1 /*
2 * Copyright (c) 2012, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "opto/arraycopynode.hpp"
28 #include "oops/objArrayKlass.hpp"
29 #include "opto/convertnode.hpp"
30 #include "opto/graphKit.hpp"
31 #include "opto/macro.hpp"
32 #include "opto/runtime.hpp"
33 #include "utilities/align.hpp"
34
35
36 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
37 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
38 mb->init_req(TypeFunc::Control, *ctrl);
39 mb->init_req(TypeFunc::Memory, *mem);
40 transform_later(mb);
41 *ctrl = new ProjNode(mb,TypeFunc::Control);
42 transform_later(*ctrl);
43 Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
44 transform_later(mem_proj);
45 *mem = mem_proj;
46 }
47
48 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
49 uint shift = exact_log2(type2aelembytes(elembt));
50 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
51 Node* base = basic_plus_adr(ary, header);
52 #ifdef _LP64
53 // see comment in GraphKit::array_element_address
54 int index_max = max_jint - 1; // array size is max_jint, index is one less
55 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
56 idx = transform_later( new ConvI2LNode(idx, lidxtype) );
57 #endif
58 Node* scale = new LShiftXNode(idx, intcon(shift));
59 transform_later(scale);
60 return basic_plus_adr(ary, base, scale);
61 }
62
63 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
64 return transform_later(new ConvI2LNode(offset));
65 }
66
67 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
68 const TypeFunc* call_type, address call_addr,
69 const char* call_name,
70 const TypePtr* adr_type,
71 Node* parm0, Node* parm1,
72 Node* parm2, Node* parm3,
73 Node* parm4, Node* parm5,
74 Node* parm6, Node* parm7) {
75 int size = call_type->domain()->cnt();
76 Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
77 call->init_req(TypeFunc::Control, ctrl);
78 call->init_req(TypeFunc::I_O , top());
79 call->init_req(TypeFunc::Memory , mem);
80 call->init_req(TypeFunc::ReturnAdr, top());
81 call->init_req(TypeFunc::FramePtr, top());
82
83 // Hook each parm in order. Stop looking at the first NULL.
84 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
85 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
86 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
87 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
88 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
89 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
90 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
91 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
92 /* close each nested if ===> */ } } } } } } } }
93 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
94
95 return call;
96 }
97
98
99 //------------------------------generate_guard---------------------------
100 // Helper function for generating guarded fast-slow graph structures.
101 // The given 'test', if true, guards a slow path. If the test fails
102 // then a fast path can be taken. (We generally hope it fails.)
103 // In all cases, GraphKit::control() is updated to the fast path.
104 // The returned value represents the control for the slow path.
105 // The return value is never 'top'; it is either a valid control
106 // or NULL if it is obvious that the slow path can never be taken.
107 // Also, if region and the slow control are not NULL, the slow edge
108 // is appended to the region.
109 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
110 if ((*ctrl)->is_top()) {
111 // Already short circuited.
112 return NULL;
113 }
114 // Build an if node and its projections.
115 // If test is true we take the slow path, which we assume is uncommon.
116 if (_igvn.type(test) == TypeInt::ZERO) {
117 // The slow branch is never taken. No need to build this guard.
118 return NULL;
119 }
120
121 IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
122 transform_later(iff);
123
124 Node* if_slow = new IfTrueNode(iff);
125 transform_later(if_slow);
126
127 if (region != NULL) {
128 region->add_req(if_slow);
129 }
130
131 Node* if_fast = new IfFalseNode(iff);
132 transform_later(if_fast);
133
134 *ctrl = if_fast;
135
136 return if_slow;
137 }
138
139 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
140 return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
141 }
142
143 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
144 if ((*ctrl)->is_top())
145 return; // already stopped
146 if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
147 return; // index is already adequately typed
148 Node* cmp_lt = new CmpINode(index, intcon(0));
149 transform_later(cmp_lt);
150 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
151 transform_later(bol_lt);
152 generate_guard(ctrl, bol_lt, region, PROB_MIN);
153 }
154
155 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
156 if ((*ctrl)->is_top())
157 return; // already stopped
158 bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
159 if (zero_offset && subseq_length->eqv_uncast(array_length))
160 return; // common case of whole-array copy
161 Node* last = subseq_length;
162 if (!zero_offset) { // last += offset
163 last = new AddINode(last, offset);
164 transform_later(last);
165 }
166 Node* cmp_lt = new CmpUNode(array_length, last);
167 transform_later(cmp_lt);
168 Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
169 transform_later(bol_lt);
170 generate_guard(ctrl, bol_lt, region, PROB_MIN);
171 }
172
173 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
174 if ((*ctrl)->is_top()) return NULL;
175
176 if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
177 return NULL; // index is already adequately typed
178 Node* cmp_le = new CmpINode(index, intcon(0));
179 transform_later(cmp_le);
180 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
181 Node* bol_le = new BoolNode(cmp_le, le_or_eq);
182 transform_later(bol_le);
183 Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
184
185 return is_notp;
186 }
187
188 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
189 transform_later(call);
190
191 *ctrl = new ProjNode(call,TypeFunc::Control);
192 transform_later(*ctrl);
193 Node* newmem = new ProjNode(call, TypeFunc::Memory);
194 transform_later(newmem);
195
196 uint alias_idx = C->get_alias_index(adr_type);
197 if (alias_idx != Compile::AliasIdxBot) {
198 *mem = MergeMemNode::make(*mem);
199 (*mem)->set_memory_at(alias_idx, newmem);
200 } else {
201 *mem = MergeMemNode::make(newmem);
202 }
203 transform_later(*mem);
204 }
205
206 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
207 Node* src_offset,
208 Node* dest_offset,
209 bool disjoint_bases,
210 const char* &name,
211 bool dest_uninitialized) {
212 const TypeInt* src_offset_inttype = _igvn.find_int_type(src_offset);;
213 const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);;
214
215 bool aligned = false;
216 bool disjoint = disjoint_bases;
217
218 // if the offsets are the same, we can treat the memory regions as
219 // disjoint, because either the memory regions are in different arrays,
220 // or they are identical (which we can treat as disjoint.) We can also
221 // treat a copy with a destination index less that the source index
222 // as disjoint since a low->high copy will work correctly in this case.
