1 /*
2 * Copyright (c) 2018, 2026, Red Hat, Inc. All rights reserved.
3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.inline.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
29 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
30 #include "gc/shenandoah/shenandoahForwarding.hpp"
31 #include "gc/shenandoah/shenandoahHeap.hpp"
32 #include "gc/shenandoah/shenandoahRuntime.hpp"
33 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/escape.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/narrowptrnode.hpp"
40 #include "opto/output.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/runtime.hpp"
43
44 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
45 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
46 }
47
48 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena) :
49 BarrierSetC2State(comp_arena),
50 _stubs(new (comp_arena) GrowableArray<ShenandoahBarrierStubC2*>(comp_arena, 8, 0, nullptr)),
51 _trampoline_stubs_count(0),
52 _stubs_start_offset(0),
53 _stubs_current_total_size(0) {
54 }
55
56 static void set_barrier_data(C2Access& access, bool load, bool store) {
57 if (!access.is_oop()) {
58 return;
59 }
60
61 DecoratorSet decorators = access.decorators();
62 bool tightly_coupled = (decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0;
63 bool in_heap = (decorators & IN_HEAP) != 0;
64 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
65 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
66
67 if (tightly_coupled) {
68 access.set_barrier_data(ShenandoahBitElided);
69 return;
70 }
71
72 uint8_t barrier_data = 0;
73
74 if (load) {
75 if (ShenandoahLoadRefBarrier) {
76 if (on_phantom) {
77 barrier_data |= ShenandoahBitPhantom;
78 } else if (on_weak) {
79 barrier_data |= ShenandoahBitWeak;
80 } else {
81 barrier_data |= ShenandoahBitStrong;
82 }
83 }
84 }
85
86 if (store) {
87 if (ShenandoahSATBBarrier) {
88 barrier_data |= ShenandoahBitKeepAlive;
89 }
90 if (ShenandoahCardBarrier && in_heap) {
91 barrier_data |= ShenandoahBitCardMark;
92 }
93 }
94
95 if (!in_heap) {
96 barrier_data |= ShenandoahBitNative;
97 }
98
99 access.set_barrier_data(barrier_data);
100 }
101
102 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
103 // 1: Non-reference load, no additional barrier is needed
104 if (!access.is_oop()) {
105 return BarrierSetC2::load_at_resolved(access, val_type);
106 }
107
108 // 2. Set barrier data for load
109 set_barrier_data(access, /* load = */ true, /* store = */ false);
110
111 // 3. Correction: If we are reading the value of the referent field of
112 // a Reference object, we need to record the referent resurrection.
113 DecoratorSet decorators = access.decorators();
114 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
115 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
116 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
117 bool needs_keepalive = ((on_weak || on_phantom) && !no_keepalive);
118 if (needs_keepalive) {
119 uint8_t barriers = access.barrier_data() | (ShenandoahSATBBarrier ? ShenandoahBitKeepAlive : 0);
120 access.set_barrier_data(barriers);
121 }
122
123 return BarrierSetC2::load_at_resolved(access, val_type);
124 }
125
126 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
127 // 1: Non-reference store, no additional barrier is needed
128 if (!access.is_oop()) {
129 return BarrierSetC2::store_at_resolved(access, val);
130 }
131
132 // 2. Set barrier data for store
133 set_barrier_data(access, /* load = */ false, /* store = */ true);
134
135 // 3. Correction: avoid keep-alive barriers that should not do keep-alive.
136 DecoratorSet decorators = access.decorators();
137 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
138 if (no_keepalive) {
139 access.set_barrier_data(access.barrier_data() & ~ShenandoahBitKeepAlive);
140 }
141
142 return BarrierSetC2::store_at_resolved(access, val);
143 }
144
145 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
146 Node* new_val, const Type* value_type) const {
147 set_barrier_data(access, /* load = */ true, /* store = */ true);
148 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
149 }
150
151 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
152 Node* new_val, const Type* value_type) const {
153 set_barrier_data(access, /* load = */ true, /* store = */ true);
154 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
155 }
156
157 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
158 set_barrier_data(access, /* load = */ true, /* store = */ true);
159 return BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
160 }
161
162 bool ShenandoahBarrierSetC2::is_Load(int opcode) {
163 switch (opcode) {
164 case Op_LoadN:
165 case Op_LoadP:
166 return true;
167 default:
168 return false;
169 }
170 }
171
172 bool ShenandoahBarrierSetC2::is_Store(int opcode) {
173 switch (opcode) {
174 case Op_StoreN:
175 case Op_StoreP:
176 return true;
177 default:
178 return false;
179 }
180 }
181
182 bool ShenandoahBarrierSetC2::is_LoadStore(int opcode) {
183 switch (opcode) {
184 case Op_CompareAndExchangeN:
185 case Op_CompareAndExchangeP:
186 case Op_WeakCompareAndSwapN:
187 case Op_WeakCompareAndSwapP:
188 case Op_CompareAndSwapN:
189 case Op_CompareAndSwapP:
190 case Op_GetAndSetP:
191 case Op_GetAndSetN:
192 return true;
193 default:
194 return false;
195 }
196 }
197
198 bool ShenandoahBarrierSetC2::can_remove_load_barrier(Node* root) {
199 // Check if all outs feed into nodes that do not expose the oops to the rest
200 // of the runtime system. In this case, we can elide the LRB barrier. We bail
201 // out with false at the first sight of trouble.
