/**
* Insert code before a return instruction to restore the volatile and volatile registers.
*
* @param inst the return instruction
*/
private void restoreVolatileRegisters(Instruction inst) {
GenericPhysicalRegisterSet phys = ir.regpool.getPhysicalRegisterSet();
// Restore every GPR
int i = 0;
for (Enumeration<Register> e = phys.enumerateVolatileGPRs(); e.hasMoreElements(); i++) {
Register r = e.nextElement();
int location = saveVolatileGPRLocation[i];
Operand M = new StackLocationOperand(true, -location, WORDSIZE);
inst.insertBefore(
MIR_Move.create(IA32_MOV, new RegisterOperand(r, PRIMITIVE_TYPE_FOR_WORD), M));
}
}
/**
* Insert code before a return instruction to restore the nonvolatile registers.
*
* @param inst the return instruction
*/
private void restoreNonVolatiles(Instruction inst) {
GenericPhysicalRegisterSet phys = ir.regpool.getPhysicalRegisterSet();
int nNonvolatileGPRS = ir.compiledMethod.getNumberOfNonvolatileGPRs();
int n = nNonvolatileGPRS - 1;
for (Enumeration<Register> e = phys.enumerateNonvolatileGPRsBackwards();
e.hasMoreElements() && n >= 0;
n--) {
Register nv = e.nextElement();
int offset = getNonvolatileGPROffset(n);
Operand M = new StackLocationOperand(true, -offset, WORDSIZE);
inst.insertBefore(
MIR_Move.create(IA32_MOV, new RegisterOperand(nv, PRIMITIVE_TYPE_FOR_WORD), M));
}
}
/**
* Insert code into the epilogue to restore the floating point state.
*
* @param inst the return instruction after the epilogue.
*/
private void restoreFloatingPointState(Instruction inst) {
if (SSE2_FULL) {
GenericPhysicalRegisterSet phys = ir.regpool.getPhysicalRegisterSet();
for (int i = 0; i < 8; i++) {
inst.insertBefore(
MIR_Move.create(
IA32_MOVQ,
new RegisterOperand(phys.getFPR(i), TypeReference.Double),
new StackLocationOperand(
true, -fsaveLocation + (i * BYTES_IN_DOUBLE), BYTES_IN_DOUBLE)));
}
} else {
Operand M = new StackLocationOperand(true, -fsaveLocation, 4);
inst.insertBefore(MIR_FSave.create(IA32_FRSTOR, M));
}
}
/**
* Insert an explicit stack overflow check in the prologue <em>before</em> buying the stack frame.
* SIDE EFFECT: mutates the plg into a trap instruction. We need to mutate so that the trap
* instruction is in the GC map data structures.
*
* @param plg the prologue instruction
*/
private void insertBigFrameStackOverflowCheck(Instruction plg) {
if (!ir.method.isInterruptible()) {
plg.remove();
return;
}
if (ir.compiledMethod.isSaveVolatile()) {
return;
}
PhysicalRegisterSet phys = (PhysicalRegisterSet) ir.regpool.getPhysicalRegisterSet();
Register ESP = phys.getESP();
Register ECX = phys.getECX();
// ECX := active Thread Stack Limit
MemoryOperand M =
MemoryOperand.BD(
ir.regpool.makeTROp(),
Entrypoints.stackLimitField.getOffset(),
(byte) WORDSIZE,
null,
null);
plg.insertBefore(
MIR_Move.create(IA32_MOV, new RegisterOperand((ECX), PRIMITIVE_TYPE_FOR_WORD), M));
// ECX += frame Size
int frameSize = getFrameFixedSize();
plg.insertBefore(
MIR_BinaryAcc.create(
IA32_ADD,
new RegisterOperand(ECX, PRIMITIVE_TYPE_FOR_WORD),
VM.BuildFor32Addr ? IC(frameSize) : LC(frameSize)));
// Trap if ESP <= ECX
MIR_TrapIf.mutate(
plg,
IA32_TRAPIF,
null,
new RegisterOperand(ESP, PRIMITIVE_TYPE_FOR_WORD),
new RegisterOperand(ECX, PRIMITIVE_TYPE_FOR_WORD),
IA32ConditionOperand.LE(),
TrapCodeOperand.StackOverflow());
}
@Override
public void insertUnspillBefore(Instruction s, Register r, Register type, int location) {
Operator move = getMoveOperator(type);
byte size = getSizeOfType(type);
RegisterOperand rOp;
if (type.isFloat()) {
rOp = F(r);
} else if (type.isDouble()) {
rOp = D(r);
} else {
if (VM.BuildFor64Addr && type.isInteger()) {
rOp = new RegisterOperand(r, TypeReference.Int);
} else {
rOp = new RegisterOperand(r, PRIMITIVE_TYPE_FOR_WORD);
}
}
StackLocationOperand spillLoc = new StackLocationOperand(true, -location, size);
Instruction unspillOp = MIR_Move.create(move, rOp, spillLoc);
if (VERBOSE_DEBUG) {
System.out.println("INSERT_UNSPILL_BEFORE: " + "Inserting " + unspillOp + " before " + s);
}
s.insertBefore(unspillOp);
}
/**
* Insert the prologue for a normal method.
