< prev index next >

src/hotspot/cpu/riscv/gc/g1/g1BarrierSetAssembler_riscv.cpp

Print this page

  1 /*
  2  * Copyright (c) 2018, 2025, Oracle and/or its affiliates. All rights reserved.
  3  * Copyright (c) 2020, 2024, Huawei Technologies Co., Ltd. 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.

120   if (UseCondCardMark) {
121     __ lbu(tmp, Address(start, 0));
122     static_assert((uint)G1CardTable::clean_card_val() == 0xff, "must be");
123     __ subi(tmp, tmp, G1CardTable::clean_card_val()); // Convert to clean_card_value() to a comparison
124                                                       // against zero to avoid use of an extra temp.
125     __ bnez(tmp, next);
126   }
127 
128   static_assert(G1CardTable::dirty_card_val() == 0, "must be to use zr");
129   __ sb(zr, Address(start, 0));
130 
131   __ bind(next);
132   __ addi(start, start, 1);
133   __ ble(start, count, loop);
134 
135   __ bind(done);
136 }
137 
138 static void generate_queue_test_and_insertion(MacroAssembler* masm, ByteSize index_offset, ByteSize buffer_offset, Label& runtime,
139                                               const Register thread, const Register value, const Register tmp1, const Register tmp2) {

140   // Can we store a value in the given thread's buffer?
141   // (The index field is typed as size_t.)
142   __ ld(tmp1, Address(thread, in_bytes(index_offset)));   // tmp1 := *(index address)
143   __ beqz(tmp1, runtime);                                 // jump to runtime if index == 0 (full buffer)
144   // The buffer is not full, store value into it.
145   __ subi(tmp1, tmp1, wordSize);                           // tmp1 := next index
146   __ sd(tmp1, Address(thread, in_bytes(index_offset)));   // *(index address) := next index
147   __ ld(tmp2, Address(thread, in_bytes(buffer_offset)));  // tmp2 := buffer address
148   __ add(tmp2, tmp2, tmp1);
149   __ sd(value, Address(tmp2));                            // *(buffer address + next index) := value
150 }
151 
152 static void generate_pre_barrier_fast_path(MacroAssembler* masm,
153                                            const Register thread,
154                                            const Register tmp1) {
155   Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
156   // Is marking active?
157   if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
158     __ lwu(tmp1, in_progress);
159   } else {
160     assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
161     __ lbu(tmp1, in_progress);
162   }
163 }
164 
165 static void generate_pre_barrier_slow_path(MacroAssembler* masm,
166                                            const Register obj,
167                                            const Register pre_val,
168                                            const Register thread,
169                                            const Register tmp1,
170                                            const Register tmp2,
171                                            Label& done,
172                                            Label& runtime) {
173   // Do we need to load the previous value?
174   if (obj != noreg) {
175     __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
176   }
177   // Is the previous value null?
178   __ beqz(pre_val, done, true);
179   generate_queue_test_and_insertion(masm,
180                                     G1ThreadLocalData::satb_mark_queue_index_offset(),
181                                     G1ThreadLocalData::satb_mark_queue_buffer_offset(),
182                                     runtime,
183                                     thread, pre_val, tmp1, tmp2);
184   __ j(done);
185 }
186 
187 void G1BarrierSetAssembler::g1_write_barrier_pre(MacroAssembler* masm,
188                                                  Register obj,
189                                                  Register pre_val,
190                                                  Register thread,
191                                                  Register tmp1,
192                                                  Register tmp2,
193                                                  bool tosca_live,
194                                                  bool expand_call) {
195   // If expand_call is true then we expand the call_VM_leaf macro
196   // directly to skip generating the check by
197   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
198 
199   assert(thread == xthread, "must be");
200 
201   Label done;
202   Label runtime;
203 
204   assert_different_registers(obj, pre_val, tmp1, tmp2);
205   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
206 
207   generate_pre_barrier_fast_path(masm, thread, tmp1);
208   // If marking is not active (*(mark queue active address) == 0), jump to done
209   __ beqz(tmp1, done);
210   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, done, runtime);
211 
212   __ bind(runtime);
213 
214   __ push_call_clobbered_registers();
215 












216   if (expand_call) {
217     assert(pre_val != c_rarg1, "smashed arg");
218     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
219   } else {
220     __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
221   }
222 
223   __ pop_call_clobbered_registers();
224 
225   __ bind(done);
226 
227 }
228 
229 static void generate_post_barrier(MacroAssembler* masm,
230                                   const Register store_addr,
231                                   const Register new_val,
232                                   const Register thread,
233                                   const Register tmp1,
234                                   const Register tmp2,
235                                   Label& done,

