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 "oops/objArrayKlass.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/convertnode.hpp"
32 #include "opto/graphKit.hpp"
33 #include "opto/macro.hpp"
34 #include "opto/runtime.hpp"
35 #include "opto/vectornode.hpp"
36 #include "runtime/stubRoutines.hpp"
37 #include "utilities/align.hpp"
38 #include "utilities/powerOfTwo.hpp"
39
40 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, int alias_idx, Node* precedent) {
41 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
42 mb->init_req(TypeFunc::Control, *ctrl);
43 mb->init_req(TypeFunc::Memory, *mem);
44 transform_later(mb);
45 *ctrl = new ProjNode(mb,TypeFunc::Control);
46 transform_later(*ctrl);
47 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
48 transform_later(mem_proj);
49 if (alias_idx == Compile::AliasIdxBot) {
50 *mem = mem_proj;
51 } else {
52 MergeMemNode* mm = (*mem)->clone()->as_MergeMem();
53 mm->set_memory_at(alias_idx, mem_proj);
54 transform_later(mm);
55 *mem = mm;
56 }
57 }
58
59 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
60 uint shift = exact_log2(type2aelembytes(elembt));
61 const TypeAryPtr* array_type = _igvn.type(ary)->isa_aryptr();
62 if (array_type != nullptr && array_type->is_aryptr()->is_flat()) {
63 // Use T_FLAT_ELEMENT to get proper alignment with COH when fetching the array element address.
64 elembt = T_FLAT_ELEMENT;
65 }
66 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
67 Node* base = basic_plus_adr(ary, header);
68 #ifdef _LP64
69 // see comment in GraphKit::array_element_address
70 int index_max = max_jint - 1; // array size is max_jint, index is one less
71 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
72 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
73 #endif
74 Node* scale = new LShiftXNode(idx, intcon(shift));
75 transform_later(scale);
76 return basic_plus_adr(ary, base, scale);
77 }
78
79 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
80 return transform_later(new ConvI2LNode(offset));
81 }
82
83 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
84 const TypeFunc* call_type, address call_addr,
85 const char* call_name,
86 const TypePtr* adr_type,
87 Node* parm0, Node* parm1,
88 Node* parm2, Node* parm3,
89 Node* parm4, Node* parm5,
90 Node* parm6, Node* parm7) {
91 Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
92 call->init_req(TypeFunc::Control, ctrl);
93 call->init_req(TypeFunc::I_O , top());
94 call->init_req(TypeFunc::Memory , mem);
95 call->init_req(TypeFunc::ReturnAdr, top());
96 call->init_req(TypeFunc::FramePtr, top());
97
98 // Hook each parm in order. Stop looking at the first null.
99 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
100 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
101 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
102 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
103 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
104 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
105 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
106 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
107 /* close each nested if ===> */ } } } } } } } }
108 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
109
110 return call;
111 }
112
113
114 //------------------------------generate_guard---------------------------
115 // Helper function for generating guarded fast-slow graph structures.
116 // The given 'test', if true, guards a slow path. If the test fails
117 // then a fast path can be taken. (We generally hope it fails.)
118 // In all cases, GraphKit::control() is updated to the fast path.
119 // The returned value represents the control for the slow path.
120 // The return value is never 'top'; it is either a valid control
121 // or null if it is obvious that the slow path can never be taken.
122 // Also, if region and the slow control are not null, the slow edge
123 // is appended to the region.
124 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
125 if ((*ctrl)->is_top()) {
126 // Already short circuited.
127 return nullptr;
128 }
129 // Build an if node and its projections.
130 // If test is true we take the slow path, which we assume is uncommon.
131 if (_igvn.type(test) == TypeInt::ZERO) {
132 // The slow branch is never taken. No need to build this guard.
133 return nullptr;
134 }
135
136 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
137 transform_later(iff);
138
139 Node* if_slow = new IfTrueNode(iff);
140 transform_later(if_slow);
141
142 if (region != nullptr) {
143 region->add_req(if_slow);
144 }
145
146 Node* if_fast = new IfFalseNode(iff);
147 transform_later(if_fast);
148
149 *ctrl = if_fast;
150
151 return if_slow;
152 }
153
154 Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
155 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
156 }
157
158 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
159 return generate_guard(ctrl, test, region, PROB_FAIR);
160 }
161
162 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
163 if ((*ctrl)->is_top())
164 return; // already stopped
165 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
166 return; // index is already adequately typed
167 Node* cmp_lt = new CmpINode(index, intcon(0));
168 transform_later(cmp_lt);
169 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
170 transform_later(bol_lt);
171 generate_guard(ctrl, bol_lt, region, PROB_MIN);
172 }
173
174 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
175 if ((*ctrl)->is_top())
176 return; // already stopped
177 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
178 if (zero_offset && subseq_length->eqv_uncast(array_length))
179 return; // common case of whole-array copy
180 Node* last = subseq_length;
181 if (!zero_offset) { // last += offset
182 last = new AddINode(last, offset);
183 transform_later(last);
184 }
185 Node* cmp_lt = new CmpUNode(array_length, last);
186 transform_later(cmp_lt);
187 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
188 transform_later(bol_lt);
189 generate_guard(ctrl, bol_lt, region, PROB_MIN);
190 }
191
192 //
193 // Partial in-lining handling for smaller conjoint/disjoint array copies having
194 // length(in bytes) less than ArrayOperationPartialInlineSize.
195 // if (length <= ArrayOperationPartialInlineSize) {
196 // partial_inlining_block:
197 // mask = Mask_Gen
198 // vload = LoadVectorMasked src , mask
199 // StoreVectorMasked dst, mask, vload
200 // } else {
201 // stub_block:
202 // callstub array_copy
203 // }
204 // exit_block:
205 // Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
206 // mem = MergeMem (Phi)
207 // control = stub_block
208 //
209 // Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
210 // edges. Remaining edges for exit_block coming from stub_block are connected by the caller
211 // post stub nodes creation.
212 //
213
214 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
215 RegionNode** exit_block, Node** result_memory, Node* length,
216 Node* src_start, Node* dst_start, BasicType type) {
217 int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
218
219 const TypeLong* length_type = _igvn.type(length)->isa_long();
220 if (length_type == nullptr) {
221 assert(_igvn.type(length) == Type::TOP, "");
222 return;
223 }
224
225 const TypeLong* inline_range = TypeLong::make(0, inline_limit, Type::WidenMin);
226 if (length_type->join(inline_range) == Type::TOP) {
227 // The ranges do not intersect, the inline check will surely fail
228 return;
229 }
230
231 // Return if the target does not supports masked load/stores.
