8322630: Remove ICStubs and related safepoints

Co-authored-by: Martin Doerr <mdoerr@openjdk.org>
Co-authored-by: Aleksey Shipilev <shade@openjdk.org>
Co-authored-by: Amit Kumar <amitkumar@openjdk.org>
Co-authored-by: Robbin Ehn <rehn@openjdk.org>
Co-authored-by: Aleksei Voitylov <avoitylov@openjdk.org>
Reviewed-by: tschatzl, aboldtch, dlong

Author: 6 people

src/hotspot/cpu/aarch64/aarch64.ad

@@ -2205,14 +2205,14 @@ void MachUEPNode::format(PhaseRegAlloc* ra_, outputStream* st) const
 {
   st->print_cr("# MachUEPNode");
   if (UseCompressedClassPointers) {
-    st->print_cr("\tldrw rscratch1, j_rarg0 + oopDesc::klass_offset_in_bytes()]\t# compressed klass");
-    if (CompressedKlassPointers::shift() != 0) {
-      st->print_cr("\tdecode_klass_not_null rscratch1, rscratch1");
-    }
+    st->print_cr("\tldrw rscratch1, [j_rarg0 + oopDesc::klass_offset_in_bytes()]\t# compressed klass");
+    st->print_cr("\tldrw r10, [rscratch2 + CompiledICData::speculated_klass_offset()]\t# compressed klass");
+    st->print_cr("\tcmpw rscratch1, r10");
   } else {
-   st->print_cr("\tldr rscratch1, j_rarg0 + oopDesc::klass_offset_in_bytes()]\t# compressed klass");
+    st->print_cr("\tldr rscratch1, [j_rarg0 + oopDesc::klass_offset_in_bytes()]\t# compressed klass");
+    st->print_cr("\tldr r10, [rscratch2 + CompiledICData::speculated_klass_offset()]\t# compressed klass");
+    st->print_cr("\tcmp rscratch1, r10");
   }
-  st->print_cr("\tcmp r0, rscratch1\t # Inline cache check");
   st->print_cr("\tbne, SharedRuntime::_ic_miss_stub");
 }
 #endif
@@ -2221,14 +2221,7 @@ void MachUEPNode::emit(CodeBuffer& cbuf, PhaseRegAlloc* ra_) const
 {
   // This is the unverified entry point.
   C2_MacroAssembler _masm(&cbuf);
-
-  __ cmp_klass(j_rarg0, rscratch2, rscratch1);
-  Label skip;
-  // TODO
-  // can we avoid this skip and still use a reloc?
-  __ br(Assembler::EQ, skip);
-  __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
-  __ bind(skip);
+  __ ic_check(InteriorEntryAlignment);
 }
 
 uint MachUEPNode::size(PhaseRegAlloc* ra_) const
@@ -3715,7 +3708,7 @@ encode %{
         cbuf.shared_stub_to_interp_for(_method, call - cbuf.insts_begin());
       } else {
         // Emit stub for static call
-        address stub = CompiledStaticCall::emit_to_interp_stub(cbuf, call);
+        address stub = CompiledDirectCall::emit_to_interp_stub(cbuf, call);
         if (stub == nullptr) {
           ciEnv::current()->record_failure("CodeCache is full");
           return;

src/hotspot/cpu/aarch64/c1_LIRAssembler_aarch64.cpp

@@ -53,7 +53,6 @@
 #endif
 
 NEEDS_CLEANUP // remove this definitions ?
-const Register IC_Klass    = rscratch2;   // where the IC klass is cached
 const Register SYNC_header = r0;   // synchronization header
 const Register SHIFT_count = r0;   // where count for shift operations must be
 
@@ -293,27 +292,7 @@ void LIR_Assembler::osr_entry() {
 
 // inline cache check; done before the frame is built.
 int LIR_Assembler::check_icache() {
-  Register receiver = FrameMap::receiver_opr->as_register();
-  Register ic_klass = IC_Klass;
-  int start_offset = __ offset();
-  __ inline_cache_check(receiver, ic_klass);
-
-  // if icache check fails, then jump to runtime routine
-  // Note: RECEIVER must still contain the receiver!
-  Label dont;
-  __ br(Assembler::EQ, dont);
-  __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
-
-  // We align the verified entry point unless the method body
-  // (including its inline cache check) will fit in a single 64-byte
-  // icache line.
-  if (! method()->is_accessor() || __ offset() - start_offset > 4 * 4) {
-    // force alignment after the cache check.
-    __ align(CodeEntryAlignment);
-  }
-
-  __ bind(dont);
-  return start_offset;
+  return __ ic_check(CodeEntryAlignment);
 }
 
 void LIR_Assembler::clinit_barrier(ciMethod* method) {
@@ -2042,7 +2021,7 @@ void LIR_Assembler::emit_static_call_stub() {
   __ relocate(static_stub_Relocation::spec(call_pc));
   __ emit_static_call_stub();
 
