Java example source code file (stubGenerator_sparc.cpp)
The stubGenerator_sparc.cpp Java example source code
/*
* Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "asm/macroAssembler.inline.hpp"
#include "interpreter/interpreter.hpp"
#include "nativeInst_sparc.hpp"
#include "oops/instanceOop.hpp"
#include "oops/method.hpp"
#include "oops/objArrayKlass.hpp"
#include "oops/oop.inline.hpp"
#include "prims/methodHandles.hpp"
#include "runtime/frame.inline.hpp"
#include "runtime/handles.inline.hpp"
#include "runtime/sharedRuntime.hpp"
#include "runtime/stubCodeGenerator.hpp"
#include "runtime/stubRoutines.hpp"
#include "runtime/thread.inline.hpp"
#include "utilities/top.hpp"
#ifdef COMPILER2
#include "opto/runtime.hpp"
#endif
// Declaration and definition of StubGenerator (no .hpp file).
// For a more detailed description of the stub routine structure
// see the comment in stubRoutines.hpp.
#define __ _masm->
#ifdef PRODUCT
#define BLOCK_COMMENT(str) /* nothing */
#else
#define BLOCK_COMMENT(str) __ block_comment(str)
#endif
#define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
// Note: The register L7 is used as L7_thread_cache, and may not be used
// any other way within this module.
static const Register& Lstub_temp = L2;
// -------------------------------------------------------------------------------------------------------------------------
// Stub Code definitions
static address handle_unsafe_access() {
JavaThread* thread = JavaThread::current();
address pc = thread->saved_exception_pc();
address npc = thread->saved_exception_npc();
// pc is the instruction which we must emulate
// doing a no-op is fine: return garbage from the load
// request an async exception
thread->set_pending_unsafe_access_error();
// return address of next instruction to execute
return npc;
}
class StubGenerator: public StubCodeGenerator {
private:
#ifdef PRODUCT
#define inc_counter_np(a,b,c) (0)
#else
#define inc_counter_np(counter, t1, t2) \
BLOCK_COMMENT("inc_counter " #counter); \
__ inc_counter(&counter, t1, t2);
#endif
//----------------------------------------------------------------------------------------------------
// Call stubs are used to call Java from C
address generate_call_stub(address& return_pc) {
StubCodeMark mark(this, "StubRoutines", "call_stub");
address start = __ pc();
// Incoming arguments:
//
// o0 : call wrapper address
// o1 : result (address)
// o2 : result type
// o3 : method
// o4 : (interpreter) entry point
// o5 : parameters (address)
// [sp + 0x5c]: parameter size (in words)
// [sp + 0x60]: thread
//
// +---------------+ <--- sp + 0
// | |
// . reg save area .
// | |
// +---------------+ <--- sp + 0x40
// | |
// . extra 7 slots .
// | |
// +---------------+ <--- sp + 0x5c
// | param. size |
// +---------------+ <--- sp + 0x60
// | thread |
// +---------------+
// | |
// note: if the link argument position changes, adjust
// the code in frame::entry_frame_call_wrapper()
const Argument link = Argument(0, false); // used only for GC
const Argument result = Argument(1, false);
const Argument result_type = Argument(2, false);
const Argument method = Argument(3, false);
const Argument entry_point = Argument(4, false);
const Argument parameters = Argument(5, false);
const Argument parameter_size = Argument(6, false);
const Argument thread = Argument(7, false);
// setup thread register
__ ld_ptr(thread.as_address(), G2_thread);
__ reinit_heapbase();
#ifdef ASSERT
// make sure we have no pending exceptions
{ const Register t = G3_scratch;
Label L;
__ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), t);
__ br_null_short(t, Assembler::pt, L);
__ stop("StubRoutines::call_stub: entered with pending exception");
__ bind(L);
}
#endif
// create activation frame & allocate space for parameters
{ const Register t = G3_scratch;
__ ld_ptr(parameter_size.as_address(), t); // get parameter size (in words)
__ add(t, frame::memory_parameter_word_sp_offset, t); // add space for save area (in words)
__ round_to(t, WordsPerLong); // make sure it is multiple of 2 (in words)
__ sll(t, Interpreter::logStackElementSize, t); // compute number of bytes
__ neg(t); // negate so it can be used with save
__ save(SP, t, SP); // setup new frame
}
// +---------------+ <--- sp + 0
// | |
// . reg save area .
// | |
// +---------------+ <--- sp + 0x40
// | |
// . extra 7 slots .
// | |
// +---------------+ <--- sp + 0x5c
// | empty slot | (only if parameter size is even)
// +---------------+
// | |
// . parameters .
// | |
// +---------------+ <--- fp + 0
// | |
// . reg save area .
// | |
// +---------------+ <--- fp + 0x40
// | |
// . extra 7 slots .
// | |
// +---------------+ <--- fp + 0x5c
// | param. size |
// +---------------+ <--- fp + 0x60
// | thread |
// +---------------+
// | |
// pass parameters if any
BLOCK_COMMENT("pass parameters if any");
{ const Register src = parameters.as_in().as_register();
const Register dst = Lentry_args;
const Register tmp = G3_scratch;
const Register cnt = G4_scratch;
// test if any parameters & setup of Lentry_args
Label exit;
__ ld_ptr(parameter_size.as_in().as_address(), cnt); // parameter counter
__ add( FP, STACK_BIAS, dst );
__ cmp_zero_and_br(Assembler::zero, cnt, exit);
__ delayed()->sub(dst, BytesPerWord, dst); // setup Lentry_args
// copy parameters if any
Label loop;
__ BIND(loop);
// Store parameter value
__ ld_ptr(src, 0, tmp);
__ add(src, BytesPerWord, src);
__ st_ptr(tmp, dst, 0);
__ deccc(cnt);
__ br(Assembler::greater, false, Assembler::pt, loop);
__ delayed()->sub(dst, Interpreter::stackElementSize, dst);
// done
__ BIND(exit);
}
// setup parameters, method & call Java function
#ifdef ASSERT
// layout_activation_impl checks it's notion of saved SP against
// this register, so if this changes update it as well.
const Register saved_SP = Lscratch;
__ mov(SP, saved_SP); // keep track of SP before call
#endif
// setup parameters
const Register t = G3_scratch;
__ ld_ptr(parameter_size.as_in().as_address(), t); // get parameter size (in words)
__ sll(t, Interpreter::logStackElementSize, t); // compute number of bytes
__ sub(FP, t, Gargs); // setup parameter pointer
#ifdef _LP64
__ add( Gargs, STACK_BIAS, Gargs ); // Account for LP64 stack bias
#endif
__ mov(SP, O5_savedSP);
// do the call
//
// the following register must be setup:
//
// G2_thread
// G5_method
// Gargs
BLOCK_COMMENT("call Java function");
__ jmpl(entry_point.as_in().as_register(), G0, O7);
__ delayed()->mov(method.as_in().as_register(), G5_method); // setup method
BLOCK_COMMENT("call_stub_return_address:");
return_pc = __ pc();
// The callee, if it wasn't interpreted, can return with SP changed so
// we can no longer assert of change of SP.
