6812831: factor duplicated assembly code for megamorphic invokeinterface (for 6655638)

Summary: Code in vtableStubs and templateTable moved into MacroAssembler.
Reviewed-by: kvn

src/cpu/sparc/vm/assembler_sparc.cpp

@@ -2638,6 +2638,135 @@ RegisterConstant MacroAssembler::delayed_value(intptr_t* delayed_value_addr,
 }
 
 
+void MacroAssembler::regcon_inc_ptr( RegisterConstant& dest, RegisterConstant src, Register temp ) {
+  assert(dest.register_or_noreg() != G0, "lost side effect");
+  if ((src.is_constant() && src.as_constant() == 0) ||
+      (src.is_register() && src.as_register() == G0)) {
+    // do nothing
+  } else if (dest.is_register()) {
+    add(dest.as_register(), ensure_rs2(src, temp), dest.as_register());
+  } else if (src.is_constant()) {
+    intptr_t res = dest.as_constant() + src.as_constant();
+    dest = RegisterConstant(res); // side effect seen by caller
+  } else {
+    assert(temp != noreg, "cannot handle constant += register");
+    add(src.as_register(), ensure_rs2(dest, temp), temp);
+    dest = RegisterConstant(temp); // side effect seen by caller
+  }
+}
+
+void MacroAssembler::regcon_sll_ptr( RegisterConstant& dest, RegisterConstant src, Register temp ) {
+  assert(dest.register_or_noreg() != G0, "lost side effect");
+  if (!is_simm13(src.constant_or_zero()))
+    src = (src.as_constant() & 0xFF);
+  if ((src.is_constant() && src.as_constant() == 0) ||
+      (src.is_register() && src.as_register() == G0)) {
+    // do nothing
+  } else if (dest.is_register()) {
+    sll_ptr(dest.as_register(), src, dest.as_register());
+  } else if (src.is_constant()) {
+    intptr_t res = dest.as_constant() << src.as_constant();
+    dest = RegisterConstant(res); // side effect seen by caller
+  } else {
+    assert(temp != noreg, "cannot handle constant <<= register");
+    set(dest.as_constant(), temp);
+    sll_ptr(temp, src, temp);
+    dest = RegisterConstant(temp); // side effect seen by caller
+  }
+}
+
+
+// Look up the method for a megamorphic invokeinterface call.
+// The target method is determined by <intf_klass, itable_index>.
+// The receiver klass is in recv_klass.
+// On success, the result will be in method_result, and execution falls through.
+// On failure, execution transfers to the given label.
+void MacroAssembler::lookup_interface_method(Register recv_klass,
+                                             Register intf_klass,
+                                             RegisterConstant itable_index,
+                                             Register method_result,
+                                             Register scan_temp,
+                                             Register sethi_temp,
+                                             Label& L_no_such_interface) {
+  assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
+  assert(itable_index.is_constant() || itable_index.as_register() == method_result,
+         "caller must use same register for non-constant itable index as for method");
+
+  // Compute start of first itableOffsetEntry (which is at the end of the vtable)
+  int vtable_base = instanceKlass::vtable_start_offset() * wordSize;
+  int scan_step   = itableOffsetEntry::size() * wordSize;
+  int vte_size    = vtableEntry::size() * wordSize;
+
+  lduw(recv_klass, instanceKlass::vtable_length_offset() * wordSize, scan_temp);
+  // %%% We should store the aligned, prescaled offset in the klassoop.
+  // Then the next several instructions would fold away.
+
+  int round_to_unit = ((HeapWordsPerLong > 1) ? BytesPerLong : 0);
+  int itb_offset = vtable_base;
+  if (round_to_unit != 0) {
+    // hoist first instruction of round_to(scan_temp, BytesPerLong):
+    itb_offset += round_to_unit - wordSize;
+  }
+  int itb_scale = exact_log2(vtableEntry::size() * wordSize);
+  sll(scan_temp, itb_scale,  scan_temp);
+  add(scan_temp, itb_offset, scan_temp);
+  if (round_to_unit != 0) {
+    // Round up to align_object_offset boundary
+    // see code for instanceKlass::start_of_itable!
+    // Was: round_to(scan_temp, BytesPerLong);
+    // Hoisted: add(scan_temp, BytesPerLong-1, scan_temp);
+    and3(scan_temp, -round_to_unit, scan_temp);
+  }
+  add(recv_klass, scan_temp, scan_temp);
+
+  // Adjust recv_klass by scaled itable_index, so we can free itable_index.
+  RegisterConstant itable_offset = itable_index;
+  regcon_sll_ptr(itable_offset, exact_log2(itableMethodEntry::size() * wordSize));
+  regcon_inc_ptr(itable_offset, itableMethodEntry::method_offset_in_bytes());
+  add(recv_klass, ensure_rs2(itable_offset, sethi_temp), recv_klass);
+
+  // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
+  //   if (scan->interface() == intf) {
+  //     result = (klass + scan->offset() + itable_index);
+  //   }
+  // }
+  Label search, found_method;
+
+  for (int peel = 1; peel >= 0; peel--) {
+    // %%%% Could load both offset and interface in one ldx, if they were
+    // in the opposite order.  This would save a load.
+    ld_ptr(scan_temp, itableOffsetEntry::interface_offset_in_bytes(), method_result);
+
+    // Check that this entry is non-null.  A null entry means that
+    // the receiver class doesn't implement the interface, and wasn't the
+    // same as when the caller was compiled.
+    bpr(Assembler::rc_z, false, Assembler::pn, method_result, L_no_such_interface);
+    delayed()->cmp(method_result, intf_klass);
+
+    if (peel) {
+      brx(Assembler::equal,    false, Assembler::pt, found_method);
+    } else {
+      brx(Assembler::notEqual, false, Assembler::pn, search);
+      // (invert the test to fall through to found_method...)
+    }
+    delayed()->add(scan_temp, scan_step, scan_temp);
+
+    if (!peel)  break;
+
+    bind(search);
+  }
+
+  bind(found_method);
+
+  // Got a hit.
+  int ito_offset = itableOffsetEntry::offset_offset_in_bytes();
+  // scan_temp[-scan_step] points to the vtable offset we need
+  ito_offset -= scan_step;
+  lduw(scan_temp, ito_offset, scan_temp);
+  ld_ptr(recv_klass, scan_temp, method_result);
+}
+
+
 void MacroAssembler::biased_locking_enter(Register obj_reg, Register mark_reg,
                                           Register temp_reg,
                                           Label& done, Label* slow_case,