202
203 ResourceMark rm;
204 VectorSet visited;
205 Node_List worklist;
206 worklist.push(root);
207
208 while (worklist.size() > 0) {
209 Node* n = worklist.pop();
210 if (visited.test_set(n->_idx)) {
211 continue;
212 }
213
214 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
215 Node* out = n->fast_out(i);
216 switch (out->Opcode()) {
217 case Op_Phi:
218 case Op_EncodeP:
219 case Op_DecodeN:
220 case Op_CastPP:
221 case Op_CheckCastPP:
222 case Op_AddP: {
223 // Transitive node, check if any other outs are doing anything troublesome.
224 worklist.push(out);
225 break;
226 }
227
228 case Op_LoadRange: {
229 // Array length is the same in all copies.
230 break;
231 }
232
233 case Op_LoadKlass: {
234 // Klass is the same in all copies.
235 // We would have liked to assert -UCOH, but there are legitimate klass
236 // loads from native Klass* instances, which are also safe under +UCOH.
237 break;
238 }
239
240 case Op_LoadNKlass: {
241 // Similar to above, but LoadNKlass is only safe without +UCOH.
242 // With +UCOH, it loads from mark word, which clashes with forwarding pointers.
243 if (!UseCompactObjectHeaders) {
244 break;
245 }
246 return false;
247 }
248
249 case Op_CmpN: {
250 if (out->in(1) == n &&
251 out->in(2)->Opcode() == Op_ConN &&
252 out->in(2)->get_narrowcon() == 0) {
253 // Null check, no oop is exposed.
254 break;
255 }
256 if (out->in(2) == n &&
257 out->in(1)->Opcode() == Op_ConN &&
258 out->in(1)->get_narrowcon() == 0) {
259 // Null check, no oop is exposed.
260 break;
261 }
262 return false;
263 }
264
265 case Op_CmpP: {
266 if (out->in(1) == n &&
267 out->in(2)->Opcode() == Op_ConP &&
268 out->in(2)->get_ptr() == 0) {
269 // Null check, no oop is exposed.
270 break;
271 }
272 if (out->in(2) == n &&
273 out->in(1)->Opcode() == Op_ConP &&
274 out->in(1)->get_ptr() == 0) {
275 // Null check, no oop is exposed.
276 break;
277 }
278 return false;
279 }
280
281 case Op_CallStaticJava: {
282 if (out->as_CallStaticJava()->is_uncommon_trap()) {
283 // Local feeds into uncommon trap. Deopt machinery handles barriers itself.
284 break;
285 }
286 return false;
287 }
288
289 default: {
290 // Paranoidly distrust any other nodes.
291 return false;
292 }
293 }
294 }
295 }
296
297 // Nothing troublesome found.
298 return true;
299 }
300
301 uint8_t ShenandoahBarrierSetC2::refine_load(Node* n, uint8_t bd) {
302 assert(ShenandoahElideIdealBarriers, "Checked by caller");
303 assert(bd != 0, "Checked by caller");
304
305 // Do not touch weak loads at all: they are responsible for shielding from
306 // Reference.referent resurrection.
307 if ((bd & (ShenandoahBitWeak | ShenandoahBitPhantom)) != 0) {
308 return bd;
309 }
310
311 if (((bd & ShenandoahBitStrong) != 0) && can_remove_load_barrier(n)) {
312 bd &= ~ShenandoahBitStrong;
313 }
314
315 return bd;
316 }
317
318 uint8_t ShenandoahBarrierSetC2::refine_store(Node* n, uint8_t bd) {
319 assert(ShenandoahElideIdealBarriers, "Checked by caller");
320 assert(bd != 0, "Checked by caller");
321 assert(n->is_Mem() || n->is_LoadStore(), "Sanity");
322
323 const Node* newval = n->in(MemNode::ValueIn);
324 assert(newval != nullptr, "Should be present");
325
326 // Type system tells us something about nullity?
327 const Type* newval_bottom = newval->bottom_type();
328 assert(newval_bottom->isa_oopptr() || newval_bottom->isa_narrowoop() ||
329 newval_bottom == TypePtr::NULL_PTR, "Should be an oop store");
330 const TypePtr* newval_type = newval_bottom->make_ptr();
331 assert(newval_type != nullptr, "Should have been filtered before");
332 TypePtr::PTR newval_type_ptr = newval_type->ptr();
333 if (newval_type_ptr == TypePtr::Null) {
334 bd &= ~ShenandoahBitNotNull;
335 // Card table barrier is not needed if we store null.
