bases,
436 bool length_never_negative,
437 RegionNode* slow_region) {
438 if (slow_region == nullptr) {
439 slow_region = new RegionNode(1);
440 transform_later(slow_region);
441 }
442
443 Node* original_dest = dest;
444 bool dest_needs_zeroing = false;
445 bool acopy_to_uninitialized = false;
446 Node* init_value = nullptr;
447 Node* raw_init_value = nullptr;
448
449 // See if this is the initialization of a newly-allocated array.
450 // If so, we will take responsibility here for initializing it to zero.
451 // (Note: Because tightly_coupled_allocation performs checks on the
452 // out-edges of the dest, we need to avoid making derived pointers
453 // from it until we have checked its uses.)
454 if (ReduceBulkZeroing
455 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
456 && basic_elem_type != T_CONFLICT // avoid corner case
457 && !src->eqv_uncast(dest)
458 && alloc != nullptr
459 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
460 assert(ac->is_alloc_tightly_coupled(), "sanity");
461 // acopy to uninitialized tightly coupled allocations
462 // needs zeroing outside the copy range
463 // and the acopy itself will be to uninitialized memory
464 acopy_to_uninitialized = true;
465 if (alloc->maybe_set_complete(&_igvn)) {
466 // "You break it, you buy it."
467 InitializeNode* init = alloc->initialization();
468 assert(init->is_complete(), "we just did this");
469 init->set_complete_with_arraycopy();
470 assert(dest->is_CheckCastPP(), "sanity");
471 assert(dest->in(0)->in(0) == init, "dest pinned");
472 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
473 // From this point on, every exit path is responsible for
474 // initializing any non-copied parts of the object to zero.
475 // Also, if this flag is set we make sure that arraycopy interacts properly
476 // with G1, eliding pre-barriers. See CR 6627983.
477 dest_needs_zeroing = true;
478 init_value = alloc->in(AllocateNode::InitValue);
479 raw_init_value = alloc->in(AllocateNode::RawInitValue);
480 } else {
481 // dest_need_zeroing = false;
482 }
483 } else {
484 // No zeroing elimination needed here.
485 alloc = nullptr;
486 acopy_to_uninitialized = false;
487 //original_dest = dest;
488 //dest_needs_zeroing = false;
489 }
490
491 uint alias_idx = C->get_alias_index(adr_type);
492
493 // Results are placed here:
494 enum { fast_path = 1, // normal void-returning assembly stub
495 checked_path = 2, // special assembly stub with cleanup
496 slow_call_path = 3, // something went wrong; call the VM
497 zero_path = 4, // bypass when length of copy is zero
498 bcopy_path = 5, // copy primitive array by 64-bit blocks
499 PATH_LIMIT = 6
500 };
501 RegionNode* result_region = new RegionNode(PATH_LIMIT);
502 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
503 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
504 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
505 transform_later(result_region);
506 transform_later(result_i_o);
507 transform_later(result_memory);
508
509 // The slow_control path:
510 Node* slow_control;
511 Node* slow_i_o = *io;
512 Node* slow_mem = mem->memory_at(alias_idx);
513 DEBUG_ONLY(slow_control = (Node*) badAddress);
514
515 // Checked control path:
516 Node* checked_control = top();
517 Node* checked_mem = nullptr;
518 Node* checked_i_o = nullptr;
519 Node* checked_value = nullptr;
520
521 if (basic_elem_type == T_CONFLICT) {
522 assert(!dest_needs_zeroing, "");
523 Node* cv = generate_generic_arraycopy(ctrl, &mem,
524 adr_type,
525 src, src_offset, dest, dest_offset,
526 copy_length, acopy_to_uninitialized);
527 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
528 checked_control = *ctrl;
529 checked_i_o = *io;
530 checked_mem = mem->memory_at(alias_idx);
531 checked_value = cv;
532 *ctrl = top();
533 }
534
535 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
536 if (not_pos != nullptr) {
537 Node* local_ctrl = not_pos, *local_io = *io;
538 MergeMemNode* local_mem = MergeMemNode::make(mem);
539 transform_later(local_mem);
540
541 // (6) length must not be negative.
542 if (!length_never_negative) {
543 generate_negative_guard(&local_ctrl, copy_length, slow_region);
544 }
545
546 // copy_length is 0.
547 if (dest_needs_zeroing) {
548 assert(!local_ctrl->is_top(), "no ctrl?");
549 if (copy_length->eqv_uncast(dest_length)
550 || _igvn.find_int_con(dest_length, 1) <= 0) {
551 // There is no zeroing to do. No need for a secondary raw memory barrier.
552 } else {
553 // Clear the whole thing since there are no source elements to copy.
554 generate_clear_array(local_ctrl, local_mem,
555 adr_type, dest,
556 init_value, raw_init_value,
557 basic_elem_type,
558 intcon(0), nullptr,
559 alloc->in(AllocateNode::AllocSize));
560 // Use a secondary InitializeNode as raw memory barrier.
561 // Currently it is needed only on this path since other
562 // paths have stub or runtime calls as raw memory barriers.