223 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
224 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
225 // both indices are constants
226 int s_offs = src_offset_inttype->get_con();
227 int d_offs = dest_offset_inttype->get_con();
228 int element_size = type2aelembytes(t);
229 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
230 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
231 if (s_offs >= d_offs) disjoint = true;
232 } else if (src_offset == dest_offset && src_offset != NULL) {
233 // This can occur if the offsets are identical non-constants.
234 disjoint = true;
235 }
236
237 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
238 }
239
240 #define XTOP LP64_ONLY(COMMA top())
241
242 // Generate an optimized call to arraycopy.
243 // Caller must guard against non-arrays.
244 // Caller must determine a common array basic-type for both arrays.
245 // Caller must validate offsets against array bounds.
246 // The slow_region has already collected guard failure paths
247 // (such as out of bounds length or non-conformable array types).
248 // The generated code has this shape, in general:
249 //
250 // if (length == 0) return // via zero_path
251 // slowval = -1
252 // if (types unknown) {
253 // slowval = call generic copy loop
254 // if (slowval == 0) return // via checked_path
255 // } else if (indexes in bounds) {
256 // if ((is object array) && !(array type check)) {
257 // slowval = call checked copy loop
258 // if (slowval == 0) return // via checked_path
259 // } else {
260 // call bulk copy loop
261 // return // via fast_path
262 // }
263 // }
264 // // adjust params for remaining work:
265 // if (slowval != -1) {
266 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
267 // }
268 // slow_region:
269 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
270 // return // via slow_call_path
271 //
272 // This routine is used from several intrinsics: System.arraycopy,
273 // Object.clone (the array subcase), and Arrays.copyOf[Range].
274 //
275 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
276 Node** ctrl, MergeMemNode* mem, Node** io,
277 const TypePtr* adr_type,
278 BasicType basic_elem_type,
279 Node* src, Node* src_offset,
280 Node* dest, Node* dest_offset,
281 Node* copy_length,
282 bool disjoint_bases,
283 bool length_never_negative,
284 RegionNode* slow_region) {
285 if (slow_region == NULL) {
286 slow_region = new RegionNode(1);
287 transform_later(slow_region);
288 }
289
290 Node* original_dest = dest;
291 bool dest_uninitialized = false;
292
293 // See if this is the initialization of a newly-allocated array.
294 // If so, we will take responsibility here for initializing it to zero.
295 // (Note: Because tightly_coupled_allocation performs checks on the
296 // out-edges of the dest, we need to avoid making derived pointers
297 // from it until we have checked its uses.)
298 if (ReduceBulkZeroing
299 && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
300 && basic_elem_type != T_CONFLICT // avoid corner case
301 && !src->eqv_uncast(dest)
302 && alloc != NULL
303 && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
304 && alloc->maybe_set_complete(&_igvn)) {
305 // "You break it, you buy it."
306 InitializeNode* init = alloc->initialization();
307 assert(init->is_complete(), "we just did this");
308 init->set_complete_with_arraycopy();
309 assert(dest->is_CheckCastPP(), "sanity");
310 assert(dest->in(0)->in(0) == init, "dest pinned");
311 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
312 // From this point on, every exit path is responsible for
313 // initializing any non-copied parts of the object to zero.
314 // Also, if this flag is set we make sure that arraycopy interacts properly
315 // with G1, eliding pre-barriers. See CR 6627983.
316 dest_uninitialized = true;
317 } else {
318 // No zeroing elimination here.
319 alloc = NULL;
320 //original_dest = dest;
321 //dest_uninitialized = false;
322 }
323
324 uint alias_idx = C->get_alias_index(adr_type);
325
326 // Results are placed here:
327 enum { fast_path = 1, // normal void-returning assembly stub
328 checked_path = 2, // special assembly stub with cleanup
329 slow_call_path = 3, // something went wrong; call the VM
330 zero_path = 4, // bypass when length of copy is zero
331 bcopy_path = 5, // copy primitive array by 64-bit blocks
332 PATH_LIMIT = 6
333 };
334 RegionNode* result_region = new RegionNode(PATH_LIMIT);
335 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
336 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
337 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
338 transform_later(result_region);
339 transform_later(result_i_o);
340 transform_later(result_memory);
341
342 // The slow_control path:
343 Node* slow_control;
344 Node* slow_i_o = *io;
345 Node* slow_mem = mem->memory_at(alias_idx);
346 DEBUG_ONLY(slow_control = (Node*) badAddress);
347
348 // Checked control path:
349 Node* checked_control = top();
350 Node* checked_mem = NULL;
351 Node* checked_i_o = NULL;
352 Node* checked_value = NULL;
353
354 if (basic_elem_type == T_CONFLICT) {
355 assert(!dest_uninitialized, "");
356 Node* cv = generate_generic_arraycopy(ctrl, &mem,
357 adr_type,
358 src, src_offset, dest, dest_offset,
359 copy_length, dest_uninitialized);
360 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
361 checked_control = *ctrl;
362 checked_i_o = *io;
363 checked_mem = mem->memory_at(alias_idx);
364 checked_value = cv;
365 *ctrl = top();
366 }
367
368 Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
369 if (not_pos != NULL) {
370 Node* local_ctrl = not_pos, *local_io = *io;
371 MergeMemNode* local_mem = MergeMemNode::make(mem);
372 transform_later(local_mem);
373
374 // (6) length must not be negative.