336 bd &= ~ShenandoahBitCardMark;
337 } else if (newval_type_ptr == TypePtr::NotNull) {
338 // Definitely not null.
339 bd |= ShenandoahBitNotNull;
340 }
341
342 return bd;
343 }
344
345 void ShenandoahBarrierSetC2::final_refinement(Compile* compile) const {
346 ResourceMark rm;
347 Unique_Node_List wq;
348
349 RootNode* root = compile->root();
350 wq.push(root);
351
352 // Also seed the outs to capture nodes are not reachable from in()-s, e.g. endless loops.
353 for (DUIterator_Fast imax, i = root->fast_outs(imax); i < imax; i++) {
354 Node* m = root->fast_out(i);
355 wq.push(m);
356 }
357
358 for (uint next = 0; next < wq.size(); next++) {
359 Node* n = wq.at(next);
360
361 assert(!n->is_Mach(), "No Mach nodes here yet");
362
363 int opc = n->Opcode();
364 bool is_load = is_Load(opc);
365 bool is_store = is_Store(opc);
366 bool is_load_store = is_LoadStore(opc);
367
368 uint8_t orig_bd = 0;
369 if (is_load_store) {
370 orig_bd = n->as_LoadStore()->barrier_data();
371 } else if (is_load || is_store) {
372 orig_bd = n->as_Mem()->barrier_data();
373 }
374
375 uint8_t bd = orig_bd;
376 if (ShenandoahElideIdealBarriers && bd != 0) {
377 // Note: we cannot apply load optimizations to LoadStores,
378 // because their load barriers are needed for fixups.
379 if (is_load) {
380 bd = refine_load(n, bd);
381 }
382 if (is_store || is_load_store) {
383 bd = refine_store(n, bd);
384 }
385 }
386
387 // If there are no real barrier flags on the node, strip away additional fluff.
388 // Matcher does not care about this, and we would like to avoid invoking "barrier_data() != 0"
389 // rules when the only flags are the irrelevant fluff.
390 if ((bd != 0) && (bd & ShenandoahBitsReal) == 0) {
391 bd = 0;
392 }
393
394 if (bd != orig_bd) {
395 if (is_load_store) {
396 n->as_LoadStore()->set_barrier_data(bd);
397 } else {
398 n->as_Mem()->set_barrier_data(bd);
399 }
400 }
401
402 for (uint j = 0; j < n->req(); j++) {
403 Node* in = n->in(j);
404 if (in != nullptr) {
405 wq.push(in);
406 }
407 }
408 }
409 }
410
411 // Support for macro expanded GC barriers
412 void ShenandoahBarrierSetC2::eliminate_gc_barrier_data(Node* node) const {
413 if (node->is_LoadStore()) {
414 LoadStoreNode* loadstore = node->as_LoadStore();
415 loadstore->set_barrier_data(0);
416 } else if (node->is_Mem()) {
417 MemNode* mem = node->as_Mem();
418 mem->set_barrier_data(0);
419 }
420 }
421
422 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
423 eliminate_gc_barrier_data(node);
424 }
425
426 void ShenandoahBarrierSetC2::elide_dominated_barrier(MachNode* node, MachNode* dominator) const {
427 uint8_t orig_bd = node->barrier_data();
428 if (orig_bd == 0) {
429 // Nothing to do.
430 return;
431 }
432
433 uint8_t bd = orig_bd;
434 int node_opcode = node->ideal_Opcode();
435
436 if (dominator == nullptr) {
437 // Must be allocation node.
438 if (is_Load(node_opcode) || is_LoadStore(node_opcode)) {
439 // Loads from recent allocations do not need LRBs.
440 bd &= ~ShenandoahBitStrong;
441 }
442 if (is_Store(node_opcode) || is_LoadStore(node_opcode)) {
443 // Stores to recent allocations do not need KA or CM.
444 bd &= ~ShenandoahBitKeepAlive;
445 bd &= ~ShenandoahBitCardMark;
446 }
447 } else {
448 // LoadStores do not get these optimizations, since their LRBs
449 // are required for fixups.
450 if (is_Load(node_opcode) || is_Store(node_opcode)) {
451 int dom_opcode = dominator->ideal_Opcode();
452 uint8_t dom_bd = dominator->barrier_data();
453
454 if (is_Load(dom_opcode) || is_LoadStore(dom_opcode)) {
455 // If dominating load is set up to perform LRB fixups, no further LRB is needed.
456 if ((dom_bd & ShenandoahBitStrong) != 0) {
457 bd &= ~ShenandoahBitStrong;
458 }
459 }
460 if (is_Store(dom_opcode)) {
461 // Dominating store has stored the good ref, no LRB is needed.
