// Check all machine operands that reference the antidependent register and must
// be replaced by NewReg. Return true if any of their parent instructions may
// clobber the new register.
//
// Note: AntiDepReg may be referenced by a two-address instruction such that
// it's use operand is tied to a def operand. We guard against the case in which
// the two-address instruction also defines NewReg, as may happen with
// pre/postincrement loads. In this case, both the use and def operands are in
// RegRefs because the def is inserted by PrescanInstruction and not erased
// during ScanInstruction. So checking for an instructions with definitions of
// both NewReg and AntiDepReg covers it.
bool
CriticalAntiDepBreaker::isNewRegClobberedByRefs(RegRefIter RegRefBegin,
RegRefIter RegRefEnd,
unsigned NewReg)
{
for (RegRefIter I = RegRefBegin; I != RegRefEnd; ++I ) {
MachineOperand *RefOper = I->second;
// Don't allow the instruction defining AntiDepReg to earlyclobber its
// operands, in case they may be assigned to NewReg. In this case antidep
// breaking must fail, but it's too rare to bother optimizing.
if (RefOper->isDef() && RefOper->isEarlyClobber())
return true;
// Handle cases in which this instructions defines NewReg.
MachineInstr *MI = RefOper->getParent();
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &CheckOper = MI->getOperand(i);
if (CheckOper.isRegMask() && CheckOper.clobbersPhysReg(NewReg))
return true;
if (!CheckOper.isReg() || !CheckOper.isDef() ||
CheckOper.getReg() != NewReg)
continue;
// Don't allow the instruction to define NewReg and AntiDepReg.
// When AntiDepReg is renamed it will be an illegal op.
if (RefOper->isDef())
return true;
// Don't allow an instruction using AntiDepReg to be earlyclobbered by
// NewReg
if (CheckOper.isEarlyClobber())
return true;
// Don't allow inline asm to define NewReg at all. Who know what it's
// doing with it.
if (MI->isInlineAsm())
return true;
}
}
return false;
}
static LaneBitmask getUsedRegMask(const MachineOperand &MO,
const MachineRegisterInfo &MRI,
const LiveIntervals &LIS) {
assert(MO.isUse() && MO.isReg() &&
TargetRegisterInfo::isVirtualRegister(MO.getReg()));
if (auto SubReg = MO.getSubReg())
return MRI.getTargetRegisterInfo()->getSubRegIndexLaneMask(SubReg);
auto MaxMask = MRI.getMaxLaneMaskForVReg(MO.getReg());
if (MaxMask == LaneBitmask::getLane(0)) // cannot have subregs
return MaxMask;
// For a tentative schedule LIS isn't updated yet but livemask should remain
// the same on any schedule. Subreg defs can be reordered but they all must
// dominate uses anyway.
auto SI = LIS.getInstructionIndex(*MO.getParent()).getBaseIndex();
return getLiveLaneMask(MO.getReg(), SI, LIS, MRI);
}
bool AMDGPUMCInstLower::lowerOperand(const MachineOperand &MO,
MCOperand &MCOp) const {
switch (MO.getType()) {
default:
llvm_unreachable("unknown operand type");
case MachineOperand::MO_Immediate:
MCOp = MCOperand::createImm(MO.getImm());
return true;
case MachineOperand::MO_Register:
MCOp = MCOperand::createReg(AMDGPU::getMCReg(MO.getReg(), ST));
return true;
case MachineOperand::MO_MachineBasicBlock: {
if (MO.getTargetFlags() != 0) {
MCOp = MCOperand::createExpr(
getLongBranchBlockExpr(*MO.getParent()->getParent(), MO));
} else {
MCOp = MCOperand::createExpr(
MCSymbolRefExpr::create(MO.getMBB()->getSymbol(), Ctx));
}
return true;
}
case MachineOperand::MO_GlobalAddress: {
const GlobalValue *GV = MO.getGlobal();
SmallString<128> SymbolName;
AP.getNameWithPrefix(SymbolName, GV);
MCSymbol *Sym = Ctx.getOrCreateSymbol(SymbolName);
const MCExpr *SymExpr =
MCSymbolRefExpr::create(Sym, getVariantKind(MO.getTargetFlags()),Ctx);
const MCExpr *Expr = MCBinaryExpr::createAdd(SymExpr,
MCConstantExpr::create(MO.getOffset(), Ctx), Ctx);
MCOp = MCOperand::createExpr(Expr);
return true;
}
case MachineOperand::MO_ExternalSymbol: {
MCSymbol *Sym = Ctx.getOrCreateSymbol(StringRef(MO.getSymbolName()));
Sym->setExternal(true);
const MCSymbolRefExpr *Expr = MCSymbolRefExpr::create(Sym, Ctx);
MCOp = MCOperand::createExpr(Expr);
return true;
}
}
}
static bool isNoReturnDef(const MachineOperand &MO) {
// Anything which is not a noreturn function is a real def.
const MachineInstr &MI = *MO.getParent();
if (!MI.isCall())
return false;
const MachineBasicBlock &MBB = *MI.getParent();
if (!MBB.succ_empty())
return false;
const MachineFunction &MF = *MBB.getParent();
// We need to keep correct unwind information even if the function will
// not return, since the runtime may need it.
if (MF.getFunction()->hasFnAttribute(Attribute::UWTable))
return false;
const Function *Called = getCalledFunction(MI);
if (Called == nullptr || !Called->hasFnAttribute(Attribute::NoReturn)
|| !Called->hasFnAttribute(Attribute::NoUnwind))
return false;
return true;
}
LaneBitmask DetectDeadLanes::transferDefinedLanes(const MachineOperand &Def,
unsigned OpNum, LaneBitmask DefinedLanes) const {
const MachineInstr &MI = *Def.getParent();
// Translate DefinedLanes if necessary.
switch (MI.getOpcode()) {
case TargetOpcode::REG_SEQUENCE: {
unsigned SubIdx = MI.getOperand(OpNum + 1).getImm();
DefinedLanes = TRI->composeSubRegIndexLaneMask(SubIdx, DefinedLanes);
DefinedLanes &= TRI->getSubRegIndexLaneMask(SubIdx);
break;
}
case TargetOpcode::INSERT_SUBREG: {
unsigned SubIdx = MI.getOperand(3).getImm();
if (OpNum == 2) {
DefinedLanes = TRI->composeSubRegIndexLaneMask(SubIdx, DefinedLanes);
DefinedLanes &= TRI->getSubRegIndexLaneMask(SubIdx);
} else {
assert(OpNum == 1 && "INSERT_SUBREG must have two operands");
// Ignore lanes defined by operand 2.
DefinedLanes &= ~TRI->getSubRegIndexLaneMask(SubIdx);
}
break;
}
case TargetOpcode::EXTRACT_SUBREG: {
unsigned SubIdx = MI.getOperand(2).getImm();
assert(OpNum == 1 && "EXTRACT_SUBREG must have one register operand only");
DefinedLanes = TRI->reverseComposeSubRegIndexLaneMask(SubIdx, DefinedLanes);
break;
}
case TargetOpcode::COPY:
case TargetOpcode::PHI:
break;
default:
llvm_unreachable("function must be called with COPY-like instruction");
}
assert(Def.getSubReg() == 0 &&
"Should not have subregister defs in machine SSA phase");
DefinedLanes &= MRI->getMaxLaneMaskForVReg(Def.getReg());
return DefinedLanes;
}
/// Returns true if the given machine operand \p MO only reads undefined lanes.
