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/g1/c2/g1BarrierSetC2.hpp" |
27 | #include "gc/g1/g1BarrierSet.hpp" |
28 | #include "gc/g1/g1BarrierSetRuntime.hpp" |
29 | #include "gc/g1/g1CardTable.hpp" |
30 | #include "gc/g1/g1ThreadLocalData.hpp" |
31 | #include "gc/g1/heapRegion.hpp" |
32 | #include "opto/arraycopynode.hpp" |
33 | #include "opto/compile.hpp" |
34 | #include "opto/escape.hpp" |
35 | #include "opto/graphKit.hpp" |
36 | #include "opto/idealKit.hpp" |
37 | #include "opto/macro.hpp" |
38 | #include "opto/rootnode.hpp" |
39 | #include "opto/type.hpp" |
40 | #include "utilities/macros.hpp" |
41 | |
42 | const TypeFunc *G1BarrierSetC2::write_ref_field_pre_entry_Type() { |
43 | const Type **fields = TypeTuple::fields(2); |
44 | fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value |
45 | fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread |
46 | const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); |
47 | |
48 | // create result type (range) |
49 | fields = TypeTuple::fields(0); |
50 | const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); |
51 | |
52 | return TypeFunc::make(domain, range); |
53 | } |
54 | |
55 | const TypeFunc *G1BarrierSetC2::write_ref_field_post_entry_Type() { |
56 | const Type **fields = TypeTuple::fields(2); |
57 | fields[TypeFunc::Parms+0] = TypeRawPtr::NOTNULL; // Card addr |
58 | fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread |
59 | const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); |
60 | |
61 | // create result type (range) |
62 | fields = TypeTuple::fields(0); |
63 | const TypeTuple *range = TypeTuple::make(TypeFunc::Parms, fields); |
64 | |
65 | return TypeFunc::make(domain, range); |
66 | } |
67 | |
68 | #define __ ideal. |
69 | /* |
70 | * Determine if the G1 pre-barrier can be removed. The pre-barrier is |
71 | * required by SATB to make sure all objects live at the start of the |
72 | * marking are kept alive, all reference updates need to any previous |
73 | * reference stored before writing. |
74 | * |
75 | * If the previous value is NULL there is no need to save the old value. |
76 | * References that are NULL are filtered during runtime by the barrier |
77 | * code to avoid unnecessary queuing. |
78 | * |
79 | * However in the case of newly allocated objects it might be possible to |
80 | * prove that the reference about to be overwritten is NULL during compile |
81 | * time and avoid adding the barrier code completely. |
82 | * |
83 | * The compiler needs to determine that the object in which a field is about |
84 | * to be written is newly allocated, and that no prior store to the same field |
85 | * has happened since the allocation. |
86 | * |
87 | * Returns true if the pre-barrier can be removed |
88 | */ |
89 | bool G1BarrierSetC2::g1_can_remove_pre_barrier(GraphKit* kit, |
90 | PhaseTransform* phase, |
91 | Node* adr, |
92 | BasicType bt, |
93 | uint adr_idx) const { |
94 | intptr_t offset = 0; |
95 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
96 | AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); |
97 | |
98 | if (offset == Type::OffsetBot) { |
99 | return false; // cannot unalias unless there are precise offsets |
100 | } |
101 | |
102 | if (alloc == NULL) { |
103 | return false; // No allocation found |
104 | } |
105 | |
106 | intptr_t size_in_bytes = type2aelembytes(bt); |
107 | |
108 | Node* mem = kit->memory(adr_idx); // start searching here... |
109 | |
110 | for (int cnt = 0; cnt < 50; cnt++) { |
111 | |
112 | if (mem->is_Store()) { |
113 | |
114 | Node* st_adr = mem->in(MemNode::Address); |
115 | intptr_t st_offset = 0; |
116 | Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); |
117 | |
118 | if (st_base == NULL) { |
119 | break; // inscrutable pointer |
120 | } |
121 | |
122 | // Break we have found a store with same base and offset as ours so break |
123 | if (st_base == base && st_offset == offset) { |
124 | break; |
125 | } |
126 | |
127 | if (st_offset != offset && st_offset != Type::OffsetBot) { |
128 | const int MAX_STORE = BytesPerLong; |
129 | if (st_offset >= offset + size_in_bytes || |
130 | st_offset <= offset - MAX_STORE || |
131 | st_offset <= offset - mem->as_Store()->memory_size()) { |
132 | // Success: The offsets are provably independent. |
