< 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.

121     __ lbu(tmp, Address(tmp, 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   __ add(tmp, start, count);
129   static_assert(G1CardTable::dirty_card_val() == 0, "must be to use zr");
130   __ sb(zr, Address(tmp, 0));
131 
132   __ bind(next);
133   __ subi(count, count, 1);
134   __ bgez(count, loop);
135 
136   __ bind(done);
137 }
138 
139 static void generate_queue_test_and_insertion(MacroAssembler* masm, ByteSize index_offset, ByteSize buffer_offset, Label& runtime,
140                                               const Register thread, const Register value, const Register tmp1, const Register 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, 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   if (expand_call) {
218     assert(pre_val != c_rarg1, "smashed arg");
219     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
220   } else {
221     __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
222   }
223 
224   __ pop_call_clobbered_registers();
225 
226   __ bind(done);
227 
228 }
229 
230 static void generate_post_barrier(MacroAssembler* masm,
231                                   const Register store_addr,
232                                   const Register new_val,
233                                   const Register thread,
234                                   const Register tmp1,
235                                   const Register tmp2,
236                                   Label& done,

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

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










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


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


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


401   }
402 }
403 
404 #ifdef COMPILER1
405 
406 #undef __
407 #define __ ce->masm()->
408 
409 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
410   G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
411 
412   // At this point we know that marking is in progress.
413   // If do_load() is true then we have to emit the
414   // load of the previous value; otherwise it has already
415   // been loaded into _pre_val.
416   __ bind(*stub->entry());
417 
418   assert(stub->pre_val()->is_register(), "Precondition.");
419 
420   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.

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

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

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