8291302: ARM32: nmethod entry barriers support

src/hotspot/cpu/arm/arm.ad

@@ -59,6 +59,9 @@ source_hpp %{
 // To keep related declarations/definitions/uses close together,
 // we switch between source %{ }% and source_hpp %{ }% freely as needed.
 
+#include "asm/macroAssembler.hpp"
+#include "gc/shared/barrierSetAssembler.hpp"
+
 // Does destination need to be loaded in a register then passed to a
 // branch instruction?
 extern bool maybe_far_call(const CallNode *n);
@@ -286,6 +289,17 @@ void MachPrologNode::format( PhaseRegAlloc *ra_, outputStream *st ) const {
   if (framesize != 0) {
     st->print   ("SUB    R_SP, R_SP, " SIZE_FORMAT,framesize);
   }
+
+  if (C->stub_function() == NULL && BarrierSet::barrier_set()->barrier_set_nmethod() != NULL) {
+    st->print("ldr  t0, [guard]\n\t");
+    st->print("ldr t1, [Rthread, #thread_disarmed_offset]\n\t");
+    st->print("cmp t0, t1\n\t");
+    st->print("beq skip\n\t");
+    st->print("blr #nmethod_entry_barrier_stub\n\t");
+    st->print("b skip\n\t");
+    st->print("guard: int\n\t");
+    st->print("skip:\n\t");
+  }
 }
 #endif
 
@@ -318,6 +332,11 @@ void MachPrologNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
     __ sub_slow(SP, SP, framesize);
   }
 
+  if (C->stub_function() == NULL) {
+    BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
+    bs->nmethod_entry_barrier(&_masm);
+  }
+
   // offset from scratch buffer is not valid
   if (strcmp(cbuf.name(), "Compile::Fill_buffer") == 0) {
     C->output()->set_frame_complete( __ offset() );

src/hotspot/cpu/arm/c1_MacroAssembler_arm.cpp

@@ -25,6 +25,8 @@
 #include "precompiled.hpp"
 #include "c1/c1_MacroAssembler.hpp"
 #include "c1/c1_Runtime1.hpp"
+#include "gc/shared/barrierSet.hpp"
+#include "gc/shared/barrierSetAssembler.hpp"
 #include "gc/shared/collectedHeap.hpp"
 #include "gc/shared/tlab_globals.hpp"
 #include "interpreter/interpreter.hpp"
@@ -62,6 +64,10 @@ void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_by
   // if this method contains a methodHandle call site
   raw_push(FP, LR);
   sub_slow(SP, SP, frame_size_in_bytes);
+
+  // Insert nmethod entry barrier into frame.
+  BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
+  bs->nmethod_entry_barrier(this);
 }
 
 void C1_MacroAssembler::remove_frame(int frame_size_in_bytes) {

src/hotspot/cpu/arm/gc/shared/barrierSetAssembler_arm.cpp

@@ -23,10 +23,13 @@
  */
 
 #include "precompiled.hpp"
+#include "gc/shared/barrierSet.hpp"
 #include "gc/shared/barrierSetAssembler.hpp"
+#include "gc/shared/barrierSetNMethod.hpp"
 #include "gc/shared/collectedHeap.hpp"
 #include "memory/universe.hpp"
 #include "runtime/javaThread.hpp"
+#include "runtime/stubRoutines.hpp"
 