375 if (!length_never_negative) {
376 generate_negative_guard(&local_ctrl, copy_length, slow_region);
377 }
378
379 // copy_length is 0.
380 if (dest_uninitialized) {
381 assert(!local_ctrl->is_top(), "no ctrl?");
382 Node* dest_length = alloc->in(AllocateNode::ALength);
383 if (copy_length->eqv_uncast(dest_length)
384 || _igvn.find_int_con(dest_length, 1) <= 0) {
385 // There is no zeroing to do. No need for a secondary raw memory barrier.
386 } else {
387 // Clear the whole thing since there are no source elements to copy.
388 generate_clear_array(local_ctrl, local_mem,
389 adr_type, dest, basic_elem_type,
390 intcon(0), NULL,
391 alloc->in(AllocateNode::AllocSize));
392 // Use a secondary InitializeNode as raw memory barrier.
393 // Currently it is needed only on this path since other
394 // paths have stub or runtime calls as raw memory barriers.
395 MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
396 Compile::AliasIdxRaw,
397 top());
398 transform_later(mb);
399 mb->set_req(TypeFunc::Control,local_ctrl);
400 mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
401 local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
402 local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
403
404 InitializeNode* init = mb->as_Initialize();
405 init->set_complete(&_igvn); // (there is no corresponding AllocateNode)
406 }
407 }
408
409 // Present the results of the fast call.
410 result_region->init_req(zero_path, local_ctrl);
411 result_i_o ->init_req(zero_path, local_io);
412 result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
413 }
414
415 if (!(*ctrl)->is_top() && dest_uninitialized) {
416 // We have to initialize the *uncopied* part of the array to zero.
417 // The copy destination is the slice dest[off..off+len]. The other slices
418 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
419 Node* dest_size = alloc->in(AllocateNode::AllocSize);
420 Node* dest_length = alloc->in(AllocateNode::ALength);
421 Node* dest_tail = transform_later( new AddINode(dest_offset, copy_length));
422
423 // If there is a head section that needs zeroing, do it now.
424 if (_igvn.find_int_con(dest_offset, -1) != 0) {
425 generate_clear_array(*ctrl, mem,
426 adr_type, dest, basic_elem_type,
427 intcon(0), dest_offset,
428 NULL);
429 }
430
431 // Next, perform a dynamic check on the tail length.
432 // It is often zero, and we can win big if we prove this.
433 // There are two wins: Avoid generating the ClearArray
434 // with its attendant messy index arithmetic, and upgrade
435 // the copy to a more hardware-friendly word size of 64 bits.
436 Node* tail_ctl = NULL;
437 if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
438 Node* cmp_lt = transform_later( new CmpINode(dest_tail, dest_length) );
439 Node* bol_lt = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
440 tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
441 assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
442 }
443
444 // At this point, let's assume there is no tail.
445 if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
446 // There is no tail. Try an upgrade to a 64-bit copy.
447 bool didit = false;
448 {
449 Node* local_ctrl = *ctrl, *local_io = *io;
450 MergeMemNode* local_mem = MergeMemNode::make(mem);
451 transform_later(local_mem);
452
453 didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
454 adr_type, basic_elem_type, alloc,
455 src, src_offset, dest, dest_offset,
456 dest_size, dest_uninitialized);
457 if (didit) {
458 // Present the results of the block-copying fast call.
459 result_region->init_req(bcopy_path, local_ctrl);
460 result_i_o ->init_req(bcopy_path, local_io);
461 result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
462 }
463 }
464 if (didit) {
465 *ctrl = top(); // no regular fast path
466 }
467 }
468
469 // Clear the tail, if any.
470 if (tail_ctl != NULL) {
471 Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
472 *ctrl = tail_ctl;
473 if (notail_ctl == NULL) {
474 generate_clear_array(*ctrl, mem,
475 adr_type, dest, basic_elem_type,
476 dest_tail, NULL,
477 dest_size);
478 } else {
479 // Make a local merge.
480 Node* done_ctl = transform_later(new RegionNode(3));
481 Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
482 done_ctl->init_req(1, notail_ctl);
483 done_mem->init_req(1, mem->memory_at(alias_idx));
484 generate_clear_array(*ctrl, mem,
485 adr_type, dest, basic_elem_type,
486 dest_tail, NULL,
487 dest_size);
488 done_ctl->init_req(2, *ctrl);
489 done_mem->init_req(2, mem->memory_at(alias_idx));
490 *ctrl = done_ctl;
491 mem->set_memory_at(alias_idx, done_mem);
492 }
493 }
494 }
495
496 BasicType copy_type = basic_elem_type;
497 assert(basic_elem_type != T_ARRAY, "caller must fix this");
498 if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
499 // If src and dest have compatible element types, we can copy bits.
500 // Types S[] and D[] are compatible if D is a supertype of S.
501 //
502 // If they are not, we will use checked_oop_disjoint_arraycopy,
503 // which performs a fast optimistic per-oop check, and backs off
504 // further to JVM_ArrayCopy on the first per-oop check that fails.
505 // (Actually, we don't move raw bits only; the GC requires card marks.)
506
507 // We don't need a subtype check for validated copies and Object[].clone()
508 bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
509 ac->is_copyofrange_validated() || ac->is_cloneoop();
510 if (!skip_subtype_check) {
511 // Get the klass* for both src and dest
512 Node* src_klass = ac->in(ArrayCopyNode::SrcKlass);
513 Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
514
515 assert(src_klass != NULL && dest_klass != NULL, "should have klasses");
516
517 // Generate the subtype check.