133 | // (You may ask, why not just test st_offset != offset and be done? |
134 | // The answer is that stores of different sizes can co-exist |
135 | // in the same sequence of RawMem effects. We sometimes initialize |
136 | // a whole 'tile' of array elements with a single jint or jlong.) |
137 | mem = mem->in(MemNode::Memory); |
138 | continue; // advance through independent store memory |
139 | } |
140 | } |
141 | |
142 | if (st_base != base |
143 | && MemNode::detect_ptr_independence(base, alloc, st_base, |
144 | AllocateNode::Ideal_allocation(st_base, phase), |
145 | phase)) { |
146 | // Success: The bases are provably independent. |
147 | mem = mem->in(MemNode::Memory); |
148 | continue; // advance through independent store memory |
149 | } |
150 | } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { |
151 | |
152 | InitializeNode* st_init = mem->in(0)->as_Initialize(); |
153 | AllocateNode* st_alloc = st_init->allocation(); |
154 | |
155 | // Make sure that we are looking at the same allocation site. |
156 | // The alloc variable is guaranteed to not be null here from earlier check. |
157 | if (alloc == st_alloc) { |
158 | // Check that the initialization is storing NULL so that no previous store |
159 | // has been moved up and directly write a reference |
160 | Node* captured_store = st_init->find_captured_store(offset, |
161 | type2aelembytes(T_OBJECT), |
162 | phase); |
163 | if (captured_store == NULL || captured_store == st_init->zero_memory()) { |
164 | return true; |
165 | } |
166 | } |
167 | } |
168 | |
169 | // Unless there is an explicit 'continue', we must bail out here, |
170 | // because 'mem' is an inscrutable memory state (e.g., a call). |
171 | break; |
172 | } |
173 | |
174 | return false; |
175 | } |
176 | |
177 | // G1 pre/post barriers |
178 | void G1BarrierSetC2::pre_barrier(GraphKit* kit, |
179 | bool do_load, |
180 | Node* ctl, |
181 | Node* obj, |
182 | Node* adr, |
183 | uint alias_idx, |
184 | Node* val, |
185 | const TypeOopPtr* val_type, |
186 | Node* pre_val, |
187 | BasicType bt) const { |
188 | // Some sanity checks |
189 | // Note: val is unused in this routine. |
190 | |
191 | if (do_load) { |
192 | // We need to generate the load of the previous value |
193 | assert(obj != NULL, "must have a base" ); |
194 | assert(adr != NULL, "where are loading from?" ); |
195 | assert(pre_val == NULL, "loaded already?" ); |
196 | assert(val_type != NULL, "need a type" ); |
197 | |
198 | if (use_ReduceInitialCardMarks() |
199 | && g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) { |
200 | return; |
201 | } |
202 | |
203 | } else { |
204 | // In this case both val_type and alias_idx are unused. |
205 | assert(pre_val != NULL, "must be loaded already" ); |
206 | // Nothing to be done if pre_val is null. |
207 | if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; |
208 | assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here" ); |
209 | } |
210 | assert(bt == T_OBJECT, "or we shouldn't be here" ); |
211 | |
212 | IdealKit ideal(kit, true); |
213 | |
214 | Node* tls = __ thread(); // ThreadLocalStorage |
215 | |
216 | Node* no_base = __ top(); |
217 | Node* zero = __ ConI(0); |
218 | Node* zeroX = __ ConX(0); |
219 | |
220 | float likely = PROB_LIKELY(0.999); |
221 | float unlikely = PROB_UNLIKELY(0.999); |
222 | |
223 | BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; |
224 | assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width" ); |
225 | |
226 | // Offsets into the thread |
227 | const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); |
228 | const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset()); |
229 | const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()); |
230 | |
231 | // Now the actual pointers into the thread |
232 | Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); |
233 | Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); |
234 | Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); |
235 | |
236 | // Now some of the values |
237 | Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); |
238 | |
239 | // if (!marking) |
240 | __ if_then(marking, BoolTest::ne, zero, unlikely); { |
241 | BasicType index_bt = TypeX_X->basic_type(); |
242 | assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size." ); |