*
* <p>Assume we are inserting the prologue for method B called from method A.
*
* <ul>
* <li>Perform a stack overflow check.
* <li>Store a back pointer to A's frame
* <li>Store B's compiled method id
* <li>Adjust frame pointer to point to B's frame
* <li>Save any used non-volatile registers
* </ul>
*/
@Override
public void insertNormalPrologue() {
PhysicalRegisterSet phys = (PhysicalRegisterSet) ir.regpool.getPhysicalRegisterSet();
Register ESP = phys.getESP();
MemoryOperand fpHome =
MemoryOperand.BD(
ir.regpool.makeTROp(),
ArchEntrypoints.framePointerField.getOffset(),
(byte) WORDSIZE,
null,
null);
// the prologue instruction
Instruction plg = ir.firstInstructionInCodeOrder().nextInstructionInCodeOrder();
// inst is the instruction immediately after the IR_PROLOGUE
// instruction
Instruction inst = plg.nextInstructionInCodeOrder();
int frameFixedSize = getFrameFixedSize();
ir.compiledMethod.setFrameFixedSize(frameFixedSize);
// I. Buy a stackframe (including overflow check)
// NOTE: We play a little game here. If the frame we are buying is
// very small (less than 256) then we can be sloppy with the
// stackoverflow check and actually allocate the frame in the guard
// region. We'll notice when this frame calls someone and take the
// stackoverflow in the callee. We can't do this if the frame is too big,
// because growing the stack in the callee and/or handling a hardware trap
// in this frame will require most of the guard region to complete.
// See libvm.C.
if (frameFixedSize >= 256) {
// 1. Insert Stack overflow check.
insertBigFrameStackOverflowCheck(plg);
// 2. Save caller's frame pointer
inst.insertBefore(MIR_UnaryNoRes.create(IA32_PUSH, fpHome));
// 3. Set my frame pointer to current value of stackpointer
inst.insertBefore(
MIR_Move.create(
IA32_MOV, fpHome.copy(), new RegisterOperand(ESP, PRIMITIVE_TYPE_FOR_WORD)));
// 4. Store my compiled method id
int cmid = ir.compiledMethod.getId();
inst.insertBefore(MIR_UnaryNoRes.create(IA32_PUSH, VM.BuildFor32Addr ? IC(cmid) : LC(cmid)));
} else {
// 1. Save caller's frame pointer
inst.insertBefore(MIR_UnaryNoRes.create(IA32_PUSH, fpHome));
// 2. Set my frame pointer to current value of stackpointer
inst.insertBefore(
MIR_Move.create(
IA32_MOV, fpHome.copy(), new RegisterOperand(ESP, PRIMITIVE_TYPE_FOR_WORD)));
// 3. Store my compiled method id
int cmid = ir.compiledMethod.getId();
inst.insertBefore(MIR_UnaryNoRes.create(IA32_PUSH, VM.BuildFor32Addr ? IC(cmid) : LC(cmid)));
// 4. Insert Stack overflow check.
insertNormalStackOverflowCheck(plg);
}
// II. Save any used volatile and non-volatile registers
if (ir.compiledMethod.isSaveVolatile()) {
saveVolatiles(inst);
saveFloatingPointState(inst);
}
saveNonVolatiles(inst);
}
/**
* expand an FMOV pseudo-insruction.
