1 /*
2 * Copyright (c) 2018, 2025, 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 ciType* type = vt->field_type(i);
69 if (!type->is_primitive_type()) {
70 ciInlineKlass* vk = vt->bottom_type()->inline_klass();
71 int field_offset = vt->field_offset(i) - vk->payload_offset();
72 Node* value = vt->field_value(i);
73 Node* field_adr = kit->basic_plus_adr(access.base(), adr, field_offset);
74 post_barrier(kit, access.base(), field_adr, value, use_precise);
75 }
76 }
77 } else {
78 post_barrier(kit, access.base(), adr, val.node(), use_precise);
79 }
80
81 return store;
82 }
83
84 Node* CardTableBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
85 Node* new_val, const Type* value_type) const {
86 if (!access.is_oop()) {
87 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
88 }
89
90 Node* result = BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
91
92 post_barrier(access.kit(), access.base(), access.addr().node(), new_val, true);
93
94 return result;
95 }
96
97 Node* CardTableBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
98 Node* new_val, const Type* value_type) const {
99 GraphKit* kit = access.kit();
100
101 if (!access.is_oop()) {
102 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
103 }
104
105 Node* load_store = BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
106
107 // Emit the post barrier only when the actual store happened. This makes sense
108 // to check only for LS_cmp_* that can fail to set the value.
109 // LS_cmp_exchange does not produce any branches by default, so there is no
110 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
111 // CompareAndExchange and move branches here, it would make sense to conditionalize
112 // post_barriers for LS_cmp_exchange as well.
113 //
114 // CAS success path is marked more likely since we anticipate this is a performance
115 // critical path, while CAS failure path can use the penalty for going through unlikely
116 // path as backoff. Which is still better than doing a store barrier there.
117 IdealKit ideal(kit);
118 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
119 kit->sync_kit(ideal);
120 post_barrier(kit, access.base(), access.addr().node(), new_val, true);
121 ideal.sync_kit(kit);
122 } ideal.end_if();
123 kit->final_sync(ideal);
124
125 return load_store;
126 }
127
128 Node* CardTableBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const {
129 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, new_val, value_type);
130 if (!access.is_oop()) {
131 return result;
132 }
133
134 post_barrier(access.kit(), access.base(), access.addr().node(), new_val, true);
135
136 return result;
137 }
138
139 Node* CardTableBarrierSetC2::byte_map_base_node(GraphKit* kit) const {
140 // Get base of card map
141 CardTable::CardValue* card_table_base = ci_card_table_address();
142 if (card_table_base != nullptr) {
143 return kit->makecon(TypeRawPtr::make((address)card_table_base));
144 } else {
145 return kit->null();
146 }
147 }
148
149 // vanilla post barrier
150 // Insert a write-barrier store. This is to let generational GC work; we have
151 // to flag all oop-stores before the next GC point.
152 void CardTableBarrierSetC2::post_barrier(GraphKit* kit,
153 Node* obj,
154 Node* adr,
155 Node* val,
156 bool use_precise) const {
157 // No store check needed if we're storing a null.
158 if (val != nullptr && val->is_Con()) {
159 const Type* t = val->bottom_type();
160 if (t == TypePtr::NULL_PTR || t == Type::TOP) {
161 return;
162 }
163 }
164
165 if (use_ReduceInitialCardMarks()
166 && obj == kit->just_allocated_object(kit->control())) {
167 // We can skip marks on a freshly-allocated object in Eden.
168 // Keep this code in sync with CardTableBarrierSet::on_slowpath_allocation_exit.
169 // That routine informs GC to take appropriate compensating steps,
170 // upon a slow-path allocation, so as to make this card-mark
171 // elision safe.
172 return;
173 }
174
175 if (!use_precise) {
176 // All card marks for a (non-array) instance are in one place:
177 adr = obj;
178 } else {
179 // Else it's an array (or unknown), and we want more precise card marks.
180 }
181
182 assert(adr != nullptr, "");
183
184 IdealKit ideal(kit, true);
185
186 // Convert the pointer to an int prior to doing math on it
187 Node* cast = __ CastPX(__ ctrl(), adr);
188
189 // Divide by card size
190 Node* card_offset = __ URShiftX(cast, __ ConI(CardTable::card_shift()));
191
192 // Combine card table base and card offset
193 Node* card_adr = __ AddP(__ top(), byte_map_base_node(kit), card_offset);
194
195 // Get the alias_index for raw card-mark memory
196 int adr_type = Compile::AliasIdxRaw;
197
198 // Dirty card value to store
199 Node* dirty = __ ConI(CardTable::dirty_card_val());
200
201 if (UseCondCardMark) {
202 // The classic GC reference write barrier is typically implemented
203 // as a store into the global card mark table. Unfortunately
204 // unconditional stores can result in false sharing and excessive
205 // coherence traffic as well as false transactional aborts.
206 // UseCondCardMark enables MP "polite" conditional card mark
207 // stores. In theory we could relax the load from ctrl() to
208 // no_ctrl, but that doesn't buy much latitude.
209 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
210 __ if_then(card_val, BoolTest::ne, dirty);
211 }
212
213 // Smash dirty value into card
214 __ store(__ ctrl(), card_adr, dirty, T_BYTE, adr_type, MemNode::unordered);
215
216 if (UseCondCardMark) {
217 __ end_if();
218 }
219
220 // Final sync IdealKit and GraphKit.
221 kit->final_sync(ideal);
222 }
223
224 bool CardTableBarrierSetC2::use_ReduceInitialCardMarks() {
225 return ReduceInitialCardMarks;
226 }
227
228 void CardTableBarrierSetC2::eliminate_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
229 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
230 for (DUIterator_Last imin, i = node->last_outs(imin); i >= imin; --i) {
231 Node* shift = node->last_out(i);
232 for (DUIterator_Last jmin, j = shift->last_outs(jmin); j >= jmin; --j) {
233 Node* addp = shift->last_out(j);
234 for (DUIterator_Last kmin, k = addp->last_outs(kmin); k >= kmin; --k) {
235 Node* mem = addp->last_out(k);
236 if (UseCondCardMark && mem->is_Load()) {
237 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
238 // The load is checking if the card has been written so
239 // replace it with zero to fold the test.
240 igvn->replace_node(mem, igvn->intcon(0));
241 continue;
242 }
243 assert(mem->is_Store(), "store required");
244 igvn->replace_node(mem, mem->in(MemNode::Memory));
245 }
246 }
247 }
248 }
249
250 bool CardTableBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
251 bool is_oop = type == T_OBJECT || type == T_ARRAY;
252 return is_oop && (!tightly_coupled_alloc || !use_ReduceInitialCardMarks());
253 }