aoqi@0: /*
aoqi@0:  * Copyright (c) 2003, 2014, Oracle and/or its affiliates. All rights reserved.
aoqi@0:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
aoqi@0:  *
aoqi@0:  * This code is free software; you can redistribute it and/or modify it
aoqi@0:  * under the terms of the GNU General Public License version 2 only, as
aoqi@0:  * published by the Free Software Foundation.
aoqi@0:  *
aoqi@0:  * This code is distributed in the hope that it will be useful, but WITHOUT
aoqi@0:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
aoqi@0:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
aoqi@0:  * version 2 for more details (a copy is included in the LICENSE file that
aoqi@0:  * accompanied this code).
aoqi@0:  *
aoqi@0:  * You should have received a copy of the GNU General Public License version
aoqi@0:  * 2 along with this work; if not, write to the Free Software Foundation,
aoqi@0:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
aoqi@0:  *
aoqi@0:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
aoqi@0:  * or visit www.oracle.com if you need additional information or have any
aoqi@0:  * questions.
aoqi@0:  *
aoqi@0:  */
aoqi@0: 
aoqi@0: #include "precompiled.hpp"
aoqi@0: #include "asm/macroAssembler.hpp"
aoqi@0: #include "code/vtableStubs.hpp"
aoqi@0: #include "interp_masm_x86.hpp"
aoqi@0: #include "memory/resourceArea.hpp"
aoqi@0: #include "oops/instanceKlass.hpp"
aoqi@0: #include "oops/klassVtable.hpp"
aoqi@0: #include "runtime/sharedRuntime.hpp"
aoqi@0: #include "vmreg_x86.inline.hpp"
aoqi@0: #ifdef COMPILER2
aoqi@0: #include "opto/runtime.hpp"
aoqi@0: #endif
aoqi@0: 
aoqi@0: PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
aoqi@0: 
aoqi@0: // machine-dependent part of VtableStubs: create VtableStub of correct size and
aoqi@0: // initialize its code
aoqi@0: 
aoqi@0: #define __ masm->
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0: extern "C" void bad_compiled_vtable_index(JavaThread* thread,
aoqi@0:                                           oop receiver,
aoqi@0:                                           int index);
aoqi@0: #endif
aoqi@0: 
aoqi@0: VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
aoqi@0:   const int amd64_code_length = VtableStub::pd_code_size_limit(true);
aoqi@0:   VtableStub* s = new(amd64_code_length) VtableStub(true, vtable_index);
aoqi@0:   // Can be NULL if there is no free space in the code cache.
aoqi@0:   if (s == NULL) {
aoqi@0:     return NULL;
aoqi@0:   }
aoqi@0: 
aoqi@0:   ResourceMark rm;
aoqi@0:   CodeBuffer cb(s->entry_point(), amd64_code_length);
aoqi@0:   MacroAssembler* masm = new MacroAssembler(&cb);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (CountCompiledCalls) {
aoqi@0:     __ incrementl(ExternalAddress((address) SharedRuntime::nof_megamorphic_calls_addr()));
aoqi@0:   }
aoqi@0: #endif
aoqi@0: 
aoqi@0:   // get receiver (need to skip return address on top of stack)
aoqi@0:   assert(VtableStub::receiver_location() == j_rarg0->as_VMReg(), "receiver expected in j_rarg0");
aoqi@0: 
aoqi@0:   // Free registers (non-args) are rax, rbx
aoqi@0: 
aoqi@0:   // get receiver klass
aoqi@0:   address npe_addr = __ pc();
aoqi@0:   __ load_klass(rax, j_rarg0);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (DebugVtables) {
aoqi@0:     Label L;
aoqi@0:     // check offset vs vtable length
aoqi@0:     __ cmpl(Address(rax, InstanceKlass::vtable_length_offset() * wordSize),
aoqi@0:             vtable_index * vtableEntry::size());
aoqi@0:     __ jcc(Assembler::greater, L);
aoqi@0:     __ movl(rbx, vtable_index);
