< prev index next >

src/hotspot/cpu/aarch64/gc/g1/g1BarrierSetAssembler_aarch64.cpp

Print this page

115   __ add(start, start, scratch);
116 
117   __ bind(loop);
118   if (UseCondCardMark) {
119     __ ldrb(scratch, Address(start, count));
120     // Instead of loading clean_card_val and comparing, we exploit the fact that
121     // the LSB of non-clean cards is always 0, and the LSB of clean cards 1.
122     __ tbz(scratch, 0, next);
123   }
124   static_assert(G1CardTable::dirty_card_val() == 0, "must be to use zr");
125   __ strb(zr, Address(start, count));
126   __ bind(next);
127   __ subs(count, count, 1);
128   __ br(Assembler::GE, loop);
129 
130   __ bind(done);
131 }
132 
133 static void generate_queue_test_and_insertion(MacroAssembler* masm, ByteSize index_offset, ByteSize buffer_offset, Label& runtime,
134                                               const Register thread, const Register value, const Register temp1, const Register temp2) {

135   // Can we store a value in the given thread's buffer?
136   // (The index field is typed as size_t.)
137   __ ldr(temp1, Address(thread, in_bytes(index_offset)));   // temp1 := *(index address)
138   __ cbz(temp1, runtime);                                   // jump to runtime if index == 0 (full buffer)
139   // The buffer is not full, store value into it.
140   __ sub(temp1, temp1, wordSize);                           // temp1 := next index
141   __ str(temp1, Address(thread, in_bytes(index_offset)));   // *(index address) := next index
142   __ ldr(temp2, Address(thread, in_bytes(buffer_offset)));  // temp2 := buffer address
143   __ str(value, Address(temp2, temp1));                     // *(buffer address + next index) := value
144 }
145 
146 static void generate_pre_barrier_fast_path(MacroAssembler* masm,
147                                            const Register thread,
148                                            const Register tmp1) {
149   Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
150   // Is marking active?
151   if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
152     __ ldrw(tmp1, in_progress);
153   } else {
154     assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");

188                                                  bool expand_call) {
189   // If expand_call is true then we expand the call_VM_leaf macro
190   // directly to skip generating the check by
191   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
192 
193   assert(thread == rthread, "must be");
194 
195   Label done;
196   Label runtime;
197 
198   assert_different_registers(obj, pre_val, tmp1, tmp2);
199   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
200 
201   generate_pre_barrier_fast_path(masm, thread, tmp1);
202   // If marking is not active (*(mark queue active address) == 0), jump to done
203   __ cbzw(tmp1, done);
204   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, done, runtime);
205 
206   __ bind(runtime);
207 
208   __ push_call_clobbered_registers();












209 
210   // Calling the runtime using the regular call_VM_leaf mechanism generates
211   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
212   // that checks that the *(rfp+frame::interpreter_frame_last_sp) == nullptr.
213   //
214   // If we care generating the pre-barrier without a frame (e.g. in the
215   // intrinsified Reference.get() routine) then rfp might be pointing to
216   // the caller frame and so this check will most likely fail at runtime.
217   //
218   // Expanding the call directly bypasses the generation of the check.
219   // So when we do not have have a full interpreter frame on the stack
220   // expand_call should be passed true.
221 
222   if (expand_call) {
223     assert(pre_val != c_rarg1, "smashed arg");
224     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
225   } else {
226     __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
227   }
228 
229   __ pop_call_clobbered_registers();
230 
231   __ bind(done);
232 
233 }
234 
235 static void generate_post_barrier(MacroAssembler* masm,
236                                   const Register store_addr,
237                                   const Register new_val,
238                                   const Register thread,
239                                   const Register tmp1,
240                                   const Register tmp2,
241                                   Label& done,
242                                   bool new_val_may_be_null) {
243   assert(thread == rthread, "must be");
244   assert_different_registers(store_addr, new_val, thread, tmp1, tmp2, noreg, rscratch1);
245 
246   // Does store cross heap regions?
247   __ eor(tmp1, store_addr, new_val);                     // tmp1 := store address ^ new value
248   __ lsr(tmp1, tmp1, G1HeapRegion::LogOfHRGrainBytes);   // tmp1 := ((store address ^ new value) >> LogOfHRGrainBytes)
249   __ cbz(tmp1, done);

350   ModRefBarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
351   if (on_oop && on_reference) {
352     // LR is live.  It must be saved around calls.
353     __ enter(/*strip_ret_addr*/true); // barrier may call runtime
354     // Generate the G1 pre-barrier code to log the value of
355     // the referent field in an SATB buffer.
356     g1_write_barrier_pre(masm /* masm */,
357                          noreg /* obj */,
358                          dst /* pre_val */,
359                          rthread /* thread */,
360                          tmp1 /* tmp1 */,
361                          tmp2 /* tmp2 */,
362                          true /* tosca_live */,
363                          true /* expand_call */);
364     __ leave();
365   }
366 }
367 
368 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
369                                          Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {










370   // flatten object address if needed
371   if (dst.index() == noreg && dst.offset() == 0) {
372     if (dst.base() != tmp3) {
373       __ mov(tmp3, dst.base());
374     }
375   } else {
376     __ lea(tmp3, dst);
377   }
378 
379   g1_write_barrier_pre(masm,
380                        tmp3 /* obj */,
381                        tmp2 /* pre_val */,
382                        rthread /* thread */,
383                        tmp1  /* tmp1 */,
384                        rscratch2  /* tmp2 */,
385                        val != noreg /* tosca_live */,
386                        false /* expand_call */);


387 
388   if (val == noreg) {
389     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), noreg, noreg, noreg, noreg);
390   } else {
391     // G1 barrier needs uncompressed oop for region cross check.
392     Register new_val = val;
393     if (UseCompressedOops) {
394       new_val = rscratch2;
395       __ mov(new_val, val);


396     }

397     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
398     g1_write_barrier_post(masm,
399                           tmp3 /* store_adr */,
400                           new_val /* new_val */,
401                           rthread /* thread */,
402                           tmp1 /* tmp1 */,
403                           tmp2 /* tmp2 */);


404   }
405 
406 }
407 
408 #ifdef COMPILER1
409 
410 #undef __
411 #define __ ce->masm()->
412 
413 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
414   G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
415   // At this point we know that marking is in progress.
416   // If do_load() is true then we have to emit the
417   // load of the previous value; otherwise it has already
418   // been loaded into _pre_val.
419 
420   __ bind(*stub->entry());
421 
422   assert(stub->pre_val()->is_register(), "Precondition.");
423 

115   __ add(start, start, scratch);
116 
117   __ bind(loop);
118   if (UseCondCardMark) {
119     __ ldrb(scratch, Address(start, count));
120     // Instead of loading clean_card_val and comparing, we exploit the fact that
121     // the LSB of non-clean cards is always 0, and the LSB of clean cards 1.
122     __ tbz(scratch, 0, next);
123   }
124   static_assert(G1CardTable::dirty_card_val() == 0, "must be to use zr");
125   __ strb(zr, Address(start, count));
126   __ bind(next);
127   __ subs(count, count, 1);
128   __ br(Assembler::GE, loop);
129 
130   __ bind(done);
131 }
132 
133 static void generate_queue_test_and_insertion(MacroAssembler* masm, ByteSize index_offset, ByteSize buffer_offset, Label& runtime,
134                                               const Register thread, const Register value, const Register temp1, const Register temp2) {
135   assert_different_registers(value, temp1, temp2);
136   // Can we store a value in the given thread's buffer?
137   // (The index field is typed as size_t.)
138   __ ldr(temp1, Address(thread, in_bytes(index_offset)));   // temp1 := *(index address)
139   __ cbz(temp1, runtime);                                   // jump to runtime if index == 0 (full buffer)
140   // The buffer is not full, store value into it.
141   __ sub(temp1, temp1, wordSize);                           // temp1 := next index
142   __ str(temp1, Address(thread, in_bytes(index_offset)));   // *(index address) := next index
143   __ ldr(temp2, Address(thread, in_bytes(buffer_offset)));  // temp2 := buffer address
144   __ str(value, Address(temp2, temp1));                     // *(buffer address + next index) := value
145 }
146 
147 static void generate_pre_barrier_fast_path(MacroAssembler* masm,
148                                            const Register thread,
149                                            const Register tmp1) {
150   Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
151   // Is marking active?
152   if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
153     __ ldrw(tmp1, in_progress);
154   } else {
155     assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");

