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 "classfile/classLoaderData.hpp"
 26 #include "gc/shared/barrierSet.hpp"
 27 #include "gc/shared/barrierSetAssembler.hpp"
 28 #include "gc/shared/barrierSetNMethod.hpp"
 29 #include "gc/shared/collectedHeap.hpp"
 30 #include "interpreter/interp_masm.hpp"
 31 #include "memory/universe.hpp"
 32 #include "runtime/javaThread.hpp"
 33 #include "runtime/jniHandles.hpp"
 34 #include "runtime/sharedRuntime.hpp"
 35 #include "runtime/stubRoutines.hpp"
 36 #ifdef COMPILER2
 37 #include "code/vmreg.inline.hpp"
 38 #include "gc/shared/c2/barrierSetC2.hpp"
 39 #endif // COMPILER2
 40 
 41 
 42 #define __ masm->
 43 
 44 void BarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 45                                   Register dst, Address src, Register tmp1, Register tmp2) {
 46 
 47   // LR is live.  It must be saved around calls.
 48 
 49   bool in_heap = (decorators & IN_HEAP) != 0;
 50   bool in_native = (decorators & IN_NATIVE) != 0;
 51   bool is_not_null = (decorators & IS_NOT_NULL) != 0;
 52   switch (type) {
 53   case T_OBJECT:
 54   case T_ARRAY: {
 55     if (in_heap) {
 56       if (UseCompressedOops) {
 57         __ ldrw(dst, src);
 58         if (is_not_null) {
 59           __ decode_heap_oop_not_null(dst);
 60         } else {
 61           __ decode_heap_oop(dst);
 62         }
 63       } else {
 64         __ ldr(dst, src);
 65       }
 66     } else {
 67       assert(in_native, "why else?");
 68       __ ldr(dst, src);
 69     }
 70     break;
 71   }
 72   case T_BOOLEAN: __ load_unsigned_byte (dst, src); break;
 73   case T_BYTE:    __ load_signed_byte   (dst, src); break;
 74   case T_CHAR:    __ load_unsigned_short(dst, src); break;
 75   case T_SHORT:   __ load_signed_short  (dst, src); break;
 76   case T_INT:     __ ldrw               (dst, src); break;
 77   case T_LONG:    __ ldr                (dst, src); break;
 78   case T_ADDRESS: __ ldr                (dst, src); break;
 79   case T_FLOAT:   __ ldrs               (v0, src);  break;
 80   case T_DOUBLE:  __ ldrd               (v0, src);  break;
 81   default: Unimplemented();
 82   }
 83 }
 84 
 85 void BarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 86                                    Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
 87   bool in_heap = (decorators & IN_HEAP) != 0;
 88   bool in_native = (decorators & IN_NATIVE) != 0;
 89   switch (type) {
 90   case T_OBJECT:
 91   case T_ARRAY: {
 92     val = val == noreg ? zr : val;
 93     if (in_heap) {
 94       if (UseCompressedOops) {
 95         assert(!dst.uses(val), "not enough registers");
 96         if (val != zr) {
 97           __ encode_heap_oop(val);
 98         }
 99         __ strw(val, dst);
100       } else {
101         __ str(val, dst);
102       }
103     } else {
104       assert(in_native, "why else?");
105       __ str(val, dst);
106     }
107     break;
108   }
109   case T_BOOLEAN:
110     __ andw(val, val, 0x1);  // boolean is true if LSB is 1
111     __ strb(val, dst);
112     break;
113   case T_BYTE:    __ strb(val, dst); break;
114   case T_CHAR:    __ strh(val, dst); break;
115   case T_SHORT:   __ strh(val, dst); break;
116   case T_INT:     __ strw(val, dst); break;
117   case T_LONG:    __ str (val, dst); break;
118   case T_ADDRESS: __ str (val, dst); break;
119   case T_FLOAT:   __ strs(v0,  dst); break;
120   case T_DOUBLE:  __ strd(v0,  dst); break;
121   default: Unimplemented();
122   }
123 }
124 
125 void BarrierSetAssembler::copy_load_at(MacroAssembler* masm,
126                                        DecoratorSet decorators,
127                                        BasicType type,
128                                        size_t bytes,
129                                        Register dst1,
130                                        Register dst2,
131                                        Address src,
132                                        Register tmp) {
133   if (bytes == 1) {
134     assert(dst2 == noreg, "invariant");
135     __ ldrb(dst1, src);
136   } else if (bytes == 2) {
137     assert(dst2 == noreg, "invariant");
138     __ ldrh(dst1, src);
139   } else if (bytes == 4) {
140     assert(dst2 == noreg, "invariant");
141     __ ldrw(dst1, src);
142   } else if (bytes == 8) {
143     assert(dst2 == noreg, "invariant");
144     __ ldr(dst1, src);
145   } else if (bytes == 16) {
146     assert(dst2 != noreg, "invariant");
147     assert(dst2 != dst1, "invariant");
148     __ ldp(dst1, dst2, src);
149   } else {
150     // Not the right size
151     ShouldNotReachHere();
152   }
153   if ((decorators & ARRAYCOPY_CHECKCAST) != 0 && UseCompressedOops) {
154     __ decode_heap_oop(dst1);
155   }
156 }
157 
158 void BarrierSetAssembler::copy_store_at(MacroAssembler* masm,
159                                         DecoratorSet decorators,
160                                         BasicType type,
161                                         size_t bytes,
162                                         Address dst,
163                                         Register src1,
164                                         Register src2,
165                                         Register tmp1,
166                                         Register tmp2,
167                                         Register tmp3) {
168   if ((decorators & ARRAYCOPY_CHECKCAST) != 0 && UseCompressedOops) {
169     __ encode_heap_oop(src1);
170   }
171   if (bytes == 1) {
172     assert(src2 == noreg, "invariant");
173     __ strb(src1, dst);
174   } else if (bytes == 2) {
175     assert(src2 == noreg, "invariant");
176     __ strh(src1, dst);
177   } else if (bytes == 4) {
178     assert(src2 == noreg, "invariant");
179     __ strw(src1, dst);
180   } else if (bytes == 8) {
181     assert(src2 == noreg, "invariant");
182     __ str(src1, dst);
183   } else if (bytes == 16) {
184     assert(src2 != noreg, "invariant");
185     assert(src2 != src1, "invariant");
186     __ stp(src1, src2, dst);
187   } else {
188     // Not the right size
189     ShouldNotReachHere();
190   }
191 }
192 
193 void BarrierSetAssembler::copy_load_at(MacroAssembler* masm,
194                                        DecoratorSet decorators,
195                                        BasicType type,
196                                        size_t bytes,
197                                        FloatRegister dst1,
198                                        FloatRegister dst2,
199                                        Address src,
200                                        Register tmp1,
201                                        Register tmp2,
202                                        FloatRegister vec_tmp) {
203   if (bytes == 32) {
204     __ ldpq(dst1, dst2, src);
205   } else {
206     ShouldNotReachHere();
207   }
208 }
209 
210 void BarrierSetAssembler::copy_store_at(MacroAssembler* masm,
211                                         DecoratorSet decorators,
212                                         BasicType type,
213                                         size_t bytes,
214                                         Address dst,
215                                         FloatRegister src1,
216                                         FloatRegister src2,
217                                         Register tmp1,
218                                         Register tmp2,
219                                         Register tmp3,
220                                         FloatRegister vec_tmp1,
221                                         FloatRegister vec_tmp2,
222                                         FloatRegister vec_tmp3) {
223   if (bytes == 32) {
224     __ stpq(src1, src2, dst);
225   } else {
226     ShouldNotReachHere();
227   }
228 }
229 
230 void BarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
231                                                         Register obj, Register tmp, Label& slowpath) {
232   // If mask changes we need to ensure that the inverse is still encodable as an immediate
233   STATIC_ASSERT(JNIHandles::tag_mask == 0b11);
234   __ andr(obj, obj, ~JNIHandles::tag_mask);
235   __ ldr(obj, Address(obj, 0));             // *obj
236 }
237 
238 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
239 void BarrierSetAssembler::tlab_allocate(MacroAssembler* masm, Register obj,
240                                         Register var_size_in_bytes,
241                                         int con_size_in_bytes,
242                                         Register t1,
243                                         Register t2,
244                                         Label& slow_case) {
245   assert_different_registers(obj, t2);
246   assert_different_registers(obj, var_size_in_bytes);
247   Register end = t2;
248 
249   // verify_tlab();
250 
251   __ ldr(obj, Address(rthread, JavaThread::tlab_top_offset()));
252   if (var_size_in_bytes == noreg) {
253     __ lea(end, Address(obj, con_size_in_bytes));
254   } else {
255     __ lea(end, Address(obj, var_size_in_bytes));
256   }
