< prev index next > src/hotspot/cpu/aarch64/gc/shenandoah/shenandoahBarrierSetAssembler_aarch64.cpp
Print this page
#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->
+ #ifdef PRODUCT
+ #define BLOCK_COMMENT(str) /* nothing */
+ #else
+ #define BLOCK_COMMENT(str) __ block_comment(str)
+ #endif
+
void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
Register src, Register dst, Register count, RegSet saved_regs) {
if (is_oop) {
bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
if ((ShenandoahSATBBarrier && !dest_uninitialized) || ShenandoahLoadRefBarrier) {
} else {
__ cset(result, Assembler::EQ);
}
}
+ #ifdef COMPILER2
+ void ShenandoahBarrierSetAssembler::load_ref_barrier_c2(const MachNode* node, MacroAssembler* masm, Register obj, Register addr, Register tmp, bool narrow, bool maybe_null) {
+ assert_different_registers(obj, addr, tmp);
+ BLOCK_COMMENT("load_ref_barrier_c2 {");
+ if (!ShenandoahLoadRefBarrierStubC2::needs_barrier(node)) {
+ return;
+ }
+ Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
+ ShenandoahLoadRefBarrierStubC2* const stub = ShenandoahLoadRefBarrierStubC2::create(node, obj, addr, tmp, noreg, noreg, narrow);
+
+ // Don't preserve the obj across the runtime call, we override it from the return value anyway.
+ stub->dont_preserve(obj);
+ if (tmp != noreg) {
+ stub->dont_preserve(tmp); // temp, no need to save
+ }
+
+ Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
+ __ ldrb(rscratch1, gc_state);
+
+ // Check if GC marking is in progress or we are handling a weak reference, otherwise we don't have to do anything.
+ bool is_strong = (node->barrier_data() & ShenandoahBarrierStrong) != 0;
+ if (is_strong) {
+ __ tbz(rscratch1, ShenandoahHeap::HAS_FORWARDED_BITPOS, *stub->continuation());
+ __ b(*stub->entry());
+ } else {
+ static_assert(ShenandoahHeap::HAS_FORWARDED_BITPOS == 0, "Relied on in LRB check below.");
+ __ orr(tmp, rscratch1, rscratch1, Assembler::LSR, ShenandoahHeap::WEAK_ROOTS_BITPOS);
+ __ tbz(tmp, ShenandoahHeap::HAS_FORWARDED_BITPOS, *stub->continuation());
+ __ b(*stub->entry());
+ }
+
+ __ bind(*stub->continuation());
+ BLOCK_COMMENT("} load_ref_barrier_c2");
+ }
+
+ void ShenandoahBarrierSetAssembler::load_ref_barrier_c3(const MachNode* node, MacroAssembler* masm, Register obj, Register addr, Register tmp, bool narrow, bool maybe_null, Register gc_state) {
+ BLOCK_COMMENT("load_ref_barrier_c3 {");
+ if (!ShenandoahLoadRefBarrierStubC2::needs_barrier(node)) {
+ return;
+ }
+ Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
+ ShenandoahLoadRefBarrierStubC2* const stub = ShenandoahLoadRefBarrierStubC2::create(node, obj, addr, tmp, noreg, noreg, narrow);
+
+ // Don't preserve the obj across the runtime call, we override it from the return value anyway.
+ stub->dont_preserve(obj);
+
+ // Check if GC marking is in progress or we are handling a weak reference,
+ // otherwise we don't have to do anything. The code below was optimized to
+ // use less registers and instructions as possible at the expense of always
+ // having a branch instruction. The reason why we use this particular branch
+ // scheme is because the stub entry may be too far for the tbnz to jump to.