// store result depending on type
// (everything that is not T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE
// is treated as T_INT)
{ const Register addr = result .as_in().as_register();
const Register type = result_type.as_in().as_register();
Label is_long, is_float, is_double, is_object, exit;
__ cmp(type, T_OBJECT); __ br(Assembler::equal, false, Assembler::pn, is_object);
__ delayed()->cmp(type, T_FLOAT); __ br(Assembler::equal, false, Assembler::pn, is_float);
__ delayed()->cmp(type, T_DOUBLE); __ br(Assembler::equal, false, Assembler::pn, is_double);
__ delayed()->cmp(type, T_LONG); __ br(Assembler::equal, false, Assembler::pn, is_long);
__ delayed()->nop();
// store int result
__ st(O0, addr, G0);
__ BIND(exit);
__ ret();
__ delayed()->restore();
__ BIND(is_object);
__ ba(exit);
__ delayed()->st_ptr(O0, addr, G0);
__ BIND(is_float);
__ ba(exit);
__ delayed()->stf(FloatRegisterImpl::S, F0, addr, G0);
__ BIND(is_double);
__ ba(exit);
__ delayed()->stf(FloatRegisterImpl::D, F0, addr, G0);
__ BIND(is_long);
#ifdef _LP64
__ ba(exit);
__ delayed()->st_long(O0, addr, G0); // store entire long
#else
#if defined(COMPILER2)
// All return values are where we want them, except for Longs. C2 returns
// longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
// Since the interpreter will return longs in G1 and O0/O1 in the 32bit
// build we simply always use G1.
// Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
// do this here. Unfortunately if we did a rethrow we'd see an machepilog node
// first which would move g1 -> O0/O1 and destroy the exception we were throwing.
__ ba(exit);
__ delayed()->stx(G1, addr, G0); // store entire long
#else
__ st(O1, addr, BytesPerInt);
__ ba(exit);
__ delayed()->st(O0, addr, G0);
#endif /* COMPILER2 */
#endif /* _LP64 */
}
return start;
}
//----------------------------------------------------------------------------------------------------
// Return point for a Java call if there's an exception thrown in Java code.
// The exception is caught and transformed into a pending exception stored in
// JavaThread that can be tested from within the VM.
//
// Oexception: exception oop
address generate_catch_exception() {
StubCodeMark mark(this, "StubRoutines", "catch_exception");
address start = __ pc();
// verify that thread corresponds
__ verify_thread();
const Register& temp_reg = Gtemp;
Address pending_exception_addr (G2_thread, Thread::pending_exception_offset());
Address exception_file_offset_addr(G2_thread, Thread::exception_file_offset ());
Address exception_line_offset_addr(G2_thread, Thread::exception_line_offset ());
// set pending exception
__ verify_oop(Oexception);
__ st_ptr(Oexception, pending_exception_addr);
__ set((intptr_t)__FILE__, temp_reg);
__ st_ptr(temp_reg, exception_file_offset_addr);
__ set((intptr_t)__LINE__, temp_reg);
__ st(temp_reg, exception_line_offset_addr);
// complete return to VM
assert(StubRoutines::_call_stub_return_address != NULL, "must have been generated before");
AddressLiteral stub_ret(StubRoutines::_call_stub_return_address);
__ jump_to(stub_ret, temp_reg);
__ delayed()->nop();
return start;
}
//----------------------------------------------------------------------------------------------------
// Continuation point for runtime calls returning with a pending exception
// The pending exception check happened in the runtime or native call stub
// The pending exception in Thread is converted into a Java-level exception
//
// Contract with Java-level exception handler: O0 = exception
// O1 = throwing pc
address generate_forward_exception() {
StubCodeMark mark(this, "StubRoutines", "forward_exception");
address start = __ pc();
// Upon entry, O7 has the return address returning into Java
// (interpreted or compiled) code; i.e. the return address
// becomes the throwing pc.
const Register& handler_reg = Gtemp;
Address exception_addr(G2_thread, Thread::pending_exception_offset());
#ifdef ASSERT
// make sure that this code is only executed if there is a pending exception
{ Label L;
__ ld_ptr(exception_addr, Gtemp);
__ br_notnull_short(Gtemp, Assembler::pt, L);
__ stop("StubRoutines::forward exception: no pending exception (1)");
__ bind(L);
}
#endif
// compute exception handler into handler_reg
__ get_thread();
__ ld_ptr(exception_addr, Oexception);
__ verify_oop(Oexception);
__ save_frame(0); // compensates for compiler weakness
__ add(O7->after_save(), frame::pc_return_offset, Lscratch); // save the issuing PC
BLOCK_COMMENT("call exception_handler_for_return_address");
__ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), G2_thread, Lscratch);
__ mov(O0, handler_reg);
__ restore(); // compensates for compiler weakness
__ ld_ptr(exception_addr, Oexception);
__ add(O7, frame::pc_return_offset, Oissuing_pc); // save the issuing PC
#ifdef ASSERT
// make sure exception is set
{ Label L;
__ br_notnull_short(Oexception, Assembler::pt, L);
__ stop("StubRoutines::forward exception: no pending exception (2)");
__ bind(L);
}
#endif
// jump to exception handler
__ jmp(handler_reg, 0);
// clear pending exception
__ delayed()->st_ptr(G0, exception_addr);
return start;
}
// Safefetch stubs.
void generate_safefetch(const char* name, int size, address* entry,
address* fault_pc, address* continuation_pc) {
// safefetch signatures:
// int SafeFetch32(int* adr, int errValue);
// intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
//
// arguments:
// o0 = adr
// o1 = errValue
//
// result:
// o0 = *adr or errValue
StubCodeMark mark(this, "StubRoutines", name);
// Entry point, pc or function descriptor.
__ align(CodeEntryAlignment);
*entry = __ pc();
__ mov(O0, G1); // g1 = o0
__ mov(O1, O0); // o0 = o1
// Load *adr into c_rarg1, may fault.
*fault_pc = __ pc();
switch (size) {
case 4:
// int32_t
__ ldsw(G1, 0, O0); // o0 = [g1]
break;
case 8:
// int64_t
__ ldx(G1, 0, O0); // o0 = [g1]
break;
default:
ShouldNotReachHere();
}
// return errValue or *adr
*continuation_pc = __ pc();
// By convention with the trap handler we ensure there is a non-CTI
// instruction in the trap shadow.
__ nop();
__ retl();
__ delayed()->nop();
}
//------------------------------------------------------------------------------------------------------------------------
// Continuation point for throwing of implicit exceptions that are not handled in
// the current activation. Fabricates an exception oop and initiates normal
// exception dispatching in this frame. Only callee-saved registers are preserved
// (through the normal register window / RegisterMap handling).
