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 "asm/macroAssembler.inline.hpp"
 26 #include "code/aotCodeCache.hpp"
 27 #include "gc/g1/g1BarrierSet.hpp"
 28 #include "gc/g1/g1BarrierSetAssembler.hpp"
 29 #include "gc/g1/g1BarrierSetRuntime.hpp"
 30 #include "gc/g1/g1CardTable.hpp"
 31 #include "gc/g1/g1HeapRegion.hpp"
 32 #include "gc/g1/g1ThreadLocalData.hpp"
 33 #include "gc/shared/collectedHeap.hpp"
 34 #include "interpreter/interp_masm.hpp"
 35 #include "runtime/javaThread.hpp"
 36 #include "runtime/sharedRuntime.hpp"
 37 #ifdef COMPILER1
 38 #include "c1/c1_LIRAssembler.hpp"
 39 #include "c1/c1_MacroAssembler.hpp"
 40 #include "gc/g1/c1/g1BarrierSetC1.hpp"
 41 #endif // COMPILER1
 42 #ifdef COMPILER2
 43 #include "gc/g1/c2/g1BarrierSetC2.hpp"
 44 #endif // COMPILER2
 45 
 46 #define __ masm->
 47 
 48 void G1BarrierSetAssembler::gen_write_ref_array_pre_barrier(MacroAssembler* masm, DecoratorSet decorators,
 49                                                             Register addr, Register count, RegSet saved_regs) {
 50   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
 51   if (!dest_uninitialized) {
 52     Label done;
 53     Address in_progress(rthread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
 54 
 55     // Is marking active?
 56     if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
 57       __ ldrw(rscratch1, in_progress);
 58     } else {
 59       assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
 60       __ ldrb(rscratch1, in_progress);
 61     }
 62     __ cbzw(rscratch1, done);
 63 
 64     __ push(saved_regs, sp);
 65     if (count == c_rarg0) {
 66       if (addr == c_rarg1) {
 67         // exactly backwards!!
 68         __ mov(rscratch1, c_rarg0);
 69         __ mov(c_rarg0, c_rarg1);
 70         __ mov(c_rarg1, rscratch1);
 71       } else {
 72         __ mov(c_rarg1, count);
 73         __ mov(c_rarg0, addr);
 74       }
 75     } else {
 76       __ mov(c_rarg0, addr);
 77       __ mov(c_rarg1, count);
 78     }
 79     if (UseCompressedOops) {
 80       __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_array_pre_narrow_oop_entry), 2);
 81     } else {
 82       __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_array_pre_oop_entry), 2);
 83     }
 84     __ pop(saved_regs, sp);
 85 
 86     __ bind(done);
 87   }
 88 }
 89 
 90 void G1BarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm,
 91                                                              DecoratorSet decorators,
 92                                                              Register start,
 93                                                              Register count,
 94                                                              Register scratch) {
 95 
 96   Label done;
 97   Label loop;
 98   Label next;
 99 
100   __ cbz(count, done);
101 
102   // Calculate the number of card marks to set. Since the object might start and
103   // end within a card, we need to calculate this via the card table indexes of
104   // the actual start and last addresses covered by the object.
105   // Temporarily use the count register for the last element address.
106   __ lea(count, Address(start, count, Address::lsl(LogBytesPerHeapOop))); // end = start + count << LogBytesPerHeapOop
107   __ sub(count, count, BytesPerHeapOop);                                  // Use last element address for end.
108 
109   __ lsr(start, start, CardTable::card_shift());
110   __ lsr(count, count, CardTable::card_shift());
111   __ sub(count, count, start);                                            // Number of bytes to mark - 1.
112 
113   // Add card table base offset to start.
114   __ ldr(scratch, Address(rthread, in_bytes(G1ThreadLocalData::card_table_base_offset())));
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");
155     __ ldrb(tmp1, in_progress);
156   }
157 }
158 
159 static void generate_pre_barrier_slow_path(MacroAssembler* masm,
160                                            const Register obj,
161                                            const Register pre_val,
162                                            const Register thread,
163                                            const Register tmp1,
164                                            const Register tmp2,
165                                            Label& done,
166                                            Label& runtime) {
167   // Do we need to load the previous value?
168   if (obj != noreg) {
169     __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
170   }
171   // Is the previous value null?