563 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
564 Compile::AliasIdxRaw,
565 top());
566 transform_later(mb);
567 mb->set_req(TypeFunc::Control,local_ctrl);
568 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
569 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
570 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
571
572 InitializeNode* init = mb->as_Initialize();
573 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
574 }
575 }
576
577 // Present the results of the fast call.
578 result_region->init_req(zero_path, local_ctrl);
579 result_i_o ->init_req(zero_path, local_io);
580 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
581 }
582
583 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
584 // We have to initialize the *uncopied* part of the array to zero.
585 // The copy destination is the slice dest[off..off+len]. The other slices
586 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
587 Node* dest_size = alloc->in(AllocateNode::AllocSize);
588 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
589
590 // If there is a head section that needs zeroing, do it now.
591 if (_igvn.find_int_con(dest_offset, -1) != 0) {
592 generate_clear_array(*ctrl, mem,
593 adr_type, dest,
594 init_value, raw_init_value,
595 basic_elem_type,
596 intcon(0), dest_offset,
597 nullptr);
598 }
599
600 // Next, perform a dynamic check on the tail length.
601 // It is often zero, and we can win big if we prove this.
602 // There are two wins: Avoid generating the ClearArray
603 // with its attendant messy index arithmetic, and upgrade
604 // the copy to a more hardware-friendly word size of 64 bits.
605 Node* tail_ctl = nullptr;
606 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
607 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
608 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
609 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
610 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
611 }
612
613 // At this point, let's assume there is no tail.
614 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
615 // There is no tail. Try an upgrade to a 64-bit copy.
616 bool didit = false;
617 {
618 Node* local_ctrl = *ctrl, *local_io = *io;
619 MergeMemNode* local_mem = MergeMemNode::make(mem);
620 transform_later(local_mem);
621
622 didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
623 adr_type, basic_elem_type, alloc,
624 src, src_offset, dest, dest_offset,
625 dest_size, acopy_to_uninitialized);
626 if (didit) {
627 // Present the results of the block-copying fast call.
628 result_region->init_req(bcopy_path, local_ctrl);
629 result_i_o ->init_req(bcopy_path, local_io);
630 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
631 }
632 }
633 if (didit) {
634 *ctrl = top(); // no regular fast path
635 }
636 }
637
638 // Clear the tail, if any.
639 if (tail_ctl != nullptr) {
640 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
641 *ctrl = tail_ctl;
642 if (notail_ctl == nullptr) {
643 generate_clear_array(*ctrl, mem,
644 adr_type, dest,
645 init_value, raw_init_value,
646 basic_elem_type,
647 dest_tail, nullptr,
648 dest_size);
649 } else {
650 // Make a local merge.
651 Node* done_ctl = transform_later(new RegionNode(3));
652 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
653 done_ctl->init_req(1, notail_ctl);
654 done_mem->init_req(1, mem->memory_at(alias_idx));
655 generate_clear_array(*ctrl, mem,
656 adr_type, dest,
657 init_value, raw_init_value,
658 basic_elem_type,
659 dest_tail, nullptr,
660 dest_size);
661 done_ctl->init_req(2, *ctrl);
662 done_mem->init_req(2, mem->memory_at(alias_idx));
663 *ctrl = done_ctl;
664 mem->set_memory_at(alias_idx, done_mem);
665 }
666 }
667 }
668
669 BasicType copy_type = basic_elem_type;
670 assert(basic_elem_type != T_ARRAY, "caller must fix this");
671 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
672 // If src and dest have compatible element types, we can copy bits.
673 // Types S[] and D[] are compatible if D is a supertype of S.
674 //
675 // If they are not, we will use checked_oop_disjoint_arraycopy,
676 // which performs a fast optimistic per-oop check, and backs off
677 // further to JVM_ArrayCopy on the first per-oop check that fails.
678 // (Actually, we don't move raw bits only; the GC requires card marks.)
679
680 // We don't need a subtype check for validated copies and Object[].clone()
681 bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
682 ac->is_copyofrange_validated() || ac->is_clone_oop_array();
683 if (!skip_subtype_check) {
684 // Get the klass* for both src and dest
685 Node* src_klass = ac->in(ArrayCopyNode::SrcKlass);
686 Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
687
688 assert(src_klass != nullptr && dest_klass != nullptr, "should have klasses");
689
690 // Generate the subtype check.
691 // This might fold up statically, or then again it might not.
692 //
693 // Non-static example: Copying List<String>.elements to a new String[].
694 // The backing store for a List<String> is always an Object[],
695 // but its elements are always type String, if the generic types
696 // are correct at the source level.
697 //
698 // Test S[] against D[], not S against D, because (probably)
699 // the secondary supertype cache is less busy for S[] than S.
700 // This usually only matters when D is an interface.
701 Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn, nullptr, -1);
702 // Plug failing path into checked_oop_disjoint_arraycopy
703 if (not_subtype_ctrl != top()) {
704 Node* local_ctrl = not_subtype_ctrl;
705 MergeMemNode* local_mem = MergeMemNode::make(mem);
706 transform_later(local_mem);
707
708 // (At this point we can assume disjoint_bases, since types differ.)