src/cpu/sparc/vm/assembler_sparc.hpp

@@ -1859,6 +1859,7 @@ class MacroAssembler: public Assembler {
   // Functions for isolating 64 bit shifts for LP64
   inline void sll_ptr( Register s1, Register s2, Register d );
   inline void sll_ptr( Register s1, int imm6a,   Register d );
+  inline void sll_ptr( Register s1, RegisterConstant s2, Register d );
   inline void srl_ptr( Register s1, Register s2, Register d );
   inline void srl_ptr( Register s1, int imm6a,   Register d );
 
@@ -1986,6 +1987,25 @@ class MacroAssembler: public Assembler {
   void load_sized_value(Register s1, RegisterConstant s2, Register d,
                         int size_in_bytes, bool is_signed);
 
+  // Helpers for address formation.
+  // They update the dest in place, whether it is a register or constant.
+  // They emit no code at all if src is a constant zero.
+  // If dest is a constant and src is a register, the temp argument
+  // is required, and becomes the result.
+  // If dest is a register and src is a non-simm13 constant,
+  // the temp argument is required, and is used to materialize the constant.
+  void regcon_inc_ptr( RegisterConstant& dest, RegisterConstant src,
+                       Register temp = noreg );
+  void regcon_sll_ptr( RegisterConstant& dest, RegisterConstant src,
+                       Register temp = noreg );
+  RegisterConstant ensure_rs2(RegisterConstant rs2, Register sethi_temp) {
+    guarantee(sethi_temp != noreg, "constant offset overflow");
+    if (is_simm13(rs2.constant_or_zero()))
+      return rs2;               // register or short constant
+    set(rs2.as_constant(), sethi_temp);
+    return sethi_temp;
+  }
+
   // --------------------------------------------------
 
  public:
@@ -2299,6 +2319,14 @@ class MacroAssembler: public Assembler {
   );
   void tlab_refill(Label& retry_tlab, Label& try_eden, Label& slow_case);
 
+  // interface method calling
+  void lookup_interface_method(Register recv_klass,
+                               Register intf_klass,
+                               RegisterConstant itable_index,
+                               Register method_result,
+                               Register temp_reg, Register temp2_reg,
+                               Label& no_such_interface);
+
   // Stack overflow checking
 
   // Note: this clobbers G3_scratch

src/cpu/sparc/vm/assembler_sparc.inline.hpp

@@ -455,6 +455,11 @@ inline void MacroAssembler::srl_ptr( Register s1, int imm6a,   Register d ) {
 #endif
 }
 
+inline void MacroAssembler::sll_ptr( Register s1, RegisterConstant s2, Register d ) {
+  if (s2.is_register())  sll_ptr(s1, s2.as_register(), d);
+  else                   sll_ptr(s1, s2.as_constant(), d);
+}
+
 // Use the right branch for the platform
 
 inline void MacroAssembler::br( Condition c, bool a, Predict p, address d, relocInfo::relocType rt ) {

src/cpu/sparc/vm/vtableStubs_sparc.cpp

@@ -106,6 +106,15 @@ VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
   __ delayed()->nop();
 
   masm->flush();
+
+  if (PrintMiscellaneous && (WizardMode || Verbose)) {
+    tty->print_cr("vtable #%d at "PTR_FORMAT"[%d] left over: %d",
+                  vtable_index, s->entry_point(),
+                  (int)(s->code_end() - s->entry_point()),
+                  (int)(s->code_end() - __ pc()));
+  }
+  guarantee(__ pc() <= s->code_end(), "overflowed buffer");
+
   s->set_exception_points(npe_addr, ame_addr);
   return s;
 }
@@ -113,9 +122,9 @@ VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
 