283   __ mv(c_rarg1, xthread);
284   __ mv(t1, runtime_path);
285   __ jalr(t1);
286 }
287 
288 void G1BarrierSetAssembler::g1_write_barrier_pre_c2(MacroAssembler* masm,
289                                                     Register obj,
290                                                     Register pre_val,
291                                                     Register thread,
292                                                     Register tmp1,
293                                                     Register tmp2,
294                                                     G1PreBarrierStubC2* stub) {
295   assert(thread == xthread, "must be");
296   assert_different_registers(obj, pre_val, tmp1, tmp2);
297   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
298 
299   stub->initialize_registers(obj, pre_val, thread, tmp1, tmp2);
300 
301   generate_pre_barrier_fast_path(masm, thread, tmp1);
302   // If marking is active (*(mark queue active address) != 0), jump to stub (slow path)
303   __ bnez(tmp1, *stub->entry(), true);
304 
305   __ bind(*stub->continuation());
306 }
307 
308 void G1BarrierSetAssembler::generate_c2_pre_barrier_stub(MacroAssembler* masm,
309                                                          G1PreBarrierStubC2* stub) const {
310   Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
311   Label runtime;
312   Register obj = stub->obj();
313   Register pre_val = stub->pre_val();
314   Register thread = stub->thread();
315   Register tmp1 = stub->tmp1();
316   Register tmp2 = stub->tmp2();
317 
318   __ bind(*stub->entry());
319   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, *stub->continuation(), runtime);
320 
321   __ bind(runtime);
322   generate_c2_barrier_runtime_call(masm, stub, pre_val, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry));
323   __ j(*stub->continuation());

346   CardTableBarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
347   if (on_oop && on_reference) {
348     // RA is live.  It must be saved around calls.
349     __ enter(); // barrier may call runtime
350     // Generate the G1 pre-barrier code to log the value of
351     // the referent field in an SATB buffer.
352     g1_write_barrier_pre(masm /* masm */,
353                          noreg /* obj */,
354                          dst /* pre_val */,
355                          xthread /* thread */,
356                          tmp1 /* tmp1 */,
357                          tmp2 /* tmp2 */,
358                          true /* tosca_live */,
359                          true /* expand_call */);
360     __ leave();
361   }
362 }
363 
364 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
365                                          Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {










366   // flatten object address if needed
367   if (dst.offset() == 0) {
368     if (dst.base() != tmp3) {
369       __ mv(tmp3, dst.base());
370     }
371   } else {
372     __ la(tmp3, dst);
373   }
374 
375   g1_write_barrier_pre(masm,
376                        tmp3 /* obj */,
377                        tmp2 /* pre_val */,
378                        xthread /* thread */,
379                        tmp1 /* tmp1 */,
380                        t1 /* tmp2 */,
381                        val != noreg /* tosca_live */,
382                        false /* expand_call */);


383 
384   if (val == noreg) {
385     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), noreg, noreg, noreg, noreg);
386   } else {
387     // G1 barrier needs uncompressed oop for region cross check.
388     Register new_val = val;
389     if (UseCompressedOops) {
390       new_val = t1;
391       __ mv(new_val, val);


392     }
393     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
394     g1_write_barrier_post(masm,
395                           tmp3 /* store_adr */,
396                           new_val /* new_val */,
397                           xthread /* thread */,
398                           tmp1 /* tmp1 */,
399                           tmp2 /* tmp2 */);


400   }
401 }
402 
403 #ifdef COMPILER1
404 
405 #undef __
406 #define __ ce->masm()->
407 
408 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
409   G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
410 
411   // At this point we know that marking is in progress.
412   // If do_load() is true then we have to emit the
413   // load of the previous value; otherwise it has already
414   // been loaded into _pre_val.
415   __ bind(*stub->entry());
416 
417   assert(stub->pre_val()->is_register(), "Precondition.");
418 
419   Register pre_val_reg = stub->pre_val()->as_register();

  1 /*
  2  * Copyright (c) 2018, 2026, Oracle and/or its affiliates. All rights reserved.
  3  * Copyright (c) 2020, 2024, Huawei Technologies Co., Ltd. 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.