243 | Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); |
244 | |
245 | if (do_load) { |
246 | // load original value |
247 | // alias_idx correct?? |
248 | pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); |
249 | } |
250 | |
251 | // if (pre_val != NULL) |
252 | __ if_then(pre_val, BoolTest::ne, kit->null()); { |
253 | Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); |
254 | |
255 | // is the queue for this thread full? |
256 | __ if_then(index, BoolTest::ne, zeroX, likely); { |
257 | |
258 | // decrement the index |
259 | Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); |
260 | |
261 | // Now get the buffer location we will log the previous value into and store it |
262 | Node *log_addr = __ AddP(no_base, buffer, next_index); |
263 | __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); |
264 | // update the index |
265 | __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); |
266 | |
267 | } __ else_(); { |
268 | |
269 | // logging buffer is full, call the runtime |
270 | const TypeFunc *tf = write_ref_field_pre_entry_Type(); |
271 | __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), "write_ref_field_pre_entry" , pre_val, tls); |
272 | } __ end_if(); // (!index) |
273 | } __ end_if(); // (pre_val != NULL) |
274 | } __ end_if(); // (!marking) |
275 | |
276 | // Final sync IdealKit and GraphKit. |
277 | kit->final_sync(ideal); |
278 | } |
279 | |
280 | /* |
281 | * G1 similar to any GC with a Young Generation requires a way to keep track of |
282 | * references from Old Generation to Young Generation to make sure all live |
283 | * objects are found. G1 also requires to keep track of object references |
284 | * between different regions to enable evacuation of old regions, which is done |
285 | * as part of mixed collections. References are tracked in remembered sets and |
286 | * is continuously updated as reference are written to with the help of the |
287 | * post-barrier. |
288 | * |
289 | * To reduce the number of updates to the remembered set the post-barrier |
290 | * filters updates to fields in objects located in the Young Generation, |
291 | * the same region as the reference, when the NULL is being written or |
292 | * if the card is already marked as dirty by an earlier write. |
293 | * |
294 | * Under certain circumstances it is possible to avoid generating the |
295 | * post-barrier completely if it is possible during compile time to prove |
296 | * the object is newly allocated and that no safepoint exists between the |
297 | * allocation and the store. |
298 | * |
299 | * In the case of slow allocation the allocation code must handle the barrier |
300 | * as part of the allocation in the case the allocated object is not located |
301 | * in the nursery, this would happen for humongous objects. This is similar to |
302 | * how CMS is required to handle this case, see the comments for the method |
303 | * CollectedHeap::new_deferred_store_barrier and OptoRuntime::new_deferred_store_barrier. |
304 | * A deferred card mark is required for these objects and handled in the above |
305 | * mentioned methods. |
306 | * |
307 | * Returns true if the post barrier can be removed |
308 | */ |
309 | bool G1BarrierSetC2::g1_can_remove_post_barrier(GraphKit* kit, |
310 | PhaseTransform* phase, Node* store, |
311 | Node* adr) const { |
312 | intptr_t offset = 0; |
313 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
314 | AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); |
315 | |
316 | if (offset == Type::OffsetBot) { |
317 | return false; // cannot unalias unless there are precise offsets |
318 | } |
319 | |
320 | if (alloc == NULL) { |
321 | return false; // No allocation found |
322 | } |
323 | |
324 | // Start search from Store node |
325 | Node* mem = store->in(MemNode::Control); |
326 | if (mem->is_Proj() && mem->in(0)->is_Initialize()) { |
327 | |
328 | InitializeNode* st_init = mem->in(0)->as_Initialize(); |
329 | AllocateNode* st_alloc = st_init->allocation(); |
330 | |
331 | // Make sure we are looking at the same allocation |
332 | if (alloc == st_alloc) { |
333 | return true; |
334 | } |
335 | } |
336 | |
337 | return false; |
338 | } |
339 | |
340 | // |
341 | // Update the card table and add card address to the queue |
342 | // |
343 | void G1BarrierSetC2::g1_mark_card(GraphKit* kit, |