/// The function only works for use operands with a subregister set.
bool VirtRegRewriter::readsUndefSubreg(const MachineOperand &MO) const {
// Shortcut if the operand is already marked undef.
if (MO.isUndef())
return true;
unsigned Reg = MO.getReg();
const LiveInterval &LI = LIS->getInterval(Reg);
const MachineInstr &MI = *MO.getParent();
SlotIndex BaseIndex = LIS->getInstructionIndex(MI);
// This code is only meant to handle reading undefined subregisters which
// we couldn't properly detect before.
assert(LI.liveAt(BaseIndex) &&
"Reads of completely dead register should be marked undef already");
unsigned SubRegIdx = MO.getSubReg();
LaneBitmask UseMask = TRI->getSubRegIndexLaneMask(SubRegIdx);
// See if any of the relevant subregister liveranges is defined at this point.
for (const LiveInterval::SubRange &SR : LI.subranges()) {
if ((SR.LaneMask & UseMask) != 0 && SR.liveAt(BaseIndex))
return false;
}
return true;
}
void MachineRegisterInfo::verifyUseList(unsigned Reg) const {
#ifndef NDEBUG
bool Valid = true;
for (MachineOperand &M : reg_operands(Reg)) {
MachineOperand *MO = &M;
MachineInstr *MI = MO->getParent();
if (!MI) {
errs() << PrintReg(Reg, getTargetRegisterInfo())
<< " use list MachineOperand " << MO
<< " has no parent instruction.\n";
Valid = false;
continue;
}
MachineOperand *MO0 = &MI->getOperand(0);
unsigned NumOps = MI->getNumOperands();
if (!(MO >= MO0 && MO < MO0+NumOps)) {
errs() << PrintReg(Reg, getTargetRegisterInfo())
<< " use list MachineOperand " << MO
<< " doesn't belong to parent MI: " << *MI;
Valid = false;
}
if (!MO->isReg()) {
errs() << PrintReg(Reg, getTargetRegisterInfo())
<< " MachineOperand " << MO << ": " << *MO
<< " is not a register\n";
Valid = false;
}
if (MO->getReg() != Reg) {
errs() << PrintReg(Reg, getTargetRegisterInfo())
<< " use-list MachineOperand " << MO << ": "
<< *MO << " is the wrong register\n";
Valid = false;
}
}
assert(Valid && "Invalid use list");
#endif
}
/// optimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
/// a single register and writes a single register and it does not modify the
/// source, and if the source value is preserved as a sub-register of the
/// result, then replace all reachable uses of the source with the subreg of the
/// result.
///
/// Do not generate an EXTRACT that is used only in a debug use, as this changes
/// the code. Since this code does not currently share EXTRACTs, just ignore all
/// debug uses.
bool PeepholeOptimizer::
optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
unsigned SrcReg, DstReg, SubIdx;
if (!TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx))
return false;
if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
TargetRegisterInfo::isPhysicalRegister(SrcReg))
return false;
if (MRI->hasOneNonDBGUse(SrcReg))
// No other uses.
return false;
// Ensure DstReg can get a register class that actually supports
// sub-registers. Don't change the class until we commit.
const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
DstRC = TM->getRegisterInfo()->getSubClassWithSubReg(DstRC, SubIdx);
if (!DstRC)
return false;
// The ext instr may be operating on a sub-register of SrcReg as well.
// PPC::EXTSW is a 32 -> 64-bit sign extension, but it reads a 64-bit
// register.
// If UseSrcSubIdx is Set, SubIdx also applies to SrcReg, and only uses of
// SrcReg:SubIdx should be replaced.
bool UseSrcSubIdx = TM->getRegisterInfo()->
getSubClassWithSubReg(MRI->getRegClass(SrcReg), SubIdx) != 0;
// The source has other uses. See if we can replace the other uses with use of
// the result of the extension.
SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
for (MachineRegisterInfo::use_nodbg_iterator
UI = MRI->use_nodbg_begin(DstReg), UE = MRI->use_nodbg_end();
UI != UE; ++UI)
ReachedBBs.insert(UI->getParent());
// Uses that are in the same BB of uses of the result of the instruction.
SmallVector<MachineOperand*, 8> Uses;
// Uses that the result of the instruction can reach.
SmallVector<MachineOperand*, 8> ExtendedUses;
bool ExtendLife = true;
for (MachineRegisterInfo::use_nodbg_iterator
UI = MRI->use_nodbg_begin(SrcReg), UE = MRI->use_nodbg_end();
UI != UE; ++UI) {
MachineOperand &UseMO = UI.getOperand();
MachineInstr *UseMI = &*UI;
if (UseMI == MI)
continue;
if (UseMI->isPHI()) {
ExtendLife = false;
continue;
}
// Only accept uses of SrcReg:SubIdx.
if (UseSrcSubIdx && UseMO.getSubReg() != SubIdx)
continue;
// It's an error to translate this:
//
// %reg1025 = <sext> %reg1024
// ...
// %reg1026 = SUBREG_TO_REG 0, %reg1024, 4
//
// into this:
//
// %reg1025 = <sext> %reg1024
// ...
// %reg1027 = COPY %reg1025:4
// %reg1026 = SUBREG_TO_REG 0, %reg1027, 4
//
// The problem here is that SUBREG_TO_REG is there to assert that an
// implicit zext occurs. It doesn't insert a zext instruction. If we allow
// the COPY here, it will give us the value after the <sext>, not the
// original value of %reg1024 before <sext>.
if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG)
continue;
MachineBasicBlock *UseMBB = UseMI->getParent();
if (UseMBB == MBB) {
// Local uses that come after the extension.
if (!LocalMIs.count(UseMI))
Uses.push_back(&UseMO);
} else if (ReachedBBs.count(UseMBB)) {
// Non-local uses where the result of the extension is used. Always
// replace these unless it's a PHI.
Uses.push_back(&UseMO);
} else if (Aggressive && DT->dominates(MBB, UseMBB)) {
// We may want to extend the live range of the extension result in order
// to replace these uses.
ExtendedUses.push_back(&UseMO);
} else {
// Both will be live out of the def MBB anyway. Don't extend live range of
// the extension result.