518 // This might fold up statically, or then again it might not.
519 //
520 // Non-static example: Copying List<String>.elements to a new String[].
521 // The backing store for a List<String> is always an Object[],
522 // but its elements are always type String, if the generic types
523 // are correct at the source level.
524 //
525 // Test S[] against D[], not S against D, because (probably)
526 // the secondary supertype cache is less busy for S[] than S.
527 // This usually only matters when D is an interface.
528 Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, &_igvn);
529 // Plug failing path into checked_oop_disjoint_arraycopy
530 if (not_subtype_ctrl != top()) {
531 Node* local_ctrl = not_subtype_ctrl;
532 MergeMemNode* local_mem = MergeMemNode::make(mem);
533 transform_later(local_mem);
534
535 // (At this point we can assume disjoint_bases, since types differ.)
536 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
537 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
538 Node* n1 = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
539 Node* dest_elem_klass = transform_later(n1);
540 Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
541 adr_type,
542 dest_elem_klass,
543 src, src_offset, dest, dest_offset,
544 ConvI2X(copy_length), dest_uninitialized);
545 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
546 checked_control = local_ctrl;
547 checked_i_o = *io;
548 checked_mem = local_mem->memory_at(alias_idx);
549 checked_value = cv;
550 }
551 }
552 // At this point we know we do not need type checks on oop stores.
553
554 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
555 if (!bs->array_copy_requires_gc_barriers(alloc != NULL, copy_type, false, BarrierSetC2::Expansion)) {
556 // If we do not need gc barriers, copy using the jint or jlong stub.
557 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
558 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
559 "sizes agree");
560 }
561 }
562
563 if (!(*ctrl)->is_top()) {
564 // Generate the fast path, if possible.
565 Node* local_ctrl = *ctrl;
566 MergeMemNode* local_mem = MergeMemNode::make(mem);
567 transform_later(local_mem);
568
569 generate_unchecked_arraycopy(&local_ctrl, &local_mem,
570 adr_type, copy_type, disjoint_bases,
571 src, src_offset, dest, dest_offset,
572 ConvI2X(copy_length), dest_uninitialized);
573
574 // Present the results of the fast call.
575 result_region->init_req(fast_path, local_ctrl);
576 result_i_o ->init_req(fast_path, *io);
577 result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
578 }
579
580 // Here are all the slow paths up to this point, in one bundle:
581 assert(slow_region != NULL, "allocated on entry");
582 slow_control = slow_region;
583 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
584
585 *ctrl = checked_control;
586 if (!(*ctrl)->is_top()) {
587 // Clean up after the checked call.
588 // The returned value is either 0 or -1^K,
589 // where K = number of partially transferred array elements.
590 Node* cmp = new CmpINode(checked_value, intcon(0));
591 transform_later(cmp);
592 Node* bol = new BoolNode(cmp, BoolTest::eq);
593 transform_later(bol);
594 IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
595 transform_later(iff);
596
597 // If it is 0, we are done, so transfer to the end.
598 Node* checks_done = new IfTrueNode(iff);
599 transform_later(checks_done);
600 result_region->init_req(checked_path, checks_done);
601 result_i_o ->init_req(checked_path, checked_i_o);
602 result_memory->init_req(checked_path, checked_mem);
603
604 // If it is not zero, merge into the slow call.
605 *ctrl = new IfFalseNode(iff);
606 transform_later(*ctrl);
607 RegionNode* slow_reg2 = new RegionNode(3);
608 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
609 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
610 transform_later(slow_reg2);
611 transform_later(slow_i_o2);
612 transform_later(slow_mem2);
613 slow_reg2 ->init_req(1, slow_control);
614 slow_i_o2 ->init_req(1, slow_i_o);
615 slow_mem2 ->init_req(1, slow_mem);
616 slow_reg2 ->init_req(2, *ctrl);
617 slow_i_o2 ->init_req(2, checked_i_o);
618 slow_mem2 ->init_req(2, checked_mem);
619
620 slow_control = slow_reg2;
621 slow_i_o = slow_i_o2;
622 slow_mem = slow_mem2;
623
624 if (alloc != NULL) {
625 // We'll restart from the very beginning, after zeroing the whole thing.
626 // This can cause double writes, but that's OK since dest is brand new.
627 // So we ignore the low 31 bits of the value returned from the stub.
628 } else {
629 // We must continue the copy exactly where it failed, or else
630 // another thread might see the wrong number of writes to dest.
631 Node* checked_offset = new XorINode(checked_value, intcon(-1));
632 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
633 transform_later(checked_offset);
634 transform_later(slow_offset);
635 slow_offset->init_req(1, intcon(0));
636 slow_offset->init_req(2, checked_offset);
637
638 // Adjust the arguments by the conditionally incoming offset.
639 Node* src_off_plus = new AddINode(src_offset, slow_offset);
640 transform_later(src_off_plus);
641 Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
642 transform_later(dest_off_plus);
643 Node* length_minus = new SubINode(copy_length, slow_offset);
644 transform_later(length_minus);
645
646 // Tweak the node variables to adjust the code produced below:
647 src_offset = src_off_plus;
648 dest_offset = dest_off_plus;
649 copy_length = length_minus;
650 }
651 }
652 *ctrl = slow_control;
653 if (!(*ctrl)->is_top()) {
654 Node* local_ctrl = *ctrl, *local_io = slow_i_o;
655 MergeMemNode* local_mem = MergeMemNode::make(mem);
656 transform_later(local_mem);
657
658 // Generate the slow path, if needed.