+ bool is_strong = (node->barrier_data() & ShenandoahBarrierStrong) != 0;
+ if (is_strong) {
+ __ tbz(gc_state, ShenandoahHeap::HAS_FORWARDED_BITPOS, *stub->continuation());
+ __ b(*stub->entry());
+ } else {
+ static_assert(ShenandoahHeap::HAS_FORWARDED_BITPOS == 0, "Relied on in LRB check below.");
+ __ orr(tmp, gc_state, gc_state, Assembler::LSR, ShenandoahHeap::WEAK_ROOTS_BITPOS);
+ __ tbz(tmp, ShenandoahHeap::HAS_FORWARDED_BITPOS, *stub->continuation());
+ __ b(*stub->entry());
+ }
+
+ __ bind(*stub->continuation());
+ BLOCK_COMMENT("} load_ref_barrier_c3");
+ }
+
+ void ShenandoahBarrierSetAssembler::satb_barrier_c3(const MachNode* node, MacroAssembler* masm, Register addr, Register pre_val, Register gc_state) {
+ assert_different_registers(addr, pre_val);
+ if (!ShenandoahSATBBarrierStubC2::needs_barrier(node)) {
+ return;
+ }
+
+ Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
+ ShenandoahSATBBarrierStubC2* const stub = ShenandoahSATBBarrierStubC2::create(node, addr, pre_val);
+
+ // Check if GC marking is in progress, otherwise we don't have to do anything.
+ __ tstw(gc_state, ShenandoahHeap::MARKING);
+ __ br(Assembler::NE, *stub->entry());
+ __ bind(*stub->continuation());
+ }
+
+ void ShenandoahBarrierSetAssembler::satb_barrier_c2(const MachNode* node, MacroAssembler* masm, Register addr, Register pre_val) {
+ assert_different_registers(addr, pre_val);
+ if (!ShenandoahSATBBarrierStubC2::needs_barrier(node)) {
+ return;
+ }
+ Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
+ ShenandoahSATBBarrierStubC2* const stub = ShenandoahSATBBarrierStubC2::create(node, addr, pre_val);
+
+ // Check if GC marking is in progress, otherwise we don't have to do anything.
+ Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
+ __ ldrb(rscratch1, gc_state);
+ __ tstw(rscratch1, ShenandoahHeap::MARKING);
+ __ br(Assembler::NE, *stub->entry());
+ __ bind(*stub->continuation());
+ }
+
+ void ShenandoahBarrierSetAssembler::card_barrier_c2(const MachNode* node, MacroAssembler* masm, Register addr, Register tmp) {
+ if (!ShenandoahCardBarrier ||
+ (node->barrier_data() & (ShenandoahBarrierCardMark | ShenandoahBarrierCardMarkNotNull)) == 0) {
+ return;
+ }
+
+ Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
+ __ lsr(tmp, addr, CardTable::card_shift());
+
+ assert(CardTable::dirty_card_val() == 0, "must be");
+
+ Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
+ __ ldr(rscratch1, curr_ct_holder_addr);
+
+ if (UseCondCardMark) {
+ Label L_already_dirty;
+ __ ldrb(rscratch2, Address(tmp, rscratch1));
+ __ cbz(rscratch2, L_already_dirty);
+ __ strb(zr, Address(tmp, rscratch1));
+ __ bind(L_already_dirty);
+ } else {
+ __ strb(zr, Address(tmp, rscratch1));
+ }
+ }
+
+ void ShenandoahBarrierSetAssembler::cmpxchg_oop_c2(const MachNode* node,
+ MacroAssembler* masm,
+ Register addr, Register oldval,
+ Register newval, Register res,
+ Register tmp1, Register tmp2,
+ bool acquire, bool release, bool weak, bool exchange) {
+ BLOCK_COMMENT("cmpxchg_oop_c2 {");
+ assert(res != noreg, "need result register");
+ assert_different_registers(oldval, addr, res, tmp1, tmp2);
+ assert_different_registers(newval, addr, res, tmp1, tmp2);
+
+ // Fast-path: Try to CAS optimistically. If successful, then we are done.
+ // EQ flag set iff success. 'tmp2' holds value fetched.
+ Assembler::operand_size size = UseCompressedOops ? Assembler::word : Assembler::xword;
+ __ cmpxchg(addr, oldval, newval, size, acquire, release, weak, tmp2);
+
+ // If we need a boolean result out of CAS, set the flag appropriately. This
+ // would be the final result if we do not go slow.
