1 /*
2 * Copyright (c) 2018, 2026, 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 "ci/ciUtilities.hpp"
26 #include "gc/shared/c2/cardTableBarrierSetC2.hpp"
27 #include "gc/shared/cardTable.hpp"
28 #include "gc/shared/cardTableBarrierSet.hpp"
29 #include "gc/shared/gc_globals.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/graphKit.hpp"
32 #include "opto/idealKit.hpp"
33 #include "opto/macro.hpp"
34 #include "utilities/macros.hpp"
35
36 #define __ ideal.
37
38 Node* CardTableBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
39 DecoratorSet decorators = access.decorators();
40
41 Node* adr = access.addr().node();
42
43 bool is_array = (decorators & IS_ARRAY) != 0;
44 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
45 bool in_heap = (decorators & IN_HEAP) != 0;
46 bool use_precise = is_array || anonymous;
47 bool tightly_coupled_alloc = (decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0;
48
49 const InlineTypeNode* vt = nullptr;
50 if (access.is_parse_access() && static_cast<C2ParseAccess&>(access).vt() != nullptr) {
51 vt = static_cast<C2ParseAccess&>(access).vt();
52 }
53
54 if (vt == nullptr && (!access.is_oop() || tightly_coupled_alloc || (!in_heap && !anonymous))) {
55 return BarrierSetC2::store_at_resolved(access, val);
56 }
57
58 assert(access.is_parse_access(), "entry not supported at optimization time");
59 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
60
61 Node* store = BarrierSetC2::store_at_resolved(access, val);
62 // TODO 8350865
63 // - We actually only need the post barrier once for non-arrays (same for C1, right)?
64 // - Value is only needed to determine if we are storing null. Maybe we can go with a simple boolean?
65 GraphKit* kit = parse_access.kit();
66 if (vt != nullptr) {
67 for (uint i = 0; i < vt->field_count(); ++i) {
68 ciField* field = vt->field(i);
69 ciType* type = field->type();
70 if (!type->is_primitive_type()) {
71 ciInlineKlass* vk = vt->bottom_type()->inline_klass();
72 int field_offset = field->offset_in_bytes() - vk->payload_offset();
73 Node* value = vt->field_value(i);
74 Node* field_adr = kit->basic_plus_adr(access.base(), adr, field_offset);
75 post_barrier(kit, access.base(), field_adr, value, use_precise);
76 }
77 }
78 } else {
79 post_barrier(kit, access.base(), adr, val.node(), use_precise);
80 }
81
82 return store;
83 }
84
85 Node* CardTableBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
86 Node* new_val, const Type* value_type) const {
87 if (!access.is_oop()) {
88 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
89 }
90
91 Node* result = BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
92
93 post_barrier(access.kit(), access.base(), access.addr().node(), new_val, true);
94
95 return result;
96 }
97
98 Node* CardTableBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
99 Node* new_val, const Type* value_type) const {
100 GraphKit* kit = access.kit();
101
102 if (!access.is_oop()) {
103 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
104 }
105
106 Node* load_store = BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
107
108 // Emit the post barrier only when the actual store happened. This makes sense
109 // to check only for LS_cmp_* that can fail to set the value.
110 // LS_cmp_exchange does not produce any branches by default, so there is no
111 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
112 // CompareAndExchange and move branches here, it would make sense to conditionalize
113 // post_barriers for LS_cmp_exchange as well.
114 //
115 // CAS success path is marked more likely since we anticipate this is a performance
116 // critical path, while CAS failure path can use the penalty for going through unlikely
117 // path as backoff. Which is still better than doing a store barrier there.
118 IdealKit ideal(kit);
119 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
120 kit->sync_kit(ideal);
121 post_barrier(kit, access.base(), access.addr().node(), new_val, true);
122 ideal.sync_kit(kit);
123 } ideal.end_if();
124 kit->final_sync(ideal);
125
126 return load_store;
127 }
128
129 Node* CardTableBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const {
130 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, new_val, value_type);
131 if (!access.is_oop()) {
132 return result;
133 }
134
135 post_barrier(access.kit(), access.base(), access.addr().node(), new_val, true);
136
137 return result;
138 }
139
140 Node* CardTableBarrierSetC2::byte_map_base_node(GraphKit* kit) const {
141 // Get base of card map
142 CardTable::CardValue* card_table_base = ci_card_table_address();
143 if (card_table_base != nullptr) {
144 return kit->makecon(TypeRawPtr::make((address)card_table_base));
145 } else {
146 return kit->null();
147 }
148 }
149
150 // vanilla post barrier
151 // Insert a write-barrier store. This is to let generational GC work; we have
152 // to flag all oop-stores before the next GC point.