709 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
710 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
711 Node* n1 = LoadKlassNode::make(_igvn, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
712 Node* dest_elem_klass = transform_later(n1);
713 Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
714 adr_type,
715 dest_elem_klass,
716 src, src_offset, dest, dest_offset,
717 ConvI2X(copy_length), acopy_to_uninitialized);
718 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
719 checked_control = local_ctrl;
720 checked_i_o = *io;
721 checked_mem = local_mem->memory_at(alias_idx);
722 checked_value = cv;
723 }
724 }
725 // At this point we know we do not need type checks on oop stores.
726
727 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
728 if (!bs->array_copy_requires_gc_barriers(alloc != nullptr, copy_type, false, false, BarrierSetC2::Expansion)) {
729 // If we do not need gc barriers, copy using the jint or jlong stub.
730 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
731 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
732 "sizes agree");
733 }
734 }
735
736 if (!(*ctrl)->is_top()) {
737 // Generate the fast path, if possible.
738 Node* local_ctrl = *ctrl;
739 MergeMemNode* local_mem = MergeMemNode::make(mem);
740 transform_later(local_mem);
741 generate_unchecked_arraycopy(&local_ctrl, &local_mem,
742 adr_type, copy_type, disjoint_bases,
743 src, src_offset, dest, dest_offset,
744 ConvI2X(copy_length), acopy_to_uninitialized);
745
746 // Present the results of the fast call.
747 result_region->init_req(fast_path, local_ctrl);
748 result_i_o ->init_req(fast_path, *io);
749 result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
750 }
751
752 // Here are all the slow paths up to this point, in one bundle:
753 assert(slow_region != nullptr, "allocated on entry");
754 slow_control = slow_region;
755 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
756
757 *ctrl = checked_control;
758 if (!(*ctrl)->is_top()) {
759 // Clean up after the checked call.
760 // The returned value is either 0 or -1^K,
761 // where K = number of partially transferred array elements.
762 Node* cmp = new CmpINode(checked_value, intcon(0));
763 transform_later(cmp);
764 Node* bol = new BoolNode(cmp, BoolTest::eq);
765 transform_later(bol);
766 IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
767 transform_later(iff);
768
769 // If it is 0, we are done, so transfer to the end.
770 Node* checks_done = new IfTrueNode(iff);
771 transform_later(checks_done);
772 result_region->init_req(checked_path, checks_done);
773 result_i_o ->init_req(checked_path, checked_i_o);
774 result_memory->init_req(checked_path, checked_mem);
775
776 // If it is not zero, merge into the slow call.
777 *ctrl = new IfFalseNode(iff);
778 transform_later(*ctrl);
779 RegionNode* slow_reg2 = new RegionNode(3);
780 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
781 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
782 transform_later(slow_reg2);
783 transform_later(slow_i_o2);
784 transform_later(slow_mem2);
785 slow_reg2 ->init_req(1, slow_control);
786 slow_i_o2 ->init_req(1, slow_i_o);
787 slow_mem2 ->init_req(1, slow_mem);
788 slow_reg2 ->init_req(2, *ctrl);
789 slow_i_o2 ->init_req(2, checked_i_o);
790 slow_mem2 ->init_req(2, checked_mem);
791
792 slow_control = slow_reg2;
793 slow_i_o = slow_i_o2;
794 slow_mem = slow_mem2;
795
796 if (alloc != nullptr) {
797 // We'll restart from the very beginning, after zeroing the whole thing.
798 // This can cause double writes, but that's OK since dest is brand new.
799 // So we ignore the low 31 bits of the value returned from the stub.
800 } else {
801 // We must continue the copy exactly where it failed, or else
802 // another thread might see the wrong number of writes to dest.
803 Node* checked_offset = new XorINode(checked_value, intcon(-1));
804 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
805 transform_later(checked_offset);
806 transform_later(slow_offset);
807 slow_offset->init_req(1, intcon(0));
808 slow_offset->init_req(2, checked_offset);
809
810 // Adjust the arguments by the conditionally incoming offset.
811 Node* src_off_plus = new AddINode(src_offset, slow_offset);
812 transform_later(src_off_plus);
813 Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
814 transform_later(dest_off_plus);
815 Node* length_minus = new SubINode(copy_length, slow_offset);
816 transform_later(length_minus);
817
818 // Tweak the node variables to adjust the code produced below:
819 src_offset = src_off_plus;
820 dest_offset = dest_off_plus;
821 copy_length = length_minus;
822 }
823 }
824 *ctrl = slow_control;
825 if (!(*ctrl)->is_top()) {
826 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
827 MergeMemNode* local_mem = MergeMemNode::make(mem);
828 transform_later(local_mem);
829
830 // Generate the slow path, if needed.
831 local_mem->set_memory_at(alias_idx, slow_mem);
832
833 if (dest_needs_zeroing) {
834 generate_clear_array(local_ctrl, local_mem,
835 adr_type, dest,
836 init_value, raw_init_value,
837 basic_elem_type,
838 intcon(0), nullptr,
839 alloc->in(AllocateNode::AllocSize));
840 }
841
842 local_mem = generate_slow_arraycopy(ac,
843 &local_ctrl, local_mem, &local_io,
844 adr_type,
845 src, src_offset, dest, dest_offset,
846 copy_length, /*dest_uninitialized*/false);
847
848 result_region->init_req(slow_call_path, local_ctrl);
849 result_i_o ->init_req(slow_call_path, local_io);
850 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
851 } else {
852 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
853 // projections were not extracted
854 }
855
856 // Remove unused edges.