+ if (!exchange) {
+ __ cset(res, Assembler::EQ);
+ } else {
+ __ mov(res, tmp2);
+ }
+
+ if (ShenandoahCASBarrier) {
+ ShenandoahCASBarrierSlowStubC2* const slow_stub =
+ ShenandoahCASBarrierSlowStubC2::create(node, addr, oldval, newval, res, tmp1, tmp2, exchange, acquire, release, weak);
+
+ slow_stub->dont_preserve(res); // set at the end, no need to save
+ slow_stub->dont_preserve(oldval); // saved explicitly
+ slow_stub->dont_preserve(tmp1); // temp, no need to save
+ slow_stub->dont_preserve(tmp2); // temp, no need to save
+
+ // On success, we do not need any additional handling.
+ __ br(Assembler::EQ, *slow_stub->continuation());
+
+ // If GC is in progress, it is likely we need additional handling for false negatives.
+ Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
+ __ ldrb(tmp1, gc_state);
+ __ tbz(tmp1, ShenandoahHeap::HAS_FORWARDED_BITPOS, *slow_stub->continuation());
+ __ b(*slow_stub->entry());
+
+ // Slow stub re-enters with result set correctly.
+ __ bind(*slow_stub->continuation());
+ }
+
+ BLOCK_COMMENT("} cmpxchg_oop_c2");
+ }
+
+ #undef __
+ #define __ masm.
+
+ void ShenandoahLoadRefBarrierStubC2::emit_code(MacroAssembler& masm) {
+ BLOCK_COMMENT("ShenandoahLoadRefBarrierStubC2::emit_code {");
+ Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
+ __ bind(*entry());
+ Register obj = _obj;
+ if (_narrow) {
+ __ decode_heap_oop(_tmp1, _obj);
+ obj = _tmp1;
+ }
+ // Weak/phantom loads always need to go to runtime.
+ if ((_node->barrier_data() & ShenandoahBarrierStrong) != 0) {
+ // Check for object in cset.
+ __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
+ __ lsr(rscratch1, obj, ShenandoahHeapRegion::region_size_bytes_shift_jint());
+ __ ldrb(rscratch2, Address(rscratch2, rscratch1));
+ __ cbz(rscratch2, *continuation());
+ }
+ {
+ SaveLiveRegisters save_registers(&masm, this);
+ if (c_rarg0 != obj) {
+ if (c_rarg0 == _addr) {
+ __ mov(rscratch1, _addr);
+ _addr = rscratch1;
+ }
+ __ mov(c_rarg0, obj);
+ }
+ __ mov(c_rarg1, _addr);
+
+ if (_narrow) {
+ if ((_node->barrier_data() & ShenandoahBarrierStrong) != 0) {
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow));
+ } else if ((_node->barrier_data() & ShenandoahBarrierWeak) != 0) {
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow));
+ } else if ((_node->barrier_data() & ShenandoahBarrierPhantom) != 0) {
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow));
+ }
+ } else {
+ if ((_node->barrier_data() & ShenandoahBarrierStrong) != 0) {
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong));
+ } else if ((_node->barrier_data() & ShenandoahBarrierWeak) != 0) {
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak));
+ } else if ((_node->barrier_data() & ShenandoahBarrierPhantom) != 0) {
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom));
+ }
+ }
+ __ blr(rscratch1);
+ __ mov(_obj, r0);
+ }
+ if (_narrow) {
+ __ encode_heap_oop(_obj);
+ }
+ __ b(*continuation());
+ BLOCK_COMMENT("} ShenandoahLoadRefBarrierStubC2::emit_code");
+ }
+
+ void ShenandoahSATBBarrierStubC2::emit_code(MacroAssembler& masm) {
+ Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
+ __ bind(*entry());
+ // Do we need to load the previous value?
+ if (_addr != noreg) {
+ __ load_heap_oop(_preval, Address(_addr, 0), noreg, noreg, AS_RAW);
+ }
+
+ Address index(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
+ Address buffer(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
+ Label runtime;
+ __ ldr(rscratch1, index);
+ // If buffer is full, call into runtime.
+ __ cbz(rscratch1, runtime);
+
+ // The buffer is not full, store value into it.