344 | IdealKit& ideal, |
345 | Node* card_adr, |
346 | Node* oop_store, |
347 | uint oop_alias_idx, |
348 | Node* index, |
349 | Node* index_adr, |
350 | Node* buffer, |
351 | const TypeFunc* tf) const { |
352 | Node* zero = __ ConI(0); |
353 | Node* zeroX = __ ConX(0); |
354 | Node* no_base = __ top(); |
355 | BasicType card_bt = T_BYTE; |
356 | // Smash zero into card. MUST BE ORDERED WRT TO STORE |
357 | __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); |
358 | |
359 | // Now do the queue work |
360 | __ if_then(index, BoolTest::ne, zeroX); { |
361 | |
362 | Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); |
363 | Node* log_addr = __ AddP(no_base, buffer, next_index); |
364 | |
365 | // Order, see storeCM. |
366 | __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); |
367 | __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); |
368 | |
369 | } __ else_(); { |
370 | __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry), "write_ref_field_post_entry" , card_adr, __ thread()); |
371 | } __ end_if(); |
372 | |
373 | } |
374 | |
375 | void G1BarrierSetC2::post_barrier(GraphKit* kit, |
376 | Node* ctl, |
377 | Node* oop_store, |
378 | Node* obj, |
379 | Node* adr, |
380 | uint alias_idx, |
381 | Node* val, |
382 | BasicType bt, |
383 | bool use_precise) const { |
384 | // If we are writing a NULL then we need no post barrier |
385 | |
386 | if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { |
387 | // Must be NULL |
388 | const Type* t = val->bottom_type(); |
389 | assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL" ); |
390 | // No post barrier if writing NULLx |
391 | return; |
392 | } |
393 | |
394 | if (use_ReduceInitialCardMarks() && obj == kit->just_allocated_object(kit->control())) { |
395 | // We can skip marks on a freshly-allocated object in Eden. |
396 | // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. |
397 | // That routine informs GC to take appropriate compensating steps, |
398 | // upon a slow-path allocation, so as to make this card-mark |
399 | // elision safe. |
400 | return; |
401 | } |
402 | |
403 | if (use_ReduceInitialCardMarks() |
404 | && g1_can_remove_post_barrier(kit, &kit->gvn(), oop_store, adr)) { |
405 | return; |
406 | } |
407 | |
408 | if (!use_precise) { |
409 | // All card marks for a (non-array) instance are in one place: |
410 | adr = obj; |
411 | } |
412 | // (Else it's an array (or unknown), and we want more precise card marks.) |
413 | assert(adr != NULL, "" ); |
414 | |
415 | IdealKit ideal(kit, true); |
416 | |
417 | Node* tls = __ thread(); // ThreadLocalStorage |
418 | |
419 | Node* no_base = __ top(); |
420 | float unlikely = PROB_UNLIKELY(0.999); |
421 | Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val()); |
422 | Node* dirty_card = __ ConI((jint)G1CardTable::dirty_card_val()); |
423 | Node* zeroX = __ ConX(0); |
424 | |
425 | const TypeFunc *tf = write_ref_field_post_entry_Type(); |
426 | |
427 | // Offsets into the thread |
428 | const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset()); |
429 | const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset()); |
430 | |
431 | // Pointers into the thread |
432 | |
433 | Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); |
434 | Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); |
435 | |
436 | // Now some values |
437 | // Use ctrl to avoid hoisting these values past a safepoint, which could |
438 | // potentially reset these fields in the JavaThread. |
439 | Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); |
440 | Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); |
441 | |
442 | // Convert the store obj pointer to an int prior to doing math on it |
443 | // Must use ctrl to prevent "integerized oop" existing across safepoint |
444 | Node* cast = __ CastPX(__ ctrl(), adr); |
445 | |
446 | // Divide pointer by card size |
447 | Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); |
448 | |
449 | // Combine card table base and card offset |
450 | Node* card_adr = __ AddP(no_base, byte_map_base_node(kit), card_offset ); |
451 | |
452 | // If we know the value being stored does it cross regions? |
453 | |
454 | if (val != NULL) { |
455 | // Does the store cause us to cross regions? |
456 | |