172   __ cbz(pre_val, done);
173   generate_queue_test_and_insertion(masm,
174                                     G1ThreadLocalData::satb_mark_queue_index_offset(),
175                                     G1ThreadLocalData::satb_mark_queue_buffer_offset(),
176                                     runtime,
177                                     thread, pre_val, tmp1, tmp2);
178   __ b(done);
179 }
180 
181 void G1BarrierSetAssembler::g1_write_barrier_pre(MacroAssembler* masm,
182                                                  Register obj,
183                                                  Register pre_val,
184                                                  Register thread,
185                                                  Register tmp1,
186                                                  Register tmp2,
187                                                  bool tosca_live,
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 #if INCLUDE_CDS
248   // AOT code needs to load the barrier grain shift from the aot
249   // runtime constants area in the code cache otherwise we can compile
250   // it as an immediate operand
251   if (AOTCodeCache::is_on_for_dump()) {
252     address grain_shift_address = (address)AOTRuntimeConstants::grain_shift_address();
253     __ eor(tmp1, store_addr, new_val);
254     __ lea(tmp2, ExternalAddress(grain_shift_address));
255     __ ldrb(tmp2, tmp2);
256     __ lsrv(tmp1, tmp1, tmp2);
257     __ cbz(tmp1, done);
258   } else
259 #endif
260   {
261     __ eor(tmp1, store_addr, new_val);                     // tmp1 := store address ^ new value
262     __ lsr(tmp1, tmp1, G1HeapRegion::LogOfHRGrainBytes);   // tmp1 := ((store address ^ new value) >> LogOfHRGrainBytes)
263     __ cbz(tmp1, done);
264   }
265 
266   // Crosses regions, storing null?
267   if (new_val_may_be_null) {
268     __ cbz(new_val, done);
269   }
270   // Storing region crossing non-null.
271   __ lsr(tmp1, store_addr, CardTable::card_shift());     // tmp1 := card address relative to card table base
272 
273   Address card_table_addr(thread, in_bytes(G1ThreadLocalData::card_table_base_offset()));
274   __ ldr(tmp2, card_table_addr);                         // tmp2 := card table base address
275   if (UseCondCardMark) {
276     __ ldrb(rscratch1, Address(tmp1, tmp2));             // rscratch1 := card
277     // Instead of loading clean_card_val and comparing, we exploit the fact that
278     // the LSB of non-clean cards is always 0, and the LSB of clean cards 1.
279     __ tbz(rscratch1, 0, done);
280   }
281   static_assert(G1CardTable::dirty_card_val() == 0, "must be to use zr");
282   __ strb(zr, Address(tmp1, tmp2));                      // *(card address) := dirty_card_val
283 }
284 
285 void G1BarrierSetAssembler::g1_write_barrier_post(MacroAssembler* masm,
286                                                   Register store_addr,
287                                                   Register new_val,
288                                                   Register thread,
289                                                   Register tmp1,
290                                                   Register tmp2) {
291   Label done;
292   generate_post_barrier(masm, store_addr, new_val, thread, tmp1, tmp2, done, false /* new_val_may_be_null */);
293   __ bind(done);
294 }
295 
296 #if defined(COMPILER2)
297 
298 static void generate_c2_barrier_runtime_call(MacroAssembler* masm, G1BarrierStubC2* stub, const Register arg, const address runtime_path) {
299   SaveLiveRegisters save_registers(masm, stub);
300   if (c_rarg0 != arg) {
301     __ mov(c_rarg0, arg);
302   }
303   __ mov(c_rarg1, rthread);
304   __ lea(rscratch1, RuntimeAddress(runtime_path));
305   __ blr(rscratch1);
306 }
307 
308 void G1BarrierSetAssembler::g1_write_barrier_pre_c2(MacroAssembler* masm,
309                                                     Register obj,
310                                                     Register pre_val,
311                                                     Register thread,
312                                                     Register tmp1,
313                                                     Register tmp2,
314                                                     G1PreBarrierStubC2* stub) {
315   assert(thread == rthread, "must be");
316   assert_different_registers(obj, pre_val, tmp1, tmp2);
317   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
318 
319   stub->initialize_registers(obj, pre_val, thread, tmp1, tmp2);
320 
321   generate_pre_barrier_fast_path(masm, thread, tmp1);
322   // If marking is active (*(mark queue active address) != 0), jump to stub (slow path)
323   __ cbnzw(tmp1, *stub->entry());
324 
325   __ bind(*stub->continuation());
326 }
327 
328 void G1BarrierSetAssembler::generate_c2_pre_barrier_stub(MacroAssembler* masm,
329                                                          G1PreBarrierStubC2* stub) const {
330   Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
331   Label runtime;
332   Register obj = stub->obj();
333   Register pre_val = stub->pre_val();
334   Register thread = stub->thread();
335   Register tmp1 = stub->tmp1();
336   Register tmp2 = stub->tmp2();
337 
338   __ bind(*stub->entry());
339   generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, *stub->continuation(), runtime);
340 
341   __ bind(runtime);
342   generate_c2_barrier_runtime_call(masm, stub, pre_val, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry));
343   __ b(*stub->continuation());
344 }
345 
346 void G1BarrierSetAssembler::g1_write_barrier_post_c2(MacroAssembler* masm,
347                                                      Register store_addr,
348                                                      Register new_val,
349                                                      Register thread,
350                                                      Register tmp1,
351                                                      Register tmp2,
352                                                      bool new_val_may_be_null) {
353   Label done;
354   generate_post_barrier(masm, store_addr, new_val, thread, tmp1, tmp2, done, new_val_may_be_null);
355   __ bind(done);
356 }
357 
358 #endif // COMPILER2
359 
360 void G1BarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
361                                     Register dst, Address src, Register tmp1, Register tmp2) {
362   bool on_oop = is_reference_type(type);
363   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
364   bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
365   bool on_reference = on_weak || on_phantom;
366   CardTableBarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
367   if (on_oop && on_reference) {
368     // LR is live.  It must be saved around calls.