189                                                  bool expand_call) {
190   // If expand_call is true then we expand the call_VM_leaf macro
191   // directly to skip generating the check by
192   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
193 
194   assert(thread == rthread, "must be");
195 
196   Label done;
197   Label runtime;
198 
199   assert_different_registers(obj, pre_val, tmp1, tmp2);
200   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
201 
202   generate_pre_barrier_fast_path(masm, thread, tmp1);
203   // If marking is not active (*(mark queue active address) == 0), jump to done
204   __ cbzw(tmp1, done);
205   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, done, runtime);
206 
207   __ bind(runtime);
208 
209   // save the live input values
210   RegSet saved = RegSet::of(pre_val);
211   FloatRegSet fsaved;
212 
213   // Barriers might be emitted when converting between (scalarized) calling
214   // conventions for inline types. Save all argument registers before calling
215   // into the runtime.
216 
217   // TODO 8366717 This came with 8284161: Implementation of Virtual Threads (Preview) later in May 2022
218   // Check if it's sufficient
219   //__ push_call_clobbered_registers();
220   assert_different_registers(rscratch1, pre_val); // push_CPU_state trashes rscratch1
221   __ push_CPU_state(true);
222 
223   // Calling the runtime using the regular call_VM_leaf mechanism generates
224   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
225   // that checks that the *(rfp+frame::interpreter_frame_last_sp) == nullptr.
226   //
227   // If we care generating the pre-barrier without a frame (e.g. in the
228   // intrinsified Reference.get() routine) then rfp might be pointing to
229   // the caller frame and so this check will most likely fail at runtime.
230   //
231   // Expanding the call directly bypasses the generation of the check.
232   // So when we do not have have a full interpreter frame on the stack
233   // expand_call should be passed true.
234 
235   if (expand_call) {
236     assert(pre_val != c_rarg1, "smashed arg");
237     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
238   } else {
239     __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
240   }
241 
242   __ pop_CPU_state(true);
243 
244   __ bind(done);
245 
246 }
247 
248 static void generate_post_barrier(MacroAssembler* masm,
249                                   const Register store_addr,
250                                   const Register new_val,
251                                   const Register thread,
252                                   const Register tmp1,
253                                   const Register tmp2,
254                                   Label& done,
255                                   bool new_val_may_be_null) {
256   assert(thread == rthread, "must be");
257   assert_different_registers(store_addr, new_val, thread, tmp1, tmp2, noreg, rscratch1);
258 
259   // Does store cross heap regions?
260   __ eor(tmp1, store_addr, new_val);                     // tmp1 := store address ^ new value
261   __ lsr(tmp1, tmp1, G1HeapRegion::LogOfHRGrainBytes);   // tmp1 := ((store address ^ new value) >> LogOfHRGrainBytes)
262   __ cbz(tmp1, done);

363   ModRefBarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
364   if (on_oop && on_reference) {
365     // LR is live.  It must be saved around calls.
366     __ enter(/*strip_ret_addr*/true); // barrier may call runtime
367     // Generate the G1 pre-barrier code to log the value of
368     // the referent field in an SATB buffer.
369     g1_write_barrier_pre(masm /* masm */,
370                          noreg /* obj */,
371                          dst /* pre_val */,
372                          rthread /* thread */,
373                          tmp1 /* tmp1 */,
374                          tmp2 /* tmp2 */,
375                          true /* tosca_live */,
376                          true /* expand_call */);
377     __ leave();
378   }
379 }
380 
381 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
382                                          Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
383 
384   bool in_heap = (decorators & IN_HEAP) != 0;
385   bool as_normal = (decorators & AS_NORMAL) != 0;
386   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
387 
388   bool needs_pre_barrier = as_normal && !dest_uninitialized;
389   bool needs_post_barrier = (val != noreg && in_heap);
390 
391   assert_different_registers(val, tmp1, tmp2, tmp3);
392 
393   // flatten object address if needed
394   if (dst.index() == noreg && dst.offset() == 0) {
395     if (dst.base() != tmp3) {
396       __ mov(tmp3, dst.base());
397     }
398   } else {
399     __ lea(tmp3, dst);
400   }
401 
402   if (needs_pre_barrier) {
403     g1_write_barrier_pre(masm,
404                          tmp3 /* obj */,
405                          tmp2 /* pre_val */,
406                          rthread /* thread */,
407                          tmp1  /* tmp1 */,
408                          rscratch2  /* tmp2 */,
409                          val != noreg /* tosca_live */,
410                          false /* expand_call */);
411   }
412 
413   if (val == noreg) {
414     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), noreg, noreg, noreg, noreg);
415   } else {
416     // G1 barrier needs uncompressed oop for region cross check.
417     Register new_val = val;
418     if (needs_post_barrier) {
419       if (UseCompressedOops) {
420         new_val = rscratch2;
421         __ mov(new_val, val);
422       }
423     }
424 
425     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
426     if (needs_post_barrier) {
427       g1_write_barrier_post(masm,
428                             tmp3 /* store_adr */,
429                             new_val /* new_val */,
430                             rthread /* thread */,
431                             tmp1 /* tmp1 */,
432                             tmp2 /* tmp2 */);
433     }
434   }
435 
436 }
437 
438 #ifdef COMPILER1
439 
440 #undef __
441 #define __ ce->masm()->
442 
443 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
444   G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
445   // At this point we know that marking is in progress.
446   // If do_load() is true then we have to emit the
447   // load of the previous value; otherwise it has already
448   // been loaded into _pre_val.
449 
450   __ bind(*stub->entry());
451 
452   assert(stub->pre_val()->is_register(), "Precondition.");
453 
< prev index next >