1 | /* |
2 | * Copyright (c) 2018, 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/c2/barrierSetC2.hpp" |
27 | #include "opto/arraycopynode.hpp" |
28 | #include "opto/convertnode.hpp" |
29 | #include "opto/graphKit.hpp" |
30 | #include "opto/idealKit.hpp" |
31 | #include "opto/macro.hpp" |
32 | #include "opto/narrowptrnode.hpp" |
33 | #include "utilities/macros.hpp" |
34 | |
35 | // By default this is a no-op. |
36 | void BarrierSetC2::resolve_address(C2Access& access) const { } |
37 | |
38 | void* C2ParseAccess::barrier_set_state() const { |
39 | return _kit->barrier_set_state(); |
40 | } |
41 | |
42 | PhaseGVN& C2ParseAccess::gvn() const { return _kit->gvn(); } |
43 | |
44 | bool C2Access::needs_cpu_membar() const { |
45 | bool mismatched = (_decorators & C2_MISMATCHED) != 0; |
46 | bool is_unordered = (_decorators & MO_UNORDERED) != 0; |
47 | bool anonymous = (_decorators & C2_UNSAFE_ACCESS) != 0; |
48 | bool in_heap = (_decorators & IN_HEAP) != 0; |
49 | |
50 | bool is_write = (_decorators & C2_WRITE_ACCESS) != 0; |
51 | bool is_read = (_decorators & C2_READ_ACCESS) != 0; |
52 | bool is_atomic = is_read && is_write; |
53 | |
54 | if (is_atomic) { |
55 | // Atomics always need to be wrapped in CPU membars |
56 | return true; |
57 | } |
58 | |
59 | if (anonymous) { |
60 | // We will need memory barriers unless we can determine a unique |
61 | // alias category for this reference. (Note: If for some reason |
62 | // the barriers get omitted and the unsafe reference begins to "pollute" |
63 | // the alias analysis of the rest of the graph, either Compile::can_alias |
64 | // or Compile::must_alias will throw a diagnostic assert.) |
65 | if (!in_heap || !is_unordered || (mismatched && !_addr.type()->isa_aryptr())) { |
66 | return true; |
67 | } |
68 | } |
69 | |
70 | return false; |
71 | } |
72 | |
73 | Node* BarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const { |
74 | DecoratorSet decorators = access.decorators(); |
75 | |
76 | bool mismatched = (decorators & C2_MISMATCHED) != 0; |
77 | bool unaligned = (decorators & C2_UNALIGNED) != 0; |
78 | bool unsafe = (decorators & C2_UNSAFE_ACCESS) != 0; |
79 | bool requires_atomic_access = (decorators & MO_UNORDERED) == 0; |
80 | |
81 | bool in_native = (decorators & IN_NATIVE) != 0; |
82 | assert(!in_native, "not supported yet" ); |
83 | |
84 | MemNode::MemOrd mo = access.mem_node_mo(); |
85 | |
86 | Node* store; |
87 | if (access.is_parse_access()) { |
88 | C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); |
89 | |
90 | GraphKit* kit = parse_access.kit(); |
91 | if (access.type() == T_DOUBLE) { |
92 | Node* new_val = kit->dstore_rounding(val.node()); |
93 | val.set_node(new_val); |
94 | } |
95 | |
96 | store = kit->store_to_memory(kit->control(), access.addr().node(), val.node(), access.type(), |
97 | access.addr().type(), mo, requires_atomic_access, unaligned, mismatched, unsafe); |
98 | access.set_raw_access(store); |
99 | } else { |
100 | assert(!requires_atomic_access, "not yet supported" ); |
101 | assert(access.is_opt_access(), "either parse or opt access" ); |
102 | C2OptAccess& opt_access = static_cast<C2OptAccess&>(access); |
103 | Node* ctl = opt_access.ctl(); |
104 | MergeMemNode* mm = opt_access.mem(); |
105 | PhaseGVN& gvn = opt_access.gvn(); |
106 | const TypePtr* adr_type = access.addr().type(); |
107 | int alias = gvn.C->get_alias_index(adr_type); |
108 | Node* mem = mm->memory_at(alias); |
109 | |
110 | StoreNode* st = StoreNode::make(gvn, ctl, mem, access.addr().node(), adr_type, val.node(), access.type(), mo); |
111 | if (unaligned) { |
112 | st->set_unaligned_access(); |
113 | } |
114 | if (mismatched) { |
115 | st->set_mismatched_access(); |
116 | } |
117 | store = gvn.transform(st); |
118 | if (store == st) { |
119 | mm->set_memory_at(alias, st); |
120 | } |
121 | } |
122 | return store; |
123 | } |
124 | |
125 | Node* BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { |
126 | DecoratorSet decorators = access.decorators(); |
127 | |