457 | // Should be able to do an unsigned compare of region_size instead of |
458 | // and extra shift. Do we have an unsigned compare?? |
459 | // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); |
460 | Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); |
461 | |
462 | // if (xor_res == 0) same region so skip |
463 | __ if_then(xor_res, BoolTest::ne, zeroX); { |
464 | |
465 | // No barrier if we are storing a NULL |
466 | __ if_then(val, BoolTest::ne, kit->null(), unlikely); { |
467 | |
468 | // Ok must mark the card if not already dirty |
469 | |
470 | // load the original value of the card |
471 | Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); |
472 | |
473 | __ if_then(card_val, BoolTest::ne, young_card); { |
474 | kit->sync_kit(ideal); |
475 | kit->insert_store_load_for_barrier(); |
476 | __ sync_kit(kit); |
477 | |
478 | Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); |
479 | __ if_then(card_val_reload, BoolTest::ne, dirty_card); { |
480 | g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); |
481 | } __ end_if(); |
482 | } __ end_if(); |
483 | } __ end_if(); |
484 | } __ end_if(); |
485 | } else { |
486 | // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. |
487 | // We don't need a barrier here if the destination is a newly allocated object |
488 | // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden |
489 | // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()). |
490 | assert(!use_ReduceInitialCardMarks(), "can only happen with card marking" ); |
491 | Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); |
492 | __ if_then(card_val, BoolTest::ne, young_card); { |
493 | g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); |
494 | } __ end_if(); |
495 | } |
496 | |
497 | // Final sync IdealKit and GraphKit. |
498 | kit->final_sync(ideal); |
499 | } |
500 | |
501 | // Helper that guards and inserts a pre-barrier. |
502 | void G1BarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset, |
503 | Node* pre_val, bool need_mem_bar) const { |
504 | // We could be accessing the referent field of a reference object. If so, when G1 |
505 | // is enabled, we need to log the value in the referent field in an SATB buffer. |
506 | // This routine performs some compile time filters and generates suitable |
507 | // runtime filters that guard the pre-barrier code. |
508 | // Also add memory barrier for non volatile load from the referent field |
509 | // to prevent commoning of loads across safepoint. |
510 | |
511 | // Some compile time checks. |
512 | |
513 | // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? |
514 | const TypeX* otype = offset->find_intptr_t_type(); |
515 | if (otype != NULL && otype->is_con() && |
516 | otype->get_con() != java_lang_ref_Reference::referent_offset) { |
517 | // Constant offset but not the reference_offset so just return |
518 | return; |
519 | } |
520 | |
521 | // We only need to generate the runtime guards for instances. |
522 | const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); |
523 | if (btype != NULL) { |
524 | if (btype->isa_aryptr()) { |
525 | // Array type so nothing to do |
526 | return; |
527 | } |
528 | |
529 | const TypeInstPtr* itype = btype->isa_instptr(); |
530 | if (itype != NULL) { |
531 | // Can the klass of base_oop be statically determined to be |
532 | // _not_ a sub-class of Reference and _not_ Object? |
533 | ciKlass* klass = itype->klass(); |
534 | if ( klass->is_loaded() && |
535 | !klass->is_subtype_of(kit->env()->Reference_klass()) && |
536 | !kit->env()->Object_klass()->is_subtype_of(klass)) { |
537 | return; |
538 | } |
539 | } |
540 | } |
541 | |
542 | // The compile time filters did not reject base_oop/offset so |
543 | // we need to generate the following runtime filters |
544 | // |
545 | // if (offset == java_lang_ref_Reference::_reference_offset) { |
546 | // if (instance_of(base, java.lang.ref.Reference)) { |
547 | // pre_barrier(_, pre_val, ...); |
548 | // } |
549 | // } |
550 | |
551 | float likely = PROB_LIKELY( 0.999); |
552 | float unlikely = PROB_UNLIKELY(0.999); |
553 | |
554 | IdealKit ideal(kit); |
555 | |
556 | Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); |
557 | |
558 | __ if_then(offset, BoolTest::eq, referent_off, unlikely); { |
559 | // Update graphKit memory and control from IdealKit. |