232 int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, length_type->_hi);
233 if (!Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type) ||
234 !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
235 !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
236 return;
237 }
238
239 Node* cmp_le = new CmpULNode(length, longcon(inline_limit));
240 transform_later(cmp_le);
241 Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
242 transform_later(bol_le);
243 Node* inline_block = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
244 Node* stub_block = *ctrl;
245
246 Node* casted_length = new CastLLNode(inline_block, length, inline_range, ConstraintCastNode::DependencyType::FloatingNarrowing);
247 transform_later(casted_length);
248 Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
249 transform_later(mask_gen);
250
251 unsigned vec_size = lane_count * type2aelembytes(type);
252 if (C->max_vector_size() < vec_size) {
253 C->set_max_vector_size(vec_size);
254 }
255
256 const TypePtr* src_adr_type = _igvn.type(src_start)->isa_ptr();
257 const TypeVect * vt = TypeVect::make(type, lane_count);
258 Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
259 Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
260 src_adr_type, vt, mask_gen);
261 transform_later(masked_load);
262
263 mm = (*mem)->memory_at(C->get_alias_index(adr_type));
264 Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
265 masked_load, adr_type, mask_gen);
266 transform_later(masked_store);
267
268 // Convergence region for inline_block and stub_block.
269 *exit_block = new RegionNode(3);
270 transform_later(*exit_block);
271 (*exit_block)->init_req(1, inline_block);
272 *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
273 transform_later(*result_memory);
274 (*result_memory)->init_req(1, masked_store);
275
276 *ctrl = stub_block;
277 }
278
279
280 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
281 if ((*ctrl)->is_top()) return nullptr;
282
283 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
284 return nullptr; // index is already adequately typed
285 Node* cmp_le = new CmpINode(index, intcon(0));
286 transform_later(cmp_le);
287 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
288 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
289 transform_later(bol_le);
290 Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
291
292 return is_notp;
293 }
294
295 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
296 // Load markword and check if obj is locked
297 Node* mark = make_load(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
298 Node* locked_bit = MakeConX(markWord::unlocked_value);
299 locked_bit = transform_later(new AndXNode(locked_bit, mark));
300 Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
301 Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
302 IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
303 Node* locked_region = transform_later(new RegionNode(3));
304 Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
305
306 // Unlocked: Use bits from mark word
307 locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
308 mark_phi->init_req(1, mark);
309
310 // Locked: Load prototype header from klass
311 *ctrl = transform_later(new IfFalseNode(iff));
312 // Make loads control dependent to make sure they are only executed if array is locked
313 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
314 Node* klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
315 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
316 Node* proto = transform_later(LoadNode::make(_igvn, *ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
317
318 locked_region->init_req(2, *ctrl);
319 mark_phi->init_req(2, proto);
320 *ctrl = locked_region;
321
322 // Now check if mark word bits are set
323 Node* mask = MakeConX(mask_val);
324 Node* masked = transform_later(new AndXNode(mark_phi, mask));
325 cmp = transform_later(new CmpXNode(masked, mask));
326 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
327 return generate_fair_guard(ctrl, bol, region);
328 }
329
330 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
331 return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
332 }
333
334 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
335 return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
336 }
337
338 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
339 transform_later(call);
340
341 *ctrl = new ProjNode(call,TypeFunc::Control);
342 transform_later(*ctrl);
343 Node* newmem = new ProjNode(call, TypeFunc::Memory);
344 transform_later(newmem);
345
346 uint alias_idx = C->get_alias_index(adr_type);
347 if (alias_idx != Compile::AliasIdxBot) {
348 *mem = MergeMemNode::make(*mem);
349 (*mem)->set_memory_at(alias_idx, newmem);
350 } else {
351 *mem = MergeMemNode::make(newmem);
352 }
353 transform_later(*mem);
354 }
355
356 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
357 Node* src_offset,
358 Node* dest_offset,
359 bool disjoint_bases,
360 const char* &name,
361 bool dest_uninitialized) {
362 const TypeInt* src_offset_inttype = _igvn.find_int_type(src_offset);
363 const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
364
365 bool aligned = false;
366 bool disjoint = disjoint_bases;
367
368 // if the offsets are the same, we can treat the memory regions as
369 // disjoint, because either the memory regions are in different arrays,
370 // or they are identical (which we can treat as disjoint.) We can also
371 // treat a copy with a destination index less that the source index
372 // as disjoint since a low->high copy will work correctly in this case.
373 if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
374 dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
375 // both indices are constants
376 int s_offs = src_offset_inttype->get_con();
377 int d_offs = dest_offset_inttype->get_con();
378 int element_size = type2aelembytes(t);
379 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
380 ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
381 if (s_offs >= d_offs) disjoint = true;
382 } else if (src_offset == dest_offset && src_offset != nullptr) {
383 // This can occur if the offsets are identical non-constants.
384 disjoint = true;
385 }
386
387 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
388 }
389
390 #define XTOP LP64_ONLY(COMMA top())
391
392 // Generate an optimized call to arraycopy.
393 // Caller must guard against non-arrays.
394 // Caller must determine a common array basic-type for both arrays.
395 // Caller must validate offsets against array bounds.
396 // The slow_region has already collected guard failure paths
397 // (such as out of bounds length or non-conformable array types).
398 // The generated code has this shape, in general:
399 //
400 // if (length == 0) return // via zero_path
401 // slowval = -1
402 // if (types unknown) {
403 // slowval = call generic copy loop
404 // if (slowval == 0) return // via checked_path
405 // } else if (indexes in bounds) {
406 // if ((is object array) && !(array type check)) {
407 // slowval = call checked copy loop
408 // if (slowval == 0) return // via checked_path
409 // } else {
410 // call bulk copy loop
411 // return // via fast_path
412 // }
413 // }
414 // // adjust params for remaining work:
415 // if (slowval != -1) {
416 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
417 // }
418 // slow_region:
419 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
420 // return // via slow_call_path
421 //
422 // This routine is used from several intrinsics: System.arraycopy,
423 // Object.clone (the array subcase), and Arrays.copyOf[Range].
424 //
425 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
426 Node** ctrl, MergeMemNode* mem, Node** io,
427 const TypePtr* adr_type,
428 BasicType basic_elem_type,
429 Node* src, Node* src_offset,
430 Node* dest, Node* dest_offset,
431 Node* copy_length,
432 Node* dest_length,
433 bool disjoint_bases,
434 bool length_never_negative,
435 RegionNode* slow_region) {
436 if (slow_region == nullptr) {
437 slow_region = new RegionNode(1);
438 transform_later(slow_region);
439 }
440
441 Node* original_dest = dest;
442 bool dest_needs_zeroing = false;
443 bool acopy_to_uninitialized = false;
444 Node* init_value = nullptr;
445 Node* raw_init_value = nullptr;
446
447 // See if this is the initialization of a newly-allocated array.
448 // If so, we will take responsibility here for initializing it to zero.
449 // (Note: Because tightly_coupled_allocation performs checks on the
450 // out-edges of the dest, we need to avoid making derived pointers
451 // from it until we have checked its uses.)