 #define __ masm->
 
@@ -195,3 +198,47 @@ void BarrierSetAssembler::incr_allocated_bytes(MacroAssembler* masm, RegisterOrC
   // Unborrow the Rthread
   __ sub(Rthread, Ralloc, in_bytes(JavaThread::allocated_bytes_offset()));
 }
+
+void BarrierSetAssembler::nmethod_entry_barrier(MacroAssembler* masm) {
+
+  BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
+
+  Register tmp0 = Rtemp;
+  Register tmp1 = R5; // wrong register here will lead to crash
+
+  if (bs_nm == NULL) {
+    return;
+  }
+
+  // The are no GCs that require memory barrier on arm32 now
+#ifndef PRODUCT
+  NMethodPatchingType patching_type = nmethod_patching_type();
+  assert(patching_type == NMethodPatchingType::stw_instruction_and_data_patch, "Unsupported patching type");
+#endif
+
+  Label skip, guard;
+  Address thread_disarmed_addr(Rthread, in_bytes(bs_nm->thread_disarmed_offset()));
+
+  __ block_comment("nmethod_barrier begin");
+  __ ldr_label(tmp0, guard);
+
+  // No memory barrier here
+  __ ldr(tmp1, thread_disarmed_addr);
+  __ cmp(tmp0, tmp1);
+  __ b(skip, eq);
+
+  __ mov_address(tmp0, StubRoutines::Arm::method_entry_barrier());
+  __ call(tmp0);
+  __ b(skip);
+
+  __ bind(guard);
+
+  // nmethod guard value. Skipped over in common case.
+  //
+  // Put a debug value to make any offsets skew
+  // clearly visible in coredump
+  __ emit_int32(0xDEADBEAF);
+
+  __ bind(skip);
+  __ block_comment("nmethod_barrier end");
+}

src/hotspot/cpu/arm/gc/shared/barrierSetAssembler_arm.hpp

@@ -29,6 +29,10 @@
 #include "memory/allocation.hpp"
 #include "oops/access.hpp"
 
+enum class NMethodPatchingType {
+  stw_instruction_and_data_patch,
+};
+
 class BarrierSetAssembler: public CHeapObj<mtGC> {
 private:
   void incr_allocated_bytes(MacroAssembler* masm,
@@ -56,6 +60,8 @@ class BarrierSetAssembler: public CHeapObj<mtGC> {
   );
 
   virtual void barrier_stubs_init() {}
+  virtual NMethodPatchingType nmethod_patching_type() { return NMethodPatchingType::stw_instruction_and_data_patch; }
+  virtual void nmethod_entry_barrier(MacroAssembler* masm);
 };
 
 #endif // CPU_ARM_GC_SHARED_BARRIERSETASSEMBLER_ARM_HPP

src/hotspot/cpu/arm/gc/shared/barrierSetNMethod_arm.cpp

@@ -23,18 +23,124 @@
  */
 
 #include "precompiled.hpp"
+#include "code/nativeInst.hpp"
+#include "gc/shared/barrierSetAssembler.hpp"
 #include "gc/shared/barrierSetNMethod.hpp"
+#include "logging/log.hpp"
+#include "memory/resourceArea.hpp"
+#include "runtime/frame.inline.hpp"
+#include "runtime/javaThread.hpp"
+#include "runtime/sharedRuntime.hpp"
+#include "runtime/registerMap.hpp"
+#include "utilities/align.hpp"
 #include "utilities/debug.hpp"
 