560 | kit->sync_kit(ideal); |
561 | |
562 | Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass())); |
563 | Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con); |
564 | |
565 | // Update IdealKit memory and control from graphKit. |
566 | __ sync_kit(kit); |
567 | |
568 | Node* one = __ ConI(1); |
569 | // is_instof == 0 if base_oop == NULL |
570 | __ if_then(is_instof, BoolTest::eq, one, unlikely); { |
571 | |
572 | // Update graphKit from IdeakKit. |
573 | kit->sync_kit(ideal); |
574 | |
575 | // Use the pre-barrier to record the value in the referent field |
576 | pre_barrier(kit, false /* do_load */, |
577 | __ ctrl(), |
578 | NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, |
579 | pre_val /* pre_val */, |
580 | T_OBJECT); |
581 | if (need_mem_bar) { |
582 | // Add memory barrier to prevent commoning reads from this field |
583 | // across safepoint since GC can change its value. |
584 | kit->insert_mem_bar(Op_MemBarCPUOrder); |
585 | } |
586 | // Update IdealKit from graphKit. |
587 | __ sync_kit(kit); |
588 | |
589 | } __ end_if(); // _ref_type != ref_none |
590 | } __ end_if(); // offset == referent_offset |
591 | |
592 | // Final sync IdealKit and GraphKit. |
593 | kit->final_sync(ideal); |
594 | } |
595 | |
596 | #undef __ |
597 | |
598 | Node* G1BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { |
599 | DecoratorSet decorators = access.decorators(); |
600 | Node* adr = access.addr().node(); |
601 | Node* obj = access.base(); |
602 | |
603 | bool mismatched = (decorators & C2_MISMATCHED) != 0; |
604 | bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0; |
605 | bool in_heap = (decorators & IN_HEAP) != 0; |
606 | bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0; |
607 | bool is_unordered = (decorators & MO_UNORDERED) != 0; |
608 | bool need_cpu_mem_bar = !is_unordered || mismatched || !in_heap; |
609 | |
610 | Node* top = Compile::current()->top(); |
611 | Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top; |
612 | Node* load = CardTableBarrierSetC2::load_at_resolved(access, val_type); |
613 | |
614 | // If we are reading the value of the referent field of a Reference |
615 | // object (either by using Unsafe directly or through reflection) |
616 | // then, if G1 is enabled, we need to record the referent in an |
617 | // SATB log buffer using the pre-barrier mechanism. |
618 | // Also we need to add memory barrier to prevent commoning reads |
619 | // from this field across safepoint since GC can change its value. |
620 | bool need_read_barrier = in_heap && (on_weak || |
621 | (unknown && offset != top && obj != top)); |
622 | |
623 | if (!access.is_oop() || !need_read_barrier) { |
624 | return load; |
625 | } |
626 | |
627 | assert(access.is_parse_access(), "entry not supported at optimization time" ); |
628 | C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); |
629 | GraphKit* kit = parse_access.kit(); |
630 | |
631 | if (on_weak) { |
632 | // Use the pre-barrier to record the value in the referent field |
633 | pre_barrier(kit, false /* do_load */, |
634 | kit->control(), |
635 | NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, |
636 | load /* pre_val */, T_OBJECT); |
637 | // Add memory barrier to prevent commoning reads from this field |
638 | // across safepoint since GC can change its value. |
639 | kit->insert_mem_bar(Op_MemBarCPUOrder); |
640 | } else if (unknown) { |
641 | // We do not require a mem bar inside pre_barrier if need_mem_bar |
642 | // is set: the barriers would be emitted by us. |
643 | insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar); |
644 | } |
645 | |
646 | return load; |
647 | } |
648 | |
649 | bool G1BarrierSetC2::is_gc_barrier_node(Node* node) const { |
650 | if (CardTableBarrierSetC2::is_gc_barrier_node(node)) { |
651 | return true; |
652 | } |
653 | if (node->Opcode() != Op_CallLeaf) { |
654 | return false; |
655 | } |
656 | CallLeafNode *call = node->as_CallLeaf(); |
657 | if (call->_name == NULL) { |
658 | return false; |
659 | } |
660 | |
661 | return strcmp(call->_name, "write_ref_field_pre_entry" ) == 0 || strcmp(call->_name, "write_ref_field_post_entry" ) == 0; |
662 | } |
663 | |
664 | void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const { |
665 | assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required" ); |