452 if (ReduceBulkZeroing
453 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
454 && basic_elem_type != T_CONFLICT // avoid corner case
455 && !src->eqv_uncast(dest)
456 && alloc != nullptr
457 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
458 assert(ac->is_alloc_tightly_coupled(), "sanity");
459 // acopy to uninitialized tightly coupled allocations
460 // needs zeroing outside the copy range
461 // and the acopy itself will be to uninitialized memory
462 acopy_to_uninitialized = true;
463 if (alloc->maybe_set_complete(&_igvn)) {
464 // "You break it, you buy it."
465 InitializeNode* init = alloc->initialization();
466 assert(init->is_complete(), "we just did this");
467 init->set_complete_with_arraycopy();
468 assert(dest->is_CheckCastPP(), "sanity");
469 assert(dest->in(0)->in(0) == init, "dest pinned");
470 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
471 // From this point on, every exit path is responsible for
472 // initializing any non-copied parts of the object to zero.
473 // Also, if this flag is set we make sure that arraycopy interacts properly
474 // with G1, eliding pre-barriers. See CR 6627983.
475 dest_needs_zeroing = true;
476 init_value = alloc->in(AllocateNode::InitValue);
477 raw_init_value = alloc->in(AllocateNode::RawInitValue);
478 } else {
479 // dest_need_zeroing = false;
480 }
481 } else {
482 // No zeroing elimination needed here.
483 alloc = nullptr;
484 acopy_to_uninitialized = false;
485 //original_dest = dest;
486 //dest_needs_zeroing = false;
487 }
488
489 uint alias_idx = C->get_alias_index(adr_type);
490
491 // Results are placed here:
492 enum { fast_path = 1, // normal void-returning assembly stub
493 checked_path = 2, // special assembly stub with cleanup
494 slow_call_path = 3, // something went wrong; call the VM
495 zero_path = 4, // bypass when length of copy is zero
496 bcopy_path = 5, // copy primitive array by 64-bit blocks
497 PATH_LIMIT = 6
498 };
499 RegionNode* result_region = new RegionNode(PATH_LIMIT);
500 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
501 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
502 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
503 transform_later(result_region);
504 transform_later(result_i_o);
505 transform_later(result_memory);
506
507 // The slow_control path:
508 Node* slow_control;
509 Node* slow_i_o = *io;
510 Node* slow_mem = mem->memory_at(alias_idx);
511 DEBUG_ONLY(slow_control = (Node*) badAddress);
512
513 // Checked control path:
514 Node* checked_control = top();
515 Node* checked_mem = nullptr;
516 Node* checked_i_o = nullptr;
517 Node* checked_value = nullptr;
518
519 if (basic_elem_type == T_CONFLICT) {
520 assert(!dest_needs_zeroing, "");
521 Node* cv = generate_generic_arraycopy(ctrl, &mem,
522 adr_type,
523 src, src_offset, dest, dest_offset,
524 copy_length, acopy_to_uninitialized);
525 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
526 checked_control = *ctrl;
527 checked_i_o = *io;
528 checked_mem = mem->memory_at(alias_idx);
529 checked_value = cv;
530 *ctrl = top();
531 }
532
533 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
534 if (not_pos != nullptr) {
535 Node* local_ctrl = not_pos, *local_io = *io;
536 MergeMemNode* local_mem = MergeMemNode::make(mem);
537 transform_later(local_mem);
538
539 // (6) length must not be negative.
540 if (!length_never_negative) {
541 generate_negative_guard(&local_ctrl, copy_length, slow_region);
542 }
543
544 // copy_length is 0.
545 if (dest_needs_zeroing) {
546 assert(!local_ctrl->is_top(), "no ctrl?");
547 if (copy_length->eqv_uncast(dest_length)
548 || _igvn.find_int_con(dest_length, 1) <= 0) {
549 // There is no zeroing to do. No need for a secondary raw memory barrier.
550 } else {
551 // Clear the whole thing since there are no source elements to copy.
552 generate_clear_array(local_ctrl, local_mem,
553 adr_type, dest,
554 init_value, raw_init_value,
555 basic_elem_type,
556 intcon(0), nullptr,
557 alloc->in(AllocateNode::AllocSize));
558 // Use a secondary InitializeNode as raw memory barrier.
559 // Currently it is needed only on this path since other
560 // paths have stub or runtime calls as raw memory barriers.
561 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
562 Compile::AliasIdxRaw,
563 top());
564 transform_later(mb);
565 mb->set_req(TypeFunc::Control,local_ctrl);
566 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
567 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
568 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
569
570 InitializeNode* init = mb->as_Initialize();
571 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
572 }
573 }
574
575 // Present the results of the fast call.
576 result_region->init_req(zero_path, local_ctrl);
577 result_i_o ->init_req(zero_path, local_io);
578 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
579 }
580
581 if (!(*ctrl)->is_top() && dest_needs_zeroing) {
582 // We have to initialize the *uncopied* part of the array to zero.
583 // The copy destination is the slice dest[off..off+len]. The other slices
584 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
585 Node* dest_size = alloc->in(AllocateNode::AllocSize);
586 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
587
588 // If there is a head section that needs zeroing, do it now.
589 if (_igvn.find_int_con(dest_offset, -1) != 0) {
590 generate_clear_array(*ctrl, mem,
591 adr_type, dest,
592 init_value, raw_init_value,
593 basic_elem_type,
594 intcon(0), dest_offset,
595 nullptr);
596 }
597
598 // Next, perform a dynamic check on the tail length.
599 // It is often zero, and we can win big if we prove this.
600 // There are two wins: Avoid generating the ClearArray
601 // with its attendant messy index arithmetic, and upgrade
602 // the copy to a more hardware-friendly word size of 64 bits.
603 Node* tail_ctl = nullptr;
604 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
605 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
606 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
607 tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
608 assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
609 }
610
611 // At this point, let's assume there is no tail.
612 if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
613 // There is no tail. Try an upgrade to a 64-bit copy.
614 bool didit = false;
615 {
616 Node* local_ctrl = *ctrl, *local_io = *io;
617 MergeMemNode* local_mem = MergeMemNode::make(mem);
618 transform_later(local_mem);
619
620 didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
621 adr_type, basic_elem_type, alloc,
622 src, src_offset, dest, dest_offset,
623 dest_size, acopy_to_uninitialized);
624 if (didit) {
625 // Present the results of the block-copying fast call.
626 result_region->init_req(bcopy_path, local_ctrl);
627 result_i_o ->init_req(bcopy_path, local_io);
628 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
629 }
630 }
631 if (didit) {
632 *ctrl = top(); // no regular fast path
633 }
634 }
635
636 // Clear the tail, if any.