+// The constant below reflects the size of the barrier
+// in barrierSetAssembler_arm.cpp
+static const int entry_barrier_bytes = 36;
+
+class NativeNMethodBarrier: public NativeInstruction {
+  address instruction_address() const { return addr_at(0); }
+
+  int *guard_addr() const {
+    // Last instruction in a barrier
+    return reinterpret_cast<int*>(instruction_address() + entry_barrier_bytes - wordSize);
+  }
+
+public:
+  int get_value() {
+    return Atomic::load_acquire(guard_addr());
+  }
+
+  void set_value(int value) {
+    Atomic::release_store(guard_addr(), value);
+  }
+
+  void verify() const;
+};
+
+// Check the first instruction of the nmethod entry barrier
+// to make sure that the offsets are not skewed.
+void NativeNMethodBarrier::verify() const {
+  NativeInstruction *ni = (NativeInstruction *) instruction_address();
+  if (!ni->is_ldr()) {
+    uint32_t *addr = (uint32_t *) ni;
+    tty->print_cr("Addr: " INTPTR_FORMAT " Code: 0x%x", (intptr_t) addr, (uint32_t) *addr);
+    fatal("not an ldr instruction.");
+  }
+}
+
+static NativeNMethodBarrier* native_nmethod_barrier(nmethod* nm) {
+  address barrier_address = nm->code_begin() + nm->frame_complete_offset() - entry_barrier_bytes;
+  NativeNMethodBarrier* barrier = reinterpret_cast<NativeNMethodBarrier*>(barrier_address);
+  debug_only(barrier->verify());
+  return barrier;
+}
+
+/* We're called from an nmethod when we need to deoptimize it. We do
+   this by throwing away the nmethod's frame and jumping to the
+   ic_miss stub. This looks like there has been an IC miss at the
+   entry of the nmethod, so we resolve the call, which will fall back
+   to the interpreter if the nmethod has been unloaded. */
 void BarrierSetNMethod::deoptimize(nmethod* nm, address* return_address_ptr) {
-  ShouldNotReachHere();
+
+  typedef struct {
+    intptr_t *sp; intptr_t *fp; address lr; address pc;
+  } frame_pointers_t;
+
+  frame_pointers_t *new_frame = (frame_pointers_t *)(return_address_ptr - 5);
+
+  JavaThread *thread = JavaThread::current();
+  RegisterMap reg_map(thread,
+                      RegisterMap::UpdateMap::skip,
+                      RegisterMap::ProcessFrames::include,
+                      RegisterMap::WalkContinuation::skip);
+  frame frame = thread->last_frame();
+
+  assert(frame.is_compiled_frame() || frame.is_native_frame(), "must be");
+  assert(frame.cb() == nm, "must be");
+  frame = frame.sender(&reg_map);
+
+  LogTarget(Trace, nmethod, barrier) out;
+  if (out.is_enabled()) {
+    ResourceMark mark;
+    log_trace(nmethod, barrier)("deoptimize(nmethod: %s(%p), return_addr: %p, osr: %d, thread: %p(%s), making rsp: %p) -> %p",
+                                nm->method()->name_and_sig_as_C_string(),
+                                nm, *(address *) return_address_ptr, nm->is_osr_method(), thread,
+                                thread->name(), frame.sp(), nm->verified_entry_point());
+  }
+
+  new_frame->sp = frame.sp();
+  new_frame->fp = frame.fp();
+  new_frame->lr = frame.pc();
+  new_frame->pc = SharedRuntime::get_handle_wrong_method_stub();
 }
 
 void BarrierSetNMethod::disarm(nmethod* nm) {
-  ShouldNotReachHere();
+  if (!supports_entry_barrier(nm)) {
+    return;
+  }
+
+  // Disarms the nmethod guard emitted by BarrierSetAssembler::nmethod_entry_barrier.
+  // Symmetric "LDR; DMB ISHLD" is in the nmethod barrier.
+  NativeNMethodBarrier* barrier = native_nmethod_barrier(nm);
+  barrier->set_value(disarmed_value());
+}
+
+void BarrierSetNMethod::arm(nmethod* nm, int arm_value) {
+  if (!supports_entry_barrier(nm)) {
+    return;
+  }
+
+  NativeNMethodBarrier* barrier = native_nmethod_barrier(nm);
+  barrier->set_value(arm_value);
 }
 
 bool BarrierSetNMethod::is_armed(nmethod* nm) {
-  ShouldNotReachHere();
-  return false;
+  if (!supports_entry_barrier(nm)) {
+    return false;
+  }
+
+  NativeNMethodBarrier* barrier = native_nmethod_barrier(nm);
+  return barrier->get_value() != disarmed_value();
 }

src/hotspot/cpu/arm/macroAssembler_arm.hpp

@@ -587,9 +587,23 @@ class MacroAssembler: public Assembler {
     AbstractAssembler::emit_address((address)L.data());
   }
 
+  void ldr_label(Register rd, Label& L) {
+    ldr(rd, Address(PC, target(L) - pc() - 8));
+  }
+
   void resolve_oop_handle(Register result);
   void load_mirror(Register mirror, Register method, Register tmp);
 
