/**
* Perform optimizations for a Goto.
*
* <p> Patterns:
* <pre>
* 1) GOTO A replaced by GOTO B
* A: GOTO B
*
* 2) GOTO A replaced by IF .. GOTO B
* A: IF .. GOTO B GOTO C
* C: ...
* 3) GOTO next instruction eliminated
* 4) GOTO A replaced by GOTO B
* A: LABEL
* BBEND
* B:
* 5) GOTO BBn where BBn has exactly one in edge
* - move BBn immediately after the GOTO in the code order,
* so that pattern 3) will create a fallthrough
* <pre>
*
* <p> Precondition: Goto.conforms(g)
*
* @param ir governing IR
* @param g the instruction to optimize
* @param bb the basic block holding g
* @return {@code true} if made a transformation
*/
private boolean processGoto(IR ir, Instruction g, BasicBlock bb) {
BasicBlock targetBlock = g.getBranchTarget();
// don't optimize jumps to a code motion landing pad
if (targetBlock.getLandingPad()) return false;
Instruction targetLabel = targetBlock.firstInstruction();
// get the first real instruction at the g target
// NOTE: this instruction is not necessarily in targetBlock,
// iff targetBlock has no real instructions
Instruction targetInst = firstRealInstructionFollowing(targetLabel);
if (targetInst == null || targetInst == g) {
return false;
}
Instruction nextLabel = firstLabelFollowing(g);
if (targetLabel == nextLabel) {
// found a GOTO to the next instruction. just remove it.
g.remove();
return true;
}
if (Goto.conforms(targetInst)) {
// unconditional branch to unconditional branch.
// replace g with goto to targetInst's target
Instruction target2 =
firstRealInstructionFollowing(targetInst.getBranchTarget().firstInstruction());
if (target2 == targetInst) {
// Avoid an infinite recursion in the following bizarre scenario:
// g: goto L
// ...
// L: goto L
// This happens in jByteMark.EmFloatPnt.denormalize() due to a while(true) {}
return false;
}
Goto.setTarget(g, (BranchOperand) Goto.getTarget(targetInst).copy());
bb.recomputeNormalOut(ir); // fix the CFG
return true;
}
if (targetBlock.isEmpty()) {
// GOTO an empty basic block. Change target to the
// next block.
BasicBlock nextBlock = targetBlock.getFallThroughBlock();
Goto.setTarget(g, nextBlock.makeJumpTarget());
bb.recomputeNormalOut(ir); // fix the CFG
return true;
}
if (mayDuplicateCondBranches && IfCmp.conforms(targetInst)) {
// unconditional branch to a conditional branch.
// If the Goto is the only branch instruction in its basic block
// and the IfCmp is the only non-GOTO branch instruction
// in its basic block then replace the goto with a copy of
// targetInst and append another GOTO to the not-taken
// target of targetInst's block.
// We impose these additional restrictions to avoid getting
// multiple conditional branches in a single basic block.
if (!g.prevInstructionInCodeOrder().isBranch()
&& (targetInst.nextInstructionInCodeOrder().operator == BBEND
|| targetInst.nextInstructionInCodeOrder().operator == GOTO)) {
Instruction copy = targetInst.copyWithoutLinks();
g.replace(copy);
Instruction newGoto = targetInst.getBasicBlock().getNotTakenNextBlock().makeGOTO();
copy.insertAfter(newGoto);
bb.recomputeNormalOut(ir); // fix the CFG
return true;
}
}
// try to create a fallthrough
if (mayReorderCode && targetBlock.getNumberOfIn() == 1) {
BasicBlock ftBlock = targetBlock.getFallThroughBlock();
if (ftBlock != null) {
BranchOperand ftTarget = ftBlock.makeJumpTarget();
targetBlock.appendInstruction(CPOS(g, Goto.create(GOTO, ftTarget)));
}
ir.cfg.removeFromCodeOrder(targetBlock);
ir.cfg.insertAfterInCodeOrder(bb, targetBlock);
targetBlock.recomputeNormalOut(ir); // fix the CFG
return true;
}
return false;
}
/**
* @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;
}