1 /*
2 * Copyright (c) 2018, 2024, 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 "classfile/javaClasses.hpp"
27 #include "code/vmreg.inline.hpp"
28 #include "gc/g1/c2/g1BarrierSetC2.hpp"
29 #include "gc/g1/g1BarrierSet.hpp"
30 #include "gc/g1/g1BarrierSetAssembler.hpp"
31 #include "gc/g1/g1BarrierSetRuntime.hpp"
32 #include "gc/g1/g1CardTable.hpp"
33 #include "gc/g1/g1ThreadLocalData.hpp"
34 #include "gc/g1/g1HeapRegion.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/node.hpp"
45 #include "opto/output.hpp"
46 #include "opto/regalloc.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/type.hpp"
50 #include "utilities/growableArray.hpp"
51 #include "utilities/macros.hpp"
52
53 /*
54 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
55 * required by SATB to make sure all objects live at the start of the
56 * marking are kept alive, all reference updates need to any previous
57 * reference stored before writing.
58 *
59 * If the previous value is null there is no need to save the old value.
60 * References that are null are filtered during runtime by the barrier
61 * code to avoid unnecessary queuing.
62 *
63 * However in the case of newly allocated objects it might be possible to
64 * prove that the reference about to be overwritten is null during compile
65 * time and avoid adding the barrier code completely.
66 *
67 * The compiler needs to determine that the object in which a field is about
68 * to be written is newly allocated, and that no prior store to the same field
69 * has happened since the allocation.
70 */
71 bool G1BarrierSetC2::g1_can_remove_pre_barrier(GraphKit* kit,
72 PhaseValues* phase,
73 Node* adr,
74 BasicType bt,
75 uint adr_idx) const {
76 intptr_t offset = 0;
77 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
78 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
79
80 if (offset == Type::OffsetBot) {
81 return false; // Cannot unalias unless there are precise offsets.
82 }
83 if (alloc == nullptr) {
84 return false; // No allocation found.
85 }
86
87 intptr_t size_in_bytes = type2aelembytes(bt);
88 Node* mem = kit->memory(adr_idx); // Start searching here.
89
90 for (int cnt = 0; cnt < 50; cnt++) {
91 if (mem->is_Store()) {
92 Node* st_adr = mem->in(MemNode::Address);
93 intptr_t st_offset = 0;
94 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
95
96 if (st_base == nullptr) {
97 break; // Inscrutable pointer.
98 }
99 if (st_base == base && st_offset == offset) {
100 // We have found a store with same base and offset as ours.
101 break;
102 }
103 if (st_offset != offset && st_offset != Type::OffsetBot) {
104 const int MAX_STORE = BytesPerLong;
105 if (st_offset >= offset + size_in_bytes ||
106 st_offset <= offset - MAX_STORE ||
107 st_offset <= offset - mem->as_Store()->memory_size()) {
108 // Success: The offsets are provably independent.
109 // (You may ask, why not just test st_offset != offset and be done?
110 // The answer is that stores of different sizes can co-exist
111 // in the same sequence of RawMem effects. We sometimes initialize
112 // a whole 'tile' of array elements with a single jint or jlong.)
113 mem = mem->in(MemNode::Memory);
114 continue; // Advance through independent store memory.
115 }
116 }
117 if (st_base != base
118 && MemNode::detect_ptr_independence(base, alloc, st_base,
119 AllocateNode::Ideal_allocation(st_base),
120 phase)) {
121 // Success: the bases are provably independent.
122 mem = mem->in(MemNode::Memory);
123 continue; // Advance through independent store memory.
124 }
125 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
126 InitializeNode* st_init = mem->in(0)->as_Initialize();
127 AllocateNode* st_alloc = st_init->allocation();
128
129 // Make sure that we are looking at the same allocation site.
130 // The alloc variable is guaranteed to not be null here from earlier check.
131 if (alloc == st_alloc) {
132 // Check that the initialization is storing null so that no previous store
133 // has been moved up and directly write a reference.
134 Node* captured_store = st_init->find_captured_store(offset,
135 type2aelembytes(T_OBJECT),
136 phase);
137 if (captured_store == nullptr || captured_store == st_init->zero_memory()) {
138 return true;
139 }
140 }
141 }
142 // Unless there is an explicit 'continue', we must bail out here,
143 // because 'mem' is an inscrutable memory state (e.g., a call).
