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