659 local_mem->set_memory_at(alias_idx, slow_mem);
660
661 if (dest_uninitialized) {
662 generate_clear_array(local_ctrl, local_mem,
663 adr_type, dest, basic_elem_type,
664 intcon(0), NULL,
665 alloc->in(AllocateNode::AllocSize));
666 }
667
668 local_mem = generate_slow_arraycopy(ac,
669 &local_ctrl, local_mem, &local_io,
670 adr_type,
671 src, src_offset, dest, dest_offset,
672 copy_length, /*dest_uninitialized*/false);
673
674 result_region->init_req(slow_call_path, local_ctrl);
675 result_i_o ->init_req(slow_call_path, local_io);
676 result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
677 } else {
678 ShouldNotReachHere(); // no call to generate_slow_arraycopy:
679 // projections were not extracted
680 }
681
682 // Remove unused edges.
683 for (uint i = 1; i < result_region->req(); i++) {
684 if (result_region->in(i) == NULL) {
685 result_region->init_req(i, top());
686 }
687 }
688
689 // Finished; return the combined state.
690 *ctrl = result_region;
691 *io = result_i_o;
692 mem->set_memory_at(alias_idx, result_memory);
693
694 // mem no longer guaranteed to stay a MergeMemNode
695 Node* out_mem = mem;
696 DEBUG_ONLY(mem = NULL);
697
698 // The memory edges above are precise in order to model effects around
699 // array copies accurately to allow value numbering of field loads around
700 // arraycopy. Such field loads, both before and after, are common in Java
701 // collections and similar classes involving header/array data structures.
702 //
703 // But with low number of register or when some registers are used or killed
704 // by arraycopy calls it causes registers spilling on stack. See 6544710.
705 // The next memory barrier is added to avoid it. If the arraycopy can be
706 // optimized away (which it can, sometimes) then we can manually remove
707 // the membar also.
708 //
709 // Do not let reads from the cloned object float above the arraycopy.
710 if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
711 // Do not let stores that initialize this object be reordered with
712 // a subsequent store that would make this object accessible by
713 // other threads.
714 insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
715 } else if (InsertMemBarAfterArraycopy) {
716 insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
717 }
718
719 _igvn.replace_node(_memproj_fallthrough, out_mem);
720 _igvn.replace_node(_ioproj_fallthrough, *io);
721 _igvn.replace_node(_fallthroughcatchproj, *ctrl);
722
723 #ifdef ASSERT
724 const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
725 if (dest_t->is_known_instance()) {
726 ArrayCopyNode* ac = NULL;
727 assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
728 assert(ac == NULL, "no arraycopy anymore");
729 }
730 #endif
731
732 return out_mem;
733 }
734
735 // Helper for initialization of arrays, creating a ClearArray.
736 // It writes zero bits in [start..end), within the body of an array object.
737 // The memory effects are all chained onto the 'adr_type' alias category.
738 //
739 // Since the object is otherwise uninitialized, we are free
740 // to put a little "slop" around the edges of the cleared area,
741 // as long as it does not go back into the array's header,
742 // or beyond the array end within the heap.
743 //
744 // The lower edge can be rounded down to the nearest jint and the
745 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
746 //
747 // Arguments:
748 // adr_type memory slice where writes are generated
749 // dest oop of the destination array
750 // basic_elem_type element type of the destination
751 // slice_idx array index of first element to store
752 // slice_len number of elements to store (or NULL)
753 // dest_size total size in bytes of the array object
754 //
755 // Exactly one of slice_len or dest_size must be non-NULL.
756 // If dest_size is non-NULL, zeroing extends to the end of the object.
757 // If slice_len is non-NULL, the slice_idx value must be a constant.
758 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
759 const TypePtr* adr_type,
760 Node* dest,
761 BasicType basic_elem_type,
762 Node* slice_idx,
763 Node* slice_len,
764 Node* dest_size) {
765 // one or the other but not both of slice_len and dest_size:
766 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
767 if (slice_len == NULL) slice_len = top();
768 if (dest_size == NULL) dest_size = top();
769
770 uint alias_idx = C->get_alias_index(adr_type);
771
772 // operate on this memory slice:
773 Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
774
775 // scaling and rounding of indexes:
776 int scale = exact_log2(type2aelembytes(basic_elem_type));
777 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
778 int clear_low = (-1 << scale) & (BytesPerInt - 1);
779 int bump_bit = (-1 << scale) & BytesPerInt;
780
781 // determine constant starts and ends
782 const intptr_t BIG_NEG = -128;
783 assert(BIG_NEG + 2*abase < 0, "neg enough");
784 intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
785 intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
786 if (slice_len_con == 0) {
787 return; // nothing to do here
788 }
789 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
790 intptr_t end_con = _igvn.find_intptr_t_con(dest_size, -1);
791 if (slice_idx_con >= 0 && slice_len_con >= 0) {
792 assert(end_con < 0, "not two cons");
793 end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
794 BytesPerLong);
795 }
796
797 if (start_con >= 0 && end_con >= 0) {
798 // Constant start and end. Simple.
799 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
800 start_con, end_con, &_igvn);
801 } else if (start_con >= 0 && dest_size != top()) {
802 // Constant start, pre-rounded end after the tail of the array.
803 Node* end = dest_size;
804 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
805 start_con, end, &_igvn);
806 } else if (start_con >= 0 && slice_len != top()) {
807 // Constant start, non-constant end. End needs rounding up.
808 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
809 intptr_t end_base = abase + (slice_idx_con << scale);
810 int end_round = (-1 << scale) & (BytesPerLong - 1);
811 Node* end = ConvI2X(slice_len);
812 if (scale != 0)
813 end = transform_later(new LShiftXNode(end, intcon(scale) ));
814 end_base += end_round;
815 end = transform_later(new AddXNode(end, MakeConX(end_base)) );
816 end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
817 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
818 start_con, end, &_igvn);
819 } else if (start_con < 0 && dest_size != top()) {
820 // Non-constant start, pre-rounded end after the tail of the array.