ExtendLife = false;
break;
}
}
if (ExtendLife && !ExtendedUses.empty())
// Extend the liveness of the extension result.
std::copy(ExtendedUses.begin(), ExtendedUses.end(),
std::back_inserter(Uses));
// Now replace all uses.
bool Changed = false;
if (!Uses.empty()) {
SmallPtrSet<MachineBasicBlock*, 4> PHIBBs;
// Look for PHI uses of the extended result, we don't want to extend the
// liveness of a PHI input. It breaks all kinds of assumptions down
// stream. A PHI use is expected to be the kill of its source values.
for (MachineRegisterInfo::use_nodbg_iterator
UI = MRI->use_nodbg_begin(DstReg), UE = MRI->use_nodbg_end();
UI != UE; ++UI)
if (UI->isPHI())
PHIBBs.insert(UI->getParent());
const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
MachineOperand *UseMO = Uses[i];
MachineInstr *UseMI = UseMO->getParent();
MachineBasicBlock *UseMBB = UseMI->getParent();
if (PHIBBs.count(UseMBB))
continue;
// About to add uses of DstReg, clear DstReg's kill flags.
if (!Changed) {
MRI->clearKillFlags(DstReg);
MRI->constrainRegClass(DstReg, DstRC);
}
unsigned NewVR = MRI->createVirtualRegister(RC);
MachineInstr *Copy = BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
TII->get(TargetOpcode::COPY), NewVR)
.addReg(DstReg, 0, SubIdx);
// SubIdx applies to both SrcReg and DstReg when UseSrcSubIdx is set.
if (UseSrcSubIdx) {
Copy->getOperand(0).setSubReg(SubIdx);
Copy->getOperand(0).setIsUndef();
}
UseMO->setReg(NewVR);
++NumReuse;
Changed = true;
}
}
return Changed;
}
static void foldOperand(MachineOperand &OpToFold, MachineInstr *UseMI,
unsigned UseOpIdx,
std::vector<FoldCandidate> &FoldList,
SmallVectorImpl<MachineInstr *> &CopiesToReplace,
const SIInstrInfo *TII, const SIRegisterInfo &TRI,
MachineRegisterInfo &MRI) {
const MachineOperand &UseOp = UseMI->getOperand(UseOpIdx);
// FIXME: Fold operands with subregs.
if (UseOp.isReg() && ((UseOp.getSubReg() && OpToFold.isReg()) ||
UseOp.isImplicit())) {
return;
}
bool FoldingImm = OpToFold.isImm();
APInt Imm;
if (FoldingImm) {
unsigned UseReg = UseOp.getReg();
const TargetRegisterClass *UseRC
= TargetRegisterInfo::isVirtualRegister(UseReg) ?
MRI.getRegClass(UseReg) :
TRI.getPhysRegClass(UseReg);
Imm = APInt(64, OpToFold.getImm());
const MCInstrDesc &FoldDesc = TII->get(OpToFold.getParent()->getOpcode());
const TargetRegisterClass *FoldRC =
TRI.getRegClass(FoldDesc.OpInfo[0].RegClass);
// Split 64-bit constants into 32-bits for folding.
if (FoldRC->getSize() == 8 && UseOp.getSubReg()) {
if (UseRC->getSize() != 8)
return;
if (UseOp.getSubReg() == AMDGPU::sub0) {
Imm = Imm.getLoBits(32);
} else {
assert(UseOp.getSubReg() == AMDGPU::sub1);
Imm = Imm.getHiBits(32);
}
}
// In order to fold immediates into copies, we need to change the
// copy to a MOV.
if (UseMI->getOpcode() == AMDGPU::COPY) {
unsigned DestReg = UseMI->getOperand(0).getReg();
const TargetRegisterClass *DestRC
= TargetRegisterInfo::isVirtualRegister(DestReg) ?
MRI.getRegClass(DestReg) :
TRI.getPhysRegClass(DestReg);
unsigned MovOp = TII->getMovOpcode(DestRC);
if (MovOp == AMDGPU::COPY)
return;
UseMI->setDesc(TII->get(MovOp));
CopiesToReplace.push_back(UseMI);
}
}
// Special case for REG_SEQUENCE: We can't fold literals into
// REG_SEQUENCE instructions, so we have to fold them into the
// uses of REG_SEQUENCE.
if (UseMI->getOpcode() == AMDGPU::REG_SEQUENCE) {
unsigned RegSeqDstReg = UseMI->getOperand(0).getReg();
unsigned RegSeqDstSubReg = UseMI->getOperand(UseOpIdx + 1).getImm();
for (MachineRegisterInfo::use_iterator
RSUse = MRI.use_begin(RegSeqDstReg),
RSE = MRI.use_end(); RSUse != RSE; ++RSUse) {
MachineInstr *RSUseMI = RSUse->getParent();
if (RSUse->getSubReg() != RegSeqDstSubReg)
continue;
foldOperand(OpToFold, RSUseMI, RSUse.getOperandNo(), FoldList,
CopiesToReplace, TII, TRI, MRI);
}
return;
}
const MCInstrDesc &UseDesc = UseMI->getDesc();
// Don't fold into target independent nodes. Target independent opcodes
// don't have defined register classes.
if (UseDesc.isVariadic() ||
UseDesc.OpInfo[UseOpIdx].RegClass == -1)
return;
if (FoldingImm) {
MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue());
tryAddToFoldList(FoldList, UseMI, UseOpIdx, &ImmOp, TII);
return;
}
tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII);
// FIXME: We could try to change the instruction from 64-bit to 32-bit
// to enable more folding opportunites. The shrink operands pass
// already does this.
return;
}
void SIFoldOperands::foldInstOperand(MachineInstr &MI,
MachineOperand &OpToFold) const {
// We need mutate the operands of new mov instructions to add implicit
// uses of EXEC, but adding them invalidates the use_iterator, so defer
// this.
SmallVector<MachineInstr *, 4> CopiesToReplace;
SmallVector<FoldCandidate, 4> FoldList;
MachineOperand &Dst = MI.getOperand(0);
bool FoldingImm = OpToFold.isImm() || OpToFold.isFI();
if (FoldingImm) {
unsigned NumLiteralUses = 0;
MachineOperand *NonInlineUse = nullptr;
int NonInlineUseOpNo = -1;
MachineRegisterInfo::use_iterator NextUse, NextInstUse;
for (MachineRegisterInfo::use_iterator
Use = MRI->use_begin(Dst.getReg()), E = MRI->use_end();
Use != E; Use = NextUse) {
NextUse = std::next(Use);
MachineInstr *UseMI = Use->getParent();
unsigned OpNo = Use.getOperandNo();
// Folding the immediate may reveal operations that can be constant
// folded or replaced with a copy. This can happen for example after
// frame indices are lowered to constants or from splitting 64-bit
// constants.
//
// We may also encounter cases where one or both operands are
// immediates materialized into a register, which would ordinarily not
// be folded due to multiple uses or operand constraints.
if (OpToFold.isImm() && tryConstantFoldOp(*MRI, TII, UseMI, &OpToFold)) {
DEBUG(dbgs() << "Constant folded " << *UseMI <<'\n');
// Some constant folding cases change the same immediate's use to a new
// instruction, e.g. and x, 0 -> 0. Make sure we re-visit the user
// again. The same constant folded instruction could also have a second
// use operand.
NextUse = MRI->use_begin(Dst.getReg());
continue;
}
// Try to fold any inline immediate uses, and then only fold other
// constants if they have one use.
//
// The legality of the inline immediate must be checked based on the use
// operand, not the defining instruction, because 32-bit instructions
// with 32-bit inline immediate sources may be used to materialize
// constants used in 16-bit operands.
//
// e.g. it is unsafe to fold:
// s_mov_b32 s0, 1.0 // materializes 0x3f800000
// v_add_f16 v0, v1, s0 // 1.0 f16 inline immediate sees 0x00003c00
// Folding immediates with more than one use will increase program size.