+ __ sub(rscratch1, rscratch1, wordSize);
+ __ str(rscratch1, index);
+ __ ldr(rscratch2, buffer);
+ __ str(_preval, Address(rscratch2, rscratch1));
+ __ b(*continuation());
+
+ // Runtime call
+ __ bind(runtime);
+ {
+ SaveLiveRegisters save_registers(&masm, this);
+ if (c_rarg0 != _preval) {
+ __ mov(c_rarg0, _preval);
+ }
+ __ mov(rscratch1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre_c2));
+ __ blr(rscratch1);
+ }
+ __ b(*continuation());
+ }
+
+ void ShenandoahCASBarrierMidStubC2::emit_code(MacroAssembler& masm) {
+ Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
+ __ bind(*entry());
+
+ // Check if CAS result is null. If it is, then we must have a legitimate failure.
+ // This makes loading the fwdptr in the slow-path simpler.
+ __ tst(_result, _result);
+ // In case of !CAE, this has the correct value for legitimate failure (0/false)
+ // in result register.
+ __ br(Assembler::EQ, *continuation());
+
+ // Check if GC is in progress, otherwise we must have a legitimate failure.
+ Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
+ __ ldrb(_tmp, gc_state);
+ __ tstw(_tmp, ShenandoahHeap::HAS_FORWARDED);
+ __ br(Assembler::NE, *_slow_stub->entry());
+
+ if (!_cae) {
+ __ mov(_result, 0); // result = false
+ }
+ __ b(*continuation());
+ }
+
+ void ShenandoahCASBarrierSlowStubC2::emit_code(MacroAssembler& masm) {
+ __ bind(*entry());
+
+ // CAS has failed because the value held at addr does not match expected.
+ // This may be a false negative because the version in memory might be
+ // the from-space version of the same object we currently hold to-space
+ // reference for.
+ //
+ // To resolve this, we need to pass the location through the LRB fixup,
+ // this will make sure that the location has only to-space pointers.
+ // To avoid calling into runtime often, we cset-check the object first.
+ // We can inline most of the work here, but there is little point,
+ // as CAS failures over cset locations must be rare. This fast-slow split
+ // matches what we do for normal LRB.
+
+ // Non-strong references should always go to runtime. We do not expect
+ // CASes over non-strong locations.
+ assert((_node->barrier_data() & ShenandoahBarrierStrong) != 0, "Only strong references for CASes");
+
+ Label L_final;
+
+ // (Compressed) failure witness is in _tmp2.
+ // Unpack it and check if it is in collection set.
+ // We need to backup the compressed version to use in the LRB.
+ __ mov(_result, _tmp2);
+ if (UseCompressedOops) {
+ __ decode_heap_oop(_tmp2);
+ }
+
+ __ mov(_tmp1, ShenandoahHeap::in_cset_fast_test_addr());
+ __ lsr(_tmp2, _tmp2, ShenandoahHeapRegion::region_size_bytes_shift_jint());
+ __ ldrb(_tmp1, Address(_tmp1, _tmp2));
+ __ cbz(_tmp1, L_final);
+
+ {
+ SaveLiveRegisters save_registers(&masm, this);
+ // Load up failure witness again.
+ __ mov(c_rarg0, _result);
+ if (UseCompressedOops) {
+ __ decode_heap_oop(c_rarg0);
+ }
+ __ mov(c_rarg1, _addr_reg);
+
+ if (UseCompressedOops) {
+ __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow), 2);
+ } else {
+ __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong), 2);
+ }
+ // We have called LRB to fix up the heap location. We do not care about its
+ // result, as we will just try to CAS the location again.
+ }
+
+ __ bind(L_final);
+
+ Assembler::operand_size size = UseCompressedOops ? Assembler::word : Assembler::xword;
+ __ cmpxchg(_addr_reg, _expected, _new_val, size, _acquire, _release, _weak, _result);
+
+ if (!_cae) {
+ __ cset(_result, Assembler::EQ);
+ }
+ __ b(*continuation());
+ }
+ #undef __
+ #define __ masm->
+ #endif // COMPILER2
+
void ShenandoahBarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
Register start, Register count, Register scratch) {
assert(ShenandoahCardBarrier, "Should have been checked by caller");
Label L_loop, L_done;
< prev index next >