637 if (tail_ctl != nullptr) {
638 Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
639 *ctrl = tail_ctl;
640 if (notail_ctl == nullptr) {
641 generate_clear_array(*ctrl, mem,
642 adr_type, dest,
643 init_value, raw_init_value,
644 basic_elem_type,
645 dest_tail, nullptr,
646 dest_size);
647 } else {
648 // Make a local merge.
649 Node* done_ctl = transform_later(new RegionNode(3));
650 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
651 done_ctl->init_req(1, notail_ctl);
652 done_mem->init_req(1, mem->memory_at(alias_idx));
653 generate_clear_array(*ctrl, mem,
654 adr_type, dest,
655 init_value, raw_init_value,
656 basic_elem_type,
657 dest_tail, nullptr,
658 dest_size);
659 done_ctl->init_req(2, *ctrl);
660 done_mem->init_req(2, mem->memory_at(alias_idx));
661 *ctrl = done_ctl;
662 mem->set_memory_at(alias_idx, done_mem);
663 }
664 }
665 }
666
667 BasicType copy_type = basic_elem_type;
668 assert(basic_elem_type != T_ARRAY, "caller must fix this");
669 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
670 // If src and dest have compatible element types, we can copy bits.
671 // Types S[] and D[] are compatible if D is a supertype of S.
672 //
673 // If they are not, we will use checked_oop_disjoint_arraycopy,
674 // which performs a fast optimistic per-oop check, and backs off
675 // further to JVM_ArrayCopy on the first per-oop check that fails.
676 // (Actually, we don't move raw bits only; the GC requires card marks.)
677
678 // We don't need a subtype check for validated copies and Object[].clone()
679 bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
680 ac->is_copyofrange_validated() || ac->is_clone_oop_array();
681 if (!skip_subtype_check) {
682 // Get the klass* for both src and dest
683 Node* src_klass = ac->in(ArrayCopyNode::SrcKlass);
684 Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
685
686 assert(src_klass != nullptr && dest_klass != nullptr, "should have klasses");
687
688 // Generate the subtype check.
689 // This might fold up statically, or then again it might not.
690 //
691 // Non-static example: Copying List<String>.elements to a new String[].
692 // The backing store for a List<String> is always an Object[],
693 // but its elements are always type String, if the generic types
694 // are correct at the source level.
695 //
696 // Test S[] against D[], not S against D, because (probably)
697 // the secondary supertype cache is less busy for S[] than S.
698 // This usually only matters when D is an interface.
699 Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn, nullptr, -1);
700 // Plug failing path into checked_oop_disjoint_arraycopy
701 if (not_subtype_ctrl != top()) {
702 Node* local_ctrl = not_subtype_ctrl;
703 MergeMemNode* local_mem = MergeMemNode::make(mem);
704 transform_later(local_mem);
705
706 // (At this point we can assume disjoint_bases, since types differ.)
707 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
708 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
709 Node* n1 = LoadKlassNode::make(_igvn, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
710 Node* dest_elem_klass = transform_later(n1);
711 Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
712 adr_type,
713 dest_elem_klass,
714 src, src_offset, dest, dest_offset,
715 ConvI2X(copy_length), acopy_to_uninitialized);
716 if (cv == nullptr) cv = intcon(-1); // failure (no stub available)
717 checked_control = local_ctrl;
718 checked_i_o = *io;
719 checked_mem = local_mem->memory_at(alias_idx);
720 checked_value = cv;
721 }
722 }
723 // At this point we know we do not need type checks on oop stores.
724
725 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
726 if (!bs->array_copy_requires_gc_barriers(alloc != nullptr, copy_type, false, false, BarrierSetC2::Expansion)) {
727 // If we do not need gc barriers, copy using the jint or jlong stub.
728 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
729 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
730 "sizes agree");
731 }
732 }
733
734 if (!(*ctrl)->is_top()) {
735 // Generate the fast path, if possible.
736 Node* local_ctrl = *ctrl;
737 MergeMemNode* local_mem = MergeMemNode::make(mem);
738 transform_later(local_mem);
739 generate_unchecked_arraycopy(&local_ctrl, &local_mem,
740 adr_type, copy_type, disjoint_bases,
741 src, src_offset, dest, dest_offset,
742 ConvI2X(copy_length), acopy_to_uninitialized);
743
744 // Present the results of the fast call.
745 result_region->init_req(fast_path, local_ctrl);
746 result_i_o ->init_req(fast_path, *io);
747 result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
748 }
749
750 // Here are all the slow paths up to this point, in one bundle:
751 assert(slow_region != nullptr, "allocated on entry");
752 slow_control = slow_region;
753 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
754
755 *ctrl = checked_control;
756 if (!(*ctrl)->is_top()) {
757 // Clean up after the checked call.
758 // The returned value is either 0 or -1^K,
759 // where K = number of partially transferred array elements.
760 Node* cmp = new CmpINode(checked_value, intcon(0));
761 transform_later(cmp);
762 Node* bol = new BoolNode(cmp, BoolTest::eq);
763 transform_later(bol);
764 IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
765 transform_later(iff);
766
767 // If it is 0, we are done, so transfer to the end.
768 Node* checks_done = new IfTrueNode(iff);
769 transform_later(checks_done);
770 result_region->init_req(checked_path, checks_done);
771 result_i_o ->init_req(checked_path, checked_i_o);
772 result_memory->init_req(checked_path, checked_mem);
773
774 // If it is not zero, merge into the slow call.
775 *ctrl = new IfFalseNode(iff);
776 transform_later(*ctrl);
777 RegionNode* slow_reg2 = new RegionNode(3);
778 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
779 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
780 transform_later(slow_reg2);
781 transform_later(slow_i_o2);
782 transform_later(slow_mem2);
783 slow_reg2 ->init_req(1, slow_control);
784 slow_i_o2 ->init_req(1, slow_i_o);
785 slow_mem2 ->init_req(1, slow_mem);
786 slow_reg2 ->init_req(2, *ctrl);
787 slow_i_o2 ->init_req(2, checked_i_o);
788 slow_mem2 ->init_req(2, checked_mem);
789
790 slow_control = slow_reg2;
791 slow_i_o = slow_i_o2;
792 slow_mem = slow_mem2;
793
794 if (alloc != nullptr) {
795 // We'll restart from the very beginning, after zeroing the whole thing.
796 // This can cause double writes, but that's OK since dest is brand new.
797 // So we ignore the low 31 bits of the value returned from the stub.
798 } else {
799 // We must continue the copy exactly where it failed, or else
800 // another thread might see the wrong number of writes to dest.
801 Node* checked_offset = new XorINode(checked_value, intcon(-1));
802 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
803 transform_later(checked_offset);
804 transform_later(slow_offset);
805 slow_offset->init_req(1, intcon(0));
806 slow_offset->init_req(2, checked_offset);
807
808 // Adjust the arguments by the conditionally incoming offset.
