aoqi@0: /*
aoqi@0:  * Copyright (c) 1997, 2013, 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.inline.hpp"
aoqi@0: #include "code/vtableStubs.hpp"
aoqi@0: #include "interp_masm_sparc.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_sparc.inline.hpp"
aoqi@0: #ifdef COMPILER2
aoqi@0: #include "opto/runtime.hpp"
aoqi@0: #endif
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: 
aoqi@0: #ifndef PRODUCT
aoqi@0: extern "C" void bad_compiled_vtable_index(JavaThread* thread, oopDesc* receiver, int index);
aoqi@0: #endif
aoqi@0: 
aoqi@0: 
aoqi@0: // Used by compiler only; may use only caller saved, non-argument registers
aoqi@0: // NOTE:  %%%% if any change is made to this stub make sure that the function
aoqi@0: //             pd_code_size_limit is changed to ensure the correct size for VtableStub
aoqi@0: VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
aoqi@0:   const int sparc_code_length = VtableStub::pd_code_size_limit(true);
aoqi@0:   VtableStub* s = new(sparc_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(), sparc_code_length);
aoqi@0:   MacroAssembler* masm = new MacroAssembler(&cb);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (CountCompiledCalls) {
aoqi@0:     __ inc_counter(SharedRuntime::nof_megamorphic_calls_addr(), G5, G3_scratch);
aoqi@0:   }
aoqi@0: #endif /* PRODUCT */
aoqi@0: 
aoqi@0:   assert(VtableStub::receiver_location() == O0->as_VMReg(), "receiver expected in O0");
aoqi@0: 
aoqi@0:   // get receiver klass
aoqi@0:   address npe_addr = __ pc();
aoqi@0:   __ load_klass(O0, G3_scratch);
aoqi@0: 
aoqi@0:   // set Method* (in case of interpreted method), and destination address
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (DebugVtables) {
aoqi@0:     Label L;
aoqi@0:     // check offset vs vtable length
aoqi@0:     __ ld(G3_scratch, InstanceKlass::vtable_length_offset()*wordSize, G5);
aoqi@0:     __ cmp_and_br_short(G5, vtable_index*vtableEntry::size(), Assembler::greaterUnsigned, Assembler::pt, L);
aoqi@0:     __ set(vtable_index, O2);
aoqi@0:     __ call_VM(noreg, CAST_FROM_FN_PTR(address, bad_compiled_vtable_index), O0, O2);
aoqi@0:     __ bind(L);
aoqi@0:   }
aoqi@0: #endif
aoqi@0: 
aoqi@0:   __ lookup_virtual_method(G3_scratch, vtable_index, G5_method);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (DebugVtables) {
aoqi@0:     Label L;
aoqi@0:     __ br_notnull_short(G5_method, Assembler::pt, L);
aoqi@0:     __ stop("Vtable entry is ZERO");
aoqi@0:     __ bind(L);
aoqi@0:   }
aoqi@0: #endif
aoqi@0: 
aoqi@0:   address ame_addr = __ pc();  // if the vtable entry is null, the method is abstract
aoqi@0:                                // NOTE: for vtable dispatches, the vtable entry will never be null.