462 bd &= ~ShenandoahBitStrong;
463 }
464 }
465 }
466
467 if (orig_bd != bd) {
468 // We are already in final output.
469 // Strip the extra barrier data if no real bits are left.
470 if ((bd & ShenandoahBitsReal) != 0) {
471 node->set_barrier_data(bd);
472 } else {
473 node->set_barrier_data(0);
474 }
475 }
476 }
477
478 void ShenandoahBarrierSetC2::analyze_dominating_barriers() const {
479 if (!ShenandoahElideMachBarriers) {
480 return;
481 }
482
483 ResourceMark rm;
484 Node_List accesses, dominators;
485
486 PhaseCFG* const cfg = Compile::current()->cfg();
487 for (uint i = 0; i < cfg->number_of_blocks(); ++i) {
488 const Block* const block = cfg->get_block(i);
489 for (uint j = 0; j < block->number_of_nodes(); ++j) {
490 Node* const node = block->get_node(j);
491
492 // Everything that happens in allocations does not need barriers.
493 // Record them for dominance analysis.
494 if (node->is_Phi() && is_allocation(node)) {
495 dominators.push(node);
496 continue;
497 }
498
499 if (!node->is_Mach()) {
500 continue;
501 }
502
503 MachNode* const mach = node->as_Mach();
504 int opcode = mach->ideal_Opcode();
505 if (is_Load(opcode) || is_Store(opcode) || is_LoadStore(opcode)) {
506 if ((mach->barrier_data() & ShenandoahBitsReal) != 0) {
507 accesses.push(mach);
508 dominators.push(mach);
509 }
510 }
511 }
512 }
513
514 elide_dominated_barriers(accesses, dominators);
515 }
516
517 uint ShenandoahBarrierSetC2::estimated_barrier_size(const Node* node) const {
518 // Barrier impact on fast-path is driven by GC state checks emitted very late.
519 // These checks are tight load-test-branch sequences, with no impact on C2 graph
520 // size. Limiting unrolling in presence of GC barriers might turn some loops
521 // tighter than with default unrolling, which may benefit performance due to denser
522 // code. Testing shows it is still counter-productive.
523 // Therefore, we report zero barrier size to let C2 do its normal thing.
524 return 0;
525 }
526
527 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
528 bool is_oop = is_reference_type(type);
529 if (!is_oop) {
530 return false;
531 }
532 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
533 if (phase == Optimization) {
534 return false;
535 }
536 return !is_clone;
537 }
538 return true;
539 }
540
541 bool ShenandoahBarrierSetC2::clone_needs_barrier(const TypeOopPtr* src_type, bool& is_oop_array) {
542 if (!ShenandoahCloneBarrier) {
543 return false;
544 }
545
546 if (src_type->isa_instptr() != nullptr) {
547 // Instance: need barrier only if there is a possibility of having an oop anywhere in it.
548 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
549 if ((src_type->klass_is_exact() || !ik->has_subklass()) &&
550 !ik->has_injected_fields() && !ik->has_object_fields()) {
551 if (!src_type->klass_is_exact()) {
552 // Class is *currently* the leaf in the hierarchy.
553 // Record the dependency so that we deopt if this does not hold in future.
554 Compile::current()->dependencies()->assert_leaf_type(ik);
555 }
556 return false;
557 }
558 } else if (src_type->isa_aryptr() != nullptr) {
559 // Array: need barrier only if array is oop-bearing.
560 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
561 if (is_reference_type(src_elem, true)) {
562 is_oop_array = true;
563 } else {
564 return false;
565 }
566 }
567
568 // Assume the worst.
569 return true;
570 }
571
572 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src_base, Node* dst_base, Node* size, bool is_array) const {
573 const TypeOopPtr* src_type = kit->gvn().type(src_base)->is_oopptr();
574
575 bool is_oop_array = false;
576 if (!clone_needs_barrier(src_type, is_oop_array)) {
577 // No barrier is needed? Just do what common BarrierSetC2 wants with it.
578 BarrierSetC2::clone(kit, src_base, dst_base, size, is_array);
579 return;
580 }
581
582 if (ShenandoahCloneRuntime || !is_array || !is_oop_array) {
583 // Looks like an instance? Prepare the instance clone. This would either
584 // be exploded into individual accesses or be left as runtime call.
585 // Common BarrierSetC2 prepares everything for both cases.
586 BarrierSetC2::clone(kit, src_base, dst_base, size, is_array);
587 return;
588 }
589
590 // We are cloning the oop array. Prepare to call the normal arraycopy stub
591 // after the expansion. Normal stub takes the number of actual type-sized
592 // elements to copy after the base, compute the count here.