809 Node* src_off_plus = new AddINode(src_offset, slow_offset);
810 transform_later(src_off_plus);
811 Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
812 transform_later(dest_off_plus);
813 Node* length_minus = new SubINode(copy_length, slow_offset);
814 transform_later(length_minus);
815
816 // Tweak the node variables to adjust the code produced below:
817 src_offset = src_off_plus;
818 dest_offset = dest_off_plus;
819 copy_length = length_minus;
820 }
821 }
822 *ctrl = slow_control;
823 if (!(*ctrl)->is_top()) {
824 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
825 MergeMemNode* local_mem = MergeMemNode::make(mem);
826 transform_later(local_mem);
827
828 // Generate the slow path, if needed.
829 local_mem->set_memory_at(alias_idx, slow_mem);
830
831 if (dest_needs_zeroing) {
832 generate_clear_array(local_ctrl, local_mem,
833 adr_type, dest,
834 init_value, raw_init_value,
835 basic_elem_type,
836 intcon(0), nullptr,
837 alloc->in(AllocateNode::AllocSize));
838 }
839
840 local_mem = generate_slow_arraycopy(ac,
841 &local_ctrl, local_mem, &local_io,
842 adr_type,
843 src, src_offset, dest, dest_offset,
844 copy_length, /*dest_uninitialized*/false);
845
846 result_region->init_req(slow_call_path, local_ctrl);
847 result_i_o ->init_req(slow_call_path, local_io);
848 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
849 } else {
850 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
851 // projections were not extracted
852 }
853
854 // Remove unused edges.
855 for (uint i = 1; i < result_region->req(); i++) {
856 if (result_region->in(i) == nullptr) {
857 result_region->init_req(i, top());
858 }
859 }
860
861 // Finished; return the combined state.
862 *ctrl = result_region;
863 *io = result_i_o;
864 mem->set_memory_at(alias_idx, result_memory);
865
866 // mem no longer guaranteed to stay a MergeMemNode
867 Node* out_mem = mem;
868 DEBUG_ONLY(mem = nullptr);
869
870 // The memory edges above are precise in order to model effects around
871 // array copies accurately to allow value numbering of field loads around
872 // arraycopy. Such field loads, both before and after, are common in Java
873 // collections and similar classes involving header/array data structures.
874 //
875 // But with low number of register or when some registers are used or killed
876 // by arraycopy calls it causes registers spilling on stack. See 6544710.
877 // The next memory barrier is added to avoid it. If the arraycopy can be
878 // optimized away (which it can, sometimes) then we can manually remove
879 // the membar also.
880 //
881 // Do not let reads from the cloned object float above the arraycopy.
882 if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
883 // Do not let stores that initialize this object be reordered with
884 // a subsequent store that would make this object accessible by
885 // other threads.
886 assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
887 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
888 } else {
889 int mem_bar_alias_idx = Compile::AliasIdxBot;
890 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
891 // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
892 // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
893 // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
894 // broken graph.
895 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
896 }
897 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
898 }
899
900 assert((*ctrl)->is_Proj(), "MemBar control projection");
901 assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
902 (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
903
904 _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
905 if (_callprojs->fallthrough_ioproj != nullptr) {
906 _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
907 }
908 _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
909
910 #ifdef ASSERT
911 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
912 if (dest_t->is_known_instance()) {
913 ArrayCopyNode* ac = nullptr;
914 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
915 assert(ac == nullptr, "no arraycopy anymore");
916 }
917 #endif
918
919 return out_mem;
920 }
921
922 // Helper for initialization of arrays, creating a ClearArray.
923 // It writes zero bits in [start..end), within the body of an array object.
924 // The memory effects are all chained onto the 'adr_type' alias category.
925 //
926 // Since the object is otherwise uninitialized, we are free
927 // to put a little "slop" around the edges of the cleared area,
928 // as long as it does not go back into the array's header,
929 // or beyond the array end within the heap.
930 //
931 // The lower edge can be rounded down to the nearest jint and the
932 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
933 //
934 // Arguments:
935 // adr_type memory slice where writes are generated
936 // dest oop of the destination array
937 // basic_elem_type element type of the destination
938 // slice_idx array index of first element to store
939 // slice_len number of elements to store (or null)
940 // dest_size total size in bytes of the array object
941 //
942 // Exactly one of slice_len or dest_size must be non-null.
943 // If dest_size is non-null, zeroing extends to the end of the object.
944 // If slice_len is non-null, the slice_idx value must be a constant.
945 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
946 const TypePtr* adr_type,
947 Node* dest,
948 Node* val,
949 Node* raw_val,
950 BasicType basic_elem_type,
951 Node* slice_idx,
952 Node* slice_len,
953 Node* dest_size) {
954 // one or the other but not both of slice_len and dest_size:
955 assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
956 if (slice_len == nullptr) slice_len = top();
957 if (dest_size == nullptr) dest_size = top();
958
959 uint alias_idx = C->get_alias_index(adr_type);
960
961 // operate on this memory slice:
962 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
963
964 // scaling and rounding of indexes:
965 int scale = exact_log2(type2aelembytes(basic_elem_type));
966 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
967 int clear_low = (-1 << scale) & (BytesPerInt - 1);
968 int bump_bit = (-1 << scale) & BytesPerInt;
969
970 // determine constant starts and ends
971 const intptr_t BIG_NEG = -128;
972 assert(BIG_NEG + 2*abase < 0, "neg enough");
973 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
974 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
975 if (slice_len_con == 0) {
976 return; // nothing to do here
977 }
978 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
979 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
980 if (slice_idx_con >= 0 && slice_len_con >= 0) {
981 assert(end_con < 0, "not two cons");
982 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
983 BytesPerLong);
984 }
985
986 if (start_con >= 0 && end_con >= 0) {
987 // Constant start and end. Simple.
988 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
989 start_con, end_con, &_igvn);
990 } else if (start_con >= 0 && dest_size != top()) {
991 // Constant start, pre-rounded end after the tail of the array.
992 Node* end = dest_size;
993 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
994 start_con, end, &_igvn);
995 } else if (start_con >= 0 && slice_len != top()) {
996 // Constant start, non-constant end. End needs rounding up.
997 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
998 intptr_t end_base = abase + (slice_idx_con << scale);
999 int end_round = (-1 << scale) & (BytesPerLong - 1);
1000 Node* end = ConvI2X(slice_len);
1001 if (scale != 0)
1002 end = transform_later(new LShiftXNode(end, intcon(scale) ));
1003 end_base += end_round;
1004 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
1005 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
1006 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1007 start_con, end, &_igvn);
1008 } else if (start_con < 0 && dest_size != top()) {
1009 // Non-constant start, pre-rounded end after the tail of the array.