128 | Node* adr = access.addr().node(); |
129 | const TypePtr* adr_type = access.addr().type(); |
130 | |
131 | bool mismatched = (decorators & C2_MISMATCHED) != 0; |
132 | bool requires_atomic_access = (decorators & MO_UNORDERED) == 0; |
133 | bool unaligned = (decorators & C2_UNALIGNED) != 0; |
134 | bool control_dependent = (decorators & C2_CONTROL_DEPENDENT_LOAD) != 0; |
135 | bool pinned = (decorators & C2_PINNED_LOAD) != 0; |
136 | bool unsafe = (decorators & C2_UNSAFE_ACCESS) != 0; |
137 | |
138 | bool in_native = (decorators & IN_NATIVE) != 0; |
139 | |
140 | MemNode::MemOrd mo = access.mem_node_mo(); |
141 | LoadNode::ControlDependency dep = pinned ? LoadNode::Pinned : LoadNode::DependsOnlyOnTest; |
142 | |
143 | Node* load; |
144 | if (access.is_parse_access()) { |
145 | C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); |
146 | GraphKit* kit = parse_access.kit(); |
147 | Node* control = control_dependent ? kit->control() : NULL; |
148 | |
149 | if (in_native) { |
150 | load = kit->make_load(control, adr, val_type, access.type(), mo); |
151 | } else { |
152 | load = kit->make_load(control, adr, val_type, access.type(), adr_type, mo, |
153 | dep, requires_atomic_access, unaligned, mismatched, unsafe); |
154 | } |
155 | access.set_raw_access(load); |
156 | } else { |
157 | assert(!requires_atomic_access, "not yet supported" ); |
158 | assert(access.is_opt_access(), "either parse or opt access" ); |
159 | C2OptAccess& opt_access = static_cast<C2OptAccess&>(access); |
160 | Node* control = control_dependent ? opt_access.ctl() : NULL; |
161 | MergeMemNode* mm = opt_access.mem(); |
162 | PhaseGVN& gvn = opt_access.gvn(); |
163 | Node* mem = mm->memory_at(gvn.C->get_alias_index(adr_type)); |
164 | load = LoadNode::make(gvn, control, mem, adr, adr_type, val_type, access.type(), mo, dep, unaligned, mismatched); |
165 | load = gvn.transform(load); |
166 | } |
167 | |
168 | return load; |
169 | } |
170 | |
171 | class C2AccessFence: public StackObj { |
172 | C2Access& _access; |
173 | Node* _leading_membar; |
174 | |
175 | public: |
176 | C2AccessFence(C2Access& access) : |
177 | _access(access), _leading_membar(NULL) { |
178 | GraphKit* kit = NULL; |
179 | if (access.is_parse_access()) { |
180 | C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); |
181 | kit = parse_access.kit(); |
182 | } |
183 | DecoratorSet decorators = access.decorators(); |
184 | |
185 | bool is_write = (decorators & C2_WRITE_ACCESS) != 0; |
186 | bool is_read = (decorators & C2_READ_ACCESS) != 0; |
187 | bool is_atomic = is_read && is_write; |
188 | |
189 | bool is_volatile = (decorators & MO_SEQ_CST) != 0; |
190 | bool is_release = (decorators & MO_RELEASE) != 0; |
191 | |
192 | if (is_atomic) { |
193 | assert(kit != NULL, "unsupported at optimization time" ); |
194 | // Memory-model-wise, a LoadStore acts like a little synchronized |
195 | // block, so needs barriers on each side. These don't translate |
196 | // into actual barriers on most machines, but we still need rest of |
197 | // compiler to respect ordering. |
198 | if (is_release) { |
199 | _leading_membar = kit->insert_mem_bar(Op_MemBarRelease); |
200 | } else if (is_volatile) { |
201 | if (support_IRIW_for_not_multiple_copy_atomic_cpu) { |
202 | _leading_membar = kit->insert_mem_bar(Op_MemBarVolatile); |
203 | } else { |
204 | _leading_membar = kit->insert_mem_bar(Op_MemBarRelease); |
205 | } |
206 | } |
207 | } else if (is_write) { |
208 | // If reference is volatile, prevent following memory ops from |
209 | // floating down past the volatile write. Also prevents commoning |
210 | // another volatile read. |
211 | if (is_volatile || is_release) { |
212 | assert(kit != NULL, "unsupported at optimization time" ); |
213 | _leading_membar = kit->insert_mem_bar(Op_MemBarRelease); |
214 | } |
215 | } else { |
216 | // Memory barrier to prevent normal and 'unsafe' accesses from |
217 | // bypassing each other. Happens after null checks, so the |