593 Node* offset = kit->MakeConX(arrayOopDesc::base_offset_in_bytes(UseCompressedOops ? T_NARROWOOP : T_OBJECT));
594 size = kit->gvn().transform(new SubXNode(size, offset));
595 size = kit->gvn().transform(new URShiftXNode(size, kit->intcon(LogBytesPerHeapOop)));
596 ArrayCopyNode* ac = ArrayCopyNode::make(kit, false, src_base, offset, dst_base, offset, size, true, false);
597 ac->set_clone_array();
598 Node* n = kit->gvn().transform(ac);
599 if (n == ac) {
600 ac->set_adr_type(TypeRawPtr::BOTTOM);
601 kit->set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), TypeRawPtr::BOTTOM);
602 } else {
603 kit->set_all_memory(n);
604 }
605 }
606
607 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
608 Node* const ctrl = ac->in(TypeFunc::Control);
609 Node* const mem = ac->in(TypeFunc::Memory);
610 Node* const src = ac->in(ArrayCopyNode::Src);
611 Node* const src_offset = ac->in(ArrayCopyNode::SrcPos);
612 Node* const dest = ac->in(ArrayCopyNode::Dest);
613 Node* const dest_offset = ac->in(ArrayCopyNode::DestPos);
614 Node* length = ac->in(ArrayCopyNode::Length);
615
616 const TypeOopPtr* src_type = phase->igvn().type(src)->is_oopptr();
617
618 bool is_oop_array = false;
619 if (!clone_needs_barrier(src_type, is_oop_array)) {
620 // No barrier is needed? Expand to normal HeapWord-sized arraycopy.
621 BarrierSetC2::clone_at_expansion(phase, ac);
622 return;
623 }
624
625 if (ShenandoahCloneRuntime || !ac->is_clone_array() || !is_oop_array) {
626 // Still looks like an instance? Likely a large instance or reflective
627 // clone with unknown length. Go to runtime and handle it there.
628 clone_in_runtime(phase, ac, ShenandoahRuntime::clone_addr(), "ShenandoahRuntime::clone");
629 return;
630 }
631
632 // We are cloning the oop array. Call into normal oop array copy stubs.
633 // Those stubs would call BarrierSetAssembler to handle GC barriers.
634
635 // This is the full clone, so offsets should equal each other and be at array base.
636 assert(src_offset == dest_offset, "should be equal");
637 const jlong offset = src_offset->get_long();
638 const TypeAryPtr* const ary_ptr = src->get_ptr_type()->isa_aryptr();
639 BasicType bt = ary_ptr->elem()->array_element_basic_type();
640 assert(offset == arrayOopDesc::base_offset_in_bytes(bt), "should match");
641
642 const char* copyfunc_name = "arraycopy";
643 const address copyfunc_addr = phase->basictype2arraycopy(T_OBJECT, nullptr, nullptr, true, copyfunc_name, true);
644
645 Node* const call = phase->make_leaf_call(ctrl, mem,
646 OptoRuntime::fast_arraycopy_Type(),
647 copyfunc_addr, copyfunc_name,
648 TypeRawPtr::BOTTOM,
649 phase->basic_plus_adr(src, src_offset),
650 phase->basic_plus_adr(dest, dest_offset),
651 length,
652 phase->top()
653 );
654 phase->transform_later(call);
655
656 phase->igvn().replace_node(ac, call);
657 }
658
659 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
660 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
661 }
662
663 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
664 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
665 }
666
667 void ShenandoahBarrierSetC2::print_barrier_data(outputStream* os, uint8_t data) {
668 os->print(" Node barriers: ");
669 if ((data & ShenandoahBitStrong) != 0) {
670 data &= ~ShenandoahBitStrong;
671 os->print("strong ");
672 }
673
674 if ((data & ShenandoahBitWeak) != 0) {
675 data &= ~ShenandoahBitWeak;
676 os->print("weak ");
677 }
678
679 if ((data & ShenandoahBitPhantom) != 0) {
680 data &= ~ShenandoahBitPhantom;
681 os->print("phantom ");
682 }
683
684 if ((data & ShenandoahBitKeepAlive) != 0) {
685 data &= ~ShenandoahBitKeepAlive;
686 os->print("keepalive ");
687 }
688
689 if ((data & ShenandoahBitCardMark) != 0) {
690 data &= ~ShenandoahBitCardMark;
691 os->print("cardmark ");
692 }
693
694 if ((data & ShenandoahBitNative) != 0) {
695 data &= ~ShenandoahBitNative;
696 os->print("native ");
697 }
698
699 if ((data & ShenandoahBitNotNull) != 0) {
700 data &= ~ShenandoahBitNotNull;
701 os->print("not-null ");
702 }
703
704 if ((data & ShenandoahBitElided) != 0) {
705 data &= ~ShenandoahBitElided;
706 os->print("elided ");
707 }
708
709 os->cr();
710
711 if (data > 0) {
712 fatal("Unknown bit!");
713 }
714
715 os->print_cr(" GC configuration: %sLRB %sSATB %sCAS %sClone %sCard",
716 (ShenandoahLoadRefBarrier ? "+" : "-"),
717 (ShenandoahSATBBarrier ? "+" : "-"),
718 (ShenandoahCASBarrier ? "+" : "-"),
719 (ShenandoahCloneBarrier ? "+" : "-"),
720 (ShenandoahCardBarrier ? "+" : "-")
721 );
722 }
723
724
725 #ifdef ASSERT
726 void ShenandoahBarrierSetC2::verify_gc_barrier_assert(bool cond, const char* msg, uint8_t bd, Node* n) {
727 if (!cond) {
728 stringStream ss;
729 ss.print_cr("%s", msg);
730 ss.print_cr("-----------------");
731 print_barrier_data(&ss, bd);
732 ss.print_cr("-----------------");
733 n->dump_bfs(1, nullptr, "", &ss);
734 report_vm_error(__FILE__, __LINE__, ss.as_string());
735 }
736 }
737
738 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
739 if (!ShenandoahVerifyOptoBarriers) {
740 return;
741 }
742
743 // Verify depending on the barriers actually enabled, allowing verification in passive mode.