// If the compiler needs all registers to be preserved between the fault
// point and the exception handler then it must assume responsibility for that in
// AbstractCompiler::continuation_for_implicit_null_exception or
// continuation_for_implicit_division_by_zero_exception. All other implicit
// exceptions (e.g., NullPointerException or AbstractMethodError on entry) are
// either at call sites or otherwise assume that stack unwinding will be initiated,
// so caller saved registers were assumed volatile in the compiler.
// Note that we generate only this stub into a RuntimeStub, because it needs to be
// properly traversed and ignored during GC, so we change the meaning of the "__"
// macro within this method.
#undef __
#define __ masm->
address generate_throw_exception(const char* name, address runtime_entry,
Register arg1 = noreg, Register arg2 = noreg) {
#ifdef ASSERT
int insts_size = VerifyThread ? 1 * K : 600;
#else
int insts_size = VerifyThread ? 1 * K : 256;
#endif /* ASSERT */
int locs_size = 32;
CodeBuffer code(name, insts_size, locs_size);
MacroAssembler* masm = new MacroAssembler(&code);
__ verify_thread();
// This is an inlined and slightly modified version of call_VM
// which has the ability to fetch the return PC out of thread-local storage
__ assert_not_delayed();
// Note that we always push a frame because on the SPARC
// architecture, for all of our implicit exception kinds at call
// sites, the implicit exception is taken before the callee frame
// is pushed.
__ save_frame(0);
int frame_complete = __ offset();
// Note that we always have a runtime stub frame on the top of stack by this point
Register last_java_sp = SP;
// 64-bit last_java_sp is biased!
__ set_last_Java_frame(last_java_sp, G0);
if (VerifyThread) __ mov(G2_thread, O0); // about to be smashed; pass early
__ save_thread(noreg);
if (arg1 != noreg) {
assert(arg2 != O1, "clobbered");
__ mov(arg1, O1);
}
if (arg2 != noreg) {
__ mov(arg2, O2);
}
// do the call
BLOCK_COMMENT("call runtime_entry");
__ call(runtime_entry, relocInfo::runtime_call_type);
if (!VerifyThread)
__ delayed()->mov(G2_thread, O0); // pass thread as first argument
else
__ delayed()->nop(); // (thread already passed)
__ restore_thread(noreg);
__ reset_last_Java_frame();
// check for pending exceptions. use Gtemp as scratch register.
#ifdef ASSERT
Label L;
Address exception_addr(G2_thread, Thread::pending_exception_offset());
Register scratch_reg = Gtemp;
__ ld_ptr(exception_addr, scratch_reg);
__ br_notnull_short(scratch_reg, Assembler::pt, L);
__ should_not_reach_here();
__ bind(L);
#endif // ASSERT
BLOCK_COMMENT("call forward_exception_entry");
__ call(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
// we use O7 linkage so that forward_exception_entry has the issuing PC
__ delayed()->restore();
RuntimeStub* stub = RuntimeStub::new_runtime_stub(name, &code, frame_complete, masm->total_frame_size_in_bytes(0), NULL, false);
return stub->entry_point();
}
#undef __
#define __ _masm->
// Generate a routine that sets all the registers so we
// can tell if the stop routine prints them correctly.
address generate_test_stop() {
StubCodeMark mark(this, "StubRoutines", "test_stop");
address start = __ pc();
int i;
__ save_frame(0);
static jfloat zero = 0.0, one = 1.0;
// put addr in L0, then load through L0 to F0
__ set((intptr_t)&zero, L0); __ ldf( FloatRegisterImpl::S, L0, 0, F0);
__ set((intptr_t)&one, L0); __ ldf( FloatRegisterImpl::S, L0, 0, F1); // 1.0 to F1
// use add to put 2..18 in F2..F18
for ( i = 2; i <= 18; ++i ) {
__ fadd( FloatRegisterImpl::S, F1, as_FloatRegister(i-1), as_FloatRegister(i));
}
// Now put double 2 in F16, double 18 in F18
__ ftof( FloatRegisterImpl::S, FloatRegisterImpl::D, F2, F16 );
__ ftof( FloatRegisterImpl::S, FloatRegisterImpl::D, F18, F18 );
// use add to put 20..32 in F20..F32
for (i = 20; i < 32; i += 2) {
__ fadd( FloatRegisterImpl::D, F16, as_FloatRegister(i-2), as_FloatRegister(i));
}
// put 0..7 in i's, 8..15 in l's, 16..23 in o's, 24..31 in g's
for ( i = 0; i < 8; ++i ) {
if (i < 6) {
__ set( i, as_iRegister(i));
__ set(16 + i, as_oRegister(i));
__ set(24 + i, as_gRegister(i));
}
__ set( 8 + i, as_lRegister(i));
}
__ stop("testing stop");
__ ret();
__ delayed()->restore();
return start;
}
address generate_stop_subroutine() {
StubCodeMark mark(this, "StubRoutines", "stop_subroutine");
address start = __ pc();
__ stop_subroutine();
return start;
}
address generate_flush_callers_register_windows() {
StubCodeMark mark(this, "StubRoutines", "flush_callers_register_windows");
address start = __ pc();
__ flushw();
__ retl(false);
__ delayed()->add( FP, STACK_BIAS, O0 );
// The returned value must be a stack pointer whose register save area
// is flushed, and will stay flushed while the caller executes.
return start;
}
// Support for jint Atomic::xchg(jint exchange_value, volatile jint* dest).
//
// Arguments:
//
// exchange_value: O0
// dest: O1
//
// Results:
//
// O0: the value previously stored in dest
//
address generate_atomic_xchg() {
StubCodeMark mark(this, "StubRoutines", "atomic_xchg");
address start = __ pc();
if (UseCASForSwap) {
// Use CAS instead of swap, just in case the MP hardware
// prefers to work with just one kind of synch. instruction.