-  assert(__ offset() - start + CompiledStaticCall::to_trampoline_stub_size()
+  assert(__ offset() - start + CompiledDirectCall::to_trampoline_stub_size()
         <= call_stub_size(), "stub too big");
   __ end_a_stub();
 }

src/hotspot/cpu/aarch64/c1_LIRAssembler_aarch64.hpp

@@ -71,8 +71,8 @@ friend class ArrayCopyStub;
   void deoptimize_trap(CodeEmitInfo *info);
 
   enum {
-    // call stub: CompiledStaticCall::to_interp_stub_size() +
-    //            CompiledStaticCall::to_trampoline_stub_size()
+    // call stub: CompiledDirectCall::to_interp_stub_size() +
+    //            CompiledDirectCall::to_trampoline_stub_size()
     _call_stub_size = 13 * NativeInstruction::instruction_size,
     _exception_handler_size = DEBUG_ONLY(1*K) NOT_DEBUG(175),
     _deopt_handler_size = 7 * NativeInstruction::instruction_size

src/hotspot/cpu/aarch64/c1_MacroAssembler_aarch64.cpp

@@ -308,17 +308,6 @@ void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1,
   verify_oop(obj);
 }
 
-
-void C1_MacroAssembler::inline_cache_check(Register receiver, Register iCache) {
-  verify_oop(receiver);
-  // explicit null check not needed since load from [klass_offset] causes a trap
-  // check against inline cache
-  assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
-
-  cmp_klass(receiver, iCache, rscratch1);
-}
-
-
 void C1_MacroAssembler::build_frame(int framesize, int bang_size_in_bytes) {
   assert(bang_size_in_bytes >= framesize, "stack bang size incorrect");
   // Make sure there is enough stack space for this method's activation.

src/hotspot/cpu/aarch64/c1_Runtime1_aarch64.cpp

@@ -38,7 +38,6 @@
 #include "interpreter/interpreter.hpp"
 #include "memory/universe.hpp"
 #include "nativeInst_aarch64.hpp"
-#include "oops/compiledICHolder.hpp"
 #include "oops/oop.inline.hpp"
 #include "prims/jvmtiExport.hpp"
 #include "register_aarch64.hpp"

src/hotspot/cpu/aarch64/compiledIC_aarch64.cpp

@@ -26,7 +26,6 @@
 #include "precompiled.hpp"
 #include "asm/macroAssembler.inline.hpp"
 #include "code/compiledIC.hpp"
-#include "code/icBuffer.hpp"
 #include "code/nmethod.hpp"
 #include "logging/log.hpp"
 #include "memory/resourceArea.hpp"
@@ -36,7 +35,7 @@
 // ----------------------------------------------------------------------------
 
 #define __ _masm.
-address CompiledStaticCall::emit_to_interp_stub(CodeBuffer &cbuf, address mark) {
+address CompiledDirectCall::emit_to_interp_stub(CodeBuffer &cbuf, address mark) {
   precond(cbuf.stubs()->start() != badAddress);
   precond(cbuf.stubs()->end() != badAddress);
 
@@ -71,33 +70,26 @@ address CompiledStaticCall::emit_to_interp_stub(CodeBuffer &cbuf, address mark)
 }
 #undef __
 
-int CompiledStaticCall::to_interp_stub_size() {
+int CompiledDirectCall::to_interp_stub_size() {
   return MacroAssembler::static_call_stub_size();
 }
 
-int CompiledStaticCall::to_trampoline_stub_size() {
+int CompiledDirectCall::to_trampoline_stub_size() {
   // Somewhat pessimistically, we count 3 instructions here (although
   // there are only two) because we sometimes emit an alignment nop.
   // Trampoline stubs are always word aligned.
   return MacroAssembler::max_trampoline_stub_size();
 }
 
 // Relocation entries for call stub, compiled java to interpreter.
-int CompiledStaticCall::reloc_to_interp_stub() {
+int CompiledDirectCall::reloc_to_interp_stub() {
   return 4; // 3 in emit_to_interp_stub + 1 in emit_call
 }
 
-void CompiledDirectStaticCall::set_to_interpreted(const methodHandle& callee, address entry) {
+void CompiledDirectCall::set_to_interpreted(const methodHandle& callee, address entry) {
   address stub = find_stub();
   guarantee(stub != nullptr, "stub not found");
 
-  {
-    ResourceMark rm;
-    log_trace(inlinecache)("CompiledDirectStaticCall@" INTPTR_FORMAT ": set_to_interpreted %s",
-                  p2i(instruction_address()),
-                  callee->name_and_sig_as_C_string());
-  }
-
   // Creation also verifies the object.
   NativeMovConstReg* method_holder
     = nativeMovConstReg_at(stub + NativeInstruction::instruction_size);
@@ -115,7 +107,7 @@ void CompiledDirectStaticCall::set_to_interpreted(const methodHandle& callee, ad
   set_destination_mt_safe(stub);
 }
 