744 // Normally, we have _some_ bits set on all accesses. Optimizations may drop some bits,
745 // but only the last optimization step eliminates all remaining metadata flags. Only then
746 // the access data can be completely blank.
747 bool final_phase = (phase == BeforeCodeGen);
748 bool expect_load_barriers = !final_phase && ShenandoahLoadRefBarrier;
749 bool expect_store_barriers = !final_phase && (ShenandoahSATBBarrier || ShenandoahCardBarrier);
750 bool expect_load_store_barriers = expect_load_barriers || expect_store_barriers;
751 bool expect_some_real = final_phase;
752
753 Unique_Node_List wq;
754
755 RootNode* root = compile->root();
756 wq.push(root);
757
758 // Also seed the outs to capture nodes are not reachable from in()-s, e.g. endless loops.
759 for (DUIterator_Fast imax, i = root->fast_outs(imax); i < imax; i++) {
760 Node* m = root->fast_out(i);
761 wq.push(m);
762 }
763
764 for (uint next = 0; next < wq.size(); next++) {
765 Node *n = wq.at(next);
766 assert(!n->is_Mach(), "No Mach nodes here yet");
767
768 int opc = n->Opcode();
769
770 uint8_t bd = 0;
771 const TypePtr* adr_type = nullptr;
772 if (is_Load(opc)) {
773 bd = n->as_Load()->barrier_data();
774 adr_type = n->as_Load()->adr_type();
775 } else if (is_Store(opc)) {
776 bd = n->as_Store()->barrier_data();
777 adr_type = n->as_Store()->adr_type();
778 } else if (is_LoadStore(opc)) {
779 bd = n->as_LoadStore()->barrier_data();
780 adr_type = n->as_LoadStore()->adr_type();
781 } else if (n->is_Mem()) {
782 bd = MemNode::barrier_data(n);
783 verify_gc_barrier_assert(bd == 0, "Other mem nodes should have no barrier data", bd, n);
784 }
785
786 bool is_weak = (bd & (ShenandoahBitWeak | ShenandoahBitPhantom)) != 0;
787 bool is_native = (bd & ShenandoahBitNative) != 0;
788
789 bool is_referent = adr_type != nullptr &&
790 adr_type->isa_instptr() &&
791 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
792 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset();
793
794 bool is_oop_addr = (adr_type != nullptr) && (adr_type->isa_oopptr() || adr_type->isa_narrowoop());
795 bool is_raw_addr = (adr_type != nullptr) && (adr_type->isa_rawptr() || adr_type->isa_klassptr());
796
797 verify_gc_barrier_assert(!expect_some_real || (bd == 0) || (bd & ShenandoahBitsReal) != 0, "Without real barriers, metadata should be stripped at this point", bd, n);
798
799 if (is_oop_addr) {
800 if (is_Load(opc)) {
801 verify_gc_barrier_assert(!expect_load_barriers || (bd != 0), "Oop load should have barrier data", bd, n);
802 verify_gc_barrier_assert(!is_weak || is_referent, "Weak load only for Reference.referent", bd, n);
803 } else if (is_Store(opc)) {
804 // Reference.referent stores can be without barriers.