*
* @param s the instruction to expand
* @param phys controlling physical register set
*/
private static void expandFmov(Instruction s, PhysicalRegisterSet phys) {
Operand result = MIR_Move.getResult(s);
Operand value = MIR_Move.getValue(s);
if (result.isRegister() && value.isRegister()) {
if (result.similar(value)) {
// eliminate useless move
s.remove();
} else {
int i = PhysicalRegisterSet.getFPRIndex(result.asRegister().getRegister());
int j = PhysicalRegisterSet.getFPRIndex(value.asRegister().getRegister());
if (j == 0) {
// We have FMOV Fi, F0
// Expand as:
// FST F(i) (copy F0 to F(i))
MIR_Move.mutate(s, IA32_FST, D(phys.getFPR(i)), D(phys.getFPR(0)));
} else {
// We have FMOV Fi, Fj
// Expand as:
// FLD Fj (push Fj on FP stack).
// FSTP F(i+1) (copy F0 to F(i+1) and then pop register stack)
s.insertBefore(MIR_Move.create(IA32_FLD, D(phys.getFPR(0)), value));
MIR_Move.mutate(s, IA32_FSTP, D(phys.getFPR(i + 1)), D(phys.getFPR(0)));
}
}
} else if (value instanceof MemoryOperand) {
if (result instanceof MemoryOperand) {
// We have FMOV M1, M2
// Expand as:
// FLD M1 (push M1 on FP stack).
// FSTP M2 (copy F0 to M2 and pop register stack)
s.insertBefore(MIR_Move.create(IA32_FLD, D(phys.getFPR(0)), value));
MIR_Move.mutate(s, IA32_FSTP, result, D(phys.getFPR(0)));
} else {
// We have FMOV Fi, M
// Expand as:
// FLD M (push M on FP stack).
// FSTP F(i+1) (copy F0 to F(i+1) and pop register stack)
if (VM.VerifyAssertions) VM._assert(result.isRegister());
int i = PhysicalRegisterSet.getFPRIndex(result.asRegister().getRegister());
s.insertBefore(MIR_Move.create(IA32_FLD, D(phys.getFPR(0)), value));
MIR_Move.mutate(s, IA32_FSTP, D(phys.getFPR(i + 1)), D(phys.getFPR(0)));
}
} else {
// We have FMOV M, Fi
if (VM.VerifyAssertions) VM._assert(value.isRegister());
if (VM.VerifyAssertions) {
VM._assert(result instanceof MemoryOperand);
}
int i = PhysicalRegisterSet.getFPRIndex(value.asRegister().getRegister());
if (i != 0) {
// Expand as:
// FLD Fi (push Fi on FP stack).
// FSTP M (store F0 in M and pop register stack);
s.insertBefore(MIR_Move.create(IA32_FLD, D(phys.getFPR(0)), value));
MIR_Move.mutate(s, IA32_FSTP, result, D(phys.getFPR(0)));
} else {
// Expand as:
// FST M (store F0 in M);
MIR_Move.mutate(s, IA32_FST, result, value);
}
}
}
/**
* @param ir the IR to expand
* @return return value is garbage for IA32
*/
public static int expand(IR ir) {
PhysicalRegisterSet phys = ir.regpool.getPhysicalRegisterSet().asIA32();
MachineCodeOffsets mcOffsets = ir.MIRInfo.mcOffsets;
for (Instruction next, p = ir.firstInstructionInCodeOrder(); p != null; p = next) {
next = p.nextInstructionInCodeOrder();
mcOffsets.setMachineCodeOffset(p, -1);
switch (p.getOpcode()) {
case IA32_MOVAPS_opcode:
// a reg-reg move turned into a memory move where we can't guarantee alignment
if (MIR_Move.getResult(p).isMemory() || MIR_Move.getValue(p).isMemory()) {
MIR_Move.mutate(p, IA32_MOVSS, MIR_Move.getClearResult(p), MIR_Move.getClearValue(p));
}
break;
case IA32_MOVAPD_opcode:
// a reg-reg move turned into a memory move where we can't guarantee alignment
if (MIR_Move.getResult(p).isMemory() || MIR_Move.getValue(p).isMemory()) {
MIR_Move.mutate(p, IA32_MOVSD, MIR_Move.getClearResult(p), MIR_Move.getClearValue(p));
}
break;
case IA32_TEST_opcode:
// don't bother telling rest of compiler that memory operand
// must be first; we can just commute it here.