aoqi@0: 
aoqi@0:   __ ld_ptr(G5_method, in_bytes(Method::from_compiled_offset()), G3_scratch);
aoqi@0: 
aoqi@0:   // jump to target (either compiled code or c2iadapter)
aoqi@0:   __ JMP(G3_scratch, 0);
aoqi@0:   // load Method* (in case we call c2iadapter)
aoqi@0:   __ delayed()->nop();
aoqi@0: 
aoqi@0:   masm->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 = 2*BytesPerInstWord;  // 32-bit offset is this much larger than a 13-bit one
aoqi@0:   assert(vtable_index > 10 || __ pc() + slop <= s->code_end(), "room for sethi;add");
aoqi@0: 
aoqi@0:   s->set_exception_points(npe_addr, ame_addr);
aoqi@0:   return s;
aoqi@0: }
aoqi@0: 
aoqi@0: 
aoqi@0: // NOTE:  %%%% if any change is made to this stub make sure that the function
aoqi@0: //             pd_code_size_limit is changed to ensure the correct size for VtableStub
aoqi@0: VtableStub* VtableStubs::create_itable_stub(int itable_index) {
aoqi@0:   const int sparc_code_length = VtableStub::pd_code_size_limit(false);
aoqi@0:   VtableStub* s = new(sparc_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(), sparc_code_length);
aoqi@0:   MacroAssembler* masm = new MacroAssembler(&cb);
aoqi@0: 
aoqi@0:   Register G3_Klass = G3_scratch;
aoqi@0:   Register G5_interface = G5;  // Passed in as an argument
aoqi@0:   Label search;
aoqi@0: 
aoqi@0:   // Entry arguments:
aoqi@0:   //  G5_interface: Interface
aoqi@0:   //  O0:           Receiver
aoqi@0:   assert(VtableStub::receiver_location() == O0->as_VMReg(), "receiver expected in O0");
aoqi@0: 
aoqi@0:   // get receiver klass (also an implicit null-check)
aoqi@0:   address npe_addr = __ pc();
aoqi@0:   __ load_klass(O0, G3_Klass);
aoqi@0: 
aoqi@0:   // Push a new window to get some temp registers.  This chops the head of all
aoqi@0:   // my 64-bit %o registers in the LION build, but this is OK because no longs
aoqi@0:   // are passed in the %o registers.  Instead, longs are passed in G1 and G4
aoqi@0:   // and so those registers are not available here.
aoqi@0:   __ save(SP,-frame::register_save_words*wordSize,SP);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (CountCompiledCalls) {
aoqi@0:     __ inc_counter(SharedRuntime::nof_megamorphic_calls_addr(), L0, L1);
aoqi@0:   }
aoqi@0: #endif /* PRODUCT */
aoqi@0: 
aoqi@0:   Label throw_icce;
aoqi@0: 
aoqi@0:   Register L5_method = L5;
aoqi@0:   __ lookup_interface_method(// inputs: rec. class, interface, itable index
aoqi@0:                              G3_Klass, G5_interface, itable_index,
aoqi@0:                              // outputs: method, scan temp. reg
aoqi@0:                              L5_method, L2, L3,
aoqi@0:                              throw_icce);
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   if (DebugVtables) {
aoqi@0:     Label L01;
aoqi@0:     __ br_notnull_short(L5_method, Assembler::pt, L01);
aoqi@0:     __ stop("Method* is null");
aoqi@0:     __ bind(L01);
aoqi@0:   }
aoqi@0: #endif
aoqi@0: 
aoqi@0:   // If the following load is through a NULL pointer, we'll take an OS
aoqi@0:   // exception that should translate into an AbstractMethodError.  We need the
aoqi@0:   // window count to be correct at that time.
aoqi@0:   __ restore(L5_method, 0, G5_method);
aoqi@0:   // Restore registers *before* the AME point.
aoqi@0: 
aoqi@0:   address ame_addr = __ pc();   // if the vtable entry is null, the method is abstract
aoqi@0:   __ ld_ptr(G5_method, in_bytes(Method::from_compiled_offset()), G3_scratch);
aoqi@0: 
aoqi@0:   // G5_method:  Method*
aoqi@0:   // O0:         Receiver
aoqi@0:   // G3_scratch: entry point
aoqi@0:   __ JMP(G3_scratch, 0);
aoqi@0:   __ delayed()->nop();
aoqi@0: 
aoqi@0:   __ bind(throw_icce);
aoqi@0:   AddressLiteral icce(StubRoutines::throw_IncompatibleClassChangeError_entry());
aoqi@0:   __ jump_to(icce, G3_scratch);
aoqi@0:   __ delayed()->restore();
aoqi@0: 
aoqi@0:   masm->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 = 2*BytesPerInstWord;  // 32-bit offset is this much larger than a 13-bit one
aoqi@0:   assert(itable_index > 10 || __ pc() + slop <= s->code_end(), "room for sethi;add");
aoqi@0: 
aoqi@0:   s->set_exception_points(npe_addr, ame_addr);
aoqi@0:   return s;
aoqi@0: }
aoqi@0: 
aoqi@0: 
aoqi@0: int VtableStub::pd_code_size_limit(bool is_vtable_stub) {
aoqi@0:   if (TraceJumps || DebugVtables || CountCompiledCalls || VerifyOops) return 1000;
aoqi@0:   else {
aoqi@0:     const int slop = 2*BytesPerInstWord; // sethi;add  (needed for long offsets)
aoqi@0:     if (is_vtable_stub) {
aoqi@0:       // ld;ld;ld,jmp,nop
aoqi@0:       const int basic = 5*BytesPerInstWord +
aoqi@0:                         // shift;add for load_klass (only shift with zero heap based)
aoqi@0:                         (UseCompressedClassPointers ?