666 | assert(node->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes" ); |
667 | // It could be only one user, URShift node, in Object.clone() intrinsic |
668 | // but the new allocation is passed to arraycopy stub and it could not |
669 | // be scalar replaced. So we don't check the case. |
670 | |
671 | // An other case of only one user (Xor) is when the value check for NULL |
672 | // in G1 post barrier is folded after CCP so the code which used URShift |
673 | // is removed. |
674 | |
675 | // Take Region node before eliminating post barrier since it also |
676 | // eliminates CastP2X node when it has only one user. |
677 | Node* this_region = node->in(0); |
678 | assert(this_region != NULL, "" ); |
679 | |
680 | // Remove G1 post barrier. |
681 | |
682 | // Search for CastP2X->Xor->URShift->Cmp path which |
683 | // checks if the store done to a different from the value's region. |
684 | // And replace Cmp with #0 (false) to collapse G1 post barrier. |
685 | Node* xorx = node->find_out_with(Op_XorX); |
686 | if (xorx != NULL) { |
687 | Node* shift = xorx->unique_out(); |
688 | Node* cmpx = shift->unique_out(); |
689 | assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && |
690 | cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, |
691 | "missing region check in G1 post barrier" ); |
692 | macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ)); |
693 | |
694 | // Remove G1 pre barrier. |
695 | |
696 | // Search "if (marking != 0)" check and set it to "false". |
697 | // There is no G1 pre barrier if previous stored value is NULL |
698 | // (for example, after initialization). |
699 | if (this_region->is_Region() && this_region->req() == 3) { |
700 | int ind = 1; |
701 | if (!this_region->in(ind)->is_IfFalse()) { |
702 | ind = 2; |
703 | } |
704 | if (this_region->in(ind)->is_IfFalse() && |
705 | this_region->in(ind)->in(0)->Opcode() == Op_If) { |
706 | Node* bol = this_region->in(ind)->in(0)->in(1); |
707 | assert(bol->is_Bool(), "" ); |
708 | cmpx = bol->in(1); |
709 | if (bol->as_Bool()->_test._test == BoolTest::ne && |
710 | cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) && |
711 | cmpx->in(1)->is_Load()) { |
712 | Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); |
713 | const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); |
714 | if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() && |
715 | adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && |
716 | adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) { |
717 | macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ)); |
718 | } |
719 | } |
720 | } |
721 | } |
722 | } else { |
723 | assert(!use_ReduceInitialCardMarks(), "can only happen with card marking" ); |
724 | // This is a G1 post barrier emitted by the Object.clone() intrinsic. |
725 | // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card |
726 | // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier. |
727 | Node* shift = node->find_out_with(Op_URShiftX); |
728 | assert(shift != NULL, "missing G1 post barrier" ); |
729 | Node* addp = shift->unique_out(); |
730 | Node* load = addp->find_out_with(Op_LoadB); |
731 | assert(load != NULL, "missing G1 post barrier" ); |
732 | Node* cmpx = load->unique_out(); |
733 | assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && |
734 | cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, |
735 | "missing card value check in G1 post barrier" ); |
736 | macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ)); |
737 | // There is no G1 pre barrier in this case |
738 | } |
739 | // Now CastP2X can be removed since it is used only on dead path |
740 | // which currently still alive until igvn optimize it. |
741 | assert(node->outcnt() == 0 || node->unique_out()->Opcode() == Op_URShiftX, "" ); |
742 | macro->replace_node(node, macro->top()); |
743 | } |
744 | |
745 | Node* G1BarrierSetC2::step_over_gc_barrier(Node* c) const { |
746 | if (!use_ReduceInitialCardMarks() && |
747 | c != NULL && c->is_Region() && c->req() == 3) { |
748 | for (uint i = 1; i < c->req(); i++) { |
749 | if (c->in(i) != NULL && c->in(i)->is_Region() && |
750 | c->in(i)->req() == 3) { |
751 | Node* r = c->in(i); |
752 | for (uint j = 1; j < r->req(); j++) { |
753 | if (r->in(j) != NULL && r->in(j)->is_Proj() && |