if (MIR_Test.getVal2(p).isMemory()) {
Operand tmp = MIR_Test.getClearVal1(p);
MIR_Test.setVal1(p, MIR_Test.getClearVal2(p));
MIR_Test.setVal2(p, tmp);
}
break;
case NULL_CHECK_opcode:
{
// mutate this into a TRAPIF, and then fall through to the the
// TRAP_IF case.
Operand ref = NullCheck.getRef(p);
MIR_TrapIf.mutate(
p,
IA32_TRAPIF,
null,
ref.copy(),
IC(0),
IA32ConditionOperand.EQ(),
TrapCodeOperand.NullPtr());
}
// There is no break statement here on purpose!
case IA32_TRAPIF_opcode:
{
// split the basic block right before the IA32_TRAPIF
BasicBlock thisBlock = p.getBasicBlock();
BasicBlock trap = thisBlock.createSubBlock(p.getBytecodeIndex(), ir, 0f);
thisBlock.insertOut(trap);
BasicBlock nextBlock = thisBlock.splitNodeWithLinksAt(p, ir);
thisBlock.insertOut(trap);
TrapCodeOperand tc = MIR_TrapIf.getClearTrapCode(p);
p.remove();
mcOffsets.setMachineCodeOffset(nextBlock.firstInstruction(), -1);
// add code to thisBlock to conditionally jump to trap
Instruction cmp =
MIR_Compare.create(IA32_CMP, MIR_TrapIf.getVal1(p), MIR_TrapIf.getVal2(p));
if (p.isMarkedAsPEI()) {
// The trap if was explictly marked, which means that it has
// a memory operand into which we've folded a null check.
// Actually need a GC map for both the compare and the INT.
cmp.markAsPEI();
cmp.copyPosition(p);
ir.MIRInfo.gcIRMap.insertTwin(p, cmp);
}
thisBlock.appendInstruction(cmp);
thisBlock.appendInstruction(
MIR_CondBranch.create(
IA32_JCC, MIR_TrapIf.getCond(p), trap.makeJumpTarget(), null));
// add block at end to hold trap instruction, and
// insert trap sequence
ir.cfg.addLastInCodeOrder(trap);
if (tc.isArrayBounds()) {
// attempt to store index expression in processor object for
// C trap handler
Operand index = MIR_TrapIf.getVal2(p);
if (!(index instanceof RegisterOperand || index instanceof IntConstantOperand)) {
index = IC(0xdeadbeef); // index was spilled, and
// we can't get it back here.
}
MemoryOperand mo =
MemoryOperand.BD(
ir.regpool.makeTROp(),
ArchEntrypoints.arrayIndexTrapParamField.getOffset(),
(byte) 4,
null,
null);
trap.appendInstruction(MIR_Move.create(IA32_MOV, mo, index.copy()));
}
// NOTE: must make p the trap instruction: it is the GC point!
// IMPORTANT: must also inform the GCMap that the instruction has
// been moved!!!
trap.appendInstruction(MIR_Trap.mutate(p, IA32_INT, null, tc));
ir.MIRInfo.gcIRMap.moveToEnd(p);
if (tc.isStackOverflow()) {
// only stackoverflow traps resume at next instruction.