857 for (uint i = 1; i < result_region->req(); i++) {
858 if (result_region->in(i) == nullptr) {
859 result_region->init_req(i, top());
860 }
861 }
862
863 // Finished; return the combined state.
864 *ctrl = result_region;
865 *io = result_i_o;
866 mem->set_memory_at(alias_idx, result_memory);
867
868 // mem no longer guaranteed to stay a MergeMemNode
869 Node* out_mem = mem;
870 DEBUG_ONLY(mem = nullptr);
871
872 // The memory edges above are precise in order to model effects around
873 // array copies accurately to allow value numbering of field loads around
874 // arraycopy. Such field loads, both before and after, are common in Java
875 // collections and similar classes involving header/array data structures.
876 //
877 // But with low number of register or when some registers are used or killed
878 // by arraycopy calls it causes registers spilling on stack. See 6544710.
879 // The next memory barrier is added to avoid it. If the arraycopy can be
880 // optimized away (which it can, sometimes) then we can manually remove
881 // the membar also.
882 //
883 // Do not let reads from the cloned object float above the arraycopy.
884 if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
885 // Do not let stores that initialize this object be reordered with
886 // a subsequent store that would make this object accessible by
887 // other threads.
888 assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
889 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
890 } else {
891 int mem_bar_alias_idx = Compile::AliasIdxBot;
892 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
893 // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
894 // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
895 // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
896 // broken graph.
897 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
898 }
899 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
900 }
901
902 assert((*ctrl)->is_Proj(), "MemBar control projection");
903 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
904 (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
905
906 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
907 if (_callprojs->fallthrough_ioproj != nullptr) {
908 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
909 }
910 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
911
912 #ifdef ASSERT
913 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
914 if (dest_t->is_known_instance()) {
915 ArrayCopyNode* ac = nullptr;
916 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
917 assert(ac == nullptr, "no arraycopy anymore");
918 }
919 #endif
920
921 return out_mem;
922 }
923
924 // Helper for initialization of arrays, creating a ClearArray.
925 // It writes zero bits in [start..end), within the body of an array object.
926 // The memory effects are all chained onto the 'adr_type' alias category.
927 //
928 // Since the object is otherwise uninitialized, we are free
929 // to put a little "slop" around the edges of the cleared area,
930 // as long as it does not go back into the array's header,
931 // or beyond the array end within the heap.
932 //
933 // The lower edge can be rounded down to the nearest jint and the
934 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
935 //
936 // Arguments:
937 // adr_type memory slice where writes are generated
938 // dest oop of the destination array
939 // basic_elem_type element type of the destination
940 // slice_idx array index of first element to store
941 // slice_len number of elements to store (or null)
942 // dest_size total size in bytes of the array object
943 //
944 // Exactly one of slice_len or dest_size must be non-null.
945 // If dest_size is non-null, zeroing extends to the end of the object.
946 // If slice_len is non-null, the slice_idx value must be a constant.
947 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
948 const TypePtr* adr_type,
949 Node* dest,
950 Node* val,
951 Node* raw_val,
952 BasicType basic_elem_type,
953 Node* slice_idx,
954 Node* slice_len,
955 Node* dest_size) {
956 // one or the other but not both of slice_len and dest_size:
957 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
958 if (slice_len == nullptr) slice_len = top();
959 if (dest_size == nullptr) dest_size = top();
960
961 uint alias_idx = C->get_alias_index(adr_type);
962
963 // operate on this memory slice:
964 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
965
966 // scaling and rounding of indexes:
967 int scale = exact_log2(type2aelembytes(basic_elem_type));
968 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
969 int clear_low = (-1 << scale) & (BytesPerInt - 1);
970 int bump_bit = (-1 << scale) & BytesPerInt;
971
972 // determine constant starts and ends
973 const intptr_t BIG_NEG = -128;
974 assert(BIG_NEG + 2*abase < 0, "neg enough");
975 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
976 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
977 if (slice_len_con == 0) {
978 return; // nothing to do here
979 }
980 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
981 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
982 if (slice_idx_con >= 0 && slice_len_con >= 0) {
983 assert(end_con < 0, "not two cons");
984 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
985 BytesPerLong);
986 }
987
988 if (start_con >= 0 && end_con >= 0) {
989 // Constant start and end. Simple.
990 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
991 start_con, end_con, &_igvn);
992 } else if (start_con >= 0 && dest_size != top()) {
993 // Constant start, pre-rounded end after the tail of the array.
994 Node* end = dest_size;
995 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
996 start_con, end, &_igvn);
997 } else if (start_con >= 0 && slice_len != top()) {
998 // Constant start, non-constant end. End needs rounding up.
999 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
1000 intptr_t end_base = abase + (slice_idx_con << scale);
1001 int end_round = (-1 << scale) & (BytesPerLong - 1);
1002 Node* end = ConvI2X(slice_len);
1003 if (scale != 0)
1004 end = transform_later(new LShiftXNode(end, intcon(scale) ));
1005 end_base += end_round;
1006 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
1007 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
1008 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1009 start_con, end, &_igvn);
1010 } else if (start_con < 0 && dest_size != top()) {
1011 // Non-constant start, pre-rounded end after the tail of the array.