Label retry;
__ BIND(retry);
__ mov(O0, O3); // scratch copy of exchange value
__ ld(O1, 0, O2); // observe the previous value
// try to replace O2 with O3
__ cas(O1, O2, O3);
__ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pn, retry);
__ retl(false);
__ delayed()->mov(O2, O0); // report previous value to caller
} else {
__ retl(false);
__ delayed()->swap(O1, 0, O0);
}
return start;
}
// Support for jint Atomic::cmpxchg(jint exchange_value, volatile jint* dest, jint compare_value)
//
// Arguments:
//
// exchange_value: O0
// dest: O1
// compare_value: O2
//
// Results:
//
// O0: the value previously stored in dest
//
address generate_atomic_cmpxchg() {
StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg");
address start = __ pc();
// cmpxchg(dest, compare_value, exchange_value)
__ cas(O1, O2, O0);
__ retl(false);
__ delayed()->nop();
return start;
}
// Support for jlong Atomic::cmpxchg(jlong exchange_value, volatile jlong *dest, jlong compare_value)
//
// Arguments:
//
// exchange_value: O1:O0
// dest: O2
// compare_value: O4:O3
//
// Results:
//
// O1:O0: the value previously stored in dest
//
// Overwrites: G1,G2,G3
//
address generate_atomic_cmpxchg_long() {
StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long");
address start = __ pc();
__ sllx(O0, 32, O0);
__ srl(O1, 0, O1);
__ or3(O0,O1,O0); // O0 holds 64-bit value from compare_value
__ sllx(O3, 32, O3);
__ srl(O4, 0, O4);
__ or3(O3,O4,O3); // O3 holds 64-bit value from exchange_value
__ casx(O2, O3, O0);
__ srl(O0, 0, O1); // unpacked return value in O1:O0
__ retl(false);
__ delayed()->srlx(O0, 32, O0);
return start;
}
// Support for jint Atomic::add(jint add_value, volatile jint* dest).
//
// Arguments:
//
// add_value: O0 (e.g., +1 or -1)
// dest: O1
//
// Results:
//
// O0: the new value stored in dest
//
// Overwrites: O3
//
address generate_atomic_add() {
StubCodeMark mark(this, "StubRoutines", "atomic_add");
address start = __ pc();
__ BIND(_atomic_add_stub);
Label(retry);
__ BIND(retry);
__ lduw(O1, 0, O2);
__ add(O0, O2, O3);
__ cas(O1, O2, O3);
__ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pn, retry);
__ retl(false);
__ delayed()->add(O0, O2, O0); // note that cas made O2==O3
return start;
}
Label _atomic_add_stub; // called from other stubs
//------------------------------------------------------------------------------------------------------------------------
// The following routine generates a subroutine to throw an asynchronous
// UnknownError when an unsafe access gets a fault that could not be
// reasonably prevented by the programmer. (Example: SIGBUS/OBJERR.)
//
// Arguments :
//
// trapping PC: O7
//
// Results:
// posts an asynchronous exception, skips the trapping instruction
//
address generate_handler_for_unsafe_access() {
StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
address start = __ pc();
const int preserve_register_words = (64 * 2);
Address preserve_addr(FP, (-preserve_register_words * wordSize) + STACK_BIAS);
Register Lthread = L7_thread_cache;
int i;
__ save_frame(0);
__ mov(G1, L1);
__ mov(G2, L2);
__ mov(G3, L3);
__ mov(G4, L4);
__ mov(G5, L5);
for (i = 0; i < 64; i += 2) {
__ stf(FloatRegisterImpl::D, as_FloatRegister(i), preserve_addr, i * wordSize);
}
address entry_point = CAST_FROM_FN_PTR(address, handle_unsafe_access);
BLOCK_COMMENT("call handle_unsafe_access");
__ call(entry_point, relocInfo::runtime_call_type);
__ delayed()->nop();
__ mov(L1, G1);
__ mov(L2, G2);
__ mov(L3, G3);
__ mov(L4, G4);
__ mov(L5, G5);
for (i = 0; i < 64; i += 2) {
__ ldf(FloatRegisterImpl::D, preserve_addr, as_FloatRegister(i), i * wordSize);
}
__ verify_thread();
__ jmp(O0, 0);
__ delayed()->restore();
return start;
}
// Support for uint StubRoutine::Sparc::partial_subtype_check( Klass sub, Klass super );
// Arguments :
//
// ret : O0, returned
// icc/xcc: set as O0 (depending on wordSize)
// sub : O1, argument, not changed
// super: O2, argument, not changed
// raddr: O7, blown by call
address generate_partial_subtype_check() {
__ align(CodeEntryAlignment);
StubCodeMark mark(this, "StubRoutines", "partial_subtype_check");
address start = __ pc();
Label miss;
#if defined(COMPILER2) && !defined(_LP64)
// Do not use a 'save' because it blows the 64-bit O registers.
__ add(SP,-4*wordSize,SP); // Make space for 4 temps (stack must be 2 words aligned)
__ st_ptr(L0,SP,(frame::register_save_words+0)*wordSize);
__ st_ptr(L1,SP,(frame::register_save_words+1)*wordSize);
__ st_ptr(L2,SP,(frame::register_save_words+2)*wordSize);
__ st_ptr(L3,SP,(frame::register_save_words+3)*wordSize);
Register Rret = O0;
Register Rsub = O1;
Register Rsuper = O2;
#else
__ save_frame(0);
Register Rret = I0;
Register Rsub = I1;
Register Rsuper = I2;
#endif
Register L0_ary_len = L0;
Register L1_ary_ptr = L1;
Register L2_super = L2;
Register L3_index = L3;
__ check_klass_subtype_slow_path(Rsub, Rsuper,
L0, L1, L2, L3,
NULL, &miss);
// Match falls through here.
__ addcc(G0,0,Rret); // set Z flags, Z result
#if defined(COMPILER2) && !defined(_LP64)
__ ld_ptr(SP,(frame::register_save_words+0)*wordSize,L0);
__ ld_ptr(SP,(frame::register_save_words+1)*wordSize,L1);
__ ld_ptr(SP,(frame::register_save_words+2)*wordSize,L2);
__ ld_ptr(SP,(frame::register_save_words+3)*wordSize,L3);
__ retl(); // Result in Rret is zero; flags set to Z
__ delayed()->add(SP,4*wordSize,SP);
#else
__ ret(); // Result in Rret is zero; flags set to Z
__ delayed()->restore();
#endif
__ BIND(miss);
__ addcc(G0,1,Rret); // set NZ flags, NZ result
#if defined(COMPILER2) && !defined(_LP64)
__ ld_ptr(SP,(frame::register_save_words+0)*wordSize,L0);
__ ld_ptr(SP,(frame::register_save_words+1)*wordSize,L1);
__ ld_ptr(SP,(frame::register_save_words+2)*wordSize,L2);
__ ld_ptr(SP,(frame::register_save_words+3)*wordSize,L3);
__ retl(); // Result in Rret is != 0; flags set to NZ
__ delayed()->add(SP,4*wordSize,SP);
#else
__ ret(); // Result in Rret is != 0; flags set to NZ
__ delayed()->restore();
#endif
return start;
}
// Called from MacroAssembler::verify_oop
//
address generate_verify_oop_subroutine() {
StubCodeMark mark(this, "StubRoutines", "verify_oop_stub");
address start = __ pc();
__ verify_oop_subroutine();
return start;
}
//
// Verify that a register contains clean 32-bits positive value
// (high 32-bits are 0) so it could be used in 64-bits shifts (sllx, srax).