821 // This is almost certainly a "round-to-end" operation.
822 Node* start = slice_idx;
823 start = ConvI2X(start);
824 if (scale != 0)
825 start = transform_later(new LShiftXNode( start, intcon(scale) ));
826 start = transform_later(new AddXNode(start, MakeConX(abase)) );
827 if ((bump_bit | clear_low) != 0) {
828 int to_clear = (bump_bit | clear_low);
829 // Align up mod 8, then store a jint zero unconditionally
830 // just before the mod-8 boundary.
831 if (((abase + bump_bit) & ~to_clear) - bump_bit
832 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
833 bump_bit = 0;
834 assert((abase & to_clear) == 0, "array base must be long-aligned");
835 } else {
836 // Bump 'start' up to (or past) the next jint boundary:
837 start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
838 assert((abase & clear_low) == 0, "array base must be int-aligned");
839 }
840 // Round bumped 'start' down to jlong boundary in body of array.
841 start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
842 if (bump_bit != 0) {
843 // Store a zero to the immediately preceding jint:
844 Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
845 Node* p1 = basic_plus_adr(dest, x1);
846 mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
847 mem = transform_later(mem);
848 }
849 }
850 Node* end = dest_size; // pre-rounded
851 mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
852 start, end, &_igvn);
853 } else {
854 // Non-constant start, unrounded non-constant end.
855 // (Nobody zeroes a random midsection of an array using this routine.)
856 ShouldNotReachHere(); // fix caller
857 }
858
859 // Done.
860 merge_mem->set_memory_at(alias_idx, mem);
861 }
862
863 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
864 const TypePtr* adr_type,
865 BasicType basic_elem_type,
866 AllocateNode* alloc,
867 Node* src, Node* src_offset,
868 Node* dest, Node* dest_offset,
869 Node* dest_size, bool dest_uninitialized) {
870 // See if there is an advantage from block transfer.
871 int scale = exact_log2(type2aelembytes(basic_elem_type));
872 if (scale >= LogBytesPerLong)
873 return false; // it is already a block transfer
874
875 // Look at the alignment of the starting offsets.
876 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
877
878 intptr_t src_off_con = (intptr_t) _igvn.find_int_con(src_offset, -1);
879 intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
880 if (src_off_con < 0 || dest_off_con < 0) {
881 // At present, we can only understand constants.
882 return false;
883 }
884
885 intptr_t src_off = abase + (src_off_con << scale);
886 intptr_t dest_off = abase + (dest_off_con << scale);
887
888 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
889 // Non-aligned; too bad.
890 // One more chance: Pick off an initial 32-bit word.
891 // This is a common case, since abase can be odd mod 8.
892 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
893 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
894 Node* sptr = basic_plus_adr(src, src_off);
895 Node* dptr = basic_plus_adr(dest, dest_off);
896 const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
897 assert(s_adr_type->isa_aryptr(), "impossible slice");
898 uint s_alias_idx = C->get_alias_index(s_adr_type);
899 uint d_alias_idx = C->get_alias_index(adr_type);
900 bool is_mismatched = (basic_elem_type != T_INT);
901 Node* sval = transform_later(
902 LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
903 TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
904 false /*unaligned*/, is_mismatched));
905 Node* st = transform_later(
906 StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
907 sval, T_INT, MemNode::unordered));
908 if (is_mismatched) {
909 st->as_Store()->set_mismatched_access();
910 }
911 (*mem)->set_memory_at(d_alias_idx, st);
912 src_off += BytesPerInt;
913 dest_off += BytesPerInt;
914 } else {
915 return false;
916 }
917 }
918 assert(src_off % BytesPerLong == 0, "");
919 assert(dest_off % BytesPerLong == 0, "");
920
921 // Do this copy by giant steps.
922 Node* sptr = basic_plus_adr(src, src_off);
923 Node* dptr = basic_plus_adr(dest, dest_off);
924 Node* countx = dest_size;
925 countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
926 countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
927
928 bool disjoint_bases = true; // since alloc != NULL
929 generate_unchecked_arraycopy(ctrl, mem,
930 adr_type, T_LONG, disjoint_bases,
931 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
932
933 return true;
934 }
935
936 // Helper function; generates code for the slow case.
937 // We make a call to a runtime method which emulates the native method,
938 // but without the native wrapper overhead.
939 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
940 Node** ctrl, Node* mem, Node** io,
941 const TypePtr* adr_type,
942 Node* src, Node* src_offset,
943 Node* dest, Node* dest_offset,
944 Node* copy_length, bool dest_uninitialized) {
945 assert(!dest_uninitialized, "Invariant");
946
947 const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
948 CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
949 "slow_arraycopy",
950 ac->jvms()->bci(), TypePtr::BOTTOM);
951
952 call->init_req(TypeFunc::Control, *ctrl);
953 call->init_req(TypeFunc::I_O , *io);
954 call->init_req(TypeFunc::Memory , mem);
955 call->init_req(TypeFunc::ReturnAdr, top());
956 call->init_req(TypeFunc::FramePtr, top());
957 call->init_req(TypeFunc::Parms+0, src);
958 call->init_req(TypeFunc::Parms+1, src_offset);
959 call->init_req(TypeFunc::Parms+2, dest);
960 call->init_req(TypeFunc::Parms+3, dest_offset);
961 call->init_req(TypeFunc::Parms+4, copy_length);
962 copy_call_debug_info(ac, call);
963
964 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
965 _igvn.replace_node(ac, call);
966 transform_later(call);
967
968 extract_call_projections(call);
969 *ctrl = _fallthroughcatchproj->clone();
970 transform_later(*ctrl);
971
972 Node* m = _memproj_fallthrough->clone();
973 transform_later(m);
974
975 uint alias_idx = C->get_alias_index(adr_type);
976 MergeMemNode* out_mem;
977 if (alias_idx != Compile::AliasIdxBot) {
978 out_mem = MergeMemNode::make(mem);
979 out_mem->set_memory_at(alias_idx, m);
980 } else {
981 out_mem = MergeMemNode::make(m);
982 }
983 transform_later(out_mem);
984
985 *io = _ioproj_fallthrough->clone();
986 transform_later(*io);
987
988 return out_mem;
989 }
990
991 // Helper function; generates code for cases requiring runtime checks.