aoqi@0:     __ call_VM(noreg,
aoqi@0:                CAST_FROM_FN_PTR(address, bad_compiled_vtable_index), j_rarg0, rbx);
aoqi@0:     __ bind(L);
aoqi@0:   }
aoqi@0: #endif // PRODUCT
aoqi@0: 
aoqi@0:   // load Method* and target address
aoqi@0:   const Register method = rbx;
aoqi@0: 
aoqi@0:   __ lookup_virtual_method(rax, vtable_index, method);
aoqi@0: 
aoqi@0:   if (DebugVtables) {
aoqi@0:     Label L;
aoqi@0:     __ cmpptr(method, (int32_t)NULL_WORD);
aoqi@0:     __ jcc(Assembler::equal, L);
aoqi@0:     __ cmpptr(Address(method, Method::from_compiled_offset()), (int32_t)NULL_WORD);
aoqi@0:     __ jcc(Assembler::notZero, L);
aoqi@0:     __ stop("Vtable entry is NULL");
aoqi@0:     __ bind(L);
aoqi@0:   }
aoqi@0:   // rax: receiver klass
aoqi@0:   // rbx: Method*
aoqi@0:   // rcx: receiver
aoqi@0:   address ame_addr = __ pc();
aoqi@0:   __ jmp( Address(rbx, Method::from_compiled_offset()));
aoqi@0: 
aoqi@0:   __ flush();
aoqi@0: 
aoqi@0:   if (PrintMiscellaneous && (WizardMode || Verbose)) {
aoqi@0:     tty->print_cr("vtable #%d at "PTR_FORMAT"[%d] left over: %d",
aoqi@0:                   vtable_index, s->entry_point(),
aoqi@0:                   (int)(s->code_end() - s->entry_point()),
aoqi@0:                   (int)(s->code_end() - __ pc()));
aoqi@0:   }
aoqi@0:   guarantee(__ pc() <= s->code_end(), "overflowed buffer");
aoqi@0:   // shut the door on sizing bugs
aoqi@0:   int slop = 3;  // 32-bit offset is this much larger than an 8-bit one
aoqi@0:   assert(vtable_index > 10 || __ pc() + slop <= s->code_end(), "room for 32-bit offset");
aoqi@0: 
aoqi@0:   s->set_exception_points(npe_addr, ame_addr);
aoqi@0:   return s;
aoqi@0: }
aoqi@0: 
aoqi@0: 
aoqi@0: VtableStub* VtableStubs::create_itable_stub(int itable_index) {
aoqi@0:   // Note well: pd_code_size_limit is the absolute minimum we can get
aoqi@0:   // away with.  If you add code here, bump the code stub size
aoqi@0:   // returned by pd_code_size_limit!
aoqi@0:   const int amd64_code_length = VtableStub::pd_code_size_limit(false);
aoqi@0:   VtableStub* s = new(amd64_code_length) VtableStub(false, itable_index);
aoqi@0:   // Can be NULL if there is no free space in the code cache.
aoqi@0:   if (s == NULL) {
aoqi@0:     return NULL;
aoqi@0:   }
aoqi@0: 
aoqi@0:   ResourceMark rm;
aoqi@0:   CodeBuffer cb(s->entry_point(), amd64_code_length);
aoqi@0:   MacroAssembler* masm = new MacroAssembler(&cb);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (CountCompiledCalls) {
aoqi@0:     __ incrementl(ExternalAddress((address) SharedRuntime::nof_megamorphic_calls_addr()));
aoqi@0:   }
aoqi@0: #endif
aoqi@0: 
aoqi@0:   // Entry arguments:
aoqi@0:   //  rax: Interface
aoqi@0:   //  j_rarg0: Receiver
aoqi@0: 
aoqi@0:   // Free registers (non-args) are rax (interface), rbx
aoqi@0: 
aoqi@0:   // get receiver (need to skip return address on top of stack)
aoqi@0: 
aoqi@0:   assert(VtableStub::receiver_location() == j_rarg0->as_VMReg(), "receiver expected in j_rarg0");
aoqi@0:   // get receiver klass (also an implicit null-check)
aoqi@0:   address npe_addr = __ pc();
aoqi@0: 
aoqi@0:   // Most registers are in use; we'll use rax, rbx, r10, r11
aoqi@0:   // (various calling sequences use r[cd]x, r[sd]i, r[89]; stay away from them)
aoqi@0:   __ load_klass(r10, j_rarg0);
aoqi@0: 
aoqi@0:   // If we take a trap while this arg is on the stack we will not
aoqi@0:   // be able to walk the stack properly. This is not an issue except
aoqi@0:   // when there are mistakes in this assembly code that could generate
aoqi@0:   // a spurious fault. Ask me how I know...