//
// Input:
// Rint - 32-bits value
// Rtmp - scratch
//
void assert_clean_int(Register Rint, Register Rtmp) {
#if defined(ASSERT) && defined(_LP64)
__ signx(Rint, Rtmp);
__ cmp(Rint, Rtmp);
__ breakpoint_trap(Assembler::notEqual, Assembler::xcc);
#endif
}
//
// Generate overlap test for array copy stubs
//
// Input:
// O0 - array1
// O1 - array2
// O2 - element count
//
// Kills temps: O3, O4
//
void array_overlap_test(address no_overlap_target, int log2_elem_size) {
assert(no_overlap_target != NULL, "must be generated");
array_overlap_test(no_overlap_target, NULL, log2_elem_size);
}
void array_overlap_test(Label& L_no_overlap, int log2_elem_size) {
array_overlap_test(NULL, &L_no_overlap, log2_elem_size);
}
void array_overlap_test(address no_overlap_target, Label* NOLp, int log2_elem_size) {
const Register from = O0;
const Register to = O1;
const Register count = O2;
const Register to_from = O3; // to - from
const Register byte_count = O4; // count << log2_elem_size
__ subcc(to, from, to_from);
__ sll_ptr(count, log2_elem_size, byte_count);
if (NOLp == NULL)
__ brx(Assembler::lessEqualUnsigned, false, Assembler::pt, no_overlap_target);
else
__ brx(Assembler::lessEqualUnsigned, false, Assembler::pt, (*NOLp));
__ delayed()->cmp(to_from, byte_count);
if (NOLp == NULL)
__ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, no_overlap_target);
else
__ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, (*NOLp));
__ delayed()->nop();
}
//
// Generate pre-write barrier for array.
//
// Input:
// addr - register containing starting address
// count - register containing element count
// tmp - scratch register
//
// The input registers are overwritten.
//
void gen_write_ref_array_pre_barrier(Register addr, Register count, bool dest_uninitialized) {
BarrierSet* bs = Universe::heap()->barrier_set();
switch (bs->kind()) {
case BarrierSet::G1SATBCT:
case BarrierSet::G1SATBCTLogging:
// With G1, don't generate the call if we statically know that the target in uninitialized
if (!dest_uninitialized) {
__ save_frame(0);
// Save the necessary global regs... will be used after.
if (addr->is_global()) {
__ mov(addr, L0);
}
if (count->is_global()) {
__ mov(count, L1);
}
__ mov(addr->after_save(), O0);
// Get the count into O1
__ call(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre));
__ delayed()->mov(count->after_save(), O1);
if (addr->is_global()) {
__ mov(L0, addr);
}
if (count->is_global()) {
__ mov(L1, count);
}
__ restore();
}
break;
case BarrierSet::CardTableModRef:
case BarrierSet::CardTableExtension:
case BarrierSet::ModRef:
break;
default:
ShouldNotReachHere();
}
}
//
// Generate post-write barrier for array.
//
// Input:
// addr - register containing starting address
// count - register containing element count
// tmp - scratch register
//
// The input registers are overwritten.
//
void gen_write_ref_array_post_barrier(Register addr, Register count,
Register tmp) {
BarrierSet* bs = Universe::heap()->barrier_set();
switch (bs->kind()) {
case BarrierSet::G1SATBCT:
case BarrierSet::G1SATBCTLogging:
{
// Get some new fresh output registers.
__ save_frame(0);
__ mov(addr->after_save(), O0);
__ call(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post));
__ delayed()->mov(count->after_save(), O1);
__ restore();
}
break;
case BarrierSet::CardTableModRef:
case BarrierSet::CardTableExtension:
{
CardTableModRefBS* ct = (CardTableModRefBS*)bs;
assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
assert_different_registers(addr, count, tmp);
Label L_loop;
__ sll_ptr(count, LogBytesPerHeapOop, count);
__ sub(count, BytesPerHeapOop, count);
__ add(count, addr, count);
// Use two shifts to clear out those low order two bits! (Cannot opt. into 1.)
__ srl_ptr(addr, CardTableModRefBS::card_shift, addr);
__ srl_ptr(count, CardTableModRefBS::card_shift, count);
__ sub(count, addr, count);
AddressLiteral rs(ct->byte_map_base);
__ set(rs, tmp);
__ BIND(L_loop);
__ stb(G0, tmp, addr);
__ subcc(count, 1, count);
__ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
__ delayed()->add(addr, 1, addr);
}
break;
case BarrierSet::ModRef:
break;
default:
ShouldNotReachHere();
}
}
//
// Generate main code for disjoint arraycopy
//
typedef void (StubGenerator::*CopyLoopFunc)(Register from, Register to, Register count, int count_dec,
Label& L_loop, bool use_prefetch, bool use_bis);
void disjoint_copy_core(Register from, Register to, Register count, int log2_elem_size,
int iter_size, CopyLoopFunc copy_loop_func) {
Label L_copy;
assert(log2_elem_size <= 3, "the following code should be changed");
int count_dec = 16>>log2_elem_size;
int prefetch_dist = MAX2(ArraycopySrcPrefetchDistance, ArraycopyDstPrefetchDistance);
assert(prefetch_dist < 4096, "invalid value");
prefetch_dist = (prefetch_dist + (iter_size-1)) & (-iter_size); // round up to one iteration copy size
int prefetch_count = (prefetch_dist >> log2_elem_size); // elements count
if (UseBlockCopy) {
Label L_block_copy, L_block_copy_prefetch, L_skip_block_copy;
// 64 bytes tail + bytes copied in one loop iteration
int tail_size = 64 + iter_size;
int block_copy_count = (MAX2(tail_size, (int)BlockCopyLowLimit)) >> log2_elem_size;
// Use BIS copy only for big arrays since it requires membar.
__ set(block_copy_count, O4);
__ cmp_and_br_short(count, O4, Assembler::lessUnsigned, Assembler::pt, L_skip_block_copy);
// This code is for disjoint source and destination:
// to <= from || to >= from+count
// but BIS will stomp over 'from' if (to > from-tail_size && to <= from)
__ sub(from, to, O4);
__ srax(O4, 4, O4); // divide by 16 since following short branch have only 5 bits for imm.
__ cmp_and_br_short(O4, (tail_size>>4), Assembler::lessEqualUnsigned, Assembler::pn, L_skip_block_copy);
__ wrasi(G0, Assembler::ASI_ST_BLKINIT_PRIMARY);
// BIS should not be used to copy tail (64 bytes+iter_size)
// to avoid zeroing of following values.
__ sub(count, (tail_size>>log2_elem_size), count); // count is still positive >= 0
if (prefetch_count > 0) { // rounded up to one iteration count
// Do prefetching only if copy size is bigger
// than prefetch distance.