218 | // exception paths do not take memory state from the memory barrier, |
219 | // so there's no problems making a strong assert about mixing users |
220 | // of safe & unsafe memory. |
221 | if (is_volatile && support_IRIW_for_not_multiple_copy_atomic_cpu) { |
222 | assert(kit != NULL, "unsupported at optimization time" ); |
223 | _leading_membar = kit->insert_mem_bar(Op_MemBarVolatile); |
224 | } |
225 | } |
226 | |
227 | if (access.needs_cpu_membar()) { |
228 | assert(kit != NULL, "unsupported at optimization time" ); |
229 | kit->insert_mem_bar(Op_MemBarCPUOrder); |
230 | } |
231 | |
232 | if (is_atomic) { |
233 | // 4984716: MemBars must be inserted before this |
234 | // memory node in order to avoid a false |
235 | // dependency which will confuse the scheduler. |
236 | access.set_memory(); |
237 | } |
238 | } |
239 | |
240 | ~C2AccessFence() { |
241 | GraphKit* kit = NULL; |
242 | if (_access.is_parse_access()) { |
243 | C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(_access); |
244 | kit = parse_access.kit(); |
245 | } |
246 | DecoratorSet decorators = _access.decorators(); |
247 | |
248 | bool is_write = (decorators & C2_WRITE_ACCESS) != 0; |
249 | bool is_read = (decorators & C2_READ_ACCESS) != 0; |
250 | bool is_atomic = is_read && is_write; |
251 | |
252 | bool is_volatile = (decorators & MO_SEQ_CST) != 0; |
253 | bool is_acquire = (decorators & MO_ACQUIRE) != 0; |
254 | |
255 | // If reference is volatile, prevent following volatiles ops from |
256 | // floating up before the volatile access. |
257 | if (_access.needs_cpu_membar()) { |
258 | kit->insert_mem_bar(Op_MemBarCPUOrder); |
259 | } |
260 | |
261 | if (is_atomic) { |
262 | assert(kit != NULL, "unsupported at optimization time" ); |
263 | if (is_acquire || is_volatile) { |
264 | Node* n = _access.raw_access(); |
265 | Node* mb = kit->insert_mem_bar(Op_MemBarAcquire, n); |
266 | if (_leading_membar != NULL) { |
267 | MemBarNode::set_load_store_pair(_leading_membar->as_MemBar(), mb->as_MemBar()); |
268 | } |
269 | } |
270 | } else if (is_write) { |
271 | // If not multiple copy atomic, we do the MemBarVolatile before the load. |
272 | if (is_volatile && !support_IRIW_for_not_multiple_copy_atomic_cpu) { |
273 | assert(kit != NULL, "unsupported at optimization time" ); |
274 | Node* n = _access.raw_access(); |
275 | Node* mb = kit->insert_mem_bar(Op_MemBarVolatile, n); // Use fat membar |
276 | if (_leading_membar != NULL) { |
277 | MemBarNode::set_store_pair(_leading_membar->as_MemBar(), mb->as_MemBar()); |
278 | } |
279 | } |
280 | } else { |
281 | if (is_volatile || is_acquire) { |
282 | assert(kit != NULL, "unsupported at optimization time" ); |
283 | Node* n = _access.raw_access(); |
284 | assert(_leading_membar == NULL || support_IRIW_for_not_multiple_copy_atomic_cpu, "no leading membar expected" ); |
285 | Node* mb = kit->insert_mem_bar(Op_MemBarAcquire, n); |
286 | mb->as_MemBar()->set_trailing_load(); |
287 | } |
288 | } |
289 | } |
290 | }; |
291 | |
292 | Node* BarrierSetC2::store_at(C2Access& access, C2AccessValue& val) const { |
293 | C2AccessFence fence(access); |
294 | resolve_address(access); |
295 | return store_at_resolved(access, val); |
296 | } |
297 | |
298 | Node* BarrierSetC2::load_at(C2Access& access, const Type* val_type) const { |
299 | C2AccessFence fence(access); |
300 | resolve_address(access); |
301 | return load_at_resolved(access, val_type); |
302 | } |
303 | |
304 | MemNode::MemOrd C2Access::mem_node_mo() const { |
305 | bool is_write = (_decorators & C2_WRITE_ACCESS) != 0; |
306 | bool is_read = (_decorators & C2_READ_ACCESS) != 0; |
307 | if ((_decorators & MO_SEQ_CST) != 0) { |
308 | if (is_write && is_read) { |
309 | // For atomic operations |
310 | return MemNode::seqcst; |
311 | } else if (is_write) { |
312 | return MemNode::release; |
313 | } else { |
314 | assert(is_read, "what else?" ); |
315 | return MemNode::acquire; |
316 | } |
317 | } else if ((_decorators & MO_RELEASE) != 0) { |
318 | return MemNode::release; |