// FIXME: This will also reduce register usage, which may be better
// in some cases. A better heuristic is needed.
if (isInlineConstantIfFolded(TII, *UseMI, OpNo, OpToFold)) {
foldOperand(OpToFold, UseMI, OpNo, FoldList, CopiesToReplace);
} else {
if (++NumLiteralUses == 1) {
NonInlineUse = &*Use;
NonInlineUseOpNo = OpNo;
}
}
}
if (NumLiteralUses == 1) {
MachineInstr *UseMI = NonInlineUse->getParent();
foldOperand(OpToFold, UseMI, NonInlineUseOpNo, FoldList, CopiesToReplace);
}
} else {
// Folding register.
for (MachineRegisterInfo::use_iterator
Use = MRI->use_begin(Dst.getReg()), E = MRI->use_end();
Use != E; ++Use) {
MachineInstr *UseMI = Use->getParent();
foldOperand(OpToFold, UseMI, Use.getOperandNo(),
FoldList, CopiesToReplace);
}
}
MachineFunction *MF = MI.getParent()->getParent();
// Make sure we add EXEC uses to any new v_mov instructions created.
for (MachineInstr *Copy : CopiesToReplace)
Copy->addImplicitDefUseOperands(*MF);
for (FoldCandidate &Fold : FoldList) {
if (updateOperand(Fold, *TRI)) {
// Clear kill flags.
if (Fold.isReg()) {
assert(Fold.OpToFold && Fold.OpToFold->isReg());
// FIXME: Probably shouldn't bother trying to fold if not an
// SGPR. PeepholeOptimizer can eliminate redundant VGPR->VGPR
// copies.
MRI->clearKillFlags(Fold.OpToFold->getReg());
}
DEBUG(dbgs() << "Folded source from " << MI << " into OpNo " <<
static_cast<int>(Fold.UseOpNo) << " of " << *Fold.UseMI << '\n');
tryFoldInst(TII, Fold.UseMI);
} else if (Fold.isCommuted()) {
// Restoring instruction's original operand order if fold has failed.
TII->commuteInstruction(*Fold.UseMI, false);
}
}
}
void SIFoldOperands::foldOperand(
MachineOperand &OpToFold,
MachineInstr *UseMI,
unsigned UseOpIdx,
SmallVectorImpl<FoldCandidate> &FoldList,
SmallVectorImpl<MachineInstr *> &CopiesToReplace) const {
const MachineOperand &UseOp = UseMI->getOperand(UseOpIdx);
if (!isUseSafeToFold(TII, *UseMI, UseOp))
return;
// FIXME: Fold operands with subregs.
if (UseOp.isReg() && OpToFold.isReg()) {
if (UseOp.isImplicit() || UseOp.getSubReg() != AMDGPU::NoSubRegister)
return;
// Don't fold subregister extracts into tied operands, only if it is a full
// copy since a subregister use tied to a full register def doesn't really
// make sense. e.g. don't fold:
//
// %vreg1 = COPY %vreg0:sub1
// %vreg2<tied3> = V_MAC_{F16, F32} %vreg3, %vreg4, %vreg1<tied0>
//
// into
// %vreg2<tied3> = V_MAC_{F16, F32} %vreg3, %vreg4, %vreg0:sub1<tied0>
if (UseOp.isTied() && OpToFold.getSubReg() != AMDGPU::NoSubRegister)
return;
}
// Special case for REG_SEQUENCE: We can't fold literals into
// REG_SEQUENCE instructions, so we have to fold them into the
// uses of REG_SEQUENCE.
if (UseMI->isRegSequence()) {
unsigned RegSeqDstReg = UseMI->getOperand(0).getReg();
unsigned RegSeqDstSubReg = UseMI->getOperand(UseOpIdx + 1).getImm();
for (MachineRegisterInfo::use_iterator
RSUse = MRI->use_begin(RegSeqDstReg), RSE = MRI->use_end();
RSUse != RSE; ++RSUse) {
MachineInstr *RSUseMI = RSUse->getParent();
if (RSUse->getSubReg() != RegSeqDstSubReg)
continue;
foldOperand(OpToFold, RSUseMI, RSUse.getOperandNo(), FoldList,
CopiesToReplace);
}
return;
}
bool FoldingImm = OpToFold.isImm();
// In order to fold immediates into copies, we need to change the
// copy to a MOV.
if (FoldingImm && UseMI->isCopy()) {
unsigned DestReg = UseMI->getOperand(0).getReg();
const TargetRegisterClass *DestRC
= TargetRegisterInfo::isVirtualRegister(DestReg) ?
MRI->getRegClass(DestReg) :
TRI->getPhysRegClass(DestReg);
unsigned MovOp = TII->getMovOpcode(DestRC);
if (MovOp == AMDGPU::COPY)
return;
UseMI->setDesc(TII->get(MovOp));
CopiesToReplace.push_back(UseMI);
} else {
const MCInstrDesc &UseDesc = UseMI->getDesc();
// Don't fold into target independent nodes. Target independent opcodes
// don't have defined register classes.
if (UseDesc.isVariadic() ||
UseDesc.OpInfo[UseOpIdx].RegClass == -1)
return;
}
if (!FoldingImm) {
tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII);
// FIXME: We could try to change the instruction from 64-bit to 32-bit
// to enable more folding opportunites. The shrink operands pass
// already does this.
return;
}
const MCInstrDesc &FoldDesc = OpToFold.getParent()->getDesc();
const TargetRegisterClass *FoldRC =
TRI->getRegClass(FoldDesc.OpInfo[0].RegClass);
// Split 64-bit constants into 32-bits for folding.
if (UseOp.getSubReg() && AMDGPU::getRegBitWidth(FoldRC->getID()) == 64) {
unsigned UseReg = UseOp.getReg();
const TargetRegisterClass *UseRC
= TargetRegisterInfo::isVirtualRegister(UseReg) ?
MRI->getRegClass(UseReg) :
TRI->getPhysRegClass(UseReg);
if (AMDGPU::getRegBitWidth(UseRC->getID()) != 64)
return;
APInt Imm(64, OpToFold.getImm());
if (UseOp.getSubReg() == AMDGPU::sub0) {
Imm = Imm.getLoBits(32);
} else {
assert(UseOp.getSubReg() == AMDGPU::sub1);
Imm = Imm.getHiBits(32);
}
MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue());
tryAddToFoldList(FoldList, UseMI, UseOpIdx, &ImmOp, TII);
return;
}
tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII);
}
static bool isKUImmOperand(const SIInstrInfo *TII, const MachineOperand &Src) {
return isUInt<16>(Src.getImm()) &&
!TII->isInlineConstant(*Src.getParent(),
Src.getParent()->getOperandNo(&Src));
}
/// Handle the direct use of a physical register. Check that the register is
/// not used by a virtreg. Kill the physreg, marking it free. This may add
/// implicit kills to MO->getParent() and invalidate MO.
void RegAllocFast::usePhysReg(MachineOperand &MO) {
// Ignore undef uses.
if (MO.isUndef())
return;
unsigned PhysReg = MO.getReg();
assert(TargetRegisterInfo::isPhysicalRegister(PhysReg) &&
"Bad usePhysReg operand");
markRegUsedInInstr(PhysReg);
switch (PhysRegState[PhysReg]) {
case regDisabled:
break;
case regReserved:
PhysRegState[PhysReg] = regFree;
LLVM_FALLTHROUGH;
case regFree:
MO.setIsKill();
return;
default:
// The physreg was allocated to a virtual register. That means the value we
// wanted has been clobbered.
llvm_unreachable("Instruction uses an allocated register");
}
// Maybe a superregister is reserved?
for (MCRegAliasIterator AI(PhysReg, TRI, false); AI.isValid(); ++AI) {
MCPhysReg Alias = *AI;
switch (PhysRegState[Alias]) {
case regDisabled:
break;
case regReserved:
// Either PhysReg is a subregister of Alias and we mark the
// whole register as free, or PhysReg is the superregister of
// Alias and we mark all the aliases as disabled before freeing
// PhysReg.