153 void CardTableBarrierSetC2::post_barrier(GraphKit* kit,
154 Node* obj,
155 Node* adr,
156 Node* val,
157 bool use_precise) const {
158 // No store check needed if we're storing a null.
159 if (val != nullptr && val->is_Con()) {
160 const Type* t = val->bottom_type();
161 if (t == TypePtr::NULL_PTR || t == Type::TOP) {
162 return;
163 }
164 }
165
166 if (use_ReduceInitialCardMarks()
167 && obj == kit->just_allocated_object(kit->control())) {
168 // We can skip marks on a freshly-allocated object in Eden.
169 // Keep this code in sync with CardTableBarrierSet::on_slowpath_allocation_exit.
170 // That routine informs GC to take appropriate compensating steps,
171 // upon a slow-path allocation, so as to make this card-mark
172 // elision safe.
173 return;
174 }
175
176 if (!use_precise) {
177 // All card marks for a (non-array) instance are in one place:
178 adr = obj;
179 } else {
180 // Else it's an array (or unknown), and we want more precise card marks.
181 }
182
183 assert(adr != nullptr, "");
184
185 IdealKit ideal(kit, true);
186
187 // Convert the pointer to an int prior to doing math on it
188 Node* cast = __ CastPX(__ ctrl(), adr);
189
190 // Divide by card size
191 Node* card_offset = __ URShiftX(cast, __ ConI(CardTable::card_shift()));
192
193 // Combine card table base and card offset
194 Node* card_adr = __ AddP(__ top(), byte_map_base_node(kit), card_offset);
195
196 // Get the alias_index for raw card-mark memory
197 int adr_type = Compile::AliasIdxRaw;
198
199 // Dirty card value to store
200 Node* dirty = __ ConI(CardTable::dirty_card_val());
201
202 if (UseCondCardMark) {
203 // The classic GC reference write barrier is typically implemented
204 // as a store into the global card mark table. Unfortunately
205 // unconditional stores can result in false sharing and excessive
206 // coherence traffic as well as false transactional aborts.
207 // UseCondCardMark enables MP "polite" conditional card mark
208 // stores. In theory we could relax the load from ctrl() to
209 // no_ctrl, but that doesn't buy much latitude.
210 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
211 __ if_then(card_val, BoolTest::ne, dirty);
212 }
213
214 // Smash dirty value into card
215 __ store(__ ctrl(), card_adr, dirty, T_BYTE, adr_type, MemNode::unordered);
216
217 if (UseCondCardMark) {
218 __ end_if();
219 }
220
221 // Final sync IdealKit and GraphKit.
222 kit->final_sync(ideal);
223 }
224
225 bool CardTableBarrierSetC2::use_ReduceInitialCardMarks() {
226 return ReduceInitialCardMarks;
227 }
228
229 void CardTableBarrierSetC2::eliminate_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
230 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
231 for (DUIterator_Last imin, i = node->last_outs(imin); i >= imin; --i) {
232 Node* shift = node->last_out(i);
233 for (DUIterator_Last jmin, j = shift->last_outs(jmin); j >= jmin; --j) {
234 Node* addp = shift->last_out(j);
235 for (DUIterator_Last kmin, k = addp->last_outs(kmin); k >= kmin; --k) {
236 Node* mem = addp->last_out(k);
237 if (UseCondCardMark && mem->is_Load()) {
238 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
239 // The load is checking if the card has been written so
240 // replace it with zero to fold the test.
241 igvn->replace_node(mem, igvn->intcon(0));
242 continue;
243 }
244 assert(mem->is_Store(), "store required");
245 igvn->replace_node(mem, mem->in(MemNode::Memory));
246 }
247 }
248 }
249 }
250
251 bool CardTableBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
252 bool is_oop = type == T_OBJECT || type == T_ARRAY;
253 return is_oop && (!tightly_coupled_alloc || !use_ReduceInitialCardMarks());
254 }