aoqi@0: 
aoqi@0:   const Register method = rbx;
aoqi@0:   Label throw_icce;
aoqi@0: 
aoqi@0:   // Get Method* and entrypoint for compiler
aoqi@0:   __ lookup_interface_method(// inputs: rec. class, interface, itable index
aoqi@0:                              r10, rax, itable_index,
aoqi@0:                              // outputs: method, scan temp. reg
aoqi@0:                              method, r11,
aoqi@0:                              throw_icce);
aoqi@0: 
aoqi@0:   // method (rbx): Method*
aoqi@0:   // j_rarg0: receiver
aoqi@0: 
aoqi@0: #ifdef ASSERT
aoqi@0:   if (DebugVtables) {
aoqi@0:     Label L2;
aoqi@0:     __ cmpptr(method, (int32_t)NULL_WORD);
aoqi@0:     __ jcc(Assembler::equal, L2);
aoqi@0:     __ cmpptr(Address(method, Method::from_compiled_offset()), (int32_t)NULL_WORD);
aoqi@0:     __ jcc(Assembler::notZero, L2);
aoqi@0:     __ stop("compiler entrypoint is null");
aoqi@0:     __ bind(L2);
aoqi@0:   }
aoqi@0: #endif // ASSERT
aoqi@0: 
aoqi@0:   // rbx: Method*
aoqi@0:   // j_rarg0: receiver
aoqi@0:   address ame_addr = __ pc();
aoqi@0:   __ jmp(Address(method, Method::from_compiled_offset()));
aoqi@0: 
aoqi@0:   __ bind(throw_icce);
aoqi@0:   __ jump(RuntimeAddress(StubRoutines::throw_IncompatibleClassChangeError_entry()));
aoqi@0: 
aoqi@0:   __ flush();
aoqi@0: 
aoqi@0:   if (PrintMiscellaneous && (WizardMode || Verbose)) {
aoqi@0:     tty->print_cr("itable #%d at "PTR_FORMAT"[%d] left over: %d",
aoqi@0:                   itable_index, s->entry_point(),
aoqi@0:                   (int)(s->code_end() - s->entry_point()),
aoqi@0:                   (int)(s->code_end() - __ pc()));
aoqi@0:   }
aoqi@0:   guarantee(__ pc() <= s->code_end(), "overflowed buffer");
aoqi@0:   // shut the door on sizing bugs
aoqi@0:   int slop = 3;  // 32-bit offset is this much larger than an 8-bit one
aoqi@0:   assert(itable_index > 10 || __ pc() + slop <= s->code_end(), "room for 32-bit offset");
aoqi@0: 
aoqi@0:   s->set_exception_points(npe_addr, ame_addr);
aoqi@0:   return s;
aoqi@0: }
aoqi@0: 
aoqi@0: int VtableStub::pd_code_size_limit(bool is_vtable_stub) {
aoqi@0:   if (is_vtable_stub) {
aoqi@0:     // Vtable stub size
aoqi@0:     return (DebugVtables ? 512 : 24) + (CountCompiledCalls ? 13 : 0) +
aoqi@0:            (UseCompressedClassPointers ?  MacroAssembler::instr_size_for_decode_klass_not_null() : 0);
aoqi@0:   } else {
aoqi@0:     // Itable stub size
aoqi@0:     return (DebugVtables ? 512 : 74) + (CountCompiledCalls ? 13 : 0) +
aoqi@0:            (UseCompressedClassPointers ?  MacroAssembler::instr_size_for_decode_klass_not_null() : 0);
aoqi@0:   }
aoqi@0:   // In order to tune these parameters, run the JVM with VM options
aoqi@0:   // +PrintMiscellaneous and +WizardMode to see information about
aoqi@0:   // actual itable stubs.  Look for lines like this:
aoqi@0:   //   itable #1 at 0x5551212[71] left over: 3
aoqi@0:   // Reduce the constants so that the "left over" number is >=3
aoqi@0:   // for the common cases.
aoqi@0:   // Do not aim at a left-over number of zero, because a
aoqi@0:   // large vtable or itable index (>= 32) will require a 32-bit
aoqi@0:   // immediate displacement instead of an 8-bit one.
aoqi@0:   //
aoqi@0:   // The JVM98 app. _202_jess has a megamorphic interface call.
aoqi@0:   // The itable code looks like this:
aoqi@0:   // Decoding VtableStub itbl[1]@12
aoqi@0:   //   mov    0x8(%rsi),%r10
aoqi@0:   //   mov    0x198(%r10),%r11d
aoqi@0:   //   lea    0x218(%r10,%r11,8),%r11
aoqi@0:   //   lea    0x8(%r10),%r10
aoqi@0:   //   mov    (%r11),%rbx
aoqi@0:   //   cmp    %rbx,%rax
aoqi@0:   //   je     success
aoqi@0:   // loop:
aoqi@0:   //   test   %rbx,%rbx
aoqi@0:   //   je     throw_icce
aoqi@0:   //   add    $0x10,%r11
aoqi@0:   //   mov    (%r11),%rbx
aoqi@0:   //   cmp    %rbx,%rax
aoqi@0:   //   jne    loop
aoqi@0:   // success:
aoqi@0:   //   mov    0x8(%r11),%r11d
aoqi@0:   //   mov    (%r10,%r11,1),%rbx
aoqi@0:   //   jmpq   *0x60(%rbx)
aoqi@0:   // throw_icce:
aoqi@0:   //   jmpq   throw_ICCE_entry
aoqi@0: }
aoqi@0: 
aoqi@0: int VtableStub::pd_code_alignment() {
aoqi@0:   return wordSize;
aoqi@0: }