-void CompiledDirectStaticCall::set_stub_to_clean(static_stub_Relocation* static_stub) {
+void CompiledDirectCall::set_stub_to_clean(static_stub_Relocation* static_stub) {
   // Reset stub.
   address stub = static_stub->addr();
   assert(stub != nullptr, "stub not found");
@@ -132,7 +124,7 @@ void CompiledDirectStaticCall::set_stub_to_clean(static_stub_Relocation* static_
 // Non-product mode code
 #ifndef PRODUCT
 
-void CompiledDirectStaticCall::verify() {
+void CompiledDirectCall::verify() {
   // Verify call.
   _call->verify();
   _call->verify_alignment();

src/hotspot/cpu/aarch64/icBuffer_aarch64.cpp

@@ -29,6 +29,7 @@
 #include "asm/assembler.hpp"
 #include "asm/assembler.inline.hpp"
 #include "ci/ciEnv.hpp"
+#include "code/compiledIC.hpp"
 #include "compiler/compileTask.hpp"
 #include "compiler/disassembler.hpp"
 #include "compiler/oopMap.hpp"
@@ -965,7 +966,7 @@ int MacroAssembler::max_trampoline_stub_size() {
 }
 
 void MacroAssembler::emit_static_call_stub() {
-  // CompiledDirectStaticCall::set_to_interpreted knows the
+  // CompiledDirectCall::set_to_interpreted knows the
   // exact layout of this stub.
 
   isb();
@@ -995,10 +996,51 @@ address MacroAssembler::ic_call(address entry, jint method_index) {
   // address const_ptr = long_constant((jlong)Universe::non_oop_word());
   // uintptr_t offset;
   // ldr_constant(rscratch2, const_ptr);
-  movptr(rscratch2, (uintptr_t)Universe::non_oop_word());
+  movptr(rscratch2, (intptr_t)Universe::non_oop_word());
   return trampoline_call(Address(entry, rh));
 }
 
+int MacroAssembler::ic_check_size() {
+  if (target_needs_far_branch(CAST_FROM_FN_PTR(address, SharedRuntime::get_ic_miss_stub()))) {
+    return NativeInstruction::instruction_size * 7;
+  } else {
+    return NativeInstruction::instruction_size * 5;
+  }
+}
+
+int MacroAssembler::ic_check(int end_alignment) {
+  Register receiver = j_rarg0;
+  Register data = rscratch2;
+  Register tmp1 = rscratch1;
+  Register tmp2 = r10;
+
+  // The UEP of a code blob ensures that the VEP is padded. However, the padding of the UEP is placed
+  // before the inline cache check, so we don't have to execute any nop instructions when dispatching
+  // through the UEP, yet we can ensure that the VEP is aligned appropriately. That's why we align
+  // before the inline cache check here, and not after
+  align(end_alignment, offset() + ic_check_size());
+
+  int uep_offset = offset();
+
+  if (UseCompressedClassPointers) {
+    ldrw(tmp1, Address(receiver, oopDesc::klass_offset_in_bytes()));
+    ldrw(tmp2, Address(data, CompiledICData::speculated_klass_offset()));
+    cmpw(tmp1, tmp2);
+  } else {
+    ldr(tmp1, Address(receiver, oopDesc::klass_offset_in_bytes()));
+    ldr(tmp2, Address(data, CompiledICData::speculated_klass_offset()));
+    cmp(tmp1, tmp2);
+  }
+
+  Label dont;
+  br(Assembler::EQ, dont);
+  far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
+  bind(dont);
+  assert((offset() % end_alignment) == 0, "Misaligned verified entry point");
+
+  return uep_offset;
+}
+
 // Implementation of call_VM versions
 
 void MacroAssembler::call_VM(Register oop_result,
@@ -1100,7 +1142,14 @@ void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thr
 }
 
 void MacroAssembler::align(int modulus) {
-  while (offset() % modulus != 0) nop();
+  align(modulus, offset());
+}
+
+// Ensure that the code at target bytes offset from the current offset() is aligned
+// according to modulus.
+void MacroAssembler::align(int modulus, int target) {
+  int delta = target - offset();
+  while ((offset() + delta) % modulus != 0) nop();
 }
 
 void MacroAssembler::post_call_nop() {
@@ -1197,7 +1246,7 @@ void MacroAssembler::lookup_interface_method(Register recv_klass,
 }
 
 // Look up the method for a megamorphic invokeinterface call in a single pass over itable:
-// - check recv_klass (actual object class) is a subtype of resolved_klass from CompiledICHolder
+// - check recv_klass (actual object class) is a subtype of resolved_klass from CompiledICData
 // - find a holder_klass (class that implements the method) vtable offset and get the method from vtable by index
 // The target method is determined by <holder_klass, itable_index>.
 // The receiver klass is in recv_klass.

src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp

@@ -720,6 +720,7 @@ class MacroAssembler: public Assembler {
 
   // Alignment
   void align(int modulus);
+  void align(int modulus, int target);
 
   // nop
   void post_call_nop();
@@ -1247,6 +1248,8 @@ class MacroAssembler: public Assembler {
 
   // Emit the CompiledIC call idiom
   address ic_call(address entry, jint method_index = 0);
+  static int ic_check_size();
+  int ic_check(int end_alignment);
 
 public: