< prev index next >

src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp

Print this page
@@ -40,10 +40,13 @@
  #ifdef COMPILER1
  #include "c1/c1_LIRAssembler.hpp"
  #include "c1/c1_MacroAssembler.hpp"
  #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  #endif
+ #ifdef COMPILER2
+ #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
+ #endif
  
  #define __ masm->
  
  void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
                                                         Register src, Register dst, Register count) {

@@ -899,5 +902,419 @@
  }
  
  #undef __
  
  #endif // COMPILER1
+ 
+ #ifdef COMPILER2
+ 
+ #undef __
+ #define __ masm->
+ 
+ void ShenandoahBarrierSetAssembler::load_c2(const MachNode* node, MacroAssembler* masm, Register dst, Address src, bool narrow) {
+   // Do the actual load. This load is the candidate for implicit null check, and MUST come first.
+   if (narrow) {
+     __ movl(dst, src);
+   } else {
+     __ movq(dst, src);
+   }
+ 
+   // Post-barrier: LRB
+   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
+     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, dst, src, narrow, false);
+     char check = 0;
+     check |= ShenandoahBarrierStubC2::needs_keep_alive_barrier(node)    ? ShenandoahHeap::MARKING : 0;
+     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier(node)      ? ShenandoahHeap::HAS_FORWARDED : 0;
+     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier_weak(node) ? ShenandoahHeap::WEAK_ROOTS : 0;
+     stub->enter_if_gc_state(*masm, check);
+   }
+ }
+ 
+ void ShenandoahBarrierSetAssembler::store_c2(const MachNode* node, MacroAssembler* masm,
+                                              Address dst, bool dst_narrow,
+                                              Register src, bool src_narrow,
+                                              Register tmp) {
+ 
+   // Pre-barrier: SATB, keep-alive the current memory value.
+   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
+     assert(!ShenandoahBarrierStubC2::needs_load_ref_barrier(node), "Should not be required for stores");
+     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, tmp, dst, dst_narrow, true);
+     stub->enter_if_gc_state(*masm, ShenandoahHeap::MARKING);
+   }
+ 
+   // Need to encode into tmp, because we cannot clobber src.
+   if (dst_narrow && !src_narrow) {
+     __ movq(tmp, src);
+     if (ShenandoahBarrierStubC2::maybe_null(node)) {
+       __ encode_heap_oop(tmp);
+     } else {
+       __ encode_heap_oop_not_null(tmp);
+     }
+     src = tmp;
+   }
+ 
+   // Do the actual store
+   if (dst_narrow) {
+     __ movl(dst, src);
+   } else {
+     __ movq(dst, src);
+   }
+ 
+   // Post-barrier: card updates.
+   if (ShenandoahBarrierStubC2::needs_card_barrier(node)) {
+     card_barrier_c2(masm, dst, tmp);
+   }
+ }
+ 
+ void ShenandoahBarrierSetAssembler::compare_and_set_c2(const MachNode* node, MacroAssembler* masm,
+                                                        Register res, Address addr,
+                                                        Register oldval, Register newval, Register tmp,
+                                                        bool narrow) {
+ 
+   assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
+ 
+   // Oldval and newval can be in the same register, but all other registers should be
+   // distinct for extra safety, as we shuffle register values around.
+   assert_different_registers(oldval, tmp, addr.base(), addr.index());
+   assert_different_registers(newval, tmp, addr.base(), addr.index());
+ 
+   // Pre-barrier covers several things:
+   //  a. Avoids false positives from CAS encountering to-space memory values.
+   //  b. Satisfies the need for LRB for the CAE result.
+   //  c. Records old value for the sake of SATB.
+   //
+   // (a) and (b) are covered because load barrier does memory location fixup.
+   // (c) is covered by KA on the current memory value.
+   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
+     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, tmp, addr, narrow, true);
+     char check = 0;
+     check |= ShenandoahBarrierStubC2::needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
+     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier(node)   ? ShenandoahHeap::HAS_FORWARDED : 0;
+     assert(!ShenandoahBarrierStubC2::needs_load_ref_barrier_weak(node), "Not supported for CAS");
+     stub->enter_if_gc_state(*masm, check);
+   }
+ 
+   // CAS!
+   __ lock();
+   if (narrow) {
+     __ cmpxchgl(newval, addr);
+   } else {
+     __ cmpxchgptr(newval, addr);
+   }
+ 
+   // If we need a boolean result out of CAS, set the flag appropriately and promote the result.
+   if (res != noreg) {
+     __ setcc(Assembler::equal, res);
+   }
+ 
+   // Post-barrier deals with card updates.
+   if (ShenandoahBarrierStubC2::needs_card_barrier(node)) {
+     card_barrier_c2(masm, addr, tmp);
+   }
+ }
+ 
+ void ShenandoahBarrierSetAssembler::get_and_set_c2(const MachNode* node, MacroAssembler* masm, Register newval, Address addr, Register tmp, bool narrow) {
+   assert_different_registers(newval, tmp, addr.base(), addr.index());
+ 
+   // Pre-barrier covers several things:
+   //  a. Satisfies the need for LRB for the GAS result.
+   //  b. Records old value for the sake of SATB.
+   //
+   // (a) is covered because load barrier does memory location fixup.
+   // (b) is covered by KA on the current memory value.
+   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
+     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, tmp, addr, narrow, true);
+     char check = 0;
+     check |= ShenandoahBarrierStubC2::needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
+     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier(node)   ? ShenandoahHeap::HAS_FORWARDED : 0;
+     assert(!ShenandoahBarrierStubC2::needs_load_ref_barrier_weak(node), "Not supported for GAS");
+     stub->enter_if_gc_state(*masm, check);
+   }
+ 
+   if (narrow) {
+     __ xchgl(newval, addr);
+   } else {
+     __ xchgq(newval, addr);
+   }
+ 
+   // Post-barrier deals with card updates.
+   if (ShenandoahBarrierStubC2::needs_card_barrier(node)) {
+     card_barrier_c2(masm, addr, tmp);
+   }
+ }
+ 
+ void ShenandoahBarrierSetAssembler::card_barrier_c2(MacroAssembler* masm, Address dst, Register tmp) {
+   Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
+ 
+   __ lea(tmp, dst);
+   __ shrptr(tmp, CardTable::card_shift());
+   __ addptr(tmp, Address(r15_thread, in_bytes(ShenandoahThreadLocalData::card_table_offset())));
+   Address card_address(tmp, 0);
+ 
+   assert(CardTable::dirty_card_val() == 0, "Encoding assumption");
+   Label L_done;
+   if (UseCondCardMark) {
+     __ cmpb(card_address, 0);
+     __ jccb(Assembler::equal, L_done);
+   }
+   if (UseCompressedOops && CompressedOops::base() == nullptr) {
+     __ movb(card_address, r12);
+   } else {
+     __ movb(card_address, 0);
+   }
+   __ bind(L_done);
+ }
+ 
+ #undef __
+ #define __ masm.
+ 
+ void ShenandoahBarrierStubC2::enter_if_gc_state(MacroAssembler& masm, const char test_state, Register tmp) {
+   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
+ 
+   Address gc_state_fast(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(test_state)));
+   __ cmpb(gc_state_fast, 0);
+   __ jcc(Assembler::notEqual, *entry());