319 | } else if ((_decorators & MO_ACQUIRE) != 0) { |
320 | return MemNode::acquire; |
321 | } else if (is_write) { |
322 | // Volatile fields need releasing stores. |
323 | // Non-volatile fields also need releasing stores if they hold an |
324 | // object reference, because the object reference might point to |
325 | // a freshly created object. |
326 | // Conservatively release stores of object references. |
327 | return StoreNode::release_if_reference(_type); |
328 | } else { |
329 | return MemNode::unordered; |
330 | } |
331 | } |
332 | |
333 | void C2Access::fixup_decorators() { |
334 | bool default_mo = (_decorators & MO_DECORATOR_MASK) == 0; |
335 | bool is_unordered = (_decorators & MO_UNORDERED) != 0 || default_mo; |
336 | bool anonymous = (_decorators & C2_UNSAFE_ACCESS) != 0; |
337 | |
338 | bool is_read = (_decorators & C2_READ_ACCESS) != 0; |
339 | bool is_write = (_decorators & C2_WRITE_ACCESS) != 0; |
340 | |
341 | if (AlwaysAtomicAccesses && is_unordered) { |
342 | _decorators &= ~MO_DECORATOR_MASK; // clear the MO bits |
343 | _decorators |= MO_RELAXED; // Force the MO_RELAXED decorator with AlwaysAtomicAccess |
344 | } |
345 | |
346 | _decorators = AccessInternal::decorator_fixup(_decorators); |
347 | |
348 | if (is_read && !is_write && anonymous) { |
349 | // To be valid, unsafe loads may depend on other conditions than |
350 | // the one that guards them: pin the Load node |
351 | _decorators |= C2_CONTROL_DEPENDENT_LOAD; |
352 | _decorators |= C2_PINNED_LOAD; |
353 | const TypePtr* adr_type = _addr.type(); |
354 | Node* adr = _addr.node(); |
355 | if (!needs_cpu_membar() && adr_type->isa_instptr()) { |
356 | assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null" ); |
357 | intptr_t offset = Type::OffsetBot; |
358 | AddPNode::Ideal_base_and_offset(adr, &gvn(), offset); |
359 | if (offset >= 0) { |
360 | int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper()); |
361 | if (offset < s) { |
362 | // Guaranteed to be a valid access, no need to pin it |
363 | _decorators ^= C2_CONTROL_DEPENDENT_LOAD; |
364 | _decorators ^= C2_PINNED_LOAD; |
365 | } |
366 | } |
367 | } |
368 | } |
369 | } |
370 | |
371 | //--------------------------- atomic operations--------------------------------- |
372 | |
373 | void BarrierSetC2::pin_atomic_op(C2AtomicParseAccess& access) const { |
374 | if (!access.needs_pinning()) { |
375 | return; |
376 | } |
377 | // SCMemProjNodes represent the memory state of a LoadStore. Their |
378 | // main role is to prevent LoadStore nodes from being optimized away |
379 | // when their results aren't used. |
380 | assert(access.is_parse_access(), "entry not supported at optimization time" ); |
381 | C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); |
382 | GraphKit* kit = parse_access.kit(); |
383 | Node* load_store = access.raw_access(); |
384 | assert(load_store != NULL, "must pin atomic op" ); |
385 | Node* proj = kit->gvn().transform(new SCMemProjNode(load_store)); |
386 | kit->set_memory(proj, access.alias_idx()); |
387 | } |
388 | |
389 | void C2AtomicParseAccess::set_memory() { |
390 | Node *mem = _kit->memory(_alias_idx); |
391 | _memory = mem; |
392 | } |
393 | |
394 | Node* BarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, |
395 | Node* new_val, const Type* value_type) const { |
396 | GraphKit* kit = access.kit(); |
397 | MemNode::MemOrd mo = access.mem_node_mo(); |
398 | Node* mem = access.memory(); |
399 | |
400 | Node* adr = access.addr().node(); |
401 | const TypePtr* adr_type = access.addr().type(); |
402 | |
403 | Node* load_store = NULL; |
404 | |
405 | if (access.is_oop()) { |
406 | #ifdef _LP64 |
407 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
408 | Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); |
409 | Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); |
410 | load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); |
411 | } else |
412 | #endif |
413 | { |
414 | load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); |
415 | } |