144 break;
145 }
146 return false;
147 }
148
149 /*
150 * G1, similar to any GC with a Young Generation, requires a way to keep track
151 * of references from Old Generation to Young Generation to make sure all live
152 * objects are found. G1 also requires to keep track of object references
153 * between different regions to enable evacuation of old regions, which is done
154 * as part of mixed collections. References are tracked in remembered sets,
155 * which are continuously updated as references are written to with the help of
156 * the post-barrier.
157 *
158 * To reduce the number of updates to the remembered set, the post-barrier
159 * filters out updates to fields in objects located in the Young Generation, the
160 * same region as the reference, when null is being written, or if the card is
161 * already marked as dirty by an earlier write.
162 *
163 * Under certain circumstances it is possible to avoid generating the
164 * post-barrier completely, if it is possible during compile time to prove the
165 * object is newly allocated and that no safepoint exists between the allocation
166 * and the store. This can be seen as a compile-time version of the
167 * above-mentioned Young Generation filter.
168 *
169 * In the case of a slow allocation, the allocation code must handle the barrier
170 * as part of the allocation if the allocated object is not located in the
171 * nursery; this would happen for humongous objects.
172 */
173 bool G1BarrierSetC2::g1_can_remove_post_barrier(GraphKit* kit,
174 PhaseValues* phase, Node* store_ctrl,
175 Node* adr) const {
176 intptr_t offset = 0;
177 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
178 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
179
180 if (offset == Type::OffsetBot) {
181 return false; // Cannot unalias unless there are precise offsets.
182 }
183 if (alloc == nullptr) {
184 return false; // No allocation found.
185 }
186
187 Node* mem = store_ctrl; // Start search from Store node.
188 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
189 InitializeNode* st_init = mem->in(0)->as_Initialize();
190 AllocateNode* st_alloc = st_init->allocation();
191 // Make sure we are looking at the same allocation
192 if (alloc == st_alloc) {
193 return true;
194 }
195 }
196
197 return false;
198 }
199
200 Node* G1BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
201 DecoratorSet decorators = access.decorators();
202 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
203 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
204 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
205 // If we are reading the value of the referent field of a Reference object, we
206 // need to record the referent in an SATB log buffer using the pre-barrier
207 // mechanism. Also we need to add a memory barrier to prevent commoning reads
208 // from this field across safepoints, since GC can change its value.
209 bool need_read_barrier = ((on_weak || on_phantom) && !no_keepalive);
210 if (access.is_oop() && need_read_barrier) {
211 access.set_barrier_data(G1C2BarrierPre);
212 }
213 return CardTableBarrierSetC2::load_at_resolved(access, val_type);
214 }
215
216 void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
217 eliminate_gc_barrier_data(node);
218 }
219
220 void G1BarrierSetC2::eliminate_gc_barrier_data(Node* node) const {
221 if (node->is_LoadStore()) {
222 LoadStoreNode* loadstore = node->as_LoadStore();
223 loadstore->set_barrier_data(0);
224 } else if (node->is_Mem()) {
225 MemNode* mem = node->as_Mem();
226 mem->set_barrier_data(0);
227 }
228 }
229
230 static void refine_barrier_by_new_val_type(const Node* n) {
231 if (n->Opcode() != Op_StoreP &&
232 n->Opcode() != Op_StoreN) {
233 return;
234 }
235 MemNode* store = n->as_Mem();
236 const Node* newval = n->in(MemNode::ValueIn);
237 assert(newval != nullptr, "");
238 const Type* newval_bottom = newval->bottom_type();
239 TypePtr::PTR newval_type = newval_bottom->make_ptr()->ptr();
240 uint8_t barrier_data = store->barrier_data();
241 if (!newval_bottom->isa_oopptr() &&
242 !newval_bottom->isa_narrowoop() &&
243 newval_type != TypePtr::Null) {
244 // newval is neither an OOP nor null, so there is no barrier to refine.
245 assert(barrier_data == 0, "non-OOP stores should have no barrier data");
246 return;
247 }
248 if (barrier_data == 0) {
249 // No barrier to refine.
250 return;
251 }
252 if (newval_type == TypePtr::Null) {
253 // Simply elide post-barrier if writing null.