trap.appendInstruction(MIR_Branch.create(IA32_JMP, nextBlock.makeJumpTarget()));
}
}
break;
case IA32_FMOV_ENDING_LIVE_RANGE_opcode:
{
Operand result = MIR_Move.getResult(p);
Operand value = MIR_Move.getValue(p);
if (result.isRegister() && value.isRegister()) {
if (result.similar(value)) {
// eliminate useless move
p.remove();
} else {
int i = PhysicalRegisterSet.getFPRIndex(result.asRegister().getRegister());
int j = PhysicalRegisterSet.getFPRIndex(value.asRegister().getRegister());
if (i == 0) {
MIR_XChng.mutate(p, IA32_FXCH, result, value);
} else if (j == 0) {
MIR_XChng.mutate(p, IA32_FXCH, value, result);
} else {
expandFmov(p, phys);
}
}
} else {
expandFmov(p, phys);
}
break;
}
case DUMMY_DEF_opcode:
case DUMMY_USE_opcode:
case REQUIRE_ESP_opcode:
case ADVISE_ESP_opcode:
p.remove();
break;
case IA32_FMOV_opcode:
expandFmov(p, phys);
break;
case IA32_MOV_opcode:
// Replace result = IA32_MOV 0 with result = IA32_XOR result, result
if (MIR_Move.getResult(p).isRegister()
&& MIR_Move.getValue(p).isIntConstant()
&& MIR_Move.getValue(p).asIntConstant().value == 0) {
// Calculate what flags are defined in coming instructions before a use of a flag or
// BBend
Instruction x = next;
int futureDefs = 0;
while (!BBend.conforms(x) && !PhysicalDefUse.usesEFLAGS(x.operator())) {
futureDefs |= x.operator().implicitDefs;
x = x.nextInstructionInCodeOrder();
}
// If the flags will be destroyed prior to use or we reached the end of the basic block
if (BBend.conforms(x)
|| (futureDefs & PhysicalDefUse.maskAF_CF_OF_PF_SF_ZF)
== PhysicalDefUse.maskAF_CF_OF_PF_SF_ZF) {
Operand result = MIR_Move.getClearResult(p);
MIR_BinaryAcc.mutate(p, IA32_XOR, result, result.copy());
}
}
break;
case IA32_SET__B_opcode:
// Replace <cmp>, set__b, movzx__b with xor, <cmp>, set__b
if (MIR_Set.getResult(p).isRegister()
&& MIR_Unary.conforms(next)
&& (next.operator() == IA32_MOVZX__B)
&& MIR_Unary.getResult(next).isRegister()
&& MIR_Unary.getVal(next).similar(MIR_Unary.getResult(next))
&& MIR_Unary.getVal(next).similar(MIR_Set.getResult(p))) {
// Find instruction in this basic block that defines flags
Instruction x = p.prevInstructionInCodeOrder();
Operand result = MIR_Unary.getResult(next);
boolean foundCmp = false;
outer:
while (!Label.conforms(x)) {
Enumeration<Operand> e = x.getUses();
while (e.hasMoreElements()) {
// We can't use an xor to clear the register if that register is
// used by the <cmp> or intervening instruction
if (e.nextElement().similar(result)) {
break outer;
}
}
if (PhysicalDefUse.definesEFLAGS(x.operator())
&& !PhysicalDefUse.usesEFLAGS(x.operator())) {
// we found a <cmp> that doesn't use the result or the flags
// that would be clobbered by the xor
foundCmp = true;
break outer;
}
x = x.prevInstructionInCodeOrder();
}
if (foundCmp) {
// We found the <cmp>, mutate the movzx__b into an xor and insert it before the <cmp>
next.remove();
MIR_BinaryAcc.mutate(next, IA32_XOR, result, MIR_Unary.getVal(next));
x.insertBefore(next);
// get ready for the next instruction
next = p.nextInstructionInCodeOrder();
}
}
break;
case IA32_LEA_opcode:
{
// Sometimes we're over eager in BURS in using LEAs and after register
// allocation we can simplify to the accumulate form
// replace reg1 = LEA [reg1 + reg2] with reg1 = reg1 + reg2
// replace reg1 = LEA [reg1 + c1] with reg1 = reg1 + c1
// replace reg1 = LEA [reg1 << c1] with reg1 = reg1 << c1
MemoryOperand value = MIR_Lea.getValue(p);