416 | } else { |
417 | switch (access.type()) { |
418 | case T_BYTE: { |
419 | load_store = kit->gvn().transform(new CompareAndExchangeBNode(kit->control(), mem, adr, new_val, expected_val, adr_type, mo)); |
420 | break; |
421 | } |
422 | case T_SHORT: { |
423 | load_store = kit->gvn().transform(new CompareAndExchangeSNode(kit->control(), mem, adr, new_val, expected_val, adr_type, mo)); |
424 | break; |
425 | } |
426 | case T_INT: { |
427 | load_store = kit->gvn().transform(new CompareAndExchangeINode(kit->control(), mem, adr, new_val, expected_val, adr_type, mo)); |
428 | break; |
429 | } |
430 | case T_LONG: { |
431 | load_store = kit->gvn().transform(new CompareAndExchangeLNode(kit->control(), mem, adr, new_val, expected_val, adr_type, mo)); |
432 | break; |
433 | } |
434 | default: |
435 | ShouldNotReachHere(); |
436 | } |
437 | } |
438 | |
439 | access.set_raw_access(load_store); |
440 | pin_atomic_op(access); |
441 | |
442 | #ifdef _LP64 |
443 | if (access.is_oop() && adr->bottom_type()->is_ptr_to_narrowoop()) { |
444 | return kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); |
445 | } |
446 | #endif |
447 | |
448 | return load_store; |
449 | } |
450 | |
451 | Node* BarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, |
452 | Node* new_val, const Type* value_type) const { |
453 | GraphKit* kit = access.kit(); |
454 | DecoratorSet decorators = access.decorators(); |
455 | MemNode::MemOrd mo = access.mem_node_mo(); |
456 | Node* mem = access.memory(); |
457 | bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0; |
458 | Node* load_store = NULL; |
459 | Node* adr = access.addr().node(); |
460 | |
461 | if (access.is_oop()) { |
462 | #ifdef _LP64 |
463 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
464 | Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); |
465 | Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); |
466 | if (is_weak_cas) { |
467 | load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); |
468 | } else { |
469 | load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); |
470 | } |
471 | } else |
472 | #endif |
473 | { |
474 | if (is_weak_cas) { |
475 | load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
476 | } else { |
477 | load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
478 | } |
479 | } |
480 | } else { |
481 | switch(access.type()) { |
482 | case T_BYTE: { |
483 | if (is_weak_cas) { |
484 | load_store = kit->gvn().transform(new WeakCompareAndSwapBNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
485 | } else { |
486 | load_store = kit->gvn().transform(new CompareAndSwapBNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
487 | } |
488 | break; |
489 | } |
490 | case T_SHORT: { |
491 | if (is_weak_cas) { |
492 | load_store = kit->gvn().transform(new WeakCompareAndSwapSNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
493 | } else { |
494 | load_store = kit->gvn().transform(new CompareAndSwapSNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
495 | } |
496 | break; |
497 | } |
498 | case T_INT: { |
499 | if (is_weak_cas) { |
500 | load_store = kit->gvn().transform(new WeakCompareAndSwapINode(kit->control(), mem, adr, new_val, expected_val, mo)); |
501 | } else { |
502 | load_store = kit->gvn().transform(new CompareAndSwapINode(kit->control(), mem, adr, new_val, expected_val, mo)); |
503 | } |
504 | break; |
505 | } |
506 | case T_LONG: { |
507 | if (is_weak_cas) { |
508 | load_store = kit->gvn().transform(new WeakCompareAndSwapLNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
509 | } else { |
510 | load_store = kit->gvn().transform(new CompareAndSwapLNode(kit->control(), mem, adr, new_val, expected_val, mo)); |
511 | } |
512 | break; |
513 | } |
514 | default: |
515 | ShouldNotReachHere(); |
516 | } |
517 | } |
518 | |
519 | access.set_raw_access(load_store); |
520 | pin_atomic_op(access); |
521 | |
522 | return load_store; |
523 | } |
524 | |
525 | Node* BarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const { |