// In the latter case, since PhysReg was disabled, this means that
// its value is defined only by physical sub-registers. This check
// is performed by the assert of the default case in this loop.
// Note: The value of the superregister may only be partial
// defined, that is why regDisabled is a valid state for aliases.
assert((TRI->isSuperRegister(PhysReg, Alias) ||
TRI->isSuperRegister(Alias, PhysReg)) &&
"Instruction is not using a subregister of a reserved register");
LLVM_FALLTHROUGH;
case regFree:
if (TRI->isSuperRegister(PhysReg, Alias)) {
// Leave the superregister in the working set.
setPhysRegState(Alias, regFree);
MO.getParent()->addRegisterKilled(Alias, TRI, true);
return;
}
// Some other alias was in the working set - clear it.
setPhysRegState(Alias, regDisabled);
break;
default:
llvm_unreachable("Instruction uses an alias of an allocated register");
}
}
// All aliases are disabled, bring register into working set.
setPhysRegState(PhysReg, regFree);
MO.setIsKill();
}
bool StrongPHIElimination::runOnMachineFunction(MachineFunction &MF) {
MRI = &MF.getRegInfo();
TII = MF.getTarget().getInstrInfo();
DT = &getAnalysis<MachineDominatorTree>();
LI = &getAnalysis<LiveIntervals>();
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
for (MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
unsigned DestReg = BBI->getOperand(0).getReg();
addReg(DestReg);
PHISrcDefs[I].push_back(BBI);
for (unsigned i = 1; i < BBI->getNumOperands(); i += 2) {
MachineOperand &SrcMO = BBI->getOperand(i);
unsigned SrcReg = SrcMO.getReg();
addReg(SrcReg);
unionRegs(DestReg, SrcReg);
MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
if (DefMI)
PHISrcDefs[DefMI->getParent()].push_back(DefMI);
}
}
}
// Perform a depth-first traversal of the dominator tree, splitting
// interferences amongst PHI-congruence classes.
DenseMap<unsigned, unsigned> CurrentDominatingParent;
DenseMap<unsigned, unsigned> ImmediateDominatingParent;
for (df_iterator<MachineDomTreeNode*> DI = df_begin(DT->getRootNode()),
DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
SplitInterferencesForBasicBlock(*DI->getBlock(),
CurrentDominatingParent,
ImmediateDominatingParent);
}
// Insert copies for all PHI source and destination registers.
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
for (MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
InsertCopiesForPHI(BBI, I);
}
}
// FIXME: Preserve the equivalence classes during copy insertion and use
// the preversed equivalence classes instead of recomputing them.
RegNodeMap.clear();
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
for (MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
unsigned DestReg = BBI->getOperand(0).getReg();
addReg(DestReg);
for (unsigned i = 1; i < BBI->getNumOperands(); i += 2) {
unsigned SrcReg = BBI->getOperand(i).getReg();
addReg(SrcReg);
unionRegs(DestReg, SrcReg);
}
}
}
DenseMap<unsigned, unsigned> RegRenamingMap;
bool Changed = false;
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
while (BBI != BBE && BBI->isPHI()) {
MachineInstr *PHI = BBI;
assert(PHI->getNumOperands() > 0);
unsigned SrcReg = PHI->getOperand(1).getReg();
unsigned SrcColor = getRegColor(SrcReg);
unsigned NewReg = RegRenamingMap[SrcColor];
if (!NewReg) {
NewReg = SrcReg;
RegRenamingMap[SrcColor] = SrcReg;
}
MergeLIsAndRename(SrcReg, NewReg);
unsigned DestReg = PHI->getOperand(0).getReg();
if (!InsertedDestCopies.count(DestReg))
MergeLIsAndRename(DestReg, NewReg);
for (unsigned i = 3; i < PHI->getNumOperands(); i += 2) {
unsigned SrcReg = PHI->getOperand(i).getReg();
MergeLIsAndRename(SrcReg, NewReg);
}
++BBI;
LI->RemoveMachineInstrFromMaps(PHI);
PHI->eraseFromParent();
Changed = true;
}
}
// Due to the insertion of copies to split live ranges, the live intervals are
// guaranteed to not overlap, except in one case: an original PHI source and a
// PHI destination copy. In this case, they have the same value and thus don't
// truly intersect, so we merge them into the value live at that point.
// FIXME: Is there some better way we can handle this?
for (DestCopyMap::iterator I = InsertedDestCopies.begin(),
E = InsertedDestCopies.end(); I != E; ++I) {
unsigned DestReg = I->first;
unsigned DestColor = getRegColor(DestReg);
unsigned NewReg = RegRenamingMap[DestColor];
LiveInterval &DestLI = LI->getInterval(DestReg);
LiveInterval &NewLI = LI->getInterval(NewReg);
assert(DestLI.ranges.size() == 1
&& "PHI destination copy's live interval should be a single live "
"range from the beginning of the BB to the copy instruction.");
LiveRange *DestLR = DestLI.begin();
VNInfo *NewVNI = NewLI.getVNInfoAt(DestLR->start);
if (!NewVNI) {
NewVNI = NewLI.createValueCopy(DestLR->valno, LI->getVNInfoAllocator());
MachineInstr *CopyInstr = I->second;
CopyInstr->getOperand(1).setIsKill(true);
}
LiveRange NewLR(DestLR->start, DestLR->end, NewVNI);
NewLI.addRange(NewLR);
LI->removeInterval(DestReg);
MRI->replaceRegWith(DestReg, NewReg);
}
// Adjust the live intervals of all PHI source registers to handle the case
// where the PHIs in successor blocks were the only later uses of the source
// register.
for (SrcCopySet::iterator I = InsertedSrcCopySet.begin(),
E = InsertedSrcCopySet.end(); I != E; ++I) {
MachineBasicBlock *MBB = I->first;
unsigned SrcReg = I->second;
if (unsigned RenamedRegister = RegRenamingMap[getRegColor(SrcReg)])
SrcReg = RenamedRegister;
LiveInterval &SrcLI = LI->getInterval(SrcReg);
bool isLiveOut = false;
for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
SE = MBB->succ_end(); SI != SE; ++SI) {
if (SrcLI.liveAt(LI->getMBBStartIdx(*SI))) {
isLiveOut = true;
break;
}
}
if (isLiveOut)
continue;
MachineOperand *LastUse = findLastUse(MBB, SrcReg);
assert(LastUse);
SlotIndex LastUseIndex = LI->getInstructionIndex(LastUse->getParent());
SrcLI.removeRange(LastUseIndex.getRegSlot(), LI->getMBBEndIdx(MBB));
LastUse->setIsKill(true);
}
Allocator.Reset();
RegNodeMap.clear();
PHISrcDefs.clear();
InsertedSrcCopySet.clear();
InsertedSrcCopyMap.clear();
InsertedDestCopies.clear();
return Changed;
}
void MIPrinter::print(const MachineOperand &Op, const TargetRegisterInfo *TRI) {
printTargetFlags(Op);
switch (Op.getType()) {
case MachineOperand::MO_Register:
// TODO: Print the other register flags.