120   if (UseCondCardMark) {
121     __ lbu(tmp, Address(start, 0));
122     static_assert((uint)G1CardTable::clean_card_val() == 0xff, "must be");
123     __ subi(tmp, tmp, G1CardTable::clean_card_val()); // Convert to clean_card_value() to a comparison
124                                                       // against zero to avoid use of an extra temp.
125     __ bnez(tmp, next);
126   }
127 
128   static_assert(G1CardTable::dirty_card_val() == 0, "must be to use zr");
129   __ sb(zr, Address(start, 0));
130 
131   __ bind(next);
132   __ addi(start, start, 1);
133   __ ble(start, count, loop);
134 
135   __ bind(done);
136 }
137 
138 static void generate_queue_test_and_insertion(MacroAssembler* masm, ByteSize index_offset, ByteSize buffer_offset, Label& runtime,
139                                               const Register thread, const Register value, const Register tmp1, const Register tmp2) {
140   assert_different_registers(value, tmp1, tmp2);
141   // Can we store a value in the given thread's buffer?
142   // (The index field is typed as size_t.)
143   __ ld(tmp1, Address(thread, in_bytes(index_offset)));   // tmp1 := *(index address)
144   __ beqz(tmp1, runtime);                                 // jump to runtime if index == 0 (full buffer)
145   // The buffer is not full, store value into it.
146   __ subi(tmp1, tmp1, wordSize);                           // tmp1 := next index
147   __ sd(tmp1, Address(thread, in_bytes(index_offset)));   // *(index address) := next index
148   __ ld(tmp2, Address(thread, in_bytes(buffer_offset)));  // tmp2 := buffer address
149   __ add(tmp2, tmp2, tmp1);
150   __ sd(value, Address(tmp2));                            // *(buffer address + next index) := value
151 }
152 
153 static void generate_pre_barrier_fast_path(MacroAssembler* masm,
154                                            const Register thread,
155                                            const Register tmp1) {
156   Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
157   // Is marking active?
158   if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
159     __ lwu(tmp1, in_progress);
160   } else {
161     assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
162     __ lbu(tmp1, in_progress);
163   }
164 }
165 
166 static void generate_pre_barrier_slow_path(MacroAssembler* masm,
167                                            const Register obj,
168                                            const Register pre_val,
169                                            const Register thread,
170                                            const Register tmp1,
171                                            const Register tmp2,
172                                            Label& done,
173                                            Label& runtime) {
174   // Do we need to load the previous value?
175   if (obj != noreg) {
176     __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
177   }
178   // Is the previous value null?
179   __ beqz(pre_val, done, /* is_far */ true);
180   generate_queue_test_and_insertion(masm,
181                                     G1ThreadLocalData::satb_mark_queue_index_offset(),
182                                     G1ThreadLocalData::satb_mark_queue_buffer_offset(),
183                                     runtime,
184                                     thread, pre_val, tmp1, tmp2);
185   __ j(done);
186 }
187 
188 void G1BarrierSetAssembler::g1_write_barrier_pre(MacroAssembler* masm,
189                                                  Register obj,
190                                                  Register pre_val,
191                                                  Register thread,
192                                                  Register tmp1,
193                                                  Register tmp2,
194                                                  bool tosca_live,
195                                                  bool expand_call) {
196   // If expand_call is true then we expand the call_VM_leaf macro
197   // directly to skip generating the check by
198   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
199 
200   assert(thread == xthread, "must be");
201 
202   Label done;
203   Label runtime;
204 
205   assert_different_registers(obj, pre_val, tmp1, tmp2);
206   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
207 
208   generate_pre_barrier_fast_path(masm, thread, tmp1);
209   // If marking is not active (*(mark queue active address) == 0), jump to done
210   __ beqz(tmp1, done);
211   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, done, runtime);
212 
213   __ bind(runtime);
214 
215   __ push_call_clobbered_registers();
216 
217   // Calling the runtime using the regular call_VM_leaf mechanism generates
218   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
219   // that checks that the *(fp+frame::interpreter_frame_last_sp) == nullptr.
220   //
221   // If we care generating the pre-barrier without a frame (e.g. in the
222   // intrinsified Reference.get() routine) then fp might be pointing to
223   // the caller frame and so this check will most likely fail at runtime.
224   //
225   // Expanding the call directly bypasses the generation of the check.
226   // So when we do not have have a full interpreter frame on the stack
227   // expand_call should be passed true.
228 
229   if (expand_call) {
230     assert(pre_val != c_rarg1, "smashed arg");
231     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
232   } else {
233     __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
234   }
235 
236   __ pop_call_clobbered_registers();
237 
238   __ bind(done);
239 
240 }
241 
242 static void generate_post_barrier(MacroAssembler* masm,
243                                   const Register store_addr,
244                                   const Register new_val,
245                                   const Register thread,
246                                   const Register tmp1,
247                                   const Register tmp2,
248                                   Label& done,