369     __ enter(/*strip_ret_addr*/true); // barrier may call runtime
370     // Generate the G1 pre-barrier code to log the value of
371     // the referent field in an SATB buffer.
372     g1_write_barrier_pre(masm /* masm */,
373                          noreg /* obj */,
374                          dst /* pre_val */,
375                          rthread /* thread */,
376                          tmp1 /* tmp1 */,
377                          tmp2 /* tmp2 */,
378                          true /* tosca_live */,
379                          true /* expand_call */);
380     __ leave();
381   }
382 }
383 
384 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
385                                          Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
386   // flatten object address if needed
387   if (dst.index() == noreg && dst.offset() == 0) {
388     if (dst.base() != tmp3) {
389       __ mov(tmp3, dst.base());
390     }
391   } else {
392     __ lea(tmp3, dst);
393   }
394 
395   g1_write_barrier_pre(masm,
396                        tmp3 /* obj */,
397                        tmp2 /* pre_val */,
398                        rthread /* thread */,
399                        tmp1  /* tmp1 */,
400                        rscratch2  /* tmp2 */,
401                        val != noreg /* tosca_live */,
402                        false /* expand_call */);
403 
404   if (val == noreg) {
405     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), noreg, noreg, noreg, noreg);
406   } else {
407     // G1 barrier needs uncompressed oop for region cross check.
408     Register new_val = val;
409     if (UseCompressedOops) {
410       new_val = rscratch2;
411       __ mov(new_val, val);
412     }
413     BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
414     g1_write_barrier_post(masm,
415                           tmp3 /* store_adr */,
416                           new_val /* new_val */,
417                           rthread /* thread */,
418                           tmp1 /* tmp1 */,
419                           tmp2 /* tmp2 */);
420   }
421 
422 }
423 
424 #ifdef COMPILER1
425 
426 #undef __
427 #define __ ce->masm()->
428 
429 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
430   G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
431   // At this point we know that marking is in progress.
432   // If do_load() is true then we have to emit the
433   // load of the previous value; otherwise it has already
434   // been loaded into _pre_val.
435 
436   __ bind(*stub->entry());
437 
438   assert(stub->pre_val()->is_register(), "Precondition.");
439 
440   Register pre_val_reg = stub->pre_val()->as_register();
441 
442   if (stub->do_load()) {
443     ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/);
444   }
445   __ cbz(pre_val_reg, *stub->continuation());
446   ce->store_parameter(stub->pre_val()->as_register(), 0);
447   __ far_call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
448   __ b(*stub->continuation());
449 }
450 
451 #undef __
452 
453 void G1BarrierSetAssembler::g1_write_barrier_post_c1(MacroAssembler* masm,
454                                                      Register store_addr,
455                                                      Register new_val,
456                                                      Register thread,
457                                                      Register tmp1,
458                                                      Register tmp2) {
459   Label done;
460   generate_post_barrier(masm, store_addr, new_val, thread, tmp1, tmp2, done, true /* new_val_may_be_null */);
461   masm->bind(done);
462 }
463 
464 #define __ sasm->
465 
466 void G1BarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
467   __ prologue("g1_pre_barrier", false);
468 
469   // arg0 : previous value of memory
470 
471   BarrierSet* bs = BarrierSet::barrier_set();
472 
473   const Register pre_val = r0;
474   const Register thread = rthread;
475   const Register tmp = rscratch1;
476 
477   Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
478   Address queue_index(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset()));
479   Address buffer(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()));
480 
481   Label done;
482   Label runtime;
483 
484   // Is marking still active?
485   if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
486     __ ldrw(tmp, in_progress);
487   } else {
488     assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
489     __ ldrb(tmp, in_progress);
490   }
491   __ cbzw(tmp, done);
492 
493   // Can we store original value in the thread's buffer?
494   __ ldr(tmp, queue_index);
495   __ cbz(tmp, runtime);
496 
497   __ sub(tmp, tmp, wordSize);
498   __ str(tmp, queue_index);
499   __ ldr(rscratch2, buffer);
500   __ add(tmp, tmp, rscratch2);
501   __ load_parameter(0, rscratch2);
502   __ str(rscratch2, Address(tmp, 0));
503   __ b(done);
504 
505   __ bind(runtime);
506   __ push_call_clobbered_registers();
507   __ load_parameter(0, pre_val);
508   __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
509   __ pop_call_clobbered_registers();
510   __ bind(done);
511 
512   __ epilogue();
513 }
514 
515 #undef __
516 
517 #endif // COMPILER1