if (Op.isImplicit())
OS << (Op.isDef() ? "implicit-def " : "implicit ");
if (Op.isDead())
OS << "dead ";
if (Op.isKill())
OS << "killed ";
if (Op.isUndef())
OS << "undef ";
if (Op.isEarlyClobber())
OS << "early-clobber ";
if (Op.isDebug())
OS << "debug-use ";
printReg(Op.getReg(), OS, TRI);
// Print the sub register.
if (Op.getSubReg() != 0)
OS << ':' << TRI->getSubRegIndexName(Op.getSubReg());
break;
case MachineOperand::MO_Immediate:
OS << Op.getImm();
break;
case MachineOperand::MO_CImmediate:
Op.getCImm()->printAsOperand(OS, /*PrintType=*/true, MST);
break;
case MachineOperand::MO_FPImmediate:
Op.getFPImm()->printAsOperand(OS, /*PrintType=*/true, MST);
break;
case MachineOperand::MO_MachineBasicBlock:
printMBBReference(*Op.getMBB());
break;
case MachineOperand::MO_FrameIndex:
printStackObjectReference(Op.getIndex());
break;
case MachineOperand::MO_ConstantPoolIndex:
OS << "%const." << Op.getIndex();
printOffset(Op.getOffset());
break;
case MachineOperand::MO_TargetIndex: {
OS << "target-index(";
if (const auto *Name = getTargetIndexName(
*Op.getParent()->getParent()->getParent(), Op.getIndex()))
OS << Name;
else
OS << "<unknown>";
OS << ')';
printOffset(Op.getOffset());
break;
}
case MachineOperand::MO_JumpTableIndex:
OS << "%jump-table." << Op.getIndex();
break;
case MachineOperand::MO_ExternalSymbol:
OS << '$';
printLLVMNameWithoutPrefix(OS, Op.getSymbolName());
printOffset(Op.getOffset());
break;
case MachineOperand::MO_GlobalAddress:
Op.getGlobal()->printAsOperand(OS, /*PrintType=*/false, MST);
printOffset(Op.getOffset());
break;
case MachineOperand::MO_BlockAddress:
OS << "blockaddress(";
Op.getBlockAddress()->getFunction()->printAsOperand(OS, /*PrintType=*/false,
MST);
OS << ", ";
printIRBlockReference(*Op.getBlockAddress()->getBasicBlock());
OS << ')';
printOffset(Op.getOffset());
break;
case MachineOperand::MO_RegisterMask: {
auto RegMaskInfo = RegisterMaskIds.find(Op.getRegMask());
if (RegMaskInfo != RegisterMaskIds.end())
OS << StringRef(TRI->getRegMaskNames()[RegMaskInfo->second]).lower();
else
llvm_unreachable("Can't print this machine register mask yet.");
break;
}
case MachineOperand::MO_Metadata:
Op.getMetadata()->printAsOperand(OS, MST);
break;
case MachineOperand::MO_CFIIndex: {
const auto &MMI = Op.getParent()->getParent()->getParent()->getMMI();
print(MMI.getFrameInstructions()[Op.getCFIIndex()], TRI);
break;
}
default:
// TODO: Print the other machine operands.
llvm_unreachable("Can't print this machine operand at the moment");
}
}
// MI is known to be dead. Figure out what instructions
// are also made dead by this and mark them for removal.
void A15SDOptimizer::eraseInstrWithNoUses(MachineInstr *MI) {
SmallVector<MachineInstr *, 8> Front;
DeadInstr.insert(MI);
DEBUG(dbgs() << "Deleting base instruction " << *MI << "\n");
Front.push_back(MI);
while (Front.size() != 0) {
MI = Front.back();
Front.pop_back();
// MI is already known to be dead. We need to see
// if other instructions can also be removed.
for (unsigned int i = 0; i < MI->getNumOperands(); ++i) {
MachineOperand &MO = MI->getOperand(i);
if ((!MO.isReg()) || (!MO.isUse()))
continue;
unsigned Reg = MO.getReg();
if (!TRI->isVirtualRegister(Reg))
continue;
MachineOperand *Op = MI->findRegisterDefOperand(Reg);
if (!Op)
continue;
MachineInstr *Def = Op->getParent();
// We don't need to do anything if we have already marked
// this instruction as being dead.
if (DeadInstr.find(Def) != DeadInstr.end())
continue;
// Check if all the uses of this instruction are marked as
// dead. If so, we can also mark this instruction as being
// dead.
bool IsDead = true;
for (unsigned int j = 0; j < Def->getNumOperands(); ++j) {
MachineOperand &MODef = Def->getOperand(j);
if ((!MODef.isReg()) || (!MODef.isDef()))
continue;
unsigned DefReg = MODef.getReg();
if (!TRI->isVirtualRegister(DefReg)) {
IsDead = false;
break;
}
for (MachineRegisterInfo::use_instr_iterator
II = MRI->use_instr_begin(Reg), EE = MRI->use_instr_end();
II != EE; ++II) {
// We don't care about self references.
if (&*II == Def)
continue;
if (DeadInstr.find(&*II) == DeadInstr.end()) {
IsDead = false;
break;
}
}
}
if (!IsDead) continue;
DEBUG(dbgs() << "Deleting instruction " << *Def << "\n");
DeadInstr.insert(Def);
}
}
}
/// OptimizeInstr - If instruction is a copy-like instruction, i.e. it reads
/// a single register and writes a single register and it does not modify
/// the source, and if the source value is preserved as a sub-register of
/// the result, then replace all reachable uses of the source with the subreg
/// of the result.
/// Do not generate an EXTRACT that is used only in a debug use, as this
/// changes the code. Since this code does not currently share EXTRACTs, just
/// ignore all debug uses.
bool OptimizeExts::OptimizeInstr(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
bool Changed = false;
LocalMIs.insert(MI);
unsigned SrcReg, DstReg, SubIdx;
if (TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx)) {
if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
TargetRegisterInfo::isPhysicalRegister(SrcReg))
return false;
MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(SrcReg);
if (++UI == MRI->use_nodbg_end())
// No other uses.
return false;
// Ok, the source has other uses. See if we can replace the other uses
// with use of the result of the extension.
SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
UI = MRI->use_nodbg_begin(DstReg);
for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
UI != UE; ++UI)
ReachedBBs.insert(UI->getParent());
bool ExtendLife = true;
// Uses that are in the same BB of uses of the result of the instruction.
SmallVector<MachineOperand*, 8> Uses;
// Uses that the result of the instruction can reach.
SmallVector<MachineOperand*, 8> ExtendedUses;
UI = MRI->use_nodbg_begin(SrcReg);
for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
UI != UE; ++UI) {
MachineOperand &UseMO = UI.getOperand();
MachineInstr *UseMI = &*UI;
if (UseMI == MI)
continue;
if (UseMI->isPHI()) {
ExtendLife = false;
continue;
}
// It's an error to translate this:
//
// %reg1025 = <sext> %reg1024
// ...
// %reg1026 = SUBREG_TO_REG 0, %reg1024, 4
//
// into this:
//
// %reg1025 = <sext> %reg1024
// ...