526 | GraphKit* kit = access.kit(); |
527 | Node* mem = access.memory(); |
528 | Node* adr = access.addr().node(); |
529 | const TypePtr* adr_type = access.addr().type(); |
530 | Node* load_store = NULL; |
531 | |
532 | if (access.is_oop()) { |
533 | #ifdef _LP64 |
534 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
535 | Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); |
536 | load_store = kit->gvn().transform(new GetAndSetNNode(kit->control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop())); |
537 | } else |
538 | #endif |
539 | { |
540 | load_store = kit->gvn().transform(new GetAndSetPNode(kit->control(), mem, adr, new_val, adr_type, value_type->is_oopptr())); |
541 | } |
542 | } else { |
543 | switch (access.type()) { |
544 | case T_BYTE: |
545 | load_store = kit->gvn().transform(new GetAndSetBNode(kit->control(), mem, adr, new_val, adr_type)); |
546 | break; |
547 | case T_SHORT: |
548 | load_store = kit->gvn().transform(new GetAndSetSNode(kit->control(), mem, adr, new_val, adr_type)); |
549 | break; |
550 | case T_INT: |
551 | load_store = kit->gvn().transform(new GetAndSetINode(kit->control(), mem, adr, new_val, adr_type)); |
552 | break; |
553 | case T_LONG: |
554 | load_store = kit->gvn().transform(new GetAndSetLNode(kit->control(), mem, adr, new_val, adr_type)); |
555 | break; |
556 | default: |
557 | ShouldNotReachHere(); |
558 | } |
559 | } |
560 | |
561 | access.set_raw_access(load_store); |
562 | pin_atomic_op(access); |
563 | |
564 | #ifdef _LP64 |
565 | if (access.is_oop() && adr->bottom_type()->is_ptr_to_narrowoop()) { |
566 | return kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); |
567 | } |
568 | #endif |
569 | |
570 | return load_store; |
571 | } |
572 | |
573 | Node* BarrierSetC2::atomic_add_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const { |
574 | Node* load_store = NULL; |
575 | GraphKit* kit = access.kit(); |
576 | Node* adr = access.addr().node(); |
577 | const TypePtr* adr_type = access.addr().type(); |
578 | Node* mem = access.memory(); |
579 | |
580 | switch(access.type()) { |
581 | case T_BYTE: |
582 | load_store = kit->gvn().transform(new GetAndAddBNode(kit->control(), mem, adr, new_val, adr_type)); |
583 | break; |
584 | case T_SHORT: |
585 | load_store = kit->gvn().transform(new GetAndAddSNode(kit->control(), mem, adr, new_val, adr_type)); |
586 | break; |
587 | case T_INT: |
588 | load_store = kit->gvn().transform(new GetAndAddINode(kit->control(), mem, adr, new_val, adr_type)); |
589 | break; |
590 | case T_LONG: |
591 | load_store = kit->gvn().transform(new GetAndAddLNode(kit->control(), mem, adr, new_val, adr_type)); |
592 | break; |
593 | default: |
594 | ShouldNotReachHere(); |
595 | } |
596 | |
597 | access.set_raw_access(load_store); |
598 | pin_atomic_op(access); |
599 | |
600 | return load_store; |
601 | } |
602 | |
603 | Node* BarrierSetC2::atomic_cmpxchg_val_at(C2AtomicParseAccess& access, Node* expected_val, |
604 | Node* new_val, const Type* value_type) const { |
605 | C2AccessFence fence(access); |
606 | resolve_address(access); |
607 | return atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type); |
608 | } |
609 | |
610 | Node* BarrierSetC2::atomic_cmpxchg_bool_at(C2AtomicParseAccess& access, Node* expected_val, |
611 | Node* new_val, const Type* value_type) const { |
612 | C2AccessFence fence(access); |
613 | resolve_address(access); |
614 | return atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); |
615 | } |
616 | |
617 | Node* BarrierSetC2::atomic_xchg_at(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const { |
618 | C2AccessFence fence(access); |
619 | resolve_address(access); |
620 | return atomic_xchg_at_resolved(access, new_val, value_type); |
621 | } |
622 | |
623 | Node* BarrierSetC2::atomic_add_at(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const { |
624 | C2AccessFence fence(access); |
625 | resolve_address(access); |
626 | return atomic_add_at_resolved(access, new_val, value_type); |
627 | } |
628 | |