1012 // This is almost certainly a "round-to-end" operation.
1013 Node* start = slice_idx;
1014 start = ConvI2X(start);
1015 if (scale != 0)
1016 start = transform_later(new LShiftXNode( start, intcon(scale) ));
1017 start = transform_later(new AddXNode(start, MakeConX(abase)) );
1018 if ((bump_bit | clear_low) != 0) {
1019 int to_clear = (bump_bit | clear_low);
1020 // Align up mod 8, then store a jint zero unconditionally
1021 // just before the mod-8 boundary.
1022 if (((abase + bump_bit) & ~to_clear) - bump_bit
1023 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1024 bump_bit = 0;
1025 assert((abase & to_clear) == 0, "array base must be long-aligned");
1026 } else {
1027 // Bump 'start' up to (or past) the next jint boundary:
1028 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1029 assert((abase & clear_low) == 0, "array base must be int-aligned");
1030 }
1031 // Round bumped 'start' down to jlong boundary in body of array.
1032 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1033 if (bump_bit != 0) {
1034 // Store a zero to the immediately preceding jint:
1035 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1036 Node* p1 = basic_plus_adr(dest, x1);
1037 if (val == nullptr) {
1038 assert(raw_val == nullptr, "val may not be null");
1039 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1040 } else {
1041 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1042 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1043 }
1044 mem = transform_later(mem);
1045 }
1046 }
1047 Node* end = dest_size; // pre-rounded
1048 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1049 start, end, &_igvn);
1050 } else {
1051 // Non-constant start, unrounded non-constant end.
1052 // (Nobody zeroes a random midsection of an array using this routine.)
1053 ShouldNotReachHere(); // fix caller
1054 }
1055
1056 // Done.
1057 merge_mem->set_memory_at(alias_idx, mem);
1058 }
1059
1060 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1061 const TypePtr* adr_type,
1062 BasicType basic_elem_type,
1063 AllocateNode* alloc,
1064 Node* src, Node* src_offset,
1065 Node* dest, Node* dest_offset,
1066 Node* dest_size, bool dest_uninitialized) {
1067 // See if there is an advantage from block transfer.
1068 int scale = exact_log2(type2aelembytes(basic_elem_type));
1069 if (scale >= LogBytesPerLong)
1070 return false; // it is already a block transfer
1071
1072 // Look at the alignment of the starting offsets.
1073 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1074
1075 intptr_t src_off_con = (intptr_t) _igvn.find_int_con(src_offset, -1);
1076 intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1077 if (src_off_con < 0 || dest_off_con < 0) {
1078 // At present, we can only understand constants.
1079 return false;
1080 }
1081
1082 intptr_t src_off = abase + (src_off_con << scale);
1083 intptr_t dest_off = abase + (dest_off_con << scale);
1084
1085 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1086 // Non-aligned; too bad.
1087 // One more chance: Pick off an initial 32-bit word.
1088 // This is a common case, since abase can be odd mod 8.
1089 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1090 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1091 Node* sptr = basic_plus_adr(src, src_off);
1092 Node* dptr = basic_plus_adr(dest, dest_off);
1093 const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1094 assert(s_adr_type->isa_aryptr(), "impossible slice");
1095 uint s_alias_idx = C->get_alias_index(s_adr_type);
1096 uint d_alias_idx = C->get_alias_index(adr_type);
1097 bool is_mismatched = (basic_elem_type != T_INT);
1098 Node* sval = transform_later(
1099 LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1100 TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1101 false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1102 Node* st = transform_later(
1103 StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1104 sval, T_INT, MemNode::unordered));
1105 if (is_mismatched) {
1106 st->as_Store()->set_mismatched_access();
1107 }
1108 (*mem)->set_memory_at(d_alias_idx, st);
1109 src_off += BytesPerInt;
1110 dest_off += BytesPerInt;
1111 } else {
1112 return false;
1113 }
1114 }
1115 assert(src_off % BytesPerLong == 0, "");
1116 assert(dest_off % BytesPerLong == 0, "");
1117
1118 // Do this copy by giant steps.
1119 Node* sptr = basic_plus_adr(src, src_off);
1120 Node* dptr = basic_plus_adr(dest, dest_off);
1121 Node* countx = dest_size;
1122 countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1123 countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1124
1125 bool disjoint_bases = true; // since alloc isn't null
1126 generate_unchecked_arraycopy(ctrl, mem,
1127 adr_type, T_LONG, disjoint_bases,
1128 sptr, nullptr, dptr, nullptr, countx, dest_uninitialized);
1129
1130 return true;
1131 }
1132
1133 // Helper function; generates code for the slow case.
1134 // We make a call to a runtime method which emulates the native method,
1135 // but without the native wrapper overhead.