1010 // This is almost certainly a "round-to-end" operation.
1011 Node* start = slice_idx;
1012 start = ConvI2X(start);
1013 if (scale != 0)
1014 start = transform_later(new LShiftXNode( start, intcon(scale) ));
1015 start = transform_later(new AddXNode(start, MakeConX(abase)) );
1016 if ((bump_bit | clear_low) != 0) {
1017 int to_clear = (bump_bit | clear_low);
1018 // Align up mod 8, then store a jint zero unconditionally
1019 // just before the mod-8 boundary.
1020 if (((abase + bump_bit) & ~to_clear) - bump_bit
1021 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1022 bump_bit = 0;
1023 assert((abase & to_clear) == 0, "array base must be long-aligned");
1024 } else {
1025 // Bump 'start' up to (or past) the next jint boundary:
1026 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1027 assert((abase & clear_low) == 0, "array base must be int-aligned");
1028 }
1029 // Round bumped 'start' down to jlong boundary in body of array.
1030 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1031 if (bump_bit != 0) {
1032 // Store a zero to the immediately preceding jint:
1033 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1034 Node* p1 = basic_plus_adr(dest, x1);
1035 if (val == nullptr) {
1036 assert(raw_val == nullptr, "val may not be null");
1037 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1038 } else {
1039 assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1040 mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1041 }
1042 mem = transform_later(mem);
1043 }
1044 }
1045 Node* end = dest_size; // pre-rounded
1046 mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1047 start, end, &_igvn);
1048 } else {
1049 // Non-constant start, unrounded non-constant end.
1050 // (Nobody zeroes a random midsection of an array using this routine.)
1051 ShouldNotReachHere(); // fix caller
1052 }
1053
1054 // Done.
1055 merge_mem->set_memory_at(alias_idx, mem);
1056 }
1057
1058 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1059 const TypePtr* adr_type,
1060 BasicType basic_elem_type,
1061 AllocateNode* alloc,
1062 Node* src, Node* src_offset,
1063 Node* dest, Node* dest_offset,
1064 Node* dest_size, bool dest_uninitialized) {
1065 // See if there is an advantage from block transfer.
1066 int scale = exact_log2(type2aelembytes(basic_elem_type));
1067 if (scale >= LogBytesPerLong)
1068 return false; // it is already a block transfer
1069
1070 // Look at the alignment of the starting offsets.
1071 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1072
1073 intptr_t src_off_con = (intptr_t) _igvn.find_int_con(src_offset, -1);
1074 intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1075 if (src_off_con < 0 || dest_off_con < 0) {
1076 // At present, we can only understand constants.
1077 return false;
1078 }
1079
1080 intptr_t src_off = abase + (src_off_con << scale);
1081 intptr_t dest_off = abase + (dest_off_con << scale);
1082
1083 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1084 // Non-aligned; too bad.
1085 // One more chance: Pick off an initial 32-bit word.
1086 // This is a common case, since abase can be odd mod 8.
1087 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1088 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1089 Node* sptr = basic_plus_adr(src, src_off);
1090 Node* dptr = basic_plus_adr(dest, dest_off);
1091 const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1092 assert(s_adr_type->isa_aryptr(), "impossible slice");
1093 uint s_alias_idx = C->get_alias_index(s_adr_type);
1094 uint d_alias_idx = C->get_alias_index(adr_type);
1095 bool is_mismatched = (basic_elem_type != T_INT);
1096 Node* sval = transform_later(
1097 LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1098 TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1099 false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1100 Node* st = transform_later(
1101 StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1102 sval, T_INT, MemNode::unordered));
1103 if (is_mismatched) {
1104 st->as_Store()->set_mismatched_access();
1105 }
1106 (*mem)->set_memory_at(d_alias_idx, st);
1107 src_off += BytesPerInt;
1108 dest_off += BytesPerInt;
1109 } else {
1110 return false;
1111 }
1112 }
1113 assert(src_off % BytesPerLong == 0, "");
1114 assert(dest_off % BytesPerLong == 0, "");
1115
1116 // Do this copy by giant steps.
1117 Node* sptr = basic_plus_adr(src, src_off);
1118 Node* dptr = basic_plus_adr(dest, dest_off);
1119 Node* countx = dest_size;
1120 countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1121 countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1122
1123 bool disjoint_bases = true; // since alloc isn't null
1124 generate_unchecked_arraycopy(ctrl, mem,
1125 adr_type, T_LONG, disjoint_bases,
1126 sptr, nullptr, dptr, nullptr, countx, dest_uninitialized);
1127
1128 return true;
1129 }
1130
1131 // Helper function; generates code for the slow case.
1132 // We make a call to a runtime method which emulates the native method,
1133 // but without the native wrapper overhead.
1134 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1135 Node** ctrl, Node* mem, Node** io,
1136 const TypePtr* adr_type,
1137 Node* src, Node* src_offset,
1138 Node* dest, Node* dest_offset,
1139 Node* copy_length, bool dest_uninitialized) {
1140 assert(!dest_uninitialized, "Invariant");
1141
1142 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1143 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1144 "slow_arraycopy", TypePtr::BOTTOM);
1145
1146 call->init_req(TypeFunc::Control, *ctrl);
1147 call->init_req(TypeFunc::I_O , *io);
1148 call->init_req(TypeFunc::Memory , mem);
1149 call->init_req(TypeFunc::ReturnAdr, top());
1150 call->init_req(TypeFunc::FramePtr, top());
1151 call->init_req(TypeFunc::Parms+0, src);
1152 call->init_req(TypeFunc::Parms+1, src_offset);
1153 call->init_req(TypeFunc::Parms+2, dest);
1154 call->init_req(TypeFunc::Parms+3, dest_offset);
1155 call->init_req(TypeFunc::Parms+4, copy_length);
1156 call->copy_call_debug_info(&_igvn, ac);
1157
1158 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1159 _igvn.replace_node(ac, call);
1160 transform_later(call);
1161
1162 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1163 *ctrl = _callprojs->fallthrough_catchproj->clone();
1164 transform_later(*ctrl);
1165
1166 Node* m = _callprojs->fallthrough_memproj->clone();
1167 transform_later(m);
1168
1169 uint alias_idx = C->get_alias_index(adr_type);
1170 MergeMemNode* out_mem;
1171 if (alias_idx != Compile::AliasIdxBot) {
1172 out_mem = MergeMemNode::make(mem);
1173 out_mem->set_memory_at(alias_idx, m);
1174 } else {
1175 out_mem = MergeMemNode::make(m);
1176 }
1177 transform_later(out_mem);
1178
1179 // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1180 // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1181 if (_callprojs->fallthrough_ioproj != nullptr) {
1182 *io = _callprojs->fallthrough_ioproj->clone();
1183 transform_later(*io);
1184 } else {
1185 *io = nullptr;
1186 }
1187
1188 return out_mem;
1189 }
1190
1191 // Helper function; generates code for cases requiring runtime checks.