629 | void BarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const { |
630 | // Exclude the header but include array length to copy by 8 bytes words. |
631 | // Can't use base_offset_in_bytes(bt) since basic type is unknown. |
632 | int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : |
633 | instanceOopDesc::base_offset_in_bytes(); |
634 | // base_off: |
635 | // 8 - 32-bit VM |
636 | // 12 - 64-bit VM, compressed klass |
637 | // 16 - 64-bit VM, normal klass |
638 | if (base_off % BytesPerLong != 0) { |
639 | assert(UseCompressedClassPointers, "" ); |
640 | if (is_array) { |
641 | // Exclude length to copy by 8 bytes words. |
642 | base_off += sizeof(int); |
643 | } else { |
644 | // Include klass to copy by 8 bytes words. |
645 | base_off = instanceOopDesc::klass_offset_in_bytes(); |
646 | } |
647 | assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment" ); |
648 | } |
649 | Node* src_base = kit->basic_plus_adr(src, base_off); |
650 | Node* dst_base = kit->basic_plus_adr(dst, base_off); |
651 | |
652 | // Compute the length also, if needed: |
653 | Node* countx = size; |
654 | countx = kit->gvn().transform(new SubXNode(countx, kit->MakeConX(base_off))); |
655 | countx = kit->gvn().transform(new URShiftXNode(countx, kit->intcon(LogBytesPerLong) )); |
656 | |
657 | const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; |
658 | |
659 | ArrayCopyNode* ac = ArrayCopyNode::make(kit, false, src_base, NULL, dst_base, NULL, countx, true, false); |
660 | ac->set_clonebasic(); |
661 | Node* n = kit->gvn().transform(ac); |
662 | if (n == ac) { |
663 | ac->_adr_type = TypeRawPtr::BOTTOM; |
664 | kit->set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); |
665 | } else { |
666 | kit->set_all_memory(n); |
667 | } |
668 | } |
669 | |
670 | Node* BarrierSetC2::obj_allocate(PhaseMacroExpand* macro, Node* ctrl, Node* mem, Node* toobig_false, Node* size_in_bytes, |
671 | Node*& i_o, Node*& needgc_ctrl, |
672 | Node*& fast_oop_ctrl, Node*& fast_oop_rawmem, |
673 | intx prefetch_lines) const { |
674 | |
675 | Node* eden_top_adr; |
676 | Node* eden_end_adr; |
677 | |
678 | macro->set_eden_pointers(eden_top_adr, eden_end_adr); |
679 | |
680 | // Load Eden::end. Loop invariant and hoisted. |
681 | // |
682 | // Note: We set the control input on "eden_end" and "old_eden_top" when using |
683 | // a TLAB to work around a bug where these values were being moved across |
684 | // a safepoint. These are not oops, so they cannot be include in the oop |
685 | // map, but they can be changed by a GC. The proper way to fix this would |
686 | // be to set the raw memory state when generating a SafepointNode. However |
687 | // this will require extensive changes to the loop optimization in order to |
688 | // prevent a degradation of the optimization. |
689 | // See comment in memnode.hpp, around line 227 in class LoadPNode. |
690 | Node *eden_end = macro->make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); |
691 | |
692 | // We need a Region for the loop-back contended case. |
693 | enum { fall_in_path = 1, contended_loopback_path = 2 }; |
694 | Node *contended_region; |
695 | Node *contended_phi_rawmem; |
696 | if (UseTLAB) { |
697 | contended_region = toobig_false; |
698 | contended_phi_rawmem = mem; |
699 | } else { |
700 | contended_region = new RegionNode(3); |
701 | contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); |
702 | // Now handle the passing-too-big test. We fall into the contended |
703 | // loop-back merge point. |
704 | contended_region ->init_req(fall_in_path, toobig_false); |
705 | contended_phi_rawmem->init_req(fall_in_path, mem); |
706 | macro->transform_later(contended_region); |
707 | macro->transform_later(contended_phi_rawmem); |
708 | } |
709 | |
710 | // Load(-locked) the heap top. |
711 | // See note above concerning the control input when using a TLAB |
712 | Node *old_eden_top = UseTLAB |
713 | ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) |
714 | : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); |