// %reg1027 = COPY %reg1025:4
// %reg1026 = SUBREG_TO_REG 0, %reg1027, 4
//
// The problem here is that SUBREG_TO_REG is there to assert that an
// implicit zext occurs. It doesn't insert a zext instruction. If we allow
// the COPY here, it will give us the value after the <sext>,
// not the original value of %reg1024 before <sext>.
if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG)
continue;
MachineBasicBlock *UseMBB = UseMI->getParent();
if (UseMBB == MBB) {
// Local uses that come after the extension.
if (!LocalMIs.count(UseMI))
Uses.push_back(&UseMO);
} else if (ReachedBBs.count(UseMBB))
// Non-local uses where the result of extension is used. Always
// replace these unless it's a PHI.
Uses.push_back(&UseMO);
else if (Aggressive && DT->dominates(MBB, UseMBB))
// We may want to extend live range of the extension result in order
// to replace these uses.
ExtendedUses.push_back(&UseMO);
else {
// Both will be live out of the def MBB anyway. Don't extend live
// range of the extension result.
ExtendLife = false;
break;
}
}
if (ExtendLife && !ExtendedUses.empty())
// Ok, we'll extend the liveness of the extension result.
std::copy(ExtendedUses.begin(), ExtendedUses.end(),
std::back_inserter(Uses));
// Now replace all uses.
if (!Uses.empty()) {
SmallPtrSet<MachineBasicBlock*, 4> PHIBBs;
// Look for PHI uses of the extended result, we don't want to extend the
// liveness of a PHI input. It breaks all kinds of assumptions down
// stream. A PHI use is expected to be the kill of its source values.
UI = MRI->use_nodbg_begin(DstReg);
for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
UI != UE; ++UI)
if (UI->isPHI())
PHIBBs.insert(UI->getParent());
const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
MachineOperand *UseMO = Uses[i];
MachineInstr *UseMI = UseMO->getParent();
MachineBasicBlock *UseMBB = UseMI->getParent();
if (PHIBBs.count(UseMBB))
continue;
unsigned NewVR = MRI->createVirtualRegister(RC);
BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
TII->get(TargetOpcode::COPY), NewVR)
.addReg(DstReg, 0, SubIdx);
UseMO->setReg(NewVR);
++NumReuse;
Changed = true;
}
}
}
return Changed;
}
static MCOperand GetSymbolRef(const MachineOperand &MO, const MCSymbol *Symbol,
AsmPrinter &Printer, bool isDarwin) {
MCContext &Ctx = Printer.OutContext;
MCSymbolRefExpr::VariantKind RefKind = MCSymbolRefExpr::VK_None;
unsigned access = MO.getTargetFlags() & PPCII::MO_ACCESS_MASK;
switch (access) {
case PPCII::MO_TPREL_LO:
RefKind = MCSymbolRefExpr::VK_PPC_TPREL_LO;
break;
case PPCII::MO_TPREL_HA:
RefKind = MCSymbolRefExpr::VK_PPC_TPREL_HA;
break;
case PPCII::MO_DTPREL_LO:
RefKind = MCSymbolRefExpr::VK_PPC_DTPREL_LO;
break;
case PPCII::MO_TLSLD_LO:
RefKind = MCSymbolRefExpr::VK_PPC_GOT_TLSLD_LO;
break;
case PPCII::MO_TOC_LO:
RefKind = MCSymbolRefExpr::VK_PPC_TOC_LO;
break;
case PPCII::MO_TLS:
RefKind = MCSymbolRefExpr::VK_PPC_TLS;
break;
case PPCII::MO_TLSGD:
RefKind = MCSymbolRefExpr::VK_PPC_TLSGD;
break;
case PPCII::MO_TLSLD:
RefKind = MCSymbolRefExpr::VK_PPC_TLSLD;
break;
}
if (MO.getTargetFlags() == PPCII::MO_PLT_OR_STUB && !isDarwin)
RefKind = MCSymbolRefExpr::VK_PLT;
const MCExpr *Expr = MCSymbolRefExpr::Create(Symbol, RefKind, Ctx);
if (!MO.isJTI() && MO.getOffset())
Expr = MCBinaryExpr::CreateAdd(Expr,
MCConstantExpr::Create(MO.getOffset(), Ctx),
Ctx);
// Subtract off the PIC base if required.
if (MO.getTargetFlags() & PPCII::MO_PIC_FLAG) {
const MachineFunction *MF = MO.getParent()->getParent()->getParent();
const MCExpr *PB = MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
Expr = MCBinaryExpr::CreateSub(Expr, PB, Ctx);
}
// Add ha16() / lo16() markers if required.
switch (access) {
case PPCII::MO_LO:
Expr = PPCMCExpr::CreateLo(Expr, isDarwin, Ctx);
break;
case PPCII::MO_HA:
Expr = PPCMCExpr::CreateHa(Expr, isDarwin, Ctx);
break;
}
return MCOperand::CreateExpr(Expr);
}
void MIPrinter::print(const MachineOperand &Op, const TargetRegisterInfo *TRI,
unsigned I, bool ShouldPrintRegisterTies, bool IsDef) {
printTargetFlags(Op);
switch (Op.getType()) {
case MachineOperand::MO_Register:
if (Op.isImplicit())
OS << (Op.isDef() ? "implicit-def " : "implicit ");
else if (!IsDef && Op.isDef())
// Print the 'def' flag only when the operand is defined after '='.
OS << "def ";
if (Op.isInternalRead())
OS << "internal ";
if (Op.isDead())
OS << "dead ";
if (Op.isKill())
OS << "killed ";
if (Op.isUndef())
OS << "undef ";
if (Op.isEarlyClobber())
OS << "early-clobber ";
if (Op.isDebug())
OS << "debug-use ";
printReg(Op.getReg(), OS, TRI);
// Print the sub register.