+   __ bind(*continuation());
+ }
+ 
+ void ShenandoahBarrierStubC2::emit_code(MacroAssembler& masm) {
+   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
+ 
+   __ align(InteriorEntryAlignment);
+   __ bind(*entry());
+ 
+   // If we need to load ourselves, do it here.
+   if (_do_load) {
+     if (_narrow) {
+       __ movl(_obj, _addr);
+     } else {
+       __ movq(_obj, _addr);
+     }
+   }
+ 
+   // If the object is null, there is no point in applying barriers.
+   if (_narrow) {
+     __ testl(_obj, _obj);
+   } else {
+     __ testq(_obj, _obj);
+   }
+   __ jcc(Assembler::zero, *continuation());
+ 
+   // We need to make sure that loads done by callers survive across slow-path calls.
+   // For self-loads, we need to care about the case when both KA and LRB are enabled (rare).
+   if (!_do_load || (_needs_keep_alive_barrier && _needs_load_ref_barrier)) {
+     preserve(_obj);
+   }
+ 
+   // Go for barriers. If both KA and LRB are needed (rare), do additional gc-state
+   // checks to verify which one is currently needed. Note that KA and LRB are *not*
+   // exclusive, because we can have an overlapping marking/evac in generational mode.
+   if (_needs_keep_alive_barrier && _needs_load_ref_barrier) {
+     Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
+ 
+     Label L_skip_keepalive;
+     __ testb(gc_state, ShenandoahHeap::MARKING);
+     __ jcc(Assembler::zero, L_skip_keepalive);
+     keepalive(masm, nullptr);
+     __ bind(L_skip_keepalive);
+ 
+     __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | (_needs_load_ref_weak_barrier ? ShenandoahHeap::WEAK_ROOTS : 0));
+     __ jcc(Assembler::zero, *continuation());
+     lrb(masm, continuation());
+   } else if (_needs_keep_alive_barrier) {
+     keepalive(masm, continuation());
+   } else if (_needs_load_ref_barrier) {
+     lrb(masm, continuation());
+   } else {
+     ShouldNotReachHere();
+   }
+ }
+ 
+ void ShenandoahBarrierStubC2::keepalive(MacroAssembler& masm, Label* L_done) {
+   Address index(r15_thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
+   Address buffer(r15_thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
+ 
+   Label L_pop_and_slow;
+ 
+   // Need temp to work, allocate one now.
+   bool tmp_live;
+   Register tmp = select_temp_register(tmp_live);
+   if (tmp_live) {
+     __ push(tmp);
+   }
+ 
+   // Fast-path: put object into buffer.
+   // If buffer is already full, go slow.
+   __ movptr(tmp, index);
+   __ subptr(tmp, wordSize);
+   __ jccb(Assembler::below, L_pop_and_slow);
+   __ movptr(index, tmp);
+   __ addptr(tmp, buffer);
+ 
+   // If object is narrow, we need to unpack it before inserting into buffer,
+   // and pack it back. We can skip the unpack if we know that object is not preserved.
+   if (_narrow) {
+     __ decode_heap_oop_not_null(_obj);
+   }
+   __ movptr(Address(tmp, 0), _obj);
+   if (_narrow && is_preserved(_obj)) {
+     __ encode_heap_oop_not_null(_obj);
+   }
+ 
+   // Fast-path exits here.
+   if (tmp_live) {
+     __ pop(tmp);
+   }
+ 
+   Label L_fallthrough;
+   if (L_done != nullptr) {
+     __ jmp(*L_done);
+   } else {
+     __ jmp(L_fallthrough);
+   }
+ 
+   // Slow-path: call runtime to handle.
+   // Need to pop tmp immediately for stack to remain aligned.
+   __ bind(L_pop_and_slow);
+   if (tmp_live) {
+     __ pop(tmp);
+   }
+   {
+     SaveLiveRegisters slr(&masm, this);
+ 
+     // Shuffle in the arguments. The end result should be:
+     //   c_rarg0 <-- obj
+     if (c_rarg0 != _obj) {
+       __ mov(c_rarg0, _obj);
+     }
+ 
+     // Go to runtime and handle the rest there.
+     __ call(RuntimeAddress(keepalive_runtime_entry_addr()));
+   }
+   if (L_done != nullptr) {