__ set(prefetch_count, O4);
__ cmp_and_brx_short(count, O4, Assembler::less, Assembler::pt, L_block_copy);
__ sub(count, prefetch_count, count);
(this->*copy_loop_func)(from, to, count, count_dec, L_block_copy_prefetch, true, true);
__ add(count, prefetch_count, count); // restore count
} // prefetch_count > 0
(this->*copy_loop_func)(from, to, count, count_dec, L_block_copy, false, true);
__ add(count, (tail_size>>log2_elem_size), count); // restore count
__ wrasi(G0, Assembler::ASI_PRIMARY_NOFAULT);
// BIS needs membar.
__ membar(Assembler::StoreLoad);
// Copy tail
__ ba_short(L_copy);
__ BIND(L_skip_block_copy);
} // UseBlockCopy
if (prefetch_count > 0) { // rounded up to one iteration count
// Do prefetching only if copy size is bigger
// than prefetch distance.
__ set(prefetch_count, O4);
__ cmp_and_brx_short(count, O4, Assembler::lessUnsigned, Assembler::pt, L_copy);
__ sub(count, prefetch_count, count);
Label L_copy_prefetch;
(this->*copy_loop_func)(from, to, count, count_dec, L_copy_prefetch, true, false);
__ add(count, prefetch_count, count); // restore count
} // prefetch_count > 0
(this->*copy_loop_func)(from, to, count, count_dec, L_copy, false, false);
}
//
// Helper methods for copy_16_bytes_forward_with_shift()
//
void copy_16_bytes_shift_loop(Register from, Register to, Register count, int count_dec,
Label& L_loop, bool use_prefetch, bool use_bis) {
const Register left_shift = G1; // left shift bit counter
const Register right_shift = G5; // right shift bit counter
__ align(OptoLoopAlignment);
__ BIND(L_loop);
if (use_prefetch) {
if (ArraycopySrcPrefetchDistance > 0) {
__ prefetch(from, ArraycopySrcPrefetchDistance, Assembler::severalReads);
}
if (ArraycopyDstPrefetchDistance > 0) {
__ prefetch(to, ArraycopyDstPrefetchDistance, Assembler::severalWritesAndPossiblyReads);
}
}
__ ldx(from, 0, O4);
__ ldx(from, 8, G4);
__ inc(to, 16);
__ inc(from, 16);
__ deccc(count, count_dec); // Can we do next iteration after this one?
__ srlx(O4, right_shift, G3);
__ bset(G3, O3);
__ sllx(O4, left_shift, O4);
__ srlx(G4, right_shift, G3);
__ bset(G3, O4);
if (use_bis) {
__ stxa(O3, to, -16);
__ stxa(O4, to, -8);
} else {
__ stx(O3, to, -16);
__ stx(O4, to, -8);
}
__ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
__ delayed()->sllx(G4, left_shift, O3);
}
// Copy big chunks forward with shift
//
// Inputs:
// from - source arrays
// to - destination array aligned to 8-bytes
// count - elements count to copy >= the count equivalent to 16 bytes
// count_dec - elements count's decrement equivalent to 16 bytes
// L_copy_bytes - copy exit label
//
void copy_16_bytes_forward_with_shift(Register from, Register to,
Register count, int log2_elem_size, Label& L_copy_bytes) {
Label L_aligned_copy, L_copy_last_bytes;
assert(log2_elem_size <= 3, "the following code should be changed");
int count_dec = 16>>log2_elem_size;
// if both arrays have the same alignment mod 8, do 8 bytes aligned copy
__ andcc(from, 7, G1); // misaligned bytes
__ br(Assembler::zero, false, Assembler::pt, L_aligned_copy);
__ delayed()->nop();
const Register left_shift = G1; // left shift bit counter
const Register right_shift = G5; // right shift bit counter
__ sll(G1, LogBitsPerByte, left_shift);
__ mov(64, right_shift);
__ sub(right_shift, left_shift, right_shift);
//
// Load 2 aligned 8-bytes chunks and use one from previous iteration
// to form 2 aligned 8-bytes chunks to store.
//
__ dec(count, count_dec); // Pre-decrement 'count'
__ andn(from, 7, from); // Align address
__ ldx(from, 0, O3);
__ inc(from, 8);
__ sllx(O3, left_shift, O3);
disjoint_copy_core(from, to, count, log2_elem_size, 16, copy_16_bytes_shift_loop);
__ inccc(count, count_dec>>1 ); // + 8 bytes
__ brx(Assembler::negative, true, Assembler::pn, L_copy_last_bytes);
__ delayed()->inc(count, count_dec>>1); // restore 'count'
// copy 8 bytes, part of them already loaded in O3
__ ldx(from, 0, O4);
__ inc(to, 8);
__ inc(from, 8);
__ srlx(O4, right_shift, G3);
__ bset(O3, G3);
__ stx(G3, to, -8);
__ BIND(L_copy_last_bytes);
__ srl(right_shift, LogBitsPerByte, right_shift); // misaligned bytes
__ br(Assembler::always, false, Assembler::pt, L_copy_bytes);
__ delayed()->sub(from, right_shift, from); // restore address
__ BIND(L_aligned_copy);
}
// Copy big chunks backward with shift
//
// Inputs:
// end_from - source arrays end address
// end_to - destination array end address aligned to 8-bytes
// count - elements count to copy >= the count equivalent to 16 bytes
// count_dec - elements count's decrement equivalent to 16 bytes
// L_aligned_copy - aligned copy exit label
// L_copy_bytes - copy exit label
//
void copy_16_bytes_backward_with_shift(Register end_from, Register end_to,
Register count, int count_dec,
Label& L_aligned_copy, Label& L_copy_bytes) {
Label L_loop, L_copy_last_bytes;
// if both arrays have the same alignment mod 8, do 8 bytes aligned copy
__ andcc(end_from, 7, G1); // misaligned bytes
__ br(Assembler::zero, false, Assembler::pt, L_aligned_copy);
__ delayed()->deccc(count, count_dec); // Pre-decrement 'count'
const Register left_shift = G1; // left shift bit counter
const Register right_shift = G5; // right shift bit counter
__ sll(G1, LogBitsPerByte, left_shift);
__ mov(64, right_shift);
__ sub(right_shift, left_shift, right_shift);
//
// Load 2 aligned 8-bytes chunks and use one from previous iteration
// to form 2 aligned 8-bytes chunks to store.