 // NOTE:  %%%% if any change is made to this stub make sure that the function
 //             pd_code_size_limit is changed to ensure the correct size for VtableStub
-VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
+VtableStub* VtableStubs::create_itable_stub(int itable_index) {
   const int sparc_code_length = VtableStub::pd_code_size_limit(false);
-  VtableStub* s = new(sparc_code_length) VtableStub(false, vtable_index);
+  VtableStub* s = new(sparc_code_length) VtableStub(false, itable_index);
   ResourceMark rm;
   CodeBuffer cb(s->entry_point(), sparc_code_length);
   MacroAssembler* masm = new MacroAssembler(&cb);
@@ -139,7 +148,6 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
   // are passed in the %o registers.  Instead, longs are passed in G1 and G4
   // and so those registers are not available here.
   __ save(SP,-frame::register_save_words*wordSize,SP);
-  Register I0_receiver = I0;    // Location of receiver after save
 
 #ifndef PRODUCT
   if (CountCompiledCalls) {
@@ -151,63 +159,31 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
   }
 #endif /* PRODUCT */
 
-  // load start of itable entries into L0 register
-  const int base = instanceKlass::vtable_start_offset() * wordSize;
-  __ ld(Address(G3_klassOop, 0, instanceKlass::vtable_length_offset() * wordSize), L0);
-
-  // %%% Could store the aligned, prescaled offset in the klassoop.
-  __ sll(L0, exact_log2(vtableEntry::size() * wordSize), L0);
-  // see code for instanceKlass::start_of_itable!
-  const int vtable_alignment = align_object_offset(1);
-  assert(vtable_alignment == 1 || vtable_alignment == 2, "");
-  const int odd_bit = vtableEntry::size() * wordSize;
-  if (vtable_alignment == 2) {
-    __ and3(L0, odd_bit, L1);   // isolate the odd bit
-  }
-  __ add(G3_klassOop, L0, L0);
-  if (vtable_alignment == 2) {
-    __ add(L0, L1, L0);         // double the odd bit, to align up
-  }
-
-  // Loop over all itable entries until desired interfaceOop (G5_interface) found
-  __ bind(search);
-
-  // %%%% Could load both offset and interface in one ldx, if they were
-  // in the opposite order.  This would save a load.
-  __ ld_ptr(L0, base + itableOffsetEntry::interface_offset_in_bytes(), L1);
-
-  // If the entry is NULL then we've reached the end of the table
-  // without finding the expected interface, so throw an exception
   Label throw_icce;
-  __ bpr(Assembler::rc_z, false, Assembler::pn, L1, throw_icce);
-  __ delayed()->cmp(G5_interface, L1);
-  __ brx(Assembler::notEqual, true, Assembler::pn, search);
-  __ delayed()->add(L0, itableOffsetEntry::size() * wordSize, L0);
 
-  // entry found and L0 points to it, move offset of vtable for interface into L0
-  __ ld(L0, base + itableOffsetEntry::offset_offset_in_bytes(), L0);
-
-  // Compute itableMethodEntry and get methodOop(G5_method) and entrypoint(L0) for compiler
-  const int method_offset = (itableMethodEntry::size() * wordSize * vtable_index) + itableMethodEntry::method_offset_in_bytes();
-  __ add(G3_klassOop, L0, L1);
-  __ ld_ptr(L1, method_offset, G5_method);
+  Register L5_method = L5;
+  __ lookup_interface_method(// inputs: rec. class, interface, itable index
+                             G3_klassOop, G5_interface, itable_index,
+                             // outputs: method, scan temp. reg
+                             L5_method, L2, L3,
+                             throw_icce);
 
 #ifndef PRODUCT
   if (DebugVtables) {
     Label L01;
-    __ ld_ptr(L1, method_offset, G5_method);
-    __ bpr(Assembler::rc_nz, false, Assembler::pt, G5_method, L01);
+    __ bpr(Assembler::rc_nz, false, Assembler::pt, L5_method, L01);
     __ delayed()->nop();
     __ stop("methodOop is null");
     __ bind(L01);
-    __ verify_oop(G5_method);
+    __ verify_oop(L5_method);
   }
 #endif
 
   // If the following load is through a NULL pointer, we'll take an OS
   // exception that should translate into an AbstractMethodError.  We need the
   // window count to be correct at that time.
-  __ restore();                 // Restore registers BEFORE the AME point
+  __ restore(L5_method, 0, G5_method);
+  // Restore registers *before* the AME point.
 
   address ame_addr = __ pc();   // if the vtable entry is null, the method is abstract
   __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3_scratch);
@@ -225,6 +201,12 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
 
   masm->flush();
 
+  if (PrintMiscellaneous && (WizardMode || Verbose)) {
+    tty->print_cr("itable #%d at "PTR_FORMAT"[%d] left over: %d",
+                  itable_index, s->entry_point(),
+                  (int)(s->code_end() - s->entry_point()),
+                  (int)(s->code_end() - __ pc()));
+  }
   guarantee(__ pc() <= s->code_end(), "overflowed buffer");
 
   s->set_exception_points(npe_addr, ame_addr);
@@ -243,8 +225,7 @@ int VtableStub::pd_code_size_limit(bool is_vtable_stub) {
                         (UseCompressedOops ? 2*BytesPerInstWord : 0);
       return basic + slop;
     } else {
-      // save, ld, ld, sll, and, add, add, ld, cmp, br, add, ld, add, ld, ld, jmp, restore, sethi, jmpl, restore
-      const int basic = (20 LP64_ONLY(+ 6)) * BytesPerInstWord +
+      const int basic = (28 LP64_ONLY(+ 6)) * BytesPerInstWord +
                         // shift;add for load_klass
                         (UseCompressedOops ? 2*BytesPerInstWord : 0);
       return (basic + slop);

src/cpu/x86/vm/assembler_x86.cpp

@@ -7076,6 +7076,81 @@ void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) {
 }
 