1136 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1137 Node** ctrl, Node* mem, Node** io,
1138 const TypePtr* adr_type,
1139 Node* src, Node* src_offset,
1140 Node* dest, Node* dest_offset,
1141 Node* copy_length, bool dest_uninitialized) {
1142 assert(!dest_uninitialized, "Invariant");
1143
1144 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1145 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1146 "slow_arraycopy", TypePtr::BOTTOM);
1147
1148 call->init_req(TypeFunc::Control, *ctrl);
1149 call->init_req(TypeFunc::I_O , *io);
1150 call->init_req(TypeFunc::Memory , mem);
1151 call->init_req(TypeFunc::ReturnAdr, top());
1152 call->init_req(TypeFunc::FramePtr, top());
1153 call->init_req(TypeFunc::Parms+0, src);
1154 call->init_req(TypeFunc::Parms+1, src_offset);
1155 call->init_req(TypeFunc::Parms+2, dest);
1156 call->init_req(TypeFunc::Parms+3, dest_offset);
1157 call->init_req(TypeFunc::Parms+4, copy_length);
1158 call->copy_call_debug_info(&_igvn, ac);
1159
1160 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1161 _igvn.replace_node(ac, call);
1162 transform_later(call);
1163
1164 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1165 *ctrl = _callprojs->fallthrough_catchproj->clone();
1166 transform_later(*ctrl);
1167
1168 Node* m = _callprojs->fallthrough_memproj->clone();
1169 transform_later(m);
1170
1171 uint alias_idx = C->get_alias_index(adr_type);
1172 MergeMemNode* out_mem;
1173 if (alias_idx != Compile::AliasIdxBot) {
1174 out_mem = MergeMemNode::make(mem);
1175 out_mem->set_memory_at(alias_idx, m);
1176 } else {
1177 out_mem = MergeMemNode::make(m);
1178 }
1179 transform_later(out_mem);
1180
1181 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1182 // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1183 if (_callprojs->fallthrough_ioproj != nullptr) {
1184 *io = _callprojs->fallthrough_ioproj->clone();
1185 transform_later(*io);
1186 } else {
1187 *io = nullptr;
1188 }
1189
1190 return out_mem;
1191 }
1192
1193 // Helper function; generates code for cases requiring runtime checks.
1194 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1195 const TypePtr* adr_type,
1196 Node* dest_elem_klass,
1197 Node* src, Node* src_offset,
1198 Node* dest, Node* dest_offset,
1199 Node* copy_length, bool dest_uninitialized) {
1200 if ((*ctrl)->is_top()) return nullptr;
1201
1202 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1203 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1204 return nullptr;
1205 }
1206
1207 // Pick out the parameters required to perform a store-check
1208 // for the target array. This is an optimistic check. It will
1209 // look in each non-null element's class, at the desired klass's
1210 // super_check_offset, for the desired klass.
1211 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1212 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1213 Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1214 Node* check_offset = ConvI2X(transform_later(n3));
1215 Node* check_value = dest_elem_klass;
1216
1217 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1218 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1219
1220 const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1221 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1222 src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1223
1224 finish_arraycopy_call(call, ctrl, mem, adr_type);
1225
1226 Node* proj = new ProjNode(call, TypeFunc::Parms);
1227 transform_later(proj);
1228
1229 return proj;
1230 }
1231
1232 // Helper function; generates code for cases requiring runtime checks.
1233 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1234 const TypePtr* adr_type,
1235 Node* src, Node* src_offset,
1236 Node* dest, Node* dest_offset,
1237 Node* copy_length, bool dest_uninitialized) {
1238 if ((*ctrl)->is_top()) return nullptr;
1239 assert(!dest_uninitialized, "Invariant");
1240
1241 address copyfunc_addr = StubRoutines::generic_arraycopy();
1242 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1243 return nullptr;
1244 }
1245
1246 const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1247 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1248 src, src_offset, dest, dest_offset, copy_length);
1249
1250 finish_arraycopy_call(call, ctrl, mem, adr_type);
1251
1252 Node* proj = new ProjNode(call, TypeFunc::Parms);
1253 transform_later(proj);
1254
1255 return proj;
1256 }
1257
1258 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1259 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1260 const TypePtr* adr_type,
1261 BasicType basic_elem_type,
1262 bool disjoint_bases,
1263 Node* src, Node* src_offset,
1264 Node* dest, Node* dest_offset,
1265 Node* copy_length, bool dest_uninitialized) {
1266 if ((*ctrl)->is_top()) {
1267 return;
1268 }
1269
1270 Node* src_start = src;
1271 Node* dest_start = dest;
1272 if (src_offset != nullptr || dest_offset != nullptr) {
1273 src_start = array_element_address(src, src_offset, basic_elem_type);
1274 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1275 }
1276
1277 // Figure out which arraycopy runtime method to call.
1278 const char* copyfunc_name = "arraycopy";
1279 address copyfunc_addr =
1280 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1281 disjoint_bases, copyfunc_name, dest_uninitialized);
1282
1283 Node* result_memory = nullptr;
1284 RegionNode* exit_block = nullptr;
1285 if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1286 Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1287 generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1288 copy_length, src_start, dest_start, basic_elem_type);
1289 }
1290
1291 const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1292 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1293 src_start, dest_start, copy_length XTOP);
1294
1295 finish_arraycopy_call(call, ctrl, mem, adr_type);
1296
1297 // Connecting remaining edges for exit_block coming from stub_block.
1298 if (exit_block) {
1299 exit_block->init_req(2, *ctrl);
1300
1301 // Memory edge corresponding to stub_region.