aoqi@0:                           MacroAssembler::instr_size_for_decode_klass_not_null() : 0);
aoqi@0:       return basic + slop;
aoqi@0:     } else {
aoqi@0:       const int basic = (28 LP64_ONLY(+ 6)) * BytesPerInstWord +
aoqi@0:                         // shift;add for load_klass (only shift with zero heap based)
aoqi@0:                         (UseCompressedClassPointers ?
aoqi@0:                           MacroAssembler::instr_size_for_decode_klass_not_null() : 0);
aoqi@0:       return (basic + slop);
aoqi@0:     }
aoqi@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[116] left over: 8
aoqi@0:   // Reduce the constants so that the "left over" number is 8
aoqi@0:   // Do not aim at a left-over number of zero, because a very
aoqi@0:   // large vtable or itable offset (> 4K) will require an extra
aoqi@0:   // sethi/or pair of instructions.
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]@16
aoqi@0:   //   ld  [ %o0 + 4 ], %g3
aoqi@0:   //   save  %sp, -64, %sp
aoqi@0:   //   ld  [ %g3 + 0xe8 ], %l2
aoqi@0:   //   sll  %l2, 2, %l2
aoqi@0:   //   add  %l2, 0x134, %l2
aoqi@0:   //   and  %l2, -8, %l2        ! NOT_LP64 only
aoqi@0:   //   add  %g3, %l2, %l2
aoqi@0:   //   add  %g3, 4, %g3
aoqi@0:   //   ld  [ %l2 ], %l5
aoqi@0:   //   brz,pn   %l5, throw_icce
aoqi@0:   //   cmp  %l5, %g5
aoqi@0:   //   be  %icc, success
aoqi@0:   //   add  %l2, 8, %l2
aoqi@0:   // loop:
aoqi@0:   //   ld  [ %l2 ], %l5
aoqi@0:   //   brz,pn   %l5, throw_icce
aoqi@0:   //   cmp  %l5, %g5
aoqi@0:   //   bne,pn   %icc, loop
aoqi@0:   //   add  %l2, 8, %l2
aoqi@0:   // success:
aoqi@0:   //   ld  [ %l2 + -4 ], %l2
aoqi@0:   //   ld  [ %g3 + %l2 ], %l5
aoqi@0:   //   restore  %l5, 0, %g5
aoqi@0:   //   ld  [ %g5 + 0x44 ], %g3
aoqi@0:   //   jmp  %g3
aoqi@0:   //   nop
aoqi@0:   // throw_icce:
aoqi@0:   //   sethi  %hi(throw_ICCE_entry), %g3
aoqi@0:   //   ! 5 more instructions here, LP64_ONLY
aoqi@0:   //   jmp  %g3 + %lo(throw_ICCE_entry)
aoqi@0:   //   restore
aoqi@0: }
aoqi@0: 
aoqi@0: 
aoqi@0: int VtableStub::pd_code_alignment() {
aoqi@0:   // UltraSPARC cache line size is 8 instructions:
aoqi@0:   const unsigned int icache_line_size = 32;
aoqi@0:   return icache_line_size;
aoqi@0: }