 
+// Look up the method for a megamorphic invokeinterface call.
+// The target method is determined by <intf_klass, itable_index>.
+// The receiver klass is in recv_klass.
+// On success, the result will be in method_result, and execution falls through.
+// On failure, execution transfers to the given label.
+void MacroAssembler::lookup_interface_method(Register recv_klass,
+                                             Register intf_klass,
+                                             RegisterConstant itable_index,
+                                             Register method_result,
+                                             Register scan_temp,
+                                             Label& L_no_such_interface) {
+  assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
+  assert(itable_index.is_constant() || itable_index.as_register() == method_result,
+         "caller must use same register for non-constant itable index as for method");
+
+  // Compute start of first itableOffsetEntry (which is at the end of the vtable)
+  int vtable_base = instanceKlass::vtable_start_offset() * wordSize;
+  int itentry_off = itableMethodEntry::method_offset_in_bytes();
+  int scan_step   = itableOffsetEntry::size() * wordSize;
+  int vte_size    = vtableEntry::size() * wordSize;
+  Address::ScaleFactor times_vte_scale = Address::times_ptr;
+  assert(vte_size == wordSize, "else adjust times_vte_scale");
+
+  movl(scan_temp, Address(recv_klass, instanceKlass::vtable_length_offset() * wordSize));
+
+  // %%% Could store the aligned, prescaled offset in the klassoop.
+  lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
+  if (HeapWordsPerLong > 1) {
+    // Round up to align_object_offset boundary
+    // see code for instanceKlass::start_of_itable!
+    round_to(scan_temp, BytesPerLong);
+  }
+
+  // Adjust recv_klass by scaled itable_index, so we can free itable_index.
+  assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
+  lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
+
+  // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
+  //   if (scan->interface() == intf) {
+  //     result = (klass + scan->offset() + itable_index);
+  //   }
+  // }
+  Label search, found_method;
+
+  for (int peel = 1; peel >= 0; peel--) {
+    movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
+    cmpptr(intf_klass, method_result);
+
+    if (peel) {
+      jccb(Assembler::equal, found_method);
+    } else {
+      jccb(Assembler::notEqual, search);
+      // (invert the test to fall through to found_method...)
+    }
+
+    if (!peel)  break;
+
+    bind(search);
+
+    // Check that the previous entry is non-null.  A null entry means that
+    // the receiver class doesn't implement the interface, and wasn't the
+    // same as when the caller was compiled.
+    testptr(method_result, method_result);
+    jcc(Assembler::zero, L_no_such_interface);
+    addptr(scan_temp, scan_step);
+  }
+
+  bind(found_method);
+
+  // Got a hit.
+  movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
+  movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
+}
+
+
 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
   ucomisd(dst, as_Address(src));
 }

src/cpu/x86/vm/assembler_x86.hpp

@@ -1765,6 +1765,14 @@ class MacroAssembler: public Assembler {
   );
   void tlab_refill(Label& retry_tlab, Label& try_eden, Label& slow_case);
 
+  // interface method calling
+  void lookup_interface_method(Register recv_klass,
+                               Register intf_klass,
+                               RegisterConstant itable_index,
+                               Register method_result,
+                               Register scan_temp,
+                               Label& no_such_interface);
+
   //----
   void set_word_if_not_zero(Register reg); // sets reg to 1 if not zero, otherwise 0
 

src/cpu/x86/vm/templateTable_x86_32.cpp

@@ -3055,35 +3055,44 @@ void TemplateTable::invokeinterface(int byte_no) {
   // profile this call
   __ profile_virtual_call(rdx, rsi, rdi);
 
-  __ mov(rdi, rdx); // Save klassOop in rdi
+  Label no_such_interface, no_such_method;
 
-  // Compute start of first itableOffsetEntry (which is at the end of the vtable)
-  const int base = instanceKlass::vtable_start_offset() * wordSize;
-  assert(vtableEntry::size() * wordSize == (1 << (int)Address::times_ptr), "adjust the scaling in the code below");
-  __ movl(rsi, Address(rdx, instanceKlass::vtable_length_offset() * wordSize)); // Get length of vtable
-  __ lea(rdx, Address(rdx, rsi, Address::times_4, base));
-  if (HeapWordsPerLong > 1) {
-    // Round up to align_object_offset boundary
-    __ round_to(rdx, BytesPerLong);
-  }
+  __ lookup_interface_method(// inputs: rec. class, interface, itable index
+                             rdx, rax, rbx,
+                             // outputs: method, scan temp. reg
+                             rbx, rsi,
+                             no_such_interface);
 
-  Label entry, search, interface_ok;
+  // rbx,: methodOop to call
+  // rcx: receiver
+  // Check for abstract method error
+  // Note: This should be done more efficiently via a throw_abstract_method_error
+  //       interpreter entry point and a conditional jump to it in case of a null
+  //       method.
+  __ testptr(rbx, rbx);
+  __ jcc(Assembler::zero, no_such_method);
 
-  __ jmpb(entry);
-  __ bind(search);
-  __ addptr(rdx, itableOffsetEntry::size() * wordSize);
+  // do the call
+  // rcx: receiver
+  // rbx,: methodOop
+  __ jump_from_interpreted(rbx, rdx);
+  __ should_not_reach_here();
 
-  __ bind(entry);
+  // exception handling code follows...
+  // note: must restore interpreter registers to canonical
+  //       state for exception handling to work correctly!
 
-  // Check that the entry is non-null.  A null entry means that the receiver
-  // class doesn't implement the interface, and wasn't the same as the
-  // receiver class checked when the interface was resolved.
-  __ push(rdx);
-  __ movptr(rdx, Address(rdx, itableOffsetEntry::interface_offset_in_bytes()));
-  __ testptr(rdx, rdx);
-  __ jcc(Assembler::notZero, interface_ok);
+  __ bind(no_such_method);
+  // throw exception
+  __ pop(rbx);           // pop return address (pushed by prepare_invoke)
+  __ restore_bcp();      // rsi must be correct for exception handler   (was destroyed)
+  __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
+  __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
+  // the call_VM checks for exception, so we should never return here.
+  __ should_not_reach_here();
+
+  __ bind(no_such_interface);
   // throw exception
-  __ pop(rdx);           // pop saved register first.
   __ pop(rbx);           // pop return address (pushed by prepare_invoke)
   __ restore_bcp();      // rsi must be correct for exception handler   (was destroyed)
   __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
@@ -3091,42 +3100,6 @@ void TemplateTable::invokeinterface(int byte_no) {
                    InterpreterRuntime::throw_IncompatibleClassChangeError));
   // the call_VM checks for exception, so we should never return here.
   __ should_not_reach_here();
-  __ bind(interface_ok);
-
-    __ pop(rdx);
-
-    __ cmpptr(rax, Address(rdx, itableOffsetEntry::interface_offset_in_bytes()));
-    __ jcc(Assembler::notEqual, search);
-
-    __ movl(rdx, Address(rdx, itableOffsetEntry::offset_offset_in_bytes()));
-    __ addptr(rdx, rdi); // Add offset to klassOop
-    assert(itableMethodEntry::size() * wordSize == (1 << (int)Address::times_ptr), "adjust the scaling in the code below");
-    __ movptr(rbx, Address(rdx, rbx, Address::times_ptr));
-    // rbx,: methodOop to call
-    // rcx: receiver
-    // Check for abstract method error
-    // Note: This should be done more efficiently via a throw_abstract_method_error
-    //       interpreter entry point and a conditional jump to it in case of a null
-    //       method.
-    { Label L;
-      __ testptr(rbx, rbx);
-      __ jcc(Assembler::notZero, L);
-      // throw exception
-          // note: must restore interpreter registers to canonical
-          //       state for exception handling to work correctly!
-          __ pop(rbx);           // pop return address (pushed by prepare_invoke)
-          __ restore_bcp();      // rsi must be correct for exception handler   (was destroyed)
-          __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
-      __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
-      // the call_VM checks for exception, so we should never return here.
-      __ should_not_reach_here();
-      __ bind(L);
-    }
-
-  // do the call
-  // rcx: receiver
-  // rbx,: methodOop
-  __ jump_from_interpreted(rbx, rdx);
 }
 
 //----------------------------------------------------------------------------------------------------

src/cpu/x86/vm/templateTable_x86_64.cpp

@@ -3010,97 +3010,55 @@ void TemplateTable::invokeinterface(int byte_no) {
   // profile this call
   __ profile_virtual_call(rdx, r13, r14);
 
-  __ mov(r14, rdx); // Save klassOop in r14
+  Label no_such_interface, no_such_method;
 
-  // Compute start of first itableOffsetEntry (which is at the end of
-  // the vtable)
-  const int base = instanceKlass::vtable_start_offset() * wordSize;
-  // Get length of vtable
-  assert(vtableEntry::size() * wordSize == 8,
-         "adjust the scaling in the code below");
-  __ movl(r13, Address(rdx,
-                       instanceKlass::vtable_length_offset() * wordSize));
-  __ lea(rdx, Address(rdx, r13, Address::times_8, base));
+  __ lookup_interface_method(// inputs: rec. class, interface, itable index
+                             rdx, rax, rbx,
+                             // outputs: method, scan temp. reg
+                             rbx, r13,
+                             no_such_interface);
 
-  if (HeapWordsPerLong > 1) {
-    // Round up to align_object_offset boundary
-    __ round_to(rdx, BytesPerLong);
-  }
+  // rbx,: methodOop to call
+  // rcx: receiver
+  // Check for abstract method error
+  // Note: This should be done more efficiently via a throw_abstract_method_error
+  //       interpreter entry point and a conditional jump to it in case of a null
+  //       method.
+  __ testptr(rbx, rbx);
+  __ jcc(Assembler::zero, no_such_method);
 
-  Label entry, search, interface_ok;
+  // do the call
+  // rcx: receiver
+  // rbx,: methodOop
+  __ jump_from_interpreted(rbx, rdx);
+  __ should_not_reach_here();
 
-  __ jmpb(entry);
-  __ bind(search);
-  __ addptr(rdx, itableOffsetEntry::size() * wordSize);
+  // exception handling code follows...
+  // note: must restore interpreter registers to canonical
+  //       state for exception handling to work correctly!
 