254 barrier_data &= ~G1C2BarrierPost;
255 barrier_data &= ~G1C2BarrierPostNotNull;
256 } else if (((barrier_data & G1C2BarrierPost) != 0) &&
257 newval_type == TypePtr::NotNull) {
258 // If the post-barrier has not been elided yet (e.g. due to newval being
259 // freshly allocated), mark it as not-null (simplifies barrier tests and
260 // compressed OOPs logic).
261 barrier_data |= G1C2BarrierPostNotNull;
262 }
263 store->set_barrier_data(barrier_data);
264 return;
265 }
266
267 // Refine (not really expand) G1 barriers by looking at the new value type
268 // (whether it is necessarily null or necessarily non-null).
269 bool G1BarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
270 ResourceMark rm;
271 VectorSet visited;
272 Node_List worklist;
273 worklist.push(C->root());
274 while (worklist.size() > 0) {
275 Node* n = worklist.pop();
276 if (visited.test_set(n->_idx)) {
277 continue;
278 }
279 refine_barrier_by_new_val_type(n);
280 for (uint j = 0; j < n->req(); j++) {
281 Node* in = n->in(j);
282 if (in != nullptr) {
283 worklist.push(in);
284 }
285 }
286 }
287 return false;
288 }
289
290 uint G1BarrierSetC2::estimated_barrier_size(const Node* node) const {
291 // These Ideal node counts are extracted from the pre-matching Ideal graph
292 // generated when compiling the following method with early barrier expansion:
293 // static void write(MyObject obj1, Object o) {
294 // obj1.o1 = o;
295 // }
296 uint8_t barrier_data = MemNode::barrier_data(node);
297 uint nodes = 0;
298 if ((barrier_data & G1C2BarrierPre) != 0) {
299 nodes += 50;
300 }
301 if ((barrier_data & G1C2BarrierPost) != 0) {
302 nodes += 60;
303 }
304 return nodes;
305 }
306
307 bool G1BarrierSetC2::can_initialize_object(const StoreNode* store) const {
308 assert(store->Opcode() == Op_StoreP || store->Opcode() == Op_StoreN, "OOP store expected");
309 // It is OK to move the store across the object initialization boundary only
310 // if it does not have any barrier, or if it has barriers that can be safely
311 // elided (because of the compensation steps taken on the allocation slow path
312 // when ReduceInitialCardMarks is enabled).
313 return (MemNode::barrier_data(store) == 0) || use_ReduceInitialCardMarks();
314 }
315
316 void G1BarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
317 if (ac->is_clone_inst() && !use_ReduceInitialCardMarks()) {
318 clone_in_runtime(phase, ac, G1BarrierSetRuntime::clone_addr(), "G1BarrierSetRuntime::clone");
319 return;
320 }
321 BarrierSetC2::clone_at_expansion(phase, ac);
322 }
323
324 Node* G1BarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
325 DecoratorSet decorators = access.decorators();
326 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
327 bool in_heap = (decorators & IN_HEAP) != 0;
328 bool tightly_coupled_alloc = (decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0;
329 bool need_store_barrier = !(tightly_coupled_alloc && use_ReduceInitialCardMarks()) && (in_heap || anonymous);
330 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
331 if (access.is_oop() && need_store_barrier) {
332 access.set_barrier_data(get_store_barrier(access));
333 if (tightly_coupled_alloc) {
334 assert(!use_ReduceInitialCardMarks(),
335 "post-barriers are only needed for tightly-coupled initialization stores when ReduceInitialCardMarks is disabled");
336 // Pre-barriers are unnecessary for tightly-coupled initialization stores.
337 access.set_barrier_data(access.barrier_data() & ~G1C2BarrierPre);
338 }
339 }
340 if (no_keepalive) {
341 // No keep-alive means no need for the pre-barrier.