//
__ andn(end_from, 7, end_from); // Align address
__ ldx(end_from, 0, O3);
__ align(OptoLoopAlignment);
__ BIND(L_loop);
__ ldx(end_from, -8, O4);
__ deccc(count, count_dec); // Can we do next iteration after this one?
__ ldx(end_from, -16, G4);
__ dec(end_to, 16);
__ dec(end_from, 16);
__ srlx(O3, right_shift, O3);
__ sllx(O4, left_shift, G3);
__ bset(G3, O3);
__ stx(O3, end_to, 8);
__ srlx(O4, right_shift, O4);
__ sllx(G4, left_shift, G3);
__ bset(G3, O4);
__ stx(O4, end_to, 0);
__ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
__ delayed()->mov(G4, O3);
__ inccc(count, count_dec>>1 ); // + 8 bytes
__ brx(Assembler::negative, true, Assembler::pn, L_copy_last_bytes);
__ delayed()->inc(count, count_dec>>1); // restore 'count'
// copy 8 bytes, part of them already loaded in O3
__ ldx(end_from, -8, O4);
__ dec(end_to, 8);
__ dec(end_from, 8);
__ srlx(O3, right_shift, O3);
__ sllx(O4, left_shift, G3);
__ bset(O3, G3);
__ stx(G3, end_to, 0);
__ BIND(L_copy_last_bytes);
__ srl(left_shift, LogBitsPerByte, left_shift); // misaligned bytes
__ br(Assembler::always, false, Assembler::pt, L_copy_bytes);
__ delayed()->add(end_from, left_shift, end_from); // restore address
}
//
// Generate stub for disjoint byte copy. If "aligned" is true, the
// "from" and "to" addresses are assumed to be heapword aligned.
//
// Arguments for generated stub:
// from: O0
// to: O1
// count: O2 treated as signed
//
address generate_disjoint_byte_copy(bool aligned, address *entry, const char *name) {
__ align(CodeEntryAlignment);
StubCodeMark mark(this, "StubRoutines", name);
address start = __ pc();
Label L_skip_alignment, L_align;
Label L_copy_byte, L_copy_byte_loop, L_exit;
const Register from = O0; // source array address
const Register to = O1; // destination array address
const Register count = O2; // elements count
const Register offset = O5; // offset from start of arrays
// O3, O4, G3, G4 are used as temp registers
assert_clean_int(count, O3); // Make sure 'count' is clean int.
if (entry != NULL) {
*entry = __ pc();
// caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
BLOCK_COMMENT("Entry:");
}
// for short arrays, just do single element copy
__ cmp(count, 23); // 16 + 7
__ brx(Assembler::less, false, Assembler::pn, L_copy_byte);
__ delayed()->mov(G0, offset);
if (aligned) {
// 'aligned' == true when it is known statically during compilation
// of this arraycopy call site that both 'from' and 'to' addresses
// are HeapWordSize aligned (see LibraryCallKit::basictype2arraycopy()).
//
// Aligned arrays have 4 bytes alignment in 32-bits VM
// and 8 bytes - in 64-bits VM. So we do it only for 32-bits VM
//
#ifndef _LP64
// copy a 4-bytes word if necessary to align 'to' to 8 bytes
__ andcc(to, 7, G0);
__ br(Assembler::zero, false, Assembler::pn, L_skip_alignment);
__ delayed()->ld(from, 0, O3);
__ inc(from, 4);
__ inc(to, 4);
__ dec(count, 4);
__ st(O3, to, -4);
__ BIND(L_skip_alignment);
#endif
} else {
// copy bytes to align 'to' on 8 byte boundary
__ andcc(to, 7, G1); // misaligned bytes
__ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
__ delayed()->neg(G1);
__ inc(G1, 8); // bytes need to copy to next 8-bytes alignment
__ sub(count, G1, count);
__ BIND(L_align);
__ ldub(from, 0, O3);
__ deccc(G1);
__ inc(from);
__ stb(O3, to, 0);
__ br(Assembler::notZero, false, Assembler::pt, L_align);
__ delayed()->inc(to);
__ BIND(L_skip_alignment);
}
#ifdef _LP64
if (!aligned)
#endif
{
// Copy with shift 16 bytes per iteration if arrays do not have
// the same alignment mod 8, otherwise fall through to the next
// code for aligned copy.
// The compare above (count >= 23) guarantes 'count' >= 16 bytes.
// Also jump over aligned copy after the copy with shift completed.
copy_16_bytes_forward_with_shift(from, to, count, 0, L_copy_byte);
}
// Both array are 8 bytes aligned, copy 16 bytes at a time
__ and3(count, 7, G4); // Save count
__ srl(count, 3, count);
generate_disjoint_long_copy_core(aligned);
__ mov(G4, count); // Restore count
// copy tailing bytes
__ BIND(L_copy_byte);
__ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
__ align(OptoLoopAlignment);
__ BIND(L_copy_byte_loop);
__ ldub(from, offset, O3);
__ deccc(count);
__ stb(O3, to, offset);
__ brx(Assembler::notZero, false, Assembler::pt, L_copy_byte_loop);
__ delayed()->inc(offset);
__ BIND(L_exit);
// O3, O4 are used as temp registers
inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr, O3, O4);
__ retl();
__ delayed()->mov(G0, O0); // return 0
return start;
}
//
// Generate stub for conjoint byte copy. If "aligned" is true, the
// "from" and "to" addresses are assumed to be heapword aligned.
//
// Arguments for generated stub:
// from: O0
// to: O1
// count: O2 treated as signed
//
address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
address *entry, const char *name) {
// Do reverse copy.
__ align(CodeEntryAlignment);
StubCodeMark mark(this, "StubRoutines", name);
address start = __ pc();
Label L_skip_alignment, L_align, L_aligned_copy;
Label L_copy_byte, L_copy_byte_loop, L_exit;
const Register from = O0; // source array address
const Register to = O1; // destination array address
const Register count = O2; // elements count
const Register end_from = from; // source array end address
const Register end_to = to; // destination array end address
assert_clean_int(count, O3); // Make sure 'count' is clean int.
if (entry != NULL) {
*entry = __ pc();
// caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
BLOCK_COMMENT("Entry:");
}
array_overlap_test(nooverlap_target, 0);
__ add(to, count, end_to); // offset after last copied element
// for short arrays, just do single element copy
__ cmp(count, 23); // 16 + 7
__ brx(Assembler::less, false, Assembler::pn, L_copy_byte);
__ delayed()->add(from, count, end_from);
{
// Align end of arrays since they could be not aligned even
// when arrays itself are aligned.
// copy bytes to align 'end_to' on 8 byte boundary
__ andcc(end_to, 7, G1); // misaligned bytes
__ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
__ delayed()->nop();
__ sub(count, G1, count);
__ BIND(L_align);
__ dec(end_from);
__ dec(end_to);
__ ldub(end_from, 0, O3);
__ deccc(G1);
__ brx(Assembler::notZero, false, Assembler::pt, L_align);
__ delayed()->stb(O3, end_to, 0);
__ BIND(L_skip_alignment);
}
#ifdef _LP64
if (aligned) {
// Both arrays are aligned to 8-bytes in 64-bits VM.
// The 'count' is decremented in copy_16_bytes_backward_with_shift()
// in unaligned case.
__ dec(count, 16);
} else
#endif
{
// Copy with shift 16 bytes per iteration if arrays do not have
// the same alignment mod 8, otherwise jump to the next
// code for aligned copy (and substracting 16 from 'count' before jump).
// The compare above (count >= 11) guarantes 'count' >= 16 bytes.