-  __ bind(entry);
+  __ bind(no_such_method);
+  // throw exception
+  __ pop(rbx);           // pop return address (pushed by prepare_invoke)
+  __ restore_bcp();      // r13 must be correct for exception handler   (was destroyed)
+  __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
+  __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
+  // the call_VM checks for exception, so we should never return here.
+  __ should_not_reach_here();
 
-  // Check that the entry is non-null.  A null entry means that the
-  // receiver class doesn't implement the interface, and wasn't the
-  // same as the receiver class checked when the interface was
-  // resolved.
-  __ push(rdx);
-  __ movptr(rdx, Address(rdx, itableOffsetEntry::interface_offset_in_bytes()));
-  __ testptr(rdx, rdx);
-  __ jcc(Assembler::notZero, interface_ok);
+  __ bind(no_such_interface);
   // throw exception
-  __ pop(rdx); // pop saved register first.
-  __ pop(rbx); // pop return address (pushed by prepare_invoke)
-  __ restore_bcp(); // r13 must be correct for exception handler (was
-                    // destroyed)
-  __ restore_locals(); // make sure locals pointer is correct as well
-                       // (was destroyed)
+  __ pop(rbx);           // pop return address (pushed by prepare_invoke)
+  __ restore_bcp();      // r13 must be correct for exception handler   (was destroyed)
+  __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
   __ call_VM(noreg, CAST_FROM_FN_PTR(address,
                    InterpreterRuntime::throw_IncompatibleClassChangeError));
   // the call_VM checks for exception, so we should never return here.
   __ should_not_reach_here();
-  __ bind(interface_ok);
-
-  __ pop(rdx);
-
-  __ cmpptr(rax, Address(rdx, itableOffsetEntry::interface_offset_in_bytes()));
-  __ jcc(Assembler::notEqual, search);
-
-  __ movl(rdx, Address(rdx, itableOffsetEntry::offset_offset_in_bytes()));
-
-  __ addptr(rdx, r14); // Add offset to klassOop
-  assert(itableMethodEntry::size() * wordSize == 8,
-         "adjust the scaling in the code below");
-  __ movptr(rbx, Address(rdx, rbx, Address::times_8));
-  // rbx: methodOop to call
-  // rcx: receiver
-  // Check for abstract method error
-  // Note: This should be done more efficiently via a
-  // throw_abstract_method_error interpreter entry point and a
-  // conditional jump to it in case of a null method.
-  {
-    Label L;
-    __ testptr(rbx, rbx);
-    __ jcc(Assembler::notZero, L);
-    // throw exception
-    // note: must restore interpreter registers to canonical
-    //       state for exception handling to work correctly!
-    __ pop(rbx);  // pop return address (pushed by prepare_invoke)
-    __ restore_bcp(); // r13 must be correct for exception handler
-                      // (was destroyed)
-    __ restore_locals(); // make sure locals pointer is correct as
-                         // well (was destroyed)
-    __ call_VM(noreg,
-               CAST_FROM_FN_PTR(address,
-                             InterpreterRuntime::throw_AbstractMethodError));
-    // the call_VM checks for exception, so we should never return here.
-    __ should_not_reach_here();
-    __ bind(L);
-  }
-
-  __ movptr(rcx, Address(rbx, methodOopDesc::interpreter_entry_offset()));
-
-  // do the call
-  // rcx: receiver
-  // rbx: methodOop
-  __ jump_from_interpreted(rbx, rdx);
+  return;
 }
 
+
 //-----------------------------------------------------------------------------
 // Allocation
 

src/cpu/x86/vm/vtableStubs_x86_32.cpp

@@ -34,10 +34,16 @@
 extern "C" void bad_compiled_vtable_index(JavaThread* thread, oop receiver, int index);
 #endif
 
-// used by compiler only; may use only caller saved registers rax, rbx, rcx.
-// rdx holds first int arg, rsi, rdi, rbp are callee-save & must be preserved.
-// Leave receiver in rcx; required behavior when +OptoArgsInRegisters
-// is modifed to put first oop in rcx.
+// These stubs are used by the compiler only.
+// Argument registers, which must be preserved:
+//   rcx - receiver (always first argument)
+//   rdx - second argument (if any)
+// Other registers that might be usable:
+//   rax - inline cache register (is interface for itable stub)
+//   rbx - method (used when calling out to interpreter)
+// Available now, but may become callee-save at some point:
+//   rsi, rdi
+// Note that rax and rdx are also used for return values.
 //
 VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
   const int i486_code_length = VtableStub::pd_code_size_limit(true);
@@ -94,16 +100,25 @@ VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
   __ jmp( Address(method, methodOopDesc::from_compiled_offset()));
 
   masm->flush();
+
+  if (PrintMiscellaneous && (WizardMode || Verbose)) {
+    tty->print_cr("vtable #%d at "PTR_FORMAT"[%d] left over: %d",
+                  vtable_index, s->entry_point(),
+                  (int)(s->code_end() - s->entry_point()),
+                  (int)(s->code_end() - __ pc()));
+  }
+  guarantee(__ pc() <= s->code_end(), "overflowed buffer");
+
   s->set_exception_points(npe_addr, ame_addr);
   return s;
 }
 