342 access.set_barrier_data(access.barrier_data() & ~G1C2BarrierPre);
343 }
344 return BarrierSetC2::store_at_resolved(access, val);
345 }
346
347 Node* G1BarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
348 Node* new_val, const Type* value_type) const {
349 GraphKit* kit = access.kit();
350 if (!access.is_oop()) {
351 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
352 }
353 access.set_barrier_data(G1C2BarrierPre | G1C2BarrierPost);
354 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
355 }
356
357 Node* G1BarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
358 Node* new_val, const Type* value_type) const {
359 GraphKit* kit = access.kit();
360 if (!access.is_oop()) {
361 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
362 }
363 access.set_barrier_data(G1C2BarrierPre | G1C2BarrierPost);
364 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
365 }
366
367 Node* G1BarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const {
368 GraphKit* kit = access.kit();
369 if (!access.is_oop()) {
370 return BarrierSetC2::atomic_xchg_at_resolved(access, new_val, value_type);
371 }
372 access.set_barrier_data(G1C2BarrierPre | G1C2BarrierPost);
373 return BarrierSetC2::atomic_xchg_at_resolved(access, new_val, value_type);
374 }
375
376 class G1BarrierSetC2State : public BarrierSetC2State {
377 private:
378 GrowableArray<G1BarrierStubC2*>* _stubs;
379
380 public:
381 G1BarrierSetC2State(Arena* arena)
382 : BarrierSetC2State(arena),
383 _stubs(new (arena) GrowableArray<G1BarrierStubC2*>(arena, 8, 0, nullptr)) {}
384
385 GrowableArray<G1BarrierStubC2*>* stubs() {
386 return _stubs;
387 }
388
389 bool needs_liveness_data(const MachNode* mach) const {
390 return G1PreBarrierStubC2::needs_barrier(mach) ||
391 G1PostBarrierStubC2::needs_barrier(mach);
392 }
393
394 bool needs_livein_data() const {
395 return false;
396 }
397 };
398
399 static G1BarrierSetC2State* barrier_set_state() {
400 return reinterpret_cast<G1BarrierSetC2State*>(Compile::current()->barrier_set_state());
401 }
402
403 G1BarrierStubC2::G1BarrierStubC2(const MachNode* node) : BarrierStubC2(node) {}
404
405 G1PreBarrierStubC2::G1PreBarrierStubC2(const MachNode* node) : G1BarrierStubC2(node) {}
406
407 bool G1PreBarrierStubC2::needs_barrier(const MachNode* node) {
408 return (node->barrier_data() & G1C2BarrierPre) != 0;
409 }
410
411 G1PreBarrierStubC2* G1PreBarrierStubC2::create(const MachNode* node) {
412 G1PreBarrierStubC2* const stub = new (Compile::current()->comp_arena()) G1PreBarrierStubC2(node);
413 if (!Compile::current()->output()->in_scratch_emit_size()) {
414 barrier_set_state()->stubs()->append(stub);
415 }
416 return stub;
417 }
418
419 void G1PreBarrierStubC2::initialize_registers(Register obj, Register pre_val, Register thread, Register tmp1, Register tmp2) {
420 _obj = obj;
421 _pre_val = pre_val;
422 _thread = thread;
423 _tmp1 = tmp1;
424 _tmp2 = tmp2;
425 }
426
427 Register G1PreBarrierStubC2::obj() const {
428 return _obj;
429 }
430
431 Register G1PreBarrierStubC2::pre_val() const {
432 return _pre_val;
433 }
434
435 Register G1PreBarrierStubC2::thread() const {
436 return _thread;
437 }
438
439 Register G1PreBarrierStubC2::tmp1() const {
440 return _tmp1;
441 }
442
443 Register G1PreBarrierStubC2::tmp2() const {
444 return _tmp2;
445 }
446
447 void G1PreBarrierStubC2::emit_code(MacroAssembler& masm) {
448 G1BarrierSetAssembler* bs = static_cast<G1BarrierSetAssembler*>(BarrierSet::barrier_set()->barrier_set_assembler());
449 bs->generate_c2_pre_barrier_stub(&masm, this);
450 }
451
452 G1PostBarrierStubC2::G1PostBarrierStubC2(const MachNode* node) : G1BarrierStubC2(node) {}
453
454 bool G1PostBarrierStubC2::needs_barrier(const MachNode* node) {
455 return (node->barrier_data() & G1C2BarrierPost) != 0;
456 }
457