257   __ ldr(rscratch1, Address(rthread, JavaThread::tlab_end_offset()));
258   __ cmp(end, rscratch1);
259   __ br(Assembler::HI, slow_case);
260 
261   // update the tlab top pointer
262   __ str(end, Address(rthread, JavaThread::tlab_top_offset()));
263 
264   // recover var_size_in_bytes if necessary
265   if (var_size_in_bytes == end) {
266     __ sub(var_size_in_bytes, var_size_in_bytes, obj);
267   }
268   // verify_tlab();
269 }
270 
271 static volatile uint32_t _patching_epoch = 0;
272 
273 address BarrierSetAssembler::patching_epoch_addr() {
274   return (address)&_patching_epoch;
275 }
276 
277 void BarrierSetAssembler::increment_patching_epoch() {
278   AtomicAccess::inc(&_patching_epoch);
279 }
280 
281 void BarrierSetAssembler::clear_patching_epoch() {
282   _patching_epoch = 0;
283 }
284 
285 void BarrierSetAssembler::nmethod_entry_barrier(MacroAssembler* masm, Label* slow_path, Label* continuation, Label* guard) {
286   BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
287 
288   Label local_guard;
289   Label skip_barrier;
290   NMethodPatchingType patching_type = nmethod_patching_type();
291 
292   if (slow_path == nullptr) {
293     guard = &local_guard;
294   }
295 
296   // If the slow path is out of line in a stub, we flip the condition
297   Assembler::Condition condition = slow_path == nullptr ? Assembler::EQ : Assembler::NE;
298   Label& barrier_target = slow_path == nullptr ? skip_barrier : *slow_path;
299 
300   __ ldrw(rscratch1, *guard);
301 
302   if (patching_type == NMethodPatchingType::stw_instruction_and_data_patch) {
303     // With STW patching, no data or instructions are updated concurrently,
304     // which means there isn't really any need for any fencing for neither
305     // data nor instruction modifications happening concurrently. The
306     // instruction patching is handled with isb fences on the way back
307     // from the safepoint to Java. So here we can do a plain conditional
308     // branch with no fencing.
309     Address thread_disarmed_addr(rthread, in_bytes(bs_nm->thread_disarmed_guard_value_offset()));
310     __ ldrw(rscratch2, thread_disarmed_addr);
311     __ cmp(rscratch1, rscratch2);
312   } else if (patching_type == NMethodPatchingType::conc_instruction_and_data_patch) {
313     // If we patch code we need both a code patching and a loadload
314     // fence. It's not super cheap, so we use a global epoch mechanism
315     // to hide them in a slow path.
316     // The high level idea of the global epoch mechanism is to detect
317     // when any thread has performed the required fencing, after the
318     // last nmethod was disarmed. This implies that the required
319     // fencing has been performed for all preceding nmethod disarms
320     // as well. Therefore, we do not need any further fencing.
321     __ lea(rscratch2, ExternalAddress((address)&_patching_epoch));
322     // Embed an artificial data dependency to order the guard load
323     // before the epoch load.
324     __ orr(rscratch2, rscratch2, rscratch1, Assembler::LSR, 32);
325     // Read the global epoch value.
326     __ ldrw(rscratch2, rscratch2);
327     // Combine the guard value (low order) with the epoch value (high order).
328     __ orr(rscratch1, rscratch1, rscratch2, Assembler::LSL, 32);
329     // Compare the global values with the thread-local values.
330     Address thread_disarmed_and_epoch_addr(rthread, in_bytes(bs_nm->thread_disarmed_guard_value_offset()));
331     __ ldr(rscratch2, thread_disarmed_and_epoch_addr);
332     __ cmp(rscratch1, rscratch2);
333   } else {
334     ShouldNotReachHere();
335   }
336   __ br(condition, barrier_target);
337 
338   if (slow_path == nullptr) {
339     __ lea(rscratch1, RuntimeAddress(StubRoutines::method_entry_barrier()));
340     __ blr(rscratch1);
341     __ b(skip_barrier);
342 
343     __ bind(local_guard);
344 
345     __ emit_int32(0);   // nmethod guard value. Skipped over in common case.
346   } else {
347     __ bind(*continuation);
348   }
349 
350   __ bind(skip_barrier);
351 }
352 
353 void BarrierSetAssembler::c2i_entry_barrier(MacroAssembler* masm) {
354   Label bad_call;
355   __ cbz(rmethod, bad_call);
356 
357   // Pointer chase to the method holder to find out if the method is concurrently unloading.
358   Label method_live;
359   __ load_method_holder_cld(rscratch1, rmethod);
360 
361   // Is it a strong CLD?
362   __ ldrw(rscratch2, Address(rscratch1, ClassLoaderData::keep_alive_ref_count_offset()));
363   __ cbnz(rscratch2, method_live);
364 
365   // Is it a weak but alive CLD?