715 | |
716 | macro->transform_later(old_eden_top); |
717 | // Add to heap top to get a new heap top |
718 | Node *new_eden_top = new AddPNode(macro->top(), old_eden_top, size_in_bytes); |
719 | macro->transform_later(new_eden_top); |
720 | // Check for needing a GC; compare against heap end |
721 | Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); |
722 | macro->transform_later(needgc_cmp); |
723 | Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); |
724 | macro->transform_later(needgc_bol); |
725 | IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); |
726 | macro->transform_later(needgc_iff); |
727 | |
728 | // Plug the failing-heap-space-need-gc test into the slow-path region |
729 | Node *needgc_true = new IfTrueNode(needgc_iff); |
730 | macro->transform_later(needgc_true); |
731 | needgc_ctrl = needgc_true; |
732 | |
733 | // No need for a GC. Setup for the Store-Conditional |
734 | Node *needgc_false = new IfFalseNode(needgc_iff); |
735 | macro->transform_later(needgc_false); |
736 | |
737 | i_o = macro->prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, |
738 | old_eden_top, new_eden_top, prefetch_lines); |
739 | |
740 | Node* fast_oop = old_eden_top; |
741 | |
742 | // Store (-conditional) the modified eden top back down. |
743 | // StorePConditional produces flags for a test PLUS a modified raw |
744 | // memory state. |
745 | if (UseTLAB) { |
746 | Node* store_eden_top = |
747 | new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, |
748 | TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); |
749 | macro->transform_later(store_eden_top); |
750 | fast_oop_ctrl = needgc_false; // No contention, so this is the fast path |
751 | fast_oop_rawmem = store_eden_top; |
752 | } else { |
753 | Node* store_eden_top = |
754 | new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, |
755 | new_eden_top, fast_oop/*old_eden_top*/); |
756 | macro->transform_later(store_eden_top); |
757 | Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); |
758 | macro->transform_later(contention_check); |
759 | store_eden_top = new SCMemProjNode(store_eden_top); |
760 | macro->transform_later(store_eden_top); |
761 | |
762 | // If not using TLABs, check to see if there was contention. |
763 | IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); |
764 | macro->transform_later(contention_iff); |
765 | Node *contention_true = new IfTrueNode(contention_iff); |
766 | macro->transform_later(contention_true); |
767 | // If contention, loopback and try again. |
768 | contended_region->init_req(contended_loopback_path, contention_true); |
769 | contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); |
770 | |
771 | // Fast-path succeeded with no contention! |
772 | Node *contention_false = new IfFalseNode(contention_iff); |
773 | macro->transform_later(contention_false); |
774 | fast_oop_ctrl = contention_false; |
775 | |
776 | // Bump total allocated bytes for this thread |
777 | Node* thread = new ThreadLocalNode(); |
778 | macro->transform_later(thread); |
779 | Node* alloc_bytes_adr = macro->basic_plus_adr(macro->top()/*not oop*/, thread, |
780 | in_bytes(JavaThread::allocated_bytes_offset())); |
781 | Node* alloc_bytes = macro->make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, |
782 | 0, TypeLong::LONG, T_LONG); |
783 | #ifdef _LP64 |
784 | Node* alloc_size = size_in_bytes; |
785 | #else |
786 | Node* alloc_size = new ConvI2LNode(size_in_bytes); |
787 | macro->transform_later(alloc_size); |
788 | #endif |
789 | Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); |
790 | macro->transform_later(new_alloc_bytes); |
791 | fast_oop_rawmem = macro->make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, |
792 | 0, new_alloc_bytes, T_LONG); |
793 | } |
794 | return fast_oop; |
795 | } |
796 | |
797 | void BarrierSetC2::clone_barrier_at_expansion(ArrayCopyNode* ac, Node* call, PhaseIterGVN& igvn) const { |
798 | // no barrier |
799 | igvn.replace_node(ac, call); |
800 | } |
801 | |