992 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
993 const TypePtr* adr_type,
994 Node* dest_elem_klass,
995 Node* src, Node* src_offset,
996 Node* dest, Node* dest_offset,
997 Node* copy_length, bool dest_uninitialized) {
998 if ((*ctrl)->is_top()) return NULL;
999
1000 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1001 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1002 return NULL;
1003 }
1004
1005 // Pick out the parameters required to perform a store-check
1006 // for the target array. This is an optimistic check. It will
1007 // look in each non-null element's class, at the desired klass's
1008 // super_check_offset, for the desired klass.
1009 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1010 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1011 Node* n3 = new LoadINode(NULL, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1012 Node* check_offset = ConvI2X(transform_later(n3));
1013 Node* check_value = dest_elem_klass;
1014
1015 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1016 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1017
1018 const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1019 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1020 src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1021
1022 finish_arraycopy_call(call, ctrl, mem, adr_type);
1023
1024 Node* proj = new ProjNode(call, TypeFunc::Parms);
1025 transform_later(proj);
1026
1027 return proj;
1028 }
1029
1030 // Helper function; generates code for cases requiring runtime checks.
1031 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1032 const TypePtr* adr_type,
1033 Node* src, Node* src_offset,
1034 Node* dest, Node* dest_offset,
1035 Node* copy_length, bool dest_uninitialized) {
1036 if ((*ctrl)->is_top()) return NULL;
1037 assert(!dest_uninitialized, "Invariant");
1038
1039 address copyfunc_addr = StubRoutines::generic_arraycopy();
1040 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1041 return NULL;
1042 }
1043
1044 const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1045 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1046 src, src_offset, dest, dest_offset, copy_length);
1047
1048 finish_arraycopy_call(call, ctrl, mem, adr_type);
1049
1050 Node* proj = new ProjNode(call, TypeFunc::Parms);
1051 transform_later(proj);
1052
1053 return proj;
1054 }
1055
1056 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1057 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1058 const TypePtr* adr_type,
1059 BasicType basic_elem_type,
1060 bool disjoint_bases,
1061 Node* src, Node* src_offset,
1062 Node* dest, Node* dest_offset,
1063 Node* copy_length, bool dest_uninitialized) {
1064 if ((*ctrl)->is_top()) return;
1065
1066 Node* src_start = src;
1067 Node* dest_start = dest;
1068 if (src_offset != NULL || dest_offset != NULL) {
1069 src_start = array_element_address(src, src_offset, basic_elem_type);
1070 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1071 }
1072
1073 // Figure out which arraycopy runtime method to call.
1074 const char* copyfunc_name = "arraycopy";
1075 address copyfunc_addr =
1076 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1077 disjoint_bases, copyfunc_name, dest_uninitialized);
1078
1079 const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1080 Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1081 src_start, dest_start, copy_length XTOP);
1082
1083 finish_arraycopy_call(call, ctrl, mem, adr_type);
1084 }
1085
1086 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1087 Node* ctrl = ac->in(TypeFunc::Control);
1088 Node* io = ac->in(TypeFunc::I_O);
1089 Node* src = ac->in(ArrayCopyNode::Src);
1090 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1091 Node* dest = ac->in(ArrayCopyNode::Dest);
1092 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1093 Node* length = ac->in(ArrayCopyNode::Length);
1094 MergeMemNode* merge_mem = NULL;
1095
1096 if (ac->is_clonebasic()) {
1097 assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
1098 Node* mem = ac->in(TypeFunc::Memory);
1099 const char* copyfunc_name = "arraycopy";
1100 address copyfunc_addr =
1101 basictype2arraycopy(T_LONG, NULL, NULL,
1102 true, copyfunc_name, true);
1103
1104 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1105 const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1106
1107 Node* call = make_leaf_call(ctrl, mem, call_type, copyfunc_addr, copyfunc_name, raw_adr_type, src, dest, length XTOP);
1108 transform_later(call);
1109
1110 _igvn.replace_node(ac, call);
1111 return;
1112 } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_cloneoop()) {
1113 Node* mem = ac->in(TypeFunc::Memory);
1114 merge_mem = MergeMemNode::make(mem);
1115 transform_later(merge_mem);
1116
1117 RegionNode* slow_region = new RegionNode(1);
1118 transform_later(slow_region);
1119
1120 AllocateArrayNode* alloc = NULL;
1121 if (ac->is_alloc_tightly_coupled()) {
1122 alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1123 assert(alloc != NULL, "expect alloc");
1124 }
1125
1126 const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1127 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1128 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1129 }
1130 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1131 adr_type, T_OBJECT,
1132 src, src_offset, dest, dest_offset, length,
1133 true, !ac->is_copyofrange());
1134
1135 return;
1136 }
1137
1138 AllocateArrayNode* alloc = NULL;
1139 if (ac->is_alloc_tightly_coupled()) {
1140 alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1141 assert(alloc != NULL, "expect alloc");
1142 }
1143
1144 assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1145
1146 // Compile time checks. If any of these checks cannot be verified at compile time,
1147 // we do not make a fast path for this call. Instead, we let the call remain as it
1148 // is. The checks we choose to mandate at compile time are:
1149 //
1150 // (1) src and dest are arrays.