366   __ push(RegSet::of(r10), sp);
367   __ ldr(r10, Address(rscratch1, ClassLoaderData::holder_offset()));
368 
369   __ resolve_weak_handle(r10, rscratch1, rscratch2);
370   __ mov(rscratch1, r10);
371   __ pop(RegSet::of(r10), sp);
372   __ cbnz(rscratch1, method_live);
373 
374   __ bind(bad_call);
375 
376   __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
377   __ bind(method_live);
378 }
379 
380 void BarrierSetAssembler::check_oop(MacroAssembler* masm, Register obj, Register tmp1, Register tmp2, Label& error) {
381   // Check if the oop is in the right area of memory
382   __ mov(tmp2, (intptr_t) Universe::verify_oop_mask());
383   __ andr(tmp1, obj, tmp2);
384   __ mov(tmp2, (intptr_t) Universe::verify_oop_bits());
385 
386   // Compare tmp1 and tmp2.  We don't use a compare
387   // instruction here because the flags register is live.
388   __ eor(tmp1, tmp1, tmp2);
389   __ cbnz(tmp1, error);
390 
391   // make sure klass is 'reasonable', which is not zero.
392   __ load_klass(obj, obj); // get klass
393   __ cbz(obj, error);      // if klass is null it is broken
394 }
395 
396 #ifdef COMPILER2
397 
398 OptoReg::Name BarrierSetAssembler::encode_float_vector_register_size(const Node* node, OptoReg::Name opto_reg) {
399   switch (node->ideal_reg()) {
400     case Op_RegF:
401     case Op_RegI: // RA may place scalar values (Op_RegI/N/L/P) in FP registers when UseFPUForSpilling is enabled
402     case Op_RegN:
403       // No need to refine. The original encoding is already fine to distinguish.
404       assert(opto_reg % 4 == 0, "32-bit register should only occupy a single slot");
405       break;
406     // Use different encoding values of the same fp/vector register to help distinguish different sizes.
407     // Such as V16. The OptoReg::name and its corresponding slot value are
408     // "V16": 64, "V16_H": 65, "V16_J": 66, "V16_K": 67.
409     case Op_RegD:
410     case Op_VecD:
411     case Op_RegL:
412     case Op_RegP:
413       opto_reg &= ~3;
414       opto_reg |= 1;
415       break;
416     case Op_VecX:
417       opto_reg &= ~3;
418       opto_reg |= 2;
419       break;
420     case Op_VecA:
421       opto_reg &= ~3;
422       opto_reg |= 3;
423       break;
424     default:
425       assert(false, "unexpected ideal register");
426       ShouldNotReachHere();
427   }
428   return opto_reg;
429 }
430 
431 OptoReg::Name BarrierSetAssembler::refine_register(const Node* node, OptoReg::Name opto_reg) {
432   if (!OptoReg::is_reg(opto_reg)) {
433     return OptoReg::Bad;
434   }
435 
436   const VMReg vm_reg = OptoReg::as_VMReg(opto_reg);
437   if (vm_reg->is_FloatRegister()) {
438     opto_reg = encode_float_vector_register_size(node, opto_reg);
439   }
440 
441   return opto_reg;
442 }
443 
444 #undef __
445 #define __ _masm->
446 
447 void SaveLiveRegisters::initialize(BarrierStubC2* stub) {
448   int index = -1;
449   GrowableArray<RegisterData> registers;
450   VMReg prev_vm_reg = VMRegImpl::Bad();
451 
452   RegMaskIterator rmi(stub->preserve_set());
453   while (rmi.has_next()) {
454     OptoReg::Name opto_reg = rmi.next();
455     VMReg vm_reg = OptoReg::as_VMReg(opto_reg);
456 
457     if (vm_reg->is_Register()) {
458       // GPR may have one or two slots in regmask
459       // Determine whether the current vm_reg is the same physical register as the previous one
460       if (is_same_register(vm_reg, prev_vm_reg)) {
461         registers.at(index)._slots++;
462       } else {
463         RegisterData reg_data = { vm_reg, 1 };
464         index = registers.append(reg_data);
465       }
466     } else if (vm_reg->is_FloatRegister()) {
467       // We have size encoding in OptoReg of stub->preserve_set()
468       // After encoding, float/neon/sve register has only one slot in regmask
469       // Decode it to get the actual size
470       VMReg vm_reg_base = vm_reg->as_FloatRegister()->as_VMReg();
471       int slots = decode_float_vector_register_size(opto_reg);
472       RegisterData reg_data = { vm_reg_base, slots };