+  void enter() {
+    raw_push(FP, LR);
+    mov(FP, SP);
+  }
+
+  void leave() {
+    mov(SP, FP);
+    raw_pop(FP, LR);
+  }
+
 #define ARM_INSTR_1(common_mnemonic, arm32_mnemonic, arg_type) \
   void common_mnemonic(arg_type arg) { \
       arm32_mnemonic(arg); \

src/hotspot/cpu/arm/sharedRuntime_arm.cpp

@@ -28,6 +28,7 @@
 #include "code/icBuffer.hpp"
 #include "code/vtableStubs.hpp"
 #include "compiler/oopMap.hpp"
+#include "gc/shared/barrierSetAssembler.hpp"
 #include "interpreter/interpreter.hpp"
 #include "logging/log.hpp"
 #include "memory/resourceArea.hpp"
@@ -873,6 +874,10 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
   __ mov(FP, SP);
   __ sub_slow(SP, SP, stack_size - 2*wordSize);
 
+  BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
+  assert(bs != NULL, "Sanity");
+  bs->nmethod_entry_barrier(masm);
+
   int frame_complete = __ pc() - start;
 
   OopMapSet* oop_maps = new OopMapSet();

src/hotspot/cpu/arm/stubGenerator_arm.cpp

@@ -27,6 +27,7 @@
 #include "compiler/oopMap.hpp"
 #include "gc/shared/barrierSet.hpp"
 #include "gc/shared/barrierSetAssembler.hpp"
+#include "gc/shared/barrierSetNMethod.hpp"
 #include "interpreter/interpreter.hpp"
 #include "memory/universe.hpp"
 #include "nativeInst_arm.hpp"
@@ -2905,6 +2906,53 @@ class StubGenerator: public StubCodeGenerator {
 
   }
 
+  address generate_method_entry_barrier() {
+    __ align(CodeEntryAlignment);
+    StubCodeMark mark(this, "StubRoutines", "nmethod_entry_barrier");
+
+    Label deoptimize_label;
+
+    address start = __ pc();
+
+    // No need to save PC on Arm
+    __ set_last_Java_frame(SP, FP, false, Rtemp);
+
+    __ enter();
+
+    __ add(Rtemp, SP, wordSize);  // Rtemp points to the saved lr
+    __ sub(SP, SP, 4 * wordSize); // four words for the returned {sp, fp, lr, pc}
+
+    const RegisterSet saved_regs = RegisterSet(R0, R10);
+    __ push(saved_regs);
+    __ fpush(FloatRegisterSet(D0, 16));
+
+    __ mov(c_rarg0, Rtemp);
+    __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSetNMethod::nmethod_stub_entry_barrier), c_rarg0);
+
+    __ reset_last_Java_frame(Rtemp);
+
+    __ mov(Rtemp, R0);
+
+    __ fpop(FloatRegisterSet(D0, 16));
+    __ pop(saved_regs);
+
+    __ cbnz(Rtemp, deoptimize_label);
+
+    __ leave();
+    __ bx(LR);
+
+    __ BIND(deoptimize_label);
+
+    __ ldr(Rtemp, Address(SP, 0));
+    __ ldr(FP, Address(SP, wordSize));
+    __ ldr(LR, Address(SP, wordSize * 2));
+    __ ldr(R5, Address(SP, wordSize * 3));
+    __ mov(SP, Rtemp);
+    __ bx(R5);
+
+    return start;
+  }
+
 #define COMPILE_CRYPTO
 #include "stubRoutinesCrypto_arm.cpp"
 
@@ -3097,6 +3145,11 @@ class StubGenerator: public StubCodeGenerator {
     // arraycopy stubs used by compilers
     generate_arraycopy_stubs();
 
+    BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
+    if (bs_nm != NULL) {
+      StubRoutines::Arm::_method_entry_barrier = generate_method_entry_barrier();
+    }
+
 #ifdef COMPILE_CRYPTO
     // generate AES intrinsics code
     if (UseAESIntrinsics) {

src/hotspot/cpu/arm/stubRoutines_arm.cpp

@@ -33,3 +33,5 @@ address StubRoutines::Arm::_partial_subtype_check = NULL;
 
 address StubRoutines::_atomic_load_long_entry = NULL;
 address StubRoutines::_atomic_store_long_entry = NULL;
+
+address StubRoutines::Arm::_method_entry_barrier = NULL;