805 verify_gc_barrier_assert(!expect_store_barriers || is_referent || (bd != 0), "Oop store should have barrier data", bd, n);
806 } else if (is_LoadStore(opc)) {
807 verify_gc_barrier_assert(!expect_load_store_barriers || (bd != 0), "Oop load-store should have barrier data", bd, n);
808 }
809 } else if (is_raw_addr) {
810 if (is_native) {
811 if (is_Load(opc)) {
812 verify_gc_barrier_assert(!expect_load_barriers || (bd != 0), "Native oop load should have barrier data", bd, n);
813 }
814 if (is_Store(opc)) {
815 verify_gc_barrier_assert(!expect_store_barriers || (bd != 0), "Native oop store should have barrier data", bd, n);
816 }
817 if (is_LoadStore(opc)) {
818 verify_gc_barrier_assert(!expect_load_store_barriers || (bd != 0), "Native oop load-store should have barrier data", bd, n);
819 }
820 } else {
821 // Some Load/Stores are used for T_ADDRESS and/or raw stores, which are supposed not to have barriers.
822 // Some other Load/Stores are emitted for real oops, but on raw addresses via Unsafe.
823 // The distinction on this level is lost, so we cannot really verify this.
824 }
825 } else {
826 if (is_Load(opc) || is_Store(opc) || is_LoadStore(opc)) {
827 verify_gc_barrier_assert(false, "Unclassified access type", bd, n);
828 }
829 }
830
831 for (uint j = 0; j < n->req(); j++) {
832 Node* in = n->in(j);
833 if (in != nullptr) {
834 wq.push(in);
835 }
836 }
837 }
838 }
839 #endif
840
841 static ShenandoahBarrierSetC2State* barrier_set_state() {
842 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
843 }
844
845 int ShenandoahBarrierSetC2::estimate_stub_size() const {
846 GrowableArray<ShenandoahBarrierStubC2*>* const stubs = barrier_set_state()->stubs();
847 assert(stubs->is_empty(), "Lifecycle: no stubs were yet created");
848 return 0;
849 }
850
851 void ShenandoahBarrierSetC2::emit_stubs(CodeBuffer& cb) const {
852 MacroAssembler masm(&cb);
853
854 PhaseOutput* const output = Compile::current()->output();
855 assert(masm.offset() <= output->buffer_sizing_data()->_code,
856 "Stubs are assumed to be emitted directly after code and code_size is a hard limit on where it can start");
857 barrier_set_state()->set_stubs_start_offset(masm.offset());
858
859 // Stub generation counts all stubs as skipped for the sake of inlining policy.
860 // This is critical for performance, check it.
861 #ifdef ASSERT
862 int offset_before = masm.offset();
863 int skipped_before = cb.total_skipped_instructions_size();
864 #endif
865
866 GrowableArray<ShenandoahBarrierStubC2*>* const stubs = barrier_set_state()->stubs();
867 for (int i = 0; i < stubs->length(); i++) {
868 // Make sure there is enough space in the code buffer
869 if (cb.insts()->maybe_expand_to_ensure_remaining(PhaseOutput::MAX_inst_size) && cb.blob() == nullptr) {
870 ciEnv::current()->record_failure("CodeCache is full");
871 return;
872 }
873 stubs->at(i)->emit_code(masm);
874 }
875
876 #ifdef ASSERT
877 int offset_after = masm.offset();
878 int skipped_after = cb.total_skipped_instructions_size();
879 assert(offset_after - offset_before == skipped_after - skipped_before,
880 "All stubs are counted as skipped. masm: %d - %d = %d, cb: %d - %d = %d",
881 offset_after, offset_before, offset_after - offset_before,
882 skipped_after, skipped_before, skipped_after - skipped_before);
883 #endif
884
885 masm.flush();
886 }
887
888 void ShenandoahBarrierStubC2::register_stub(ShenandoahBarrierStubC2* stub) {
889 if (!Compile::current()->output()->in_scratch_emit_size()) {
890 barrier_set_state()->stubs()->append(stub);
891 }
892 }
893
894 ShenandoahBarrierStubC2* ShenandoahBarrierStubC2::create(const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow, bool do_load) {
895 auto* stub = new (Compile::current()->comp_arena()) ShenandoahBarrierStubC2(node, obj, addr, tmp1, tmp2, narrow, do_load);
896 register_stub(stub);
897 return stub;
898 }
899
900 void ShenandoahBarrierStubC2::load_post(MacroAssembler* masm, const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow) {
901 // Load post-barrier:
902 // a. Satisfies the need for LRB for normal loads
903 // b. Passes a weak load through LRB-weak
904 // c. Keep-alives a weak load
905 if (needs_slow_barrier(node)) {
906 ShenandoahBarrierStubC2* const stub = create(node, obj, addr, tmp1, tmp2, narrow, /* do_load = */ false);
907 char check = 0;
908 check |= needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
909 check |= needs_load_ref_barrier(node) ? ShenandoahHeap::HAS_FORWARDED : 0;
910 check |= needs_load_ref_barrier_weak(node) ? ShenandoahHeap::WEAK_ROOTS : 0;
911 stub->enter_if_gc_state(*masm, check, tmp1);
912 }
913 }
914
915 void ShenandoahBarrierStubC2::store_pre(MacroAssembler* masm, const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow) {
916 // Store pre-barrier: SATB, keep-alive the current memory value.