RegisterOperand result = MIR_Lea.getResult(p);
if ((value.base != null && value.base.getRegister() == result.getRegister())
|| (value.index != null && value.index.getRegister() == result.getRegister())) {
// Calculate what flags are defined in coming instructions before a use of a flag or
// BBend
Instruction x = next;
int futureDefs = 0;
while (!BBend.conforms(x) && !PhysicalDefUse.usesEFLAGS(x.operator())) {
futureDefs |= x.operator().implicitDefs;
x = x.nextInstructionInCodeOrder();
}
// If the flags will be destroyed prior to use or we reached the end of the basic
// block
if (BBend.conforms(x)
|| (futureDefs & PhysicalDefUse.maskAF_CF_OF_PF_SF_ZF)
== PhysicalDefUse.maskAF_CF_OF_PF_SF_ZF) {
if (value.base != null
&& value.index != null
&& value.index.getRegister() == result.getRegister()
&& value.disp.isZero()
&& value.scale == 0) {
// reg1 = lea [base + reg1] -> add reg1, base
MIR_BinaryAcc.mutate(p, IA32_ADD, result, value.base);
} else if (value.base != null
&& value.base.getRegister() == result.getRegister()
&& value.index != null
&& value.disp.isZero()
&& value.scale == 0) {
// reg1 = lea [reg1 + index] -> add reg1, index
MIR_BinaryAcc.mutate(p, IA32_ADD, result, value.index);
} else if (value.base != null
&& value.base.getRegister() == result.getRegister()
&& value.index == null) {
if (VM.VerifyAssertions) VM._assert(fits(value.disp, 32));
// reg1 = lea [reg1 + disp] -> add reg1, disp
MIR_BinaryAcc.mutate(p, IA32_ADD, result, IC(value.disp.toInt()));
} else if (value.base == null
&& value.index != null
&& value.index.getRegister() == result.getRegister()
&& value.scale == 0) {
if (VM.VerifyAssertions) VM._assert(fits(value.disp, 32));
// reg1 = lea [reg1 + disp] -> add reg1, disp
MIR_BinaryAcc.mutate(p, IA32_ADD, result, IC(value.disp.toInt()));
} else if (value.base == null
&& value.index != null
&& value.index.getRegister() == result.getRegister()
&& value.disp.isZero()) {
// reg1 = lea [reg1 << scale] -> shl reg1, scale
if (value.scale == 0) {
p.remove();
} else if (value.scale == 1) {
MIR_BinaryAcc.mutate(p, IA32_ADD, result, value.index);
} else {
MIR_BinaryAcc.mutate(p, IA32_SHL, result, IC(value.scale));
}
}
}
}
}
break;
case IA32_FCLEAR_opcode:
expandFClear(p, ir);
break;
case IA32_JCC2_opcode:
p.insertBefore(
MIR_CondBranch.create(
IA32_JCC,
MIR_CondBranch2.getCond1(p),
MIR_CondBranch2.getTarget1(p),
MIR_CondBranch2.getBranchProfile1(p)));
MIR_CondBranch.mutate(
p,
IA32_JCC,
MIR_CondBranch2.getCond2(p),
MIR_CondBranch2.getTarget2(p),
MIR_CondBranch2.getBranchProfile2(p));
break;
case CALL_SAVE_VOLATILE_opcode:
p.changeOperatorTo(IA32_CALL);
break;
case IA32_LOCK_CMPXCHG_opcode:
p.insertBefore(MIR_Empty.create(IA32_LOCK));
p.changeOperatorTo(IA32_CMPXCHG);
break;
case IA32_LOCK_CMPXCHG8B_opcode:
p.insertBefore(MIR_Empty.create(IA32_LOCK));
p.changeOperatorTo(IA32_CMPXCHG8B);
break;
case YIELDPOINT_PROLOGUE_opcode:
expandYieldpoint(
p, ir, Entrypoints.optThreadSwitchFromPrologueMethod, IA32ConditionOperand.NE());
break;
case YIELDPOINT_EPILOGUE_opcode:
expandYieldpoint(
p, ir, Entrypoints.optThreadSwitchFromEpilogueMethod, IA32ConditionOperand.NE());
break;
case YIELDPOINT_BACKEDGE_opcode:
expandYieldpoint(
p, ir, Entrypoints.optThreadSwitchFromBackedgeMethod, IA32ConditionOperand.GT());
break;
case YIELDPOINT_OSR_opcode:
// must yield, does not check threadSwitch request
expandUnconditionalYieldpoint(p, ir, Entrypoints.optThreadSwitchFromOsrOptMethod);
break;
}
}
return 0;
}