1151 const Type* src_type = src->Value(&_igvn);
1152 const Type* dest_type = dest->Value(&_igvn);
1153 const TypeAryPtr* top_src = src_type->isa_aryptr();
1154 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1155
1156 BasicType src_elem = T_CONFLICT;
1157 BasicType dest_elem = T_CONFLICT;
1158
1159 if (top_dest != NULL && top_dest->klass() != NULL) {
1160 dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1161 }
1162 if (top_src != NULL && top_src->klass() != NULL) {
1163 src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
1164 }
1165 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
1166 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
1167
1168 if (ac->is_arraycopy_validated() &&
1169 dest_elem != T_CONFLICT &&
1170 src_elem == T_CONFLICT) {
1171 src_elem = dest_elem;
1172 }
1173
1174 if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1175 // Conservatively insert a memory barrier on all memory slices.
1176 // Do not let writes into the source float below the arraycopy.
1177 {
1178 Node* mem = ac->in(TypeFunc::Memory);
1179 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1180
1181 merge_mem = MergeMemNode::make(mem);
1182 transform_later(merge_mem);
1183 }
1184
1185 // Call StubRoutines::generic_arraycopy stub.
1186 Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1187 TypeRawPtr::BOTTOM, T_CONFLICT,
1188 src, src_offset, dest, dest_offset, length,
1189 // If a negative length guard was generated for the ArrayCopyNode,
1190 // the length of the array can never be negative.
1191 false, ac->has_negative_length_guard());
1192
1193 // Do not let reads from the destination float above the arraycopy.
1194 // Since we cannot type the arrays, we don't know which slices
1195 // might be affected. We could restrict this barrier only to those
1196 // memory slices which pertain to array elements--but don't bother.
1197 if (!InsertMemBarAfterArraycopy) {
1198 // (If InsertMemBarAfterArraycopy, there is already one in place.)
1199 insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1200 }
1201 return;
1202 }
1203
1204 assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1205
1206 // (2) src and dest arrays must have elements of the same BasicType
1207 // Figure out the size and type of the elements we will be copying.
1208 if (src_elem != dest_elem || dest_elem == T_VOID) {
1209 // The component types are not the same or are not recognized. Punt.
1210 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1211 {
1212 Node* mem = ac->in(TypeFunc::Memory);
1213 merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1214 }
1215
1216 _igvn.replace_node(_memproj_fallthrough, merge_mem);
1217 _igvn.replace_node(_ioproj_fallthrough, io);
1218 _igvn.replace_node(_fallthroughcatchproj, ctrl);
1219 return;
1220 }
1221
1222 //---------------------------------------------------------------------------
1223 // We will make a fast path for this call to arraycopy.
1224
1225 // We have the following tests left to perform:
1226 //
1227 // (3) src and dest must not be null.
1228 // (4) src_offset must not be negative.
1229 // (5) dest_offset must not be negative.
1230 // (6) length must not be negative.
1231 // (7) src_offset + length must not exceed length of src.
1232 // (8) dest_offset + length must not exceed length of dest.
1233 // (9) each element of an oop array must be assignable
1234
1235 {
1236 Node* mem = ac->in(TypeFunc::Memory);
1237 merge_mem = MergeMemNode::make(mem);
1238 transform_later(merge_mem);
1239 }
1240
1241 RegionNode* slow_region = new RegionNode(1);
1242 transform_later(slow_region);
1243
1244 if (!ac->is_arraycopy_validated()) {
1245 // (3) operands must not be null
1246 // We currently perform our null checks with the null_check routine.
1247 // This means that the null exceptions will be reported in the caller
1248 // rather than (correctly) reported inside of the native arraycopy call.
1249 // This should be corrected, given time. We do our null check with the
1250 // stack pointer restored.
1251 // null checks done library_call.cpp
1252
1253 // (4) src_offset must not be negative.
1254 generate_negative_guard(&ctrl, src_offset, slow_region);
1255
1256 // (5) dest_offset must not be negative.
1257 generate_negative_guard(&ctrl, dest_offset, slow_region);
1258
1259 // (6) length must not be negative (moved to generate_arraycopy()).
1260 // generate_negative_guard(length, slow_region);
1261
1262 // (7) src_offset + length must not exceed length of src.
1263 Node* alen = ac->in(ArrayCopyNode::SrcLen);
1264 assert(alen != NULL, "need src len");
1265 generate_limit_guard(&ctrl,
1266 src_offset, length,
1267 alen,
1268 slow_region);
1269
1270 // (8) dest_offset + length must not exceed length of dest.
1271 alen = ac->in(ArrayCopyNode::DestLen);
1272 assert(alen != NULL, "need dest len");
1273 generate_limit_guard(&ctrl,
1274 dest_offset, length,
1275 alen,
1276 slow_region);
1277
1278 // (9) each element of an oop array must be assignable
1279 // The generate_arraycopy subroutine checks this.
1280 }
1281 // This is where the memory effects are placed:
1282 const TypePtr* adr_type = NULL;
1283 if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1284 adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1285 } else {
1286 adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1287 }
1288
1289 generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1290 adr_type, dest_elem,
1291 src, src_offset, dest, dest_offset, length,
1292 // If a negative length guard was generated for the ArrayCopyNode,
1293 // the length of the array can never be negative.
1294 false, ac->has_negative_length_guard(), slow_region);
1295 }