1192 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1193 const TypePtr* adr_type,
1194 Node* dest_elem_klass,
1195 Node* src, Node* src_offset,
1196 Node* dest, Node* dest_offset,
1197 Node* copy_length, bool dest_uninitialized) {
1198 if ((*ctrl)->is_top()) return nullptr;
1199
1200 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1201 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1202 return nullptr;
1203 }
1204
1205 // Pick out the parameters required to perform a store-check
1206 // for the target array. This is an optimistic check. It will
1207 // look in each non-null element's class, at the desired klass's
1208 // super_check_offset, for the desired klass.
1209 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1210 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1211 Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1212 Node* check_offset = ConvI2X(transform_later(n3));
1213 Node* check_value = dest_elem_klass;
1214
1215 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1216 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1217
1218 const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1219 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1220 src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1221
1222 finish_arraycopy_call(call, ctrl, mem, adr_type);
1223
1224 Node* proj = new ProjNode(call, TypeFunc::Parms);
1225 transform_later(proj);
1226
1227 return proj;
1228 }
1229
1230 // Helper function; generates code for cases requiring runtime checks.
1231 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1232 const TypePtr* adr_type,
1233 Node* src, Node* src_offset,
1234 Node* dest, Node* dest_offset,
1235 Node* copy_length, bool dest_uninitialized) {
1236 if ((*ctrl)->is_top()) return nullptr;
1237 assert(!dest_uninitialized, "Invariant");
1238
1239 address copyfunc_addr = StubRoutines::generic_arraycopy();
1240 if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1241 return nullptr;
1242 }
1243
1244 const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1245 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1246 src, src_offset, dest, dest_offset, copy_length);
1247
1248 finish_arraycopy_call(call, ctrl, mem, adr_type);
1249
1250 Node* proj = new ProjNode(call, TypeFunc::Parms);
1251 transform_later(proj);
1252
1253 return proj;
1254 }
1255
1256 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1257 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1258 const TypePtr* adr_type,
1259 BasicType basic_elem_type,
1260 bool disjoint_bases,
1261 Node* src, Node* src_offset,
1262 Node* dest, Node* dest_offset,
1263 Node* copy_length, bool dest_uninitialized) {
1264 if ((*ctrl)->is_top()) {
1265 return;
1266 }
1267
1268 Node* src_start = src;
1269 Node* dest_start = dest;
1270 if (src_offset != nullptr || dest_offset != nullptr) {
1271 src_start = array_element_address(src, src_offset, basic_elem_type);
1272 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1273 }
1274
1275 // Figure out which arraycopy runtime method to call.
1276 const char* copyfunc_name = "arraycopy";
1277 address copyfunc_addr =
1278 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1279 disjoint_bases, copyfunc_name, dest_uninitialized);
1280
1281 Node* result_memory = nullptr;
1282 RegionNode* exit_block = nullptr;
1283 if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1284 Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1285 generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1286 copy_length, src_start, dest_start, basic_elem_type);
1287 }
1288
1289 const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1290 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1291 src_start, dest_start, copy_length XTOP);
1292
1293 finish_arraycopy_call(call, ctrl, mem, adr_type);
1294
1295 // Connecting remaining edges for exit_block coming from stub_block.
1296 if (exit_block) {
1297 exit_block->init_req(2, *ctrl);
1298
1299 // Memory edge corresponding to stub_region.
1300 result_memory->init_req(2, *mem);
1301
1302 uint alias_idx = C->get_alias_index(adr_type);
1303 if (alias_idx != Compile::AliasIdxBot) {
1304 *mem = MergeMemNode::make(*mem);
1305 (*mem)->set_memory_at(alias_idx, result_memory);
1306 } else {
1307 *mem = MergeMemNode::make(result_memory);
1308 }
1309 transform_later(*mem);
1310 *ctrl = exit_block;
1311 }
1312 }
1313
1314 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1315 Node*& dest_offset, Node*& length, BasicType& dest_elem,
1316 Node*& dest_length) {
1317 #ifdef ASSERT
1318 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1319 bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1320 bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1321 assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1322 #endif
1323 int elem_size = top_dest->flat_elem_size();
1324 if (elem_size >= 8) {
1325 if (elem_size > 8) {
1326 // treat as array of long but scale length, src offset and dest offset
1327 assert((elem_size % 8) == 0, "not a power of 2?");
1328 int factor = elem_size / 8;
1329 length = transform_later(new MulINode(length, intcon(factor)));
1330 src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1331 dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1332 if (dest_length != nullptr) {
1333 dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1334 }
1335 elem_size = 8;
1336 }
1337 dest_elem = T_LONG;
1338 } else if (elem_size == 4) {
1339 dest_elem = T_INT;
1340 } else if (elem_size == 2) {
1341 dest_elem = T_CHAR;
1342 } else if (elem_size == 1) {
1343 dest_elem = T_BYTE;
1344 } else {
1345 ShouldNotReachHere();
1346 }
1347 return TypeRawPtr::BOTTOM;
1348 }
1349
1350 #undef XTOP
1351
1352 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1353 Node* ctrl = ac->in(TypeFunc::Control);
1354 Node* io = ac->in(TypeFunc::I_O);
1355 Node* src = ac->in(ArrayCopyNode::Src);
1356 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1357 Node* dest = ac->in(ArrayCopyNode::Dest);
1358 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1359 Node* length = ac->in(ArrayCopyNode::Length);
1360 MergeMemNode* merge_mem = nullptr;
1361
1362 if (ac->is_clonebasic()) {
1363 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1364 bs->clone_at_expansion(this, ac);
1365 return;
1366 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1367 const Type* src_type = _igvn.type(src);
1368 const Type* dest_type = _igvn.type(dest);
1369 const TypeAryPtr* top_src = src_type->isa_aryptr();
1370 // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1371 // Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1372 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1373 BasicType dest_elem = T_OBJECT;
1374 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1375 dest_elem = top_dest->elem()->array_element_basic_type();
1376 }
1377 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1378
1379 if (top_src != nullptr && top_src->is_flat()) {
1380 // If src is flat, dest is guaranteed to be flat as well
1381 top_dest = top_src;
1382 }
1383
1384 AllocateArrayNode* alloc = nullptr;
1385 Node* dest_length = nullptr;
1386 if (ac->is_alloc_tightly_coupled()) {
1387 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1388 assert(alloc != nullptr, "expect alloc");
1389 dest_length = alloc->in(AllocateNode::ALength);
1390 }
1391
1392 Node* mem = ac->in(TypeFunc::Memory);
1393 const TypePtr* adr_type = nullptr;
1394 if (top_dest != nullptr && top_dest->is_flat()) {
1395 assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1396 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1397 // on all slices to prevent writes into the source from floating below the arraycopy.