458 G1PostBarrierStubC2* G1PostBarrierStubC2::create(const MachNode* node) {
459 G1PostBarrierStubC2* const stub = new (Compile::current()->comp_arena()) G1PostBarrierStubC2(node);
460 if (!Compile::current()->output()->in_scratch_emit_size()) {
461 barrier_set_state()->stubs()->append(stub);
462 }
463 return stub;
464 }
465
466 void G1PostBarrierStubC2::initialize_registers(Register thread, Register tmp1, Register tmp2, Register tmp3) {
467 _thread = thread;
468 _tmp1 = tmp1;
469 _tmp2 = tmp2;
470 _tmp3 = tmp3;
471 }
472
473 Register G1PostBarrierStubC2::thread() const {
474 return _thread;
475 }
476
477 Register G1PostBarrierStubC2::tmp1() const {
478 return _tmp1;
479 }
480
481 Register G1PostBarrierStubC2::tmp2() const {
482 return _tmp2;
483 }
484
485 Register G1PostBarrierStubC2::tmp3() const {
486 return _tmp3;
487 }
488
489 void G1PostBarrierStubC2::emit_code(MacroAssembler& masm) {
490 G1BarrierSetAssembler* bs = static_cast<G1BarrierSetAssembler*>(BarrierSet::barrier_set()->barrier_set_assembler());
491 bs->generate_c2_post_barrier_stub(&masm, this);
492 }
493
494 void* G1BarrierSetC2::create_barrier_state(Arena* comp_arena) const {
495 return new (comp_arena) G1BarrierSetC2State(comp_arena);
496 }
497
498 int G1BarrierSetC2::get_store_barrier(C2Access& access) const {
499 if (!access.is_parse_access()) {
500 // Only support for eliding barriers at parse time for now.
501 return G1C2BarrierPre | G1C2BarrierPost;
502 }
503 GraphKit* kit = (static_cast<C2ParseAccess&>(access)).kit();
504 Node* ctl = kit->control();
505 Node* adr = access.addr().node();
506 uint adr_idx = kit->C->get_alias_index(access.addr().type());
507 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory");
508
509 bool can_remove_pre_barrier = g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, access.type(), adr_idx);
510
511 // We can skip marks on a freshly-allocated object in Eden. Keep this code in
512 // sync with CardTableBarrierSet::on_slowpath_allocation_exit. That routine
513 // informs GC to take appropriate compensating steps, upon a slow-path
514 // allocation, so as to make this card-mark elision safe.
515 // The post-barrier can also be removed if null is written. This case is
516 // handled by G1BarrierSetC2::expand_barriers, which runs at the end of C2's
517 // platform-independent optimizations to exploit stronger type information.
518 bool can_remove_post_barrier = use_ReduceInitialCardMarks() &&
519 ((access.base() == kit->just_allocated_object(ctl)) ||
520 g1_can_remove_post_barrier(kit, &kit->gvn(), ctl, adr));
521
522 int barriers = 0;
523 if (!can_remove_pre_barrier) {
524 barriers |= G1C2BarrierPre;
525 }
526 if (!can_remove_post_barrier) {
527 barriers |= G1C2BarrierPost;
528 }
529
530 return barriers;
531 }
532
533 void G1BarrierSetC2::late_barrier_analysis() const {
534 compute_liveness_at_stubs();
535 }
536
537 void G1BarrierSetC2::emit_stubs(CodeBuffer& cb) const {
538 MacroAssembler masm(&cb);
539 GrowableArray<G1BarrierStubC2*>* const stubs = barrier_set_state()->stubs();
540 for (int i = 0; i < stubs->length(); i++) {
541 // Make sure there is enough space in the code buffer
542 if (cb.insts()->maybe_expand_to_ensure_remaining(PhaseOutput::MAX_inst_size) && cb.blob() == nullptr) {
543 ciEnv::current()->record_failure("CodeCache is full");
544 return;
545 }
546 stubs->at(i)->emit_code(masm);
547 }
548 masm.flush();
549 }
550
551 #ifndef PRODUCT
552 void G1BarrierSetC2::dump_barrier_data(const MachNode* mach, outputStream* st) const {
553 if ((mach->barrier_data() & G1C2BarrierPre) != 0) {
554 st->print("pre ");
555 }
556 if ((mach->barrier_data() & G1C2BarrierPost) != 0) {
557 st->print("post ");
558 }
559 if ((mach->barrier_data() & G1C2BarrierPostNotNull) != 0) {
560 st->print("notnull ");
561 }
562 }
563 #endif // !PRODUCT