917 if (needs_slow_barrier(node)) {
918 assert(!needs_load_ref_barrier(node), "Should not be required for stores");
919 ShenandoahBarrierStubC2* const stub = create(node, obj, addr, tmp1, tmp2, narrow, /* do_load = */ true);
920 stub->enter_if_gc_state(*masm, ShenandoahHeap::MARKING, tmp1);
921 }
922 }
923
924 void ShenandoahBarrierStubC2::load_store_pre(MacroAssembler* masm, const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow) {
925 // Load/Store pre-barrier:
926 // a. Avoids false positives from CAS encountering to-space memory values.
927 // b. Satisfies the need for LRB for the CAE result.
928 // c. Records old value for the sake of SATB.
929 //
930 // (a) and (b) are covered because load barrier does memory location fixup.
931 // (c) is covered by KA on the current memory value.
932 if (needs_slow_barrier(node)) {
933 ShenandoahBarrierStubC2* const stub = create(node, obj, addr, tmp1, tmp2, narrow, /* do_load = */ true);
934 char check = 0;
935 check |= needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
936 check |= needs_load_ref_barrier(node) ? ShenandoahHeap::HAS_FORWARDED : 0;
937 assert(!needs_load_ref_barrier_weak(node), "Not supported for Load/Stores");
938 stub->enter_if_gc_state(*masm, check, tmp1);
939 }
940 }
941
942 void ShenandoahBarrierStubC2::store_post(MacroAssembler* masm, const MachNode* node, Address addr, Register tmp1, Register tmp2) {
943 if (needs_card_barrier(node)) {
944 cardtable(*masm, addr, tmp1, tmp2);
945 }
946 }
947
948 void ShenandoahBarrierStubC2::load_store_post(MacroAssembler* masm, const MachNode* node, Address addr, Register tmp1, Register tmp2) {
949 store_post(masm, node, addr, tmp1, tmp2);
950 }
951
952 bool ShenandoahBarrierStubC2::is_live_register(Register reg) {
953 return preserve_set().member(OptoReg::as_OptoReg(reg->as_VMReg()));
954 }
955
956 Register ShenandoahBarrierStubC2::select_temp_register(bool& selected_live, Register skip_reg1, Register skip_reg2) {
957 Register tmp = noreg;
958 Register fallback_live = noreg;
959
960 // Try to select non-live first:
961 for (int i = 0; i < available_gp_registers(); i++) {
962 Register r = as_Register(i);
963 if (r != _obj && r != _addr.base() && r != _addr.index() &&
964 r != skip_reg1 && r != skip_reg2 && !is_special_register(r)) {
965 if (!is_live_register(r)) {
966 tmp = r;
967 break;
968 } else if (fallback_live == noreg) {
969 fallback_live = r;
970 }
971 }
972 }
973
974 // If we could not find a non-live register, select the live fallback:
975 if (tmp == noreg) {
976 tmp = fallback_live;
977 selected_live = true;
978 } else {
979 selected_live = false;
980 }
981
982 assert(tmp != noreg, "successfully selected");
983 assert_different_registers(tmp, skip_reg1);
984 assert_different_registers(tmp, skip_reg2);
985 assert_different_registers(tmp, _obj);
986 assert_different_registers(tmp, _addr.base());
987 assert_different_registers(tmp, _addr.index());
988 return tmp;
989 }
990
991 address ShenandoahBarrierStubC2::keepalive_runtime_entry_addr() {
992 if (_narrow) {
993 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre_narrow);
994 } else {
995 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre);
996 }
997 }
998
999 address ShenandoahBarrierStubC2::lrb_runtime_entry_addr() {
1000 bool is_strong = (_node->barrier_data() & ShenandoahBitStrong) != 0;
1001 bool is_weak = (_node->barrier_data() & ShenandoahBitWeak) != 0;
1002 bool is_phantom = (_node->barrier_data() & ShenandoahBitPhantom) != 0;
1003
1004 if (_narrow) {
1005 if (is_strong) {
1006 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow_narrow);
1007 } else if (is_weak) {
1008 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow_narrow);
1009 } else if (is_phantom) {
1010 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow_narrow);
1011 }
1012 } else {
1013 if (is_strong) {
1014 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
1015 } else if (is_weak) {
1016 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
1017 } else if (is_phantom) {
1018 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1019 }
1020 }
1021
1022 ShouldNotReachHere();
1023 return nullptr;
1024 }
1025
1026 bool ShenandoahBarrierSetC2State::needs_liveness_data(const MachNode* mach) const {
1027 // Nodes that require slow-path stubs need liveness data.
1028 return ShenandoahBarrierStubC2::needs_slow_barrier(mach);
1029 }
1030
1031 bool ShenandoahBarrierSetC2State::needs_livein_data() const {
1032 return true;
1033 }