 
-VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
+VtableStub* VtableStubs::create_itable_stub(int itable_index) {
   // Note well: pd_code_size_limit is the absolute minimum we can get away with.  If you
   //            add code here, bump the code stub size returned by pd_code_size_limit!
   const int i486_code_length = VtableStub::pd_code_size_limit(false);
-  VtableStub* s = new(i486_code_length) VtableStub(false, vtable_index);
+  VtableStub* s = new(i486_code_length) VtableStub(false, itable_index);
   ResourceMark rm;
   CodeBuffer cb(s->entry_point(), i486_code_length);
   MacroAssembler* masm = new MacroAssembler(&cb);
@@ -123,50 +138,19 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
 
   // get receiver klass (also an implicit null-check)
   address npe_addr = __ pc();
-  __ movptr(rbx, Address(rcx, oopDesc::klass_offset_in_bytes()));
-
-  __ mov(rsi, rbx);   // Save klass in free register
-  // Most registers are in use, so save a few
-  __ push(rdx);
-  // compute itable entry offset (in words)
-  const int base = instanceKlass::vtable_start_offset() * wordSize;
-  assert(vtableEntry::size() * wordSize == 4, "adjust the scaling in the code below");
-  __ movl(rdx, Address(rbx, instanceKlass::vtable_length_offset() * wordSize)); // Get length of vtable
-  __ lea(rbx, Address(rbx, rdx, Address::times_ptr, base));
-  if (HeapWordsPerLong > 1) {
-    // Round up to align_object_offset boundary
-    __ round_to(rbx, BytesPerLong);
-  }
-
-  Label hit, next, entry, throw_icce;
-
-  __ jmpb(entry);
-
-  __ bind(next);
-  __ addptr(rbx, itableOffsetEntry::size() * wordSize);
+  __ movptr(rsi, Address(rcx, oopDesc::klass_offset_in_bytes()));
 
-  __ bind(entry);
-
-  // If the entry is NULL then we've reached the end of the table
-  // without finding the expected interface, so throw an exception
-  __ movptr(rdx, Address(rbx, itableOffsetEntry::interface_offset_in_bytes()));
-  __ testptr(rdx, rdx);
-  __ jcc(Assembler::zero, throw_icce);
-  __ cmpptr(rax, rdx);
-  __ jcc(Assembler::notEqual, next);
-
-  // We found a hit, move offset into rbx,
-  __ movl(rdx, Address(rbx, itableOffsetEntry::offset_offset_in_bytes()));
-
-  // Compute itableMethodEntry.
-  const int method_offset = (itableMethodEntry::size() * wordSize * vtable_index) + itableMethodEntry::method_offset_in_bytes();
-
-  // Get methodOop and entrypoint for compiler
+  // Most registers are in use; we'll use rax, rbx, rsi, rdi
+  // (If we need to make rsi, rdi callee-save, do a push/pop here.)
   const Register method = rbx;
-  __ movptr(method, Address(rsi, rdx, Address::times_1, method_offset));
+  Label throw_icce;
 
-  // Restore saved register, before possible trap.
-  __ pop(rdx);
+  // Get methodOop and entrypoint for compiler
+  __ lookup_interface_method(// inputs: rec. class, interface, itable index
+                             rsi, rax, itable_index,
+                             // outputs: method, scan temp. reg
+                             method, rdi,
+                             throw_icce);
 
   // method (rbx): methodOop
   // rcx: receiver
@@ -187,12 +171,15 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
   __ jmp(Address(method, methodOopDesc::from_compiled_offset()));
 
   __ bind(throw_icce);
-  // Restore saved register
-  __ pop(rdx);
   __ jump(RuntimeAddress(StubRoutines::throw_IncompatibleClassChangeError_entry()));
-
   masm->flush();
 
+  if (PrintMiscellaneous && (WizardMode || Verbose)) {
+    tty->print_cr("itable #%d at "PTR_FORMAT"[%d] left over: %d",
+                  itable_index, s->entry_point(),
+                  (int)(s->code_end() - s->entry_point()),
+                  (int)(s->code_end() - __ pc()));
+  }
   guarantee(__ pc() <= s->code_end(), "overflowed buffer");
 
   s->set_exception_points(npe_addr, ame_addr);
@@ -207,7 +194,7 @@ int VtableStub::pd_code_size_limit(bool is_vtable_stub) {
     return (DebugVtables ? 210 : 16) + (CountCompiledCalls ? 6 : 0);
   } else {
     // Itable stub size
-    return (DebugVtables ? 144 : 64) + (CountCompiledCalls ? 6 : 0);
+    return (DebugVtables ? 256 : 66) + (CountCompiledCalls ? 6 : 0);
   }
 }
 

src/cpu/x86/vm/vtableStubs_x86_64.cpp

@@ -98,17 +98,26 @@ VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
   __ jmp( Address(rbx, methodOopDesc::from_compiled_offset()));
 
   __ flush();
+
+  if (PrintMiscellaneous && (WizardMode || Verbose)) {
+    tty->print_cr("vtable #%d at "PTR_FORMAT"[%d] left over: %d",
+                  vtable_index, s->entry_point(),
+                  (int)(s->code_end() - s->entry_point()),
+                  (int)(s->code_end() - __ pc()));
+  }
+  guarantee(__ pc() <= s->code_end(), "overflowed buffer");
+
   s->set_exception_points(npe_addr, ame_addr);
   return s;
 }
 