if (Op.getSubReg() != 0)
OS << ':' << TRI->getSubRegIndexName(Op.getSubReg());
if (ShouldPrintRegisterTies && Op.isTied() && !Op.isDef())
OS << "(tied-def " << Op.getParent()->findTiedOperandIdx(I) << ")";
break;
case MachineOperand::MO_Immediate:
OS << Op.getImm();
break;
case MachineOperand::MO_CImmediate:
Op.getCImm()->printAsOperand(OS, /*PrintType=*/true, MST);
break;
case MachineOperand::MO_FPImmediate:
Op.getFPImm()->printAsOperand(OS, /*PrintType=*/true, MST);
break;
case MachineOperand::MO_MachineBasicBlock:
printMBBReference(*Op.getMBB());
break;
case MachineOperand::MO_FrameIndex:
printStackObjectReference(Op.getIndex());
break;
case MachineOperand::MO_ConstantPoolIndex:
OS << "%const." << Op.getIndex();
printOffset(Op.getOffset());
break;
case MachineOperand::MO_TargetIndex: {
OS << "target-index(";
if (const auto *Name = getTargetIndexName(
*Op.getParent()->getParent()->getParent(), Op.getIndex()))
OS << Name;
else
OS << "<unknown>";
OS << ')';
printOffset(Op.getOffset());
break;
}
case MachineOperand::MO_JumpTableIndex:
OS << "%jump-table." << Op.getIndex();
break;
case MachineOperand::MO_ExternalSymbol:
OS << '$';
printLLVMNameWithoutPrefix(OS, Op.getSymbolName());
printOffset(Op.getOffset());
break;
case MachineOperand::MO_GlobalAddress:
Op.getGlobal()->printAsOperand(OS, /*PrintType=*/false, MST);
printOffset(Op.getOffset());
break;
case MachineOperand::MO_BlockAddress:
OS << "blockaddress(";
Op.getBlockAddress()->getFunction()->printAsOperand(OS, /*PrintType=*/false,
MST);
OS << ", ";
printIRBlockReference(*Op.getBlockAddress()->getBasicBlock());
OS << ')';
printOffset(Op.getOffset());
break;
case MachineOperand::MO_RegisterMask: {
auto RegMaskInfo = RegisterMaskIds.find(Op.getRegMask());
if (RegMaskInfo != RegisterMaskIds.end())
OS << StringRef(TRI->getRegMaskNames()[RegMaskInfo->second]).lower();
else
llvm_unreachable("Can't print this machine register mask yet.");
break;
}
case MachineOperand::MO_RegisterLiveOut: {
const uint32_t *RegMask = Op.getRegLiveOut();
OS << "liveout(";
bool IsCommaNeeded = false;
for (unsigned Reg = 0, E = TRI->getNumRegs(); Reg < E; ++Reg) {
if (RegMask[Reg / 32] & (1U << (Reg % 32))) {
if (IsCommaNeeded)
OS << ", ";
printReg(Reg, OS, TRI);
IsCommaNeeded = true;
}
}
OS << ")";
break;
}
case MachineOperand::MO_Metadata:
Op.getMetadata()->printAsOperand(OS, MST);
break;
case MachineOperand::MO_MCSymbol:
OS << "<mcsymbol " << *Op.getMCSymbol() << ">";
break;
case MachineOperand::MO_CFIIndex: {
const auto &MMI = Op.getParent()->getParent()->getParent()->getMMI();
print(MMI.getFrameInstructions()[Op.getCFIIndex()], TRI);
break;
}
}
}
void MIPrinter::print(const MachineOperand &Op, const TargetRegisterInfo *TRI,
unsigned I, bool ShouldPrintRegisterTies, LLT TypeToPrint,
bool IsDef) {
printTargetFlags(Op);
switch (Op.getType()) {
case MachineOperand::MO_Register:
if (Op.isImplicit())
OS << (Op.isDef() ? "implicit-def " : "implicit ");
else if (!IsDef && Op.isDef())
// Print the 'def' flag only when the operand is defined after '='.
OS << "def ";
if (Op.isInternalRead())
OS << "internal ";
if (Op.isDead())
OS << "dead ";
if (Op.isKill())
OS << "killed ";
if (Op.isUndef())
OS << "undef ";
if (Op.isEarlyClobber())
OS << "early-clobber ";
if (Op.isDebug())
OS << "debug-use ";
printReg(Op.getReg(), OS, TRI);
// Print the sub register.
if (Op.getSubReg() != 0)
OS << '.' << TRI->getSubRegIndexName(Op.getSubReg());
if (ShouldPrintRegisterTies && Op.isTied() && !Op.isDef())
OS << "(tied-def " << Op.getParent()->findTiedOperandIdx(I) << ")";
if (TypeToPrint.isValid())
OS << '(' << TypeToPrint << ')';
break;
case MachineOperand::MO_Immediate:
OS << Op.getImm();
break;
case MachineOperand::MO_CImmediate:
Op.getCImm()->printAsOperand(OS, /*PrintType=*/true, MST);
break;
case MachineOperand::MO_FPImmediate:
Op.getFPImm()->printAsOperand(OS, /*PrintType=*/true, MST);
break;
case MachineOperand::MO_MachineBasicBlock:
printMBBReference(*Op.getMBB());
break;
case MachineOperand::MO_FrameIndex:
printStackObjectReference(Op.getIndex());
break;
case MachineOperand::MO_ConstantPoolIndex:
OS << "%const." << Op.getIndex();
printOffset(Op.getOffset());
break;
case MachineOperand::MO_TargetIndex:
OS << "target-index(";
if (const auto *Name = getTargetIndexName(
*Op.getParent()->getParent()->getParent(), Op.getIndex()))
OS << Name;
else
OS << "<unknown>";
OS << ')';
printOffset(Op.getOffset());
break;
case MachineOperand::MO_JumpTableIndex:
OS << "%jump-table." << Op.getIndex();
break;
case MachineOperand::MO_ExternalSymbol: {
StringRef Name = Op.getSymbolName();
OS << '$';
if (Name.empty()) {
OS << "\"\"";
} else {
printLLVMNameWithoutPrefix(OS, Name);
}
printOffset(Op.getOffset());
break;
}
case MachineOperand::MO_GlobalAddress:
Op.getGlobal()->printAsOperand(OS, /*PrintType=*/false, MST);
printOffset(Op.getOffset());
break;
case MachineOperand::MO_BlockAddress:
OS << "blockaddress(";
Op.getBlockAddress()->getFunction()->printAsOperand(OS, /*PrintType=*/false,
MST);
OS << ", ";
printIRBlockReference(*Op.getBlockAddress()->getBasicBlock());
OS << ')';
printOffset(Op.getOffset());
break;
case MachineOperand::MO_RegisterMask: {
auto RegMaskInfo = RegisterMaskIds.find(Op.getRegMask());
if (RegMaskInfo != RegisterMaskIds.end())
OS << StringRef(TRI->getRegMaskNames()[RegMaskInfo->second]).lower();
else
printCustomRegMask(Op.getRegMask(), OS, TRI);
break;
}
case MachineOperand::MO_RegisterLiveOut: {
const uint32_t *RegMask = Op.getRegLiveOut();
OS << "liveout(";
bool IsCommaNeeded = false;
for (unsigned Reg = 0, E = TRI->getNumRegs(); Reg < E; ++Reg) {
if (RegMask[Reg / 32] & (1U << (Reg % 32))) {
if (IsCommaNeeded)
OS << ", ";
printReg(Reg, OS, TRI);
IsCommaNeeded = true;
}
}
OS << ")";
break;
}
case MachineOperand::MO_Metadata:
Op.getMetadata()->printAsOperand(OS, MST);
break;
case MachineOperand::MO_MCSymbol:
OS << "<mcsymbol " << *Op.getMCSymbol() << ">";
break;
case MachineOperand::MO_CFIIndex: {
const MachineFunction &MF = *Op.getParent()->getParent()->getParent();
print(MF.getFrameInstructions()[Op.getCFIIndex()], TRI);
break;
}
case MachineOperand::MO_IntrinsicID: {
Intrinsic::ID ID = Op.getIntrinsicID();
if (ID < Intrinsic::num_intrinsics)
OS << "intrinsic(@" << Intrinsic::getName(ID, None) << ')';
else {
const MachineFunction &MF = *Op.getParent()->getParent()->getParent();
const TargetIntrinsicInfo *TII = MF.getTarget().getIntrinsicInfo();
OS << "intrinsic(@" << TII->getName(ID) << ')';
}
break;
}
case MachineOperand::MO_Predicate: {
auto Pred = static_cast<CmpInst::Predicate>(Op.getPredicate());
OS << (CmpInst::isIntPredicate(Pred) ? "int" : "float") << "pred("
<< CmpInst::getPredicateName(Pred) << ')';
break;
}
}
}