754 | r->in(j)->in(0) != NULL && |
755 | r->in(j)->in(0)->Opcode() == Op_CallLeaf && |
756 | r->in(j)->in(0)->as_Call()->entry_point() == CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry)) { |
757 | Node* call = r->in(j)->in(0); |
758 | c = c->in(i == 1 ? 2 : 1); |
759 | if (c != NULL) { |
760 | c = c->in(0); |
761 | if (c != NULL) { |
762 | c = c->in(0); |
763 | assert(call->in(0) == NULL || |
764 | call->in(0)->in(0) == NULL || |
765 | call->in(0)->in(0)->in(0) == NULL || |
766 | call->in(0)->in(0)->in(0)->in(0) == NULL || |
767 | call->in(0)->in(0)->in(0)->in(0)->in(0) == NULL || |
768 | c == call->in(0)->in(0)->in(0)->in(0)->in(0), "bad barrier shape" ); |
769 | return c; |
770 | } |
771 | } |
772 | } |
773 | } |
774 | } |
775 | } |
776 | } |
777 | return c; |
778 | } |
779 | |
780 | #ifdef ASSERT |
781 | void G1BarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const { |
782 | if (phase != BarrierSetC2::BeforeCodeGen) { |
783 | return; |
784 | } |
785 | // Verify G1 pre-barriers |
786 | const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); |
787 | |
788 | ResourceArea *area = Thread::current()->resource_area(); |
789 | Unique_Node_List visited(area); |
790 | Node_List worklist(area); |
791 | // We're going to walk control flow backwards starting from the Root |
792 | worklist.push(compile->root()); |
793 | while (worklist.size() > 0) { |
794 | Node* x = worklist.pop(); |
795 | if (x == NULL || x == compile->top()) continue; |
796 | if (visited.member(x)) { |
797 | continue; |
798 | } else { |
799 | visited.push(x); |
800 | } |
801 | |
802 | if (x->is_Region()) { |
803 | for (uint i = 1; i < x->req(); i++) { |
804 | worklist.push(x->in(i)); |
805 | } |
806 | } else { |
807 | worklist.push(x->in(0)); |
808 | // We are looking for the pattern: |
809 | // /->ThreadLocal |
810 | // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) |
811 | // \->ConI(0) |
812 | // We want to verify that the If and the LoadB have the same control |
813 | // See GraphKit::g1_write_barrier_pre() |
814 | if (x->is_If()) { |
815 | IfNode *iff = x->as_If(); |
816 | if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { |
817 | CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); |
818 | if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 |
819 | && cmp->in(1)->is_Load()) { |
820 | LoadNode* load = cmp->in(1)->as_Load(); |
821 | if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal |
822 | && load->in(2)->in(3)->is_Con() |
823 | && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { |
824 | |
825 | Node* if_ctrl = iff->in(0); |
826 | Node* load_ctrl = load->in(0); |
827 | |
828 | if (if_ctrl != load_ctrl) { |
829 | // Skip possible CProj->NeverBranch in infinite loops |
830 | if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) |
831 | && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { |
832 | if_ctrl = if_ctrl->in(0)->in(0); |
833 | } |
834 | } |
835 | assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match" ); |
836 | } |
837 | } |
838 | } |
839 | } |
840 | } |
841 | } |
842 | } |
843 | #endif |
844 | |
845 | bool G1BarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const { |
846 | if (opcode == Op_StoreP) { |
847 | Node* adr = n->in(MemNode::Address); |
848 | const Type* adr_type = gvn->type(adr); |
849 | // Pointer stores in G1 barriers looks like unsafe access. |
850 | // Ignore such stores to be able scalar replace non-escaping |
851 | // allocations. |
852 | if (adr_type->isa_rawptr() && adr->is_AddP()) { |
853 | Node* base = conn_graph->get_addp_base(adr); |
854 | if (base->Opcode() == Op_LoadP && |
855 | base->in(MemNode::Address)->is_AddP()) { |
856 | adr = base->in(MemNode::Address); |
857 | Node* tls = conn_graph->get_addp_base(adr); |
858 | if (tls->Opcode() == Op_ThreadLocal) { |
859 | int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); |
860 | const int buf_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()); |
861 | if (offs == buf_offset) { |
862 | return true; // G1 pre barrier previous oop value store. |
863 | } |
864 | if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) { |
865 | return true; // G1 post barrier card address store. |
866 | } |
867 | } |
868 | } |
869 | } |
870 | } |
871 | return false; |
872 | } |
873 | |