1302 result_memory->init_req(2, *mem);
1303
1304 uint alias_idx = C->get_alias_index(adr_type);
1305 if (alias_idx != Compile::AliasIdxBot) {
1306 *mem = MergeMemNode::make(*mem);
1307 (*mem)->set_memory_at(alias_idx, result_memory);
1308 } else {
1309 *mem = MergeMemNode::make(result_memory);
1310 }
1311 transform_later(*mem);
1312 *ctrl = exit_block;
1313 }
1314 }
1315
1316 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1317 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1318 Node*& dest_length) {
1319 #ifdef ASSERT
1320 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1321 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1322 bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1323 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1324 #endif
1325 int elem_size = top_dest->flat_elem_size();
1326 if (elem_size >= 8) {
1327 if (elem_size > 8) {
1328 // treat as array of long but scale length, src offset and dest offset
1329 assert((elem_size % 8) == 0, "not a power of 2?");
1330 int factor = elem_size / 8;
1331 length = transform_later(new MulINode(length, intcon(factor)));
1332 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1333 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1334 if (dest_length != nullptr) {
1335 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1336 }
1337 elem_size = 8;
1338 }
1339 dest_elem = T_LONG;
1340 } else if (elem_size == 4) {
1341 dest_elem = T_INT;
1342 } else if (elem_size == 2) {
1343 dest_elem = T_CHAR;
1344 } else if (elem_size == 1) {
1345 dest_elem = T_BYTE;
1346 } else {
1347 ShouldNotReachHere();
1348 }
1349 return TypeRawPtr::BOTTOM;
1350 }
1351
1352 #undef XTOP
1353
1354 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1355 Node* ctrl = ac->in(TypeFunc::Control);
1356 Node* io = ac->in(TypeFunc::I_O);
1357 Node* src = ac->in(ArrayCopyNode::Src);
1358 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1359 Node* dest = ac->in(ArrayCopyNode::Dest);
1360 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1361 Node* length = ac->in(ArrayCopyNode::Length);
1362 MergeMemNode* merge_mem = nullptr;
1363
1364 if (ac->is_clonebasic()) {
1365 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1366 bs->clone_at_expansion(this, ac);
1367 return;
1368 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1369 const Type* src_type = _igvn.type(src);
1370 const Type* dest_type = _igvn.type(dest);
1371 const TypeAryPtr* top_src = src_type->isa_aryptr();
1372 // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1373 // Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1374 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1375 BasicType dest_elem = T_OBJECT;
1376 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1377 dest_elem = top_dest->elem()->array_element_basic_type();
1378 }
1379 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1380
1381 if (top_src != nullptr && top_src->is_flat()) {
1382 // If src is flat, dest is guaranteed to be flat as well
1383 top_dest = top_src;
1384 }
1385
1386 AllocateArrayNode* alloc = nullptr;
1387 Node* dest_length = nullptr;
1388 if (ac->is_alloc_tightly_coupled()) {
1389 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1390 assert(alloc != nullptr, "expect alloc");
1391 dest_length = alloc->in(AllocateNode::ALength);
1392 }
1393
1394 Node* mem = ac->in(TypeFunc::Memory);
1395 const TypePtr* adr_type = nullptr;
1396 if (top_dest != nullptr && top_dest->is_flat()) {
1397 assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1398 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1399 // on all slices to prevent writes into the source from floating below the arraycopy.
1400 int mem_bar_alias_idx = Compile::AliasIdxBot;
1401 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1402 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1403 }
1404 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1405 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1406 } else {
1407 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1408 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1409 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1410 }
1411 if (ac->_src_type != ac->_dest_type) {
1412 adr_type = TypeRawPtr::BOTTOM;
1413 }
1414 }
1415 merge_mem = MergeMemNode::make(mem);
1416 transform_later(merge_mem);
1417
1418 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1419 adr_type, dest_elem,
1420 src, src_offset, dest, dest_offset, length,
1421 dest_length,
1422 true, ac->has_negative_length_guard());
1423
1424 return;
1425 }
1426
1427 AllocateArrayNode* alloc = nullptr;
1428 if (ac->is_alloc_tightly_coupled()) {
1429 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1430 assert(alloc != nullptr, "expect alloc");
1431 }
1432
1433 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1434
1435 // Compile time checks. If any of these checks cannot be verified at compile time,
1436 // we do not make a fast path for this call. Instead, we let the call remain as it
1437 // is. The checks we choose to mandate at compile time are:
1438 //
1439 // (1) src and dest are arrays.
1440 const Type* src_type = src->Value(&_igvn);
1441 const Type* dest_type = dest->Value(&_igvn);
1442 const TypeAryPtr* top_src = src_type->isa_aryptr();
1443 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1444
1445 BasicType src_elem = T_CONFLICT;
1446 BasicType dest_elem = T_CONFLICT;
1447
1448 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1449 src_elem = top_src->elem()->array_element_basic_type();
1450 }
1451 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1452 dest_elem = top_dest->elem()->array_element_basic_type();
1453 }
1454 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1455 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1456
1457 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1458 src_elem = dest_elem;
1459 }
1460
1461 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1462 // Conservatively insert a memory barrier on all memory slices.