296   __ mv(c_rarg1, xthread);
297   __ mv(t1, runtime_path);
298   __ jalr(t1);
299 }
300 
301 void G1BarrierSetAssembler::g1_write_barrier_pre_c2(MacroAssembler* masm,
302                                                     Register obj,
303                                                     Register pre_val,
304                                                     Register thread,
305                                                     Register tmp1,
306                                                     Register tmp2,
307                                                     G1PreBarrierStubC2* stub) {
308   assert(thread == xthread, "must be");
309   assert_different_registers(obj, pre_val, tmp1, tmp2);
310   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
311 
312   stub->initialize_registers(obj, pre_val, thread, tmp1, tmp2);
313 
314   generate_pre_barrier_fast_path(masm, thread, tmp1);
315   // If marking is active (*(mark queue active address) != 0), jump to stub (slow path)
316   __ bnez(tmp1, *stub->entry(), /* is_far */ true);
317 
318   __ bind(*stub->continuation());
319 }
320 
321 void G1BarrierSetAssembler::generate_c2_pre_barrier_stub(MacroAssembler* masm,
322                                                          G1PreBarrierStubC2* stub) const {
323   Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
324   Label runtime;
325   Register obj = stub->obj();
326   Register pre_val = stub->pre_val();
327   Register thread = stub->thread();
328   Register tmp1 = stub->tmp1();
329   Register tmp2 = stub->tmp2();
330 
331   __ bind(*stub->entry());
332   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, *stub->continuation(), runtime);
333 
334   __ bind(runtime);
335   generate_c2_barrier_runtime_call(masm, stub, pre_val, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry));
336   __ j(*stub->continuation());

359   CardTableBarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
360   if (on_oop && on_reference) {
361     // RA is live.  It must be saved around calls.
362     __ enter(); // barrier may call runtime
363     // Generate the G1 pre-barrier code to log the value of
364     // the referent field in an SATB buffer.
365     g1_write_barrier_pre(masm /* masm */,
366                          noreg /* obj */,
367                          dst /* pre_val */,
368                          xthread /* thread */,
369                          tmp1 /* tmp1 */,
370                          tmp2 /* tmp2 */,
371                          true /* tosca_live */,
372                          true /* expand_call */);
373     __ leave();
374   }
375 }
376 
377 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
378                                          Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
379 
380   bool in_heap = (decorators & IN_HEAP) != 0;
381   bool as_normal = (decorators & AS_NORMAL) != 0;
382   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
383 
384   bool needs_pre_barrier = as_normal && !dest_uninitialized;
385   bool needs_post_barrier = (val != noreg && in_heap);
386 
387   assert_different_registers(val, tmp1, tmp2, tmp3);
388 
389   // flatten object address if needed
390   if (dst.offset() == 0) {
391     if (dst.base() != tmp3) {
392       __ mv(tmp3, dst.base());
393     }
394   } else {
395     __ la(tmp3, dst);
396   }
397 
398   if (needs_pre_barrier) {
399     g1_write_barrier_pre(masm,
400                          tmp3 /* obj */,
401                          tmp2 /* pre_val */,
402                          xthread /* thread */,
403                          tmp1 /* tmp1 */,
404                          t1 /* tmp2 */,
405                          val != noreg /* tosca_live */,
406                          false /* expand_call */);
407   }
408 
409   if (val == noreg) {
410     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), noreg, noreg, noreg, noreg);
411   } else {
412     // G1 barrier needs uncompressed oop for region cross check.
413     Register new_val = val;
414     if (needs_post_barrier) {
415       if (UseCompressedOops) {
416         new_val = t1;
417         __ mv(new_val, val);
418       }
419     }
420     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
421     if (needs_post_barrier) {
422       g1_write_barrier_post(masm,
423                             tmp3 /* store_adr */,
424                             new_val /* new_val */,
425                             xthread /* thread */,
426                             tmp1 /* tmp1 */,
427                             tmp2 /* tmp2 */);
428     }
429   }
430 }
431 
432 #ifdef COMPILER1
433 
434 #undef __
435 #define __ ce->masm()->
436 
437 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
438   G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
439 
440   // At this point we know that marking is in progress.
441   // If do_load() is true then we have to emit the
442   // load of the previous value; otherwise it has already
443   // been loaded into _pre_val.
444   __ bind(*stub->entry());
445 
446   assert(stub->pre_val()->is_register(), "Precondition.");
447 
448   Register pre_val_reg = stub->pre_val()->as_register();
< prev index next >