1398 int mem_bar_alias_idx = Compile::AliasIdxBot;
1399 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1400 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1401 }
1402 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1403 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1404 } else {
1405 adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1406 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1407 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1408 }
1409 if (ac->_src_type != ac->_dest_type) {
1410 adr_type = TypeRawPtr::BOTTOM;
1411 }
1412 }
1413 merge_mem = MergeMemNode::make(mem);
1414 transform_later(merge_mem);
1415
1416 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1417 adr_type, dest_elem,
1418 src, src_offset, dest, dest_offset, length,
1419 dest_length,
1420 true, ac->has_negative_length_guard());
1421
1422 return;
1423 }
1424
1425 AllocateArrayNode* alloc = nullptr;
1426 if (ac->is_alloc_tightly_coupled()) {
1427 alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1428 assert(alloc != nullptr, "expect alloc");
1429 }
1430
1431 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1432
1433 // Compile time checks. If any of these checks cannot be verified at compile time,
1434 // we do not make a fast path for this call. Instead, we let the call remain as it
1435 // is. The checks we choose to mandate at compile time are:
1436 //
1437 // (1) src and dest are arrays.
1438 const Type* src_type = src->Value(&_igvn);
1439 const Type* dest_type = dest->Value(&_igvn);
1440 const TypeAryPtr* top_src = src_type->isa_aryptr();
1441 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1442
1443 BasicType src_elem = T_CONFLICT;
1444 BasicType dest_elem = T_CONFLICT;
1445
1446 if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1447 src_elem = top_src->elem()->array_element_basic_type();
1448 }
1449 if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1450 dest_elem = top_dest->elem()->array_element_basic_type();
1451 }
1452 if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1453 if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1454
1455 if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1456 src_elem = dest_elem;
1457 }
1458
1459 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1460 // Conservatively insert a memory barrier on all memory slices.
1461 // Do not let writes into the source float below the arraycopy.
1462 {
1463 Node* mem = ac->in(TypeFunc::Memory);
1464 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1465
1466 merge_mem = MergeMemNode::make(mem);
1467 transform_later(merge_mem);
1468 }
1469
1470 // Call StubRoutines::generic_arraycopy stub.
1471 generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1472 TypeRawPtr::BOTTOM, T_CONFLICT,
1473 src, src_offset, dest, dest_offset, length,
1474 nullptr,
1475 // If a negative length guard was generated for the ArrayCopyNode,
1476 // the length of the array can never be negative.
1477 false, ac->has_negative_length_guard());
1478 return;
1479 }
1480
1481 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1482
1483 // (2) src and dest arrays must have elements of the same BasicType
1484 // Figure out the size and type of the elements we will be copying.
1485 //
1486 // We have no stub to copy flat inline type arrays with oop
1487 // fields if we need to emit write barriers.
1488 //
1489 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1490 if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1491 (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1492 bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1493 // The component types are not the same or are not recognized. Punt.
1494 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1495 {
1496 Node* mem = ac->in(TypeFunc::Memory);
1497 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1498 }
1499
1500 _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1501 if (_callprojs->fallthrough_ioproj != nullptr) {
1502 _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1503 }
1504 _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1505 return;
1506 }
1507
1508 //---------------------------------------------------------------------------
1509 // We will make a fast path for this call to arraycopy.
1510
1511 // We have the following tests left to perform:
1512 //
1513 // (3) src and dest must not be null.
1514 // (4) src_offset must not be negative.
1515 // (5) dest_offset must not be negative.
1516 // (6) length must not be negative.
1517 // (7) src_offset + length must not exceed length of src.
1518 // (8) dest_offset + length must not exceed length of dest.
1519 // (9) each element of an oop array must be assignable
1520
1521 Node* mem = ac->in(TypeFunc::Memory);
1522 if (top_dest->is_flat()) {
1523 // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1524 // on all slices to prevent writes into the source from floating below the arraycopy.
1525 int mem_bar_alias_idx = Compile::AliasIdxBot;
1526 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1527 mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1528 }
1529 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1530 }
1531 merge_mem = MergeMemNode::make(mem);
1532 transform_later(merge_mem);
1533
1534 RegionNode* slow_region = new RegionNode(1);
1535 transform_later(slow_region);
1536
1537 if (!ac->is_arraycopy_validated()) {
1538 // (3) operands must not be null
1539 // We currently perform our null checks with the null_check routine.
1540 // This means that the null exceptions will be reported in the caller
1541 // rather than (correctly) reported inside of the native arraycopy call.
1542 // This should be corrected, given time. We do our null check with the
1543 // stack pointer restored.
1544 // null checks done library_call.cpp
1545
1546 // (4) src_offset must not be negative.
1547 generate_negative_guard(&ctrl, src_offset, slow_region);
1548
1549 // (5) dest_offset must not be negative.
1550 generate_negative_guard(&ctrl, dest_offset, slow_region);
1551
1552 // (6) length must not be negative (moved to generate_arraycopy()).
1553 // generate_negative_guard(length, slow_region);
1554
1555 // (7) src_offset + length must not exceed length of src.
1556 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1557 assert(alen != nullptr, "need src len");
1558 generate_limit_guard(&ctrl,
1559 src_offset, length,
1560 alen,
1561 slow_region);
1562
1563 // (8) dest_offset + length must not exceed length of dest.
1564 alen = ac->in(ArrayCopyNode::DestLen);
1565 assert(alen != nullptr, "need dest len");
1566 generate_limit_guard(&ctrl,
1567 dest_offset, length,
1568 alen,
1569 slow_region);
1570
1571 // (9) each element of an oop array must be assignable
1572 // The generate_arraycopy subroutine checks this.
1573
1574 // TODO 8350865 This is too strong
1575 // We need to be careful here because 'adjust_for_flat_array' will adjust offsets/length etc. which then does not work anymore for the slow call to SharedRuntime::slow_arraycopy_C.
1576 if (!(top_src->is_flat() && top_dest->is_flat() && top_src->is_null_free() == top_dest->is_null_free())) {
1577 generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1578 generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1579 generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1580 }
1581 }
1582
1583 // This is where the memory effects are placed:
1584 const TypePtr* adr_type = nullptr;
1585 Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1586
1587 if (top_src->is_flat() && top_dest->is_flat() &&
1588 top_src->is_null_free() == top_dest->is_null_free()) {
1589 adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1590 } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1591 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1592 } else {
1593 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1594 }
1595
1596 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1597 adr_type, dest_elem,
1598 src, src_offset, dest, dest_offset, length,
1599 dest_length,
1600 // If a negative length guard was generated for the ArrayCopyNode,
1601 // the length of the array can never be negative.
1602 false, ac->has_negative_length_guard(),
1603 slow_region);
1604 }