+     __ jmp(*L_done);
+   } else {
+     __ bind(L_fallthrough);
+   }
+ }
+ 
+ void ShenandoahBarrierStubC2::lrb(MacroAssembler& masm, Label* L_done) {
+   assert(L_done != nullptr, "Must be set");
+ 
+   Label L_pop_and_slow, L_slow;
+ 
+   // If weak references are being processed, weak/phantom loads need to go slow,
+   // regardless of their cset status.
+   if (_needs_load_ref_weak_barrier) {
+     Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
+     __ testb(gc_state, ShenandoahHeap::WEAK_ROOTS);
+     __ jccb(Assembler::notZero, L_slow);
+   }
+ 
+   // Need temp to work, allocate one now.
+   bool tmp_live;
+   Register tmp = select_temp_register(tmp_live);
+   if (tmp_live) {
+     __ push(tmp);
+   }
+ 
+   // Compute the cset bitmap index
+   if (_narrow) {
+     __ decode_heap_oop_not_null(tmp, _obj);
+   } else {
+     __ movptr(tmp, _obj);
+   }
+   __ shrptr(tmp, ShenandoahHeapRegion::region_size_bytes_shift_jint());
+ 
+   Address cset_addr_arg;
+   intptr_t cset_addr = reinterpret_cast<intptr_t>(ShenandoahHeap::in_cset_fast_test_addr());
+   if ((cset_addr >> 3) < INT32_MAX) {
+     // Cset bitmap is at easily encodeable address. Just use it as offset.
+     assert(is_aligned(cset_addr, 8), "Sanity");
+     cset_addr_arg = Address(tmp, cset_addr >> 3, Address::times_8);
+   } else {
+     // Cset bitmap is way further than our encoding limit. Add its address fully.
+     bool tmp2_live;
+     Register tmp2 = select_temp_register(tmp2_live, /* skip_reg1 = */ tmp);
+     if (tmp2_live) {
+       __ push(tmp2);
+     }
+     __ movptr(tmp2, cset_addr);
+     __ addptr(tmp, tmp2);
+     if (tmp2_live) {
+       __ pop(tmp2);
+     }
+     cset_addr_arg = Address(tmp, 0);
+   }
+ 
+   // Cset-check. Fall-through to slow if in collection set.
+   __ cmpb(cset_addr_arg, 0);
+   if (tmp_live) {
+     __ jccb(Assembler::notEqual, L_pop_and_slow);
+     __ pop(tmp);
+     __ jmp(*L_done);
+   } else {
+     // Nothing else to do, jump back
+     __ jcc(Assembler::equal, *L_done);
+   }
+ 
+   // Slow path
+   __ bind(L_pop_and_slow);
+   // Need to pop tmp immediately for stack to remain aligned.
+   if (tmp_live) {
+     __ pop(tmp);
+   }
+   __ bind(L_slow);
+ 
+   // Obj is the result, need to temporarily stop preserving it.
+   bool is_obj_preserved = is_preserved(_obj);
+   if (is_obj_preserved) {
+     dont_preserve(_obj);
+   }
+   {
+     SaveLiveRegisters slr(&masm, this);
+ 
+     assert_different_registers(rax, c_rarg0, c_rarg1);
+ 
+     // Shuffle in the arguments. The end result should be:
+     //   c_rarg0 <-- obj
+     //   c_rarg1 <-- lea(addr)
+     if (_obj == c_rarg0) {
+       __ lea(c_rarg1, _addr);
+     } else if (_obj == c_rarg1) {
+       // Set up arguments in reverse, and then flip them
+       __ lea(c_rarg0, _addr);
+       __ xchgptr(c_rarg0, c_rarg1);
+     } else {
+       assert_different_registers(_obj, c_rarg0, c_rarg1);
+       __ lea(c_rarg1, _addr);
+       __ movptr(c_rarg0, _obj);
+     }
+ 
+     // Go to runtime and handle the rest there.
+     __ call(RuntimeAddress(lrb_runtime_entry_addr()));
+ 
+     // Save the result where needed.
+     if (_obj != rax) {
+       __ movptr(_obj, rax);
+     }
+   }
+   if (is_obj_preserved) {
+     preserve(_obj);
+   }
+ 
+   __ jmp(*L_done);
+ }
+ 
+ int ShenandoahBarrierStubC2::available_gp_registers() {
+   return Register::available_gp_registers();
+ }
+ 
+ bool ShenandoahBarrierStubC2::is_special_register(Register r) {
+   return r == rsp || r == rbp || r == r12_heapbase || r == r15_thread;
+ }
+ 
+ void ShenandoahBarrierStubC2::post_init() {
+   // Do nothing.
+ }
+ #undef __
+ #endif
< prev index next >