1463 // Do not let writes into the source float below the arraycopy.
1464 {
1465 Node* mem = ac->in(TypeFunc::Memory);
1466 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1467
1468 merge_mem = MergeMemNode::make(mem);
1469 transform_later(merge_mem);
1470 }
1471
1472 // Call StubRoutines::generic_arraycopy stub.
1473 generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1474 TypeRawPtr::BOTTOM, T_CONFLICT,
1475 src, src_offset, dest, dest_offset, length,
1476 nullptr,
1477 // If a negative length guard was generated for the ArrayCopyNode,
1478 // the length of the array can never be negative.
1479 false, ac->has_negative_length_guard());
1480 return;
1481 }
1482
1483 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1484
1485 // (2) src and dest arrays must have elements of the same BasicType
1486 // Figure out the size and type of the elements we will be copying.
1487 //
1488 // We have no stub to copy flat inline type arrays with oop
1489 // fields if we need to emit write barriers.
1490 //
1491 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1492 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1493 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1494 bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1495 // The component types are not the same or are not recognized. Punt.
1496 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1497 {
1498 Node* mem = ac->in(TypeFunc::Memory);
1499 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1500 }
1501
1502 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1503 if (_callprojs->fallthrough_ioproj != nullptr) {
1504 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1505 }
1506 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1507 return;
1508 }
1509
1510 //---------------------------------------------------------------------------
1511 // We will make a fast path for this call to arraycopy.
1512
1513 // We have the following tests left to perform:
1514 //
1515 // (3) src and dest must not be null.
1516 // (4) src_offset must not be negative.
1517 // (5) dest_offset must not be negative.
1518 // (6) length must not be negative.
1519 // (7) src_offset + length must not exceed length of src.
1520 // (8) dest_offset + length must not exceed length of dest.
1521 // (9) each element of an oop array must be assignable
1522
1523 Node* mem = ac->in(TypeFunc::Memory);
1524 if (top_dest->is_flat()) {
1525 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1526 // on all slices to prevent writes into the source from floating below the arraycopy.
1527 int mem_bar_alias_idx = Compile::AliasIdxBot;
1528 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1529 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1530 }
1531 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1532 }
1533 merge_mem = MergeMemNode::make(mem);
1534 transform_later(merge_mem);
1535
1536 RegionNode* slow_region = new RegionNode(1);
1537 transform_later(slow_region);
1538
1539 if (!ac->is_arraycopy_validated()) {
1540 // (3) operands must not be null
1541 // We currently perform our null checks with the null_check routine.
1542 // This means that the null exceptions will be reported in the caller
1543 // rather than (correctly) reported inside of the native arraycopy call.
1544 // This should be corrected, given time. We do our null check with the
1545 // stack pointer restored.
1546 // null checks done library_call.cpp
1547
1548 // (4) src_offset must not be negative.
1549 generate_negative_guard(&ctrl, src_offset, slow_region);
1550
1551 // (5) dest_offset must not be negative.
1552 generate_negative_guard(&ctrl, dest_offset, slow_region);
1553
1554 // (6) length must not be negative (moved to generate_arraycopy()).
1555 // generate_negative_guard(length, slow_region);
1556
1557 // (7) src_offset + length must not exceed length of src.
1558 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1559 assert(alen != nullptr, "need src len");
1560 generate_limit_guard(&ctrl,
1561 src_offset, length,
1562 alen,
1563 slow_region);
1564
1565 // (8) dest_offset + length must not exceed length of dest.
1566 alen = ac->in(ArrayCopyNode::DestLen);
1567 assert(alen != nullptr, "need dest len");
1568 generate_limit_guard(&ctrl,
1569 dest_offset, length,
1570 alen,
1571 slow_region);
1572
1573 // (9) each element of an oop array must be assignable
1574 // The generate_arraycopy subroutine checks this.
1575
1576 // TODO 8350865 Fix below logic. Also handle atomicity.
1577 // 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.
1578 if (!(top_src->is_flat() && top_dest->is_flat())) {
1579 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1580 generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1581 }
1582
1583 // Handle inline type arrays
1584 if (!top_src->is_flat()) {
1585 if (UseArrayFlattening && !top_src->is_not_flat()) {
1586 // Src might be flat and dest might not be flat. Go to the slow path if src is flat.
1587 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1588 }
1589 if (EnableValhalla) {
1590 // No validation. The subtype check emitted at macro expansion time will not go to the slow
1591 // path but call checkcast_arraycopy which can not handle flat/null-free inline type arrays.
1592 generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1593 }
1594 } else {
1595 assert(top_dest->is_flat(), "dest array must be flat");
1596 }
1597 }
1598
1599 // This is where the memory effects are placed:
1600 const TypePtr* adr_type = nullptr;
1601 Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1602
1603 if (top_src->is_flat() && top_dest->is_flat()) {
1604 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1605 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1606 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1607 } else {
1608 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1609 }
1610
1611 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1612 adr_type, dest_elem,
1613 src, src_offset, dest, dest_offset, length,
1614 dest_length,
1615 // If a negative length guard was generated for the ArrayCopyNode,
1616 // the length of the array can never be negative.
1617 false, ac->has_negative_length_guard(),
1618 slow_region);
1619 }