// Also jump over aligned copy after the copy with shift completed.
copy_16_bytes_backward_with_shift(end_from, end_to, count, 16,
L_aligned_copy, L_copy_byte);
}
// copy 4 elements (16 bytes) at a time
__ align(OptoLoopAlignment);
__ BIND(L_aligned_copy);
__ dec(end_from, 16);
__ ldx(end_from, 8, O3);
__ ldx(end_from, 0, O4);
__ dec(end_to, 16);
__ deccc(count, 16);
__ stx(O3, end_to, 8);
__ brx(Assembler::greaterEqual, false, Assembler::pt, L_aligned_copy);
__ delayed()->stx(O4, end_to, 0);
__ inc(count, 16);
// copy 1 element (2 bytes) at a time
__ BIND(L_copy_byte);
__ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
__ align(OptoLoopAlignment);
__ BIND(L_copy_byte_loop);
__ dec(end_from);
__ dec(end_to);
__ ldub(end_from, 0, O4);
__ deccc(count);
__ brx(Assembler::greater, false, Assembler::pt, L_copy_byte_loop);
__ delayed()->stb(O4, end_to, 0);
__ BIND(L_exit);
// O3, O4 are used as temp registers
inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr, O3, O4);
__ retl();
__ delayed()->mov(G0, O0); // return 0
return start;
}
//
// Generate stub for disjoint short copy. If "aligned" is true, the
// "from" and "to" addresses are assumed to be heapword aligned.
//
// Arguments for generated stub:
// from: O0
// to: O1
// count: O2 treated as signed
//
address generate_disjoint_short_copy(bool aligned, address *entry, const char * name) {
__ align(CodeEntryAlignment);
StubCodeMark mark(this, "StubRoutines", name);
address start = __ pc();
Label L_skip_alignment, L_skip_alignment2;
Label L_copy_2_bytes, L_copy_2_bytes_loop, L_exit;
const Register from = O0; // source array address
const Register to = O1; // destination array address
const Register count = O2; // elements count
const Register offset = O5; // offset from start of arrays
// O3, O4, G3, G4 are used as temp registers
assert_clean_int(count, O3); // Make sure 'count' is clean int.
if (entry != NULL) {
*entry = __ pc();
// caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
BLOCK_COMMENT("Entry:");
}
// for short arrays, just do single element copy
__ cmp(count, 11); // 8 + 3 (22 bytes)
__ brx(Assembler::less, false, Assembler::pn, L_copy_2_bytes);
__ delayed()->mov(G0, offset);
if (aligned) {
// 'aligned' == true when it is known statically during compilation
// of this arraycopy call site that both 'from' and 'to' addresses
// are HeapWordSize aligned (see LibraryCallKit::basictype2arraycopy()).
//
// Aligned arrays have 4 bytes alignment in 32-bits VM
// and 8 bytes - in 64-bits VM.
//
#ifndef _LP64
// copy a 2-elements word if necessary to align 'to' to 8 bytes
__ andcc(to, 7, G0);
__ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
__ delayed()->ld(from, 0, O3);
__ inc(from, 4);
__ inc(to, 4);
__ dec(count, 2);
__ st(O3, to, -4);
__ BIND(L_skip_alignment);
#endif
} else {
// copy 1 element if necessary to align 'to' on an 4 bytes
__ andcc(to, 3, G0);
__ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
__ delayed()->lduh(from, 0, O3);
__ inc(from, 2);
__ inc(to, 2);
__ dec(count);
__ sth(O3, to, -2);
__ BIND(L_skip_alignment);
// copy 2 elements to align 'to' on an 8 byte boundary
__ andcc(to, 7, G0);
__ br(Assembler::zero, false, Assembler::pn, L_skip_alignment2);
__ delayed()->lduh(from, 0, O3);
__ dec(count, 2);
__ lduh(from, 2, O4);
__ inc(from, 4);
__ inc(to, 4);
__ sth(O3, to, -4);
__ sth(O4, to, -2);
__ BIND(L_skip_alignment2);
}
#ifdef _LP64
if (!aligned)
#endif
{
// Copy with shift 16 bytes per iteration if arrays do not have
// the same alignment mod 8, otherwise fall through to the next
// code for aligned copy.
// The compare above (count >= 11) guarantes 'count' >= 16 bytes.
// Also jump over aligned copy after the copy with shift completed.
copy_16_bytes_forward_with_shift(from, to, count, 1, L_copy_2_bytes);
}
// Both array are 8 bytes aligned, copy 16 bytes at a time
__ and3(count, 3, G4); // Save
__ srl(count, 2, count);
generate_disjoint_long_copy_core(aligned);
__ mov(G4, count); // restore
// copy 1 element at a time
__ BIND(L_copy_2_bytes);
__ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
__ align(OptoLoopAlignment);
__ BIND(L_copy_2_bytes_loop);
__ lduh(from, offset, O3);
__ deccc(count);
__ sth(O3, to, offset);
__ brx(Assembler::notZero, false, Assembler::pt, L_copy_2_bytes_loop);
__ delayed()->inc(offset, 2);
__ BIND(L_exit);
// O3, O4 are used as temp registers
inc_counter_np(SharedRuntime::_jshort_array_copy_ctr, O3, O4);
__ retl();
__ delayed()->mov(G0, O0); // return 0
return start;
}
//
// Generate stub for disjoint short fill. If "aligned" is true, the
// "to" address is assumed to be heapword aligned.
//
// Arguments for generated stub:
// to: O0
// value: O1
// count: O2 treated as signed
//
address generate_fill(BasicType t, bool aligned, const char* name) {
__ align(CodeEntryAlignment);
StubCodeMark mark(this, "StubRoutines", name);
address start = __ pc();
const Register to = O0; // source array address
const Register value = O1; // fill value
const Register count = O2; // elements count
// O3 is used as a temp register
assert_clean_int(count, O3); // Make sure 'count' is clean int.
Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
Label L_fill_2_bytes, L_fill_elements, L_fill_32_bytes;
int shift = -1;
switch (t) {
case T_BYTE:
shift = 2;
break;
case T_SHORT:
shift = 1;
break;
case T_INT:
shift = 0;
break;
default: ShouldNotReachHere();
}
BLOCK_COMMENT("Entry:");
if (t == T_BYTE) {
// Zero extend value
__ and3(value, 0xff, value);
__ sllx(value, 8, O3);
__ or3(value, O3, value);
}
if (t == T_SHORT) {
// Zero extend value
__ sllx(value, 48, value);
__ srlx(value, 48, value);
}
if (t == T_BYTE || t == T_SHORT) {
__ sllx(value, 16, O3);
__ or3(value, O3, value);
}
__ cmp(count, 2<
Other Java examples (source code examples)Here is a short list of links related to this Java stubGenerator_sparc.cpp source code file: |
The search page
Other Java source code examples at this package level
Click here to learn more about this project