473       index = registers.append(reg_data);
474     } else if (vm_reg->is_PRegister()) {
475       // PRegister has only one slot in regmask
476       RegisterData reg_data = { vm_reg, 1 };
477       index = registers.append(reg_data);
478     } else {
479       assert(false, "Unknown register type");
480       ShouldNotReachHere();
481     }
482     prev_vm_reg = vm_reg;
483   }
484 
485   // Record registers that needs to be saved/restored
486   for (GrowableArrayIterator<RegisterData> it = registers.begin(); it != registers.end(); ++it) {
487     RegisterData reg_data = *it;
488     VMReg vm_reg = reg_data._reg;
489     int slots = reg_data._slots;
490     if (vm_reg->is_Register()) {
491       assert(slots == 1 || slots == 2, "Unexpected register save size");
492       _gp_regs += RegSet::of(vm_reg->as_Register());
493     } else if (vm_reg->is_FloatRegister()) {
494       if (slots == 1 || slots == 2) {
495         _fp_regs += FloatRegSet::of(vm_reg->as_FloatRegister());
496       } else if (slots == 4) {
497         _neon_regs += FloatRegSet::of(vm_reg->as_FloatRegister());
498       } else {
499         assert(slots == Matcher::scalable_vector_reg_size(T_FLOAT), "Unexpected register save size");
500         _sve_regs += FloatRegSet::of(vm_reg->as_FloatRegister());
501       }
502     } else {
503       assert(vm_reg->is_PRegister() && slots == 1, "Unknown register type");
504       _p_regs += PRegSet::of(vm_reg->as_PRegister());
505     }
506   }
507 
508   // Remove C-ABI SOE registers and scratch regs
509   _gp_regs -= RegSet::range(r19, r30) + RegSet::of(r8, r9);
510 
511   // Remove C-ABI SOE fp registers
512   _fp_regs -= FloatRegSet::range(v8, v15);
513 }
514 
515 enum RC SaveLiveRegisters::rc_class(VMReg reg) {
516   if (reg->is_reg()) {
517     if (reg->is_Register()) {
518       return rc_int;
519     } else if (reg->is_FloatRegister()) {
520       return rc_float;
521     } else if (reg->is_PRegister()) {
522       return rc_predicate;
523     }
524   }
525   if (reg->is_stack()) {
526     return rc_stack;
527   }
528   return rc_bad;
529 }
530 
531 bool SaveLiveRegisters::is_same_register(VMReg reg1, VMReg reg2) {
532   if (reg1 == reg2) {
533     return true;
534   }
535   if (rc_class(reg1) == rc_class(reg2)) {
536     if (reg1->is_Register()) {
537       return reg1->as_Register() == reg2->as_Register();
538     } else if (reg1->is_FloatRegister()) {
539       return reg1->as_FloatRegister() == reg2->as_FloatRegister();
540     } else if (reg1->is_PRegister()) {
541       return reg1->as_PRegister() == reg2->as_PRegister();
542     }
543   }
544   return false;
545 }
546 
547 int SaveLiveRegisters::decode_float_vector_register_size(OptoReg::Name opto_reg) {
548   switch (opto_reg & 3) {
549     case 0:
550       return 1;
551     case 1:
552       return 2;
553     case 2:
554       return 4;
555     case 3:
556       return Matcher::scalable_vector_reg_size(T_FLOAT);
557     default:
558       ShouldNotReachHere();
559       return 0;
560   }
561 }
562 
563 SaveLiveRegisters::SaveLiveRegisters(MacroAssembler* masm, BarrierStubC2* stub)
564   : _masm(masm),
565     _gp_regs(),
566     _fp_regs(),
567     _neon_regs(),
568     _sve_regs(),
569     _p_regs() {
570 
571   // Figure out what registers to save/restore
572   initialize(stub);
573 
574   // Save registers
575   __ push(_gp_regs, sp);
576   __ push_fp(_fp_regs, sp, MacroAssembler::PushPopFp);
577   __ push_fp(_neon_regs, sp, MacroAssembler::PushPopNeon);
578   __ push_fp(_sve_regs, sp, MacroAssembler::PushPopSVE);
579   __ push_p(_p_regs, sp);
580 }
581 
582 SaveLiveRegisters::~SaveLiveRegisters() {
583   // Restore registers
584   __ pop_p(_p_regs, sp);
585   __ pop_fp(_sve_regs, sp, MacroAssembler::PushPopSVE);
586   __ pop_fp(_neon_regs, sp, MacroAssembler::PushPopNeon);
587   __ pop_fp(_fp_regs, sp, MacroAssembler::PushPopFp);
588 
589   // External runtime call may clobber ptrue reg
590   __ reinitialize_ptrue();
591 
592   __ pop(_gp_regs, sp);
593 }
594 
595 #endif // COMPILER2