 
-VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
+VtableStub* VtableStubs::create_itable_stub(int itable_index) {
   // Note well: pd_code_size_limit is the absolute minimum we can get
   // away with.  If you add code here, bump the code stub size
   // returned by pd_code_size_limit!
   const int amd64_code_length = VtableStub::pd_code_size_limit(false);
-  VtableStub* s = new(amd64_code_length) VtableStub(false, vtable_index);
+  VtableStub* s = new(amd64_code_length) VtableStub(false, itable_index);
   ResourceMark rm;
   CodeBuffer cb(s->entry_point(), amd64_code_length);
   MacroAssembler* masm = new MacroAssembler(&cb);
@@ -131,68 +140,28 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
   // get receiver klass (also an implicit null-check)
   address npe_addr = __ pc();
 
-  __ load_klass(rbx, j_rarg0);
+  // Most registers are in use; we'll use rax, rbx, r10, r11
+  // (various calling sequences use r[cd]x, r[sd]i, r[89]; stay away from them)
+  __ load_klass(r10, j_rarg0);
 
   // If we take a trap while this arg is on the stack we will not
   // be able to walk the stack properly. This is not an issue except
   // when there are mistakes in this assembly code that could generate
   // a spurious fault. Ask me how I know...
 
-  __ push(j_rarg1);     // Most registers are in use, so save one
-
-  // compute itable entry offset (in words)
-  const int base = instanceKlass::vtable_start_offset() * wordSize;
-  assert(vtableEntry::size() * wordSize == 8,
-         "adjust the scaling in the code below");
-  // Get length of vtable
-  __ movl(j_rarg1,
-          Address(rbx, instanceKlass::vtable_length_offset() * wordSize));
-  __ lea(rbx, Address(rbx, j_rarg1, Address::times_8, base));
-
-  if (HeapWordsPerLong > 1) {
-    // Round up to align_object_offset boundary
-    __ round_to(rbx, BytesPerLong);
-  }
-  Label hit, next, entry, throw_icce;
-
-  __ jmpb(entry);
-
-  __ bind(next);
-  __ addptr(rbx, itableOffsetEntry::size() * wordSize);
-
-  __ bind(entry);
-
-  // If the entry is NULL then we've reached the end of the table
-  // without finding the expected interface, so throw an exception
-  __ movptr(j_rarg1, Address(rbx, itableOffsetEntry::interface_offset_in_bytes()));
-  __ testptr(j_rarg1, j_rarg1);
-  __ jcc(Assembler::zero, throw_icce);
-  __ cmpptr(rax, j_rarg1);
-  __ jccb(Assembler::notEqual, next);
-
-  // We found a hit, move offset into j_rarg1
-  __ movl(j_rarg1, Address(rbx, itableOffsetEntry::offset_offset_in_bytes()));
-
-  // Compute itableMethodEntry
-  const int method_offset =
-    (itableMethodEntry::size() * wordSize * vtable_index) +
-    itableMethodEntry::method_offset_in_bytes();
-
-  // Get methodOop and entrypoint for compiler
-
-  // Get klass pointer again
-  __ load_klass(rax, j_rarg0);
-
   const Register method = rbx;
-  __ movptr(method, Address(rax, j_rarg1, Address::times_1, method_offset));
+  Label throw_icce;
 
-  // Restore saved register, before possible trap.
-  __ pop(j_rarg1);
+  // Get methodOop and entrypoint for compiler
+  __ lookup_interface_method(// inputs: rec. class, interface, itable index
+                             r10, rax, itable_index,
+                             // outputs: method, scan temp. reg
+                             method, r11,
+                             throw_icce);
 
   // method (rbx): methodOop
   // j_rarg0: receiver
 
-
 #ifdef ASSERT
   if (DebugVtables) {
     Label L2;
@@ -211,12 +180,16 @@ VtableStub* VtableStubs::create_itable_stub(int vtable_index) {
   __ jmp(Address(method, methodOopDesc::from_compiled_offset()));
 
   __ bind(throw_icce);
-  // Restore saved register
-  __ pop(j_rarg1);
   __ jump(RuntimeAddress(StubRoutines::throw_IncompatibleClassChangeError_entry()));
 
   __ flush();
 
+  if (PrintMiscellaneous && (WizardMode || Verbose)) {
+    tty->print_cr("itable #%d at "PTR_FORMAT"[%d] left over: %d",
+                  itable_index, s->entry_point(),
+                  (int)(s->code_end() - s->entry_point()),
+                  (int)(s->code_end() - __ pc()));
+  }
   guarantee(__ pc() <= s->code_end(), "overflowed buffer");
 
   s->set_exception_points(npe_addr, ame_addr);
@@ -230,7 +203,7 @@ int VtableStub::pd_code_size_limit(bool is_vtable_stub) {
            (UseCompressedOops ? 16 : 0);  // 1 leaq can be 3 bytes + 1 long
   } else {
     // Itable stub size
-    return (DebugVtables ? 636 : 72) + (CountCompiledCalls ? 13 : 0) +
+    return (DebugVtables ? 512 : 72) + (CountCompiledCalls ? 13 : 0) +
            (UseCompressedOops ? 32 : 0);  // 2 leaqs
   }
 }